SemaChecking.cpp revision 4e4b30ec62d78b24e6556fea2624855c193d0e3e
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_fpclassify:
145    if (SemaBuiltinFPClassification(TheCall, 6))
146      return ExprError();
147    break;
148  case Builtin::BI__builtin_isfinite:
149  case Builtin::BI__builtin_isinf:
150  case Builtin::BI__builtin_isinf_sign:
151  case Builtin::BI__builtin_isnan:
152  case Builtin::BI__builtin_isnormal:
153    if (SemaBuiltinFPClassification(TheCall))
154      return ExprError();
155    break;
156  case Builtin::BI__builtin_return_address:
157  case Builtin::BI__builtin_frame_address:
158    if (SemaBuiltinStackAddress(TheCall))
159      return ExprError();
160    break;
161  case Builtin::BI__builtin_eh_return_data_regno:
162    if (SemaBuiltinEHReturnDataRegNo(TheCall))
163      return ExprError();
164    break;
165  case Builtin::BI__builtin_shufflevector:
166    return SemaBuiltinShuffleVector(TheCall);
167    // TheCall will be freed by the smart pointer here, but that's fine, since
168    // SemaBuiltinShuffleVector guts it, but then doesn't release it.
169  case Builtin::BI__builtin_prefetch:
170    if (SemaBuiltinPrefetch(TheCall))
171      return ExprError();
172    break;
173  case Builtin::BI__builtin_object_size:
174    if (SemaBuiltinObjectSize(TheCall))
175      return ExprError();
176    break;
177  case Builtin::BI__builtin_longjmp:
178    if (SemaBuiltinLongjmp(TheCall))
179      return ExprError();
180    break;
181  case Builtin::BI__sync_fetch_and_add:
182  case Builtin::BI__sync_fetch_and_sub:
183  case Builtin::BI__sync_fetch_and_or:
184  case Builtin::BI__sync_fetch_and_and:
185  case Builtin::BI__sync_fetch_and_xor:
186  case Builtin::BI__sync_fetch_and_nand:
187  case Builtin::BI__sync_add_and_fetch:
188  case Builtin::BI__sync_sub_and_fetch:
189  case Builtin::BI__sync_and_and_fetch:
190  case Builtin::BI__sync_or_and_fetch:
191  case Builtin::BI__sync_xor_and_fetch:
192  case Builtin::BI__sync_nand_and_fetch:
193  case Builtin::BI__sync_val_compare_and_swap:
194  case Builtin::BI__sync_bool_compare_and_swap:
195  case Builtin::BI__sync_lock_test_and_set:
196  case Builtin::BI__sync_lock_release:
197    if (SemaBuiltinAtomicOverloaded(TheCall))
198      return ExprError();
199    break;
200  }
201
202  return move(TheCallResult);
203}
204
205/// CheckFunctionCall - Check a direct function call for various correctness
206/// and safety properties not strictly enforced by the C type system.
207bool Sema::CheckFunctionCall(FunctionDecl *FDecl, CallExpr *TheCall) {
208  // Get the IdentifierInfo* for the called function.
209  IdentifierInfo *FnInfo = FDecl->getIdentifier();
210
211  // None of the checks below are needed for functions that don't have
212  // simple names (e.g., C++ conversion functions).
213  if (!FnInfo)
214    return false;
215
216  // FIXME: This mechanism should be abstracted to be less fragile and
217  // more efficient. For example, just map function ids to custom
218  // handlers.
219
220  // Printf checking.
221  if (const FormatAttr *Format = FDecl->getAttr<FormatAttr>()) {
222    if (CheckablePrintfAttr(Format, TheCall)) {
223      bool HasVAListArg = Format->getFirstArg() == 0;
224      if (!HasVAListArg) {
225        if (const FunctionProtoType *Proto
226            = FDecl->getType()->getAs<FunctionProtoType>())
227          HasVAListArg = !Proto->isVariadic();
228      }
229      CheckPrintfArguments(TheCall, HasVAListArg, Format->getFormatIdx() - 1,
230                           HasVAListArg ? 0 : Format->getFirstArg() - 1);
231    }
232  }
233
234  for (const NonNullAttr *NonNull = FDecl->getAttr<NonNullAttr>(); NonNull;
235       NonNull = NonNull->getNext<NonNullAttr>())
236    CheckNonNullArguments(NonNull, TheCall);
237
238  return false;
239}
240
241bool Sema::CheckBlockCall(NamedDecl *NDecl, CallExpr *TheCall) {
242  // Printf checking.
243  const FormatAttr *Format = NDecl->getAttr<FormatAttr>();
244  if (!Format)
245    return false;
246
247  const VarDecl *V = dyn_cast<VarDecl>(NDecl);
248  if (!V)
249    return false;
250
251  QualType Ty = V->getType();
252  if (!Ty->isBlockPointerType())
253    return false;
254
255  if (!CheckablePrintfAttr(Format, TheCall))
256    return false;
257
258  bool HasVAListArg = Format->getFirstArg() == 0;
259  if (!HasVAListArg) {
260    const FunctionType *FT =
261      Ty->getAs<BlockPointerType>()->getPointeeType()->getAs<FunctionType>();
262    if (const FunctionProtoType *Proto = dyn_cast<FunctionProtoType>(FT))
263      HasVAListArg = !Proto->isVariadic();
264  }
265  CheckPrintfArguments(TheCall, HasVAListArg, Format->getFormatIdx() - 1,
266                       HasVAListArg ? 0 : Format->getFirstArg() - 1);
267
268  return false;
269}
270
271/// SemaBuiltinAtomicOverloaded - We have a call to a function like
272/// __sync_fetch_and_add, which is an overloaded function based on the pointer
273/// type of its first argument.  The main ActOnCallExpr routines have already
274/// promoted the types of arguments because all of these calls are prototyped as
275/// void(...).
276///
277/// This function goes through and does final semantic checking for these
278/// builtins,
279bool Sema::SemaBuiltinAtomicOverloaded(CallExpr *TheCall) {
280  DeclRefExpr *DRE =cast<DeclRefExpr>(TheCall->getCallee()->IgnoreParenCasts());
281  FunctionDecl *FDecl = cast<FunctionDecl>(DRE->getDecl());
282
283  // Ensure that we have at least one argument to do type inference from.
284  if (TheCall->getNumArgs() < 1)
285    return Diag(TheCall->getLocEnd(), diag::err_typecheck_call_too_few_args)
286              << 0 << TheCall->getCallee()->getSourceRange();
287
288  // Inspect the first argument of the atomic builtin.  This should always be
289  // a pointer type, whose element is an integral scalar or pointer type.
290  // Because it is a pointer type, we don't have to worry about any implicit
291  // casts here.
292  Expr *FirstArg = TheCall->getArg(0);
293  if (!FirstArg->getType()->isPointerType())
294    return Diag(DRE->getLocStart(), diag::err_atomic_builtin_must_be_pointer)
295             << FirstArg->getType() << FirstArg->getSourceRange();
296
297  QualType ValType = FirstArg->getType()->getAs<PointerType>()->getPointeeType();
298  if (!ValType->isIntegerType() && !ValType->isPointerType() &&
299      !ValType->isBlockPointerType())
300    return Diag(DRE->getLocStart(),
301                diag::err_atomic_builtin_must_be_pointer_intptr)
302             << FirstArg->getType() << FirstArg->getSourceRange();
303
304  // We need to figure out which concrete builtin this maps onto.  For example,
305  // __sync_fetch_and_add with a 2 byte object turns into
306  // __sync_fetch_and_add_2.
307#define BUILTIN_ROW(x) \
308  { Builtin::BI##x##_1, Builtin::BI##x##_2, Builtin::BI##x##_4, \
309    Builtin::BI##x##_8, Builtin::BI##x##_16 }
310
311  static const unsigned BuiltinIndices[][5] = {
312    BUILTIN_ROW(__sync_fetch_and_add),
313    BUILTIN_ROW(__sync_fetch_and_sub),
314    BUILTIN_ROW(__sync_fetch_and_or),
315    BUILTIN_ROW(__sync_fetch_and_and),
316    BUILTIN_ROW(__sync_fetch_and_xor),
317    BUILTIN_ROW(__sync_fetch_and_nand),
318
319    BUILTIN_ROW(__sync_add_and_fetch),
320    BUILTIN_ROW(__sync_sub_and_fetch),
321    BUILTIN_ROW(__sync_and_and_fetch),
322    BUILTIN_ROW(__sync_or_and_fetch),
323    BUILTIN_ROW(__sync_xor_and_fetch),
324    BUILTIN_ROW(__sync_nand_and_fetch),
325
326    BUILTIN_ROW(__sync_val_compare_and_swap),
327    BUILTIN_ROW(__sync_bool_compare_and_swap),
328    BUILTIN_ROW(__sync_lock_test_and_set),
329    BUILTIN_ROW(__sync_lock_release)
330  };
331#undef BUILTIN_ROW
332
333  // Determine the index of the size.
334  unsigned SizeIndex;
335  switch (Context.getTypeSizeInChars(ValType).getQuantity()) {
336  case 1: SizeIndex = 0; break;
337  case 2: SizeIndex = 1; break;
338  case 4: SizeIndex = 2; break;
339  case 8: SizeIndex = 3; break;
340  case 16: SizeIndex = 4; break;
341  default:
342    return Diag(DRE->getLocStart(), diag::err_atomic_builtin_pointer_size)
343             << FirstArg->getType() << FirstArg->getSourceRange();
344  }
345
346  // Each of these builtins has one pointer argument, followed by some number of
347  // values (0, 1 or 2) followed by a potentially empty varags list of stuff
348  // that we ignore.  Find out which row of BuiltinIndices to read from as well
349  // as the number of fixed args.
350  unsigned BuiltinID = FDecl->getBuiltinID();
351  unsigned BuiltinIndex, NumFixed = 1;
352  switch (BuiltinID) {
353  default: assert(0 && "Unknown overloaded atomic builtin!");
354  case Builtin::BI__sync_fetch_and_add: BuiltinIndex = 0; break;
355  case Builtin::BI__sync_fetch_and_sub: BuiltinIndex = 1; break;
356  case Builtin::BI__sync_fetch_and_or:  BuiltinIndex = 2; break;
357  case Builtin::BI__sync_fetch_and_and: BuiltinIndex = 3; break;
358  case Builtin::BI__sync_fetch_and_xor: BuiltinIndex = 4; break;
359  case Builtin::BI__sync_fetch_and_nand:BuiltinIndex = 5; break;
360
361  case Builtin::BI__sync_add_and_fetch: BuiltinIndex = 6; break;
362  case Builtin::BI__sync_sub_and_fetch: BuiltinIndex = 7; break;
363  case Builtin::BI__sync_and_and_fetch: BuiltinIndex = 8; break;
364  case Builtin::BI__sync_or_and_fetch:  BuiltinIndex = 9; break;
365  case Builtin::BI__sync_xor_and_fetch: BuiltinIndex =10; break;
366  case Builtin::BI__sync_nand_and_fetch:BuiltinIndex =11; break;
367
368  case Builtin::BI__sync_val_compare_and_swap:
369    BuiltinIndex = 12;
370    NumFixed = 2;
371    break;
372  case Builtin::BI__sync_bool_compare_and_swap:
373    BuiltinIndex = 13;
374    NumFixed = 2;
375    break;
376  case Builtin::BI__sync_lock_test_and_set: BuiltinIndex = 14; break;
377  case Builtin::BI__sync_lock_release:
378    BuiltinIndex = 15;
379    NumFixed = 0;
380    break;
381  }
382
383  // Now that we know how many fixed arguments we expect, first check that we
384  // have at least that many.
385  if (TheCall->getNumArgs() < 1+NumFixed)
386    return Diag(TheCall->getLocEnd(), diag::err_typecheck_call_too_few_args)
387            << 0 << TheCall->getCallee()->getSourceRange();
388
389
390  // Get the decl for the concrete builtin from this, we can tell what the
391  // concrete integer type we should convert to is.
392  unsigned NewBuiltinID = BuiltinIndices[BuiltinIndex][SizeIndex];
393  const char *NewBuiltinName = Context.BuiltinInfo.GetName(NewBuiltinID);
394  IdentifierInfo *NewBuiltinII = PP.getIdentifierInfo(NewBuiltinName);
395  FunctionDecl *NewBuiltinDecl =
396    cast<FunctionDecl>(LazilyCreateBuiltin(NewBuiltinII, NewBuiltinID,
397                                           TUScope, false, DRE->getLocStart()));
398  const FunctionProtoType *BuiltinFT =
399    NewBuiltinDecl->getType()->getAs<FunctionProtoType>();
400  ValType = BuiltinFT->getArgType(0)->getAs<PointerType>()->getPointeeType();
401
402  // If the first type needs to be converted (e.g. void** -> int*), do it now.
403  if (BuiltinFT->getArgType(0) != FirstArg->getType()) {
404    ImpCastExprToType(FirstArg, BuiltinFT->getArgType(0), CastExpr::CK_BitCast);
405    TheCall->setArg(0, FirstArg);
406  }
407
408  // Next, walk the valid ones promoting to the right type.
409  for (unsigned i = 0; i != NumFixed; ++i) {
410    Expr *Arg = TheCall->getArg(i+1);
411
412    // If the argument is an implicit cast, then there was a promotion due to
413    // "...", just remove it now.
414    if (ImplicitCastExpr *ICE = dyn_cast<ImplicitCastExpr>(Arg)) {
415      Arg = ICE->getSubExpr();
416      ICE->setSubExpr(0);
417      ICE->Destroy(Context);
418      TheCall->setArg(i+1, Arg);
419    }
420
421    // GCC does an implicit conversion to the pointer or integer ValType.  This
422    // can fail in some cases (1i -> int**), check for this error case now.
423    CastExpr::CastKind Kind = CastExpr::CK_Unknown;
424    CXXMethodDecl *ConversionDecl = 0;
425    if (CheckCastTypes(Arg->getSourceRange(), ValType, Arg, Kind,
426                       ConversionDecl))
427      return true;
428
429    // Okay, we have something that *can* be converted to the right type.  Check
430    // to see if there is a potentially weird extension going on here.  This can
431    // happen when you do an atomic operation on something like an char* and
432    // pass in 42.  The 42 gets converted to char.  This is even more strange
433    // for things like 45.123 -> char, etc.
434    // FIXME: Do this check.
435    ImpCastExprToType(Arg, ValType, Kind, /*isLvalue=*/false);
436    TheCall->setArg(i+1, Arg);
437  }
438
439  // Switch the DeclRefExpr to refer to the new decl.
440  DRE->setDecl(NewBuiltinDecl);
441  DRE->setType(NewBuiltinDecl->getType());
442
443  // Set the callee in the CallExpr.
444  // FIXME: This leaks the original parens and implicit casts.
445  Expr *PromotedCall = DRE;
446  UsualUnaryConversions(PromotedCall);
447  TheCall->setCallee(PromotedCall);
448
449
450  // Change the result type of the call to match the result type of the decl.
451  TheCall->setType(NewBuiltinDecl->getResultType());
452  return false;
453}
454
455
456/// CheckObjCString - Checks that the argument to the builtin
457/// CFString constructor is correct
458/// FIXME: GCC currently emits the following warning:
459/// "warning: input conversion stopped due to an input byte that does not
460///           belong to the input codeset UTF-8"
461/// Note: It might also make sense to do the UTF-16 conversion here (would
462/// simplify the backend).
463bool Sema::CheckObjCString(Expr *Arg) {
464  Arg = Arg->IgnoreParenCasts();
465  StringLiteral *Literal = dyn_cast<StringLiteral>(Arg);
466
467  if (!Literal || Literal->isWide()) {
468    Diag(Arg->getLocStart(), diag::err_cfstring_literal_not_string_constant)
469      << Arg->getSourceRange();
470    return true;
471  }
472
473  const char *Data = Literal->getStrData();
474  unsigned Length = Literal->getByteLength();
475
476  for (unsigned i = 0; i < Length; ++i) {
477    if (!Data[i]) {
478      Diag(getLocationOfStringLiteralByte(Literal, i),
479           diag::warn_cfstring_literal_contains_nul_character)
480        << Arg->getSourceRange();
481      break;
482    }
483  }
484
485  return false;
486}
487
488/// SemaBuiltinVAStart - Check the arguments to __builtin_va_start for validity.
489/// Emit an error and return true on failure, return false on success.
490bool Sema::SemaBuiltinVAStart(CallExpr *TheCall) {
491  Expr *Fn = TheCall->getCallee();
492  if (TheCall->getNumArgs() > 2) {
493    Diag(TheCall->getArg(2)->getLocStart(),
494         diag::err_typecheck_call_too_many_args)
495      << 0 /*function call*/ << Fn->getSourceRange()
496      << SourceRange(TheCall->getArg(2)->getLocStart(),
497                     (*(TheCall->arg_end()-1))->getLocEnd());
498    return true;
499  }
500
501  if (TheCall->getNumArgs() < 2) {
502    return Diag(TheCall->getLocEnd(), diag::err_typecheck_call_too_few_args)
503      << 0 /*function call*/;
504  }
505
506  // Determine whether the current function is variadic or not.
507  bool isVariadic;
508  if (CurBlock)
509    isVariadic = CurBlock->isVariadic;
510  else if (getCurFunctionDecl()) {
511    if (FunctionProtoType* FTP =
512            dyn_cast<FunctionProtoType>(getCurFunctionDecl()->getType()))
513      isVariadic = FTP->isVariadic();
514    else
515      isVariadic = false;
516  } else {
517    isVariadic = getCurMethodDecl()->isVariadic();
518  }
519
520  if (!isVariadic) {
521    Diag(Fn->getLocStart(), diag::err_va_start_used_in_non_variadic_function);
522    return true;
523  }
524
525  // Verify that the second argument to the builtin is the last argument of the
526  // current function or method.
527  bool SecondArgIsLastNamedArgument = false;
528  const Expr *Arg = TheCall->getArg(1)->IgnoreParenCasts();
529
530  if (const DeclRefExpr *DR = dyn_cast<DeclRefExpr>(Arg)) {
531    if (const ParmVarDecl *PV = dyn_cast<ParmVarDecl>(DR->getDecl())) {
532      // FIXME: This isn't correct for methods (results in bogus warning).
533      // Get the last formal in the current function.
534      const ParmVarDecl *LastArg;
535      if (CurBlock)
536        LastArg = *(CurBlock->TheDecl->param_end()-1);
537      else if (FunctionDecl *FD = getCurFunctionDecl())
538        LastArg = *(FD->param_end()-1);
539      else
540        LastArg = *(getCurMethodDecl()->param_end()-1);
541      SecondArgIsLastNamedArgument = PV == LastArg;
542    }
543  }
544
545  if (!SecondArgIsLastNamedArgument)
546    Diag(TheCall->getArg(1)->getLocStart(),
547         diag::warn_second_parameter_of_va_start_not_last_named_argument);
548  return false;
549}
550
551/// SemaBuiltinUnorderedCompare - Handle functions like __builtin_isgreater and
552/// friends.  This is declared to take (...), so we have to check everything.
553bool Sema::SemaBuiltinUnorderedCompare(CallExpr *TheCall) {
554  if (TheCall->getNumArgs() < 2)
555    return Diag(TheCall->getLocEnd(), diag::err_typecheck_call_too_few_args)
556      << 0 /*function call*/;
557  if (TheCall->getNumArgs() > 2)
558    return Diag(TheCall->getArg(2)->getLocStart(),
559                diag::err_typecheck_call_too_many_args)
560      << 0 /*function call*/
561      << SourceRange(TheCall->getArg(2)->getLocStart(),
562                     (*(TheCall->arg_end()-1))->getLocEnd());
563
564  Expr *OrigArg0 = TheCall->getArg(0);
565  Expr *OrigArg1 = TheCall->getArg(1);
566
567  // Do standard promotions between the two arguments, returning their common
568  // type.
569  QualType Res = UsualArithmeticConversions(OrigArg0, OrigArg1, false);
570
571  // Make sure any conversions are pushed back into the call; this is
572  // type safe since unordered compare builtins are declared as "_Bool
573  // foo(...)".
574  TheCall->setArg(0, OrigArg0);
575  TheCall->setArg(1, OrigArg1);
576
577  if (OrigArg0->isTypeDependent() || OrigArg1->isTypeDependent())
578    return false;
579
580  // If the common type isn't a real floating type, then the arguments were
581  // invalid for this operation.
582  if (!Res->isRealFloatingType())
583    return Diag(OrigArg0->getLocStart(),
584                diag::err_typecheck_call_invalid_ordered_compare)
585      << OrigArg0->getType() << OrigArg1->getType()
586      << SourceRange(OrigArg0->getLocStart(), OrigArg1->getLocEnd());
587
588  return false;
589}
590
591/// SemaBuiltinSemaBuiltinFPClassification - Handle functions like
592/// __builtin_isnan and friends.  This is declared to take (...), so we have
593/// to check everything.
594bool Sema::SemaBuiltinFPClassification(CallExpr *TheCall, unsigned LastArg) {
595  if (TheCall->getNumArgs() < LastArg)
596    return Diag(TheCall->getLocEnd(), diag::err_typecheck_call_too_few_args)
597      << 0 /*function call*/;
598  if (TheCall->getNumArgs() > LastArg)
599    return Diag(TheCall->getArg(LastArg)->getLocStart(),
600                diag::err_typecheck_call_too_many_args)
601      << 0 /*function call*/
602      << SourceRange(TheCall->getArg(LastArg)->getLocStart(),
603                     (*(TheCall->arg_end()-1))->getLocEnd());
604
605  Expr *OrigArg = TheCall->getArg(LastArg-1);
606
607  if (OrigArg->isTypeDependent())
608    return false;
609
610  // This operation requires a floating-point number
611  if (!OrigArg->getType()->isRealFloatingType())
612    return Diag(OrigArg->getLocStart(),
613                diag::err_typecheck_call_invalid_unary_fp)
614      << OrigArg->getType() << OrigArg->getSourceRange();
615
616  return false;
617}
618
619bool Sema::SemaBuiltinStackAddress(CallExpr *TheCall) {
620  // The signature for these builtins is exact; the only thing we need
621  // to check is that the argument is a constant.
622  SourceLocation Loc;
623  if (!TheCall->getArg(0)->isTypeDependent() &&
624      !TheCall->getArg(0)->isValueDependent() &&
625      !TheCall->getArg(0)->isIntegerConstantExpr(Context, &Loc))
626    return Diag(Loc, diag::err_stack_const_level) << TheCall->getSourceRange();
627
628  return false;
629}
630
631/// SemaBuiltinShuffleVector - Handle __builtin_shufflevector.
632// This is declared to take (...), so we have to check everything.
633Action::OwningExprResult Sema::SemaBuiltinShuffleVector(CallExpr *TheCall) {
634  if (TheCall->getNumArgs() < 3)
635    return ExprError(Diag(TheCall->getLocEnd(),
636                          diag::err_typecheck_call_too_few_args)
637      << 0 /*function call*/ << TheCall->getSourceRange());
638
639  unsigned numElements = std::numeric_limits<unsigned>::max();
640  if (!TheCall->getArg(0)->isTypeDependent() &&
641      !TheCall->getArg(1)->isTypeDependent()) {
642    QualType FAType = TheCall->getArg(0)->getType();
643    QualType SAType = TheCall->getArg(1)->getType();
644
645    if (!FAType->isVectorType() || !SAType->isVectorType()) {
646      Diag(TheCall->getLocStart(), diag::err_shufflevector_non_vector)
647        << SourceRange(TheCall->getArg(0)->getLocStart(),
648                       TheCall->getArg(1)->getLocEnd());
649      return ExprError();
650    }
651
652    if (!Context.hasSameUnqualifiedType(FAType, SAType)) {
653      Diag(TheCall->getLocStart(), diag::err_shufflevector_incompatible_vector)
654        << SourceRange(TheCall->getArg(0)->getLocStart(),
655                       TheCall->getArg(1)->getLocEnd());
656      return ExprError();
657    }
658
659    numElements = FAType->getAs<VectorType>()->getNumElements();
660    if (TheCall->getNumArgs() != numElements+2) {
661      if (TheCall->getNumArgs() < numElements+2)
662        return ExprError(Diag(TheCall->getLocEnd(),
663                              diag::err_typecheck_call_too_few_args)
664                 << 0 /*function call*/ << TheCall->getSourceRange());
665      return ExprError(Diag(TheCall->getLocEnd(),
666                            diag::err_typecheck_call_too_many_args)
667                 << 0 /*function call*/ << TheCall->getSourceRange());
668    }
669  }
670
671  for (unsigned i = 2; i < TheCall->getNumArgs(); i++) {
672    if (TheCall->getArg(i)->isTypeDependent() ||
673        TheCall->getArg(i)->isValueDependent())
674      continue;
675
676    llvm::APSInt Result(32);
677    if (!TheCall->getArg(i)->isIntegerConstantExpr(Result, Context))
678      return ExprError(Diag(TheCall->getLocStart(),
679                  diag::err_shufflevector_nonconstant_argument)
680                << TheCall->getArg(i)->getSourceRange());
681
682    if (Result.getActiveBits() > 64 || Result.getZExtValue() >= numElements*2)
683      return ExprError(Diag(TheCall->getLocStart(),
684                  diag::err_shufflevector_argument_too_large)
685               << TheCall->getArg(i)->getSourceRange());
686  }
687
688  llvm::SmallVector<Expr*, 32> exprs;
689
690  for (unsigned i = 0, e = TheCall->getNumArgs(); i != e; i++) {
691    exprs.push_back(TheCall->getArg(i));
692    TheCall->setArg(i, 0);
693  }
694
695  return Owned(new (Context) ShuffleVectorExpr(Context, exprs.begin(),
696                                            exprs.size(), exprs[0]->getType(),
697                                            TheCall->getCallee()->getLocStart(),
698                                            TheCall->getRParenLoc()));
699}
700
701/// SemaBuiltinPrefetch - Handle __builtin_prefetch.
702// This is declared to take (const void*, ...) and can take two
703// optional constant int args.
704bool Sema::SemaBuiltinPrefetch(CallExpr *TheCall) {
705  unsigned NumArgs = TheCall->getNumArgs();
706
707  if (NumArgs > 3)
708    return Diag(TheCall->getLocEnd(), diag::err_typecheck_call_too_many_args)
709             << 0 /*function call*/ << TheCall->getSourceRange();
710
711  // Argument 0 is checked for us and the remaining arguments must be
712  // constant integers.
713  for (unsigned i = 1; i != NumArgs; ++i) {
714    Expr *Arg = TheCall->getArg(i);
715    if (Arg->isTypeDependent())
716      continue;
717
718    if (!Arg->getType()->isIntegralType())
719      return Diag(TheCall->getLocStart(), diag::err_prefetch_invalid_arg_type)
720              << Arg->getSourceRange();
721
722    ImpCastExprToType(Arg, Context.IntTy, CastExpr::CK_IntegralCast);
723    TheCall->setArg(i, Arg);
724
725    if (Arg->isValueDependent())
726      continue;
727
728    llvm::APSInt Result;
729    if (!Arg->isIntegerConstantExpr(Result, Context))
730      return Diag(TheCall->getLocStart(), diag::err_prefetch_invalid_arg_ice)
731        << SourceRange(Arg->getLocStart(), Arg->getLocEnd());
732
733    // FIXME: gcc issues a warning and rewrites these to 0. These
734    // seems especially odd for the third argument since the default
735    // is 3.
736    if (i == 1) {
737      if (Result.getLimitedValue() > 1)
738        return Diag(TheCall->getLocStart(), diag::err_argument_invalid_range)
739             << "0" << "1" << Arg->getSourceRange();
740    } else {
741      if (Result.getLimitedValue() > 3)
742        return Diag(TheCall->getLocStart(), diag::err_argument_invalid_range)
743            << "0" << "3" << Arg->getSourceRange();
744    }
745  }
746
747  return false;
748}
749
750/// SemaBuiltinEHReturnDataRegNo - Handle __builtin_eh_return_data_regno, the
751/// operand must be an integer constant.
752bool Sema::SemaBuiltinEHReturnDataRegNo(CallExpr *TheCall) {
753  llvm::APSInt Result;
754  if (!TheCall->getArg(0)->isIntegerConstantExpr(Result, Context))
755    return Diag(TheCall->getLocStart(), diag::err_expr_not_ice)
756      << TheCall->getArg(0)->getSourceRange();
757
758  return false;
759}
760
761
762/// SemaBuiltinObjectSize - Handle __builtin_object_size(void *ptr,
763/// int type). This simply type checks that type is one of the defined
764/// constants (0-3).
765// For compatability check 0-3, llvm only handles 0 and 2.
766bool Sema::SemaBuiltinObjectSize(CallExpr *TheCall) {
767  Expr *Arg = TheCall->getArg(1);
768  if (Arg->isTypeDependent())
769    return false;
770
771  QualType ArgType = Arg->getType();
772  const BuiltinType *BT = ArgType->getAs<BuiltinType>();
773  llvm::APSInt Result(32);
774  if (!BT || BT->getKind() != BuiltinType::Int)
775    return Diag(TheCall->getLocStart(), diag::err_object_size_invalid_argument)
776             << SourceRange(Arg->getLocStart(), Arg->getLocEnd());
777
778  if (Arg->isValueDependent())
779    return false;
780
781  if (!Arg->isIntegerConstantExpr(Result, Context)) {
782    return Diag(TheCall->getLocStart(), diag::err_object_size_invalid_argument)
783             << SourceRange(Arg->getLocStart(), Arg->getLocEnd());
784  }
785
786  if (Result.getSExtValue() < 0 || Result.getSExtValue() > 3) {
787    return Diag(TheCall->getLocStart(), diag::err_argument_invalid_range)
788             << "0" << "3" << SourceRange(Arg->getLocStart(), Arg->getLocEnd());
789  }
790
791  return false;
792}
793
794/// SemaBuiltinLongjmp - Handle __builtin_longjmp(void *env[5], int val).
795/// This checks that val is a constant 1.
796bool Sema::SemaBuiltinLongjmp(CallExpr *TheCall) {
797  Expr *Arg = TheCall->getArg(1);
798  if (Arg->isTypeDependent() || Arg->isValueDependent())
799    return false;
800
801  llvm::APSInt Result(32);
802  if (!Arg->isIntegerConstantExpr(Result, Context) || Result != 1)
803    return Diag(TheCall->getLocStart(), diag::err_builtin_longjmp_invalid_val)
804             << SourceRange(Arg->getLocStart(), Arg->getLocEnd());
805
806  return false;
807}
808
809// Handle i > 1 ? "x" : "y", recursivelly
810bool Sema::SemaCheckStringLiteral(const Expr *E, const CallExpr *TheCall,
811                                  bool HasVAListArg,
812                                  unsigned format_idx, unsigned firstDataArg) {
813  if (E->isTypeDependent() || E->isValueDependent())
814    return false;
815
816  switch (E->getStmtClass()) {
817  case Stmt::ConditionalOperatorClass: {
818    const ConditionalOperator *C = cast<ConditionalOperator>(E);
819    return SemaCheckStringLiteral(C->getTrueExpr(), TheCall,
820                                  HasVAListArg, format_idx, firstDataArg)
821        && SemaCheckStringLiteral(C->getRHS(), TheCall,
822                                  HasVAListArg, format_idx, firstDataArg);
823  }
824
825  case Stmt::ImplicitCastExprClass: {
826    const ImplicitCastExpr *Expr = cast<ImplicitCastExpr>(E);
827    return SemaCheckStringLiteral(Expr->getSubExpr(), TheCall, HasVAListArg,
828                                  format_idx, firstDataArg);
829  }
830
831  case Stmt::ParenExprClass: {
832    const ParenExpr *Expr = cast<ParenExpr>(E);
833    return SemaCheckStringLiteral(Expr->getSubExpr(), TheCall, HasVAListArg,
834                                  format_idx, firstDataArg);
835  }
836
837  case Stmt::DeclRefExprClass: {
838    const DeclRefExpr *DR = cast<DeclRefExpr>(E);
839
840    // As an exception, do not flag errors for variables binding to
841    // const string literals.
842    if (const VarDecl *VD = dyn_cast<VarDecl>(DR->getDecl())) {
843      bool isConstant = false;
844      QualType T = DR->getType();
845
846      if (const ArrayType *AT = Context.getAsArrayType(T)) {
847        isConstant = AT->getElementType().isConstant(Context);
848      } else if (const PointerType *PT = T->getAs<PointerType>()) {
849        isConstant = T.isConstant(Context) &&
850                     PT->getPointeeType().isConstant(Context);
851      }
852
853      if (isConstant) {
854        if (const Expr *Init = VD->getAnyInitializer())
855          return SemaCheckStringLiteral(Init, TheCall,
856                                        HasVAListArg, format_idx, firstDataArg);
857      }
858
859      // For vprintf* functions (i.e., HasVAListArg==true), we add a
860      // special check to see if the format string is a function parameter
861      // of the function calling the printf function.  If the function
862      // has an attribute indicating it is a printf-like function, then we
863      // should suppress warnings concerning non-literals being used in a call
864      // to a vprintf function.  For example:
865      //
866      // void
867      // logmessage(char const *fmt __attribute__ (format (printf, 1, 2)), ...){
868      //      va_list ap;
869      //      va_start(ap, fmt);
870      //      vprintf(fmt, ap);  // Do NOT emit a warning about "fmt".
871      //      ...
872      //
873      //
874      //  FIXME: We don't have full attribute support yet, so just check to see
875      //    if the argument is a DeclRefExpr that references a parameter.  We'll
876      //    add proper support for checking the attribute later.
877      if (HasVAListArg)
878        if (isa<ParmVarDecl>(VD))
879          return true;
880    }
881
882    return false;
883  }
884
885  case Stmt::CallExprClass: {
886    const CallExpr *CE = cast<CallExpr>(E);
887    if (const ImplicitCastExpr *ICE
888          = dyn_cast<ImplicitCastExpr>(CE->getCallee())) {
889      if (const DeclRefExpr *DRE = dyn_cast<DeclRefExpr>(ICE->getSubExpr())) {
890        if (const FunctionDecl *FD = dyn_cast<FunctionDecl>(DRE->getDecl())) {
891          if (const FormatArgAttr *FA = FD->getAttr<FormatArgAttr>()) {
892            unsigned ArgIndex = FA->getFormatIdx();
893            const Expr *Arg = CE->getArg(ArgIndex - 1);
894
895            return SemaCheckStringLiteral(Arg, TheCall, HasVAListArg,
896                                          format_idx, firstDataArg);
897          }
898        }
899      }
900    }
901
902    return false;
903  }
904  case Stmt::ObjCStringLiteralClass:
905  case Stmt::StringLiteralClass: {
906    const StringLiteral *StrE = NULL;
907
908    if (const ObjCStringLiteral *ObjCFExpr = dyn_cast<ObjCStringLiteral>(E))
909      StrE = ObjCFExpr->getString();
910    else
911      StrE = cast<StringLiteral>(E);
912
913    if (StrE) {
914      CheckPrintfString(StrE, E, TheCall, HasVAListArg, format_idx,
915                        firstDataArg);
916      return true;
917    }
918
919    return false;
920  }
921
922  default:
923    return false;
924  }
925}
926
927void
928Sema::CheckNonNullArguments(const NonNullAttr *NonNull,
929                            const CallExpr *TheCall) {
930  for (NonNullAttr::iterator i = NonNull->begin(), e = NonNull->end();
931       i != e; ++i) {
932    const Expr *ArgExpr = TheCall->getArg(*i);
933    if (ArgExpr->isNullPointerConstant(Context,
934                                       Expr::NPC_ValueDependentIsNotNull))
935      Diag(TheCall->getCallee()->getLocStart(), diag::warn_null_arg)
936        << ArgExpr->getSourceRange();
937  }
938}
939
940/// CheckPrintfArguments - Check calls to printf (and similar functions) for
941/// correct use of format strings.
942///
943///  HasVAListArg - A predicate indicating whether the printf-like
944///    function is passed an explicit va_arg argument (e.g., vprintf)
945///
946///  format_idx - The index into Args for the format string.
947///
948/// Improper format strings to functions in the printf family can be
949/// the source of bizarre bugs and very serious security holes.  A
950/// good source of information is available in the following paper
951/// (which includes additional references):
952///
953///  FormatGuard: Automatic Protection From printf Format String
954///  Vulnerabilities, Proceedings of the 10th USENIX Security Symposium, 2001.
955///
956/// Functionality implemented:
957///
958///  We can statically check the following properties for string
959///  literal format strings for non v.*printf functions (where the
960///  arguments are passed directly):
961//
962///  (1) Are the number of format conversions equal to the number of
963///      data arguments?
964///
965///  (2) Does each format conversion correctly match the type of the
966///      corresponding data argument?  (TODO)
967///
968/// Moreover, for all printf functions we can:
969///
970///  (3) Check for a missing format string (when not caught by type checking).
971///
972///  (4) Check for no-operation flags; e.g. using "#" with format
973///      conversion 'c'  (TODO)
974///
975///  (5) Check the use of '%n', a major source of security holes.
976///
977///  (6) Check for malformed format conversions that don't specify anything.
978///
979///  (7) Check for empty format strings.  e.g: printf("");
980///
981///  (8) Check that the format string is a wide literal.
982///
983/// All of these checks can be done by parsing the format string.
984///
985/// For now, we ONLY do (1), (3), (5), (6), (7), and (8).
986void
987Sema::CheckPrintfArguments(const CallExpr *TheCall, bool HasVAListArg,
988                           unsigned format_idx, unsigned firstDataArg) {
989  const Expr *Fn = TheCall->getCallee();
990
991  // The way the format attribute works in GCC, the implicit this argument
992  // of member functions is counted. However, it doesn't appear in our own
993  // lists, so decrement format_idx in that case.
994  if (isa<CXXMemberCallExpr>(TheCall)) {
995    // Catch a format attribute mistakenly referring to the object argument.
996    if (format_idx == 0)
997      return;
998    --format_idx;
999    if(firstDataArg != 0)
1000      --firstDataArg;
1001  }
1002
1003  // CHECK: printf-like function is called with no format string.
1004  if (format_idx >= TheCall->getNumArgs()) {
1005    Diag(TheCall->getRParenLoc(), diag::warn_printf_missing_format_string)
1006      << Fn->getSourceRange();
1007    return;
1008  }
1009
1010  const Expr *OrigFormatExpr = TheCall->getArg(format_idx)->IgnoreParenCasts();
1011
1012  // CHECK: format string is not a string literal.
1013  //
1014  // Dynamically generated format strings are difficult to
1015  // automatically vet at compile time.  Requiring that format strings
1016  // are string literals: (1) permits the checking of format strings by
1017  // the compiler and thereby (2) can practically remove the source of
1018  // many format string exploits.
1019
1020  // Format string can be either ObjC string (e.g. @"%d") or
1021  // C string (e.g. "%d")
1022  // ObjC string uses the same format specifiers as C string, so we can use
1023  // the same format string checking logic for both ObjC and C strings.
1024  if (SemaCheckStringLiteral(OrigFormatExpr, TheCall, HasVAListArg, format_idx,
1025                             firstDataArg))
1026    return;  // Literal format string found, check done!
1027
1028  // If there are no arguments specified, warn with -Wformat-security, otherwise
1029  // warn only with -Wformat-nonliteral.
1030  if (TheCall->getNumArgs() == format_idx+1)
1031    Diag(TheCall->getArg(format_idx)->getLocStart(),
1032         diag::warn_printf_nonliteral_noargs)
1033      << OrigFormatExpr->getSourceRange();
1034  else
1035    Diag(TheCall->getArg(format_idx)->getLocStart(),
1036         diag::warn_printf_nonliteral)
1037           << OrigFormatExpr->getSourceRange();
1038}
1039
1040namespace {
1041class CheckPrintfHandler : public analyze_printf::FormatStringHandler {
1042  Sema &S;
1043  const StringLiteral *FExpr;
1044  const Expr *OrigFormatExpr;
1045  unsigned NumConversions;
1046  const unsigned NumDataArgs;
1047  const bool IsObjCLiteral;
1048  const char *Beg; // Start of format string.
1049  const bool HasVAListArg;
1050  const CallExpr *TheCall;
1051  unsigned FormatIdx;
1052public:
1053  CheckPrintfHandler(Sema &s, const StringLiteral *fexpr,
1054                     const Expr *origFormatExpr,
1055                     unsigned numDataArgs, bool isObjCLiteral,
1056                     const char *beg, bool hasVAListArg,
1057                     const CallExpr *theCall, unsigned formatIdx)
1058    : S(s), FExpr(fexpr), OrigFormatExpr(origFormatExpr),
1059      NumConversions(0), NumDataArgs(numDataArgs),
1060      IsObjCLiteral(isObjCLiteral), Beg(beg),
1061      HasVAListArg(hasVAListArg),
1062      TheCall(theCall), FormatIdx(formatIdx) {}
1063
1064  void DoneProcessing();
1065
1066  void HandleIncompleteFormatSpecifier(const char *startSpecifier,
1067                                       unsigned specifierLen);
1068
1069  void
1070  HandleInvalidConversionSpecifier(const analyze_printf::FormatSpecifier &FS,
1071                                   const char *startSpecifier,
1072                                   unsigned specifierLen);
1073
1074  void HandleNullChar(const char *nullCharacter);
1075
1076  bool HandleFormatSpecifier(const analyze_printf::FormatSpecifier &FS,
1077                             const char *startSpecifier,
1078                             unsigned specifierLen);
1079private:
1080  SourceRange getFormatStringRange();
1081  SourceRange getFormatSpecifierRange(const char *startSpecifier,
1082                                      unsigned specifierLen);
1083  SourceLocation getLocationOfByte(const char *x);
1084
1085  bool HandleAmount(const analyze_printf::OptionalAmount &Amt,
1086                    unsigned MissingArgDiag, unsigned BadTypeDiag,
1087          const char *startSpecifier, unsigned specifierLen);
1088  void HandleFlags(const analyze_printf::FormatSpecifier &FS,
1089                   llvm::StringRef flag, llvm::StringRef cspec,
1090                   const char *startSpecifier, unsigned specifierLen);
1091
1092  const Expr *getDataArg(unsigned i) const;
1093};
1094}
1095
1096SourceRange CheckPrintfHandler::getFormatStringRange() {
1097  return OrigFormatExpr->getSourceRange();
1098}
1099
1100SourceRange CheckPrintfHandler::
1101getFormatSpecifierRange(const char *startSpecifier, unsigned specifierLen) {
1102  return SourceRange(getLocationOfByte(startSpecifier),
1103                     getLocationOfByte(startSpecifier+specifierLen-1));
1104}
1105
1106SourceLocation CheckPrintfHandler::getLocationOfByte(const char *x) {
1107  return S.getLocationOfStringLiteralByte(FExpr, x - Beg);
1108}
1109
1110void CheckPrintfHandler::
1111HandleIncompleteFormatSpecifier(const char *startSpecifier,
1112                                unsigned specifierLen) {
1113  SourceLocation Loc = getLocationOfByte(startSpecifier);
1114  S.Diag(Loc, diag::warn_printf_incomplete_specifier)
1115    << getFormatSpecifierRange(startSpecifier, specifierLen);
1116}
1117
1118void CheckPrintfHandler::
1119HandleInvalidConversionSpecifier(const analyze_printf::FormatSpecifier &FS,
1120                                 const char *startSpecifier,
1121                                 unsigned specifierLen) {
1122
1123  ++NumConversions;
1124  const analyze_printf::ConversionSpecifier &CS =
1125    FS.getConversionSpecifier();
1126  SourceLocation Loc = getLocationOfByte(CS.getStart());
1127  S.Diag(Loc, diag::warn_printf_invalid_conversion)
1128      << llvm::StringRef(CS.getStart(), CS.getLength())
1129      << getFormatSpecifierRange(startSpecifier, specifierLen);
1130}
1131
1132void CheckPrintfHandler::HandleNullChar(const char *nullCharacter) {
1133  // The presence of a null character is likely an error.
1134  S.Diag(getLocationOfByte(nullCharacter),
1135         diag::warn_printf_format_string_contains_null_char)
1136    << getFormatStringRange();
1137}
1138
1139const Expr *CheckPrintfHandler::getDataArg(unsigned i) const {
1140  return TheCall->getArg(FormatIdx + i);
1141}
1142
1143
1144
1145void CheckPrintfHandler::HandleFlags(const analyze_printf::FormatSpecifier &FS,
1146                                     llvm::StringRef flag,
1147                                     llvm::StringRef cspec,
1148                                     const char *startSpecifier,
1149                                     unsigned specifierLen) {
1150  const analyze_printf::ConversionSpecifier &CS = FS.getConversionSpecifier();
1151  S.Diag(getLocationOfByte(CS.getStart()), diag::warn_printf_nonsensical_flag)
1152    << flag << cspec << getFormatSpecifierRange(startSpecifier, specifierLen);
1153}
1154
1155bool
1156CheckPrintfHandler::HandleAmount(const analyze_printf::OptionalAmount &Amt,
1157                                 unsigned MissingArgDiag,
1158                                 unsigned BadTypeDiag,
1159                                 const char *startSpecifier,
1160                                 unsigned specifierLen) {
1161
1162  if (Amt.hasDataArgument()) {
1163    ++NumConversions;
1164    if (!HasVAListArg) {
1165      if (NumConversions > NumDataArgs) {
1166        S.Diag(getLocationOfByte(Amt.getStart()), MissingArgDiag)
1167          << getFormatSpecifierRange(startSpecifier, specifierLen);
1168        // Don't do any more checking.  We will just emit
1169        // spurious errors.
1170        return false;
1171      }
1172
1173      // Type check the data argument.  It should be an 'int'.
1174      // Although not in conformance with C99, we also allow the argument to be
1175      // an 'unsigned int' as that is a reasonably safe case.  GCC also
1176      // doesn't emit a warning for that case.
1177      const Expr *Arg = getDataArg(NumConversions);
1178      QualType T = Arg->getType();
1179
1180      const analyze_printf::ArgTypeResult &ATR = Amt.getArgType(S.Context);
1181      assert(ATR.isValid());
1182
1183      if (!ATR.matchesType(S.Context, T)) {
1184        S.Diag(getLocationOfByte(Amt.getStart()), BadTypeDiag)
1185          << ATR.getRepresentativeType(S.Context) << T
1186          << getFormatSpecifierRange(startSpecifier, specifierLen)
1187          << Arg->getSourceRange();
1188        // Don't do any more checking.  We will just emit
1189        // spurious errors.
1190        return false;
1191      }
1192    }
1193  }
1194  return true;
1195}
1196
1197bool
1198CheckPrintfHandler::HandleFormatSpecifier(const analyze_printf::FormatSpecifier
1199                                            &FS,
1200                                          const char *startSpecifier,
1201                                          unsigned specifierLen) {
1202
1203  using namespace analyze_printf;
1204  const ConversionSpecifier &CS = FS.getConversionSpecifier();
1205
1206  // First check if the field width, precision, and conversion specifier
1207  // have matching data arguments.
1208  if (!HandleAmount(FS.getFieldWidth(),
1209                    diag::warn_printf_asterisk_width_missing_arg,
1210                    diag::warn_printf_asterisk_width_wrong_type,
1211          startSpecifier, specifierLen)) {
1212    return false;
1213  }
1214
1215  if (!HandleAmount(FS.getPrecision(),
1216                    diag::warn_printf_asterisk_precision_missing_arg,
1217                    diag::warn_printf_asterisk_precision_wrong_type,
1218          startSpecifier, specifierLen)) {
1219    return false;
1220  }
1221
1222  // Check for using an Objective-C specific conversion specifier
1223  // in a non-ObjC literal.
1224  if (!IsObjCLiteral && CS.isObjCArg()) {
1225    HandleInvalidConversionSpecifier(FS, startSpecifier, specifierLen);
1226
1227    // Continue checking the other format specifiers.
1228    return true;
1229  }
1230
1231  if (!CS.consumesDataArgument()) {
1232    // FIXME: Technically specifying a precision or field width here
1233    // makes no sense.  Worth issuing a warning at some point.
1234    return true;
1235  }
1236
1237  ++NumConversions;
1238
1239  // Are we using '%n'?  Issue a warning about this being
1240  // a possible security issue.
1241  if (CS.getKind() == ConversionSpecifier::OutIntPtrArg) {
1242    S.Diag(getLocationOfByte(CS.getStart()), diag::warn_printf_write_back)
1243      << getFormatSpecifierRange(startSpecifier, specifierLen);
1244    // Continue checking the other format specifiers.
1245    return true;
1246  }
1247
1248  if (CS.getKind() == ConversionSpecifier::VoidPtrArg) {
1249    if (FS.getPrecision().getHowSpecified() != OptionalAmount::NotSpecified)
1250      S.Diag(getLocationOfByte(CS.getStart()),
1251             diag::warn_printf_nonsensical_precision)
1252        << CS.getCharacters()
1253        << getFormatSpecifierRange(startSpecifier, specifierLen);
1254  }
1255  if (CS.getKind() == ConversionSpecifier::VoidPtrArg ||
1256      CS.getKind() == ConversionSpecifier::CStrArg) {
1257    // FIXME: Instead of using "0", "+", etc., eventually get them from
1258    // the FormatSpecifier.
1259    if (FS.hasLeadingZeros())
1260      HandleFlags(FS, "0", CS.getCharacters(), startSpecifier, specifierLen);
1261    if (FS.hasPlusPrefix())
1262      HandleFlags(FS, "+", CS.getCharacters(), startSpecifier, specifierLen);
1263    if (FS.hasSpacePrefix())
1264      HandleFlags(FS, " ", CS.getCharacters(), startSpecifier, specifierLen);
1265  }
1266
1267  // The remaining checks depend on the data arguments.
1268  if (HasVAListArg)
1269    return true;
1270
1271  if (NumConversions > NumDataArgs) {
1272    S.Diag(getLocationOfByte(CS.getStart()),
1273           diag::warn_printf_insufficient_data_args)
1274      << getFormatSpecifierRange(startSpecifier, specifierLen);
1275    // Don't do any more checking.
1276    return false;
1277  }
1278
1279  // Now type check the data expression that matches the
1280  // format specifier.
1281  const Expr *Ex = getDataArg(NumConversions);
1282  const analyze_printf::ArgTypeResult &ATR = FS.getArgType(S.Context);
1283  if (ATR.isValid() && !ATR.matchesType(S.Context, Ex->getType())) {
1284    // Check if we didn't match because of an implicit cast from a 'char'
1285    // or 'short' to an 'int'.  This is done because printf is a varargs
1286    // function.
1287    if (const ImplicitCastExpr *ICE = dyn_cast<ImplicitCastExpr>(Ex))
1288      if (ICE->getType() == S.Context.IntTy)
1289        if (ATR.matchesType(S.Context, ICE->getSubExpr()->getType()))
1290          return true;
1291
1292    S.Diag(getLocationOfByte(CS.getStart()),
1293           diag::warn_printf_conversion_argument_type_mismatch)
1294      << ATR.getRepresentativeType(S.Context) << Ex->getType()
1295      << getFormatSpecifierRange(startSpecifier, specifierLen)
1296      << Ex->getSourceRange();
1297  }
1298
1299  return true;
1300}
1301
1302void CheckPrintfHandler::DoneProcessing() {
1303  // Does the number of data arguments exceed the number of
1304  // format conversions in the format string?
1305  if (!HasVAListArg && NumConversions < NumDataArgs)
1306    S.Diag(getDataArg(NumConversions+1)->getLocStart(),
1307           diag::warn_printf_too_many_data_args)
1308      << getFormatStringRange();
1309}
1310
1311void Sema::CheckPrintfString(const StringLiteral *FExpr,
1312                             const Expr *OrigFormatExpr,
1313                             const CallExpr *TheCall, bool HasVAListArg,
1314                             unsigned format_idx, unsigned firstDataArg) {
1315
1316  // CHECK: is the format string a wide literal?
1317  if (FExpr->isWide()) {
1318    Diag(FExpr->getLocStart(),
1319         diag::warn_printf_format_string_is_wide_literal)
1320    << OrigFormatExpr->getSourceRange();
1321    return;
1322  }
1323
1324  // Str - The format string.  NOTE: this is NOT null-terminated!
1325  const char *Str = FExpr->getStrData();
1326
1327  // CHECK: empty format string?
1328  unsigned StrLen = FExpr->getByteLength();
1329
1330  if (StrLen == 0) {
1331    Diag(FExpr->getLocStart(), diag::warn_printf_empty_format_string)
1332    << OrigFormatExpr->getSourceRange();
1333    return;
1334  }
1335
1336  CheckPrintfHandler H(*this, FExpr, OrigFormatExpr,
1337                       TheCall->getNumArgs() - firstDataArg,
1338                       isa<ObjCStringLiteral>(OrigFormatExpr), Str,
1339                       HasVAListArg, TheCall, format_idx);
1340
1341  if (!analyze_printf::ParseFormatString(H, Str, Str + StrLen))
1342    H.DoneProcessing();
1343}
1344
1345//===--- CHECK: Return Address of Stack Variable --------------------------===//
1346
1347static DeclRefExpr* EvalVal(Expr *E);
1348static DeclRefExpr* EvalAddr(Expr* E);
1349
1350/// CheckReturnStackAddr - Check if a return statement returns the address
1351///   of a stack variable.
1352void
1353Sema::CheckReturnStackAddr(Expr *RetValExp, QualType lhsType,
1354                           SourceLocation ReturnLoc) {
1355
1356  // Perform checking for returned stack addresses.
1357  if (lhsType->isPointerType() || lhsType->isBlockPointerType()) {
1358    if (DeclRefExpr *DR = EvalAddr(RetValExp))
1359      Diag(DR->getLocStart(), diag::warn_ret_stack_addr)
1360       << DR->getDecl()->getDeclName() << RetValExp->getSourceRange();
1361
1362    // Skip over implicit cast expressions when checking for block expressions.
1363    RetValExp = RetValExp->IgnoreParenCasts();
1364
1365    if (BlockExpr *C = dyn_cast<BlockExpr>(RetValExp))
1366      if (C->hasBlockDeclRefExprs())
1367        Diag(C->getLocStart(), diag::err_ret_local_block)
1368          << C->getSourceRange();
1369
1370    if (AddrLabelExpr *ALE = dyn_cast<AddrLabelExpr>(RetValExp))
1371      Diag(ALE->getLocStart(), diag::warn_ret_addr_label)
1372        << ALE->getSourceRange();
1373
1374  } else if (lhsType->isReferenceType()) {
1375    // Perform checking for stack values returned by reference.
1376    // Check for a reference to the stack
1377    if (DeclRefExpr *DR = EvalVal(RetValExp))
1378      Diag(DR->getLocStart(), diag::warn_ret_stack_ref)
1379        << DR->getDecl()->getDeclName() << RetValExp->getSourceRange();
1380  }
1381}
1382
1383/// EvalAddr - EvalAddr and EvalVal are mutually recursive functions that
1384///  check if the expression in a return statement evaluates to an address
1385///  to a location on the stack.  The recursion is used to traverse the
1386///  AST of the return expression, with recursion backtracking when we
1387///  encounter a subexpression that (1) clearly does not lead to the address
1388///  of a stack variable or (2) is something we cannot determine leads to
1389///  the address of a stack variable based on such local checking.
1390///
1391///  EvalAddr processes expressions that are pointers that are used as
1392///  references (and not L-values).  EvalVal handles all other values.
1393///  At the base case of the recursion is a check for a DeclRefExpr* in
1394///  the refers to a stack variable.
1395///
1396///  This implementation handles:
1397///
1398///   * pointer-to-pointer casts
1399///   * implicit conversions from array references to pointers
1400///   * taking the address of fields
1401///   * arbitrary interplay between "&" and "*" operators
1402///   * pointer arithmetic from an address of a stack variable
1403///   * taking the address of an array element where the array is on the stack
1404static DeclRefExpr* EvalAddr(Expr *E) {
1405  // We should only be called for evaluating pointer expressions.
1406  assert((E->getType()->isAnyPointerType() ||
1407          E->getType()->isBlockPointerType() ||
1408          E->getType()->isObjCQualifiedIdType()) &&
1409         "EvalAddr only works on pointers");
1410
1411  // Our "symbolic interpreter" is just a dispatch off the currently
1412  // viewed AST node.  We then recursively traverse the AST by calling
1413  // EvalAddr and EvalVal appropriately.
1414  switch (E->getStmtClass()) {
1415  case Stmt::ParenExprClass:
1416    // Ignore parentheses.
1417    return EvalAddr(cast<ParenExpr>(E)->getSubExpr());
1418
1419  case Stmt::UnaryOperatorClass: {
1420    // The only unary operator that make sense to handle here
1421    // is AddrOf.  All others don't make sense as pointers.
1422    UnaryOperator *U = cast<UnaryOperator>(E);
1423
1424    if (U->getOpcode() == UnaryOperator::AddrOf)
1425      return EvalVal(U->getSubExpr());
1426    else
1427      return NULL;
1428  }
1429
1430  case Stmt::BinaryOperatorClass: {
1431    // Handle pointer arithmetic.  All other binary operators are not valid
1432    // in this context.
1433    BinaryOperator *B = cast<BinaryOperator>(E);
1434    BinaryOperator::Opcode op = B->getOpcode();
1435
1436    if (op != BinaryOperator::Add && op != BinaryOperator::Sub)
1437      return NULL;
1438
1439    Expr *Base = B->getLHS();
1440
1441    // Determine which argument is the real pointer base.  It could be
1442    // the RHS argument instead of the LHS.
1443    if (!Base->getType()->isPointerType()) Base = B->getRHS();
1444
1445    assert (Base->getType()->isPointerType());
1446    return EvalAddr(Base);
1447  }
1448
1449  // For conditional operators we need to see if either the LHS or RHS are
1450  // valid DeclRefExpr*s.  If one of them is valid, we return it.
1451  case Stmt::ConditionalOperatorClass: {
1452    ConditionalOperator *C = cast<ConditionalOperator>(E);
1453
1454    // Handle the GNU extension for missing LHS.
1455    if (Expr *lhsExpr = C->getLHS())
1456      if (DeclRefExpr* LHS = EvalAddr(lhsExpr))
1457        return LHS;
1458
1459     return EvalAddr(C->getRHS());
1460  }
1461
1462  // For casts, we need to handle conversions from arrays to
1463  // pointer values, and pointer-to-pointer conversions.
1464  case Stmt::ImplicitCastExprClass:
1465  case Stmt::CStyleCastExprClass:
1466  case Stmt::CXXFunctionalCastExprClass: {
1467    Expr* SubExpr = cast<CastExpr>(E)->getSubExpr();
1468    QualType T = SubExpr->getType();
1469
1470    if (SubExpr->getType()->isPointerType() ||
1471        SubExpr->getType()->isBlockPointerType() ||
1472        SubExpr->getType()->isObjCQualifiedIdType())
1473      return EvalAddr(SubExpr);
1474    else if (T->isArrayType())
1475      return EvalVal(SubExpr);
1476    else
1477      return 0;
1478  }
1479
1480  // C++ casts.  For dynamic casts, static casts, and const casts, we
1481  // are always converting from a pointer-to-pointer, so we just blow
1482  // through the cast.  In the case the dynamic cast doesn't fail (and
1483  // return NULL), we take the conservative route and report cases
1484  // where we return the address of a stack variable.  For Reinterpre
1485  // FIXME: The comment about is wrong; we're not always converting
1486  // from pointer to pointer. I'm guessing that this code should also
1487  // handle references to objects.
1488  case Stmt::CXXStaticCastExprClass:
1489  case Stmt::CXXDynamicCastExprClass:
1490  case Stmt::CXXConstCastExprClass:
1491  case Stmt::CXXReinterpretCastExprClass: {
1492      Expr *S = cast<CXXNamedCastExpr>(E)->getSubExpr();
1493      if (S->getType()->isPointerType() || S->getType()->isBlockPointerType())
1494        return EvalAddr(S);
1495      else
1496        return NULL;
1497  }
1498
1499  // Everything else: we simply don't reason about them.
1500  default:
1501    return NULL;
1502  }
1503}
1504
1505
1506///  EvalVal - This function is complements EvalAddr in the mutual recursion.
1507///   See the comments for EvalAddr for more details.
1508static DeclRefExpr* EvalVal(Expr *E) {
1509
1510  // We should only be called for evaluating non-pointer expressions, or
1511  // expressions with a pointer type that are not used as references but instead
1512  // are l-values (e.g., DeclRefExpr with a pointer type).
1513
1514  // Our "symbolic interpreter" is just a dispatch off the currently
1515  // viewed AST node.  We then recursively traverse the AST by calling
1516  // EvalAddr and EvalVal appropriately.
1517  switch (E->getStmtClass()) {
1518  case Stmt::DeclRefExprClass: {
1519    // DeclRefExpr: the base case.  When we hit a DeclRefExpr we are looking
1520    //  at code that refers to a variable's name.  We check if it has local
1521    //  storage within the function, and if so, return the expression.
1522    DeclRefExpr *DR = cast<DeclRefExpr>(E);
1523
1524    if (VarDecl *V = dyn_cast<VarDecl>(DR->getDecl()))
1525      if (V->hasLocalStorage() && !V->getType()->isReferenceType()) return DR;
1526
1527    return NULL;
1528  }
1529
1530  case Stmt::ParenExprClass:
1531    // Ignore parentheses.
1532    return EvalVal(cast<ParenExpr>(E)->getSubExpr());
1533
1534  case Stmt::UnaryOperatorClass: {
1535    // The only unary operator that make sense to handle here
1536    // is Deref.  All others don't resolve to a "name."  This includes
1537    // handling all sorts of rvalues passed to a unary operator.
1538    UnaryOperator *U = cast<UnaryOperator>(E);
1539
1540    if (U->getOpcode() == UnaryOperator::Deref)
1541      return EvalAddr(U->getSubExpr());
1542
1543    return NULL;
1544  }
1545
1546  case Stmt::ArraySubscriptExprClass: {
1547    // Array subscripts are potential references to data on the stack.  We
1548    // retrieve the DeclRefExpr* for the array variable if it indeed
1549    // has local storage.
1550    return EvalAddr(cast<ArraySubscriptExpr>(E)->getBase());
1551  }
1552
1553  case Stmt::ConditionalOperatorClass: {
1554    // For conditional operators we need to see if either the LHS or RHS are
1555    // non-NULL DeclRefExpr's.  If one is non-NULL, we return it.
1556    ConditionalOperator *C = cast<ConditionalOperator>(E);
1557
1558    // Handle the GNU extension for missing LHS.
1559    if (Expr *lhsExpr = C->getLHS())
1560      if (DeclRefExpr *LHS = EvalVal(lhsExpr))
1561        return LHS;
1562
1563    return EvalVal(C->getRHS());
1564  }
1565
1566  // Accesses to members are potential references to data on the stack.
1567  case Stmt::MemberExprClass: {
1568    MemberExpr *M = cast<MemberExpr>(E);
1569
1570    // Check for indirect access.  We only want direct field accesses.
1571    if (!M->isArrow())
1572      return EvalVal(M->getBase());
1573    else
1574      return NULL;
1575  }
1576
1577  // Everything else: we simply don't reason about them.
1578  default:
1579    return NULL;
1580  }
1581}
1582
1583//===--- CHECK: Floating-Point comparisons (-Wfloat-equal) ---------------===//
1584
1585/// Check for comparisons of floating point operands using != and ==.
1586/// Issue a warning if these are no self-comparisons, as they are not likely
1587/// to do what the programmer intended.
1588void Sema::CheckFloatComparison(SourceLocation loc, Expr* lex, Expr *rex) {
1589  bool EmitWarning = true;
1590
1591  Expr* LeftExprSansParen = lex->IgnoreParens();
1592  Expr* RightExprSansParen = rex->IgnoreParens();
1593
1594  // Special case: check for x == x (which is OK).
1595  // Do not emit warnings for such cases.
1596  if (DeclRefExpr* DRL = dyn_cast<DeclRefExpr>(LeftExprSansParen))
1597    if (DeclRefExpr* DRR = dyn_cast<DeclRefExpr>(RightExprSansParen))
1598      if (DRL->getDecl() == DRR->getDecl())
1599        EmitWarning = false;
1600
1601
1602  // Special case: check for comparisons against literals that can be exactly
1603  //  represented by APFloat.  In such cases, do not emit a warning.  This
1604  //  is a heuristic: often comparison against such literals are used to
1605  //  detect if a value in a variable has not changed.  This clearly can
1606  //  lead to false negatives.
1607  if (EmitWarning) {
1608    if (FloatingLiteral* FLL = dyn_cast<FloatingLiteral>(LeftExprSansParen)) {
1609      if (FLL->isExact())
1610        EmitWarning = false;
1611    } else
1612      if (FloatingLiteral* FLR = dyn_cast<FloatingLiteral>(RightExprSansParen)){
1613        if (FLR->isExact())
1614          EmitWarning = false;
1615    }
1616  }
1617
1618  // Check for comparisons with builtin types.
1619  if (EmitWarning)
1620    if (CallExpr* CL = dyn_cast<CallExpr>(LeftExprSansParen))
1621      if (CL->isBuiltinCall(Context))
1622        EmitWarning = false;
1623
1624  if (EmitWarning)
1625    if (CallExpr* CR = dyn_cast<CallExpr>(RightExprSansParen))
1626      if (CR->isBuiltinCall(Context))
1627        EmitWarning = false;
1628
1629  // Emit the diagnostic.
1630  if (EmitWarning)
1631    Diag(loc, diag::warn_floatingpoint_eq)
1632      << lex->getSourceRange() << rex->getSourceRange();
1633}
1634
1635//===--- CHECK: Integer mixed-sign comparisons (-Wsign-compare) --------===//
1636//===--- CHECK: Lossy implicit conversions (-Wconversion) --------------===//
1637
1638namespace {
1639
1640/// Structure recording the 'active' range of an integer-valued
1641/// expression.
1642struct IntRange {
1643  /// The number of bits active in the int.
1644  unsigned Width;
1645
1646  /// True if the int is known not to have negative values.
1647  bool NonNegative;
1648
1649  IntRange() {}
1650  IntRange(unsigned Width, bool NonNegative)
1651    : Width(Width), NonNegative(NonNegative)
1652  {}
1653
1654  // Returns the range of the bool type.
1655  static IntRange forBoolType() {
1656    return IntRange(1, true);
1657  }
1658
1659  // Returns the range of an integral type.
1660  static IntRange forType(ASTContext &C, QualType T) {
1661    return forCanonicalType(C, T->getCanonicalTypeInternal().getTypePtr());
1662  }
1663
1664  // Returns the range of an integeral type based on its canonical
1665  // representation.
1666  static IntRange forCanonicalType(ASTContext &C, const Type *T) {
1667    assert(T->isCanonicalUnqualified());
1668
1669    if (const VectorType *VT = dyn_cast<VectorType>(T))
1670      T = VT->getElementType().getTypePtr();
1671    if (const ComplexType *CT = dyn_cast<ComplexType>(T))
1672      T = CT->getElementType().getTypePtr();
1673    if (const EnumType *ET = dyn_cast<EnumType>(T))
1674      T = ET->getDecl()->getIntegerType().getTypePtr();
1675
1676    const BuiltinType *BT = cast<BuiltinType>(T);
1677    assert(BT->isInteger());
1678
1679    return IntRange(C.getIntWidth(QualType(T, 0)), BT->isUnsignedInteger());
1680  }
1681
1682  // Returns the supremum of two ranges: i.e. their conservative merge.
1683  static IntRange join(const IntRange &L, const IntRange &R) {
1684    return IntRange(std::max(L.Width, R.Width),
1685                    L.NonNegative && R.NonNegative);
1686  }
1687
1688  // Returns the infinum of two ranges: i.e. their aggressive merge.
1689  static IntRange meet(const IntRange &L, const IntRange &R) {
1690    return IntRange(std::min(L.Width, R.Width),
1691                    L.NonNegative || R.NonNegative);
1692  }
1693};
1694
1695IntRange GetValueRange(ASTContext &C, llvm::APSInt &value, unsigned MaxWidth) {
1696  if (value.isSigned() && value.isNegative())
1697    return IntRange(value.getMinSignedBits(), false);
1698
1699  if (value.getBitWidth() > MaxWidth)
1700    value.trunc(MaxWidth);
1701
1702  // isNonNegative() just checks the sign bit without considering
1703  // signedness.
1704  return IntRange(value.getActiveBits(), true);
1705}
1706
1707IntRange GetValueRange(ASTContext &C, APValue &result, QualType Ty,
1708                       unsigned MaxWidth) {
1709  if (result.isInt())
1710    return GetValueRange(C, result.getInt(), MaxWidth);
1711
1712  if (result.isVector()) {
1713    IntRange R = GetValueRange(C, result.getVectorElt(0), Ty, MaxWidth);
1714    for (unsigned i = 1, e = result.getVectorLength(); i != e; ++i) {
1715      IntRange El = GetValueRange(C, result.getVectorElt(i), Ty, MaxWidth);
1716      R = IntRange::join(R, El);
1717    }
1718    return R;
1719  }
1720
1721  if (result.isComplexInt()) {
1722    IntRange R = GetValueRange(C, result.getComplexIntReal(), MaxWidth);
1723    IntRange I = GetValueRange(C, result.getComplexIntImag(), MaxWidth);
1724    return IntRange::join(R, I);
1725  }
1726
1727  // This can happen with lossless casts to intptr_t of "based" lvalues.
1728  // Assume it might use arbitrary bits.
1729  // FIXME: The only reason we need to pass the type in here is to get
1730  // the sign right on this one case.  It would be nice if APValue
1731  // preserved this.
1732  assert(result.isLValue());
1733  return IntRange(MaxWidth, Ty->isUnsignedIntegerType());
1734}
1735
1736/// Pseudo-evaluate the given integer expression, estimating the
1737/// range of values it might take.
1738///
1739/// \param MaxWidth - the width to which the value will be truncated
1740IntRange GetExprRange(ASTContext &C, Expr *E, unsigned MaxWidth) {
1741  E = E->IgnoreParens();
1742
1743  // Try a full evaluation first.
1744  Expr::EvalResult result;
1745  if (E->Evaluate(result, C))
1746    return GetValueRange(C, result.Val, E->getType(), MaxWidth);
1747
1748  // I think we only want to look through implicit casts here; if the
1749  // user has an explicit widening cast, we should treat the value as
1750  // being of the new, wider type.
1751  if (ImplicitCastExpr *CE = dyn_cast<ImplicitCastExpr>(E)) {
1752    if (CE->getCastKind() == CastExpr::CK_NoOp)
1753      return GetExprRange(C, CE->getSubExpr(), MaxWidth);
1754
1755    IntRange OutputTypeRange = IntRange::forType(C, CE->getType());
1756
1757    bool isIntegerCast = (CE->getCastKind() == CastExpr::CK_IntegralCast);
1758    if (!isIntegerCast && CE->getCastKind() == CastExpr::CK_Unknown)
1759      isIntegerCast = CE->getSubExpr()->getType()->isIntegerType();
1760
1761    // Assume that non-integer casts can span the full range of the type.
1762    if (!isIntegerCast)
1763      return OutputTypeRange;
1764
1765    IntRange SubRange
1766      = GetExprRange(C, CE->getSubExpr(),
1767                     std::min(MaxWidth, OutputTypeRange.Width));
1768
1769    // Bail out if the subexpr's range is as wide as the cast type.
1770    if (SubRange.Width >= OutputTypeRange.Width)
1771      return OutputTypeRange;
1772
1773    // Otherwise, we take the smaller width, and we're non-negative if
1774    // either the output type or the subexpr is.
1775    return IntRange(SubRange.Width,
1776                    SubRange.NonNegative || OutputTypeRange.NonNegative);
1777  }
1778
1779  if (ConditionalOperator *CO = dyn_cast<ConditionalOperator>(E)) {
1780    // If we can fold the condition, just take that operand.
1781    bool CondResult;
1782    if (CO->getCond()->EvaluateAsBooleanCondition(CondResult, C))
1783      return GetExprRange(C, CondResult ? CO->getTrueExpr()
1784                                        : CO->getFalseExpr(),
1785                          MaxWidth);
1786
1787    // Otherwise, conservatively merge.
1788    IntRange L = GetExprRange(C, CO->getTrueExpr(), MaxWidth);
1789    IntRange R = GetExprRange(C, CO->getFalseExpr(), MaxWidth);
1790    return IntRange::join(L, R);
1791  }
1792
1793  if (BinaryOperator *BO = dyn_cast<BinaryOperator>(E)) {
1794    switch (BO->getOpcode()) {
1795
1796    // Boolean-valued operations are single-bit and positive.
1797    case BinaryOperator::LAnd:
1798    case BinaryOperator::LOr:
1799    case BinaryOperator::LT:
1800    case BinaryOperator::GT:
1801    case BinaryOperator::LE:
1802    case BinaryOperator::GE:
1803    case BinaryOperator::EQ:
1804    case BinaryOperator::NE:
1805      return IntRange::forBoolType();
1806
1807    // Operations with opaque sources are black-listed.
1808    case BinaryOperator::PtrMemD:
1809    case BinaryOperator::PtrMemI:
1810      return IntRange::forType(C, E->getType());
1811
1812    // Bitwise-and uses the *infinum* of the two source ranges.
1813    case BinaryOperator::And:
1814      return IntRange::meet(GetExprRange(C, BO->getLHS(), MaxWidth),
1815                            GetExprRange(C, BO->getRHS(), MaxWidth));
1816
1817    // Left shift gets black-listed based on a judgement call.
1818    case BinaryOperator::Shl:
1819      return IntRange::forType(C, E->getType());
1820
1821    // Right shift by a constant can narrow its left argument.
1822    case BinaryOperator::Shr: {
1823      IntRange L = GetExprRange(C, BO->getLHS(), MaxWidth);
1824
1825      // If the shift amount is a positive constant, drop the width by
1826      // that much.
1827      llvm::APSInt shift;
1828      if (BO->getRHS()->isIntegerConstantExpr(shift, C) &&
1829          shift.isNonNegative()) {
1830        unsigned zext = shift.getZExtValue();
1831        if (zext >= L.Width)
1832          L.Width = (L.NonNegative ? 0 : 1);
1833        else
1834          L.Width -= zext;
1835      }
1836
1837      return L;
1838    }
1839
1840    // Comma acts as its right operand.
1841    case BinaryOperator::Comma:
1842      return GetExprRange(C, BO->getRHS(), MaxWidth);
1843
1844    // Black-list pointer subtractions.
1845    case BinaryOperator::Sub:
1846      if (BO->getLHS()->getType()->isPointerType())
1847        return IntRange::forType(C, E->getType());
1848      // fallthrough
1849
1850    default:
1851      break;
1852    }
1853
1854    // Treat every other operator as if it were closed on the
1855    // narrowest type that encompasses both operands.
1856    IntRange L = GetExprRange(C, BO->getLHS(), MaxWidth);
1857    IntRange R = GetExprRange(C, BO->getRHS(), MaxWidth);
1858    return IntRange::join(L, R);
1859  }
1860
1861  if (UnaryOperator *UO = dyn_cast<UnaryOperator>(E)) {
1862    switch (UO->getOpcode()) {
1863    // Boolean-valued operations are white-listed.
1864    case UnaryOperator::LNot:
1865      return IntRange::forBoolType();
1866
1867    // Operations with opaque sources are black-listed.
1868    case UnaryOperator::Deref:
1869    case UnaryOperator::AddrOf: // should be impossible
1870    case UnaryOperator::OffsetOf:
1871      return IntRange::forType(C, E->getType());
1872
1873    default:
1874      return GetExprRange(C, UO->getSubExpr(), MaxWidth);
1875    }
1876  }
1877
1878  FieldDecl *BitField = E->getBitField();
1879  if (BitField) {
1880    llvm::APSInt BitWidthAP = BitField->getBitWidth()->EvaluateAsInt(C);
1881    unsigned BitWidth = BitWidthAP.getZExtValue();
1882
1883    return IntRange(BitWidth, BitField->getType()->isUnsignedIntegerType());
1884  }
1885
1886  return IntRange::forType(C, E->getType());
1887}
1888
1889/// Checks whether the given value, which currently has the given
1890/// source semantics, has the same value when coerced through the
1891/// target semantics.
1892bool IsSameFloatAfterCast(const llvm::APFloat &value,
1893                          const llvm::fltSemantics &Src,
1894                          const llvm::fltSemantics &Tgt) {
1895  llvm::APFloat truncated = value;
1896
1897  bool ignored;
1898  truncated.convert(Src, llvm::APFloat::rmNearestTiesToEven, &ignored);
1899  truncated.convert(Tgt, llvm::APFloat::rmNearestTiesToEven, &ignored);
1900
1901  return truncated.bitwiseIsEqual(value);
1902}
1903
1904/// Checks whether the given value, which currently has the given
1905/// source semantics, has the same value when coerced through the
1906/// target semantics.
1907///
1908/// The value might be a vector of floats (or a complex number).
1909bool IsSameFloatAfterCast(const APValue &value,
1910                          const llvm::fltSemantics &Src,
1911                          const llvm::fltSemantics &Tgt) {
1912  if (value.isFloat())
1913    return IsSameFloatAfterCast(value.getFloat(), Src, Tgt);
1914
1915  if (value.isVector()) {
1916    for (unsigned i = 0, e = value.getVectorLength(); i != e; ++i)
1917      if (!IsSameFloatAfterCast(value.getVectorElt(i), Src, Tgt))
1918        return false;
1919    return true;
1920  }
1921
1922  assert(value.isComplexFloat());
1923  return (IsSameFloatAfterCast(value.getComplexFloatReal(), Src, Tgt) &&
1924          IsSameFloatAfterCast(value.getComplexFloatImag(), Src, Tgt));
1925}
1926
1927} // end anonymous namespace
1928
1929/// \brief Implements -Wsign-compare.
1930///
1931/// \param lex the left-hand expression
1932/// \param rex the right-hand expression
1933/// \param OpLoc the location of the joining operator
1934/// \param Equality whether this is an "equality-like" join, which
1935///   suppresses the warning in some cases
1936void Sema::CheckSignCompare(Expr *lex, Expr *rex, SourceLocation OpLoc,
1937                            const PartialDiagnostic &PD, bool Equality) {
1938  // Don't warn if we're in an unevaluated context.
1939  if (ExprEvalContexts.back().Context == Unevaluated)
1940    return;
1941
1942  // If either expression is value-dependent, don't warn. We'll get another
1943  // chance at instantiation time.
1944  if (lex->isValueDependent() || rex->isValueDependent())
1945    return;
1946
1947  QualType lt = lex->getType(), rt = rex->getType();
1948
1949  // Only warn if both operands are integral.
1950  if (!lt->isIntegerType() || !rt->isIntegerType())
1951    return;
1952
1953  // In C, the width of a bitfield determines its type, and the
1954  // declared type only contributes the signedness.  This duplicates
1955  // the work that will later be done by UsualUnaryConversions.
1956  // Eventually, this check will be reorganized in a way that avoids
1957  // this duplication.
1958  if (!getLangOptions().CPlusPlus) {
1959    QualType tmp;
1960    tmp = Context.isPromotableBitField(lex);
1961    if (!tmp.isNull()) lt = tmp;
1962    tmp = Context.isPromotableBitField(rex);
1963    if (!tmp.isNull()) rt = tmp;
1964  }
1965
1966  // The rule is that the signed operand becomes unsigned, so isolate the
1967  // signed operand.
1968  Expr *signedOperand = lex, *unsignedOperand = rex;
1969  QualType signedType = lt, unsignedType = rt;
1970  if (lt->isSignedIntegerType()) {
1971    if (rt->isSignedIntegerType()) return;
1972  } else {
1973    if (!rt->isSignedIntegerType()) return;
1974    std::swap(signedOperand, unsignedOperand);
1975    std::swap(signedType, unsignedType);
1976  }
1977
1978  unsigned unsignedWidth = Context.getIntWidth(unsignedType);
1979  unsigned signedWidth = Context.getIntWidth(signedType);
1980
1981  // If the unsigned type is strictly smaller than the signed type,
1982  // then (1) the result type will be signed and (2) the unsigned
1983  // value will fit fully within the signed type, and thus the result
1984  // of the comparison will be exact.
1985  if (signedWidth > unsignedWidth)
1986    return;
1987
1988  // Otherwise, calculate the effective ranges.
1989  IntRange signedRange = GetExprRange(Context, signedOperand, signedWidth);
1990  IntRange unsignedRange = GetExprRange(Context, unsignedOperand, unsignedWidth);
1991
1992  // We should never be unable to prove that the unsigned operand is
1993  // non-negative.
1994  assert(unsignedRange.NonNegative && "unsigned range includes negative?");
1995
1996  // If the signed operand is non-negative, then the signed->unsigned
1997  // conversion won't change it.
1998  if (signedRange.NonNegative)
1999    return;
2000
2001  // For (in)equality comparisons, if the unsigned operand is a
2002  // constant which cannot collide with a overflowed signed operand,
2003  // then reinterpreting the signed operand as unsigned will not
2004  // change the result of the comparison.
2005  if (Equality && unsignedRange.Width < unsignedWidth)
2006    return;
2007
2008  Diag(OpLoc, PD)
2009    << lt << rt << lex->getSourceRange() << rex->getSourceRange();
2010}
2011
2012/// Diagnose an implicit cast;  purely a helper for CheckImplicitConversion.
2013static void DiagnoseImpCast(Sema &S, Expr *E, QualType T, unsigned diag) {
2014  S.Diag(E->getExprLoc(), diag) << E->getType() << T << E->getSourceRange();
2015}
2016
2017/// Implements -Wconversion.
2018void Sema::CheckImplicitConversion(Expr *E, QualType T) {
2019  // Don't diagnose in unevaluated contexts.
2020  if (ExprEvalContexts.back().Context == Sema::Unevaluated)
2021    return;
2022
2023  // Don't diagnose for value-dependent expressions.
2024  if (E->isValueDependent())
2025    return;
2026
2027  const Type *Source = Context.getCanonicalType(E->getType()).getTypePtr();
2028  const Type *Target = Context.getCanonicalType(T).getTypePtr();
2029
2030  // Never diagnose implicit casts to bool.
2031  if (Target->isSpecificBuiltinType(BuiltinType::Bool))
2032    return;
2033
2034  // Strip vector types.
2035  if (isa<VectorType>(Source)) {
2036    if (!isa<VectorType>(Target))
2037      return DiagnoseImpCast(*this, E, T, diag::warn_impcast_vector_scalar);
2038
2039    Source = cast<VectorType>(Source)->getElementType().getTypePtr();
2040    Target = cast<VectorType>(Target)->getElementType().getTypePtr();
2041  }
2042
2043  // Strip complex types.
2044  if (isa<ComplexType>(Source)) {
2045    if (!isa<ComplexType>(Target))
2046      return DiagnoseImpCast(*this, E, T, diag::warn_impcast_complex_scalar);
2047
2048    Source = cast<ComplexType>(Source)->getElementType().getTypePtr();
2049    Target = cast<ComplexType>(Target)->getElementType().getTypePtr();
2050  }
2051
2052  const BuiltinType *SourceBT = dyn_cast<BuiltinType>(Source);
2053  const BuiltinType *TargetBT = dyn_cast<BuiltinType>(Target);
2054
2055  // If the source is floating point...
2056  if (SourceBT && SourceBT->isFloatingPoint()) {
2057    // ...and the target is floating point...
2058    if (TargetBT && TargetBT->isFloatingPoint()) {
2059      // ...then warn if we're dropping FP rank.
2060
2061      // Builtin FP kinds are ordered by increasing FP rank.
2062      if (SourceBT->getKind() > TargetBT->getKind()) {
2063        // Don't warn about float constants that are precisely
2064        // representable in the target type.
2065        Expr::EvalResult result;
2066        if (E->Evaluate(result, Context)) {
2067          // Value might be a float, a float vector, or a float complex.
2068          if (IsSameFloatAfterCast(result.Val,
2069                     Context.getFloatTypeSemantics(QualType(TargetBT, 0)),
2070                     Context.getFloatTypeSemantics(QualType(SourceBT, 0))))
2071            return;
2072        }
2073
2074        DiagnoseImpCast(*this, E, T, diag::warn_impcast_float_precision);
2075      }
2076      return;
2077    }
2078
2079    // If the target is integral, always warn.
2080    if ((TargetBT && TargetBT->isInteger()))
2081      // TODO: don't warn for integer values?
2082      return DiagnoseImpCast(*this, E, T, diag::warn_impcast_float_integer);
2083
2084    return;
2085  }
2086
2087  if (!Source->isIntegerType() || !Target->isIntegerType())
2088    return;
2089
2090  IntRange SourceRange = GetExprRange(Context, E, Context.getIntWidth(E->getType()));
2091  IntRange TargetRange = IntRange::forCanonicalType(Context, Target);
2092
2093  // FIXME: also signed<->unsigned?
2094
2095  if (SourceRange.Width > TargetRange.Width) {
2096    // People want to build with -Wshorten-64-to-32 and not -Wconversion
2097    // and by god we'll let them.
2098    if (SourceRange.Width == 64 && TargetRange.Width == 32)
2099      return DiagnoseImpCast(*this, E, T, diag::warn_impcast_integer_64_32);
2100    return DiagnoseImpCast(*this, E, T, diag::warn_impcast_integer_precision);
2101  }
2102
2103  return;
2104}
2105
2106// MarkLive - Mark all the blocks reachable from e as live.  Returns the total
2107// number of blocks just marked live.
2108static unsigned MarkLive(CFGBlock *e, llvm::BitVector &live) {
2109  unsigned count = 0;
2110  std::queue<CFGBlock*> workq;
2111  // Prep work queue
2112  live.set(e->getBlockID());
2113  ++count;
2114  workq.push(e);
2115  // Solve
2116  while (!workq.empty()) {
2117    CFGBlock *item = workq.front();
2118    workq.pop();
2119    for (CFGBlock::succ_iterator I=item->succ_begin(),
2120           E=item->succ_end();
2121         I != E;
2122         ++I) {
2123      if ((*I) && !live[(*I)->getBlockID()]) {
2124        live.set((*I)->getBlockID());
2125        ++count;
2126        workq.push(*I);
2127      }
2128    }
2129  }
2130  return count;
2131}
2132
2133static SourceLocation GetUnreachableLoc(CFGBlock &b, SourceRange &R1,
2134                                        SourceRange &R2) {
2135  Stmt *S;
2136  unsigned sn = 0;
2137  R1 = R2 = SourceRange();
2138
2139  top:
2140  if (sn < b.size())
2141    S = b[sn].getStmt();
2142  else if (b.getTerminator())
2143    S = b.getTerminator();
2144  else
2145    return SourceLocation();
2146
2147  switch (S->getStmtClass()) {
2148  case Expr::BinaryOperatorClass: {
2149    BinaryOperator *BO = cast<BinaryOperator>(S);
2150    if (BO->getOpcode() == BinaryOperator::Comma) {
2151      if (sn+1 < b.size())
2152        return b[sn+1].getStmt()->getLocStart();
2153      CFGBlock *n = &b;
2154      while (1) {
2155        if (n->getTerminator())
2156          return n->getTerminator()->getLocStart();
2157        if (n->succ_size() != 1)
2158          return SourceLocation();
2159        n = n[0].succ_begin()[0];
2160        if (n->pred_size() != 1)
2161          return SourceLocation();
2162        if (!n->empty())
2163          return n[0][0].getStmt()->getLocStart();
2164      }
2165    }
2166    R1 = BO->getLHS()->getSourceRange();
2167    R2 = BO->getRHS()->getSourceRange();
2168    return BO->getOperatorLoc();
2169  }
2170  case Expr::UnaryOperatorClass: {
2171    const UnaryOperator *UO = cast<UnaryOperator>(S);
2172    R1 = UO->getSubExpr()->getSourceRange();
2173    return UO->getOperatorLoc();
2174  }
2175  case Expr::CompoundAssignOperatorClass: {
2176    const CompoundAssignOperator *CAO = cast<CompoundAssignOperator>(S);
2177    R1 = CAO->getLHS()->getSourceRange();
2178    R2 = CAO->getRHS()->getSourceRange();
2179    return CAO->getOperatorLoc();
2180  }
2181  case Expr::ConditionalOperatorClass: {
2182    const ConditionalOperator *CO = cast<ConditionalOperator>(S);
2183    return CO->getQuestionLoc();
2184  }
2185  case Expr::MemberExprClass: {
2186    const MemberExpr *ME = cast<MemberExpr>(S);
2187    R1 = ME->getSourceRange();
2188    return ME->getMemberLoc();
2189  }
2190  case Expr::ArraySubscriptExprClass: {
2191    const ArraySubscriptExpr *ASE = cast<ArraySubscriptExpr>(S);
2192    R1 = ASE->getLHS()->getSourceRange();
2193    R2 = ASE->getRHS()->getSourceRange();
2194    return ASE->getRBracketLoc();
2195  }
2196  case Expr::CStyleCastExprClass: {
2197    const CStyleCastExpr *CSC = cast<CStyleCastExpr>(S);
2198    R1 = CSC->getSubExpr()->getSourceRange();
2199    return CSC->getLParenLoc();
2200  }
2201  case Expr::CXXFunctionalCastExprClass: {
2202    const CXXFunctionalCastExpr *CE = cast <CXXFunctionalCastExpr>(S);
2203    R1 = CE->getSubExpr()->getSourceRange();
2204    return CE->getTypeBeginLoc();
2205  }
2206  case Expr::ImplicitCastExprClass:
2207    ++sn;
2208    goto top;
2209  case Stmt::CXXTryStmtClass: {
2210    return cast<CXXTryStmt>(S)->getHandler(0)->getCatchLoc();
2211  }
2212  default: ;
2213  }
2214  R1 = S->getSourceRange();
2215  return S->getLocStart();
2216}
2217
2218static SourceLocation MarkLiveTop(CFGBlock *e, llvm::BitVector &live,
2219                               SourceManager &SM) {
2220  std::queue<CFGBlock*> workq;
2221  // Prep work queue
2222  workq.push(e);
2223  SourceRange R1, R2;
2224  SourceLocation top = GetUnreachableLoc(*e, R1, R2);
2225  bool FromMainFile = false;
2226  bool FromSystemHeader = false;
2227  bool TopValid = false;
2228  if (top.isValid()) {
2229    FromMainFile = SM.isFromMainFile(top);
2230    FromSystemHeader = SM.isInSystemHeader(top);
2231    TopValid = true;
2232  }
2233  // Solve
2234  while (!workq.empty()) {
2235    CFGBlock *item = workq.front();
2236    workq.pop();
2237    SourceLocation c = GetUnreachableLoc(*item, R1, R2);
2238    if (c.isValid()
2239        && (!TopValid
2240            || (SM.isFromMainFile(c) && !FromMainFile)
2241            || (FromSystemHeader && !SM.isInSystemHeader(c))
2242            || SM.isBeforeInTranslationUnit(c, top))) {
2243      top = c;
2244      FromMainFile = SM.isFromMainFile(top);
2245      FromSystemHeader = SM.isInSystemHeader(top);
2246    }
2247    live.set(item->getBlockID());
2248    for (CFGBlock::succ_iterator I=item->succ_begin(),
2249           E=item->succ_end();
2250         I != E;
2251         ++I) {
2252      if ((*I) && !live[(*I)->getBlockID()]) {
2253        live.set((*I)->getBlockID());
2254        workq.push(*I);
2255      }
2256    }
2257  }
2258  return top;
2259}
2260
2261static int LineCmp(const void *p1, const void *p2) {
2262  SourceLocation *Line1 = (SourceLocation *)p1;
2263  SourceLocation *Line2 = (SourceLocation *)p2;
2264  return !(*Line1 < *Line2);
2265}
2266
2267namespace {
2268  struct ErrLoc {
2269    SourceLocation Loc;
2270    SourceRange R1;
2271    SourceRange R2;
2272    ErrLoc(SourceLocation l, SourceRange r1, SourceRange r2)
2273      : Loc(l), R1(r1), R2(r2) { }
2274  };
2275}
2276
2277/// CheckUnreachable - Check for unreachable code.
2278void Sema::CheckUnreachable(AnalysisContext &AC) {
2279  unsigned count;
2280  // We avoid checking when there are errors, as the CFG won't faithfully match
2281  // the user's code.
2282  if (getDiagnostics().hasErrorOccurred())
2283    return;
2284  if (Diags.getDiagnosticLevel(diag::warn_unreachable) == Diagnostic::Ignored)
2285    return;
2286
2287  CFG *cfg = AC.getCFG();
2288  if (cfg == 0)
2289    return;
2290
2291  llvm::BitVector live(cfg->getNumBlockIDs());
2292  // Mark all live things first.
2293  count = MarkLive(&cfg->getEntry(), live);
2294
2295  if (count == cfg->getNumBlockIDs())
2296    // If there are no dead blocks, we're done.
2297    return;
2298
2299  SourceRange R1, R2;
2300
2301  llvm::SmallVector<ErrLoc, 24> lines;
2302  bool AddEHEdges = AC.getAddEHEdges();
2303  // First, give warnings for blocks with no predecessors, as they
2304  // can't be part of a loop.
2305  for (CFG::iterator I = cfg->begin(), E = cfg->end(); I != E; ++I) {
2306    CFGBlock &b = **I;
2307    if (!live[b.getBlockID()]) {
2308      if (b.pred_begin() == b.pred_end()) {
2309        if (!AddEHEdges && b.getTerminator()
2310            && isa<CXXTryStmt>(b.getTerminator())) {
2311          // When not adding EH edges from calls, catch clauses
2312          // can otherwise seem dead.  Avoid noting them as dead.
2313          count += MarkLive(&b, live);
2314          continue;
2315        }
2316        SourceLocation c = GetUnreachableLoc(b, R1, R2);
2317        if (!c.isValid()) {
2318          // Blocks without a location can't produce a warning, so don't mark
2319          // reachable blocks from here as live.
2320          live.set(b.getBlockID());
2321          ++count;
2322          continue;
2323        }
2324        lines.push_back(ErrLoc(c, R1, R2));
2325        // Avoid excessive errors by marking everything reachable from here
2326        count += MarkLive(&b, live);
2327      }
2328    }
2329  }
2330
2331  if (count < cfg->getNumBlockIDs()) {
2332    // And then give warnings for the tops of loops.
2333    for (CFG::iterator I = cfg->begin(), E = cfg->end(); I != E; ++I) {
2334      CFGBlock &b = **I;
2335      if (!live[b.getBlockID()])
2336        // Avoid excessive errors by marking everything reachable from here
2337        lines.push_back(ErrLoc(MarkLiveTop(&b, live,
2338                                           Context.getSourceManager()),
2339                               SourceRange(), SourceRange()));
2340    }
2341  }
2342
2343  llvm::array_pod_sort(lines.begin(), lines.end(), LineCmp);
2344  for (llvm::SmallVector<ErrLoc, 24>::iterator I = lines.begin(),
2345         E = lines.end();
2346       I != E;
2347       ++I)
2348    if (I->Loc.isValid())
2349      Diag(I->Loc, diag::warn_unreachable) << I->R1 << I->R2;
2350}
2351
2352/// CheckFallThrough - Check that we don't fall off the end of a
2353/// Statement that should return a value.
2354///
2355/// \returns AlwaysFallThrough iff we always fall off the end of the statement,
2356/// MaybeFallThrough iff we might or might not fall off the end,
2357/// NeverFallThroughOrReturn iff we never fall off the end of the statement or
2358/// return.  We assume NeverFallThrough iff we never fall off the end of the
2359/// statement but we may return.  We assume that functions not marked noreturn
2360/// will return.
2361Sema::ControlFlowKind Sema::CheckFallThrough(AnalysisContext &AC) {
2362  CFG *cfg = AC.getCFG();
2363  if (cfg == 0)
2364    // FIXME: This should be NeverFallThrough
2365    return NeverFallThroughOrReturn;
2366
2367  // The CFG leaves in dead things, and we don't want the dead code paths to
2368  // confuse us, so we mark all live things first.
2369  std::queue<CFGBlock*> workq;
2370  llvm::BitVector live(cfg->getNumBlockIDs());
2371  unsigned count = MarkLive(&cfg->getEntry(), live);
2372
2373  bool AddEHEdges = AC.getAddEHEdges();
2374  if (!AddEHEdges && count != cfg->getNumBlockIDs())
2375    // When there are things remaining dead, and we didn't add EH edges
2376    // from CallExprs to the catch clauses, we have to go back and
2377    // mark them as live.
2378    for (CFG::iterator I = cfg->begin(), E = cfg->end(); I != E; ++I) {
2379      CFGBlock &b = **I;
2380      if (!live[b.getBlockID()]) {
2381        if (b.pred_begin() == b.pred_end()) {
2382          if (b.getTerminator() && isa<CXXTryStmt>(b.getTerminator()))
2383            // When not adding EH edges from calls, catch clauses
2384            // can otherwise seem dead.  Avoid noting them as dead.
2385            count += MarkLive(&b, live);
2386          continue;
2387        }
2388      }
2389    }
2390
2391  // Now we know what is live, we check the live precessors of the exit block
2392  // and look for fall through paths, being careful to ignore normal returns,
2393  // and exceptional paths.
2394  bool HasLiveReturn = false;
2395  bool HasFakeEdge = false;
2396  bool HasPlainEdge = false;
2397  bool HasAbnormalEdge = false;
2398  for (CFGBlock::pred_iterator I=cfg->getExit().pred_begin(),
2399         E = cfg->getExit().pred_end();
2400       I != E;
2401       ++I) {
2402    CFGBlock& B = **I;
2403    if (!live[B.getBlockID()])
2404      continue;
2405    if (B.size() == 0) {
2406      if (B.getTerminator() && isa<CXXTryStmt>(B.getTerminator())) {
2407        HasAbnormalEdge = true;
2408        continue;
2409      }
2410
2411      // A labeled empty statement, or the entry block...
2412      HasPlainEdge = true;
2413      continue;
2414    }
2415    Stmt *S = B[B.size()-1];
2416    if (isa<ReturnStmt>(S)) {
2417      HasLiveReturn = true;
2418      continue;
2419    }
2420    if (isa<ObjCAtThrowStmt>(S)) {
2421      HasFakeEdge = true;
2422      continue;
2423    }
2424    if (isa<CXXThrowExpr>(S)) {
2425      HasFakeEdge = true;
2426      continue;
2427    }
2428    if (const AsmStmt *AS = dyn_cast<AsmStmt>(S)) {
2429      if (AS->isMSAsm()) {
2430        HasFakeEdge = true;
2431        HasLiveReturn = true;
2432        continue;
2433      }
2434    }
2435    if (isa<CXXTryStmt>(S)) {
2436      HasAbnormalEdge = true;
2437      continue;
2438    }
2439
2440    bool NoReturnEdge = false;
2441    if (CallExpr *C = dyn_cast<CallExpr>(S)) {
2442      if (B.succ_begin()[0] != &cfg->getExit()) {
2443        HasAbnormalEdge = true;
2444        continue;
2445      }
2446      Expr *CEE = C->getCallee()->IgnoreParenCasts();
2447      if (CEE->getType().getNoReturnAttr()) {
2448        NoReturnEdge = true;
2449        HasFakeEdge = true;
2450      } else if (DeclRefExpr *DRE = dyn_cast<DeclRefExpr>(CEE)) {
2451        ValueDecl *VD = DRE->getDecl();
2452        if (VD->hasAttr<NoReturnAttr>()) {
2453          NoReturnEdge = true;
2454          HasFakeEdge = true;
2455        }
2456      }
2457    }
2458    // FIXME: Add noreturn message sends.
2459    if (NoReturnEdge == false)
2460      HasPlainEdge = true;
2461  }
2462  if (!HasPlainEdge) {
2463    if (HasLiveReturn)
2464      return NeverFallThrough;
2465    return NeverFallThroughOrReturn;
2466  }
2467  if (HasAbnormalEdge || HasFakeEdge || HasLiveReturn)
2468    return MaybeFallThrough;
2469  // This says AlwaysFallThrough for calls to functions that are not marked
2470  // noreturn, that don't return.  If people would like this warning to be more
2471  // accurate, such functions should be marked as noreturn.
2472  return AlwaysFallThrough;
2473}
2474
2475/// CheckFallThroughForFunctionDef - Check that we don't fall off the end of a
2476/// function that should return a value.  Check that we don't fall off the end
2477/// of a noreturn function.  We assume that functions and blocks not marked
2478/// noreturn will return.
2479void Sema::CheckFallThroughForFunctionDef(Decl *D, Stmt *Body,
2480                                          AnalysisContext &AC) {
2481  // FIXME: Would be nice if we had a better way to control cascading errors,
2482  // but for now, avoid them.  The problem is that when Parse sees:
2483  //   int foo() { return a; }
2484  // The return is eaten and the Sema code sees just:
2485  //   int foo() { }
2486  // which this code would then warn about.
2487  if (getDiagnostics().hasErrorOccurred())
2488    return;
2489
2490  bool ReturnsVoid = false;
2491  bool HasNoReturn = false;
2492  if (FunctionDecl *FD = dyn_cast<FunctionDecl>(D)) {
2493    // For function templates, class templates and member function templates
2494    // we'll do the analysis at instantiation time.
2495    if (FD->isDependentContext())
2496      return;
2497
2498    if (FD->getResultType()->isVoidType())
2499      ReturnsVoid = true;
2500    if (FD->hasAttr<NoReturnAttr>() ||
2501        FD->getType()->getAs<FunctionType>()->getNoReturnAttr())
2502      HasNoReturn = true;
2503  } else if (ObjCMethodDecl *MD = dyn_cast<ObjCMethodDecl>(D)) {
2504    if (MD->getResultType()->isVoidType())
2505      ReturnsVoid = true;
2506    if (MD->hasAttr<NoReturnAttr>())
2507      HasNoReturn = true;
2508  }
2509
2510  // Short circuit for compilation speed.
2511  if ((Diags.getDiagnosticLevel(diag::warn_maybe_falloff_nonvoid_function)
2512       == Diagnostic::Ignored || ReturnsVoid)
2513      && (Diags.getDiagnosticLevel(diag::warn_noreturn_function_has_return_expr)
2514          == Diagnostic::Ignored || !HasNoReturn)
2515      && (Diags.getDiagnosticLevel(diag::warn_suggest_noreturn_block)
2516          == Diagnostic::Ignored || !ReturnsVoid))
2517    return;
2518  // FIXME: Function try block
2519  if (CompoundStmt *Compound = dyn_cast<CompoundStmt>(Body)) {
2520    switch (CheckFallThrough(AC)) {
2521    case MaybeFallThrough:
2522      if (HasNoReturn)
2523        Diag(Compound->getRBracLoc(), diag::warn_falloff_noreturn_function);
2524      else if (!ReturnsVoid)
2525        Diag(Compound->getRBracLoc(),diag::warn_maybe_falloff_nonvoid_function);
2526      break;
2527    case AlwaysFallThrough:
2528      if (HasNoReturn)
2529        Diag(Compound->getRBracLoc(), diag::warn_falloff_noreturn_function);
2530      else if (!ReturnsVoid)
2531        Diag(Compound->getRBracLoc(), diag::warn_falloff_nonvoid_function);
2532      break;
2533    case NeverFallThroughOrReturn:
2534      if (ReturnsVoid && !HasNoReturn)
2535        Diag(Compound->getLBracLoc(), diag::warn_suggest_noreturn_function);
2536      break;
2537    case NeverFallThrough:
2538      break;
2539    }
2540  }
2541}
2542
2543/// CheckFallThroughForBlock - Check that we don't fall off the end of a block
2544/// that should return a value.  Check that we don't fall off the end of a
2545/// noreturn block.  We assume that functions and blocks not marked noreturn
2546/// will return.
2547void Sema::CheckFallThroughForBlock(QualType BlockTy, Stmt *Body,
2548                                    AnalysisContext &AC) {
2549  // FIXME: Would be nice if we had a better way to control cascading errors,
2550  // but for now, avoid them.  The problem is that when Parse sees:
2551  //   int foo() { return a; }
2552  // The return is eaten and the Sema code sees just:
2553  //   int foo() { }
2554  // which this code would then warn about.
2555  if (getDiagnostics().hasErrorOccurred())
2556    return;
2557  bool ReturnsVoid = false;
2558  bool HasNoReturn = false;
2559  if (const FunctionType *FT =BlockTy->getPointeeType()->getAs<FunctionType>()){
2560    if (FT->getResultType()->isVoidType())
2561      ReturnsVoid = true;
2562    if (FT->getNoReturnAttr())
2563      HasNoReturn = true;
2564  }
2565
2566  // Short circuit for compilation speed.
2567  if (ReturnsVoid
2568      && !HasNoReturn
2569      && (Diags.getDiagnosticLevel(diag::warn_suggest_noreturn_block)
2570          == Diagnostic::Ignored || !ReturnsVoid))
2571    return;
2572  // FIXME: Funtion try block
2573  if (CompoundStmt *Compound = dyn_cast<CompoundStmt>(Body)) {
2574    switch (CheckFallThrough(AC)) {
2575    case MaybeFallThrough:
2576      if (HasNoReturn)
2577        Diag(Compound->getRBracLoc(), diag::err_noreturn_block_has_return_expr);
2578      else if (!ReturnsVoid)
2579        Diag(Compound->getRBracLoc(), diag::err_maybe_falloff_nonvoid_block);
2580      break;
2581    case AlwaysFallThrough:
2582      if (HasNoReturn)
2583        Diag(Compound->getRBracLoc(), diag::err_noreturn_block_has_return_expr);
2584      else if (!ReturnsVoid)
2585        Diag(Compound->getRBracLoc(), diag::err_falloff_nonvoid_block);
2586      break;
2587    case NeverFallThroughOrReturn:
2588      if (ReturnsVoid)
2589        Diag(Compound->getLBracLoc(), diag::warn_suggest_noreturn_block);
2590      break;
2591    case NeverFallThrough:
2592      break;
2593    }
2594  }
2595}
2596
2597/// CheckParmsForFunctionDef - Check that the parameters of the given
2598/// function are appropriate for the definition of a function. This
2599/// takes care of any checks that cannot be performed on the
2600/// declaration itself, e.g., that the types of each of the function
2601/// parameters are complete.
2602bool Sema::CheckParmsForFunctionDef(FunctionDecl *FD) {
2603  bool HasInvalidParm = false;
2604  for (unsigned p = 0, NumParams = FD->getNumParams(); p < NumParams; ++p) {
2605    ParmVarDecl *Param = FD->getParamDecl(p);
2606
2607    // C99 6.7.5.3p4: the parameters in a parameter type list in a
2608    // function declarator that is part of a function definition of
2609    // that function shall not have incomplete type.
2610    //
2611    // This is also C++ [dcl.fct]p6.
2612    if (!Param->isInvalidDecl() &&
2613        RequireCompleteType(Param->getLocation(), Param->getType(),
2614                               diag::err_typecheck_decl_incomplete_type)) {
2615      Param->setInvalidDecl();
2616      HasInvalidParm = true;
2617    }
2618
2619    // C99 6.9.1p5: If the declarator includes a parameter type list, the
2620    // declaration of each parameter shall include an identifier.
2621    if (Param->getIdentifier() == 0 &&
2622        !Param->isImplicit() &&
2623        !getLangOptions().CPlusPlus)
2624      Diag(Param->getLocation(), diag::err_parameter_name_omitted);
2625
2626    // C99 6.7.5.3p12:
2627    //   If the function declarator is not part of a definition of that
2628    //   function, parameters may have incomplete type and may use the [*]
2629    //   notation in their sequences of declarator specifiers to specify
2630    //   variable length array types.
2631    QualType PType = Param->getOriginalType();
2632    if (const ArrayType *AT = Context.getAsArrayType(PType)) {
2633      if (AT->getSizeModifier() == ArrayType::Star) {
2634        // FIXME: This diagnosic should point the the '[*]' if source-location
2635        // information is added for it.
2636        Diag(Param->getLocation(), diag::err_array_star_in_function_definition);
2637      }
2638    }
2639
2640    if (getLangOptions().CPlusPlus)
2641      if (const RecordType *RT = Param->getType()->getAs<RecordType>())
2642        FinalizeVarWithDestructor(Param, RT);
2643  }
2644
2645  return HasInvalidParm;
2646}
2647