SemaChecking.cpp revision d635c5fcc45c952b75743dd2f4c86d6950e4954e
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        const VarDecl *Def = 0;
850        if (const Expr *Init = VD->getDefinition(Def))
851          return SemaCheckStringLiteral(Init, TheCall,
852                                        HasVAListArg, format_idx, firstDataArg);
853      }
854
855      // For vprintf* functions (i.e., HasVAListArg==true), we add a
856      // special check to see if the format string is a function parameter
857      // of the function calling the printf function.  If the function
858      // has an attribute indicating it is a printf-like function, then we
859      // should suppress warnings concerning non-literals being used in a call
860      // to a vprintf function.  For example:
861      //
862      // void
863      // logmessage(char const *fmt __attribute__ (format (printf, 1, 2)), ...){
864      //      va_list ap;
865      //      va_start(ap, fmt);
866      //      vprintf(fmt, ap);  // Do NOT emit a warning about "fmt".
867      //      ...
868      //
869      //
870      //  FIXME: We don't have full attribute support yet, so just check to see
871      //    if the argument is a DeclRefExpr that references a parameter.  We'll
872      //    add proper support for checking the attribute later.
873      if (HasVAListArg)
874        if (isa<ParmVarDecl>(VD))
875          return true;
876    }
877
878    return false;
879  }
880
881  case Stmt::CallExprClass: {
882    const CallExpr *CE = cast<CallExpr>(E);
883    if (const ImplicitCastExpr *ICE
884          = dyn_cast<ImplicitCastExpr>(CE->getCallee())) {
885      if (const DeclRefExpr *DRE = dyn_cast<DeclRefExpr>(ICE->getSubExpr())) {
886        if (const FunctionDecl *FD = dyn_cast<FunctionDecl>(DRE->getDecl())) {
887          if (const FormatArgAttr *FA = FD->getAttr<FormatArgAttr>()) {
888            unsigned ArgIndex = FA->getFormatIdx();
889            const Expr *Arg = CE->getArg(ArgIndex - 1);
890
891            return SemaCheckStringLiteral(Arg, TheCall, HasVAListArg,
892                                          format_idx, firstDataArg);
893          }
894        }
895      }
896    }
897
898    return false;
899  }
900  case Stmt::ObjCStringLiteralClass:
901  case Stmt::StringLiteralClass: {
902    const StringLiteral *StrE = NULL;
903
904    if (const ObjCStringLiteral *ObjCFExpr = dyn_cast<ObjCStringLiteral>(E))
905      StrE = ObjCFExpr->getString();
906    else
907      StrE = cast<StringLiteral>(E);
908
909    if (StrE) {
910      CheckPrintfString(StrE, E, TheCall, HasVAListArg, format_idx,
911                        firstDataArg);
912      return true;
913    }
914
915    return false;
916  }
917
918  default:
919    return false;
920  }
921}
922
923void
924Sema::CheckNonNullArguments(const NonNullAttr *NonNull,
925                            const CallExpr *TheCall) {
926  for (NonNullAttr::iterator i = NonNull->begin(), e = NonNull->end();
927       i != e; ++i) {
928    const Expr *ArgExpr = TheCall->getArg(*i);
929    if (ArgExpr->isNullPointerConstant(Context,
930                                       Expr::NPC_ValueDependentIsNotNull))
931      Diag(TheCall->getCallee()->getLocStart(), diag::warn_null_arg)
932        << ArgExpr->getSourceRange();
933  }
934}
935
936/// CheckPrintfArguments - Check calls to printf (and similar functions) for
937/// correct use of format strings.
938///
939///  HasVAListArg - A predicate indicating whether the printf-like
940///    function is passed an explicit va_arg argument (e.g., vprintf)
941///
942///  format_idx - The index into Args for the format string.
943///
944/// Improper format strings to functions in the printf family can be
945/// the source of bizarre bugs and very serious security holes.  A
946/// good source of information is available in the following paper
947/// (which includes additional references):
948///
949///  FormatGuard: Automatic Protection From printf Format String
950///  Vulnerabilities, Proceedings of the 10th USENIX Security Symposium, 2001.
951///
952/// Functionality implemented:
953///
954///  We can statically check the following properties for string
955///  literal format strings for non v.*printf functions (where the
956///  arguments are passed directly):
957//
958///  (1) Are the number of format conversions equal to the number of
959///      data arguments?
960///
961///  (2) Does each format conversion correctly match the type of the
962///      corresponding data argument?  (TODO)
963///
964/// Moreover, for all printf functions we can:
965///
966///  (3) Check for a missing format string (when not caught by type checking).
967///
968///  (4) Check for no-operation flags; e.g. using "#" with format
969///      conversion 'c'  (TODO)
970///
971///  (5) Check the use of '%n', a major source of security holes.
972///
973///  (6) Check for malformed format conversions that don't specify anything.
974///
975///  (7) Check for empty format strings.  e.g: printf("");
976///
977///  (8) Check that the format string is a wide literal.
978///
979/// All of these checks can be done by parsing the format string.
980///
981/// For now, we ONLY do (1), (3), (5), (6), (7), and (8).
982void
983Sema::CheckPrintfArguments(const CallExpr *TheCall, bool HasVAListArg,
984                           unsigned format_idx, unsigned firstDataArg) {
985  const Expr *Fn = TheCall->getCallee();
986
987  // The way the format attribute works in GCC, the implicit this argument
988  // of member functions is counted. However, it doesn't appear in our own
989  // lists, so decrement format_idx in that case.
990  if (isa<CXXMemberCallExpr>(TheCall)) {
991    // Catch a format attribute mistakenly referring to the object argument.
992    if (format_idx == 0)
993      return;
994    --format_idx;
995    if(firstDataArg != 0)
996      --firstDataArg;
997  }
998
999  // CHECK: printf-like function is called with no format string.
1000  if (format_idx >= TheCall->getNumArgs()) {
1001    Diag(TheCall->getRParenLoc(), diag::warn_printf_missing_format_string)
1002      << Fn->getSourceRange();
1003    return;
1004  }
1005
1006  const Expr *OrigFormatExpr = TheCall->getArg(format_idx)->IgnoreParenCasts();
1007
1008  // CHECK: format string is not a string literal.
1009  //
1010  // Dynamically generated format strings are difficult to
1011  // automatically vet at compile time.  Requiring that format strings
1012  // are string literals: (1) permits the checking of format strings by
1013  // the compiler and thereby (2) can practically remove the source of
1014  // many format string exploits.
1015
1016  // Format string can be either ObjC string (e.g. @"%d") or
1017  // C string (e.g. "%d")
1018  // ObjC string uses the same format specifiers as C string, so we can use
1019  // the same format string checking logic for both ObjC and C strings.
1020  if (SemaCheckStringLiteral(OrigFormatExpr, TheCall, HasVAListArg, format_idx,
1021                             firstDataArg))
1022    return;  // Literal format string found, check done!
1023
1024  // If there are no arguments specified, warn with -Wformat-security, otherwise
1025  // warn only with -Wformat-nonliteral.
1026  if (TheCall->getNumArgs() == format_idx+1)
1027    Diag(TheCall->getArg(format_idx)->getLocStart(),
1028         diag::warn_printf_nonliteral_noargs)
1029      << OrigFormatExpr->getSourceRange();
1030  else
1031    Diag(TheCall->getArg(format_idx)->getLocStart(),
1032         diag::warn_printf_nonliteral)
1033           << OrigFormatExpr->getSourceRange();
1034}
1035
1036namespace {
1037class CheckPrintfHandler : public FormatStringHandler {
1038  Sema &S;
1039  const StringLiteral *FExpr;
1040  const Expr *OrigFormatExpr;
1041  unsigned NumConversions;
1042  const unsigned NumDataArgs;
1043  const bool IsObjCLiteral;
1044  const char *Beg; // Start of format string.
1045  const bool HasVAListArg;
1046  const CallExpr *TheCall;
1047  unsigned FormatIdx;
1048public:
1049  CheckPrintfHandler(Sema &s, const StringLiteral *fexpr,
1050                     const Expr *origFormatExpr,
1051                     unsigned numDataArgs, bool isObjCLiteral,
1052                     const char *beg, bool hasVAListArg,
1053                     const CallExpr *theCall, unsigned formatIdx)
1054    : S(s), FExpr(fexpr), OrigFormatExpr(origFormatExpr),
1055      NumConversions(0), NumDataArgs(numDataArgs),
1056      IsObjCLiteral(isObjCLiteral), Beg(beg),
1057      HasVAListArg(hasVAListArg),
1058      TheCall(theCall), FormatIdx(formatIdx) {}
1059
1060  void DoneProcessing();
1061
1062  void HandleIncompleteFormatSpecifier(const char *startSpecifier,
1063                                       unsigned specifierLen);
1064
1065  void 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
1084  bool MatchType(QualType A, QualType B, bool ignoreSign);
1085
1086  const Expr *getDataArg(unsigned i) const;
1087};
1088}
1089
1090SourceRange CheckPrintfHandler::getFormatStringRange() {
1091  return OrigFormatExpr->getSourceRange();
1092}
1093
1094SourceRange CheckPrintfHandler::
1095getFormatSpecifierRange(const char *startSpecifier, unsigned specifierLen) {
1096  return SourceRange(getLocationOfByte(startSpecifier),
1097					 getLocationOfByte(startSpecifier+specifierLen-1));
1098}
1099
1100SourceLocation CheckPrintfHandler::getLocationOfByte(const char *x) {
1101  return S.getLocationOfStringLiteralByte(FExpr, x - Beg);
1102}
1103
1104void CheckPrintfHandler::
1105HandleIncompleteFormatSpecifier(const char *startSpecifier,
1106                                unsigned specifierLen) {
1107  SourceLocation Loc = getLocationOfByte(startSpecifier);
1108  S.Diag(Loc, diag::warn_printf_incomplete_specifier)
1109    << getFormatSpecifierRange(startSpecifier, specifierLen);
1110}
1111
1112void CheckPrintfHandler::
1113HandleInvalidConversionSpecifier(const analyze_printf::FormatSpecifier &FS,
1114                                 const char *startSpecifier,
1115                                 unsigned specifierLen) {
1116
1117  ++NumConversions;
1118  const analyze_printf::ConversionSpecifier &CS =
1119    FS.getConversionSpecifier();
1120  SourceLocation Loc = getLocationOfByte(CS.getStart());
1121  S.Diag(Loc, diag::warn_printf_invalid_conversion)
1122      << llvm::StringRef(CS.getStart(), CS.getLength())
1123      << getFormatSpecifierRange(startSpecifier, specifierLen);
1124}
1125
1126void CheckPrintfHandler::HandleNullChar(const char *nullCharacter) {
1127  // The presence of a null character is likely an error.
1128  S.Diag(getLocationOfByte(nullCharacter),
1129         diag::warn_printf_format_string_contains_null_char)
1130    << getFormatStringRange();
1131}
1132
1133const Expr *CheckPrintfHandler::getDataArg(unsigned i) const {
1134  return TheCall->getArg(FormatIdx + i);
1135}
1136
1137bool CheckPrintfHandler::MatchType(QualType A, QualType B, bool ignoreSign) {
1138  A = S.Context.getCanonicalType(A).getUnqualifiedType();
1139  B = S.Context.getCanonicalType(B).getUnqualifiedType();
1140
1141  if (A == B)
1142    return true;
1143
1144  if (ignoreSign) {
1145    if (const BuiltinType *BT = B->getAs<BuiltinType>()) {
1146      switch (BT->getKind()) {
1147        default:
1148          return false;
1149        case BuiltinType::Char_S:
1150        case BuiltinType::SChar:
1151          return A == S.Context.UnsignedCharTy;
1152        case BuiltinType::Char_U:
1153        case BuiltinType::UChar:
1154          return A == S.Context.SignedCharTy;
1155        case BuiltinType::Short:
1156          return A == S.Context.UnsignedShortTy;
1157        case BuiltinType::UShort:
1158          return A == S.Context.ShortTy;
1159        case BuiltinType::Int:
1160          return A == S.Context.UnsignedIntTy;
1161        case BuiltinType::UInt:
1162          return A == S.Context.IntTy;
1163        case BuiltinType::Long:
1164          return A == S.Context.UnsignedLongTy;
1165        case BuiltinType::ULong:
1166          return A == S.Context.LongTy;
1167        case BuiltinType::LongLong:
1168          return A == S.Context.UnsignedLongLongTy;
1169        case BuiltinType::ULongLong:
1170          return A == S.Context.LongLongTy;
1171      }
1172      return A == B;
1173    }
1174  }
1175  return false;
1176}
1177
1178bool
1179CheckPrintfHandler::HandleAmount(const analyze_printf::OptionalAmount &Amt,
1180                                 unsigned MissingArgDiag,
1181                                 unsigned BadTypeDiag,
1182								 const char *startSpecifier,
1183								 unsigned specifierLen) {
1184
1185  if (Amt.hasDataArgument()) {
1186    ++NumConversions;
1187    if (!HasVAListArg) {
1188      if (NumConversions > NumDataArgs) {
1189        S.Diag(getLocationOfByte(Amt.getStart()), MissingArgDiag)
1190          << getFormatSpecifierRange(startSpecifier, specifierLen);
1191        // Don't do any more checking.  We will just emit
1192        // spurious errors.
1193        return false;
1194      }
1195
1196      // Type check the data argument.  It should be an 'int'.
1197      // Although not in conformance with C99, we also allow the argument to be
1198      // an 'unsigned int' as that is a reasonably safe case.  GCC also
1199      // doesn't emit a warning for that case.
1200      const Expr *Arg = getDataArg(NumConversions);
1201      QualType T = Arg->getType();
1202      if (!MatchType(T, S.Context.IntTy, true)) {
1203        S.Diag(getLocationOfByte(Amt.getStart()), BadTypeDiag)
1204          << S.Context.IntTy << T
1205          << getFormatSpecifierRange(startSpecifier, specifierLen)
1206          << Arg->getSourceRange();
1207        // Don't do any more checking.  We will just emit
1208        // spurious errors.
1209        return false;
1210      }
1211    }
1212  }
1213  return true;
1214}
1215
1216bool
1217CheckPrintfHandler::HandleFormatSpecifier(const analyze_printf::FormatSpecifier &FS,
1218                                          const char *startSpecifier,
1219                                          unsigned specifierLen) {
1220
1221  using namespace analyze_printf;
1222  const ConversionSpecifier &CS = FS.getConversionSpecifier();
1223
1224  // First check if the field width, precision, and conversion specifier
1225  // have matching data arguments.
1226  if (!HandleAmount(FS.getFieldWidth(),
1227                    diag::warn_printf_asterisk_width_missing_arg,
1228                    diag::warn_printf_asterisk_width_wrong_type,
1229					startSpecifier, specifierLen)) {
1230    return false;
1231  }
1232
1233  if (!HandleAmount(FS.getPrecision(),
1234                    diag::warn_printf_asterisk_precision_missing_arg,
1235                    diag::warn_printf_asterisk_precision_wrong_type,
1236					startSpecifier, specifierLen)) {
1237    return false;
1238  }
1239
1240  // Check for using an Objective-C specific conversion specifier
1241  // in a non-ObjC literal.
1242  if (!IsObjCLiteral && CS.isObjCArg()) {
1243    HandleInvalidConversionSpecifier(FS, startSpecifier, specifierLen);
1244
1245    // Continue checking the other format specifiers.
1246    return true;
1247  }
1248
1249  if (!CS.consumesDataArgument()) {
1250    // FIXME: Technically specifying a precision or field width here
1251    // makes no sense.  Worth issuing a warning at some point.
1252	return true;
1253  }
1254
1255  ++NumConversions;
1256
1257  // Are we using '%n'?  Issue a warning about this being
1258  // a possible security issue.
1259  if (CS.getKind() == ConversionSpecifier::OutIntPtrArg) {
1260    S.Diag(getLocationOfByte(CS.getStart()), diag::warn_printf_write_back)
1261      << getFormatSpecifierRange(startSpecifier, specifierLen);
1262    // Continue checking the other format specifiers.
1263    return true;
1264  }
1265
1266
1267  // The remaining checks depend on the data arguments.
1268  if (HasVAListArg)
1269    return true;
1270
1271  if (NumConversions > NumDataArgs) {
1272    S.Diag(getLocationOfByte(CS.getStart()),
1273           diag::warn_printf_insufficient_data_args)
1274      << getFormatSpecifierRange(startSpecifier, specifierLen);
1275    // Don't do any more checking.
1276    return false;
1277  }
1278
1279  // Now type check the data expression that matches the
1280  // format specifier.
1281  const Expr *Ex = getDataArg(NumConversions);
1282  const analyze_printf::ArgTypeResult &ATR = FS.getArgType(S.Context);
1283
1284  if (const QualType *T = ATR.getSpecificType()) {
1285    if (!MatchType(*T, Ex->getType(), true)) {
1286      S.Diag(getLocationOfByte(CS.getStart()),
1287             diag::warn_printf_conversion_argument_type_mismatch)
1288        << *T << Ex->getType()
1289        << getFormatSpecifierRange(startSpecifier, specifierLen)
1290        << Ex->getSourceRange();
1291    }
1292    return true;
1293  }
1294
1295  return true;
1296}
1297
1298void CheckPrintfHandler::DoneProcessing() {
1299  // Does the number of data arguments exceed the number of
1300  // format conversions in the format string?
1301  if (!HasVAListArg && NumConversions < NumDataArgs)
1302    S.Diag(getDataArg(NumConversions+1)->getLocStart(),
1303           diag::warn_printf_too_many_data_args)
1304      << getFormatStringRange();
1305}
1306
1307void Sema::CheckPrintfString(const StringLiteral *FExpr,
1308							 const Expr *OrigFormatExpr,
1309							 const CallExpr *TheCall, bool HasVAListArg,
1310							 unsigned format_idx, unsigned firstDataArg) {
1311
1312  // CHECK: is the format string a wide literal?
1313  if (FExpr->isWide()) {
1314    Diag(FExpr->getLocStart(),
1315         diag::warn_printf_format_string_is_wide_literal)
1316    << OrigFormatExpr->getSourceRange();
1317    return;
1318  }
1319
1320  // Str - The format string.  NOTE: this is NOT null-terminated!
1321  const char *Str = FExpr->getStrData();
1322
1323  // CHECK: empty format string?
1324  unsigned StrLen = FExpr->getByteLength();
1325
1326  if (StrLen == 0) {
1327    Diag(FExpr->getLocStart(), diag::warn_printf_empty_format_string)
1328    << OrigFormatExpr->getSourceRange();
1329    return;
1330  }
1331
1332  CheckPrintfHandler H(*this, FExpr, OrigFormatExpr,
1333                       TheCall->getNumArgs() - firstDataArg,
1334                       isa<ObjCStringLiteral>(OrigFormatExpr), Str,
1335                       HasVAListArg, TheCall, format_idx);
1336
1337  if (!ParseFormatString(H, Str, Str + StrLen))
1338    H.DoneProcessing();
1339}
1340
1341//===--- CHECK: Return Address of Stack Variable --------------------------===//
1342
1343static DeclRefExpr* EvalVal(Expr *E);
1344static DeclRefExpr* EvalAddr(Expr* E);
1345
1346/// CheckReturnStackAddr - Check if a return statement returns the address
1347///   of a stack variable.
1348void
1349Sema::CheckReturnStackAddr(Expr *RetValExp, QualType lhsType,
1350                           SourceLocation ReturnLoc) {
1351
1352  // Perform checking for returned stack addresses.
1353  if (lhsType->isPointerType() || lhsType->isBlockPointerType()) {
1354    if (DeclRefExpr *DR = EvalAddr(RetValExp))
1355      Diag(DR->getLocStart(), diag::warn_ret_stack_addr)
1356       << DR->getDecl()->getDeclName() << RetValExp->getSourceRange();
1357
1358    // Skip over implicit cast expressions when checking for block expressions.
1359    RetValExp = RetValExp->IgnoreParenCasts();
1360
1361    if (BlockExpr *C = dyn_cast<BlockExpr>(RetValExp))
1362      if (C->hasBlockDeclRefExprs())
1363        Diag(C->getLocStart(), diag::err_ret_local_block)
1364          << C->getSourceRange();
1365
1366    if (AddrLabelExpr *ALE = dyn_cast<AddrLabelExpr>(RetValExp))
1367      Diag(ALE->getLocStart(), diag::warn_ret_addr_label)
1368        << ALE->getSourceRange();
1369
1370  } else if (lhsType->isReferenceType()) {
1371    // Perform checking for stack values returned by reference.
1372    // Check for a reference to the stack
1373    if (DeclRefExpr *DR = EvalVal(RetValExp))
1374      Diag(DR->getLocStart(), diag::warn_ret_stack_ref)
1375        << DR->getDecl()->getDeclName() << RetValExp->getSourceRange();
1376  }
1377}
1378
1379/// EvalAddr - EvalAddr and EvalVal are mutually recursive functions that
1380///  check if the expression in a return statement evaluates to an address
1381///  to a location on the stack.  The recursion is used to traverse the
1382///  AST of the return expression, with recursion backtracking when we
1383///  encounter a subexpression that (1) clearly does not lead to the address
1384///  of a stack variable or (2) is something we cannot determine leads to
1385///  the address of a stack variable based on such local checking.
1386///
1387///  EvalAddr processes expressions that are pointers that are used as
1388///  references (and not L-values).  EvalVal handles all other values.
1389///  At the base case of the recursion is a check for a DeclRefExpr* in
1390///  the refers to a stack variable.
1391///
1392///  This implementation handles:
1393///
1394///   * pointer-to-pointer casts
1395///   * implicit conversions from array references to pointers
1396///   * taking the address of fields
1397///   * arbitrary interplay between "&" and "*" operators
1398///   * pointer arithmetic from an address of a stack variable
1399///   * taking the address of an array element where the array is on the stack
1400static DeclRefExpr* EvalAddr(Expr *E) {
1401  // We should only be called for evaluating pointer expressions.
1402  assert((E->getType()->isAnyPointerType() ||
1403          E->getType()->isBlockPointerType() ||
1404          E->getType()->isObjCQualifiedIdType()) &&
1405         "EvalAddr only works on pointers");
1406
1407  // Our "symbolic interpreter" is just a dispatch off the currently
1408  // viewed AST node.  We then recursively traverse the AST by calling
1409  // EvalAddr and EvalVal appropriately.
1410  switch (E->getStmtClass()) {
1411  case Stmt::ParenExprClass:
1412    // Ignore parentheses.
1413    return EvalAddr(cast<ParenExpr>(E)->getSubExpr());
1414
1415  case Stmt::UnaryOperatorClass: {
1416    // The only unary operator that make sense to handle here
1417    // is AddrOf.  All others don't make sense as pointers.
1418    UnaryOperator *U = cast<UnaryOperator>(E);
1419
1420    if (U->getOpcode() == UnaryOperator::AddrOf)
1421      return EvalVal(U->getSubExpr());
1422    else
1423      return NULL;
1424  }
1425
1426  case Stmt::BinaryOperatorClass: {
1427    // Handle pointer arithmetic.  All other binary operators are not valid
1428    // in this context.
1429    BinaryOperator *B = cast<BinaryOperator>(E);
1430    BinaryOperator::Opcode op = B->getOpcode();
1431
1432    if (op != BinaryOperator::Add && op != BinaryOperator::Sub)
1433      return NULL;
1434
1435    Expr *Base = B->getLHS();
1436
1437    // Determine which argument is the real pointer base.  It could be
1438    // the RHS argument instead of the LHS.
1439    if (!Base->getType()->isPointerType()) Base = B->getRHS();
1440
1441    assert (Base->getType()->isPointerType());
1442    return EvalAddr(Base);
1443  }
1444
1445  // For conditional operators we need to see if either the LHS or RHS are
1446  // valid DeclRefExpr*s.  If one of them is valid, we return it.
1447  case Stmt::ConditionalOperatorClass: {
1448    ConditionalOperator *C = cast<ConditionalOperator>(E);
1449
1450    // Handle the GNU extension for missing LHS.
1451    if (Expr *lhsExpr = C->getLHS())
1452      if (DeclRefExpr* LHS = EvalAddr(lhsExpr))
1453        return LHS;
1454
1455     return EvalAddr(C->getRHS());
1456  }
1457
1458  // For casts, we need to handle conversions from arrays to
1459  // pointer values, and pointer-to-pointer conversions.
1460  case Stmt::ImplicitCastExprClass:
1461  case Stmt::CStyleCastExprClass:
1462  case Stmt::CXXFunctionalCastExprClass: {
1463    Expr* SubExpr = cast<CastExpr>(E)->getSubExpr();
1464    QualType T = SubExpr->getType();
1465
1466    if (SubExpr->getType()->isPointerType() ||
1467        SubExpr->getType()->isBlockPointerType() ||
1468        SubExpr->getType()->isObjCQualifiedIdType())
1469      return EvalAddr(SubExpr);
1470    else if (T->isArrayType())
1471      return EvalVal(SubExpr);
1472    else
1473      return 0;
1474  }
1475
1476  // C++ casts.  For dynamic casts, static casts, and const casts, we
1477  // are always converting from a pointer-to-pointer, so we just blow
1478  // through the cast.  In the case the dynamic cast doesn't fail (and
1479  // return NULL), we take the conservative route and report cases
1480  // where we return the address of a stack variable.  For Reinterpre
1481  // FIXME: The comment about is wrong; we're not always converting
1482  // from pointer to pointer. I'm guessing that this code should also
1483  // handle references to objects.
1484  case Stmt::CXXStaticCastExprClass:
1485  case Stmt::CXXDynamicCastExprClass:
1486  case Stmt::CXXConstCastExprClass:
1487  case Stmt::CXXReinterpretCastExprClass: {
1488      Expr *S = cast<CXXNamedCastExpr>(E)->getSubExpr();
1489      if (S->getType()->isPointerType() || S->getType()->isBlockPointerType())
1490        return EvalAddr(S);
1491      else
1492        return NULL;
1493  }
1494
1495  // Everything else: we simply don't reason about them.
1496  default:
1497    return NULL;
1498  }
1499}
1500
1501
1502///  EvalVal - This function is complements EvalAddr in the mutual recursion.
1503///   See the comments for EvalAddr for more details.
1504static DeclRefExpr* EvalVal(Expr *E) {
1505
1506  // We should only be called for evaluating non-pointer expressions, or
1507  // expressions with a pointer type that are not used as references but instead
1508  // are l-values (e.g., DeclRefExpr with a pointer type).
1509
1510  // Our "symbolic interpreter" is just a dispatch off the currently
1511  // viewed AST node.  We then recursively traverse the AST by calling
1512  // EvalAddr and EvalVal appropriately.
1513  switch (E->getStmtClass()) {
1514  case Stmt::DeclRefExprClass: {
1515    // DeclRefExpr: the base case.  When we hit a DeclRefExpr we are looking
1516    //  at code that refers to a variable's name.  We check if it has local
1517    //  storage within the function, and if so, return the expression.
1518    DeclRefExpr *DR = cast<DeclRefExpr>(E);
1519
1520    if (VarDecl *V = dyn_cast<VarDecl>(DR->getDecl()))
1521      if (V->hasLocalStorage() && !V->getType()->isReferenceType()) return DR;
1522
1523    return NULL;
1524  }
1525
1526  case Stmt::ParenExprClass:
1527    // Ignore parentheses.
1528    return EvalVal(cast<ParenExpr>(E)->getSubExpr());
1529
1530  case Stmt::UnaryOperatorClass: {
1531    // The only unary operator that make sense to handle here
1532    // is Deref.  All others don't resolve to a "name."  This includes
1533    // handling all sorts of rvalues passed to a unary operator.
1534    UnaryOperator *U = cast<UnaryOperator>(E);
1535
1536    if (U->getOpcode() == UnaryOperator::Deref)
1537      return EvalAddr(U->getSubExpr());
1538
1539    return NULL;
1540  }
1541
1542  case Stmt::ArraySubscriptExprClass: {
1543    // Array subscripts are potential references to data on the stack.  We
1544    // retrieve the DeclRefExpr* for the array variable if it indeed
1545    // has local storage.
1546    return EvalAddr(cast<ArraySubscriptExpr>(E)->getBase());
1547  }
1548
1549  case Stmt::ConditionalOperatorClass: {
1550    // For conditional operators we need to see if either the LHS or RHS are
1551    // non-NULL DeclRefExpr's.  If one is non-NULL, we return it.
1552    ConditionalOperator *C = cast<ConditionalOperator>(E);
1553
1554    // Handle the GNU extension for missing LHS.
1555    if (Expr *lhsExpr = C->getLHS())
1556      if (DeclRefExpr *LHS = EvalVal(lhsExpr))
1557        return LHS;
1558
1559    return EvalVal(C->getRHS());
1560  }
1561
1562  // Accesses to members are potential references to data on the stack.
1563  case Stmt::MemberExprClass: {
1564    MemberExpr *M = cast<MemberExpr>(E);
1565
1566    // Check for indirect access.  We only want direct field accesses.
1567    if (!M->isArrow())
1568      return EvalVal(M->getBase());
1569    else
1570      return NULL;
1571  }
1572
1573  // Everything else: we simply don't reason about them.
1574  default:
1575    return NULL;
1576  }
1577}
1578
1579//===--- CHECK: Floating-Point comparisons (-Wfloat-equal) ---------------===//
1580
1581/// Check for comparisons of floating point operands using != and ==.
1582/// Issue a warning if these are no self-comparisons, as they are not likely
1583/// to do what the programmer intended.
1584void Sema::CheckFloatComparison(SourceLocation loc, Expr* lex, Expr *rex) {
1585  bool EmitWarning = true;
1586
1587  Expr* LeftExprSansParen = lex->IgnoreParens();
1588  Expr* RightExprSansParen = rex->IgnoreParens();
1589
1590  // Special case: check for x == x (which is OK).
1591  // Do not emit warnings for such cases.
1592  if (DeclRefExpr* DRL = dyn_cast<DeclRefExpr>(LeftExprSansParen))
1593    if (DeclRefExpr* DRR = dyn_cast<DeclRefExpr>(RightExprSansParen))
1594      if (DRL->getDecl() == DRR->getDecl())
1595        EmitWarning = false;
1596
1597
1598  // Special case: check for comparisons against literals that can be exactly
1599  //  represented by APFloat.  In such cases, do not emit a warning.  This
1600  //  is a heuristic: often comparison against such literals are used to
1601  //  detect if a value in a variable has not changed.  This clearly can
1602  //  lead to false negatives.
1603  if (EmitWarning) {
1604    if (FloatingLiteral* FLL = dyn_cast<FloatingLiteral>(LeftExprSansParen)) {
1605      if (FLL->isExact())
1606        EmitWarning = false;
1607    } else
1608      if (FloatingLiteral* FLR = dyn_cast<FloatingLiteral>(RightExprSansParen)){
1609        if (FLR->isExact())
1610          EmitWarning = false;
1611    }
1612  }
1613
1614  // Check for comparisons with builtin types.
1615  if (EmitWarning)
1616    if (CallExpr* CL = dyn_cast<CallExpr>(LeftExprSansParen))
1617      if (CL->isBuiltinCall(Context))
1618        EmitWarning = false;
1619
1620  if (EmitWarning)
1621    if (CallExpr* CR = dyn_cast<CallExpr>(RightExprSansParen))
1622      if (CR->isBuiltinCall(Context))
1623        EmitWarning = false;
1624
1625  // Emit the diagnostic.
1626  if (EmitWarning)
1627    Diag(loc, diag::warn_floatingpoint_eq)
1628      << lex->getSourceRange() << rex->getSourceRange();
1629}
1630
1631//===--- CHECK: Integer mixed-sign comparisons (-Wsign-compare) --------===//
1632//===--- CHECK: Lossy implicit conversions (-Wconversion) --------------===//
1633
1634namespace {
1635
1636/// Structure recording the 'active' range of an integer-valued
1637/// expression.
1638struct IntRange {
1639  /// The number of bits active in the int.
1640  unsigned Width;
1641
1642  /// True if the int is known not to have negative values.
1643  bool NonNegative;
1644
1645  IntRange() {}
1646  IntRange(unsigned Width, bool NonNegative)
1647    : Width(Width), NonNegative(NonNegative)
1648  {}
1649
1650  // Returns the range of the bool type.
1651  static IntRange forBoolType() {
1652    return IntRange(1, true);
1653  }
1654
1655  // Returns the range of an integral type.
1656  static IntRange forType(ASTContext &C, QualType T) {
1657    return forCanonicalType(C, T->getCanonicalTypeInternal().getTypePtr());
1658  }
1659
1660  // Returns the range of an integeral type based on its canonical
1661  // representation.
1662  static IntRange forCanonicalType(ASTContext &C, const Type *T) {
1663    assert(T->isCanonicalUnqualified());
1664
1665    if (const VectorType *VT = dyn_cast<VectorType>(T))
1666      T = VT->getElementType().getTypePtr();
1667    if (const ComplexType *CT = dyn_cast<ComplexType>(T))
1668      T = CT->getElementType().getTypePtr();
1669    if (const EnumType *ET = dyn_cast<EnumType>(T))
1670      T = ET->getDecl()->getIntegerType().getTypePtr();
1671
1672    const BuiltinType *BT = cast<BuiltinType>(T);
1673    assert(BT->isInteger());
1674
1675    return IntRange(C.getIntWidth(QualType(T, 0)), BT->isUnsignedInteger());
1676  }
1677
1678  // Returns the supremum of two ranges: i.e. their conservative merge.
1679  static IntRange join(const IntRange &L, const IntRange &R) {
1680    return IntRange(std::max(L.Width, R.Width),
1681                    L.NonNegative && R.NonNegative);
1682  }
1683
1684  // Returns the infinum of two ranges: i.e. their aggressive merge.
1685  static IntRange meet(const IntRange &L, const IntRange &R) {
1686    return IntRange(std::min(L.Width, R.Width),
1687                    L.NonNegative || R.NonNegative);
1688  }
1689};
1690
1691IntRange GetValueRange(ASTContext &C, llvm::APSInt &value, unsigned MaxWidth) {
1692  if (value.isSigned() && value.isNegative())
1693    return IntRange(value.getMinSignedBits(), false);
1694
1695  if (value.getBitWidth() > MaxWidth)
1696    value.trunc(MaxWidth);
1697
1698  // isNonNegative() just checks the sign bit without considering
1699  // signedness.
1700  return IntRange(value.getActiveBits(), true);
1701}
1702
1703IntRange GetValueRange(ASTContext &C, APValue &result, QualType Ty,
1704                       unsigned MaxWidth) {
1705  if (result.isInt())
1706    return GetValueRange(C, result.getInt(), MaxWidth);
1707
1708  if (result.isVector()) {
1709    IntRange R = GetValueRange(C, result.getVectorElt(0), Ty, MaxWidth);
1710    for (unsigned i = 1, e = result.getVectorLength(); i != e; ++i) {
1711      IntRange El = GetValueRange(C, result.getVectorElt(i), Ty, MaxWidth);
1712      R = IntRange::join(R, El);
1713    }
1714    return R;
1715  }
1716
1717  if (result.isComplexInt()) {
1718    IntRange R = GetValueRange(C, result.getComplexIntReal(), MaxWidth);
1719    IntRange I = GetValueRange(C, result.getComplexIntImag(), MaxWidth);
1720    return IntRange::join(R, I);
1721  }
1722
1723  // This can happen with lossless casts to intptr_t of "based" lvalues.
1724  // Assume it might use arbitrary bits.
1725  // FIXME: The only reason we need to pass the type in here is to get
1726  // the sign right on this one case.  It would be nice if APValue
1727  // preserved this.
1728  assert(result.isLValue());
1729  return IntRange(MaxWidth, Ty->isUnsignedIntegerType());
1730}
1731
1732/// Pseudo-evaluate the given integer expression, estimating the
1733/// range of values it might take.
1734///
1735/// \param MaxWidth - the width to which the value will be truncated
1736IntRange GetExprRange(ASTContext &C, Expr *E, unsigned MaxWidth) {
1737  E = E->IgnoreParens();
1738
1739  // Try a full evaluation first.
1740  Expr::EvalResult result;
1741  if (E->Evaluate(result, C))
1742    return GetValueRange(C, result.Val, E->getType(), MaxWidth);
1743
1744  // I think we only want to look through implicit casts here; if the
1745  // user has an explicit widening cast, we should treat the value as
1746  // being of the new, wider type.
1747  if (ImplicitCastExpr *CE = dyn_cast<ImplicitCastExpr>(E)) {
1748    if (CE->getCastKind() == CastExpr::CK_NoOp)
1749      return GetExprRange(C, CE->getSubExpr(), MaxWidth);
1750
1751    IntRange OutputTypeRange = IntRange::forType(C, CE->getType());
1752
1753    bool isIntegerCast = (CE->getCastKind() == CastExpr::CK_IntegralCast);
1754    if (!isIntegerCast && CE->getCastKind() == CastExpr::CK_Unknown)
1755      isIntegerCast = CE->getSubExpr()->getType()->isIntegerType();
1756
1757    // Assume that non-integer casts can span the full range of the type.
1758    if (!isIntegerCast)
1759      return OutputTypeRange;
1760
1761    IntRange SubRange
1762      = GetExprRange(C, CE->getSubExpr(),
1763                     std::min(MaxWidth, OutputTypeRange.Width));
1764
1765    // Bail out if the subexpr's range is as wide as the cast type.
1766    if (SubRange.Width >= OutputTypeRange.Width)
1767      return OutputTypeRange;
1768
1769    // Otherwise, we take the smaller width, and we're non-negative if
1770    // either the output type or the subexpr is.
1771    return IntRange(SubRange.Width,
1772                    SubRange.NonNegative || OutputTypeRange.NonNegative);
1773  }
1774
1775  if (ConditionalOperator *CO = dyn_cast<ConditionalOperator>(E)) {
1776    // If we can fold the condition, just take that operand.
1777    bool CondResult;
1778    if (CO->getCond()->EvaluateAsBooleanCondition(CondResult, C))
1779      return GetExprRange(C, CondResult ? CO->getTrueExpr()
1780                                        : CO->getFalseExpr(),
1781                          MaxWidth);
1782
1783    // Otherwise, conservatively merge.
1784    IntRange L = GetExprRange(C, CO->getTrueExpr(), MaxWidth);
1785    IntRange R = GetExprRange(C, CO->getFalseExpr(), MaxWidth);
1786    return IntRange::join(L, R);
1787  }
1788
1789  if (BinaryOperator *BO = dyn_cast<BinaryOperator>(E)) {
1790    switch (BO->getOpcode()) {
1791
1792    // Boolean-valued operations are single-bit and positive.
1793    case BinaryOperator::LAnd:
1794    case BinaryOperator::LOr:
1795    case BinaryOperator::LT:
1796    case BinaryOperator::GT:
1797    case BinaryOperator::LE:
1798    case BinaryOperator::GE:
1799    case BinaryOperator::EQ:
1800    case BinaryOperator::NE:
1801      return IntRange::forBoolType();
1802
1803    // Operations with opaque sources are black-listed.
1804    case BinaryOperator::PtrMemD:
1805    case BinaryOperator::PtrMemI:
1806      return IntRange::forType(C, E->getType());
1807
1808    // Bitwise-and uses the *infinum* of the two source ranges.
1809    case BinaryOperator::And:
1810      return IntRange::meet(GetExprRange(C, BO->getLHS(), MaxWidth),
1811                            GetExprRange(C, BO->getRHS(), MaxWidth));
1812
1813    // Left shift gets black-listed based on a judgement call.
1814    case BinaryOperator::Shl:
1815      return IntRange::forType(C, E->getType());
1816
1817    // Right shift by a constant can narrow its left argument.
1818    case BinaryOperator::Shr: {
1819      IntRange L = GetExprRange(C, BO->getLHS(), MaxWidth);
1820
1821      // If the shift amount is a positive constant, drop the width by
1822      // that much.
1823      llvm::APSInt shift;
1824      if (BO->getRHS()->isIntegerConstantExpr(shift, C) &&
1825          shift.isNonNegative()) {
1826        unsigned zext = shift.getZExtValue();
1827        if (zext >= L.Width)
1828          L.Width = (L.NonNegative ? 0 : 1);
1829        else
1830          L.Width -= zext;
1831      }
1832
1833      return L;
1834    }
1835
1836    // Comma acts as its right operand.
1837    case BinaryOperator::Comma:
1838      return GetExprRange(C, BO->getRHS(), MaxWidth);
1839
1840    // Black-list pointer subtractions.
1841    case BinaryOperator::Sub:
1842      if (BO->getLHS()->getType()->isPointerType())
1843        return IntRange::forType(C, E->getType());
1844      // fallthrough
1845
1846    default:
1847      break;
1848    }
1849
1850    // Treat every other operator as if it were closed on the
1851    // narrowest type that encompasses both operands.
1852    IntRange L = GetExprRange(C, BO->getLHS(), MaxWidth);
1853    IntRange R = GetExprRange(C, BO->getRHS(), MaxWidth);
1854    return IntRange::join(L, R);
1855  }
1856
1857  if (UnaryOperator *UO = dyn_cast<UnaryOperator>(E)) {
1858    switch (UO->getOpcode()) {
1859    // Boolean-valued operations are white-listed.
1860    case UnaryOperator::LNot:
1861      return IntRange::forBoolType();
1862
1863    // Operations with opaque sources are black-listed.
1864    case UnaryOperator::Deref:
1865    case UnaryOperator::AddrOf: // should be impossible
1866    case UnaryOperator::OffsetOf:
1867      return IntRange::forType(C, E->getType());
1868
1869    default:
1870      return GetExprRange(C, UO->getSubExpr(), MaxWidth);
1871    }
1872  }
1873
1874  FieldDecl *BitField = E->getBitField();
1875  if (BitField) {
1876    llvm::APSInt BitWidthAP = BitField->getBitWidth()->EvaluateAsInt(C);
1877    unsigned BitWidth = BitWidthAP.getZExtValue();
1878
1879    return IntRange(BitWidth, BitField->getType()->isUnsignedIntegerType());
1880  }
1881
1882  return IntRange::forType(C, E->getType());
1883}
1884
1885/// Checks whether the given value, which currently has the given
1886/// source semantics, has the same value when coerced through the
1887/// target semantics.
1888bool IsSameFloatAfterCast(const llvm::APFloat &value,
1889                          const llvm::fltSemantics &Src,
1890                          const llvm::fltSemantics &Tgt) {
1891  llvm::APFloat truncated = value;
1892
1893  bool ignored;
1894  truncated.convert(Src, llvm::APFloat::rmNearestTiesToEven, &ignored);
1895  truncated.convert(Tgt, llvm::APFloat::rmNearestTiesToEven, &ignored);
1896
1897  return truncated.bitwiseIsEqual(value);
1898}
1899
1900/// Checks whether the given value, which currently has the given
1901/// source semantics, has the same value when coerced through the
1902/// target semantics.
1903///
1904/// The value might be a vector of floats (or a complex number).
1905bool IsSameFloatAfterCast(const APValue &value,
1906                          const llvm::fltSemantics &Src,
1907                          const llvm::fltSemantics &Tgt) {
1908  if (value.isFloat())
1909    return IsSameFloatAfterCast(value.getFloat(), Src, Tgt);
1910
1911  if (value.isVector()) {
1912    for (unsigned i = 0, e = value.getVectorLength(); i != e; ++i)
1913      if (!IsSameFloatAfterCast(value.getVectorElt(i), Src, Tgt))
1914        return false;
1915    return true;
1916  }
1917
1918  assert(value.isComplexFloat());
1919  return (IsSameFloatAfterCast(value.getComplexFloatReal(), Src, Tgt) &&
1920          IsSameFloatAfterCast(value.getComplexFloatImag(), Src, Tgt));
1921}
1922
1923} // end anonymous namespace
1924
1925/// \brief Implements -Wsign-compare.
1926///
1927/// \param lex the left-hand expression
1928/// \param rex the right-hand expression
1929/// \param OpLoc the location of the joining operator
1930/// \param Equality whether this is an "equality-like" join, which
1931///   suppresses the warning in some cases
1932void Sema::CheckSignCompare(Expr *lex, Expr *rex, SourceLocation OpLoc,
1933                            const PartialDiagnostic &PD, bool Equality) {
1934  // Don't warn if we're in an unevaluated context.
1935  if (ExprEvalContexts.back().Context == Unevaluated)
1936    return;
1937
1938  // If either expression is value-dependent, don't warn. We'll get another
1939  // chance at instantiation time.
1940  if (lex->isValueDependent() || rex->isValueDependent())
1941    return;
1942
1943  QualType lt = lex->getType(), rt = rex->getType();
1944
1945  // Only warn if both operands are integral.
1946  if (!lt->isIntegerType() || !rt->isIntegerType())
1947    return;
1948
1949  // In C, the width of a bitfield determines its type, and the
1950  // declared type only contributes the signedness.  This duplicates
1951  // the work that will later be done by UsualUnaryConversions.
1952  // Eventually, this check will be reorganized in a way that avoids
1953  // this duplication.
1954  if (!getLangOptions().CPlusPlus) {
1955    QualType tmp;
1956    tmp = Context.isPromotableBitField(lex);
1957    if (!tmp.isNull()) lt = tmp;
1958    tmp = Context.isPromotableBitField(rex);
1959    if (!tmp.isNull()) rt = tmp;
1960  }
1961
1962  // The rule is that the signed operand becomes unsigned, so isolate the
1963  // signed operand.
1964  Expr *signedOperand = lex, *unsignedOperand = rex;
1965  QualType signedType = lt, unsignedType = rt;
1966  if (lt->isSignedIntegerType()) {
1967    if (rt->isSignedIntegerType()) return;
1968  } else {
1969    if (!rt->isSignedIntegerType()) return;
1970    std::swap(signedOperand, unsignedOperand);
1971    std::swap(signedType, unsignedType);
1972  }
1973
1974  unsigned unsignedWidth = Context.getIntWidth(unsignedType);
1975  unsigned signedWidth = Context.getIntWidth(signedType);
1976
1977  // If the unsigned type is strictly smaller than the signed type,
1978  // then (1) the result type will be signed and (2) the unsigned
1979  // value will fit fully within the signed type, and thus the result
1980  // of the comparison will be exact.
1981  if (signedWidth > unsignedWidth)
1982    return;
1983
1984  // Otherwise, calculate the effective ranges.
1985  IntRange signedRange = GetExprRange(Context, signedOperand, signedWidth);
1986  IntRange unsignedRange = GetExprRange(Context, unsignedOperand, unsignedWidth);
1987
1988  // We should never be unable to prove that the unsigned operand is
1989  // non-negative.
1990  assert(unsignedRange.NonNegative && "unsigned range includes negative?");
1991
1992  // If the signed operand is non-negative, then the signed->unsigned
1993  // conversion won't change it.
1994  if (signedRange.NonNegative)
1995    return;
1996
1997  // For (in)equality comparisons, if the unsigned operand is a
1998  // constant which cannot collide with a overflowed signed operand,
1999  // then reinterpreting the signed operand as unsigned will not
2000  // change the result of the comparison.
2001  if (Equality && unsignedRange.Width < unsignedWidth)
2002    return;
2003
2004  Diag(OpLoc, PD)
2005    << lt << rt << lex->getSourceRange() << rex->getSourceRange();
2006}
2007
2008/// Diagnose an implicit cast;  purely a helper for CheckImplicitConversion.
2009static void DiagnoseImpCast(Sema &S, Expr *E, QualType T, unsigned diag) {
2010  S.Diag(E->getExprLoc(), diag) << E->getType() << T << E->getSourceRange();
2011}
2012
2013/// Implements -Wconversion.
2014void Sema::CheckImplicitConversion(Expr *E, QualType T) {
2015  // Don't diagnose in unevaluated contexts.
2016  if (ExprEvalContexts.back().Context == Sema::Unevaluated)
2017    return;
2018
2019  // Don't diagnose for value-dependent expressions.
2020  if (E->isValueDependent())
2021    return;
2022
2023  const Type *Source = Context.getCanonicalType(E->getType()).getTypePtr();
2024  const Type *Target = Context.getCanonicalType(T).getTypePtr();
2025
2026  // Never diagnose implicit casts to bool.
2027  if (Target->isSpecificBuiltinType(BuiltinType::Bool))
2028    return;
2029
2030  // Strip vector types.
2031  if (isa<VectorType>(Source)) {
2032    if (!isa<VectorType>(Target))
2033      return DiagnoseImpCast(*this, E, T, diag::warn_impcast_vector_scalar);
2034
2035    Source = cast<VectorType>(Source)->getElementType().getTypePtr();
2036    Target = cast<VectorType>(Target)->getElementType().getTypePtr();
2037  }
2038
2039  // Strip complex types.
2040  if (isa<ComplexType>(Source)) {
2041    if (!isa<ComplexType>(Target))
2042      return DiagnoseImpCast(*this, E, T, diag::warn_impcast_complex_scalar);
2043
2044    Source = cast<ComplexType>(Source)->getElementType().getTypePtr();
2045    Target = cast<ComplexType>(Target)->getElementType().getTypePtr();
2046  }
2047
2048  const BuiltinType *SourceBT = dyn_cast<BuiltinType>(Source);
2049  const BuiltinType *TargetBT = dyn_cast<BuiltinType>(Target);
2050
2051  // If the source is floating point...
2052  if (SourceBT && SourceBT->isFloatingPoint()) {
2053    // ...and the target is floating point...
2054    if (TargetBT && TargetBT->isFloatingPoint()) {
2055      // ...then warn if we're dropping FP rank.
2056
2057      // Builtin FP kinds are ordered by increasing FP rank.
2058      if (SourceBT->getKind() > TargetBT->getKind()) {
2059        // Don't warn about float constants that are precisely
2060        // representable in the target type.
2061        Expr::EvalResult result;
2062        if (E->Evaluate(result, Context)) {
2063          // Value might be a float, a float vector, or a float complex.
2064          if (IsSameFloatAfterCast(result.Val,
2065                     Context.getFloatTypeSemantics(QualType(TargetBT, 0)),
2066                     Context.getFloatTypeSemantics(QualType(SourceBT, 0))))
2067            return;
2068        }
2069
2070        DiagnoseImpCast(*this, E, T, diag::warn_impcast_float_precision);
2071      }
2072      return;
2073    }
2074
2075    // If the target is integral, always warn.
2076    if ((TargetBT && TargetBT->isInteger()))
2077      // TODO: don't warn for integer values?
2078      return DiagnoseImpCast(*this, E, T, diag::warn_impcast_float_integer);
2079
2080    return;
2081  }
2082
2083  if (!Source->isIntegerType() || !Target->isIntegerType())
2084    return;
2085
2086  IntRange SourceRange = GetExprRange(Context, E, Context.getIntWidth(E->getType()));
2087  IntRange TargetRange = IntRange::forCanonicalType(Context, Target);
2088
2089  // FIXME: also signed<->unsigned?
2090
2091  if (SourceRange.Width > TargetRange.Width) {
2092    // People want to build with -Wshorten-64-to-32 and not -Wconversion
2093    // and by god we'll let them.
2094    if (SourceRange.Width == 64 && TargetRange.Width == 32)
2095      return DiagnoseImpCast(*this, E, T, diag::warn_impcast_integer_64_32);
2096    return DiagnoseImpCast(*this, E, T, diag::warn_impcast_integer_precision);
2097  }
2098
2099  return;
2100}
2101
2102// MarkLive - Mark all the blocks reachable from e as live.  Returns the total
2103// number of blocks just marked live.
2104static unsigned MarkLive(CFGBlock *e, llvm::BitVector &live) {
2105  unsigned count = 0;
2106  std::queue<CFGBlock*> workq;
2107  // Prep work queue
2108  live.set(e->getBlockID());
2109  ++count;
2110  workq.push(e);
2111  // Solve
2112  while (!workq.empty()) {
2113    CFGBlock *item = workq.front();
2114    workq.pop();
2115    for (CFGBlock::succ_iterator I=item->succ_begin(),
2116           E=item->succ_end();
2117         I != E;
2118         ++I) {
2119      if ((*I) && !live[(*I)->getBlockID()]) {
2120        live.set((*I)->getBlockID());
2121        ++count;
2122        workq.push(*I);
2123      }
2124    }
2125  }
2126  return count;
2127}
2128
2129static SourceLocation GetUnreachableLoc(CFGBlock &b, SourceRange &R1,
2130                                        SourceRange &R2) {
2131  Stmt *S;
2132  unsigned sn = 0;
2133  R1 = R2 = SourceRange();
2134
2135  top:
2136  if (sn < b.size())
2137    S = b[sn].getStmt();
2138  else if (b.getTerminator())
2139    S = b.getTerminator();
2140  else
2141    return SourceLocation();
2142
2143  switch (S->getStmtClass()) {
2144  case Expr::BinaryOperatorClass: {
2145    BinaryOperator *BO = cast<BinaryOperator>(S);
2146    if (BO->getOpcode() == BinaryOperator::Comma) {
2147      if (sn+1 < b.size())
2148        return b[sn+1].getStmt()->getLocStart();
2149      CFGBlock *n = &b;
2150      while (1) {
2151        if (n->getTerminator())
2152          return n->getTerminator()->getLocStart();
2153        if (n->succ_size() != 1)
2154          return SourceLocation();
2155        n = n[0].succ_begin()[0];
2156        if (n->pred_size() != 1)
2157          return SourceLocation();
2158        if (!n->empty())
2159          return n[0][0].getStmt()->getLocStart();
2160      }
2161    }
2162    R1 = BO->getLHS()->getSourceRange();
2163    R2 = BO->getRHS()->getSourceRange();
2164    return BO->getOperatorLoc();
2165  }
2166  case Expr::UnaryOperatorClass: {
2167    const UnaryOperator *UO = cast<UnaryOperator>(S);
2168    R1 = UO->getSubExpr()->getSourceRange();
2169    return UO->getOperatorLoc();
2170  }
2171  case Expr::CompoundAssignOperatorClass: {
2172    const CompoundAssignOperator *CAO = cast<CompoundAssignOperator>(S);
2173    R1 = CAO->getLHS()->getSourceRange();
2174    R2 = CAO->getRHS()->getSourceRange();
2175    return CAO->getOperatorLoc();
2176  }
2177  case Expr::ConditionalOperatorClass: {
2178    const ConditionalOperator *CO = cast<ConditionalOperator>(S);
2179    return CO->getQuestionLoc();
2180  }
2181  case Expr::MemberExprClass: {
2182    const MemberExpr *ME = cast<MemberExpr>(S);
2183    R1 = ME->getSourceRange();
2184    return ME->getMemberLoc();
2185  }
2186  case Expr::ArraySubscriptExprClass: {
2187    const ArraySubscriptExpr *ASE = cast<ArraySubscriptExpr>(S);
2188    R1 = ASE->getLHS()->getSourceRange();
2189    R2 = ASE->getRHS()->getSourceRange();
2190    return ASE->getRBracketLoc();
2191  }
2192  case Expr::CStyleCastExprClass: {
2193    const CStyleCastExpr *CSC = cast<CStyleCastExpr>(S);
2194    R1 = CSC->getSubExpr()->getSourceRange();
2195    return CSC->getLParenLoc();
2196  }
2197  case Expr::CXXFunctionalCastExprClass: {
2198    const CXXFunctionalCastExpr *CE = cast <CXXFunctionalCastExpr>(S);
2199    R1 = CE->getSubExpr()->getSourceRange();
2200    return CE->getTypeBeginLoc();
2201  }
2202  case Expr::ImplicitCastExprClass:
2203    ++sn;
2204    goto top;
2205  case Stmt::CXXTryStmtClass: {
2206    return cast<CXXTryStmt>(S)->getHandler(0)->getCatchLoc();
2207  }
2208  default: ;
2209  }
2210  R1 = S->getSourceRange();
2211  return S->getLocStart();
2212}
2213
2214static SourceLocation MarkLiveTop(CFGBlock *e, llvm::BitVector &live,
2215                               SourceManager &SM) {
2216  std::queue<CFGBlock*> workq;
2217  // Prep work queue
2218  workq.push(e);
2219  SourceRange R1, R2;
2220  SourceLocation top = GetUnreachableLoc(*e, R1, R2);
2221  bool FromMainFile = false;
2222  bool FromSystemHeader = false;
2223  bool TopValid = false;
2224  if (top.isValid()) {
2225    FromMainFile = SM.isFromMainFile(top);
2226    FromSystemHeader = SM.isInSystemHeader(top);
2227    TopValid = true;
2228  }
2229  // Solve
2230  while (!workq.empty()) {
2231    CFGBlock *item = workq.front();
2232    workq.pop();
2233    SourceLocation c = GetUnreachableLoc(*item, R1, R2);
2234    if (c.isValid()
2235        && (!TopValid
2236            || (SM.isFromMainFile(c) && !FromMainFile)
2237            || (FromSystemHeader && !SM.isInSystemHeader(c))
2238            || SM.isBeforeInTranslationUnit(c, top))) {
2239      top = c;
2240      FromMainFile = SM.isFromMainFile(top);
2241      FromSystemHeader = SM.isInSystemHeader(top);
2242    }
2243    live.set(item->getBlockID());
2244    for (CFGBlock::succ_iterator I=item->succ_begin(),
2245           E=item->succ_end();
2246         I != E;
2247         ++I) {
2248      if ((*I) && !live[(*I)->getBlockID()]) {
2249        live.set((*I)->getBlockID());
2250        workq.push(*I);
2251      }
2252    }
2253  }
2254  return top;
2255}
2256
2257static int LineCmp(const void *p1, const void *p2) {
2258  SourceLocation *Line1 = (SourceLocation *)p1;
2259  SourceLocation *Line2 = (SourceLocation *)p2;
2260  return !(*Line1 < *Line2);
2261}
2262
2263namespace {
2264  struct ErrLoc {
2265    SourceLocation Loc;
2266    SourceRange R1;
2267    SourceRange R2;
2268    ErrLoc(SourceLocation l, SourceRange r1, SourceRange r2)
2269      : Loc(l), R1(r1), R2(r2) { }
2270  };
2271}
2272
2273/// CheckUnreachable - Check for unreachable code.
2274void Sema::CheckUnreachable(AnalysisContext &AC) {
2275  unsigned count;
2276  // We avoid checking when there are errors, as the CFG won't faithfully match
2277  // the user's code.
2278  if (getDiagnostics().hasErrorOccurred())
2279    return;
2280  if (Diags.getDiagnosticLevel(diag::warn_unreachable) == Diagnostic::Ignored)
2281    return;
2282
2283  CFG *cfg = AC.getCFG();
2284  if (cfg == 0)
2285    return;
2286
2287  llvm::BitVector live(cfg->getNumBlockIDs());
2288  // Mark all live things first.
2289  count = MarkLive(&cfg->getEntry(), live);
2290
2291  if (count == cfg->getNumBlockIDs())
2292    // If there are no dead blocks, we're done.
2293    return;
2294
2295  SourceRange R1, R2;
2296
2297  llvm::SmallVector<ErrLoc, 24> lines;
2298  bool AddEHEdges = AC.getAddEHEdges();
2299  // First, give warnings for blocks with no predecessors, as they
2300  // can't be part of a loop.
2301  for (CFG::iterator I = cfg->begin(), E = cfg->end(); I != E; ++I) {
2302    CFGBlock &b = **I;
2303    if (!live[b.getBlockID()]) {
2304      if (b.pred_begin() == b.pred_end()) {
2305        if (!AddEHEdges && b.getTerminator()
2306            && isa<CXXTryStmt>(b.getTerminator())) {
2307          // When not adding EH edges from calls, catch clauses
2308          // can otherwise seem dead.  Avoid noting them as dead.
2309          count += MarkLive(&b, live);
2310          continue;
2311        }
2312        SourceLocation c = GetUnreachableLoc(b, R1, R2);
2313        if (!c.isValid()) {
2314          // Blocks without a location can't produce a warning, so don't mark
2315          // reachable blocks from here as live.
2316          live.set(b.getBlockID());
2317          ++count;
2318          continue;
2319        }
2320        lines.push_back(ErrLoc(c, R1, R2));
2321        // Avoid excessive errors by marking everything reachable from here
2322        count += MarkLive(&b, live);
2323      }
2324    }
2325  }
2326
2327  if (count < cfg->getNumBlockIDs()) {
2328    // And then give warnings for the tops of loops.
2329    for (CFG::iterator I = cfg->begin(), E = cfg->end(); I != E; ++I) {
2330      CFGBlock &b = **I;
2331      if (!live[b.getBlockID()])
2332        // Avoid excessive errors by marking everything reachable from here
2333        lines.push_back(ErrLoc(MarkLiveTop(&b, live,
2334                                           Context.getSourceManager()),
2335                               SourceRange(), SourceRange()));
2336    }
2337  }
2338
2339  llvm::array_pod_sort(lines.begin(), lines.end(), LineCmp);
2340  for (llvm::SmallVector<ErrLoc, 24>::iterator I = lines.begin(),
2341         E = lines.end();
2342       I != E;
2343       ++I)
2344    if (I->Loc.isValid())
2345      Diag(I->Loc, diag::warn_unreachable) << I->R1 << I->R2;
2346}
2347
2348/// CheckFallThrough - Check that we don't fall off the end of a
2349/// Statement that should return a value.
2350///
2351/// \returns AlwaysFallThrough iff we always fall off the end of the statement,
2352/// MaybeFallThrough iff we might or might not fall off the end,
2353/// NeverFallThroughOrReturn iff we never fall off the end of the statement or
2354/// return.  We assume NeverFallThrough iff we never fall off the end of the
2355/// statement but we may return.  We assume that functions not marked noreturn
2356/// will return.
2357Sema::ControlFlowKind Sema::CheckFallThrough(AnalysisContext &AC) {
2358  CFG *cfg = AC.getCFG();
2359  if (cfg == 0)
2360    // FIXME: This should be NeverFallThrough
2361    return NeverFallThroughOrReturn;
2362
2363  // The CFG leaves in dead things, and we don't want the dead code paths to
2364  // confuse us, so we mark all live things first.
2365  std::queue<CFGBlock*> workq;
2366  llvm::BitVector live(cfg->getNumBlockIDs());
2367  unsigned count = MarkLive(&cfg->getEntry(), live);
2368
2369  bool AddEHEdges = AC.getAddEHEdges();
2370  if (!AddEHEdges && count != cfg->getNumBlockIDs())
2371    // When there are things remaining dead, and we didn't add EH edges
2372    // from CallExprs to the catch clauses, we have to go back and
2373    // mark them as live.
2374    for (CFG::iterator I = cfg->begin(), E = cfg->end(); I != E; ++I) {
2375      CFGBlock &b = **I;
2376      if (!live[b.getBlockID()]) {
2377        if (b.pred_begin() == b.pred_end()) {
2378          if (b.getTerminator() && isa<CXXTryStmt>(b.getTerminator()))
2379            // When not adding EH edges from calls, catch clauses
2380            // can otherwise seem dead.  Avoid noting them as dead.
2381            count += MarkLive(&b, live);
2382          continue;
2383        }
2384      }
2385    }
2386
2387  // Now we know what is live, we check the live precessors of the exit block
2388  // and look for fall through paths, being careful to ignore normal returns,
2389  // and exceptional paths.
2390  bool HasLiveReturn = false;
2391  bool HasFakeEdge = false;
2392  bool HasPlainEdge = false;
2393  bool HasAbnormalEdge = false;
2394  for (CFGBlock::pred_iterator I=cfg->getExit().pred_begin(),
2395         E = cfg->getExit().pred_end();
2396       I != E;
2397       ++I) {
2398    CFGBlock& B = **I;
2399    if (!live[B.getBlockID()])
2400      continue;
2401    if (B.size() == 0) {
2402      if (B.getTerminator() && isa<CXXTryStmt>(B.getTerminator())) {
2403        HasAbnormalEdge = true;
2404        continue;
2405      }
2406
2407      // A labeled empty statement, or the entry block...
2408      HasPlainEdge = true;
2409      continue;
2410    }
2411    Stmt *S = B[B.size()-1];
2412    if (isa<ReturnStmt>(S)) {
2413      HasLiveReturn = true;
2414      continue;
2415    }
2416    if (isa<ObjCAtThrowStmt>(S)) {
2417      HasFakeEdge = true;
2418      continue;
2419    }
2420    if (isa<CXXThrowExpr>(S)) {
2421      HasFakeEdge = true;
2422      continue;
2423    }
2424    if (const AsmStmt *AS = dyn_cast<AsmStmt>(S)) {
2425      if (AS->isMSAsm()) {
2426        HasFakeEdge = true;
2427        HasLiveReturn = true;
2428        continue;
2429      }
2430    }
2431    if (isa<CXXTryStmt>(S)) {
2432      HasAbnormalEdge = true;
2433      continue;
2434    }
2435
2436    bool NoReturnEdge = false;
2437    if (CallExpr *C = dyn_cast<CallExpr>(S)) {
2438      if (B.succ_begin()[0] != &cfg->getExit()) {
2439        HasAbnormalEdge = true;
2440        continue;
2441      }
2442      Expr *CEE = C->getCallee()->IgnoreParenCasts();
2443      if (CEE->getType().getNoReturnAttr()) {
2444        NoReturnEdge = true;
2445        HasFakeEdge = true;
2446      } else if (DeclRefExpr *DRE = dyn_cast<DeclRefExpr>(CEE)) {
2447        ValueDecl *VD = DRE->getDecl();
2448        if (VD->hasAttr<NoReturnAttr>()) {
2449          NoReturnEdge = true;
2450          HasFakeEdge = true;
2451        }
2452      }
2453    }
2454    // FIXME: Add noreturn message sends.
2455    if (NoReturnEdge == false)
2456      HasPlainEdge = true;
2457  }
2458  if (!HasPlainEdge) {
2459    if (HasLiveReturn)
2460      return NeverFallThrough;
2461    return NeverFallThroughOrReturn;
2462  }
2463  if (HasAbnormalEdge || HasFakeEdge || HasLiveReturn)
2464    return MaybeFallThrough;
2465  // This says AlwaysFallThrough for calls to functions that are not marked
2466  // noreturn, that don't return.  If people would like this warning to be more
2467  // accurate, such functions should be marked as noreturn.
2468  return AlwaysFallThrough;
2469}
2470
2471/// CheckFallThroughForFunctionDef - Check that we don't fall off the end of a
2472/// function that should return a value.  Check that we don't fall off the end
2473/// of a noreturn function.  We assume that functions and blocks not marked
2474/// noreturn will return.
2475void Sema::CheckFallThroughForFunctionDef(Decl *D, Stmt *Body,
2476                                          AnalysisContext &AC) {
2477  // FIXME: Would be nice if we had a better way to control cascading errors,
2478  // but for now, avoid them.  The problem is that when Parse sees:
2479  //   int foo() { return a; }
2480  // The return is eaten and the Sema code sees just:
2481  //   int foo() { }
2482  // which this code would then warn about.
2483  if (getDiagnostics().hasErrorOccurred())
2484    return;
2485
2486  bool ReturnsVoid = false;
2487  bool HasNoReturn = false;
2488  if (FunctionDecl *FD = dyn_cast<FunctionDecl>(D)) {
2489    // If the result type of the function is a dependent type, we don't know
2490    // whether it will be void or not, so don't
2491    if (FD->getResultType()->isDependentType())
2492      return;
2493    if (FD->getResultType()->isVoidType())
2494      ReturnsVoid = true;
2495    if (FD->hasAttr<NoReturnAttr>())
2496      HasNoReturn = true;
2497  } else if (ObjCMethodDecl *MD = dyn_cast<ObjCMethodDecl>(D)) {
2498    if (MD->getResultType()->isVoidType())
2499      ReturnsVoid = true;
2500    if (MD->hasAttr<NoReturnAttr>())
2501      HasNoReturn = true;
2502  }
2503
2504  // Short circuit for compilation speed.
2505  if ((Diags.getDiagnosticLevel(diag::warn_maybe_falloff_nonvoid_function)
2506       == Diagnostic::Ignored || ReturnsVoid)
2507      && (Diags.getDiagnosticLevel(diag::warn_noreturn_function_has_return_expr)
2508          == Diagnostic::Ignored || !HasNoReturn)
2509      && (Diags.getDiagnosticLevel(diag::warn_suggest_noreturn_block)
2510          == Diagnostic::Ignored || !ReturnsVoid))
2511    return;
2512  // FIXME: Function try block
2513  if (CompoundStmt *Compound = dyn_cast<CompoundStmt>(Body)) {
2514    switch (CheckFallThrough(AC)) {
2515    case MaybeFallThrough:
2516      if (HasNoReturn)
2517        Diag(Compound->getRBracLoc(), diag::warn_falloff_noreturn_function);
2518      else if (!ReturnsVoid)
2519        Diag(Compound->getRBracLoc(),diag::warn_maybe_falloff_nonvoid_function);
2520      break;
2521    case AlwaysFallThrough:
2522      if (HasNoReturn)
2523        Diag(Compound->getRBracLoc(), diag::warn_falloff_noreturn_function);
2524      else if (!ReturnsVoid)
2525        Diag(Compound->getRBracLoc(), diag::warn_falloff_nonvoid_function);
2526      break;
2527    case NeverFallThroughOrReturn:
2528      if (ReturnsVoid && !HasNoReturn)
2529        Diag(Compound->getLBracLoc(), diag::warn_suggest_noreturn_function);
2530      break;
2531    case NeverFallThrough:
2532      break;
2533    }
2534  }
2535}
2536
2537/// CheckFallThroughForBlock - Check that we don't fall off the end of a block
2538/// that should return a value.  Check that we don't fall off the end of a
2539/// noreturn block.  We assume that functions and blocks not marked noreturn
2540/// will return.
2541void Sema::CheckFallThroughForBlock(QualType BlockTy, Stmt *Body,
2542                                    AnalysisContext &AC) {
2543  // FIXME: Would be nice if we had a better way to control cascading errors,
2544  // but for now, avoid them.  The problem is that when Parse sees:
2545  //   int foo() { return a; }
2546  // The return is eaten and the Sema code sees just:
2547  //   int foo() { }
2548  // which this code would then warn about.
2549  if (getDiagnostics().hasErrorOccurred())
2550    return;
2551  bool ReturnsVoid = false;
2552  bool HasNoReturn = false;
2553  if (const FunctionType *FT =BlockTy->getPointeeType()->getAs<FunctionType>()){
2554    if (FT->getResultType()->isVoidType())
2555      ReturnsVoid = true;
2556    if (FT->getNoReturnAttr())
2557      HasNoReturn = true;
2558  }
2559
2560  // Short circuit for compilation speed.
2561  if (ReturnsVoid
2562      && !HasNoReturn
2563      && (Diags.getDiagnosticLevel(diag::warn_suggest_noreturn_block)
2564          == Diagnostic::Ignored || !ReturnsVoid))
2565    return;
2566  // FIXME: Funtion try block
2567  if (CompoundStmt *Compound = dyn_cast<CompoundStmt>(Body)) {
2568    switch (CheckFallThrough(AC)) {
2569    case MaybeFallThrough:
2570      if (HasNoReturn)
2571        Diag(Compound->getRBracLoc(), diag::err_noreturn_block_has_return_expr);
2572      else if (!ReturnsVoid)
2573        Diag(Compound->getRBracLoc(), diag::err_maybe_falloff_nonvoid_block);
2574      break;
2575    case AlwaysFallThrough:
2576      if (HasNoReturn)
2577        Diag(Compound->getRBracLoc(), diag::err_noreturn_block_has_return_expr);
2578      else if (!ReturnsVoid)
2579        Diag(Compound->getRBracLoc(), diag::err_falloff_nonvoid_block);
2580      break;
2581    case NeverFallThroughOrReturn:
2582      if (ReturnsVoid)
2583        Diag(Compound->getLBracLoc(), diag::warn_suggest_noreturn_block);
2584      break;
2585    case NeverFallThrough:
2586      break;
2587    }
2588  }
2589}
2590
2591/// CheckParmsForFunctionDef - Check that the parameters of the given
2592/// function are appropriate for the definition of a function. This
2593/// takes care of any checks that cannot be performed on the
2594/// declaration itself, e.g., that the types of each of the function
2595/// parameters are complete.
2596bool Sema::CheckParmsForFunctionDef(FunctionDecl *FD) {
2597  bool HasInvalidParm = false;
2598  for (unsigned p = 0, NumParams = FD->getNumParams(); p < NumParams; ++p) {
2599    ParmVarDecl *Param = FD->getParamDecl(p);
2600
2601    // C99 6.7.5.3p4: the parameters in a parameter type list in a
2602    // function declarator that is part of a function definition of
2603    // that function shall not have incomplete type.
2604    //
2605    // This is also C++ [dcl.fct]p6.
2606    if (!Param->isInvalidDecl() &&
2607        RequireCompleteType(Param->getLocation(), Param->getType(),
2608                               diag::err_typecheck_decl_incomplete_type)) {
2609      Param->setInvalidDecl();
2610      HasInvalidParm = true;
2611    }
2612
2613    // C99 6.9.1p5: If the declarator includes a parameter type list, the
2614    // declaration of each parameter shall include an identifier.
2615    if (Param->getIdentifier() == 0 &&
2616        !Param->isImplicit() &&
2617        !getLangOptions().CPlusPlus)
2618      Diag(Param->getLocation(), diag::err_parameter_name_omitted);
2619  }
2620
2621  return HasInvalidParm;
2622}
2623