SemaChecking.cpp revision 0d27735c51f5bd392e673cf39a675e14e9442387
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
1036void Sema::CheckPrintfString(const StringLiteral *FExpr,
1037                             const Expr *OrigFormatExpr,
1038                             const CallExpr *TheCall, bool HasVAListArg,
1039                             unsigned format_idx, unsigned firstDataArg) {
1040
1041  static bool UseAlternatePrintfChecking = false;
1042  if (UseAlternatePrintfChecking) {
1043    AlternateCheckPrintfString(FExpr, OrigFormatExpr, TheCall,
1044                               HasVAListArg, format_idx, firstDataArg);
1045    return;
1046  }
1047
1048
1049  const ObjCStringLiteral *ObjCFExpr =
1050    dyn_cast<ObjCStringLiteral>(OrigFormatExpr);
1051
1052  // CHECK: is the format string a wide literal?
1053  if (FExpr->isWide()) {
1054    Diag(FExpr->getLocStart(),
1055         diag::warn_printf_format_string_is_wide_literal)
1056      << OrigFormatExpr->getSourceRange();
1057    return;
1058  }
1059
1060  // Str - The format string.  NOTE: this is NOT null-terminated!
1061  const char *Str = FExpr->getStrData();
1062
1063  // CHECK: empty format string?
1064  unsigned StrLen = FExpr->getByteLength();
1065
1066  if (StrLen == 0) {
1067    Diag(FExpr->getLocStart(), diag::warn_printf_empty_format_string)
1068      << OrigFormatExpr->getSourceRange();
1069    return;
1070  }
1071
1072  // We process the format string using a binary state machine.  The
1073  // current state is stored in CurrentState.
1074  enum {
1075    state_OrdChr,
1076    state_Conversion
1077  } CurrentState = state_OrdChr;
1078
1079  // numConversions - The number of conversions seen so far.  This is
1080  //  incremented as we traverse the format string.
1081  unsigned numConversions = 0;
1082
1083  // numDataArgs - The number of data arguments after the format
1084  //  string.  This can only be determined for non vprintf-like
1085  //  functions.  For those functions, this value is 1 (the sole
1086  //  va_arg argument).
1087  unsigned numDataArgs = TheCall->getNumArgs()-firstDataArg;
1088
1089  // Inspect the format string.
1090  unsigned StrIdx = 0;
1091
1092  // LastConversionIdx - Index within the format string where we last saw
1093  //  a '%' character that starts a new format conversion.
1094  unsigned LastConversionIdx = 0;
1095
1096  for (; StrIdx < StrLen; ++StrIdx) {
1097
1098    // Is the number of detected conversion conversions greater than
1099    // the number of matching data arguments?  If so, stop.
1100    if (!HasVAListArg && numConversions > numDataArgs) break;
1101
1102    // Handle "\0"
1103    if (Str[StrIdx] == '\0') {
1104      // The string returned by getStrData() is not null-terminated,
1105      // so the presence of a null character is likely an error.
1106      Diag(getLocationOfStringLiteralByte(FExpr, StrIdx),
1107           diag::warn_printf_format_string_contains_null_char)
1108        <<  OrigFormatExpr->getSourceRange();
1109      return;
1110    }
1111
1112    // Ordinary characters (not processing a format conversion).
1113    if (CurrentState == state_OrdChr) {
1114      if (Str[StrIdx] == '%') {
1115        CurrentState = state_Conversion;
1116        LastConversionIdx = StrIdx;
1117      }
1118      continue;
1119    }
1120
1121    // Seen '%'.  Now processing a format conversion.
1122    switch (Str[StrIdx]) {
1123    // Handle dynamic precision or width specifier.
1124    case '*': {
1125      ++numConversions;
1126
1127      if (!HasVAListArg) {
1128        if (numConversions > numDataArgs) {
1129          SourceLocation Loc = getLocationOfStringLiteralByte(FExpr, StrIdx);
1130
1131          if (Str[StrIdx-1] == '.')
1132            Diag(Loc, diag::warn_printf_asterisk_precision_missing_arg)
1133              << OrigFormatExpr->getSourceRange();
1134          else
1135            Diag(Loc, diag::warn_printf_asterisk_width_missing_arg)
1136              << OrigFormatExpr->getSourceRange();
1137
1138          // Don't do any more checking.  We'll just emit spurious errors.
1139          return;
1140        }
1141
1142        // Perform type checking on width/precision specifier.
1143        const Expr *E = TheCall->getArg(format_idx+numConversions);
1144        if (const BuiltinType *BT = E->getType()->getAs<BuiltinType>())
1145          if (BT->getKind() == BuiltinType::Int)
1146            break;
1147
1148        SourceLocation Loc = getLocationOfStringLiteralByte(FExpr, StrIdx);
1149
1150        if (Str[StrIdx-1] == '.')
1151          Diag(Loc, diag::warn_printf_asterisk_precision_wrong_type)
1152          << E->getType() << E->getSourceRange();
1153        else
1154          Diag(Loc, diag::warn_printf_asterisk_width_wrong_type)
1155          << E->getType() << E->getSourceRange();
1156
1157        break;
1158      }
1159    }
1160
1161    // Characters which can terminate a format conversion
1162    // (e.g. "%d").  Characters that specify length modifiers or
1163    // other flags are handled by the default case below.
1164    //
1165    // FIXME: additional checks will go into the following cases.
1166    case 'i':
1167    case 'd':
1168    case 'o':
1169    case 'u':
1170    case 'x':
1171    case 'X':
1172    case 'e':
1173    case 'E':
1174    case 'f':
1175    case 'F':
1176    case 'g':
1177    case 'G':
1178    case 'a':
1179    case 'A':
1180    case 'c':
1181    case 's':
1182    case 'p':
1183      ++numConversions;
1184      CurrentState = state_OrdChr;
1185      break;
1186
1187    case 'm':
1188      // FIXME: Warn in situations where this isn't supported!
1189      CurrentState = state_OrdChr;
1190      break;
1191
1192    // CHECK: Are we using "%n"?  Issue a warning.
1193    case 'n': {
1194      ++numConversions;
1195      CurrentState = state_OrdChr;
1196      SourceLocation Loc = getLocationOfStringLiteralByte(FExpr,
1197                                                          LastConversionIdx);
1198
1199      Diag(Loc, diag::warn_printf_write_back)<<OrigFormatExpr->getSourceRange();
1200      break;
1201    }
1202
1203    // Handle "%@"
1204    case '@':
1205      // %@ is allowed in ObjC format strings only.
1206      if (ObjCFExpr != NULL)
1207        CurrentState = state_OrdChr;
1208      else {
1209        // Issue a warning: invalid format conversion.
1210        SourceLocation Loc =
1211          getLocationOfStringLiteralByte(FExpr, LastConversionIdx);
1212
1213        Diag(Loc, diag::warn_printf_invalid_conversion)
1214          <<  std::string(Str+LastConversionIdx,
1215                          Str+std::min(LastConversionIdx+2, StrLen))
1216          << OrigFormatExpr->getSourceRange();
1217      }
1218      ++numConversions;
1219      break;
1220
1221    // Handle "%%"
1222    case '%':
1223      // Sanity check: Was the first "%" character the previous one?
1224      // If not, we will assume that we have a malformed format
1225      // conversion, and that the current "%" character is the start
1226      // of a new conversion.
1227      if (StrIdx - LastConversionIdx == 1)
1228        CurrentState = state_OrdChr;
1229      else {
1230        // Issue a warning: invalid format conversion.
1231        SourceLocation Loc =
1232          getLocationOfStringLiteralByte(FExpr, LastConversionIdx);
1233
1234        Diag(Loc, diag::warn_printf_invalid_conversion)
1235          << std::string(Str+LastConversionIdx, Str+StrIdx)
1236          << OrigFormatExpr->getSourceRange();
1237
1238        // This conversion is broken.  Advance to the next format
1239        // conversion.
1240        LastConversionIdx = StrIdx;
1241        ++numConversions;
1242      }
1243      break;
1244
1245    default:
1246      // This case catches all other characters: flags, widths, etc.
1247      // We should eventually process those as well.
1248      break;
1249    }
1250  }
1251
1252  if (CurrentState == state_Conversion) {
1253    // Issue a warning: invalid format conversion.
1254    SourceLocation Loc =
1255      getLocationOfStringLiteralByte(FExpr, LastConversionIdx);
1256
1257    Diag(Loc, diag::warn_printf_invalid_conversion)
1258      << std::string(Str+LastConversionIdx,
1259                     Str+std::min(LastConversionIdx+2, StrLen))
1260      << OrigFormatExpr->getSourceRange();
1261    return;
1262  }
1263
1264  if (!HasVAListArg) {
1265    // CHECK: Does the number of format conversions exceed the number
1266    //        of data arguments?
1267    if (numConversions > numDataArgs) {
1268      SourceLocation Loc =
1269        getLocationOfStringLiteralByte(FExpr, LastConversionIdx);
1270
1271      Diag(Loc, diag::warn_printf_insufficient_data_args)
1272        << OrigFormatExpr->getSourceRange();
1273    }
1274    // CHECK: Does the number of data arguments exceed the number of
1275    //        format conversions in the format string?
1276    else if (numConversions < numDataArgs)
1277      Diag(TheCall->getArg(format_idx+numConversions+1)->getLocStart(),
1278           diag::warn_printf_too_many_data_args)
1279        << OrigFormatExpr->getSourceRange();
1280  }
1281}
1282
1283
1284namespace {
1285class CheckPrintfHandler : public analyze_printf::FormatStringHandler {
1286  Sema &S;
1287  const StringLiteral *FExpr;
1288  const Expr *OrigFormatExpr;
1289  unsigned NumConversions;
1290  const unsigned NumDataArgs;
1291  const bool IsObjCLiteral;
1292  const char *Beg; // Start of format string.
1293  const bool HasVAListArg;
1294  const CallExpr *TheCall;
1295  unsigned FormatIdx;
1296public:
1297  CheckPrintfHandler(Sema &s, const StringLiteral *fexpr,
1298                     const Expr *origFormatExpr,
1299                     unsigned numDataArgs, bool isObjCLiteral,
1300                     const char *beg, bool hasVAListArg,
1301                     const CallExpr *theCall, unsigned formatIdx)
1302    : S(s), FExpr(fexpr), OrigFormatExpr(origFormatExpr),
1303      NumConversions(0), NumDataArgs(numDataArgs),
1304      IsObjCLiteral(isObjCLiteral), Beg(beg),
1305      HasVAListArg(hasVAListArg),
1306      TheCall(theCall), FormatIdx(formatIdx) {}
1307
1308  void HandleNullChar(const char *nullCharacter);
1309
1310  bool HandleFormatSpecifier(const analyze_printf::FormatSpecifier &FS,
1311                             const char *startSpecifier,
1312                             unsigned specifierLen);
1313private:
1314  SourceRange getFormatRange();
1315  SourceLocation getLocationOfByte(const char *x);
1316
1317  bool HandleAmount(const analyze_printf::OptionalAmount &Amt,
1318                    unsigned MissingArgDiag, unsigned BadTypeDiag);
1319
1320  const Expr *getDataArg(unsigned i) const;
1321};
1322}
1323
1324SourceRange CheckPrintfHandler::getFormatRange() {
1325  return OrigFormatExpr->getSourceRange();
1326}
1327
1328SourceLocation CheckPrintfHandler::getLocationOfByte(const char *x) {
1329  return S.getLocationOfStringLiteralByte(FExpr, x - Beg);
1330}
1331
1332void CheckPrintfHandler::HandleNullChar(const char *nullCharacter) {
1333  // The presence of a null character is likely an error.
1334  S.Diag(getLocationOfByte(nullCharacter),
1335         diag::warn_printf_format_string_contains_null_char)
1336    << getFormatRange();
1337}
1338
1339const Expr *CheckPrintfHandler::getDataArg(unsigned i) const {
1340  return TheCall->getArg(FormatIdx + i);
1341}
1342
1343bool
1344CheckPrintfHandler::HandleAmount(const analyze_printf::OptionalAmount &Amt,
1345                                 unsigned MissingArgDiag,
1346                                 unsigned BadTypeDiag) {
1347
1348  if (Amt.hasDataArgument()) {
1349    ++NumConversions;
1350    if (!HasVAListArg) {
1351      if (NumConversions > NumDataArgs) {
1352        S.Diag(getLocationOfByte(Amt.getStart()), MissingArgDiag)
1353          << getFormatRange();
1354        // Don't do any more checking.  We will just emit
1355        // spurious errors.
1356        return false;
1357      }
1358
1359      // Type check the data argument.  It should be an 'int'.
1360      const Expr *Arg = getDataArg(NumConversions);
1361      QualType T = Arg->getType();
1362      const BuiltinType *BT = T->getAs<BuiltinType>();
1363      if (!BT || BT->getKind() != BuiltinType::Int) {
1364        S.Diag(getLocationOfByte(Amt.getStart()), BadTypeDiag)
1365          << T << getFormatRange() << Arg->getSourceRange();
1366        // Don't do any more checking.  We will just emit
1367        // spurious errors.
1368        return false;
1369      }
1370    }
1371  }
1372  return true;
1373}
1374
1375
1376bool
1377CheckPrintfHandler::HandleFormatSpecifier(const analyze_printf::FormatSpecifier &FS,
1378                                          const char *startSpecifier,
1379                                          unsigned specifierLen) {
1380
1381  using namespace analyze_printf;
1382  const ConversionSpecifier &CS = FS.getConversionSpecifier();
1383
1384  // First check if the field width, precision, and conversion specifier
1385  // have matching data arguments.
1386  if (!HandleAmount(FS.getFieldWidth(),
1387                    diag::warn_printf_asterisk_width_missing_arg,
1388                    diag::warn_printf_asterisk_width_wrong_type)) {
1389    return false;
1390  }
1391
1392  if (!HandleAmount(FS.getPrecision(),
1393                    diag::warn_printf_asterisk_precision_missing_arg,
1394                    diag::warn_printf_asterisk_precision_wrong_type)) {
1395    return false;
1396  }
1397
1398  ++NumConversions;
1399
1400  // Check for using an Objective-C specific conversion specifier
1401  // in a non-ObjC literal.
1402  if (!IsObjCLiteral && CS.isObjCArg()) {
1403    SourceLocation Loc = getLocationOfByte(CS.getConversionStart());
1404    S.Diag(Loc, diag::warn_printf_invalid_conversion)
1405      << llvm::StringRef(startSpecifier, specifierLen)
1406      << getFormatRange();
1407
1408    // Continue checking the other format specifiers.
1409    return true;
1410  }
1411
1412  return true;
1413}
1414
1415
1416void
1417Sema::AlternateCheckPrintfString(const StringLiteral *FExpr,
1418                                 const Expr *OrigFormatExpr,
1419                                 const CallExpr *TheCall, bool HasVAListArg,
1420                                 unsigned format_idx, unsigned firstDataArg) {
1421
1422  // CHECK: is the format string a wide literal?
1423  if (FExpr->isWide()) {
1424    Diag(FExpr->getLocStart(),
1425         diag::warn_printf_format_string_is_wide_literal)
1426    << OrigFormatExpr->getSourceRange();
1427    return;
1428  }
1429
1430  // Str - The format string.  NOTE: this is NOT null-terminated!
1431  const char *Str = FExpr->getStrData();
1432
1433  // CHECK: empty format string?
1434  unsigned StrLen = FExpr->getByteLength();
1435
1436  if (StrLen == 0) {
1437    Diag(FExpr->getLocStart(), diag::warn_printf_empty_format_string)
1438    << OrigFormatExpr->getSourceRange();
1439    return;
1440  }
1441
1442  CheckPrintfHandler H(*this, FExpr, OrigFormatExpr,
1443                       TheCall->getNumArgs() - firstDataArg,
1444                       isa<ObjCStringLiteral>(OrigFormatExpr), Str,
1445                       HasVAListArg, TheCall, format_idx);
1446
1447  analyze_printf::ParseFormatString(H, Str, Str + StrLen);
1448}
1449
1450//===--- CHECK: Return Address of Stack Variable --------------------------===//
1451
1452static DeclRefExpr* EvalVal(Expr *E);
1453static DeclRefExpr* EvalAddr(Expr* E);
1454
1455/// CheckReturnStackAddr - Check if a return statement returns the address
1456///   of a stack variable.
1457void
1458Sema::CheckReturnStackAddr(Expr *RetValExp, QualType lhsType,
1459                           SourceLocation ReturnLoc) {
1460
1461  // Perform checking for returned stack addresses.
1462  if (lhsType->isPointerType() || lhsType->isBlockPointerType()) {
1463    if (DeclRefExpr *DR = EvalAddr(RetValExp))
1464      Diag(DR->getLocStart(), diag::warn_ret_stack_addr)
1465       << DR->getDecl()->getDeclName() << RetValExp->getSourceRange();
1466
1467    // Skip over implicit cast expressions when checking for block expressions.
1468    RetValExp = RetValExp->IgnoreParenCasts();
1469
1470    if (BlockExpr *C = dyn_cast<BlockExpr>(RetValExp))
1471      if (C->hasBlockDeclRefExprs())
1472        Diag(C->getLocStart(), diag::err_ret_local_block)
1473          << C->getSourceRange();
1474
1475    if (AddrLabelExpr *ALE = dyn_cast<AddrLabelExpr>(RetValExp))
1476      Diag(ALE->getLocStart(), diag::warn_ret_addr_label)
1477        << ALE->getSourceRange();
1478
1479  } else if (lhsType->isReferenceType()) {
1480    // Perform checking for stack values returned by reference.
1481    // Check for a reference to the stack
1482    if (DeclRefExpr *DR = EvalVal(RetValExp))
1483      Diag(DR->getLocStart(), diag::warn_ret_stack_ref)
1484        << DR->getDecl()->getDeclName() << RetValExp->getSourceRange();
1485  }
1486}
1487
1488/// EvalAddr - EvalAddr and EvalVal are mutually recursive functions that
1489///  check if the expression in a return statement evaluates to an address
1490///  to a location on the stack.  The recursion is used to traverse the
1491///  AST of the return expression, with recursion backtracking when we
1492///  encounter a subexpression that (1) clearly does not lead to the address
1493///  of a stack variable or (2) is something we cannot determine leads to
1494///  the address of a stack variable based on such local checking.
1495///
1496///  EvalAddr processes expressions that are pointers that are used as
1497///  references (and not L-values).  EvalVal handles all other values.
1498///  At the base case of the recursion is a check for a DeclRefExpr* in
1499///  the refers to a stack variable.
1500///
1501///  This implementation handles:
1502///
1503///   * pointer-to-pointer casts
1504///   * implicit conversions from array references to pointers
1505///   * taking the address of fields
1506///   * arbitrary interplay between "&" and "*" operators
1507///   * pointer arithmetic from an address of a stack variable
1508///   * taking the address of an array element where the array is on the stack
1509static DeclRefExpr* EvalAddr(Expr *E) {
1510  // We should only be called for evaluating pointer expressions.
1511  assert((E->getType()->isAnyPointerType() ||
1512          E->getType()->isBlockPointerType() ||
1513          E->getType()->isObjCQualifiedIdType()) &&
1514         "EvalAddr only works on pointers");
1515
1516  // Our "symbolic interpreter" is just a dispatch off the currently
1517  // viewed AST node.  We then recursively traverse the AST by calling
1518  // EvalAddr and EvalVal appropriately.
1519  switch (E->getStmtClass()) {
1520  case Stmt::ParenExprClass:
1521    // Ignore parentheses.
1522    return EvalAddr(cast<ParenExpr>(E)->getSubExpr());
1523
1524  case Stmt::UnaryOperatorClass: {
1525    // The only unary operator that make sense to handle here
1526    // is AddrOf.  All others don't make sense as pointers.
1527    UnaryOperator *U = cast<UnaryOperator>(E);
1528
1529    if (U->getOpcode() == UnaryOperator::AddrOf)
1530      return EvalVal(U->getSubExpr());
1531    else
1532      return NULL;
1533  }
1534
1535  case Stmt::BinaryOperatorClass: {
1536    // Handle pointer arithmetic.  All other binary operators are not valid
1537    // in this context.
1538    BinaryOperator *B = cast<BinaryOperator>(E);
1539    BinaryOperator::Opcode op = B->getOpcode();
1540
1541    if (op != BinaryOperator::Add && op != BinaryOperator::Sub)
1542      return NULL;
1543
1544    Expr *Base = B->getLHS();
1545
1546    // Determine which argument is the real pointer base.  It could be
1547    // the RHS argument instead of the LHS.
1548    if (!Base->getType()->isPointerType()) Base = B->getRHS();
1549
1550    assert (Base->getType()->isPointerType());
1551    return EvalAddr(Base);
1552  }
1553
1554  // For conditional operators we need to see if either the LHS or RHS are
1555  // valid DeclRefExpr*s.  If one of them is valid, we return it.
1556  case Stmt::ConditionalOperatorClass: {
1557    ConditionalOperator *C = cast<ConditionalOperator>(E);
1558
1559    // Handle the GNU extension for missing LHS.
1560    if (Expr *lhsExpr = C->getLHS())
1561      if (DeclRefExpr* LHS = EvalAddr(lhsExpr))
1562        return LHS;
1563
1564     return EvalAddr(C->getRHS());
1565  }
1566
1567  // For casts, we need to handle conversions from arrays to
1568  // pointer values, and pointer-to-pointer conversions.
1569  case Stmt::ImplicitCastExprClass:
1570  case Stmt::CStyleCastExprClass:
1571  case Stmt::CXXFunctionalCastExprClass: {
1572    Expr* SubExpr = cast<CastExpr>(E)->getSubExpr();
1573    QualType T = SubExpr->getType();
1574
1575    if (SubExpr->getType()->isPointerType() ||
1576        SubExpr->getType()->isBlockPointerType() ||
1577        SubExpr->getType()->isObjCQualifiedIdType())
1578      return EvalAddr(SubExpr);
1579    else if (T->isArrayType())
1580      return EvalVal(SubExpr);
1581    else
1582      return 0;
1583  }
1584
1585  // C++ casts.  For dynamic casts, static casts, and const casts, we
1586  // are always converting from a pointer-to-pointer, so we just blow
1587  // through the cast.  In the case the dynamic cast doesn't fail (and
1588  // return NULL), we take the conservative route and report cases
1589  // where we return the address of a stack variable.  For Reinterpre
1590  // FIXME: The comment about is wrong; we're not always converting
1591  // from pointer to pointer. I'm guessing that this code should also
1592  // handle references to objects.
1593  case Stmt::CXXStaticCastExprClass:
1594  case Stmt::CXXDynamicCastExprClass:
1595  case Stmt::CXXConstCastExprClass:
1596  case Stmt::CXXReinterpretCastExprClass: {
1597      Expr *S = cast<CXXNamedCastExpr>(E)->getSubExpr();
1598      if (S->getType()->isPointerType() || S->getType()->isBlockPointerType())
1599        return EvalAddr(S);
1600      else
1601        return NULL;
1602  }
1603
1604  // Everything else: we simply don't reason about them.
1605  default:
1606    return NULL;
1607  }
1608}
1609
1610
1611///  EvalVal - This function is complements EvalAddr in the mutual recursion.
1612///   See the comments for EvalAddr for more details.
1613static DeclRefExpr* EvalVal(Expr *E) {
1614
1615  // We should only be called for evaluating non-pointer expressions, or
1616  // expressions with a pointer type that are not used as references but instead
1617  // are l-values (e.g., DeclRefExpr with a pointer type).
1618
1619  // Our "symbolic interpreter" is just a dispatch off the currently
1620  // viewed AST node.  We then recursively traverse the AST by calling
1621  // EvalAddr and EvalVal appropriately.
1622  switch (E->getStmtClass()) {
1623  case Stmt::DeclRefExprClass: {
1624    // DeclRefExpr: the base case.  When we hit a DeclRefExpr we are looking
1625    //  at code that refers to a variable's name.  We check if it has local
1626    //  storage within the function, and if so, return the expression.
1627    DeclRefExpr *DR = cast<DeclRefExpr>(E);
1628
1629    if (VarDecl *V = dyn_cast<VarDecl>(DR->getDecl()))
1630      if (V->hasLocalStorage() && !V->getType()->isReferenceType()) return DR;
1631
1632    return NULL;
1633  }
1634
1635  case Stmt::ParenExprClass:
1636    // Ignore parentheses.
1637    return EvalVal(cast<ParenExpr>(E)->getSubExpr());
1638
1639  case Stmt::UnaryOperatorClass: {
1640    // The only unary operator that make sense to handle here
1641    // is Deref.  All others don't resolve to a "name."  This includes
1642    // handling all sorts of rvalues passed to a unary operator.
1643    UnaryOperator *U = cast<UnaryOperator>(E);
1644
1645    if (U->getOpcode() == UnaryOperator::Deref)
1646      return EvalAddr(U->getSubExpr());
1647
1648    return NULL;
1649  }
1650
1651  case Stmt::ArraySubscriptExprClass: {
1652    // Array subscripts are potential references to data on the stack.  We
1653    // retrieve the DeclRefExpr* for the array variable if it indeed
1654    // has local storage.
1655    return EvalAddr(cast<ArraySubscriptExpr>(E)->getBase());
1656  }
1657
1658  case Stmt::ConditionalOperatorClass: {
1659    // For conditional operators we need to see if either the LHS or RHS are
1660    // non-NULL DeclRefExpr's.  If one is non-NULL, we return it.
1661    ConditionalOperator *C = cast<ConditionalOperator>(E);
1662
1663    // Handle the GNU extension for missing LHS.
1664    if (Expr *lhsExpr = C->getLHS())
1665      if (DeclRefExpr *LHS = EvalVal(lhsExpr))
1666        return LHS;
1667
1668    return EvalVal(C->getRHS());
1669  }
1670
1671  // Accesses to members are potential references to data on the stack.
1672  case Stmt::MemberExprClass: {
1673    MemberExpr *M = cast<MemberExpr>(E);
1674
1675    // Check for indirect access.  We only want direct field accesses.
1676    if (!M->isArrow())
1677      return EvalVal(M->getBase());
1678    else
1679      return NULL;
1680  }
1681
1682  // Everything else: we simply don't reason about them.
1683  default:
1684    return NULL;
1685  }
1686}
1687
1688//===--- CHECK: Floating-Point comparisons (-Wfloat-equal) ---------------===//
1689
1690/// Check for comparisons of floating point operands using != and ==.
1691/// Issue a warning if these are no self-comparisons, as they are not likely
1692/// to do what the programmer intended.
1693void Sema::CheckFloatComparison(SourceLocation loc, Expr* lex, Expr *rex) {
1694  bool EmitWarning = true;
1695
1696  Expr* LeftExprSansParen = lex->IgnoreParens();
1697  Expr* RightExprSansParen = rex->IgnoreParens();
1698
1699  // Special case: check for x == x (which is OK).
1700  // Do not emit warnings for such cases.
1701  if (DeclRefExpr* DRL = dyn_cast<DeclRefExpr>(LeftExprSansParen))
1702    if (DeclRefExpr* DRR = dyn_cast<DeclRefExpr>(RightExprSansParen))
1703      if (DRL->getDecl() == DRR->getDecl())
1704        EmitWarning = false;
1705
1706
1707  // Special case: check for comparisons against literals that can be exactly
1708  //  represented by APFloat.  In such cases, do not emit a warning.  This
1709  //  is a heuristic: often comparison against such literals are used to
1710  //  detect if a value in a variable has not changed.  This clearly can
1711  //  lead to false negatives.
1712  if (EmitWarning) {
1713    if (FloatingLiteral* FLL = dyn_cast<FloatingLiteral>(LeftExprSansParen)) {
1714      if (FLL->isExact())
1715        EmitWarning = false;
1716    } else
1717      if (FloatingLiteral* FLR = dyn_cast<FloatingLiteral>(RightExprSansParen)){
1718        if (FLR->isExact())
1719          EmitWarning = false;
1720    }
1721  }
1722
1723  // Check for comparisons with builtin types.
1724  if (EmitWarning)
1725    if (CallExpr* CL = dyn_cast<CallExpr>(LeftExprSansParen))
1726      if (CL->isBuiltinCall(Context))
1727        EmitWarning = false;
1728
1729  if (EmitWarning)
1730    if (CallExpr* CR = dyn_cast<CallExpr>(RightExprSansParen))
1731      if (CR->isBuiltinCall(Context))
1732        EmitWarning = false;
1733
1734  // Emit the diagnostic.
1735  if (EmitWarning)
1736    Diag(loc, diag::warn_floatingpoint_eq)
1737      << lex->getSourceRange() << rex->getSourceRange();
1738}
1739
1740//===--- CHECK: Integer mixed-sign comparisons (-Wsign-compare) --------===//
1741//===--- CHECK: Lossy implicit conversions (-Wconversion) --------------===//
1742
1743namespace {
1744
1745/// Structure recording the 'active' range of an integer-valued
1746/// expression.
1747struct IntRange {
1748  /// The number of bits active in the int.
1749  unsigned Width;
1750
1751  /// True if the int is known not to have negative values.
1752  bool NonNegative;
1753
1754  IntRange() {}
1755  IntRange(unsigned Width, bool NonNegative)
1756    : Width(Width), NonNegative(NonNegative)
1757  {}
1758
1759  // Returns the range of the bool type.
1760  static IntRange forBoolType() {
1761    return IntRange(1, true);
1762  }
1763
1764  // Returns the range of an integral type.
1765  static IntRange forType(ASTContext &C, QualType T) {
1766    return forCanonicalType(C, T->getCanonicalTypeInternal().getTypePtr());
1767  }
1768
1769  // Returns the range of an integeral type based on its canonical
1770  // representation.
1771  static IntRange forCanonicalType(ASTContext &C, const Type *T) {
1772    assert(T->isCanonicalUnqualified());
1773
1774    if (const VectorType *VT = dyn_cast<VectorType>(T))
1775      T = VT->getElementType().getTypePtr();
1776    if (const ComplexType *CT = dyn_cast<ComplexType>(T))
1777      T = CT->getElementType().getTypePtr();
1778    if (const EnumType *ET = dyn_cast<EnumType>(T))
1779      T = ET->getDecl()->getIntegerType().getTypePtr();
1780
1781    const BuiltinType *BT = cast<BuiltinType>(T);
1782    assert(BT->isInteger());
1783
1784    return IntRange(C.getIntWidth(QualType(T, 0)), BT->isUnsignedInteger());
1785  }
1786
1787  // Returns the supremum of two ranges: i.e. their conservative merge.
1788  static IntRange join(const IntRange &L, const IntRange &R) {
1789    return IntRange(std::max(L.Width, R.Width),
1790                    L.NonNegative && R.NonNegative);
1791  }
1792
1793  // Returns the infinum of two ranges: i.e. their aggressive merge.
1794  static IntRange meet(const IntRange &L, const IntRange &R) {
1795    return IntRange(std::min(L.Width, R.Width),
1796                    L.NonNegative || R.NonNegative);
1797  }
1798};
1799
1800IntRange GetValueRange(ASTContext &C, llvm::APSInt &value, unsigned MaxWidth) {
1801  if (value.isSigned() && value.isNegative())
1802    return IntRange(value.getMinSignedBits(), false);
1803
1804  if (value.getBitWidth() > MaxWidth)
1805    value.trunc(MaxWidth);
1806
1807  // isNonNegative() just checks the sign bit without considering
1808  // signedness.
1809  return IntRange(value.getActiveBits(), true);
1810}
1811
1812IntRange GetValueRange(ASTContext &C, APValue &result, QualType Ty,
1813                       unsigned MaxWidth) {
1814  if (result.isInt())
1815    return GetValueRange(C, result.getInt(), MaxWidth);
1816
1817  if (result.isVector()) {
1818    IntRange R = GetValueRange(C, result.getVectorElt(0), Ty, MaxWidth);
1819    for (unsigned i = 1, e = result.getVectorLength(); i != e; ++i) {
1820      IntRange El = GetValueRange(C, result.getVectorElt(i), Ty, MaxWidth);
1821      R = IntRange::join(R, El);
1822    }
1823    return R;
1824  }
1825
1826  if (result.isComplexInt()) {
1827    IntRange R = GetValueRange(C, result.getComplexIntReal(), MaxWidth);
1828    IntRange I = GetValueRange(C, result.getComplexIntImag(), MaxWidth);
1829    return IntRange::join(R, I);
1830  }
1831
1832  // This can happen with lossless casts to intptr_t of "based" lvalues.
1833  // Assume it might use arbitrary bits.
1834  // FIXME: The only reason we need to pass the type in here is to get
1835  // the sign right on this one case.  It would be nice if APValue
1836  // preserved this.
1837  assert(result.isLValue());
1838  return IntRange(MaxWidth, Ty->isUnsignedIntegerType());
1839}
1840
1841/// Pseudo-evaluate the given integer expression, estimating the
1842/// range of values it might take.
1843///
1844/// \param MaxWidth - the width to which the value will be truncated
1845IntRange GetExprRange(ASTContext &C, Expr *E, unsigned MaxWidth) {
1846  E = E->IgnoreParens();
1847
1848  // Try a full evaluation first.
1849  Expr::EvalResult result;
1850  if (E->Evaluate(result, C))
1851    return GetValueRange(C, result.Val, E->getType(), MaxWidth);
1852
1853  // I think we only want to look through implicit casts here; if the
1854  // user has an explicit widening cast, we should treat the value as
1855  // being of the new, wider type.
1856  if (ImplicitCastExpr *CE = dyn_cast<ImplicitCastExpr>(E)) {
1857    if (CE->getCastKind() == CastExpr::CK_NoOp)
1858      return GetExprRange(C, CE->getSubExpr(), MaxWidth);
1859
1860    IntRange OutputTypeRange = IntRange::forType(C, CE->getType());
1861
1862    bool isIntegerCast = (CE->getCastKind() == CastExpr::CK_IntegralCast);
1863    if (!isIntegerCast && CE->getCastKind() == CastExpr::CK_Unknown)
1864      isIntegerCast = CE->getSubExpr()->getType()->isIntegerType();
1865
1866    // Assume that non-integer casts can span the full range of the type.
1867    if (!isIntegerCast)
1868      return OutputTypeRange;
1869
1870    IntRange SubRange
1871      = GetExprRange(C, CE->getSubExpr(),
1872                     std::min(MaxWidth, OutputTypeRange.Width));
1873
1874    // Bail out if the subexpr's range is as wide as the cast type.
1875    if (SubRange.Width >= OutputTypeRange.Width)
1876      return OutputTypeRange;
1877
1878    // Otherwise, we take the smaller width, and we're non-negative if
1879    // either the output type or the subexpr is.
1880    return IntRange(SubRange.Width,
1881                    SubRange.NonNegative || OutputTypeRange.NonNegative);
1882  }
1883
1884  if (ConditionalOperator *CO = dyn_cast<ConditionalOperator>(E)) {
1885    // If we can fold the condition, just take that operand.
1886    bool CondResult;
1887    if (CO->getCond()->EvaluateAsBooleanCondition(CondResult, C))
1888      return GetExprRange(C, CondResult ? CO->getTrueExpr()
1889                                        : CO->getFalseExpr(),
1890                          MaxWidth);
1891
1892    // Otherwise, conservatively merge.
1893    IntRange L = GetExprRange(C, CO->getTrueExpr(), MaxWidth);
1894    IntRange R = GetExprRange(C, CO->getFalseExpr(), MaxWidth);
1895    return IntRange::join(L, R);
1896  }
1897
1898  if (BinaryOperator *BO = dyn_cast<BinaryOperator>(E)) {
1899    switch (BO->getOpcode()) {
1900
1901    // Boolean-valued operations are single-bit and positive.
1902    case BinaryOperator::LAnd:
1903    case BinaryOperator::LOr:
1904    case BinaryOperator::LT:
1905    case BinaryOperator::GT:
1906    case BinaryOperator::LE:
1907    case BinaryOperator::GE:
1908    case BinaryOperator::EQ:
1909    case BinaryOperator::NE:
1910      return IntRange::forBoolType();
1911
1912    // Operations with opaque sources are black-listed.
1913    case BinaryOperator::PtrMemD:
1914    case BinaryOperator::PtrMemI:
1915      return IntRange::forType(C, E->getType());
1916
1917    // Bitwise-and uses the *infinum* of the two source ranges.
1918    case BinaryOperator::And:
1919      return IntRange::meet(GetExprRange(C, BO->getLHS(), MaxWidth),
1920                            GetExprRange(C, BO->getRHS(), MaxWidth));
1921
1922    // Left shift gets black-listed based on a judgement call.
1923    case BinaryOperator::Shl:
1924      return IntRange::forType(C, E->getType());
1925
1926    // Right shift by a constant can narrow its left argument.
1927    case BinaryOperator::Shr: {
1928      IntRange L = GetExprRange(C, BO->getLHS(), MaxWidth);
1929
1930      // If the shift amount is a positive constant, drop the width by
1931      // that much.
1932      llvm::APSInt shift;
1933      if (BO->getRHS()->isIntegerConstantExpr(shift, C) &&
1934          shift.isNonNegative()) {
1935        unsigned zext = shift.getZExtValue();
1936        if (zext >= L.Width)
1937          L.Width = (L.NonNegative ? 0 : 1);
1938        else
1939          L.Width -= zext;
1940      }
1941
1942      return L;
1943    }
1944
1945    // Comma acts as its right operand.
1946    case BinaryOperator::Comma:
1947      return GetExprRange(C, BO->getRHS(), MaxWidth);
1948
1949    // Black-list pointer subtractions.
1950    case BinaryOperator::Sub:
1951      if (BO->getLHS()->getType()->isPointerType())
1952        return IntRange::forType(C, E->getType());
1953      // fallthrough
1954
1955    default:
1956      break;
1957    }
1958
1959    // Treat every other operator as if it were closed on the
1960    // narrowest type that encompasses both operands.
1961    IntRange L = GetExprRange(C, BO->getLHS(), MaxWidth);
1962    IntRange R = GetExprRange(C, BO->getRHS(), MaxWidth);
1963    return IntRange::join(L, R);
1964  }
1965
1966  if (UnaryOperator *UO = dyn_cast<UnaryOperator>(E)) {
1967    switch (UO->getOpcode()) {
1968    // Boolean-valued operations are white-listed.
1969    case UnaryOperator::LNot:
1970      return IntRange::forBoolType();
1971
1972    // Operations with opaque sources are black-listed.
1973    case UnaryOperator::Deref:
1974    case UnaryOperator::AddrOf: // should be impossible
1975    case UnaryOperator::OffsetOf:
1976      return IntRange::forType(C, E->getType());
1977
1978    default:
1979      return GetExprRange(C, UO->getSubExpr(), MaxWidth);
1980    }
1981  }
1982
1983  FieldDecl *BitField = E->getBitField();
1984  if (BitField) {
1985    llvm::APSInt BitWidthAP = BitField->getBitWidth()->EvaluateAsInt(C);
1986    unsigned BitWidth = BitWidthAP.getZExtValue();
1987
1988    return IntRange(BitWidth, BitField->getType()->isUnsignedIntegerType());
1989  }
1990
1991  return IntRange::forType(C, E->getType());
1992}
1993
1994/// Checks whether the given value, which currently has the given
1995/// source semantics, has the same value when coerced through the
1996/// target semantics.
1997bool IsSameFloatAfterCast(const llvm::APFloat &value,
1998                          const llvm::fltSemantics &Src,
1999                          const llvm::fltSemantics &Tgt) {
2000  llvm::APFloat truncated = value;
2001
2002  bool ignored;
2003  truncated.convert(Src, llvm::APFloat::rmNearestTiesToEven, &ignored);
2004  truncated.convert(Tgt, llvm::APFloat::rmNearestTiesToEven, &ignored);
2005
2006  return truncated.bitwiseIsEqual(value);
2007}
2008
2009/// Checks whether the given value, which currently has the given
2010/// source semantics, has the same value when coerced through the
2011/// target semantics.
2012///
2013/// The value might be a vector of floats (or a complex number).
2014bool IsSameFloatAfterCast(const APValue &value,
2015                          const llvm::fltSemantics &Src,
2016                          const llvm::fltSemantics &Tgt) {
2017  if (value.isFloat())
2018    return IsSameFloatAfterCast(value.getFloat(), Src, Tgt);
2019
2020  if (value.isVector()) {
2021    for (unsigned i = 0, e = value.getVectorLength(); i != e; ++i)
2022      if (!IsSameFloatAfterCast(value.getVectorElt(i), Src, Tgt))
2023        return false;
2024    return true;
2025  }
2026
2027  assert(value.isComplexFloat());
2028  return (IsSameFloatAfterCast(value.getComplexFloatReal(), Src, Tgt) &&
2029          IsSameFloatAfterCast(value.getComplexFloatImag(), Src, Tgt));
2030}
2031
2032} // end anonymous namespace
2033
2034/// \brief Implements -Wsign-compare.
2035///
2036/// \param lex the left-hand expression
2037/// \param rex the right-hand expression
2038/// \param OpLoc the location of the joining operator
2039/// \param Equality whether this is an "equality-like" join, which
2040///   suppresses the warning in some cases
2041void Sema::CheckSignCompare(Expr *lex, Expr *rex, SourceLocation OpLoc,
2042                            const PartialDiagnostic &PD, bool Equality) {
2043  // Don't warn if we're in an unevaluated context.
2044  if (ExprEvalContexts.back().Context == Unevaluated)
2045    return;
2046
2047  // If either expression is value-dependent, don't warn. We'll get another
2048  // chance at instantiation time.
2049  if (lex->isValueDependent() || rex->isValueDependent())
2050    return;
2051
2052  QualType lt = lex->getType(), rt = rex->getType();
2053
2054  // Only warn if both operands are integral.
2055  if (!lt->isIntegerType() || !rt->isIntegerType())
2056    return;
2057
2058  // In C, the width of a bitfield determines its type, and the
2059  // declared type only contributes the signedness.  This duplicates
2060  // the work that will later be done by UsualUnaryConversions.
2061  // Eventually, this check will be reorganized in a way that avoids
2062  // this duplication.
2063  if (!getLangOptions().CPlusPlus) {
2064    QualType tmp;
2065    tmp = Context.isPromotableBitField(lex);
2066    if (!tmp.isNull()) lt = tmp;
2067    tmp = Context.isPromotableBitField(rex);
2068    if (!tmp.isNull()) rt = tmp;
2069  }
2070
2071  // The rule is that the signed operand becomes unsigned, so isolate the
2072  // signed operand.
2073  Expr *signedOperand = lex, *unsignedOperand = rex;
2074  QualType signedType = lt, unsignedType = rt;
2075  if (lt->isSignedIntegerType()) {
2076    if (rt->isSignedIntegerType()) return;
2077  } else {
2078    if (!rt->isSignedIntegerType()) return;
2079    std::swap(signedOperand, unsignedOperand);
2080    std::swap(signedType, unsignedType);
2081  }
2082
2083  unsigned unsignedWidth = Context.getIntWidth(unsignedType);
2084  unsigned signedWidth = Context.getIntWidth(signedType);
2085
2086  // If the unsigned type is strictly smaller than the signed type,
2087  // then (1) the result type will be signed and (2) the unsigned
2088  // value will fit fully within the signed type, and thus the result
2089  // of the comparison will be exact.
2090  if (signedWidth > unsignedWidth)
2091    return;
2092
2093  // Otherwise, calculate the effective ranges.
2094  IntRange signedRange = GetExprRange(Context, signedOperand, signedWidth);
2095  IntRange unsignedRange = GetExprRange(Context, unsignedOperand, unsignedWidth);
2096
2097  // We should never be unable to prove that the unsigned operand is
2098  // non-negative.
2099  assert(unsignedRange.NonNegative && "unsigned range includes negative?");
2100
2101  // If the signed operand is non-negative, then the signed->unsigned
2102  // conversion won't change it.
2103  if (signedRange.NonNegative)
2104    return;
2105
2106  // For (in)equality comparisons, if the unsigned operand is a
2107  // constant which cannot collide with a overflowed signed operand,
2108  // then reinterpreting the signed operand as unsigned will not
2109  // change the result of the comparison.
2110  if (Equality && unsignedRange.Width < unsignedWidth)
2111    return;
2112
2113  Diag(OpLoc, PD)
2114    << lt << rt << lex->getSourceRange() << rex->getSourceRange();
2115}
2116
2117/// Diagnose an implicit cast;  purely a helper for CheckImplicitConversion.
2118static void DiagnoseImpCast(Sema &S, Expr *E, QualType T, unsigned diag) {
2119  S.Diag(E->getExprLoc(), diag) << E->getType() << T << E->getSourceRange();
2120}
2121
2122/// Implements -Wconversion.
2123void Sema::CheckImplicitConversion(Expr *E, QualType T) {
2124  // Don't diagnose in unevaluated contexts.
2125  if (ExprEvalContexts.back().Context == Sema::Unevaluated)
2126    return;
2127
2128  // Don't diagnose for value-dependent expressions.
2129  if (E->isValueDependent())
2130    return;
2131
2132  const Type *Source = Context.getCanonicalType(E->getType()).getTypePtr();
2133  const Type *Target = Context.getCanonicalType(T).getTypePtr();
2134
2135  // Never diagnose implicit casts to bool.
2136  if (Target->isSpecificBuiltinType(BuiltinType::Bool))
2137    return;
2138
2139  // Strip vector types.
2140  if (isa<VectorType>(Source)) {
2141    if (!isa<VectorType>(Target))
2142      return DiagnoseImpCast(*this, E, T, diag::warn_impcast_vector_scalar);
2143
2144    Source = cast<VectorType>(Source)->getElementType().getTypePtr();
2145    Target = cast<VectorType>(Target)->getElementType().getTypePtr();
2146  }
2147
2148  // Strip complex types.
2149  if (isa<ComplexType>(Source)) {
2150    if (!isa<ComplexType>(Target))
2151      return DiagnoseImpCast(*this, E, T, diag::warn_impcast_complex_scalar);
2152
2153    Source = cast<ComplexType>(Source)->getElementType().getTypePtr();
2154    Target = cast<ComplexType>(Target)->getElementType().getTypePtr();
2155  }
2156
2157  const BuiltinType *SourceBT = dyn_cast<BuiltinType>(Source);
2158  const BuiltinType *TargetBT = dyn_cast<BuiltinType>(Target);
2159
2160  // If the source is floating point...
2161  if (SourceBT && SourceBT->isFloatingPoint()) {
2162    // ...and the target is floating point...
2163    if (TargetBT && TargetBT->isFloatingPoint()) {
2164      // ...then warn if we're dropping FP rank.
2165
2166      // Builtin FP kinds are ordered by increasing FP rank.
2167      if (SourceBT->getKind() > TargetBT->getKind()) {
2168        // Don't warn about float constants that are precisely
2169        // representable in the target type.
2170        Expr::EvalResult result;
2171        if (E->Evaluate(result, Context)) {
2172          // Value might be a float, a float vector, or a float complex.
2173          if (IsSameFloatAfterCast(result.Val,
2174                     Context.getFloatTypeSemantics(QualType(TargetBT, 0)),
2175                     Context.getFloatTypeSemantics(QualType(SourceBT, 0))))
2176            return;
2177        }
2178
2179        DiagnoseImpCast(*this, E, T, diag::warn_impcast_float_precision);
2180      }
2181      return;
2182    }
2183
2184    // If the target is integral, always warn.
2185    if ((TargetBT && TargetBT->isInteger()))
2186      // TODO: don't warn for integer values?
2187      return DiagnoseImpCast(*this, E, T, diag::warn_impcast_float_integer);
2188
2189    return;
2190  }
2191
2192  if (!Source->isIntegerType() || !Target->isIntegerType())
2193    return;
2194
2195  IntRange SourceRange = GetExprRange(Context, E, Context.getIntWidth(E->getType()));
2196  IntRange TargetRange = IntRange::forCanonicalType(Context, Target);
2197
2198  // FIXME: also signed<->unsigned?
2199
2200  if (SourceRange.Width > TargetRange.Width) {
2201    // People want to build with -Wshorten-64-to-32 and not -Wconversion
2202    // and by god we'll let them.
2203    if (SourceRange.Width == 64 && TargetRange.Width == 32)
2204      return DiagnoseImpCast(*this, E, T, diag::warn_impcast_integer_64_32);
2205    return DiagnoseImpCast(*this, E, T, diag::warn_impcast_integer_precision);
2206  }
2207
2208  return;
2209}
2210
2211// MarkLive - Mark all the blocks reachable from e as live.  Returns the total
2212// number of blocks just marked live.
2213static unsigned MarkLive(CFGBlock *e, llvm::BitVector &live) {
2214  unsigned count = 0;
2215  std::queue<CFGBlock*> workq;
2216  // Prep work queue
2217  live.set(e->getBlockID());
2218  ++count;
2219  workq.push(e);
2220  // Solve
2221  while (!workq.empty()) {
2222    CFGBlock *item = workq.front();
2223    workq.pop();
2224    for (CFGBlock::succ_iterator I=item->succ_begin(),
2225           E=item->succ_end();
2226         I != E;
2227         ++I) {
2228      if ((*I) && !live[(*I)->getBlockID()]) {
2229        live.set((*I)->getBlockID());
2230        ++count;
2231        workq.push(*I);
2232      }
2233    }
2234  }
2235  return count;
2236}
2237
2238static SourceLocation GetUnreachableLoc(CFGBlock &b, SourceRange &R1,
2239                                        SourceRange &R2) {
2240  Stmt *S;
2241  unsigned sn = 0;
2242  R1 = R2 = SourceRange();
2243
2244  top:
2245  if (sn < b.size())
2246    S = b[sn].getStmt();
2247  else if (b.getTerminator())
2248    S = b.getTerminator();
2249  else
2250    return SourceLocation();
2251
2252  switch (S->getStmtClass()) {
2253  case Expr::BinaryOperatorClass: {
2254    BinaryOperator *BO = cast<BinaryOperator>(S);
2255    if (BO->getOpcode() == BinaryOperator::Comma) {
2256      if (sn+1 < b.size())
2257        return b[sn+1].getStmt()->getLocStart();
2258      CFGBlock *n = &b;
2259      while (1) {
2260        if (n->getTerminator())
2261          return n->getTerminator()->getLocStart();
2262        if (n->succ_size() != 1)
2263          return SourceLocation();
2264        n = n[0].succ_begin()[0];
2265        if (n->pred_size() != 1)
2266          return SourceLocation();
2267        if (!n->empty())
2268          return n[0][0].getStmt()->getLocStart();
2269      }
2270    }
2271    R1 = BO->getLHS()->getSourceRange();
2272    R2 = BO->getRHS()->getSourceRange();
2273    return BO->getOperatorLoc();
2274  }
2275  case Expr::UnaryOperatorClass: {
2276    const UnaryOperator *UO = cast<UnaryOperator>(S);
2277    R1 = UO->getSubExpr()->getSourceRange();
2278    return UO->getOperatorLoc();
2279  }
2280  case Expr::CompoundAssignOperatorClass: {
2281    const CompoundAssignOperator *CAO = cast<CompoundAssignOperator>(S);
2282    R1 = CAO->getLHS()->getSourceRange();
2283    R2 = CAO->getRHS()->getSourceRange();
2284    return CAO->getOperatorLoc();
2285  }
2286  case Expr::ConditionalOperatorClass: {
2287    const ConditionalOperator *CO = cast<ConditionalOperator>(S);
2288    return CO->getQuestionLoc();
2289  }
2290  case Expr::MemberExprClass: {
2291    const MemberExpr *ME = cast<MemberExpr>(S);
2292    R1 = ME->getSourceRange();
2293    return ME->getMemberLoc();
2294  }
2295  case Expr::ArraySubscriptExprClass: {
2296    const ArraySubscriptExpr *ASE = cast<ArraySubscriptExpr>(S);
2297    R1 = ASE->getLHS()->getSourceRange();
2298    R2 = ASE->getRHS()->getSourceRange();
2299    return ASE->getRBracketLoc();
2300  }
2301  case Expr::CStyleCastExprClass: {
2302    const CStyleCastExpr *CSC = cast<CStyleCastExpr>(S);
2303    R1 = CSC->getSubExpr()->getSourceRange();
2304    return CSC->getLParenLoc();
2305  }
2306  case Expr::CXXFunctionalCastExprClass: {
2307    const CXXFunctionalCastExpr *CE = cast <CXXFunctionalCastExpr>(S);
2308    R1 = CE->getSubExpr()->getSourceRange();
2309    return CE->getTypeBeginLoc();
2310  }
2311  case Expr::ImplicitCastExprClass:
2312    ++sn;
2313    goto top;
2314  case Stmt::CXXTryStmtClass: {
2315    return cast<CXXTryStmt>(S)->getHandler(0)->getCatchLoc();
2316  }
2317  default: ;
2318  }
2319  R1 = S->getSourceRange();
2320  return S->getLocStart();
2321}
2322
2323static SourceLocation MarkLiveTop(CFGBlock *e, llvm::BitVector &live,
2324                               SourceManager &SM) {
2325  std::queue<CFGBlock*> workq;
2326  // Prep work queue
2327  workq.push(e);
2328  SourceRange R1, R2;
2329  SourceLocation top = GetUnreachableLoc(*e, R1, R2);
2330  bool FromMainFile = false;
2331  bool FromSystemHeader = false;
2332  bool TopValid = false;
2333  if (top.isValid()) {
2334    FromMainFile = SM.isFromMainFile(top);
2335    FromSystemHeader = SM.isInSystemHeader(top);
2336    TopValid = true;
2337  }
2338  // Solve
2339  while (!workq.empty()) {
2340    CFGBlock *item = workq.front();
2341    workq.pop();
2342    SourceLocation c = GetUnreachableLoc(*item, R1, R2);
2343    if (c.isValid()
2344        && (!TopValid
2345            || (SM.isFromMainFile(c) && !FromMainFile)
2346            || (FromSystemHeader && !SM.isInSystemHeader(c))
2347            || SM.isBeforeInTranslationUnit(c, top))) {
2348      top = c;
2349      FromMainFile = SM.isFromMainFile(top);
2350      FromSystemHeader = SM.isInSystemHeader(top);
2351    }
2352    live.set(item->getBlockID());
2353    for (CFGBlock::succ_iterator I=item->succ_begin(),
2354           E=item->succ_end();
2355         I != E;
2356         ++I) {
2357      if ((*I) && !live[(*I)->getBlockID()]) {
2358        live.set((*I)->getBlockID());
2359        workq.push(*I);
2360      }
2361    }
2362  }
2363  return top;
2364}
2365
2366static int LineCmp(const void *p1, const void *p2) {
2367  SourceLocation *Line1 = (SourceLocation *)p1;
2368  SourceLocation *Line2 = (SourceLocation *)p2;
2369  return !(*Line1 < *Line2);
2370}
2371
2372namespace {
2373  struct ErrLoc {
2374    SourceLocation Loc;
2375    SourceRange R1;
2376    SourceRange R2;
2377    ErrLoc(SourceLocation l, SourceRange r1, SourceRange r2)
2378      : Loc(l), R1(r1), R2(r2) { }
2379  };
2380}
2381
2382/// CheckUnreachable - Check for unreachable code.
2383void Sema::CheckUnreachable(AnalysisContext &AC) {
2384  unsigned count;
2385  // We avoid checking when there are errors, as the CFG won't faithfully match
2386  // the user's code.
2387  if (getDiagnostics().hasErrorOccurred())
2388    return;
2389  if (Diags.getDiagnosticLevel(diag::warn_unreachable) == Diagnostic::Ignored)
2390    return;
2391
2392  CFG *cfg = AC.getCFG();
2393  if (cfg == 0)
2394    return;
2395
2396  llvm::BitVector live(cfg->getNumBlockIDs());
2397  // Mark all live things first.
2398  count = MarkLive(&cfg->getEntry(), live);
2399
2400  if (count == cfg->getNumBlockIDs())
2401    // If there are no dead blocks, we're done.
2402    return;
2403
2404  SourceRange R1, R2;
2405
2406  llvm::SmallVector<ErrLoc, 24> lines;
2407  bool AddEHEdges = AC.getAddEHEdges();
2408  // First, give warnings for blocks with no predecessors, as they
2409  // can't be part of a loop.
2410  for (CFG::iterator I = cfg->begin(), E = cfg->end(); I != E; ++I) {
2411    CFGBlock &b = **I;
2412    if (!live[b.getBlockID()]) {
2413      if (b.pred_begin() == b.pred_end()) {
2414        if (!AddEHEdges && b.getTerminator()
2415            && isa<CXXTryStmt>(b.getTerminator())) {
2416          // When not adding EH edges from calls, catch clauses
2417          // can otherwise seem dead.  Avoid noting them as dead.
2418          count += MarkLive(&b, live);
2419          continue;
2420        }
2421        SourceLocation c = GetUnreachableLoc(b, R1, R2);
2422        if (!c.isValid()) {
2423          // Blocks without a location can't produce a warning, so don't mark
2424          // reachable blocks from here as live.
2425          live.set(b.getBlockID());
2426          ++count;
2427          continue;
2428        }
2429        lines.push_back(ErrLoc(c, R1, R2));
2430        // Avoid excessive errors by marking everything reachable from here
2431        count += MarkLive(&b, live);
2432      }
2433    }
2434  }
2435
2436  if (count < cfg->getNumBlockIDs()) {
2437    // And then give warnings for the tops of loops.
2438    for (CFG::iterator I = cfg->begin(), E = cfg->end(); I != E; ++I) {
2439      CFGBlock &b = **I;
2440      if (!live[b.getBlockID()])
2441        // Avoid excessive errors by marking everything reachable from here
2442        lines.push_back(ErrLoc(MarkLiveTop(&b, live,
2443                                           Context.getSourceManager()),
2444                               SourceRange(), SourceRange()));
2445    }
2446  }
2447
2448  llvm::array_pod_sort(lines.begin(), lines.end(), LineCmp);
2449  for (llvm::SmallVector<ErrLoc, 24>::iterator I = lines.begin(),
2450         E = lines.end();
2451       I != E;
2452       ++I)
2453    if (I->Loc.isValid())
2454      Diag(I->Loc, diag::warn_unreachable) << I->R1 << I->R2;
2455}
2456
2457/// CheckFallThrough - Check that we don't fall off the end of a
2458/// Statement that should return a value.
2459///
2460/// \returns AlwaysFallThrough iff we always fall off the end of the statement,
2461/// MaybeFallThrough iff we might or might not fall off the end,
2462/// NeverFallThroughOrReturn iff we never fall off the end of the statement or
2463/// return.  We assume NeverFallThrough iff we never fall off the end of the
2464/// statement but we may return.  We assume that functions not marked noreturn
2465/// will return.
2466Sema::ControlFlowKind Sema::CheckFallThrough(AnalysisContext &AC) {
2467  CFG *cfg = AC.getCFG();
2468  if (cfg == 0)
2469    // FIXME: This should be NeverFallThrough
2470    return NeverFallThroughOrReturn;
2471
2472  // The CFG leaves in dead things, and we don't want the dead code paths to
2473  // confuse us, so we mark all live things first.
2474  std::queue<CFGBlock*> workq;
2475  llvm::BitVector live(cfg->getNumBlockIDs());
2476  unsigned count = MarkLive(&cfg->getEntry(), live);
2477
2478  bool AddEHEdges = AC.getAddEHEdges();
2479  if (!AddEHEdges && count != cfg->getNumBlockIDs())
2480    // When there are things remaining dead, and we didn't add EH edges
2481    // from CallExprs to the catch clauses, we have to go back and
2482    // mark them as live.
2483    for (CFG::iterator I = cfg->begin(), E = cfg->end(); I != E; ++I) {
2484      CFGBlock &b = **I;
2485      if (!live[b.getBlockID()]) {
2486        if (b.pred_begin() == b.pred_end()) {
2487          if (b.getTerminator() && isa<CXXTryStmt>(b.getTerminator()))
2488            // When not adding EH edges from calls, catch clauses
2489            // can otherwise seem dead.  Avoid noting them as dead.
2490            count += MarkLive(&b, live);
2491          continue;
2492        }
2493      }
2494    }
2495
2496  // Now we know what is live, we check the live precessors of the exit block
2497  // and look for fall through paths, being careful to ignore normal returns,
2498  // and exceptional paths.
2499  bool HasLiveReturn = false;
2500  bool HasFakeEdge = false;
2501  bool HasPlainEdge = false;
2502  bool HasAbnormalEdge = false;
2503  for (CFGBlock::pred_iterator I=cfg->getExit().pred_begin(),
2504         E = cfg->getExit().pred_end();
2505       I != E;
2506       ++I) {
2507    CFGBlock& B = **I;
2508    if (!live[B.getBlockID()])
2509      continue;
2510    if (B.size() == 0) {
2511      if (B.getTerminator() && isa<CXXTryStmt>(B.getTerminator())) {
2512        HasAbnormalEdge = true;
2513        continue;
2514      }
2515
2516      // A labeled empty statement, or the entry block...
2517      HasPlainEdge = true;
2518      continue;
2519    }
2520    Stmt *S = B[B.size()-1];
2521    if (isa<ReturnStmt>(S)) {
2522      HasLiveReturn = true;
2523      continue;
2524    }
2525    if (isa<ObjCAtThrowStmt>(S)) {
2526      HasFakeEdge = true;
2527      continue;
2528    }
2529    if (isa<CXXThrowExpr>(S)) {
2530      HasFakeEdge = true;
2531      continue;
2532    }
2533    if (const AsmStmt *AS = dyn_cast<AsmStmt>(S)) {
2534      if (AS->isMSAsm()) {
2535        HasFakeEdge = true;
2536        HasLiveReturn = true;
2537        continue;
2538      }
2539    }
2540    if (isa<CXXTryStmt>(S)) {
2541      HasAbnormalEdge = true;
2542      continue;
2543    }
2544
2545    bool NoReturnEdge = false;
2546    if (CallExpr *C = dyn_cast<CallExpr>(S)) {
2547      if (B.succ_begin()[0] != &cfg->getExit()) {
2548        HasAbnormalEdge = true;
2549        continue;
2550      }
2551      Expr *CEE = C->getCallee()->IgnoreParenCasts();
2552      if (CEE->getType().getNoReturnAttr()) {
2553        NoReturnEdge = true;
2554        HasFakeEdge = true;
2555      } else if (DeclRefExpr *DRE = dyn_cast<DeclRefExpr>(CEE)) {
2556        ValueDecl *VD = DRE->getDecl();
2557        if (VD->hasAttr<NoReturnAttr>()) {
2558          NoReturnEdge = true;
2559          HasFakeEdge = true;
2560        }
2561      }
2562    }
2563    // FIXME: Add noreturn message sends.
2564    if (NoReturnEdge == false)
2565      HasPlainEdge = true;
2566  }
2567  if (!HasPlainEdge) {
2568    if (HasLiveReturn)
2569      return NeverFallThrough;
2570    return NeverFallThroughOrReturn;
2571  }
2572  if (HasAbnormalEdge || HasFakeEdge || HasLiveReturn)
2573    return MaybeFallThrough;
2574  // This says AlwaysFallThrough for calls to functions that are not marked
2575  // noreturn, that don't return.  If people would like this warning to be more
2576  // accurate, such functions should be marked as noreturn.
2577  return AlwaysFallThrough;
2578}
2579
2580/// CheckFallThroughForFunctionDef - Check that we don't fall off the end of a
2581/// function that should return a value.  Check that we don't fall off the end
2582/// of a noreturn function.  We assume that functions and blocks not marked
2583/// noreturn will return.
2584void Sema::CheckFallThroughForFunctionDef(Decl *D, Stmt *Body,
2585                                          AnalysisContext &AC) {
2586  // FIXME: Would be nice if we had a better way to control cascading errors,
2587  // but for now, avoid them.  The problem is that when Parse sees:
2588  //   int foo() { return a; }
2589  // The return is eaten and the Sema code sees just:
2590  //   int foo() { }
2591  // which this code would then warn about.
2592  if (getDiagnostics().hasErrorOccurred())
2593    return;
2594
2595  bool ReturnsVoid = false;
2596  bool HasNoReturn = false;
2597  if (FunctionDecl *FD = dyn_cast<FunctionDecl>(D)) {
2598    // If the result type of the function is a dependent type, we don't know
2599    // whether it will be void or not, so don't
2600    if (FD->getResultType()->isDependentType())
2601      return;
2602    if (FD->getResultType()->isVoidType())
2603      ReturnsVoid = true;
2604    if (FD->hasAttr<NoReturnAttr>())
2605      HasNoReturn = true;
2606  } else if (ObjCMethodDecl *MD = dyn_cast<ObjCMethodDecl>(D)) {
2607    if (MD->getResultType()->isVoidType())
2608      ReturnsVoid = true;
2609    if (MD->hasAttr<NoReturnAttr>())
2610      HasNoReturn = true;
2611  }
2612
2613  // Short circuit for compilation speed.
2614  if ((Diags.getDiagnosticLevel(diag::warn_maybe_falloff_nonvoid_function)
2615       == Diagnostic::Ignored || ReturnsVoid)
2616      && (Diags.getDiagnosticLevel(diag::warn_noreturn_function_has_return_expr)
2617          == Diagnostic::Ignored || !HasNoReturn)
2618      && (Diags.getDiagnosticLevel(diag::warn_suggest_noreturn_block)
2619          == Diagnostic::Ignored || !ReturnsVoid))
2620    return;
2621  // FIXME: Function try block
2622  if (CompoundStmt *Compound = dyn_cast<CompoundStmt>(Body)) {
2623    switch (CheckFallThrough(AC)) {
2624    case MaybeFallThrough:
2625      if (HasNoReturn)
2626        Diag(Compound->getRBracLoc(), diag::warn_falloff_noreturn_function);
2627      else if (!ReturnsVoid)
2628        Diag(Compound->getRBracLoc(),diag::warn_maybe_falloff_nonvoid_function);
2629      break;
2630    case AlwaysFallThrough:
2631      if (HasNoReturn)
2632        Diag(Compound->getRBracLoc(), diag::warn_falloff_noreturn_function);
2633      else if (!ReturnsVoid)
2634        Diag(Compound->getRBracLoc(), diag::warn_falloff_nonvoid_function);
2635      break;
2636    case NeverFallThroughOrReturn:
2637      if (ReturnsVoid && !HasNoReturn)
2638        Diag(Compound->getLBracLoc(), diag::warn_suggest_noreturn_function);
2639      break;
2640    case NeverFallThrough:
2641      break;
2642    }
2643  }
2644}
2645
2646/// CheckFallThroughForBlock - Check that we don't fall off the end of a block
2647/// that should return a value.  Check that we don't fall off the end of a
2648/// noreturn block.  We assume that functions and blocks not marked noreturn
2649/// will return.
2650void Sema::CheckFallThroughForBlock(QualType BlockTy, Stmt *Body,
2651                                    AnalysisContext &AC) {
2652  // FIXME: Would be nice if we had a better way to control cascading errors,
2653  // but for now, avoid them.  The problem is that when Parse sees:
2654  //   int foo() { return a; }
2655  // The return is eaten and the Sema code sees just:
2656  //   int foo() { }
2657  // which this code would then warn about.
2658  if (getDiagnostics().hasErrorOccurred())
2659    return;
2660  bool ReturnsVoid = false;
2661  bool HasNoReturn = false;
2662  if (const FunctionType *FT =BlockTy->getPointeeType()->getAs<FunctionType>()){
2663    if (FT->getResultType()->isVoidType())
2664      ReturnsVoid = true;
2665    if (FT->getNoReturnAttr())
2666      HasNoReturn = true;
2667  }
2668
2669  // Short circuit for compilation speed.
2670  if (ReturnsVoid
2671      && !HasNoReturn
2672      && (Diags.getDiagnosticLevel(diag::warn_suggest_noreturn_block)
2673          == Diagnostic::Ignored || !ReturnsVoid))
2674    return;
2675  // FIXME: Funtion try block
2676  if (CompoundStmt *Compound = dyn_cast<CompoundStmt>(Body)) {
2677    switch (CheckFallThrough(AC)) {
2678    case MaybeFallThrough:
2679      if (HasNoReturn)
2680        Diag(Compound->getRBracLoc(), diag::err_noreturn_block_has_return_expr);
2681      else if (!ReturnsVoid)
2682        Diag(Compound->getRBracLoc(), diag::err_maybe_falloff_nonvoid_block);
2683      break;
2684    case AlwaysFallThrough:
2685      if (HasNoReturn)
2686        Diag(Compound->getRBracLoc(), diag::err_noreturn_block_has_return_expr);
2687      else if (!ReturnsVoid)
2688        Diag(Compound->getRBracLoc(), diag::err_falloff_nonvoid_block);
2689      break;
2690    case NeverFallThroughOrReturn:
2691      if (ReturnsVoid)
2692        Diag(Compound->getLBracLoc(), diag::warn_suggest_noreturn_block);
2693      break;
2694    case NeverFallThrough:
2695      break;
2696    }
2697  }
2698}
2699
2700/// CheckParmsForFunctionDef - Check that the parameters of the given
2701/// function are appropriate for the definition of a function. This
2702/// takes care of any checks that cannot be performed on the
2703/// declaration itself, e.g., that the types of each of the function
2704/// parameters are complete.
2705bool Sema::CheckParmsForFunctionDef(FunctionDecl *FD) {
2706  bool HasInvalidParm = false;
2707  for (unsigned p = 0, NumParams = FD->getNumParams(); p < NumParams; ++p) {
2708    ParmVarDecl *Param = FD->getParamDecl(p);
2709
2710    // C99 6.7.5.3p4: the parameters in a parameter type list in a
2711    // function declarator that is part of a function definition of
2712    // that function shall not have incomplete type.
2713    //
2714    // This is also C++ [dcl.fct]p6.
2715    if (!Param->isInvalidDecl() &&
2716        RequireCompleteType(Param->getLocation(), Param->getType(),
2717                               diag::err_typecheck_decl_incomplete_type)) {
2718      Param->setInvalidDecl();
2719      HasInvalidParm = true;
2720    }
2721
2722    // C99 6.9.1p5: If the declarator includes a parameter type list, the
2723    // declaration of each parameter shall include an identifier.
2724    if (Param->getIdentifier() == 0 &&
2725        !Param->isImplicit() &&
2726        !getLangOptions().CPlusPlus)
2727      Diag(Param->getLocation(), diag::err_parameter_name_omitted);
2728  }
2729
2730  return HasInvalidParm;
2731}
2732