SemaChecking.cpp revision 1497bffe829d082dd1d1927dc80ea08dcf1fcb49
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/Analysis/CFG.h"
17#include "clang/Analysis/AnalysisContext.h"
18#include "clang/Analysis/Analyses/PrintfFormatString.h"
19#include "clang/AST/ASTContext.h"
20#include "clang/AST/CharUnits.h"
21#include "clang/AST/DeclObjC.h"
22#include "clang/AST/ExprCXX.h"
23#include "clang/AST/ExprObjC.h"
24#include "clang/AST/DeclObjC.h"
25#include "clang/AST/StmtCXX.h"
26#include "clang/AST/StmtObjC.h"
27#include "clang/Lex/LiteralSupport.h"
28#include "clang/Lex/Preprocessor.h"
29#include "llvm/ADT/BitVector.h"
30#include "llvm/ADT/STLExtras.h"
31#include <limits>
32#include <queue>
33using namespace clang;
34
35/// getLocationOfStringLiteralByte - Return a source location that points to the
36/// specified byte of the specified string literal.
37///
38/// Strings are amazingly complex.  They can be formed from multiple tokens and
39/// can have escape sequences in them in addition to the usual trigraph and
40/// escaped newline business.  This routine handles this complexity.
41///
42SourceLocation Sema::getLocationOfStringLiteralByte(const StringLiteral *SL,
43                                                    unsigned ByteNo) const {
44  assert(!SL->isWide() && "This doesn't work for wide strings yet");
45
46  // Loop over all of the tokens in this string until we find the one that
47  // contains the byte we're looking for.
48  unsigned TokNo = 0;
49  while (1) {
50    assert(TokNo < SL->getNumConcatenated() && "Invalid byte number!");
51    SourceLocation StrTokLoc = SL->getStrTokenLoc(TokNo);
52
53    // Get the spelling of the string so that we can get the data that makes up
54    // the string literal, not the identifier for the macro it is potentially
55    // expanded through.
56    SourceLocation StrTokSpellingLoc = SourceMgr.getSpellingLoc(StrTokLoc);
57
58    // Re-lex the token to get its length and original spelling.
59    std::pair<FileID, unsigned> LocInfo =
60      SourceMgr.getDecomposedLoc(StrTokSpellingLoc);
61    std::pair<const char *,const char *> Buffer =
62      SourceMgr.getBufferData(LocInfo.first);
63    const char *StrData = Buffer.first+LocInfo.second;
64
65    // Create a langops struct and enable trigraphs.  This is sufficient for
66    // relexing tokens.
67    LangOptions LangOpts;
68    LangOpts.Trigraphs = true;
69
70    // Create a lexer starting at the beginning of this token.
71    Lexer TheLexer(StrTokSpellingLoc, LangOpts, Buffer.first, StrData,
72                   Buffer.second);
73    Token TheTok;
74    TheLexer.LexFromRawLexer(TheTok);
75
76    // Use the StringLiteralParser to compute the length of the string in bytes.
77    StringLiteralParser SLP(&TheTok, 1, PP);
78    unsigned TokNumBytes = SLP.GetStringLength();
79
80    // If the byte is in this token, return the location of the byte.
81    if (ByteNo < TokNumBytes ||
82        (ByteNo == TokNumBytes && TokNo == SL->getNumConcatenated())) {
83      unsigned Offset =
84        StringLiteralParser::getOffsetOfStringByte(TheTok, ByteNo, PP);
85
86      // Now that we know the offset of the token in the spelling, use the
87      // preprocessor to get the offset in the original source.
88      return PP.AdvanceToTokenCharacter(StrTokLoc, Offset);
89    }
90
91    // Move to the next string token.
92    ++TokNo;
93    ByteNo -= TokNumBytes;
94  }
95}
96
97/// CheckablePrintfAttr - does a function call have a "printf" attribute
98/// and arguments that merit checking?
99bool Sema::CheckablePrintfAttr(const FormatAttr *Format, CallExpr *TheCall) {
100  if (Format->getType() == "printf") return true;
101  if (Format->getType() == "printf0") {
102    // printf0 allows null "format" string; if so don't check format/args
103    unsigned format_idx = Format->getFormatIdx() - 1;
104    // Does the index refer to the implicit object argument?
105    if (isa<CXXMemberCallExpr>(TheCall)) {
106      if (format_idx == 0)
107        return false;
108      --format_idx;
109    }
110    if (format_idx < TheCall->getNumArgs()) {
111      Expr *Format = TheCall->getArg(format_idx)->IgnoreParenCasts();
112      if (!Format->isNullPointerConstant(Context, Expr::NPC_ValueDependentIsNull))
113        return true;
114    }
115  }
116  return false;
117}
118
119Action::OwningExprResult
120Sema::CheckBuiltinFunctionCall(unsigned BuiltinID, CallExpr *TheCall) {
121  OwningExprResult TheCallResult(Owned(TheCall));
122
123  switch (BuiltinID) {
124  case Builtin::BI__builtin___CFStringMakeConstantString:
125    assert(TheCall->getNumArgs() == 1 &&
126           "Wrong # arguments to builtin CFStringMakeConstantString");
127    if (CheckObjCString(TheCall->getArg(0)))
128      return ExprError();
129    break;
130  case Builtin::BI__builtin_stdarg_start:
131  case Builtin::BI__builtin_va_start:
132    if (SemaBuiltinVAStart(TheCall))
133      return ExprError();
134    break;
135  case Builtin::BI__builtin_isgreater:
136  case Builtin::BI__builtin_isgreaterequal:
137  case Builtin::BI__builtin_isless:
138  case Builtin::BI__builtin_islessequal:
139  case Builtin::BI__builtin_islessgreater:
140  case Builtin::BI__builtin_isunordered:
141    if (SemaBuiltinUnorderedCompare(TheCall))
142      return ExprError();
143    break;
144  case Builtin::BI__builtin_isfinite:
145  case Builtin::BI__builtin_isinf:
146  case Builtin::BI__builtin_isinf_sign:
147  case Builtin::BI__builtin_isnan:
148  case Builtin::BI__builtin_isnormal:
149    if (SemaBuiltinUnaryFP(TheCall))
150      return ExprError();
151    break;
152  case Builtin::BI__builtin_return_address:
153  case Builtin::BI__builtin_frame_address:
154    if (SemaBuiltinStackAddress(TheCall))
155      return ExprError();
156    break;
157  case Builtin::BI__builtin_eh_return_data_regno:
158    if (SemaBuiltinEHReturnDataRegNo(TheCall))
159      return ExprError();
160    break;
161  case Builtin::BI__builtin_shufflevector:
162    return SemaBuiltinShuffleVector(TheCall);
163    // TheCall will be freed by the smart pointer here, but that's fine, since
164    // SemaBuiltinShuffleVector guts it, but then doesn't release it.
165  case Builtin::BI__builtin_prefetch:
166    if (SemaBuiltinPrefetch(TheCall))
167      return ExprError();
168    break;
169  case Builtin::BI__builtin_object_size:
170    if (SemaBuiltinObjectSize(TheCall))
171      return ExprError();
172    break;
173  case Builtin::BI__builtin_longjmp:
174    if (SemaBuiltinLongjmp(TheCall))
175      return ExprError();
176    break;
177  case Builtin::BI__sync_fetch_and_add:
178  case Builtin::BI__sync_fetch_and_sub:
179  case Builtin::BI__sync_fetch_and_or:
180  case Builtin::BI__sync_fetch_and_and:
181  case Builtin::BI__sync_fetch_and_xor:
182  case Builtin::BI__sync_fetch_and_nand:
183  case Builtin::BI__sync_add_and_fetch:
184  case Builtin::BI__sync_sub_and_fetch:
185  case Builtin::BI__sync_and_and_fetch:
186  case Builtin::BI__sync_or_and_fetch:
187  case Builtin::BI__sync_xor_and_fetch:
188  case Builtin::BI__sync_nand_and_fetch:
189  case Builtin::BI__sync_val_compare_and_swap:
190  case Builtin::BI__sync_bool_compare_and_swap:
191  case Builtin::BI__sync_lock_test_and_set:
192  case Builtin::BI__sync_lock_release:
193    if (SemaBuiltinAtomicOverloaded(TheCall))
194      return ExprError();
195    break;
196  }
197
198  return move(TheCallResult);
199}
200
201/// CheckFunctionCall - Check a direct function call for various correctness
202/// and safety properties not strictly enforced by the C type system.
203bool Sema::CheckFunctionCall(FunctionDecl *FDecl, CallExpr *TheCall) {
204  // Get the IdentifierInfo* for the called function.
205  IdentifierInfo *FnInfo = FDecl->getIdentifier();
206
207  // None of the checks below are needed for functions that don't have
208  // simple names (e.g., C++ conversion functions).
209  if (!FnInfo)
210    return false;
211
212  // FIXME: This mechanism should be abstracted to be less fragile and
213  // more efficient. For example, just map function ids to custom
214  // handlers.
215
216  // Printf checking.
217  if (const FormatAttr *Format = FDecl->getAttr<FormatAttr>()) {
218    if (CheckablePrintfAttr(Format, TheCall)) {
219      bool HasVAListArg = Format->getFirstArg() == 0;
220      if (!HasVAListArg) {
221        if (const FunctionProtoType *Proto
222            = FDecl->getType()->getAs<FunctionProtoType>())
223          HasVAListArg = !Proto->isVariadic();
224      }
225      CheckPrintfArguments(TheCall, HasVAListArg, Format->getFormatIdx() - 1,
226                           HasVAListArg ? 0 : Format->getFirstArg() - 1);
227    }
228  }
229
230  for (const NonNullAttr *NonNull = FDecl->getAttr<NonNullAttr>(); NonNull;
231       NonNull = NonNull->getNext<NonNullAttr>())
232    CheckNonNullArguments(NonNull, TheCall);
233
234  return false;
235}
236
237bool Sema::CheckBlockCall(NamedDecl *NDecl, CallExpr *TheCall) {
238  // Printf checking.
239  const FormatAttr *Format = NDecl->getAttr<FormatAttr>();
240  if (!Format)
241    return false;
242
243  const VarDecl *V = dyn_cast<VarDecl>(NDecl);
244  if (!V)
245    return false;
246
247  QualType Ty = V->getType();
248  if (!Ty->isBlockPointerType())
249    return false;
250
251  if (!CheckablePrintfAttr(Format, TheCall))
252    return false;
253
254  bool HasVAListArg = Format->getFirstArg() == 0;
255  if (!HasVAListArg) {
256    const FunctionType *FT =
257      Ty->getAs<BlockPointerType>()->getPointeeType()->getAs<FunctionType>();
258    if (const FunctionProtoType *Proto = dyn_cast<FunctionProtoType>(FT))
259      HasVAListArg = !Proto->isVariadic();
260  }
261  CheckPrintfArguments(TheCall, HasVAListArg, Format->getFormatIdx() - 1,
262                       HasVAListArg ? 0 : Format->getFirstArg() - 1);
263
264  return false;
265}
266
267/// SemaBuiltinAtomicOverloaded - We have a call to a function like
268/// __sync_fetch_and_add, which is an overloaded function based on the pointer
269/// type of its first argument.  The main ActOnCallExpr routines have already
270/// promoted the types of arguments because all of these calls are prototyped as
271/// void(...).
272///
273/// This function goes through and does final semantic checking for these
274/// builtins,
275bool Sema::SemaBuiltinAtomicOverloaded(CallExpr *TheCall) {
276  DeclRefExpr *DRE =cast<DeclRefExpr>(TheCall->getCallee()->IgnoreParenCasts());
277  FunctionDecl *FDecl = cast<FunctionDecl>(DRE->getDecl());
278
279  // Ensure that we have at least one argument to do type inference from.
280  if (TheCall->getNumArgs() < 1)
281    return Diag(TheCall->getLocEnd(), diag::err_typecheck_call_too_few_args)
282              << 0 << TheCall->getCallee()->getSourceRange();
283
284  // Inspect the first argument of the atomic builtin.  This should always be
285  // a pointer type, whose element is an integral scalar or pointer type.
286  // Because it is a pointer type, we don't have to worry about any implicit
287  // casts here.
288  Expr *FirstArg = TheCall->getArg(0);
289  if (!FirstArg->getType()->isPointerType())
290    return Diag(DRE->getLocStart(), diag::err_atomic_builtin_must_be_pointer)
291             << FirstArg->getType() << FirstArg->getSourceRange();
292
293  QualType ValType = FirstArg->getType()->getAs<PointerType>()->getPointeeType();
294  if (!ValType->isIntegerType() && !ValType->isPointerType() &&
295      !ValType->isBlockPointerType())
296    return Diag(DRE->getLocStart(),
297                diag::err_atomic_builtin_must_be_pointer_intptr)
298             << FirstArg->getType() << FirstArg->getSourceRange();
299
300  // We need to figure out which concrete builtin this maps onto.  For example,
301  // __sync_fetch_and_add with a 2 byte object turns into
302  // __sync_fetch_and_add_2.
303#define BUILTIN_ROW(x) \
304  { Builtin::BI##x##_1, Builtin::BI##x##_2, Builtin::BI##x##_4, \
305    Builtin::BI##x##_8, Builtin::BI##x##_16 }
306
307  static const unsigned BuiltinIndices[][5] = {
308    BUILTIN_ROW(__sync_fetch_and_add),
309    BUILTIN_ROW(__sync_fetch_and_sub),
310    BUILTIN_ROW(__sync_fetch_and_or),
311    BUILTIN_ROW(__sync_fetch_and_and),
312    BUILTIN_ROW(__sync_fetch_and_xor),
313    BUILTIN_ROW(__sync_fetch_and_nand),
314
315    BUILTIN_ROW(__sync_add_and_fetch),
316    BUILTIN_ROW(__sync_sub_and_fetch),
317    BUILTIN_ROW(__sync_and_and_fetch),
318    BUILTIN_ROW(__sync_or_and_fetch),
319    BUILTIN_ROW(__sync_xor_and_fetch),
320    BUILTIN_ROW(__sync_nand_and_fetch),
321
322    BUILTIN_ROW(__sync_val_compare_and_swap),
323    BUILTIN_ROW(__sync_bool_compare_and_swap),
324    BUILTIN_ROW(__sync_lock_test_and_set),
325    BUILTIN_ROW(__sync_lock_release)
326  };
327#undef BUILTIN_ROW
328
329  // Determine the index of the size.
330  unsigned SizeIndex;
331  switch (Context.getTypeSizeInChars(ValType).getQuantity()) {
332  case 1: SizeIndex = 0; break;
333  case 2: SizeIndex = 1; break;
334  case 4: SizeIndex = 2; break;
335  case 8: SizeIndex = 3; break;
336  case 16: SizeIndex = 4; break;
337  default:
338    return Diag(DRE->getLocStart(), diag::err_atomic_builtin_pointer_size)
339             << FirstArg->getType() << FirstArg->getSourceRange();
340  }
341
342  // Each of these builtins has one pointer argument, followed by some number of
343  // values (0, 1 or 2) followed by a potentially empty varags list of stuff
344  // that we ignore.  Find out which row of BuiltinIndices to read from as well
345  // as the number of fixed args.
346  unsigned BuiltinID = FDecl->getBuiltinID();
347  unsigned BuiltinIndex, NumFixed = 1;
348  switch (BuiltinID) {
349  default: assert(0 && "Unknown overloaded atomic builtin!");
350  case Builtin::BI__sync_fetch_and_add: BuiltinIndex = 0; break;
351  case Builtin::BI__sync_fetch_and_sub: BuiltinIndex = 1; break;
352  case Builtin::BI__sync_fetch_and_or:  BuiltinIndex = 2; break;
353  case Builtin::BI__sync_fetch_and_and: BuiltinIndex = 3; break;
354  case Builtin::BI__sync_fetch_and_xor: BuiltinIndex = 4; break;
355  case Builtin::BI__sync_fetch_and_nand:BuiltinIndex = 5; break;
356
357  case Builtin::BI__sync_add_and_fetch: BuiltinIndex = 6; break;
358  case Builtin::BI__sync_sub_and_fetch: BuiltinIndex = 7; break;
359  case Builtin::BI__sync_and_and_fetch: BuiltinIndex = 8; break;
360  case Builtin::BI__sync_or_and_fetch:  BuiltinIndex = 9; break;
361  case Builtin::BI__sync_xor_and_fetch: BuiltinIndex =10; break;
362  case Builtin::BI__sync_nand_and_fetch:BuiltinIndex =11; break;
363
364  case Builtin::BI__sync_val_compare_and_swap:
365    BuiltinIndex = 12;
366    NumFixed = 2;
367    break;
368  case Builtin::BI__sync_bool_compare_and_swap:
369    BuiltinIndex = 13;
370    NumFixed = 2;
371    break;
372  case Builtin::BI__sync_lock_test_and_set: BuiltinIndex = 14; break;
373  case Builtin::BI__sync_lock_release:
374    BuiltinIndex = 15;
375    NumFixed = 0;
376    break;
377  }
378
379  // Now that we know how many fixed arguments we expect, first check that we
380  // have at least that many.
381  if (TheCall->getNumArgs() < 1+NumFixed)
382    return Diag(TheCall->getLocEnd(), diag::err_typecheck_call_too_few_args)
383            << 0 << TheCall->getCallee()->getSourceRange();
384
385
386  // Get the decl for the concrete builtin from this, we can tell what the
387  // concrete integer type we should convert to is.
388  unsigned NewBuiltinID = BuiltinIndices[BuiltinIndex][SizeIndex];
389  const char *NewBuiltinName = Context.BuiltinInfo.GetName(NewBuiltinID);
390  IdentifierInfo *NewBuiltinII = PP.getIdentifierInfo(NewBuiltinName);
391  FunctionDecl *NewBuiltinDecl =
392    cast<FunctionDecl>(LazilyCreateBuiltin(NewBuiltinII, NewBuiltinID,
393                                           TUScope, false, DRE->getLocStart()));
394  const FunctionProtoType *BuiltinFT =
395    NewBuiltinDecl->getType()->getAs<FunctionProtoType>();
396  ValType = BuiltinFT->getArgType(0)->getAs<PointerType>()->getPointeeType();
397
398  // If the first type needs to be converted (e.g. void** -> int*), do it now.
399  if (BuiltinFT->getArgType(0) != FirstArg->getType()) {
400    ImpCastExprToType(FirstArg, BuiltinFT->getArgType(0), CastExpr::CK_BitCast);
401    TheCall->setArg(0, FirstArg);
402  }
403
404  // Next, walk the valid ones promoting to the right type.
405  for (unsigned i = 0; i != NumFixed; ++i) {
406    Expr *Arg = TheCall->getArg(i+1);
407
408    // If the argument is an implicit cast, then there was a promotion due to
409    // "...", just remove it now.
410    if (ImplicitCastExpr *ICE = dyn_cast<ImplicitCastExpr>(Arg)) {
411      Arg = ICE->getSubExpr();
412      ICE->setSubExpr(0);
413      ICE->Destroy(Context);
414      TheCall->setArg(i+1, Arg);
415    }
416
417    // GCC does an implicit conversion to the pointer or integer ValType.  This
418    // can fail in some cases (1i -> int**), check for this error case now.
419    CastExpr::CastKind Kind = CastExpr::CK_Unknown;
420    CXXMethodDecl *ConversionDecl = 0;
421    if (CheckCastTypes(Arg->getSourceRange(), ValType, Arg, Kind,
422                       ConversionDecl))
423      return true;
424
425    // Okay, we have something that *can* be converted to the right type.  Check
426    // to see if there is a potentially weird extension going on here.  This can
427    // happen when you do an atomic operation on something like an char* and
428    // pass in 42.  The 42 gets converted to char.  This is even more strange
429    // for things like 45.123 -> char, etc.
430    // FIXME: Do this check.
431    ImpCastExprToType(Arg, ValType, Kind, /*isLvalue=*/false);
432    TheCall->setArg(i+1, Arg);
433  }
434
435  // Switch the DeclRefExpr to refer to the new decl.
436  DRE->setDecl(NewBuiltinDecl);
437  DRE->setType(NewBuiltinDecl->getType());
438
439  // Set the callee in the CallExpr.
440  // FIXME: This leaks the original parens and implicit casts.
441  Expr *PromotedCall = DRE;
442  UsualUnaryConversions(PromotedCall);
443  TheCall->setCallee(PromotedCall);
444
445
446  // Change the result type of the call to match the result type of the decl.
447  TheCall->setType(NewBuiltinDecl->getResultType());
448  return false;
449}
450
451
452/// CheckObjCString - Checks that the argument to the builtin
453/// CFString constructor is correct
454/// FIXME: GCC currently emits the following warning:
455/// "warning: input conversion stopped due to an input byte that does not
456///           belong to the input codeset UTF-8"
457/// Note: It might also make sense to do the UTF-16 conversion here (would
458/// simplify the backend).
459bool Sema::CheckObjCString(Expr *Arg) {
460  Arg = Arg->IgnoreParenCasts();
461  StringLiteral *Literal = dyn_cast<StringLiteral>(Arg);
462
463  if (!Literal || Literal->isWide()) {
464    Diag(Arg->getLocStart(), diag::err_cfstring_literal_not_string_constant)
465      << Arg->getSourceRange();
466    return true;
467  }
468
469  const char *Data = Literal->getStrData();
470  unsigned Length = Literal->getByteLength();
471
472  for (unsigned i = 0; i < Length; ++i) {
473    if (!Data[i]) {
474      Diag(getLocationOfStringLiteralByte(Literal, i),
475           diag::warn_cfstring_literal_contains_nul_character)
476        << Arg->getSourceRange();
477      break;
478    }
479  }
480
481  return false;
482}
483
484/// SemaBuiltinVAStart - Check the arguments to __builtin_va_start for validity.
485/// Emit an error and return true on failure, return false on success.
486bool Sema::SemaBuiltinVAStart(CallExpr *TheCall) {
487  Expr *Fn = TheCall->getCallee();
488  if (TheCall->getNumArgs() > 2) {
489    Diag(TheCall->getArg(2)->getLocStart(),
490         diag::err_typecheck_call_too_many_args)
491      << 0 /*function call*/ << Fn->getSourceRange()
492      << SourceRange(TheCall->getArg(2)->getLocStart(),
493                     (*(TheCall->arg_end()-1))->getLocEnd());
494    return true;
495  }
496
497  if (TheCall->getNumArgs() < 2) {
498    return Diag(TheCall->getLocEnd(), diag::err_typecheck_call_too_few_args)
499      << 0 /*function call*/;
500  }
501
502  // Determine whether the current function is variadic or not.
503  bool isVariadic;
504  if (CurBlock)
505    isVariadic = CurBlock->isVariadic;
506  else if (getCurFunctionDecl()) {
507    if (FunctionProtoType* FTP =
508            dyn_cast<FunctionProtoType>(getCurFunctionDecl()->getType()))
509      isVariadic = FTP->isVariadic();
510    else
511      isVariadic = false;
512  } else {
513    isVariadic = getCurMethodDecl()->isVariadic();
514  }
515
516  if (!isVariadic) {
517    Diag(Fn->getLocStart(), diag::err_va_start_used_in_non_variadic_function);
518    return true;
519  }
520
521  // Verify that the second argument to the builtin is the last argument of the
522  // current function or method.
523  bool SecondArgIsLastNamedArgument = false;
524  const Expr *Arg = TheCall->getArg(1)->IgnoreParenCasts();
525
526  if (const DeclRefExpr *DR = dyn_cast<DeclRefExpr>(Arg)) {
527    if (const ParmVarDecl *PV = dyn_cast<ParmVarDecl>(DR->getDecl())) {
528      // FIXME: This isn't correct for methods (results in bogus warning).
529      // Get the last formal in the current function.
530      const ParmVarDecl *LastArg;
531      if (CurBlock)
532        LastArg = *(CurBlock->TheDecl->param_end()-1);
533      else if (FunctionDecl *FD = getCurFunctionDecl())
534        LastArg = *(FD->param_end()-1);
535      else
536        LastArg = *(getCurMethodDecl()->param_end()-1);
537      SecondArgIsLastNamedArgument = PV == LastArg;
538    }
539  }
540
541  if (!SecondArgIsLastNamedArgument)
542    Diag(TheCall->getArg(1)->getLocStart(),
543         diag::warn_second_parameter_of_va_start_not_last_named_argument);
544  return false;
545}
546
547/// SemaBuiltinUnorderedCompare - Handle functions like __builtin_isgreater and
548/// friends.  This is declared to take (...), so we have to check everything.
549bool Sema::SemaBuiltinUnorderedCompare(CallExpr *TheCall) {
550  if (TheCall->getNumArgs() < 2)
551    return Diag(TheCall->getLocEnd(), diag::err_typecheck_call_too_few_args)
552      << 0 /*function call*/;
553  if (TheCall->getNumArgs() > 2)
554    return Diag(TheCall->getArg(2)->getLocStart(),
555                diag::err_typecheck_call_too_many_args)
556      << 0 /*function call*/
557      << SourceRange(TheCall->getArg(2)->getLocStart(),
558                     (*(TheCall->arg_end()-1))->getLocEnd());
559
560  Expr *OrigArg0 = TheCall->getArg(0);
561  Expr *OrigArg1 = TheCall->getArg(1);
562
563  // Do standard promotions between the two arguments, returning their common
564  // type.
565  QualType Res = UsualArithmeticConversions(OrigArg0, OrigArg1, false);
566
567  // Make sure any conversions are pushed back into the call; this is
568  // type safe since unordered compare builtins are declared as "_Bool
569  // foo(...)".
570  TheCall->setArg(0, OrigArg0);
571  TheCall->setArg(1, OrigArg1);
572
573  if (OrigArg0->isTypeDependent() || OrigArg1->isTypeDependent())
574    return false;
575
576  // If the common type isn't a real floating type, then the arguments were
577  // invalid for this operation.
578  if (!Res->isRealFloatingType())
579    return Diag(OrigArg0->getLocStart(),
580                diag::err_typecheck_call_invalid_ordered_compare)
581      << OrigArg0->getType() << OrigArg1->getType()
582      << SourceRange(OrigArg0->getLocStart(), OrigArg1->getLocEnd());
583
584  return false;
585}
586
587/// SemaBuiltinUnorderedCompare - Handle functions like __builtin_isnan and
588/// friends.  This is declared to take (...), so we have to check everything.
589bool Sema::SemaBuiltinUnaryFP(CallExpr *TheCall) {
590  if (TheCall->getNumArgs() < 1)
591    return Diag(TheCall->getLocEnd(), diag::err_typecheck_call_too_few_args)
592      << 0 /*function call*/;
593  if (TheCall->getNumArgs() > 1)
594    return Diag(TheCall->getArg(1)->getLocStart(),
595                diag::err_typecheck_call_too_many_args)
596      << 0 /*function call*/
597      << SourceRange(TheCall->getArg(1)->getLocStart(),
598                     (*(TheCall->arg_end()-1))->getLocEnd());
599
600  Expr *OrigArg = TheCall->getArg(0);
601
602  if (OrigArg->isTypeDependent())
603    return false;
604
605  // This operation requires a floating-point number
606  if (!OrigArg->getType()->isRealFloatingType())
607    return Diag(OrigArg->getLocStart(),
608                diag::err_typecheck_call_invalid_unary_fp)
609      << OrigArg->getType() << OrigArg->getSourceRange();
610
611  return false;
612}
613
614bool Sema::SemaBuiltinStackAddress(CallExpr *TheCall) {
615  // The signature for these builtins is exact; the only thing we need
616  // to check is that the argument is a constant.
617  SourceLocation Loc;
618  if (!TheCall->getArg(0)->isTypeDependent() &&
619      !TheCall->getArg(0)->isValueDependent() &&
620      !TheCall->getArg(0)->isIntegerConstantExpr(Context, &Loc))
621    return Diag(Loc, diag::err_stack_const_level) << TheCall->getSourceRange();
622
623  return false;
624}
625
626/// SemaBuiltinShuffleVector - Handle __builtin_shufflevector.
627// This is declared to take (...), so we have to check everything.
628Action::OwningExprResult Sema::SemaBuiltinShuffleVector(CallExpr *TheCall) {
629  if (TheCall->getNumArgs() < 3)
630    return ExprError(Diag(TheCall->getLocEnd(),
631                          diag::err_typecheck_call_too_few_args)
632      << 0 /*function call*/ << TheCall->getSourceRange());
633
634  unsigned numElements = std::numeric_limits<unsigned>::max();
635  if (!TheCall->getArg(0)->isTypeDependent() &&
636      !TheCall->getArg(1)->isTypeDependent()) {
637    QualType FAType = TheCall->getArg(0)->getType();
638    QualType SAType = TheCall->getArg(1)->getType();
639
640    if (!FAType->isVectorType() || !SAType->isVectorType()) {
641      Diag(TheCall->getLocStart(), diag::err_shufflevector_non_vector)
642        << SourceRange(TheCall->getArg(0)->getLocStart(),
643                       TheCall->getArg(1)->getLocEnd());
644      return ExprError();
645    }
646
647    if (!Context.hasSameUnqualifiedType(FAType, SAType)) {
648      Diag(TheCall->getLocStart(), diag::err_shufflevector_incompatible_vector)
649        << SourceRange(TheCall->getArg(0)->getLocStart(),
650                       TheCall->getArg(1)->getLocEnd());
651      return ExprError();
652    }
653
654    numElements = FAType->getAs<VectorType>()->getNumElements();
655    if (TheCall->getNumArgs() != numElements+2) {
656      if (TheCall->getNumArgs() < numElements+2)
657        return ExprError(Diag(TheCall->getLocEnd(),
658                              diag::err_typecheck_call_too_few_args)
659                 << 0 /*function call*/ << TheCall->getSourceRange());
660      return ExprError(Diag(TheCall->getLocEnd(),
661                            diag::err_typecheck_call_too_many_args)
662                 << 0 /*function call*/ << TheCall->getSourceRange());
663    }
664  }
665
666  for (unsigned i = 2; i < TheCall->getNumArgs(); i++) {
667    if (TheCall->getArg(i)->isTypeDependent() ||
668        TheCall->getArg(i)->isValueDependent())
669      continue;
670
671    llvm::APSInt Result(32);
672    if (!TheCall->getArg(i)->isIntegerConstantExpr(Result, Context))
673      return ExprError(Diag(TheCall->getLocStart(),
674                  diag::err_shufflevector_nonconstant_argument)
675                << TheCall->getArg(i)->getSourceRange());
676
677    if (Result.getActiveBits() > 64 || Result.getZExtValue() >= numElements*2)
678      return ExprError(Diag(TheCall->getLocStart(),
679                  diag::err_shufflevector_argument_too_large)
680               << TheCall->getArg(i)->getSourceRange());
681  }
682
683  llvm::SmallVector<Expr*, 32> exprs;
684
685  for (unsigned i = 0, e = TheCall->getNumArgs(); i != e; i++) {
686    exprs.push_back(TheCall->getArg(i));
687    TheCall->setArg(i, 0);
688  }
689
690  return Owned(new (Context) ShuffleVectorExpr(Context, exprs.begin(),
691                                            exprs.size(), exprs[0]->getType(),
692                                            TheCall->getCallee()->getLocStart(),
693                                            TheCall->getRParenLoc()));
694}
695
696/// SemaBuiltinPrefetch - Handle __builtin_prefetch.
697// This is declared to take (const void*, ...) and can take two
698// optional constant int args.
699bool Sema::SemaBuiltinPrefetch(CallExpr *TheCall) {
700  unsigned NumArgs = TheCall->getNumArgs();
701
702  if (NumArgs > 3)
703    return Diag(TheCall->getLocEnd(), diag::err_typecheck_call_too_many_args)
704             << 0 /*function call*/ << TheCall->getSourceRange();
705
706  // Argument 0 is checked for us and the remaining arguments must be
707  // constant integers.
708  for (unsigned i = 1; i != NumArgs; ++i) {
709    Expr *Arg = TheCall->getArg(i);
710    if (Arg->isTypeDependent())
711      continue;
712
713    if (!Arg->getType()->isIntegralType())
714      return Diag(TheCall->getLocStart(), diag::err_prefetch_invalid_arg_type)
715              << Arg->getSourceRange();
716
717    ImpCastExprToType(Arg, Context.IntTy, CastExpr::CK_IntegralCast);
718    TheCall->setArg(i, Arg);
719
720    if (Arg->isValueDependent())
721      continue;
722
723    llvm::APSInt Result;
724    if (!Arg->isIntegerConstantExpr(Result, Context))
725      return Diag(TheCall->getLocStart(), diag::err_prefetch_invalid_arg_ice)
726        << SourceRange(Arg->getLocStart(), Arg->getLocEnd());
727
728    // FIXME: gcc issues a warning and rewrites these to 0. These
729    // seems especially odd for the third argument since the default
730    // is 3.
731    if (i == 1) {
732      if (Result.getLimitedValue() > 1)
733        return Diag(TheCall->getLocStart(), diag::err_argument_invalid_range)
734             << "0" << "1" << Arg->getSourceRange();
735    } else {
736      if (Result.getLimitedValue() > 3)
737        return Diag(TheCall->getLocStart(), diag::err_argument_invalid_range)
738            << "0" << "3" << Arg->getSourceRange();
739    }
740  }
741
742  return false;
743}
744
745/// SemaBuiltinEHReturnDataRegNo - Handle __builtin_eh_return_data_regno, the
746/// operand must be an integer constant.
747bool Sema::SemaBuiltinEHReturnDataRegNo(CallExpr *TheCall) {
748  llvm::APSInt Result;
749  if (!TheCall->getArg(0)->isIntegerConstantExpr(Result, Context))
750    return Diag(TheCall->getLocStart(), diag::err_expr_not_ice)
751      << TheCall->getArg(0)->getSourceRange();
752
753  return false;
754}
755
756
757/// SemaBuiltinObjectSize - Handle __builtin_object_size(void *ptr,
758/// int type). This simply type checks that type is one of the defined
759/// constants (0-3).
760// For compatability check 0-3, llvm only handles 0 and 2.
761bool Sema::SemaBuiltinObjectSize(CallExpr *TheCall) {
762  Expr *Arg = TheCall->getArg(1);
763  if (Arg->isTypeDependent())
764    return false;
765
766  QualType ArgType = Arg->getType();
767  const BuiltinType *BT = ArgType->getAs<BuiltinType>();
768  llvm::APSInt Result(32);
769  if (!BT || BT->getKind() != BuiltinType::Int)
770    return Diag(TheCall->getLocStart(), diag::err_object_size_invalid_argument)
771             << SourceRange(Arg->getLocStart(), Arg->getLocEnd());
772
773  if (Arg->isValueDependent())
774    return false;
775
776  if (!Arg->isIntegerConstantExpr(Result, Context)) {
777    return Diag(TheCall->getLocStart(), diag::err_object_size_invalid_argument)
778             << SourceRange(Arg->getLocStart(), Arg->getLocEnd());
779  }
780
781  if (Result.getSExtValue() < 0 || Result.getSExtValue() > 3) {
782    return Diag(TheCall->getLocStart(), diag::err_argument_invalid_range)
783             << "0" << "3" << SourceRange(Arg->getLocStart(), Arg->getLocEnd());
784  }
785
786  return false;
787}
788
789/// SemaBuiltinLongjmp - Handle __builtin_longjmp(void *env[5], int val).
790/// This checks that val is a constant 1.
791bool Sema::SemaBuiltinLongjmp(CallExpr *TheCall) {
792  Expr *Arg = TheCall->getArg(1);
793  if (Arg->isTypeDependent() || Arg->isValueDependent())
794    return false;
795
796  llvm::APSInt Result(32);
797  if (!Arg->isIntegerConstantExpr(Result, Context) || Result != 1)
798    return Diag(TheCall->getLocStart(), diag::err_builtin_longjmp_invalid_val)
799             << SourceRange(Arg->getLocStart(), Arg->getLocEnd());
800
801  return false;
802}
803
804// Handle i > 1 ? "x" : "y", recursivelly
805bool Sema::SemaCheckStringLiteral(const Expr *E, const CallExpr *TheCall,
806                                  bool HasVAListArg,
807                                  unsigned format_idx, unsigned firstDataArg) {
808  if (E->isTypeDependent() || E->isValueDependent())
809    return false;
810
811  switch (E->getStmtClass()) {
812  case Stmt::ConditionalOperatorClass: {
813    const ConditionalOperator *C = cast<ConditionalOperator>(E);
814    return SemaCheckStringLiteral(C->getTrueExpr(), TheCall,
815                                  HasVAListArg, format_idx, firstDataArg)
816        && SemaCheckStringLiteral(C->getRHS(), TheCall,
817                                  HasVAListArg, format_idx, firstDataArg);
818  }
819
820  case Stmt::ImplicitCastExprClass: {
821    const ImplicitCastExpr *Expr = cast<ImplicitCastExpr>(E);
822    return SemaCheckStringLiteral(Expr->getSubExpr(), TheCall, HasVAListArg,
823                                  format_idx, firstDataArg);
824  }
825
826  case Stmt::ParenExprClass: {
827    const ParenExpr *Expr = cast<ParenExpr>(E);
828    return SemaCheckStringLiteral(Expr->getSubExpr(), TheCall, HasVAListArg,
829                                  format_idx, firstDataArg);
830  }
831
832  case Stmt::DeclRefExprClass: {
833    const DeclRefExpr *DR = cast<DeclRefExpr>(E);
834
835    // As an exception, do not flag errors for variables binding to
836    // const string literals.
837    if (const VarDecl *VD = dyn_cast<VarDecl>(DR->getDecl())) {
838      bool isConstant = false;
839      QualType T = DR->getType();
840
841      if (const ArrayType *AT = Context.getAsArrayType(T)) {
842        isConstant = AT->getElementType().isConstant(Context);
843      } else if (const PointerType *PT = T->getAs<PointerType>()) {
844        isConstant = T.isConstant(Context) &&
845                     PT->getPointeeType().isConstant(Context);
846      }
847
848      if (isConstant) {
849        if (const Expr *Init = VD->getAnyInitializer())
850          return SemaCheckStringLiteral(Init, TheCall,
851                                        HasVAListArg, format_idx, firstDataArg);
852      }
853
854      // For vprintf* functions (i.e., HasVAListArg==true), we add a
855      // special check to see if the format string is a function parameter
856      // of the function calling the printf function.  If the function
857      // has an attribute indicating it is a printf-like function, then we
858      // should suppress warnings concerning non-literals being used in a call
859      // to a vprintf function.  For example:
860      //
861      // void
862      // logmessage(char const *fmt __attribute__ (format (printf, 1, 2)), ...){
863      //      va_list ap;
864      //      va_start(ap, fmt);
865      //      vprintf(fmt, ap);  // Do NOT emit a warning about "fmt".
866      //      ...
867      //
868      //
869      //  FIXME: We don't have full attribute support yet, so just check to see
870      //    if the argument is a DeclRefExpr that references a parameter.  We'll
871      //    add proper support for checking the attribute later.
872      if (HasVAListArg)
873        if (isa<ParmVarDecl>(VD))
874          return true;
875    }
876
877    return false;
878  }
879
880  case Stmt::CallExprClass: {
881    const CallExpr *CE = cast<CallExpr>(E);
882    if (const ImplicitCastExpr *ICE
883          = dyn_cast<ImplicitCastExpr>(CE->getCallee())) {
884      if (const DeclRefExpr *DRE = dyn_cast<DeclRefExpr>(ICE->getSubExpr())) {
885        if (const FunctionDecl *FD = dyn_cast<FunctionDecl>(DRE->getDecl())) {
886          if (const FormatArgAttr *FA = FD->getAttr<FormatArgAttr>()) {
887            unsigned ArgIndex = FA->getFormatIdx();
888            const Expr *Arg = CE->getArg(ArgIndex - 1);
889
890            return SemaCheckStringLiteral(Arg, TheCall, HasVAListArg,
891                                          format_idx, firstDataArg);
892          }
893        }
894      }
895    }
896
897    return false;
898  }
899  case Stmt::ObjCStringLiteralClass:
900  case Stmt::StringLiteralClass: {
901    const StringLiteral *StrE = NULL;
902
903    if (const ObjCStringLiteral *ObjCFExpr = dyn_cast<ObjCStringLiteral>(E))
904      StrE = ObjCFExpr->getString();
905    else
906      StrE = cast<StringLiteral>(E);
907
908    if (StrE) {
909      CheckPrintfString(StrE, E, TheCall, HasVAListArg, format_idx,
910                        firstDataArg);
911      return true;
912    }
913
914    return false;
915  }
916
917  default:
918    return false;
919  }
920}
921
922void
923Sema::CheckNonNullArguments(const NonNullAttr *NonNull,
924                            const CallExpr *TheCall) {
925  for (NonNullAttr::iterator i = NonNull->begin(), e = NonNull->end();
926       i != e; ++i) {
927    const Expr *ArgExpr = TheCall->getArg(*i);
928    if (ArgExpr->isNullPointerConstant(Context,
929                                       Expr::NPC_ValueDependentIsNotNull))
930      Diag(TheCall->getCallee()->getLocStart(), diag::warn_null_arg)
931        << ArgExpr->getSourceRange();
932  }
933}
934
935/// CheckPrintfArguments - Check calls to printf (and similar functions) for
936/// correct use of format strings.
937///
938///  HasVAListArg - A predicate indicating whether the printf-like
939///    function is passed an explicit va_arg argument (e.g., vprintf)
940///
941///  format_idx - The index into Args for the format string.
942///
943/// Improper format strings to functions in the printf family can be
944/// the source of bizarre bugs and very serious security holes.  A
945/// good source of information is available in the following paper
946/// (which includes additional references):
947///
948///  FormatGuard: Automatic Protection From printf Format String
949///  Vulnerabilities, Proceedings of the 10th USENIX Security Symposium, 2001.
950///
951/// Functionality implemented:
952///
953///  We can statically check the following properties for string
954///  literal format strings for non v.*printf functions (where the
955///  arguments are passed directly):
956//
957///  (1) Are the number of format conversions equal to the number of
958///      data arguments?
959///
960///  (2) Does each format conversion correctly match the type of the
961///      corresponding data argument?  (TODO)
962///
963/// Moreover, for all printf functions we can:
964///
965///  (3) Check for a missing format string (when not caught by type checking).
966///
967///  (4) Check for no-operation flags; e.g. using "#" with format
968///      conversion 'c'  (TODO)
969///
970///  (5) Check the use of '%n', a major source of security holes.
971///
972///  (6) Check for malformed format conversions that don't specify anything.
973///
974///  (7) Check for empty format strings.  e.g: printf("");
975///
976///  (8) Check that the format string is a wide literal.
977///
978/// All of these checks can be done by parsing the format string.
979///
980/// For now, we ONLY do (1), (3), (5), (6), (7), and (8).
981void
982Sema::CheckPrintfArguments(const CallExpr *TheCall, bool HasVAListArg,
983                           unsigned format_idx, unsigned firstDataArg) {
984  const Expr *Fn = TheCall->getCallee();
985
986  // The way the format attribute works in GCC, the implicit this argument
987  // of member functions is counted. However, it doesn't appear in our own
988  // lists, so decrement format_idx in that case.
989  if (isa<CXXMemberCallExpr>(TheCall)) {
990    // Catch a format attribute mistakenly referring to the object argument.
991    if (format_idx == 0)
992      return;
993    --format_idx;
994    if(firstDataArg != 0)
995      --firstDataArg;
996  }
997
998  // CHECK: printf-like function is called with no format string.
999  if (format_idx >= TheCall->getNumArgs()) {
1000    Diag(TheCall->getRParenLoc(), diag::warn_printf_missing_format_string)
1001      << Fn->getSourceRange();
1002    return;
1003  }
1004
1005  const Expr *OrigFormatExpr = TheCall->getArg(format_idx)->IgnoreParenCasts();
1006
1007  // CHECK: format string is not a string literal.
1008  //
1009  // Dynamically generated format strings are difficult to
1010  // automatically vet at compile time.  Requiring that format strings
1011  // are string literals: (1) permits the checking of format strings by
1012  // the compiler and thereby (2) can practically remove the source of
1013  // many format string exploits.
1014
1015  // Format string can be either ObjC string (e.g. @"%d") or
1016  // C string (e.g. "%d")
1017  // ObjC string uses the same format specifiers as C string, so we can use
1018  // the same format string checking logic for both ObjC and C strings.
1019  if (SemaCheckStringLiteral(OrigFormatExpr, TheCall, HasVAListArg, format_idx,
1020                             firstDataArg))
1021    return;  // Literal format string found, check done!
1022
1023  // If there are no arguments specified, warn with -Wformat-security, otherwise
1024  // warn only with -Wformat-nonliteral.
1025  if (TheCall->getNumArgs() == format_idx+1)
1026    Diag(TheCall->getArg(format_idx)->getLocStart(),
1027         diag::warn_printf_nonliteral_noargs)
1028      << OrigFormatExpr->getSourceRange();
1029  else
1030    Diag(TheCall->getArg(format_idx)->getLocStart(),
1031         diag::warn_printf_nonliteral)
1032           << OrigFormatExpr->getSourceRange();
1033}
1034
1035namespace {
1036class CheckPrintfHandler : public analyze_printf::FormatStringHandler {
1037  Sema &S;
1038  const StringLiteral *FExpr;
1039  const Expr *OrigFormatExpr;
1040  unsigned NumConversions;
1041  const unsigned NumDataArgs;
1042  const bool IsObjCLiteral;
1043  const char *Beg; // Start of format string.
1044  const bool HasVAListArg;
1045  const CallExpr *TheCall;
1046  unsigned FormatIdx;
1047public:
1048  CheckPrintfHandler(Sema &s, const StringLiteral *fexpr,
1049                     const Expr *origFormatExpr,
1050                     unsigned numDataArgs, bool isObjCLiteral,
1051                     const char *beg, bool hasVAListArg,
1052                     const CallExpr *theCall, unsigned formatIdx)
1053    : S(s), FExpr(fexpr), OrigFormatExpr(origFormatExpr),
1054      NumConversions(0), NumDataArgs(numDataArgs),
1055      IsObjCLiteral(isObjCLiteral), Beg(beg),
1056      HasVAListArg(hasVAListArg),
1057      TheCall(theCall), FormatIdx(formatIdx) {}
1058
1059  void DoneProcessing();
1060
1061  void HandleIncompleteFormatSpecifier(const char *startSpecifier,
1062                                       unsigned specifierLen);
1063
1064  void
1065  HandleInvalidConversionSpecifier(const analyze_printf::FormatSpecifier &FS,
1066                                   const char *startSpecifier,
1067                                   unsigned specifierLen);
1068
1069  void HandleNullChar(const char *nullCharacter);
1070
1071  bool HandleFormatSpecifier(const analyze_printf::FormatSpecifier &FS,
1072                             const char *startSpecifier,
1073                             unsigned specifierLen);
1074private:
1075  SourceRange getFormatStringRange();
1076  SourceRange getFormatSpecifierRange(const char *startSpecifier,
1077                                      unsigned specifierLen);
1078  SourceLocation getLocationOfByte(const char *x);
1079
1080  bool HandleAmount(const analyze_printf::OptionalAmount &Amt,
1081                    unsigned MissingArgDiag, unsigned BadTypeDiag,
1082          const char *startSpecifier, unsigned specifierLen);
1083  void HandleFlags(const analyze_printf::FormatSpecifier &FS,
1084                   llvm::StringRef flag, llvm::StringRef cspec,
1085                   const char *startSpecifier, unsigned specifierLen);
1086
1087  bool MatchType(QualType A, QualType B, bool ignoreSign);
1088
1089  const Expr *getDataArg(unsigned i) const;
1090};
1091}
1092
1093SourceRange CheckPrintfHandler::getFormatStringRange() {
1094  return OrigFormatExpr->getSourceRange();
1095}
1096
1097SourceRange CheckPrintfHandler::
1098getFormatSpecifierRange(const char *startSpecifier, unsigned specifierLen) {
1099  return SourceRange(getLocationOfByte(startSpecifier),
1100                     getLocationOfByte(startSpecifier+specifierLen-1));
1101}
1102
1103SourceLocation CheckPrintfHandler::getLocationOfByte(const char *x) {
1104  return S.getLocationOfStringLiteralByte(FExpr, x - Beg);
1105}
1106
1107void CheckPrintfHandler::
1108HandleIncompleteFormatSpecifier(const char *startSpecifier,
1109                                unsigned specifierLen) {
1110  SourceLocation Loc = getLocationOfByte(startSpecifier);
1111  S.Diag(Loc, diag::warn_printf_incomplete_specifier)
1112    << getFormatSpecifierRange(startSpecifier, specifierLen);
1113}
1114
1115void CheckPrintfHandler::
1116HandleInvalidConversionSpecifier(const analyze_printf::FormatSpecifier &FS,
1117                                 const char *startSpecifier,
1118                                 unsigned specifierLen) {
1119
1120  ++NumConversions;
1121  const analyze_printf::ConversionSpecifier &CS =
1122    FS.getConversionSpecifier();
1123  SourceLocation Loc = getLocationOfByte(CS.getStart());
1124  S.Diag(Loc, diag::warn_printf_invalid_conversion)
1125      << llvm::StringRef(CS.getStart(), CS.getLength())
1126      << getFormatSpecifierRange(startSpecifier, specifierLen);
1127}
1128
1129void CheckPrintfHandler::HandleNullChar(const char *nullCharacter) {
1130  // The presence of a null character is likely an error.
1131  S.Diag(getLocationOfByte(nullCharacter),
1132         diag::warn_printf_format_string_contains_null_char)
1133    << getFormatStringRange();
1134}
1135
1136const Expr *CheckPrintfHandler::getDataArg(unsigned i) const {
1137  return TheCall->getArg(FormatIdx + i);
1138}
1139
1140bool CheckPrintfHandler::MatchType(QualType A, QualType B, bool ignoreSign) {
1141  A = S.Context.getCanonicalType(A).getUnqualifiedType();
1142  B = S.Context.getCanonicalType(B).getUnqualifiedType();
1143
1144  if (A == B)
1145    return true;
1146
1147  if (ignoreSign) {
1148    if (const BuiltinType *BT = B->getAs<BuiltinType>()) {
1149      switch (BT->getKind()) {
1150        default:
1151          return false;
1152        case BuiltinType::Char_S:
1153        case BuiltinType::SChar:
1154          return A == S.Context.UnsignedCharTy;
1155        case BuiltinType::Char_U:
1156        case BuiltinType::UChar:
1157          return A == S.Context.SignedCharTy;
1158        case BuiltinType::Short:
1159          return A == S.Context.UnsignedShortTy;
1160        case BuiltinType::UShort:
1161          return A == S.Context.ShortTy;
1162        case BuiltinType::Int:
1163          return A == S.Context.UnsignedIntTy;
1164        case BuiltinType::UInt:
1165          return A == S.Context.IntTy;
1166        case BuiltinType::Long:
1167          return A == S.Context.UnsignedLongTy;
1168        case BuiltinType::ULong:
1169          return A == S.Context.LongTy;
1170        case BuiltinType::LongLong:
1171          return A == S.Context.UnsignedLongLongTy;
1172        case BuiltinType::ULongLong:
1173          return A == S.Context.LongLongTy;
1174      }
1175      return A == B;
1176    }
1177  }
1178  return false;
1179}
1180
1181void CheckPrintfHandler::HandleFlags(const analyze_printf::FormatSpecifier &FS,
1182                                     llvm::StringRef flag,
1183                                     llvm::StringRef cspec,
1184                                     const char *startSpecifier,
1185                                     unsigned specifierLen) {
1186  const analyze_printf::ConversionSpecifier &CS = FS.getConversionSpecifier();
1187  S.Diag(getLocationOfByte(CS.getStart()), diag::warn_printf_nonsensical_flag)
1188    << flag << cspec << getFormatSpecifierRange(startSpecifier, specifierLen);
1189}
1190
1191bool
1192CheckPrintfHandler::HandleAmount(const analyze_printf::OptionalAmount &Amt,
1193                                 unsigned MissingArgDiag,
1194                                 unsigned BadTypeDiag,
1195                                 const char *startSpecifier,
1196                                 unsigned specifierLen) {
1197
1198  if (Amt.hasDataArgument()) {
1199    ++NumConversions;
1200    if (!HasVAListArg) {
1201      if (NumConversions > NumDataArgs) {
1202        S.Diag(getLocationOfByte(Amt.getStart()), MissingArgDiag)
1203          << getFormatSpecifierRange(startSpecifier, specifierLen);
1204        // Don't do any more checking.  We will just emit
1205        // spurious errors.
1206        return false;
1207      }
1208
1209      // Type check the data argument.  It should be an 'int'.
1210      // Although not in conformance with C99, we also allow the argument to be
1211      // an 'unsigned int' as that is a reasonably safe case.  GCC also
1212      // doesn't emit a warning for that case.
1213      const Expr *Arg = getDataArg(NumConversions);
1214      QualType T = Arg->getType();
1215      if (!MatchType(T, S.Context.IntTy, true)) {
1216        S.Diag(getLocationOfByte(Amt.getStart()), BadTypeDiag)
1217          << S.Context.IntTy << T
1218          << getFormatSpecifierRange(startSpecifier, specifierLen)
1219          << Arg->getSourceRange();
1220        // Don't do any more checking.  We will just emit
1221        // spurious errors.
1222        return false;
1223      }
1224    }
1225  }
1226  return true;
1227}
1228
1229bool
1230CheckPrintfHandler::HandleFormatSpecifier(const analyze_printf::FormatSpecifier
1231                                            &FS,
1232                                          const char *startSpecifier,
1233                                          unsigned specifierLen) {
1234
1235  using namespace analyze_printf;
1236  const ConversionSpecifier &CS = FS.getConversionSpecifier();
1237
1238  // First check if the field width, precision, and conversion specifier
1239  // have matching data arguments.
1240  if (!HandleAmount(FS.getFieldWidth(),
1241                    diag::warn_printf_asterisk_width_missing_arg,
1242                    diag::warn_printf_asterisk_width_wrong_type,
1243          startSpecifier, specifierLen)) {
1244    return false;
1245  }
1246
1247  if (!HandleAmount(FS.getPrecision(),
1248                    diag::warn_printf_asterisk_precision_missing_arg,
1249                    diag::warn_printf_asterisk_precision_wrong_type,
1250          startSpecifier, specifierLen)) {
1251    return false;
1252  }
1253
1254  // Check for using an Objective-C specific conversion specifier
1255  // in a non-ObjC literal.
1256  if (!IsObjCLiteral && CS.isObjCArg()) {
1257    HandleInvalidConversionSpecifier(FS, startSpecifier, specifierLen);
1258
1259    // Continue checking the other format specifiers.
1260    return true;
1261  }
1262
1263  if (!CS.consumesDataArgument()) {
1264    // FIXME: Technically specifying a precision or field width here
1265    // makes no sense.  Worth issuing a warning at some point.
1266    return true;
1267  }
1268
1269  ++NumConversions;
1270
1271  // Are we using '%n'?  Issue a warning about this being
1272  // a possible security issue.
1273  if (CS.getKind() == ConversionSpecifier::OutIntPtrArg) {
1274    S.Diag(getLocationOfByte(CS.getStart()), diag::warn_printf_write_back)
1275      << getFormatSpecifierRange(startSpecifier, specifierLen);
1276    // Continue checking the other format specifiers.
1277    return true;
1278  }
1279
1280  if (CS.getKind() == ConversionSpecifier::VoidPtrArg) {
1281    if (FS.getPrecision().getHowSpecified() != OptionalAmount::NotSpecified)
1282      S.Diag(getLocationOfByte(CS.getStart()),
1283             diag::warn_printf_nonsensical_precision)
1284        << CS.getCharacters()
1285        << getFormatSpecifierRange(startSpecifier, specifierLen);
1286  }
1287  if (CS.getKind() == ConversionSpecifier::VoidPtrArg ||
1288      CS.getKind() == ConversionSpecifier::CStrArg) {
1289    // FIXME: Instead of using "0", "+", etc., eventually get them from
1290    // the FormatSpecifier.
1291    if (FS.hasLeadingZeros())
1292      HandleFlags(FS, "0", CS.getCharacters(), startSpecifier, specifierLen);
1293    if (FS.hasPlusPrefix())
1294      HandleFlags(FS, "+", CS.getCharacters(), startSpecifier, specifierLen);
1295    if (FS.hasSpacePrefix())
1296      HandleFlags(FS, " ", CS.getCharacters(), startSpecifier, specifierLen);
1297  }
1298
1299  // The remaining checks depend on the data arguments.
1300  if (HasVAListArg)
1301    return true;
1302
1303  if (NumConversions > NumDataArgs) {
1304    S.Diag(getLocationOfByte(CS.getStart()),
1305           diag::warn_printf_insufficient_data_args)
1306      << getFormatSpecifierRange(startSpecifier, specifierLen);
1307    // Don't do any more checking.
1308    return false;
1309  }
1310
1311  // Now type check the data expression that matches the
1312  // format specifier.
1313  const Expr *Ex = getDataArg(NumConversions);
1314  const analyze_printf::ArgTypeResult &ATR = FS.getArgType(S.Context);
1315
1316  if (const QualType *T = ATR.getSpecificType()) {
1317    if (!MatchType(*T, Ex->getType(), true)) {
1318      // Check if we didn't match because of an implicit cast from a 'char'
1319      // or 'short' to an 'int'.  This is done because printf is a varargs
1320      // function.
1321      if (const ImplicitCastExpr *ICE = dyn_cast<ImplicitCastExpr>(Ex))
1322        if (ICE->getType() == S.Context.IntTy)
1323          if (MatchType(*T, ICE->getSubExpr()->getType(), true))
1324            return true;
1325
1326      S.Diag(getLocationOfByte(CS.getStart()),
1327             diag::warn_printf_conversion_argument_type_mismatch)
1328      << *T << Ex->getType()
1329      << getFormatSpecifierRange(startSpecifier, specifierLen)
1330      << Ex->getSourceRange();
1331    }
1332    return true;
1333  }
1334
1335  return true;
1336}
1337
1338void CheckPrintfHandler::DoneProcessing() {
1339  // Does the number of data arguments exceed the number of
1340  // format conversions in the format string?
1341  if (!HasVAListArg && NumConversions < NumDataArgs)
1342    S.Diag(getDataArg(NumConversions+1)->getLocStart(),
1343           diag::warn_printf_too_many_data_args)
1344      << getFormatStringRange();
1345}
1346
1347void Sema::CheckPrintfString(const StringLiteral *FExpr,
1348                             const Expr *OrigFormatExpr,
1349                             const CallExpr *TheCall, bool HasVAListArg,
1350                             unsigned format_idx, unsigned firstDataArg) {
1351
1352  // CHECK: is the format string a wide literal?
1353  if (FExpr->isWide()) {
1354    Diag(FExpr->getLocStart(),
1355         diag::warn_printf_format_string_is_wide_literal)
1356    << OrigFormatExpr->getSourceRange();
1357    return;
1358  }
1359
1360  // Str - The format string.  NOTE: this is NOT null-terminated!
1361  const char *Str = FExpr->getStrData();
1362
1363  // CHECK: empty format string?
1364  unsigned StrLen = FExpr->getByteLength();
1365
1366  if (StrLen == 0) {
1367    Diag(FExpr->getLocStart(), diag::warn_printf_empty_format_string)
1368    << OrigFormatExpr->getSourceRange();
1369    return;
1370  }
1371
1372  CheckPrintfHandler H(*this, FExpr, OrigFormatExpr,
1373                       TheCall->getNumArgs() - firstDataArg,
1374                       isa<ObjCStringLiteral>(OrigFormatExpr), Str,
1375                       HasVAListArg, TheCall, format_idx);
1376
1377  if (!analyze_printf::ParseFormatString(H, Str, Str + StrLen))
1378    H.DoneProcessing();
1379}
1380
1381//===--- CHECK: Return Address of Stack Variable --------------------------===//
1382
1383static DeclRefExpr* EvalVal(Expr *E);
1384static DeclRefExpr* EvalAddr(Expr* E);
1385
1386/// CheckReturnStackAddr - Check if a return statement returns the address
1387///   of a stack variable.
1388void
1389Sema::CheckReturnStackAddr(Expr *RetValExp, QualType lhsType,
1390                           SourceLocation ReturnLoc) {
1391
1392  // Perform checking for returned stack addresses.
1393  if (lhsType->isPointerType() || lhsType->isBlockPointerType()) {
1394    if (DeclRefExpr *DR = EvalAddr(RetValExp))
1395      Diag(DR->getLocStart(), diag::warn_ret_stack_addr)
1396       << DR->getDecl()->getDeclName() << RetValExp->getSourceRange();
1397
1398    // Skip over implicit cast expressions when checking for block expressions.
1399    RetValExp = RetValExp->IgnoreParenCasts();
1400
1401    if (BlockExpr *C = dyn_cast<BlockExpr>(RetValExp))
1402      if (C->hasBlockDeclRefExprs())
1403        Diag(C->getLocStart(), diag::err_ret_local_block)
1404          << C->getSourceRange();
1405
1406    if (AddrLabelExpr *ALE = dyn_cast<AddrLabelExpr>(RetValExp))
1407      Diag(ALE->getLocStart(), diag::warn_ret_addr_label)
1408        << ALE->getSourceRange();
1409
1410  } else if (lhsType->isReferenceType()) {
1411    // Perform checking for stack values returned by reference.
1412    // Check for a reference to the stack
1413    if (DeclRefExpr *DR = EvalVal(RetValExp))
1414      Diag(DR->getLocStart(), diag::warn_ret_stack_ref)
1415        << DR->getDecl()->getDeclName() << RetValExp->getSourceRange();
1416  }
1417}
1418
1419/// EvalAddr - EvalAddr and EvalVal are mutually recursive functions that
1420///  check if the expression in a return statement evaluates to an address
1421///  to a location on the stack.  The recursion is used to traverse the
1422///  AST of the return expression, with recursion backtracking when we
1423///  encounter a subexpression that (1) clearly does not lead to the address
1424///  of a stack variable or (2) is something we cannot determine leads to
1425///  the address of a stack variable based on such local checking.
1426///
1427///  EvalAddr processes expressions that are pointers that are used as
1428///  references (and not L-values).  EvalVal handles all other values.
1429///  At the base case of the recursion is a check for a DeclRefExpr* in
1430///  the refers to a stack variable.
1431///
1432///  This implementation handles:
1433///
1434///   * pointer-to-pointer casts
1435///   * implicit conversions from array references to pointers
1436///   * taking the address of fields
1437///   * arbitrary interplay between "&" and "*" operators
1438///   * pointer arithmetic from an address of a stack variable
1439///   * taking the address of an array element where the array is on the stack
1440static DeclRefExpr* EvalAddr(Expr *E) {
1441  // We should only be called for evaluating pointer expressions.
1442  assert((E->getType()->isAnyPointerType() ||
1443          E->getType()->isBlockPointerType() ||
1444          E->getType()->isObjCQualifiedIdType()) &&
1445         "EvalAddr only works on pointers");
1446
1447  // Our "symbolic interpreter" is just a dispatch off the currently
1448  // viewed AST node.  We then recursively traverse the AST by calling
1449  // EvalAddr and EvalVal appropriately.
1450  switch (E->getStmtClass()) {
1451  case Stmt::ParenExprClass:
1452    // Ignore parentheses.
1453    return EvalAddr(cast<ParenExpr>(E)->getSubExpr());
1454
1455  case Stmt::UnaryOperatorClass: {
1456    // The only unary operator that make sense to handle here
1457    // is AddrOf.  All others don't make sense as pointers.
1458    UnaryOperator *U = cast<UnaryOperator>(E);
1459
1460    if (U->getOpcode() == UnaryOperator::AddrOf)
1461      return EvalVal(U->getSubExpr());
1462    else
1463      return NULL;
1464  }
1465
1466  case Stmt::BinaryOperatorClass: {
1467    // Handle pointer arithmetic.  All other binary operators are not valid
1468    // in this context.
1469    BinaryOperator *B = cast<BinaryOperator>(E);
1470    BinaryOperator::Opcode op = B->getOpcode();
1471
1472    if (op != BinaryOperator::Add && op != BinaryOperator::Sub)
1473      return NULL;
1474
1475    Expr *Base = B->getLHS();
1476
1477    // Determine which argument is the real pointer base.  It could be
1478    // the RHS argument instead of the LHS.
1479    if (!Base->getType()->isPointerType()) Base = B->getRHS();
1480
1481    assert (Base->getType()->isPointerType());
1482    return EvalAddr(Base);
1483  }
1484
1485  // For conditional operators we need to see if either the LHS or RHS are
1486  // valid DeclRefExpr*s.  If one of them is valid, we return it.
1487  case Stmt::ConditionalOperatorClass: {
1488    ConditionalOperator *C = cast<ConditionalOperator>(E);
1489
1490    // Handle the GNU extension for missing LHS.
1491    if (Expr *lhsExpr = C->getLHS())
1492      if (DeclRefExpr* LHS = EvalAddr(lhsExpr))
1493        return LHS;
1494
1495     return EvalAddr(C->getRHS());
1496  }
1497
1498  // For casts, we need to handle conversions from arrays to
1499  // pointer values, and pointer-to-pointer conversions.
1500  case Stmt::ImplicitCastExprClass:
1501  case Stmt::CStyleCastExprClass:
1502  case Stmt::CXXFunctionalCastExprClass: {
1503    Expr* SubExpr = cast<CastExpr>(E)->getSubExpr();
1504    QualType T = SubExpr->getType();
1505
1506    if (SubExpr->getType()->isPointerType() ||
1507        SubExpr->getType()->isBlockPointerType() ||
1508        SubExpr->getType()->isObjCQualifiedIdType())
1509      return EvalAddr(SubExpr);
1510    else if (T->isArrayType())
1511      return EvalVal(SubExpr);
1512    else
1513      return 0;
1514  }
1515
1516  // C++ casts.  For dynamic casts, static casts, and const casts, we
1517  // are always converting from a pointer-to-pointer, so we just blow
1518  // through the cast.  In the case the dynamic cast doesn't fail (and
1519  // return NULL), we take the conservative route and report cases
1520  // where we return the address of a stack variable.  For Reinterpre
1521  // FIXME: The comment about is wrong; we're not always converting
1522  // from pointer to pointer. I'm guessing that this code should also
1523  // handle references to objects.
1524  case Stmt::CXXStaticCastExprClass:
1525  case Stmt::CXXDynamicCastExprClass:
1526  case Stmt::CXXConstCastExprClass:
1527  case Stmt::CXXReinterpretCastExprClass: {
1528      Expr *S = cast<CXXNamedCastExpr>(E)->getSubExpr();
1529      if (S->getType()->isPointerType() || S->getType()->isBlockPointerType())
1530        return EvalAddr(S);
1531      else
1532        return NULL;
1533  }
1534
1535  // Everything else: we simply don't reason about them.
1536  default:
1537    return NULL;
1538  }
1539}
1540
1541
1542///  EvalVal - This function is complements EvalAddr in the mutual recursion.
1543///   See the comments for EvalAddr for more details.
1544static DeclRefExpr* EvalVal(Expr *E) {
1545
1546  // We should only be called for evaluating non-pointer expressions, or
1547  // expressions with a pointer type that are not used as references but instead
1548  // are l-values (e.g., DeclRefExpr with a pointer type).
1549
1550  // Our "symbolic interpreter" is just a dispatch off the currently
1551  // viewed AST node.  We then recursively traverse the AST by calling
1552  // EvalAddr and EvalVal appropriately.
1553  switch (E->getStmtClass()) {
1554  case Stmt::DeclRefExprClass: {
1555    // DeclRefExpr: the base case.  When we hit a DeclRefExpr we are looking
1556    //  at code that refers to a variable's name.  We check if it has local
1557    //  storage within the function, and if so, return the expression.
1558    DeclRefExpr *DR = cast<DeclRefExpr>(E);
1559
1560    if (VarDecl *V = dyn_cast<VarDecl>(DR->getDecl()))
1561      if (V->hasLocalStorage() && !V->getType()->isReferenceType()) return DR;
1562
1563    return NULL;
1564  }
1565
1566  case Stmt::ParenExprClass:
1567    // Ignore parentheses.
1568    return EvalVal(cast<ParenExpr>(E)->getSubExpr());
1569
1570  case Stmt::UnaryOperatorClass: {
1571    // The only unary operator that make sense to handle here
1572    // is Deref.  All others don't resolve to a "name."  This includes
1573    // handling all sorts of rvalues passed to a unary operator.
1574    UnaryOperator *U = cast<UnaryOperator>(E);
1575
1576    if (U->getOpcode() == UnaryOperator::Deref)
1577      return EvalAddr(U->getSubExpr());
1578
1579    return NULL;
1580  }
1581
1582  case Stmt::ArraySubscriptExprClass: {
1583    // Array subscripts are potential references to data on the stack.  We
1584    // retrieve the DeclRefExpr* for the array variable if it indeed
1585    // has local storage.
1586    return EvalAddr(cast<ArraySubscriptExpr>(E)->getBase());
1587  }
1588
1589  case Stmt::ConditionalOperatorClass: {
1590    // For conditional operators we need to see if either the LHS or RHS are
1591    // non-NULL DeclRefExpr's.  If one is non-NULL, we return it.
1592    ConditionalOperator *C = cast<ConditionalOperator>(E);
1593
1594    // Handle the GNU extension for missing LHS.
1595    if (Expr *lhsExpr = C->getLHS())
1596      if (DeclRefExpr *LHS = EvalVal(lhsExpr))
1597        return LHS;
1598
1599    return EvalVal(C->getRHS());
1600  }
1601
1602  // Accesses to members are potential references to data on the stack.
1603  case Stmt::MemberExprClass: {
1604    MemberExpr *M = cast<MemberExpr>(E);
1605
1606    // Check for indirect access.  We only want direct field accesses.
1607    if (!M->isArrow())
1608      return EvalVal(M->getBase());
1609    else
1610      return NULL;
1611  }
1612
1613  // Everything else: we simply don't reason about them.
1614  default:
1615    return NULL;
1616  }
1617}
1618
1619//===--- CHECK: Floating-Point comparisons (-Wfloat-equal) ---------------===//
1620
1621/// Check for comparisons of floating point operands using != and ==.
1622/// Issue a warning if these are no self-comparisons, as they are not likely
1623/// to do what the programmer intended.
1624void Sema::CheckFloatComparison(SourceLocation loc, Expr* lex, Expr *rex) {
1625  bool EmitWarning = true;
1626
1627  Expr* LeftExprSansParen = lex->IgnoreParens();
1628  Expr* RightExprSansParen = rex->IgnoreParens();
1629
1630  // Special case: check for x == x (which is OK).
1631  // Do not emit warnings for such cases.
1632  if (DeclRefExpr* DRL = dyn_cast<DeclRefExpr>(LeftExprSansParen))
1633    if (DeclRefExpr* DRR = dyn_cast<DeclRefExpr>(RightExprSansParen))
1634      if (DRL->getDecl() == DRR->getDecl())
1635        EmitWarning = false;
1636
1637
1638  // Special case: check for comparisons against literals that can be exactly
1639  //  represented by APFloat.  In such cases, do not emit a warning.  This
1640  //  is a heuristic: often comparison against such literals are used to
1641  //  detect if a value in a variable has not changed.  This clearly can
1642  //  lead to false negatives.
1643  if (EmitWarning) {
1644    if (FloatingLiteral* FLL = dyn_cast<FloatingLiteral>(LeftExprSansParen)) {
1645      if (FLL->isExact())
1646        EmitWarning = false;
1647    } else
1648      if (FloatingLiteral* FLR = dyn_cast<FloatingLiteral>(RightExprSansParen)){
1649        if (FLR->isExact())
1650          EmitWarning = false;
1651    }
1652  }
1653
1654  // Check for comparisons with builtin types.
1655  if (EmitWarning)
1656    if (CallExpr* CL = dyn_cast<CallExpr>(LeftExprSansParen))
1657      if (CL->isBuiltinCall(Context))
1658        EmitWarning = false;
1659
1660  if (EmitWarning)
1661    if (CallExpr* CR = dyn_cast<CallExpr>(RightExprSansParen))
1662      if (CR->isBuiltinCall(Context))
1663        EmitWarning = false;
1664
1665  // Emit the diagnostic.
1666  if (EmitWarning)
1667    Diag(loc, diag::warn_floatingpoint_eq)
1668      << lex->getSourceRange() << rex->getSourceRange();
1669}
1670
1671//===--- CHECK: Integer mixed-sign comparisons (-Wsign-compare) --------===//
1672//===--- CHECK: Lossy implicit conversions (-Wconversion) --------------===//
1673
1674namespace {
1675
1676/// Structure recording the 'active' range of an integer-valued
1677/// expression.
1678struct IntRange {
1679  /// The number of bits active in the int.
1680  unsigned Width;
1681
1682  /// True if the int is known not to have negative values.
1683  bool NonNegative;
1684
1685  IntRange() {}
1686  IntRange(unsigned Width, bool NonNegative)
1687    : Width(Width), NonNegative(NonNegative)
1688  {}
1689
1690  // Returns the range of the bool type.
1691  static IntRange forBoolType() {
1692    return IntRange(1, true);
1693  }
1694
1695  // Returns the range of an integral type.
1696  static IntRange forType(ASTContext &C, QualType T) {
1697    return forCanonicalType(C, T->getCanonicalTypeInternal().getTypePtr());
1698  }
1699
1700  // Returns the range of an integeral type based on its canonical
1701  // representation.
1702  static IntRange forCanonicalType(ASTContext &C, const Type *T) {
1703    assert(T->isCanonicalUnqualified());
1704
1705    if (const VectorType *VT = dyn_cast<VectorType>(T))
1706      T = VT->getElementType().getTypePtr();
1707    if (const ComplexType *CT = dyn_cast<ComplexType>(T))
1708      T = CT->getElementType().getTypePtr();
1709    if (const EnumType *ET = dyn_cast<EnumType>(T))
1710      T = ET->getDecl()->getIntegerType().getTypePtr();
1711
1712    const BuiltinType *BT = cast<BuiltinType>(T);
1713    assert(BT->isInteger());
1714
1715    return IntRange(C.getIntWidth(QualType(T, 0)), BT->isUnsignedInteger());
1716  }
1717
1718  // Returns the supremum of two ranges: i.e. their conservative merge.
1719  static IntRange join(const IntRange &L, const IntRange &R) {
1720    return IntRange(std::max(L.Width, R.Width),
1721                    L.NonNegative && R.NonNegative);
1722  }
1723
1724  // Returns the infinum of two ranges: i.e. their aggressive merge.
1725  static IntRange meet(const IntRange &L, const IntRange &R) {
1726    return IntRange(std::min(L.Width, R.Width),
1727                    L.NonNegative || R.NonNegative);
1728  }
1729};
1730
1731IntRange GetValueRange(ASTContext &C, llvm::APSInt &value, unsigned MaxWidth) {
1732  if (value.isSigned() && value.isNegative())
1733    return IntRange(value.getMinSignedBits(), false);
1734
1735  if (value.getBitWidth() > MaxWidth)
1736    value.trunc(MaxWidth);
1737
1738  // isNonNegative() just checks the sign bit without considering
1739  // signedness.
1740  return IntRange(value.getActiveBits(), true);
1741}
1742
1743IntRange GetValueRange(ASTContext &C, APValue &result, QualType Ty,
1744                       unsigned MaxWidth) {
1745  if (result.isInt())
1746    return GetValueRange(C, result.getInt(), MaxWidth);
1747
1748  if (result.isVector()) {
1749    IntRange R = GetValueRange(C, result.getVectorElt(0), Ty, MaxWidth);
1750    for (unsigned i = 1, e = result.getVectorLength(); i != e; ++i) {
1751      IntRange El = GetValueRange(C, result.getVectorElt(i), Ty, MaxWidth);
1752      R = IntRange::join(R, El);
1753    }
1754    return R;
1755  }
1756
1757  if (result.isComplexInt()) {
1758    IntRange R = GetValueRange(C, result.getComplexIntReal(), MaxWidth);
1759    IntRange I = GetValueRange(C, result.getComplexIntImag(), MaxWidth);
1760    return IntRange::join(R, I);
1761  }
1762
1763  // This can happen with lossless casts to intptr_t of "based" lvalues.
1764  // Assume it might use arbitrary bits.
1765  // FIXME: The only reason we need to pass the type in here is to get
1766  // the sign right on this one case.  It would be nice if APValue
1767  // preserved this.
1768  assert(result.isLValue());
1769  return IntRange(MaxWidth, Ty->isUnsignedIntegerType());
1770}
1771
1772/// Pseudo-evaluate the given integer expression, estimating the
1773/// range of values it might take.
1774///
1775/// \param MaxWidth - the width to which the value will be truncated
1776IntRange GetExprRange(ASTContext &C, Expr *E, unsigned MaxWidth) {
1777  E = E->IgnoreParens();
1778
1779  // Try a full evaluation first.
1780  Expr::EvalResult result;
1781  if (E->Evaluate(result, C))
1782    return GetValueRange(C, result.Val, E->getType(), MaxWidth);
1783
1784  // I think we only want to look through implicit casts here; if the
1785  // user has an explicit widening cast, we should treat the value as
1786  // being of the new, wider type.
1787  if (ImplicitCastExpr *CE = dyn_cast<ImplicitCastExpr>(E)) {
1788    if (CE->getCastKind() == CastExpr::CK_NoOp)
1789      return GetExprRange(C, CE->getSubExpr(), MaxWidth);
1790
1791    IntRange OutputTypeRange = IntRange::forType(C, CE->getType());
1792
1793    bool isIntegerCast = (CE->getCastKind() == CastExpr::CK_IntegralCast);
1794    if (!isIntegerCast && CE->getCastKind() == CastExpr::CK_Unknown)
1795      isIntegerCast = CE->getSubExpr()->getType()->isIntegerType();
1796
1797    // Assume that non-integer casts can span the full range of the type.
1798    if (!isIntegerCast)
1799      return OutputTypeRange;
1800
1801    IntRange SubRange
1802      = GetExprRange(C, CE->getSubExpr(),
1803                     std::min(MaxWidth, OutputTypeRange.Width));
1804
1805    // Bail out if the subexpr's range is as wide as the cast type.
1806    if (SubRange.Width >= OutputTypeRange.Width)
1807      return OutputTypeRange;
1808
1809    // Otherwise, we take the smaller width, and we're non-negative if
1810    // either the output type or the subexpr is.
1811    return IntRange(SubRange.Width,
1812                    SubRange.NonNegative || OutputTypeRange.NonNegative);
1813  }
1814
1815  if (ConditionalOperator *CO = dyn_cast<ConditionalOperator>(E)) {
1816    // If we can fold the condition, just take that operand.
1817    bool CondResult;
1818    if (CO->getCond()->EvaluateAsBooleanCondition(CondResult, C))
1819      return GetExprRange(C, CondResult ? CO->getTrueExpr()
1820                                        : CO->getFalseExpr(),
1821                          MaxWidth);
1822
1823    // Otherwise, conservatively merge.
1824    IntRange L = GetExprRange(C, CO->getTrueExpr(), MaxWidth);
1825    IntRange R = GetExprRange(C, CO->getFalseExpr(), MaxWidth);
1826    return IntRange::join(L, R);
1827  }
1828
1829  if (BinaryOperator *BO = dyn_cast<BinaryOperator>(E)) {
1830    switch (BO->getOpcode()) {
1831
1832    // Boolean-valued operations are single-bit and positive.
1833    case BinaryOperator::LAnd:
1834    case BinaryOperator::LOr:
1835    case BinaryOperator::LT:
1836    case BinaryOperator::GT:
1837    case BinaryOperator::LE:
1838    case BinaryOperator::GE:
1839    case BinaryOperator::EQ:
1840    case BinaryOperator::NE:
1841      return IntRange::forBoolType();
1842
1843    // Operations with opaque sources are black-listed.
1844    case BinaryOperator::PtrMemD:
1845    case BinaryOperator::PtrMemI:
1846      return IntRange::forType(C, E->getType());
1847
1848    // Bitwise-and uses the *infinum* of the two source ranges.
1849    case BinaryOperator::And:
1850      return IntRange::meet(GetExprRange(C, BO->getLHS(), MaxWidth),
1851                            GetExprRange(C, BO->getRHS(), MaxWidth));
1852
1853    // Left shift gets black-listed based on a judgement call.
1854    case BinaryOperator::Shl:
1855      return IntRange::forType(C, E->getType());
1856
1857    // Right shift by a constant can narrow its left argument.
1858    case BinaryOperator::Shr: {
1859      IntRange L = GetExprRange(C, BO->getLHS(), MaxWidth);
1860
1861      // If the shift amount is a positive constant, drop the width by
1862      // that much.
1863      llvm::APSInt shift;
1864      if (BO->getRHS()->isIntegerConstantExpr(shift, C) &&
1865          shift.isNonNegative()) {
1866        unsigned zext = shift.getZExtValue();
1867        if (zext >= L.Width)
1868          L.Width = (L.NonNegative ? 0 : 1);
1869        else
1870          L.Width -= zext;
1871      }
1872
1873      return L;
1874    }
1875
1876    // Comma acts as its right operand.
1877    case BinaryOperator::Comma:
1878      return GetExprRange(C, BO->getRHS(), MaxWidth);
1879
1880    // Black-list pointer subtractions.
1881    case BinaryOperator::Sub:
1882      if (BO->getLHS()->getType()->isPointerType())
1883        return IntRange::forType(C, E->getType());
1884      // fallthrough
1885
1886    default:
1887      break;
1888    }
1889
1890    // Treat every other operator as if it were closed on the
1891    // narrowest type that encompasses both operands.
1892    IntRange L = GetExprRange(C, BO->getLHS(), MaxWidth);
1893    IntRange R = GetExprRange(C, BO->getRHS(), MaxWidth);
1894    return IntRange::join(L, R);
1895  }
1896
1897  if (UnaryOperator *UO = dyn_cast<UnaryOperator>(E)) {
1898    switch (UO->getOpcode()) {
1899    // Boolean-valued operations are white-listed.
1900    case UnaryOperator::LNot:
1901      return IntRange::forBoolType();
1902
1903    // Operations with opaque sources are black-listed.
1904    case UnaryOperator::Deref:
1905    case UnaryOperator::AddrOf: // should be impossible
1906    case UnaryOperator::OffsetOf:
1907      return IntRange::forType(C, E->getType());
1908
1909    default:
1910      return GetExprRange(C, UO->getSubExpr(), MaxWidth);
1911    }
1912  }
1913
1914  FieldDecl *BitField = E->getBitField();
1915  if (BitField) {
1916    llvm::APSInt BitWidthAP = BitField->getBitWidth()->EvaluateAsInt(C);
1917    unsigned BitWidth = BitWidthAP.getZExtValue();
1918
1919    return IntRange(BitWidth, BitField->getType()->isUnsignedIntegerType());
1920  }
1921
1922  return IntRange::forType(C, E->getType());
1923}
1924
1925/// Checks whether the given value, which currently has the given
1926/// source semantics, has the same value when coerced through the
1927/// target semantics.
1928bool IsSameFloatAfterCast(const llvm::APFloat &value,
1929                          const llvm::fltSemantics &Src,
1930                          const llvm::fltSemantics &Tgt) {
1931  llvm::APFloat truncated = value;
1932
1933  bool ignored;
1934  truncated.convert(Src, llvm::APFloat::rmNearestTiesToEven, &ignored);
1935  truncated.convert(Tgt, llvm::APFloat::rmNearestTiesToEven, &ignored);
1936
1937  return truncated.bitwiseIsEqual(value);
1938}
1939
1940/// Checks whether the given value, which currently has the given
1941/// source semantics, has the same value when coerced through the
1942/// target semantics.
1943///
1944/// The value might be a vector of floats (or a complex number).
1945bool IsSameFloatAfterCast(const APValue &value,
1946                          const llvm::fltSemantics &Src,
1947                          const llvm::fltSemantics &Tgt) {
1948  if (value.isFloat())
1949    return IsSameFloatAfterCast(value.getFloat(), Src, Tgt);
1950
1951  if (value.isVector()) {
1952    for (unsigned i = 0, e = value.getVectorLength(); i != e; ++i)
1953      if (!IsSameFloatAfterCast(value.getVectorElt(i), Src, Tgt))
1954        return false;
1955    return true;
1956  }
1957
1958  assert(value.isComplexFloat());
1959  return (IsSameFloatAfterCast(value.getComplexFloatReal(), Src, Tgt) &&
1960          IsSameFloatAfterCast(value.getComplexFloatImag(), Src, Tgt));
1961}
1962
1963} // end anonymous namespace
1964
1965/// \brief Implements -Wsign-compare.
1966///
1967/// \param lex the left-hand expression
1968/// \param rex the right-hand expression
1969/// \param OpLoc the location of the joining operator
1970/// \param Equality whether this is an "equality-like" join, which
1971///   suppresses the warning in some cases
1972void Sema::CheckSignCompare(Expr *lex, Expr *rex, SourceLocation OpLoc,
1973                            const PartialDiagnostic &PD, bool Equality) {
1974  // Don't warn if we're in an unevaluated context.
1975  if (ExprEvalContexts.back().Context == Unevaluated)
1976    return;
1977
1978  // If either expression is value-dependent, don't warn. We'll get another
1979  // chance at instantiation time.
1980  if (lex->isValueDependent() || rex->isValueDependent())
1981    return;
1982
1983  QualType lt = lex->getType(), rt = rex->getType();
1984
1985  // Only warn if both operands are integral.
1986  if (!lt->isIntegerType() || !rt->isIntegerType())
1987    return;
1988
1989  // In C, the width of a bitfield determines its type, and the
1990  // declared type only contributes the signedness.  This duplicates
1991  // the work that will later be done by UsualUnaryConversions.
1992  // Eventually, this check will be reorganized in a way that avoids
1993  // this duplication.
1994  if (!getLangOptions().CPlusPlus) {
1995    QualType tmp;
1996    tmp = Context.isPromotableBitField(lex);
1997    if (!tmp.isNull()) lt = tmp;
1998    tmp = Context.isPromotableBitField(rex);
1999    if (!tmp.isNull()) rt = tmp;
2000  }
2001
2002  // The rule is that the signed operand becomes unsigned, so isolate the
2003  // signed operand.
2004  Expr *signedOperand = lex, *unsignedOperand = rex;
2005  QualType signedType = lt, unsignedType = rt;
2006  if (lt->isSignedIntegerType()) {
2007    if (rt->isSignedIntegerType()) return;
2008  } else {
2009    if (!rt->isSignedIntegerType()) return;
2010    std::swap(signedOperand, unsignedOperand);
2011    std::swap(signedType, unsignedType);
2012  }
2013
2014  unsigned unsignedWidth = Context.getIntWidth(unsignedType);
2015  unsigned signedWidth = Context.getIntWidth(signedType);
2016
2017  // If the unsigned type is strictly smaller than the signed type,
2018  // then (1) the result type will be signed and (2) the unsigned
2019  // value will fit fully within the signed type, and thus the result
2020  // of the comparison will be exact.
2021  if (signedWidth > unsignedWidth)
2022    return;
2023
2024  // Otherwise, calculate the effective ranges.
2025  IntRange signedRange = GetExprRange(Context, signedOperand, signedWidth);
2026  IntRange unsignedRange = GetExprRange(Context, unsignedOperand, unsignedWidth);
2027
2028  // We should never be unable to prove that the unsigned operand is
2029  // non-negative.
2030  assert(unsignedRange.NonNegative && "unsigned range includes negative?");
2031
2032  // If the signed operand is non-negative, then the signed->unsigned
2033  // conversion won't change it.
2034  if (signedRange.NonNegative)
2035    return;
2036
2037  // For (in)equality comparisons, if the unsigned operand is a
2038  // constant which cannot collide with a overflowed signed operand,
2039  // then reinterpreting the signed operand as unsigned will not
2040  // change the result of the comparison.
2041  if (Equality && unsignedRange.Width < unsignedWidth)
2042    return;
2043
2044  Diag(OpLoc, PD)
2045    << lt << rt << lex->getSourceRange() << rex->getSourceRange();
2046}
2047
2048/// Diagnose an implicit cast;  purely a helper for CheckImplicitConversion.
2049static void DiagnoseImpCast(Sema &S, Expr *E, QualType T, unsigned diag) {
2050  S.Diag(E->getExprLoc(), diag) << E->getType() << T << E->getSourceRange();
2051}
2052
2053/// Implements -Wconversion.
2054void Sema::CheckImplicitConversion(Expr *E, QualType T) {
2055  // Don't diagnose in unevaluated contexts.
2056  if (ExprEvalContexts.back().Context == Sema::Unevaluated)
2057    return;
2058
2059  // Don't diagnose for value-dependent expressions.
2060  if (E->isValueDependent())
2061    return;
2062
2063  const Type *Source = Context.getCanonicalType(E->getType()).getTypePtr();
2064  const Type *Target = Context.getCanonicalType(T).getTypePtr();
2065
2066  // Never diagnose implicit casts to bool.
2067  if (Target->isSpecificBuiltinType(BuiltinType::Bool))
2068    return;
2069
2070  // Strip vector types.
2071  if (isa<VectorType>(Source)) {
2072    if (!isa<VectorType>(Target))
2073      return DiagnoseImpCast(*this, E, T, diag::warn_impcast_vector_scalar);
2074
2075    Source = cast<VectorType>(Source)->getElementType().getTypePtr();
2076    Target = cast<VectorType>(Target)->getElementType().getTypePtr();
2077  }
2078
2079  // Strip complex types.
2080  if (isa<ComplexType>(Source)) {
2081    if (!isa<ComplexType>(Target))
2082      return DiagnoseImpCast(*this, E, T, diag::warn_impcast_complex_scalar);
2083
2084    Source = cast<ComplexType>(Source)->getElementType().getTypePtr();
2085    Target = cast<ComplexType>(Target)->getElementType().getTypePtr();
2086  }
2087
2088  const BuiltinType *SourceBT = dyn_cast<BuiltinType>(Source);
2089  const BuiltinType *TargetBT = dyn_cast<BuiltinType>(Target);
2090
2091  // If the source is floating point...
2092  if (SourceBT && SourceBT->isFloatingPoint()) {
2093    // ...and the target is floating point...
2094    if (TargetBT && TargetBT->isFloatingPoint()) {
2095      // ...then warn if we're dropping FP rank.
2096
2097      // Builtin FP kinds are ordered by increasing FP rank.
2098      if (SourceBT->getKind() > TargetBT->getKind()) {
2099        // Don't warn about float constants that are precisely
2100        // representable in the target type.
2101        Expr::EvalResult result;
2102        if (E->Evaluate(result, Context)) {
2103          // Value might be a float, a float vector, or a float complex.
2104          if (IsSameFloatAfterCast(result.Val,
2105                     Context.getFloatTypeSemantics(QualType(TargetBT, 0)),
2106                     Context.getFloatTypeSemantics(QualType(SourceBT, 0))))
2107            return;
2108        }
2109
2110        DiagnoseImpCast(*this, E, T, diag::warn_impcast_float_precision);
2111      }
2112      return;
2113    }
2114
2115    // If the target is integral, always warn.
2116    if ((TargetBT && TargetBT->isInteger()))
2117      // TODO: don't warn for integer values?
2118      return DiagnoseImpCast(*this, E, T, diag::warn_impcast_float_integer);
2119
2120    return;
2121  }
2122
2123  if (!Source->isIntegerType() || !Target->isIntegerType())
2124    return;
2125
2126  IntRange SourceRange = GetExprRange(Context, E, Context.getIntWidth(E->getType()));
2127  IntRange TargetRange = IntRange::forCanonicalType(Context, Target);
2128
2129  // FIXME: also signed<->unsigned?
2130
2131  if (SourceRange.Width > TargetRange.Width) {
2132    // People want to build with -Wshorten-64-to-32 and not -Wconversion
2133    // and by god we'll let them.
2134    if (SourceRange.Width == 64 && TargetRange.Width == 32)
2135      return DiagnoseImpCast(*this, E, T, diag::warn_impcast_integer_64_32);
2136    return DiagnoseImpCast(*this, E, T, diag::warn_impcast_integer_precision);
2137  }
2138
2139  return;
2140}
2141
2142// MarkLive - Mark all the blocks reachable from e as live.  Returns the total
2143// number of blocks just marked live.
2144static unsigned MarkLive(CFGBlock *e, llvm::BitVector &live) {
2145  unsigned count = 0;
2146  std::queue<CFGBlock*> workq;
2147  // Prep work queue
2148  live.set(e->getBlockID());
2149  ++count;
2150  workq.push(e);
2151  // Solve
2152  while (!workq.empty()) {
2153    CFGBlock *item = workq.front();
2154    workq.pop();
2155    for (CFGBlock::succ_iterator I=item->succ_begin(),
2156           E=item->succ_end();
2157         I != E;
2158         ++I) {
2159      if ((*I) && !live[(*I)->getBlockID()]) {
2160        live.set((*I)->getBlockID());
2161        ++count;
2162        workq.push(*I);
2163      }
2164    }
2165  }
2166  return count;
2167}
2168
2169static SourceLocation GetUnreachableLoc(CFGBlock &b, SourceRange &R1,
2170                                        SourceRange &R2) {
2171  Stmt *S;
2172  unsigned sn = 0;
2173  R1 = R2 = SourceRange();
2174
2175  top:
2176  if (sn < b.size())
2177    S = b[sn].getStmt();
2178  else if (b.getTerminator())
2179    S = b.getTerminator();
2180  else
2181    return SourceLocation();
2182
2183  switch (S->getStmtClass()) {
2184  case Expr::BinaryOperatorClass: {
2185    BinaryOperator *BO = cast<BinaryOperator>(S);
2186    if (BO->getOpcode() == BinaryOperator::Comma) {
2187      if (sn+1 < b.size())
2188        return b[sn+1].getStmt()->getLocStart();
2189      CFGBlock *n = &b;
2190      while (1) {
2191        if (n->getTerminator())
2192          return n->getTerminator()->getLocStart();
2193        if (n->succ_size() != 1)
2194          return SourceLocation();
2195        n = n[0].succ_begin()[0];
2196        if (n->pred_size() != 1)
2197          return SourceLocation();
2198        if (!n->empty())
2199          return n[0][0].getStmt()->getLocStart();
2200      }
2201    }
2202    R1 = BO->getLHS()->getSourceRange();
2203    R2 = BO->getRHS()->getSourceRange();
2204    return BO->getOperatorLoc();
2205  }
2206  case Expr::UnaryOperatorClass: {
2207    const UnaryOperator *UO = cast<UnaryOperator>(S);
2208    R1 = UO->getSubExpr()->getSourceRange();
2209    return UO->getOperatorLoc();
2210  }
2211  case Expr::CompoundAssignOperatorClass: {
2212    const CompoundAssignOperator *CAO = cast<CompoundAssignOperator>(S);
2213    R1 = CAO->getLHS()->getSourceRange();
2214    R2 = CAO->getRHS()->getSourceRange();
2215    return CAO->getOperatorLoc();
2216  }
2217  case Expr::ConditionalOperatorClass: {
2218    const ConditionalOperator *CO = cast<ConditionalOperator>(S);
2219    return CO->getQuestionLoc();
2220  }
2221  case Expr::MemberExprClass: {
2222    const MemberExpr *ME = cast<MemberExpr>(S);
2223    R1 = ME->getSourceRange();
2224    return ME->getMemberLoc();
2225  }
2226  case Expr::ArraySubscriptExprClass: {
2227    const ArraySubscriptExpr *ASE = cast<ArraySubscriptExpr>(S);
2228    R1 = ASE->getLHS()->getSourceRange();
2229    R2 = ASE->getRHS()->getSourceRange();
2230    return ASE->getRBracketLoc();
2231  }
2232  case Expr::CStyleCastExprClass: {
2233    const CStyleCastExpr *CSC = cast<CStyleCastExpr>(S);
2234    R1 = CSC->getSubExpr()->getSourceRange();
2235    return CSC->getLParenLoc();
2236  }
2237  case Expr::CXXFunctionalCastExprClass: {
2238    const CXXFunctionalCastExpr *CE = cast <CXXFunctionalCastExpr>(S);
2239    R1 = CE->getSubExpr()->getSourceRange();
2240    return CE->getTypeBeginLoc();
2241  }
2242  case Expr::ImplicitCastExprClass:
2243    ++sn;
2244    goto top;
2245  case Stmt::CXXTryStmtClass: {
2246    return cast<CXXTryStmt>(S)->getHandler(0)->getCatchLoc();
2247  }
2248  default: ;
2249  }
2250  R1 = S->getSourceRange();
2251  return S->getLocStart();
2252}
2253
2254static SourceLocation MarkLiveTop(CFGBlock *e, llvm::BitVector &live,
2255                               SourceManager &SM) {
2256  std::queue<CFGBlock*> workq;
2257  // Prep work queue
2258  workq.push(e);
2259  SourceRange R1, R2;
2260  SourceLocation top = GetUnreachableLoc(*e, R1, R2);
2261  bool FromMainFile = false;
2262  bool FromSystemHeader = false;
2263  bool TopValid = false;
2264  if (top.isValid()) {
2265    FromMainFile = SM.isFromMainFile(top);
2266    FromSystemHeader = SM.isInSystemHeader(top);
2267    TopValid = true;
2268  }
2269  // Solve
2270  while (!workq.empty()) {
2271    CFGBlock *item = workq.front();
2272    workq.pop();
2273    SourceLocation c = GetUnreachableLoc(*item, R1, R2);
2274    if (c.isValid()
2275        && (!TopValid
2276            || (SM.isFromMainFile(c) && !FromMainFile)
2277            || (FromSystemHeader && !SM.isInSystemHeader(c))
2278            || SM.isBeforeInTranslationUnit(c, top))) {
2279      top = c;
2280      FromMainFile = SM.isFromMainFile(top);
2281      FromSystemHeader = SM.isInSystemHeader(top);
2282    }
2283    live.set(item->getBlockID());
2284    for (CFGBlock::succ_iterator I=item->succ_begin(),
2285           E=item->succ_end();
2286         I != E;
2287         ++I) {
2288      if ((*I) && !live[(*I)->getBlockID()]) {
2289        live.set((*I)->getBlockID());
2290        workq.push(*I);
2291      }
2292    }
2293  }
2294  return top;
2295}
2296
2297static int LineCmp(const void *p1, const void *p2) {
2298  SourceLocation *Line1 = (SourceLocation *)p1;
2299  SourceLocation *Line2 = (SourceLocation *)p2;
2300  return !(*Line1 < *Line2);
2301}
2302
2303namespace {
2304  struct ErrLoc {
2305    SourceLocation Loc;
2306    SourceRange R1;
2307    SourceRange R2;
2308    ErrLoc(SourceLocation l, SourceRange r1, SourceRange r2)
2309      : Loc(l), R1(r1), R2(r2) { }
2310  };
2311}
2312
2313/// CheckUnreachable - Check for unreachable code.
2314void Sema::CheckUnreachable(AnalysisContext &AC) {
2315  unsigned count;
2316  // We avoid checking when there are errors, as the CFG won't faithfully match
2317  // the user's code.
2318  if (getDiagnostics().hasErrorOccurred())
2319    return;
2320  if (Diags.getDiagnosticLevel(diag::warn_unreachable) == Diagnostic::Ignored)
2321    return;
2322
2323  CFG *cfg = AC.getCFG();
2324  if (cfg == 0)
2325    return;
2326
2327  llvm::BitVector live(cfg->getNumBlockIDs());
2328  // Mark all live things first.
2329  count = MarkLive(&cfg->getEntry(), live);
2330
2331  if (count == cfg->getNumBlockIDs())
2332    // If there are no dead blocks, we're done.
2333    return;
2334
2335  SourceRange R1, R2;
2336
2337  llvm::SmallVector<ErrLoc, 24> lines;
2338  bool AddEHEdges = AC.getAddEHEdges();
2339  // First, give warnings for blocks with no predecessors, as they
2340  // can't be part of a loop.
2341  for (CFG::iterator I = cfg->begin(), E = cfg->end(); I != E; ++I) {
2342    CFGBlock &b = **I;
2343    if (!live[b.getBlockID()]) {
2344      if (b.pred_begin() == b.pred_end()) {
2345        if (!AddEHEdges && b.getTerminator()
2346            && isa<CXXTryStmt>(b.getTerminator())) {
2347          // When not adding EH edges from calls, catch clauses
2348          // can otherwise seem dead.  Avoid noting them as dead.
2349          count += MarkLive(&b, live);
2350          continue;
2351        }
2352        SourceLocation c = GetUnreachableLoc(b, R1, R2);
2353        if (!c.isValid()) {
2354          // Blocks without a location can't produce a warning, so don't mark
2355          // reachable blocks from here as live.
2356          live.set(b.getBlockID());
2357          ++count;
2358          continue;
2359        }
2360        lines.push_back(ErrLoc(c, R1, R2));
2361        // Avoid excessive errors by marking everything reachable from here
2362        count += MarkLive(&b, live);
2363      }
2364    }
2365  }
2366
2367  if (count < cfg->getNumBlockIDs()) {
2368    // And then give warnings for the tops of loops.
2369    for (CFG::iterator I = cfg->begin(), E = cfg->end(); I != E; ++I) {
2370      CFGBlock &b = **I;
2371      if (!live[b.getBlockID()])
2372        // Avoid excessive errors by marking everything reachable from here
2373        lines.push_back(ErrLoc(MarkLiveTop(&b, live,
2374                                           Context.getSourceManager()),
2375                               SourceRange(), SourceRange()));
2376    }
2377  }
2378
2379  llvm::array_pod_sort(lines.begin(), lines.end(), LineCmp);
2380  for (llvm::SmallVector<ErrLoc, 24>::iterator I = lines.begin(),
2381         E = lines.end();
2382       I != E;
2383       ++I)
2384    if (I->Loc.isValid())
2385      Diag(I->Loc, diag::warn_unreachable) << I->R1 << I->R2;
2386}
2387
2388/// CheckFallThrough - Check that we don't fall off the end of a
2389/// Statement that should return a value.
2390///
2391/// \returns AlwaysFallThrough iff we always fall off the end of the statement,
2392/// MaybeFallThrough iff we might or might not fall off the end,
2393/// NeverFallThroughOrReturn iff we never fall off the end of the statement or
2394/// return.  We assume NeverFallThrough iff we never fall off the end of the
2395/// statement but we may return.  We assume that functions not marked noreturn
2396/// will return.
2397Sema::ControlFlowKind Sema::CheckFallThrough(AnalysisContext &AC) {
2398  CFG *cfg = AC.getCFG();
2399  if (cfg == 0)
2400    // FIXME: This should be NeverFallThrough
2401    return NeverFallThroughOrReturn;
2402
2403  // The CFG leaves in dead things, and we don't want the dead code paths to
2404  // confuse us, so we mark all live things first.
2405  std::queue<CFGBlock*> workq;
2406  llvm::BitVector live(cfg->getNumBlockIDs());
2407  unsigned count = MarkLive(&cfg->getEntry(), live);
2408
2409  bool AddEHEdges = AC.getAddEHEdges();
2410  if (!AddEHEdges && count != cfg->getNumBlockIDs())
2411    // When there are things remaining dead, and we didn't add EH edges
2412    // from CallExprs to the catch clauses, we have to go back and
2413    // mark them as live.
2414    for (CFG::iterator I = cfg->begin(), E = cfg->end(); I != E; ++I) {
2415      CFGBlock &b = **I;
2416      if (!live[b.getBlockID()]) {
2417        if (b.pred_begin() == b.pred_end()) {
2418          if (b.getTerminator() && isa<CXXTryStmt>(b.getTerminator()))
2419            // When not adding EH edges from calls, catch clauses
2420            // can otherwise seem dead.  Avoid noting them as dead.
2421            count += MarkLive(&b, live);
2422          continue;
2423        }
2424      }
2425    }
2426
2427  // Now we know what is live, we check the live precessors of the exit block
2428  // and look for fall through paths, being careful to ignore normal returns,
2429  // and exceptional paths.
2430  bool HasLiveReturn = false;
2431  bool HasFakeEdge = false;
2432  bool HasPlainEdge = false;
2433  bool HasAbnormalEdge = false;
2434  for (CFGBlock::pred_iterator I=cfg->getExit().pred_begin(),
2435         E = cfg->getExit().pred_end();
2436       I != E;
2437       ++I) {
2438    CFGBlock& B = **I;
2439    if (!live[B.getBlockID()])
2440      continue;
2441    if (B.size() == 0) {
2442      if (B.getTerminator() && isa<CXXTryStmt>(B.getTerminator())) {
2443        HasAbnormalEdge = true;
2444        continue;
2445      }
2446
2447      // A labeled empty statement, or the entry block...
2448      HasPlainEdge = true;
2449      continue;
2450    }
2451    Stmt *S = B[B.size()-1];
2452    if (isa<ReturnStmt>(S)) {
2453      HasLiveReturn = true;
2454      continue;
2455    }
2456    if (isa<ObjCAtThrowStmt>(S)) {
2457      HasFakeEdge = true;
2458      continue;
2459    }
2460    if (isa<CXXThrowExpr>(S)) {
2461      HasFakeEdge = true;
2462      continue;
2463    }
2464    if (const AsmStmt *AS = dyn_cast<AsmStmt>(S)) {
2465      if (AS->isMSAsm()) {
2466        HasFakeEdge = true;
2467        HasLiveReturn = true;
2468        continue;
2469      }
2470    }
2471    if (isa<CXXTryStmt>(S)) {
2472      HasAbnormalEdge = true;
2473      continue;
2474    }
2475
2476    bool NoReturnEdge = false;
2477    if (CallExpr *C = dyn_cast<CallExpr>(S)) {
2478      if (B.succ_begin()[0] != &cfg->getExit()) {
2479        HasAbnormalEdge = true;
2480        continue;
2481      }
2482      Expr *CEE = C->getCallee()->IgnoreParenCasts();
2483      if (CEE->getType().getNoReturnAttr()) {
2484        NoReturnEdge = true;
2485        HasFakeEdge = true;
2486      } else if (DeclRefExpr *DRE = dyn_cast<DeclRefExpr>(CEE)) {
2487        ValueDecl *VD = DRE->getDecl();
2488        if (VD->hasAttr<NoReturnAttr>()) {
2489          NoReturnEdge = true;
2490          HasFakeEdge = true;
2491        }
2492      }
2493    }
2494    // FIXME: Add noreturn message sends.
2495    if (NoReturnEdge == false)
2496      HasPlainEdge = true;
2497  }
2498  if (!HasPlainEdge) {
2499    if (HasLiveReturn)
2500      return NeverFallThrough;
2501    return NeverFallThroughOrReturn;
2502  }
2503  if (HasAbnormalEdge || HasFakeEdge || HasLiveReturn)
2504    return MaybeFallThrough;
2505  // This says AlwaysFallThrough for calls to functions that are not marked
2506  // noreturn, that don't return.  If people would like this warning to be more
2507  // accurate, such functions should be marked as noreturn.
2508  return AlwaysFallThrough;
2509}
2510
2511/// CheckFallThroughForFunctionDef - Check that we don't fall off the end of a
2512/// function that should return a value.  Check that we don't fall off the end
2513/// of a noreturn function.  We assume that functions and blocks not marked
2514/// noreturn will return.
2515void Sema::CheckFallThroughForFunctionDef(Decl *D, Stmt *Body,
2516                                          AnalysisContext &AC) {
2517  // FIXME: Would be nice if we had a better way to control cascading errors,
2518  // but for now, avoid them.  The problem is that when Parse sees:
2519  //   int foo() { return a; }
2520  // The return is eaten and the Sema code sees just:
2521  //   int foo() { }
2522  // which this code would then warn about.
2523  if (getDiagnostics().hasErrorOccurred())
2524    return;
2525
2526  bool ReturnsVoid = false;
2527  bool HasNoReturn = false;
2528  if (FunctionDecl *FD = dyn_cast<FunctionDecl>(D)) {
2529    // For function templates, class templates and member function templates
2530    // we'll do the analysis at instantiation time.
2531    if (FD->isDependentContext())
2532      return;
2533
2534    if (FD->getResultType()->isVoidType())
2535      ReturnsVoid = true;
2536    if (FD->hasAttr<NoReturnAttr>() ||
2537        FD->getType()->getAs<FunctionType>()->getNoReturnAttr())
2538      HasNoReturn = true;
2539  } else if (ObjCMethodDecl *MD = dyn_cast<ObjCMethodDecl>(D)) {
2540    if (MD->getResultType()->isVoidType())
2541      ReturnsVoid = true;
2542    if (MD->hasAttr<NoReturnAttr>())
2543      HasNoReturn = true;
2544  }
2545
2546  // Short circuit for compilation speed.
2547  if ((Diags.getDiagnosticLevel(diag::warn_maybe_falloff_nonvoid_function)
2548       == Diagnostic::Ignored || ReturnsVoid)
2549      && (Diags.getDiagnosticLevel(diag::warn_noreturn_function_has_return_expr)
2550          == Diagnostic::Ignored || !HasNoReturn)
2551      && (Diags.getDiagnosticLevel(diag::warn_suggest_noreturn_block)
2552          == Diagnostic::Ignored || !ReturnsVoid))
2553    return;
2554  // FIXME: Function try block
2555  if (CompoundStmt *Compound = dyn_cast<CompoundStmt>(Body)) {
2556    switch (CheckFallThrough(AC)) {
2557    case MaybeFallThrough:
2558      if (HasNoReturn)
2559        Diag(Compound->getRBracLoc(), diag::warn_falloff_noreturn_function);
2560      else if (!ReturnsVoid)
2561        Diag(Compound->getRBracLoc(),diag::warn_maybe_falloff_nonvoid_function);
2562      break;
2563    case AlwaysFallThrough:
2564      if (HasNoReturn)
2565        Diag(Compound->getRBracLoc(), diag::warn_falloff_noreturn_function);
2566      else if (!ReturnsVoid)
2567        Diag(Compound->getRBracLoc(), diag::warn_falloff_nonvoid_function);
2568      break;
2569    case NeverFallThroughOrReturn:
2570      if (ReturnsVoid && !HasNoReturn)
2571        Diag(Compound->getLBracLoc(), diag::warn_suggest_noreturn_function);
2572      break;
2573    case NeverFallThrough:
2574      break;
2575    }
2576  }
2577}
2578
2579/// CheckFallThroughForBlock - Check that we don't fall off the end of a block
2580/// that should return a value.  Check that we don't fall off the end of a
2581/// noreturn block.  We assume that functions and blocks not marked noreturn
2582/// will return.
2583void Sema::CheckFallThroughForBlock(QualType BlockTy, Stmt *Body,
2584                                    AnalysisContext &AC) {
2585  // FIXME: Would be nice if we had a better way to control cascading errors,
2586  // but for now, avoid them.  The problem is that when Parse sees:
2587  //   int foo() { return a; }
2588  // The return is eaten and the Sema code sees just:
2589  //   int foo() { }
2590  // which this code would then warn about.
2591  if (getDiagnostics().hasErrorOccurred())
2592    return;
2593  bool ReturnsVoid = false;
2594  bool HasNoReturn = false;
2595  if (const FunctionType *FT =BlockTy->getPointeeType()->getAs<FunctionType>()){
2596    if (FT->getResultType()->isVoidType())
2597      ReturnsVoid = true;
2598    if (FT->getNoReturnAttr())
2599      HasNoReturn = true;
2600  }
2601
2602  // Short circuit for compilation speed.
2603  if (ReturnsVoid
2604      && !HasNoReturn
2605      && (Diags.getDiagnosticLevel(diag::warn_suggest_noreturn_block)
2606          == Diagnostic::Ignored || !ReturnsVoid))
2607    return;
2608  // FIXME: Funtion try block
2609  if (CompoundStmt *Compound = dyn_cast<CompoundStmt>(Body)) {
2610    switch (CheckFallThrough(AC)) {
2611    case MaybeFallThrough:
2612      if (HasNoReturn)
2613        Diag(Compound->getRBracLoc(), diag::err_noreturn_block_has_return_expr);
2614      else if (!ReturnsVoid)
2615        Diag(Compound->getRBracLoc(), diag::err_maybe_falloff_nonvoid_block);
2616      break;
2617    case AlwaysFallThrough:
2618      if (HasNoReturn)
2619        Diag(Compound->getRBracLoc(), diag::err_noreturn_block_has_return_expr);
2620      else if (!ReturnsVoid)
2621        Diag(Compound->getRBracLoc(), diag::err_falloff_nonvoid_block);
2622      break;
2623    case NeverFallThroughOrReturn:
2624      if (ReturnsVoid)
2625        Diag(Compound->getLBracLoc(), diag::warn_suggest_noreturn_block);
2626      break;
2627    case NeverFallThrough:
2628      break;
2629    }
2630  }
2631}
2632
2633/// CheckParmsForFunctionDef - Check that the parameters of the given
2634/// function are appropriate for the definition of a function. This
2635/// takes care of any checks that cannot be performed on the
2636/// declaration itself, e.g., that the types of each of the function
2637/// parameters are complete.
2638bool Sema::CheckParmsForFunctionDef(FunctionDecl *FD) {
2639  bool HasInvalidParm = false;
2640  for (unsigned p = 0, NumParams = FD->getNumParams(); p < NumParams; ++p) {
2641    ParmVarDecl *Param = FD->getParamDecl(p);
2642
2643    // C99 6.7.5.3p4: the parameters in a parameter type list in a
2644    // function declarator that is part of a function definition of
2645    // that function shall not have incomplete type.
2646    //
2647    // This is also C++ [dcl.fct]p6.
2648    if (!Param->isInvalidDecl() &&
2649        RequireCompleteType(Param->getLocation(), Param->getType(),
2650                               diag::err_typecheck_decl_incomplete_type)) {
2651      Param->setInvalidDecl();
2652      HasInvalidParm = true;
2653    }
2654
2655    // C99 6.9.1p5: If the declarator includes a parameter type list, the
2656    // declaration of each parameter shall include an identifier.
2657    if (Param->getIdentifier() == 0 &&
2658        !Param->isImplicit() &&
2659        !getLangOptions().CPlusPlus)
2660      Diag(Param->getLocation(), diag::err_parameter_name_omitted);
2661
2662    // C99 6.7.5.3p12:
2663    //   If the function declarator is not part of a definition of that
2664    //   function, parameters may have incomplete type and may use the [*]
2665    //   notation in their sequences of declarator specifiers to specify
2666    //   variable length array types.
2667    QualType PType = Param->getOriginalType();
2668    if (const ArrayType *AT = Context.getAsArrayType(PType)) {
2669      if (AT->getSizeModifier() == ArrayType::Star) {
2670        // FIXME: This diagnosic should point the the '[*]' if source-location
2671        // information is added for it.
2672        Diag(Param->getLocation(), diag::err_array_star_in_function_definition);
2673      }
2674    }
2675
2676    if (getLangOptions().CPlusPlus)
2677      if (const RecordType *RT = Param->getType()->getAs<RecordType>())
2678        FinalizeVarWithDestructor(Param, RT);
2679  }
2680
2681  return HasInvalidParm;
2682}
2683