SemaChecking.cpp revision 909a70d8a185f701c20bacc3c76f8f5979cbbc7c
15821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)//===--- SemaChecking.cpp - Extra Semantic Checking -----------------------===//
25821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)//
35821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)//                     The LLVM Compiler Infrastructure
45821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)//
55821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// This file is distributed under the University of Illinois Open Source
65821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)// License. See LICENSE.TXT for details.
75821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)//
85821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)//===----------------------------------------------------------------------===//
95821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)//
105821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)//  This file implements extra semantic analysis beyond what is enforced
115821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)//  by the C type system.
12f8ee788a64d60abd8f2d742a5fdedde054ecd910Torne (Richard Coles)//
131320f92c476a1ad9d19dba2a48c72b75566198e9Primiano Tucci//===----------------------------------------------------------------------===//
145821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)
155821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)#include "clang/Sema/Sema.h"
165821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)#include "clang/Sema/SemaInternal.h"
175821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)#include "clang/Sema/ScopeInfo.h"
18f8ee788a64d60abd8f2d742a5fdedde054ecd910Torne (Richard Coles)#include "clang/Analysis/Analyses/FormatString.h"
19f8ee788a64d60abd8f2d742a5fdedde054ecd910Torne (Richard Coles)#include "clang/AST/ASTContext.h"
20f8ee788a64d60abd8f2d742a5fdedde054ecd910Torne (Richard Coles)#include "clang/AST/CharUnits.h"
21f8ee788a64d60abd8f2d742a5fdedde054ecd910Torne (Richard Coles)#include "clang/AST/DeclCXX.h"
22f8ee788a64d60abd8f2d742a5fdedde054ecd910Torne (Richard Coles)#include "clang/AST/DeclObjC.h"
235821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)#include "clang/AST/ExprCXX.h"
245821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)#include "clang/AST/ExprObjC.h"
255821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)#include "clang/AST/DeclObjC.h"
265821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)#include "clang/AST/StmtCXX.h"
275821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)#include "clang/AST/StmtObjC.h"
285821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)#include "clang/Lex/Preprocessor.h"
295821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)#include "llvm/ADT/BitVector.h"
305821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)#include "llvm/ADT/STLExtras.h"
315821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)#include "llvm/Support/raw_ostream.h"
325821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)#include "clang/Basic/TargetBuiltins.h"
335821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)#include "clang/Basic/TargetInfo.h"
345821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)#include "clang/Basic/ConvertUTF.h"
355821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)#include <limits>
365821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)using namespace clang;
375821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)using namespace sema;
385821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)
395821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)SourceLocation Sema::getLocationOfStringLiteralByte(const StringLiteral *SL,
405821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)                                                    unsigned ByteNo) const {
415821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)  return SL->getLocationOfByte(ByteNo, PP.getSourceManager(),
425821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)                               PP.getLangOptions(), PP.getTargetInfo());
435821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)}
445821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)
455821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)
465821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)/// CheckablePrintfAttr - does a function call have a "printf" attribute
475821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)/// and arguments that merit checking?
485821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)bool Sema::CheckablePrintfAttr(const FormatAttr *Format, CallExpr *TheCall) {
495821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)  if (Format->getType() == "printf") return true;
505821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)  if (Format->getType() == "printf0") {
515821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)    // printf0 allows null "format" string; if so don't check format/args
525821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)    unsigned format_idx = Format->getFormatIdx() - 1;
535821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)    // Does the index refer to the implicit object argument?
545821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)    if (isa<CXXMemberCallExpr>(TheCall)) {
555821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)      if (format_idx == 0)
565821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)        return false;
575821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)      --format_idx;
585821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)    }
59a3f6a49ab37290eeeb8db0f41ec0f1cb74a68be7Torne (Richard Coles)    if (format_idx < TheCall->getNumArgs()) {
60a3f6a49ab37290eeeb8db0f41ec0f1cb74a68be7Torne (Richard Coles)      Expr *Format = TheCall->getArg(format_idx)->IgnoreParenCasts();
615821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)      if (!Format->isNullPointerConstant(Context,
625821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)                                         Expr::NPC_ValueDependentIsNull))
635821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)        return true;
645821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)    }
655821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)  }
665821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)  return false;
675821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)}
685821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)
695821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)/// Checks that a call expression's argument count is the desired number.
705821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)/// This is useful when doing custom type-checking.  Returns true on error.
715821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)static bool checkArgCount(Sema &S, CallExpr *call, unsigned desiredArgCount) {
72f8ee788a64d60abd8f2d742a5fdedde054ecd910Torne (Richard Coles)  unsigned argCount = call->getNumArgs();
73f8ee788a64d60abd8f2d742a5fdedde054ecd910Torne (Richard Coles)  if (argCount == desiredArgCount) return false;
745821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)
75f8ee788a64d60abd8f2d742a5fdedde054ecd910Torne (Richard Coles)  if (argCount < desiredArgCount)
76f8ee788a64d60abd8f2d742a5fdedde054ecd910Torne (Richard Coles)    return S.Diag(call->getLocEnd(), diag::err_typecheck_call_too_few_args)
775821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)        << 0 /*function call*/ << desiredArgCount << argCount
785821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)        << call->getSourceRange();
79a3f6a49ab37290eeeb8db0f41ec0f1cb74a68be7Torne (Richard Coles)
805821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)  // Highlight all the excess arguments.
815821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)  SourceRange range(call->getArg(desiredArgCount)->getLocStart(),
825821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)                    call->getArg(argCount - 1)->getLocEnd());
835821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)
845821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)  return S.Diag(range.getBegin(), diag::err_typecheck_call_too_many_args)
855821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)    << 0 /*function call*/ << desiredArgCount << argCount
865821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)    << call->getArg(1)->getSourceRange();
875821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)}
885821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)
895821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)ExprResult
90f8ee788a64d60abd8f2d742a5fdedde054ecd910Torne (Richard Coles)Sema::CheckBuiltinFunctionCall(unsigned BuiltinID, CallExpr *TheCall) {
915821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)  ExprResult TheCallResult(Owned(TheCall));
925821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)
935821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)  // Find out if any arguments are required to be integer constant expressions.
945821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)  unsigned ICEArguments = 0;
955821806d5e7f356e8fa4b058a389a808ea183019Torne (Richard Coles)  ASTContext::GetBuiltinTypeError Error;
96  Context.GetBuiltinType(BuiltinID, Error, &ICEArguments);
97  if (Error != ASTContext::GE_None)
98    ICEArguments = 0;  // Don't diagnose previously diagnosed errors.
99
100  // If any arguments are required to be ICE's, check and diagnose.
101  for (unsigned ArgNo = 0; ICEArguments != 0; ++ArgNo) {
102    // Skip arguments not required to be ICE's.
103    if ((ICEArguments & (1 << ArgNo)) == 0) continue;
104
105    llvm::APSInt Result;
106    if (SemaBuiltinConstantArg(TheCall, ArgNo, Result))
107      return true;
108    ICEArguments &= ~(1 << ArgNo);
109  }
110
111  switch (BuiltinID) {
112  case Builtin::BI__builtin___CFStringMakeConstantString:
113    assert(TheCall->getNumArgs() == 1 &&
114           "Wrong # arguments to builtin CFStringMakeConstantString");
115    if (CheckObjCString(TheCall->getArg(0)))
116      return ExprError();
117    break;
118  case Builtin::BI__builtin_stdarg_start:
119  case Builtin::BI__builtin_va_start:
120    if (SemaBuiltinVAStart(TheCall))
121      return ExprError();
122    break;
123  case Builtin::BI__builtin_isgreater:
124  case Builtin::BI__builtin_isgreaterequal:
125  case Builtin::BI__builtin_isless:
126  case Builtin::BI__builtin_islessequal:
127  case Builtin::BI__builtin_islessgreater:
128  case Builtin::BI__builtin_isunordered:
129    if (SemaBuiltinUnorderedCompare(TheCall))
130      return ExprError();
131    break;
132  case Builtin::BI__builtin_fpclassify:
133    if (SemaBuiltinFPClassification(TheCall, 6))
134      return ExprError();
135    break;
136  case Builtin::BI__builtin_isfinite:
137  case Builtin::BI__builtin_isinf:
138  case Builtin::BI__builtin_isinf_sign:
139  case Builtin::BI__builtin_isnan:
140  case Builtin::BI__builtin_isnormal:
141    if (SemaBuiltinFPClassification(TheCall, 1))
142      return ExprError();
143    break;
144  case Builtin::BI__builtin_shufflevector:
145    return SemaBuiltinShuffleVector(TheCall);
146    // TheCall will be freed by the smart pointer here, but that's fine, since
147    // SemaBuiltinShuffleVector guts it, but then doesn't release it.
148  case Builtin::BI__builtin_prefetch:
149    if (SemaBuiltinPrefetch(TheCall))
150      return ExprError();
151    break;
152  case Builtin::BI__builtin_object_size:
153    if (SemaBuiltinObjectSize(TheCall))
154      return ExprError();
155    break;
156  case Builtin::BI__builtin_longjmp:
157    if (SemaBuiltinLongjmp(TheCall))
158      return ExprError();
159    break;
160
161  case Builtin::BI__builtin_classify_type:
162    if (checkArgCount(*this, TheCall, 1)) return true;
163    TheCall->setType(Context.IntTy);
164    break;
165  case Builtin::BI__builtin_constant_p:
166    if (checkArgCount(*this, TheCall, 1)) return true;
167    TheCall->setType(Context.IntTy);
168    break;
169  case Builtin::BI__sync_fetch_and_add:
170  case Builtin::BI__sync_fetch_and_sub:
171  case Builtin::BI__sync_fetch_and_or:
172  case Builtin::BI__sync_fetch_and_and:
173  case Builtin::BI__sync_fetch_and_xor:
174  case Builtin::BI__sync_add_and_fetch:
175  case Builtin::BI__sync_sub_and_fetch:
176  case Builtin::BI__sync_and_and_fetch:
177  case Builtin::BI__sync_or_and_fetch:
178  case Builtin::BI__sync_xor_and_fetch:
179  case Builtin::BI__sync_val_compare_and_swap:
180  case Builtin::BI__sync_bool_compare_and_swap:
181  case Builtin::BI__sync_lock_test_and_set:
182  case Builtin::BI__sync_lock_release:
183    return SemaBuiltinAtomicOverloaded(move(TheCallResult));
184  }
185
186  // Since the target specific builtins for each arch overlap, only check those
187  // of the arch we are compiling for.
188  if (BuiltinID >= Builtin::FirstTSBuiltin) {
189    switch (Context.Target.getTriple().getArch()) {
190      case llvm::Triple::arm:
191      case llvm::Triple::thumb:
192        if (CheckARMBuiltinFunctionCall(BuiltinID, TheCall))
193          return ExprError();
194        break;
195      default:
196        break;
197    }
198  }
199
200  return move(TheCallResult);
201}
202
203// Get the valid immediate range for the specified NEON type code.
204static unsigned RFT(unsigned t, bool shift = false) {
205  bool quad = t & 0x10;
206
207  switch (t & 0x7) {
208    case 0: // i8
209      return shift ? 7 : (8 << (int)quad) - 1;
210    case 1: // i16
211      return shift ? 15 : (4 << (int)quad) - 1;
212    case 2: // i32
213      return shift ? 31 : (2 << (int)quad) - 1;
214    case 3: // i64
215      return shift ? 63 : (1 << (int)quad) - 1;
216    case 4: // f32
217      assert(!shift && "cannot shift float types!");
218      return (2 << (int)quad) - 1;
219    case 5: // poly8
220      return shift ? 7 : (8 << (int)quad) - 1;
221    case 6: // poly16
222      return shift ? 15 : (4 << (int)quad) - 1;
223    case 7: // float16
224      assert(!shift && "cannot shift float types!");
225      return (4 << (int)quad) - 1;
226  }
227  return 0;
228}
229
230bool Sema::CheckARMBuiltinFunctionCall(unsigned BuiltinID, CallExpr *TheCall) {
231  llvm::APSInt Result;
232
233  unsigned mask = 0;
234  unsigned TV = 0;
235  switch (BuiltinID) {
236#define GET_NEON_OVERLOAD_CHECK
237#include "clang/Basic/arm_neon.inc"
238#undef GET_NEON_OVERLOAD_CHECK
239  }
240
241  // For NEON intrinsics which are overloaded on vector element type, validate
242  // the immediate which specifies which variant to emit.
243  if (mask) {
244    unsigned ArgNo = TheCall->getNumArgs()-1;
245    if (SemaBuiltinConstantArg(TheCall, ArgNo, Result))
246      return true;
247
248    TV = Result.getLimitedValue(32);
249    if ((TV > 31) || (mask & (1 << TV)) == 0)
250      return Diag(TheCall->getLocStart(), diag::err_invalid_neon_type_code)
251        << TheCall->getArg(ArgNo)->getSourceRange();
252  }
253
254  // For NEON intrinsics which take an immediate value as part of the
255  // instruction, range check them here.
256  unsigned i = 0, l = 0, u = 0;
257  switch (BuiltinID) {
258  default: return false;
259  case ARM::BI__builtin_arm_ssat: i = 1; l = 1; u = 31; break;
260  case ARM::BI__builtin_arm_usat: i = 1; u = 31; break;
261  case ARM::BI__builtin_arm_vcvtr_f:
262  case ARM::BI__builtin_arm_vcvtr_d: i = 1; u = 1; break;
263#define GET_NEON_IMMEDIATE_CHECK
264#include "clang/Basic/arm_neon.inc"
265#undef GET_NEON_IMMEDIATE_CHECK
266  };
267
268  // Check that the immediate argument is actually a constant.
269  if (SemaBuiltinConstantArg(TheCall, i, Result))
270    return true;
271
272  // Range check against the upper/lower values for this isntruction.
273  unsigned Val = Result.getZExtValue();
274  if (Val < l || Val > (u + l))
275    return Diag(TheCall->getLocStart(), diag::err_argument_invalid_range)
276      << l << u+l << TheCall->getArg(i)->getSourceRange();
277
278  // FIXME: VFP Intrinsics should error if VFP not present.
279  return false;
280}
281
282/// CheckFunctionCall - Check a direct function call for various correctness
283/// and safety properties not strictly enforced by the C type system.
284bool Sema::CheckFunctionCall(FunctionDecl *FDecl, CallExpr *TheCall) {
285  // Get the IdentifierInfo* for the called function.
286  IdentifierInfo *FnInfo = FDecl->getIdentifier();
287
288  // None of the checks below are needed for functions that don't have
289  // simple names (e.g., C++ conversion functions).
290  if (!FnInfo)
291    return false;
292
293  // FIXME: This mechanism should be abstracted to be less fragile and
294  // more efficient. For example, just map function ids to custom
295  // handlers.
296
297  // Printf and scanf checking.
298  for (specific_attr_iterator<FormatAttr>
299         i = FDecl->specific_attr_begin<FormatAttr>(),
300         e = FDecl->specific_attr_end<FormatAttr>(); i != e ; ++i) {
301
302    const FormatAttr *Format = *i;
303    const bool b = Format->getType() == "scanf";
304    if (b || CheckablePrintfAttr(Format, TheCall)) {
305      bool HasVAListArg = Format->getFirstArg() == 0;
306      CheckPrintfScanfArguments(TheCall, HasVAListArg,
307                                Format->getFormatIdx() - 1,
308                                HasVAListArg ? 0 : Format->getFirstArg() - 1,
309                                !b);
310    }
311  }
312
313  for (specific_attr_iterator<NonNullAttr>
314         i = FDecl->specific_attr_begin<NonNullAttr>(),
315         e = FDecl->specific_attr_end<NonNullAttr>(); i != e; ++i) {
316    CheckNonNullArguments(*i, TheCall->getArgs(),
317                          TheCall->getCallee()->getLocStart());
318  }
319
320  return false;
321}
322
323bool Sema::CheckBlockCall(NamedDecl *NDecl, CallExpr *TheCall) {
324  // Printf checking.
325  const FormatAttr *Format = NDecl->getAttr<FormatAttr>();
326  if (!Format)
327    return false;
328
329  const VarDecl *V = dyn_cast<VarDecl>(NDecl);
330  if (!V)
331    return false;
332
333  QualType Ty = V->getType();
334  if (!Ty->isBlockPointerType())
335    return false;
336
337  const bool b = Format->getType() == "scanf";
338  if (!b && !CheckablePrintfAttr(Format, TheCall))
339    return false;
340
341  bool HasVAListArg = Format->getFirstArg() == 0;
342  CheckPrintfScanfArguments(TheCall, HasVAListArg, Format->getFormatIdx() - 1,
343                            HasVAListArg ? 0 : Format->getFirstArg() - 1, !b);
344
345  return false;
346}
347
348/// SemaBuiltinAtomicOverloaded - We have a call to a function like
349/// __sync_fetch_and_add, which is an overloaded function based on the pointer
350/// type of its first argument.  The main ActOnCallExpr routines have already
351/// promoted the types of arguments because all of these calls are prototyped as
352/// void(...).
353///
354/// This function goes through and does final semantic checking for these
355/// builtins,
356ExprResult
357Sema::SemaBuiltinAtomicOverloaded(ExprResult TheCallResult) {
358  CallExpr *TheCall = (CallExpr *)TheCallResult.get();
359  DeclRefExpr *DRE =cast<DeclRefExpr>(TheCall->getCallee()->IgnoreParenCasts());
360  FunctionDecl *FDecl = cast<FunctionDecl>(DRE->getDecl());
361
362  // Ensure that we have at least one argument to do type inference from.
363  if (TheCall->getNumArgs() < 1) {
364    Diag(TheCall->getLocEnd(), diag::err_typecheck_call_too_few_args_at_least)
365      << 0 << 1 << TheCall->getNumArgs()
366      << TheCall->getCallee()->getSourceRange();
367    return ExprError();
368  }
369
370  // Inspect the first argument of the atomic builtin.  This should always be
371  // a pointer type, whose element is an integral scalar or pointer type.
372  // Because it is a pointer type, we don't have to worry about any implicit
373  // casts here.
374  // FIXME: We don't allow floating point scalars as input.
375  Expr *FirstArg = TheCall->getArg(0);
376  if (!FirstArg->getType()->isPointerType()) {
377    Diag(DRE->getLocStart(), diag::err_atomic_builtin_must_be_pointer)
378      << FirstArg->getType() << FirstArg->getSourceRange();
379    return ExprError();
380  }
381
382  QualType ValType =
383    FirstArg->getType()->getAs<PointerType>()->getPointeeType();
384  if (!ValType->isIntegerType() && !ValType->isAnyPointerType() &&
385      !ValType->isBlockPointerType()) {
386    Diag(DRE->getLocStart(), diag::err_atomic_builtin_must_be_pointer_intptr)
387      << FirstArg->getType() << FirstArg->getSourceRange();
388    return ExprError();
389  }
390
391  // The majority of builtins return a value, but a few have special return
392  // types, so allow them to override appropriately below.
393  QualType ResultType = ValType;
394
395  // We need to figure out which concrete builtin this maps onto.  For example,
396  // __sync_fetch_and_add with a 2 byte object turns into
397  // __sync_fetch_and_add_2.
398#define BUILTIN_ROW(x) \
399  { Builtin::BI##x##_1, Builtin::BI##x##_2, Builtin::BI##x##_4, \
400    Builtin::BI##x##_8, Builtin::BI##x##_16 }
401
402  static const unsigned BuiltinIndices[][5] = {
403    BUILTIN_ROW(__sync_fetch_and_add),
404    BUILTIN_ROW(__sync_fetch_and_sub),
405    BUILTIN_ROW(__sync_fetch_and_or),
406    BUILTIN_ROW(__sync_fetch_and_and),
407    BUILTIN_ROW(__sync_fetch_and_xor),
408
409    BUILTIN_ROW(__sync_add_and_fetch),
410    BUILTIN_ROW(__sync_sub_and_fetch),
411    BUILTIN_ROW(__sync_and_and_fetch),
412    BUILTIN_ROW(__sync_or_and_fetch),
413    BUILTIN_ROW(__sync_xor_and_fetch),
414
415    BUILTIN_ROW(__sync_val_compare_and_swap),
416    BUILTIN_ROW(__sync_bool_compare_and_swap),
417    BUILTIN_ROW(__sync_lock_test_and_set),
418    BUILTIN_ROW(__sync_lock_release)
419  };
420#undef BUILTIN_ROW
421
422  // Determine the index of the size.
423  unsigned SizeIndex;
424  switch (Context.getTypeSizeInChars(ValType).getQuantity()) {
425  case 1: SizeIndex = 0; break;
426  case 2: SizeIndex = 1; break;
427  case 4: SizeIndex = 2; break;
428  case 8: SizeIndex = 3; break;
429  case 16: SizeIndex = 4; break;
430  default:
431    Diag(DRE->getLocStart(), diag::err_atomic_builtin_pointer_size)
432      << FirstArg->getType() << FirstArg->getSourceRange();
433    return ExprError();
434  }
435
436  // Each of these builtins has one pointer argument, followed by some number of
437  // values (0, 1 or 2) followed by a potentially empty varags list of stuff
438  // that we ignore.  Find out which row of BuiltinIndices to read from as well
439  // as the number of fixed args.
440  unsigned BuiltinID = FDecl->getBuiltinID();
441  unsigned BuiltinIndex, NumFixed = 1;
442  switch (BuiltinID) {
443  default: assert(0 && "Unknown overloaded atomic builtin!");
444  case Builtin::BI__sync_fetch_and_add: BuiltinIndex = 0; break;
445  case Builtin::BI__sync_fetch_and_sub: BuiltinIndex = 1; break;
446  case Builtin::BI__sync_fetch_and_or:  BuiltinIndex = 2; break;
447  case Builtin::BI__sync_fetch_and_and: BuiltinIndex = 3; break;
448  case Builtin::BI__sync_fetch_and_xor: BuiltinIndex = 4; break;
449
450  case Builtin::BI__sync_add_and_fetch: BuiltinIndex = 5; break;
451  case Builtin::BI__sync_sub_and_fetch: BuiltinIndex = 6; break;
452  case Builtin::BI__sync_and_and_fetch: BuiltinIndex = 7; break;
453  case Builtin::BI__sync_or_and_fetch:  BuiltinIndex = 8; break;
454  case Builtin::BI__sync_xor_and_fetch: BuiltinIndex = 9; break;
455
456  case Builtin::BI__sync_val_compare_and_swap:
457    BuiltinIndex = 10;
458    NumFixed = 2;
459    break;
460  case Builtin::BI__sync_bool_compare_and_swap:
461    BuiltinIndex = 11;
462    NumFixed = 2;
463    ResultType = Context.BoolTy;
464    break;
465  case Builtin::BI__sync_lock_test_and_set: BuiltinIndex = 12; break;
466  case Builtin::BI__sync_lock_release:
467    BuiltinIndex = 13;
468    NumFixed = 0;
469    ResultType = Context.VoidTy;
470    break;
471  }
472
473  // Now that we know how many fixed arguments we expect, first check that we
474  // have at least that many.
475  if (TheCall->getNumArgs() < 1+NumFixed) {
476    Diag(TheCall->getLocEnd(), diag::err_typecheck_call_too_few_args_at_least)
477      << 0 << 1+NumFixed << TheCall->getNumArgs()
478      << TheCall->getCallee()->getSourceRange();
479    return ExprError();
480  }
481
482  // Get the decl for the concrete builtin from this, we can tell what the
483  // concrete integer type we should convert to is.
484  unsigned NewBuiltinID = BuiltinIndices[BuiltinIndex][SizeIndex];
485  const char *NewBuiltinName = Context.BuiltinInfo.GetName(NewBuiltinID);
486  IdentifierInfo *NewBuiltinII = PP.getIdentifierInfo(NewBuiltinName);
487  FunctionDecl *NewBuiltinDecl =
488    cast<FunctionDecl>(LazilyCreateBuiltin(NewBuiltinII, NewBuiltinID,
489                                           TUScope, false, DRE->getLocStart()));
490
491  // The first argument --- the pointer --- has a fixed type; we
492  // deduce the types of the rest of the arguments accordingly.  Walk
493  // the remaining arguments, converting them to the deduced value type.
494  for (unsigned i = 0; i != NumFixed; ++i) {
495    Expr *Arg = TheCall->getArg(i+1);
496
497    // If the argument is an implicit cast, then there was a promotion due to
498    // "...", just remove it now.
499    if (ImplicitCastExpr *ICE = dyn_cast<ImplicitCastExpr>(Arg)) {
500      Arg = ICE->getSubExpr();
501      ICE->setSubExpr(0);
502      TheCall->setArg(i+1, Arg);
503    }
504
505    // GCC does an implicit conversion to the pointer or integer ValType.  This
506    // can fail in some cases (1i -> int**), check for this error case now.
507    CastKind Kind = CK_Invalid;
508    ExprValueKind VK = VK_RValue;
509    CXXCastPath BasePath;
510    if (CheckCastTypes(Arg->getSourceRange(), ValType, Arg, Kind, VK, BasePath))
511      return ExprError();
512
513    // Okay, we have something that *can* be converted to the right type.  Check
514    // to see if there is a potentially weird extension going on here.  This can
515    // happen when you do an atomic operation on something like an char* and
516    // pass in 42.  The 42 gets converted to char.  This is even more strange
517    // for things like 45.123 -> char, etc.
518    // FIXME: Do this check.
519    ImpCastExprToType(Arg, ValType, Kind, VK, &BasePath);
520    TheCall->setArg(i+1, Arg);
521  }
522
523  // Switch the DeclRefExpr to refer to the new decl.
524  DRE->setDecl(NewBuiltinDecl);
525  DRE->setType(NewBuiltinDecl->getType());
526
527  // Set the callee in the CallExpr.
528  // FIXME: This leaks the original parens and implicit casts.
529  Expr *PromotedCall = DRE;
530  UsualUnaryConversions(PromotedCall);
531  TheCall->setCallee(PromotedCall);
532
533  // Change the result type of the call to match the original value type. This
534  // is arbitrary, but the codegen for these builtins ins design to handle it
535  // gracefully.
536  TheCall->setType(ResultType);
537
538  return move(TheCallResult);
539}
540
541
542/// CheckObjCString - Checks that the argument to the builtin
543/// CFString constructor is correct
544/// Note: It might also make sense to do the UTF-16 conversion here (would
545/// simplify the backend).
546bool Sema::CheckObjCString(Expr *Arg) {
547  Arg = Arg->IgnoreParenCasts();
548  StringLiteral *Literal = dyn_cast<StringLiteral>(Arg);
549
550  if (!Literal || Literal->isWide()) {
551    Diag(Arg->getLocStart(), diag::err_cfstring_literal_not_string_constant)
552      << Arg->getSourceRange();
553    return true;
554  }
555
556  if (Literal->containsNonAsciiOrNull()) {
557    llvm::StringRef String = Literal->getString();
558    unsigned NumBytes = String.size();
559    llvm::SmallVector<UTF16, 128> ToBuf(NumBytes);
560    const UTF8 *FromPtr = (UTF8 *)String.data();
561    UTF16 *ToPtr = &ToBuf[0];
562
563    ConversionResult Result = ConvertUTF8toUTF16(&FromPtr, FromPtr + NumBytes,
564                                                 &ToPtr, ToPtr + NumBytes,
565                                                 strictConversion);
566    // Check for conversion failure.
567    if (Result != conversionOK)
568      Diag(Arg->getLocStart(),
569           diag::warn_cfstring_truncated) << Arg->getSourceRange();
570  }
571  return false;
572}
573
574/// SemaBuiltinVAStart - Check the arguments to __builtin_va_start for validity.
575/// Emit an error and return true on failure, return false on success.
576bool Sema::SemaBuiltinVAStart(CallExpr *TheCall) {
577  Expr *Fn = TheCall->getCallee();
578  if (TheCall->getNumArgs() > 2) {
579    Diag(TheCall->getArg(2)->getLocStart(),
580         diag::err_typecheck_call_too_many_args)
581      << 0 /*function call*/ << 2 << TheCall->getNumArgs()
582      << Fn->getSourceRange()
583      << SourceRange(TheCall->getArg(2)->getLocStart(),
584                     (*(TheCall->arg_end()-1))->getLocEnd());
585    return true;
586  }
587
588  if (TheCall->getNumArgs() < 2) {
589    return Diag(TheCall->getLocEnd(),
590      diag::err_typecheck_call_too_few_args_at_least)
591      << 0 /*function call*/ << 2 << TheCall->getNumArgs();
592  }
593
594  // Determine whether the current function is variadic or not.
595  BlockScopeInfo *CurBlock = getCurBlock();
596  bool isVariadic;
597  if (CurBlock)
598    isVariadic = CurBlock->TheDecl->isVariadic();
599  else if (FunctionDecl *FD = getCurFunctionDecl())
600    isVariadic = FD->isVariadic();
601  else
602    isVariadic = getCurMethodDecl()->isVariadic();
603
604  if (!isVariadic) {
605    Diag(Fn->getLocStart(), diag::err_va_start_used_in_non_variadic_function);
606    return true;
607  }
608
609  // Verify that the second argument to the builtin is the last argument of the
610  // current function or method.
611  bool SecondArgIsLastNamedArgument = false;
612  const Expr *Arg = TheCall->getArg(1)->IgnoreParenCasts();
613
614  if (const DeclRefExpr *DR = dyn_cast<DeclRefExpr>(Arg)) {
615    if (const ParmVarDecl *PV = dyn_cast<ParmVarDecl>(DR->getDecl())) {
616      // FIXME: This isn't correct for methods (results in bogus warning).
617      // Get the last formal in the current function.
618      const ParmVarDecl *LastArg;
619      if (CurBlock)
620        LastArg = *(CurBlock->TheDecl->param_end()-1);
621      else if (FunctionDecl *FD = getCurFunctionDecl())
622        LastArg = *(FD->param_end()-1);
623      else
624        LastArg = *(getCurMethodDecl()->param_end()-1);
625      SecondArgIsLastNamedArgument = PV == LastArg;
626    }
627  }
628
629  if (!SecondArgIsLastNamedArgument)
630    Diag(TheCall->getArg(1)->getLocStart(),
631         diag::warn_second_parameter_of_va_start_not_last_named_argument);
632  return false;
633}
634
635/// SemaBuiltinUnorderedCompare - Handle functions like __builtin_isgreater and
636/// friends.  This is declared to take (...), so we have to check everything.
637bool Sema::SemaBuiltinUnorderedCompare(CallExpr *TheCall) {
638  if (TheCall->getNumArgs() < 2)
639    return Diag(TheCall->getLocEnd(), diag::err_typecheck_call_too_few_args)
640      << 0 << 2 << TheCall->getNumArgs()/*function call*/;
641  if (TheCall->getNumArgs() > 2)
642    return Diag(TheCall->getArg(2)->getLocStart(),
643                diag::err_typecheck_call_too_many_args)
644      << 0 /*function call*/ << 2 << TheCall->getNumArgs()
645      << SourceRange(TheCall->getArg(2)->getLocStart(),
646                     (*(TheCall->arg_end()-1))->getLocEnd());
647
648  Expr *OrigArg0 = TheCall->getArg(0);
649  Expr *OrigArg1 = TheCall->getArg(1);
650
651  // Do standard promotions between the two arguments, returning their common
652  // type.
653  QualType Res = UsualArithmeticConversions(OrigArg0, OrigArg1, false);
654
655  // Make sure any conversions are pushed back into the call; this is
656  // type safe since unordered compare builtins are declared as "_Bool
657  // foo(...)".
658  TheCall->setArg(0, OrigArg0);
659  TheCall->setArg(1, OrigArg1);
660
661  if (OrigArg0->isTypeDependent() || OrigArg1->isTypeDependent())
662    return false;
663
664  // If the common type isn't a real floating type, then the arguments were
665  // invalid for this operation.
666  if (!Res->isRealFloatingType())
667    return Diag(OrigArg0->getLocStart(),
668                diag::err_typecheck_call_invalid_ordered_compare)
669      << OrigArg0->getType() << OrigArg1->getType()
670      << SourceRange(OrigArg0->getLocStart(), OrigArg1->getLocEnd());
671
672  return false;
673}
674
675/// SemaBuiltinSemaBuiltinFPClassification - Handle functions like
676/// __builtin_isnan and friends.  This is declared to take (...), so we have
677/// to check everything. We expect the last argument to be a floating point
678/// value.
679bool Sema::SemaBuiltinFPClassification(CallExpr *TheCall, unsigned NumArgs) {
680  if (TheCall->getNumArgs() < NumArgs)
681    return Diag(TheCall->getLocEnd(), diag::err_typecheck_call_too_few_args)
682      << 0 << NumArgs << TheCall->getNumArgs()/*function call*/;
683  if (TheCall->getNumArgs() > NumArgs)
684    return Diag(TheCall->getArg(NumArgs)->getLocStart(),
685                diag::err_typecheck_call_too_many_args)
686      << 0 /*function call*/ << NumArgs << TheCall->getNumArgs()
687      << SourceRange(TheCall->getArg(NumArgs)->getLocStart(),
688                     (*(TheCall->arg_end()-1))->getLocEnd());
689
690  Expr *OrigArg = TheCall->getArg(NumArgs-1);
691
692  if (OrigArg->isTypeDependent())
693    return false;
694
695  // This operation requires a non-_Complex floating-point number.
696  if (!OrigArg->getType()->isRealFloatingType())
697    return Diag(OrigArg->getLocStart(),
698                diag::err_typecheck_call_invalid_unary_fp)
699      << OrigArg->getType() << OrigArg->getSourceRange();
700
701  // If this is an implicit conversion from float -> double, remove it.
702  if (ImplicitCastExpr *Cast = dyn_cast<ImplicitCastExpr>(OrigArg)) {
703    Expr *CastArg = Cast->getSubExpr();
704    if (CastArg->getType()->isSpecificBuiltinType(BuiltinType::Float)) {
705      assert(Cast->getType()->isSpecificBuiltinType(BuiltinType::Double) &&
706             "promotion from float to double is the only expected cast here");
707      Cast->setSubExpr(0);
708      TheCall->setArg(NumArgs-1, CastArg);
709      OrigArg = CastArg;
710    }
711  }
712
713  return false;
714}
715
716/// SemaBuiltinShuffleVector - Handle __builtin_shufflevector.
717// This is declared to take (...), so we have to check everything.
718ExprResult Sema::SemaBuiltinShuffleVector(CallExpr *TheCall) {
719  if (TheCall->getNumArgs() < 2)
720    return ExprError(Diag(TheCall->getLocEnd(),
721                          diag::err_typecheck_call_too_few_args_at_least)
722      << 0 /*function call*/ << 2 << TheCall->getNumArgs()
723      << TheCall->getSourceRange());
724
725  // Determine which of the following types of shufflevector we're checking:
726  // 1) unary, vector mask: (lhs, mask)
727  // 2) binary, vector mask: (lhs, rhs, mask)
728  // 3) binary, scalar mask: (lhs, rhs, index, ..., index)
729  QualType resType = TheCall->getArg(0)->getType();
730  unsigned numElements = 0;
731
732  if (!TheCall->getArg(0)->isTypeDependent() &&
733      !TheCall->getArg(1)->isTypeDependent()) {
734    QualType LHSType = TheCall->getArg(0)->getType();
735    QualType RHSType = TheCall->getArg(1)->getType();
736
737    if (!LHSType->isVectorType() || !RHSType->isVectorType()) {
738      Diag(TheCall->getLocStart(), diag::err_shufflevector_non_vector)
739        << SourceRange(TheCall->getArg(0)->getLocStart(),
740                       TheCall->getArg(1)->getLocEnd());
741      return ExprError();
742    }
743
744    numElements = LHSType->getAs<VectorType>()->getNumElements();
745    unsigned numResElements = TheCall->getNumArgs() - 2;
746
747    // Check to see if we have a call with 2 vector arguments, the unary shuffle
748    // with mask.  If so, verify that RHS is an integer vector type with the
749    // same number of elts as lhs.
750    if (TheCall->getNumArgs() == 2) {
751      if (!RHSType->hasIntegerRepresentation() ||
752          RHSType->getAs<VectorType>()->getNumElements() != numElements)
753        Diag(TheCall->getLocStart(), diag::err_shufflevector_incompatible_vector)
754          << SourceRange(TheCall->getArg(1)->getLocStart(),
755                         TheCall->getArg(1)->getLocEnd());
756      numResElements = numElements;
757    }
758    else if (!Context.hasSameUnqualifiedType(LHSType, RHSType)) {
759      Diag(TheCall->getLocStart(), diag::err_shufflevector_incompatible_vector)
760        << SourceRange(TheCall->getArg(0)->getLocStart(),
761                       TheCall->getArg(1)->getLocEnd());
762      return ExprError();
763    } else if (numElements != numResElements) {
764      QualType eltType = LHSType->getAs<VectorType>()->getElementType();
765      resType = Context.getVectorType(eltType, numResElements,
766                                      VectorType::GenericVector);
767    }
768  }
769
770  for (unsigned i = 2; i < TheCall->getNumArgs(); i++) {
771    if (TheCall->getArg(i)->isTypeDependent() ||
772        TheCall->getArg(i)->isValueDependent())
773      continue;
774
775    llvm::APSInt Result(32);
776    if (!TheCall->getArg(i)->isIntegerConstantExpr(Result, Context))
777      return ExprError(Diag(TheCall->getLocStart(),
778                  diag::err_shufflevector_nonconstant_argument)
779                << TheCall->getArg(i)->getSourceRange());
780
781    if (Result.getActiveBits() > 64 || Result.getZExtValue() >= numElements*2)
782      return ExprError(Diag(TheCall->getLocStart(),
783                  diag::err_shufflevector_argument_too_large)
784               << TheCall->getArg(i)->getSourceRange());
785  }
786
787  llvm::SmallVector<Expr*, 32> exprs;
788
789  for (unsigned i = 0, e = TheCall->getNumArgs(); i != e; i++) {
790    exprs.push_back(TheCall->getArg(i));
791    TheCall->setArg(i, 0);
792  }
793
794  return Owned(new (Context) ShuffleVectorExpr(Context, exprs.begin(),
795                                            exprs.size(), resType,
796                                            TheCall->getCallee()->getLocStart(),
797                                            TheCall->getRParenLoc()));
798}
799
800/// SemaBuiltinPrefetch - Handle __builtin_prefetch.
801// This is declared to take (const void*, ...) and can take two
802// optional constant int args.
803bool Sema::SemaBuiltinPrefetch(CallExpr *TheCall) {
804  unsigned NumArgs = TheCall->getNumArgs();
805
806  if (NumArgs > 3)
807    return Diag(TheCall->getLocEnd(),
808             diag::err_typecheck_call_too_many_args_at_most)
809             << 0 /*function call*/ << 3 << NumArgs
810             << TheCall->getSourceRange();
811
812  // Argument 0 is checked for us and the remaining arguments must be
813  // constant integers.
814  for (unsigned i = 1; i != NumArgs; ++i) {
815    Expr *Arg = TheCall->getArg(i);
816
817    llvm::APSInt Result;
818    if (SemaBuiltinConstantArg(TheCall, i, Result))
819      return true;
820
821    // FIXME: gcc issues a warning and rewrites these to 0. These
822    // seems especially odd for the third argument since the default
823    // is 3.
824    if (i == 1) {
825      if (Result.getLimitedValue() > 1)
826        return Diag(TheCall->getLocStart(), diag::err_argument_invalid_range)
827             << "0" << "1" << Arg->getSourceRange();
828    } else {
829      if (Result.getLimitedValue() > 3)
830        return Diag(TheCall->getLocStart(), diag::err_argument_invalid_range)
831            << "0" << "3" << Arg->getSourceRange();
832    }
833  }
834
835  return false;
836}
837
838/// SemaBuiltinConstantArg - Handle a check if argument ArgNum of CallExpr
839/// TheCall is a constant expression.
840bool Sema::SemaBuiltinConstantArg(CallExpr *TheCall, int ArgNum,
841                                  llvm::APSInt &Result) {
842  Expr *Arg = TheCall->getArg(ArgNum);
843  DeclRefExpr *DRE =cast<DeclRefExpr>(TheCall->getCallee()->IgnoreParenCasts());
844  FunctionDecl *FDecl = cast<FunctionDecl>(DRE->getDecl());
845
846  if (Arg->isTypeDependent() || Arg->isValueDependent()) return false;
847
848  if (!Arg->isIntegerConstantExpr(Result, Context))
849    return Diag(TheCall->getLocStart(), diag::err_constant_integer_arg_type)
850                << FDecl->getDeclName() <<  Arg->getSourceRange();
851
852  return false;
853}
854
855/// SemaBuiltinObjectSize - Handle __builtin_object_size(void *ptr,
856/// int type). This simply type checks that type is one of the defined
857/// constants (0-3).
858// For compatability check 0-3, llvm only handles 0 and 2.
859bool Sema::SemaBuiltinObjectSize(CallExpr *TheCall) {
860  llvm::APSInt Result;
861
862  // Check constant-ness first.
863  if (SemaBuiltinConstantArg(TheCall, 1, Result))
864    return true;
865
866  Expr *Arg = TheCall->getArg(1);
867  if (Result.getSExtValue() < 0 || Result.getSExtValue() > 3) {
868    return Diag(TheCall->getLocStart(), diag::err_argument_invalid_range)
869             << "0" << "3" << SourceRange(Arg->getLocStart(), Arg->getLocEnd());
870  }
871
872  return false;
873}
874
875/// SemaBuiltinLongjmp - Handle __builtin_longjmp(void *env[5], int val).
876/// This checks that val is a constant 1.
877bool Sema::SemaBuiltinLongjmp(CallExpr *TheCall) {
878  Expr *Arg = TheCall->getArg(1);
879  llvm::APSInt Result;
880
881  // TODO: This is less than ideal. Overload this to take a value.
882  if (SemaBuiltinConstantArg(TheCall, 1, Result))
883    return true;
884
885  if (Result != 1)
886    return Diag(TheCall->getLocStart(), diag::err_builtin_longjmp_invalid_val)
887             << SourceRange(Arg->getLocStart(), Arg->getLocEnd());
888
889  return false;
890}
891
892// Handle i > 1 ? "x" : "y", recursively.
893bool Sema::SemaCheckStringLiteral(const Expr *E, const CallExpr *TheCall,
894                                  bool HasVAListArg,
895                                  unsigned format_idx, unsigned firstDataArg,
896                                  bool isPrintf) {
897 tryAgain:
898  if (E->isTypeDependent() || E->isValueDependent())
899    return false;
900
901  switch (E->getStmtClass()) {
902  case Stmt::BinaryConditionalOperatorClass:
903  case Stmt::ConditionalOperatorClass: {
904    const AbstractConditionalOperator *C = cast<AbstractConditionalOperator>(E);
905    return SemaCheckStringLiteral(C->getTrueExpr(), TheCall, HasVAListArg,
906                                  format_idx, firstDataArg, isPrintf)
907        && SemaCheckStringLiteral(C->getFalseExpr(), TheCall, HasVAListArg,
908                                  format_idx, firstDataArg, isPrintf);
909  }
910
911  case Stmt::IntegerLiteralClass:
912    // Technically -Wformat-nonliteral does not warn about this case.
913    // The behavior of printf and friends in this case is implementation
914    // dependent.  Ideally if the format string cannot be null then
915    // it should have a 'nonnull' attribute in the function prototype.
916    return true;
917
918  case Stmt::ImplicitCastExprClass: {
919    E = cast<ImplicitCastExpr>(E)->getSubExpr();
920    goto tryAgain;
921  }
922
923  case Stmt::ParenExprClass: {
924    E = cast<ParenExpr>(E)->getSubExpr();
925    goto tryAgain;
926  }
927
928  case Stmt::OpaqueValueExprClass:
929    if (const Expr *src = cast<OpaqueValueExpr>(E)->getSourceExpr()) {
930      E = src;
931      goto tryAgain;
932    }
933    return false;
934
935  case Stmt::PredefinedExprClass:
936    // While __func__, etc., are technically not string literals, they
937    // cannot contain format specifiers and thus are not a security
938    // liability.
939    return true;
940
941  case Stmt::DeclRefExprClass: {
942    const DeclRefExpr *DR = cast<DeclRefExpr>(E);
943
944    // As an exception, do not flag errors for variables binding to
945    // const string literals.
946    if (const VarDecl *VD = dyn_cast<VarDecl>(DR->getDecl())) {
947      bool isConstant = false;
948      QualType T = DR->getType();
949
950      if (const ArrayType *AT = Context.getAsArrayType(T)) {
951        isConstant = AT->getElementType().isConstant(Context);
952      } else if (const PointerType *PT = T->getAs<PointerType>()) {
953        isConstant = T.isConstant(Context) &&
954                     PT->getPointeeType().isConstant(Context);
955      }
956
957      if (isConstant) {
958        if (const Expr *Init = VD->getAnyInitializer())
959          return SemaCheckStringLiteral(Init, TheCall,
960                                        HasVAListArg, format_idx, firstDataArg,
961                                        isPrintf);
962      }
963
964      // For vprintf* functions (i.e., HasVAListArg==true), we add a
965      // special check to see if the format string is a function parameter
966      // of the function calling the printf function.  If the function
967      // has an attribute indicating it is a printf-like function, then we
968      // should suppress warnings concerning non-literals being used in a call
969      // to a vprintf function.  For example:
970      //
971      // void
972      // logmessage(char const *fmt __attribute__ (format (printf, 1, 2)), ...){
973      //      va_list ap;
974      //      va_start(ap, fmt);
975      //      vprintf(fmt, ap);  // Do NOT emit a warning about "fmt".
976      //      ...
977      //
978      //
979      //  FIXME: We don't have full attribute support yet, so just check to see
980      //    if the argument is a DeclRefExpr that references a parameter.  We'll
981      //    add proper support for checking the attribute later.
982      if (HasVAListArg)
983        if (isa<ParmVarDecl>(VD))
984          return true;
985    }
986
987    return false;
988  }
989
990  case Stmt::CallExprClass: {
991    const CallExpr *CE = cast<CallExpr>(E);
992    if (const ImplicitCastExpr *ICE
993          = dyn_cast<ImplicitCastExpr>(CE->getCallee())) {
994      if (const DeclRefExpr *DRE = dyn_cast<DeclRefExpr>(ICE->getSubExpr())) {
995        if (const FunctionDecl *FD = dyn_cast<FunctionDecl>(DRE->getDecl())) {
996          if (const FormatArgAttr *FA = FD->getAttr<FormatArgAttr>()) {
997            unsigned ArgIndex = FA->getFormatIdx();
998            const Expr *Arg = CE->getArg(ArgIndex - 1);
999
1000            return SemaCheckStringLiteral(Arg, TheCall, HasVAListArg,
1001                                          format_idx, firstDataArg, isPrintf);
1002          }
1003        }
1004      }
1005    }
1006
1007    return false;
1008  }
1009  case Stmt::ObjCStringLiteralClass:
1010  case Stmt::StringLiteralClass: {
1011    const StringLiteral *StrE = NULL;
1012
1013    if (const ObjCStringLiteral *ObjCFExpr = dyn_cast<ObjCStringLiteral>(E))
1014      StrE = ObjCFExpr->getString();
1015    else
1016      StrE = cast<StringLiteral>(E);
1017
1018    if (StrE) {
1019      CheckFormatString(StrE, E, TheCall, HasVAListArg, format_idx,
1020                        firstDataArg, isPrintf);
1021      return true;
1022    }
1023
1024    return false;
1025  }
1026
1027  default:
1028    return false;
1029  }
1030}
1031
1032void
1033Sema::CheckNonNullArguments(const NonNullAttr *NonNull,
1034                            const Expr * const *ExprArgs,
1035                            SourceLocation CallSiteLoc) {
1036  for (NonNullAttr::args_iterator i = NonNull->args_begin(),
1037                                  e = NonNull->args_end();
1038       i != e; ++i) {
1039    const Expr *ArgExpr = ExprArgs[*i];
1040    if (ArgExpr->isNullPointerConstant(Context,
1041                                       Expr::NPC_ValueDependentIsNotNull))
1042      Diag(CallSiteLoc, diag::warn_null_arg) << ArgExpr->getSourceRange();
1043  }
1044}
1045
1046/// CheckPrintfScanfArguments - Check calls to printf and scanf (and similar
1047/// functions) for correct use of format strings.
1048void
1049Sema::CheckPrintfScanfArguments(const CallExpr *TheCall, bool HasVAListArg,
1050                                unsigned format_idx, unsigned firstDataArg,
1051                                bool isPrintf) {
1052
1053  const Expr *Fn = TheCall->getCallee();
1054
1055  // The way the format attribute works in GCC, the implicit this argument
1056  // of member functions is counted. However, it doesn't appear in our own
1057  // lists, so decrement format_idx in that case.
1058  if (isa<CXXMemberCallExpr>(TheCall)) {
1059    const CXXMethodDecl *method_decl =
1060      dyn_cast<CXXMethodDecl>(TheCall->getCalleeDecl());
1061    if (method_decl && method_decl->isInstance()) {
1062      // Catch a format attribute mistakenly referring to the object argument.
1063      if (format_idx == 0)
1064        return;
1065      --format_idx;
1066      if(firstDataArg != 0)
1067        --firstDataArg;
1068    }
1069  }
1070
1071  // CHECK: printf/scanf-like function is called with no format string.
1072  if (format_idx >= TheCall->getNumArgs()) {
1073    Diag(TheCall->getRParenLoc(), diag::warn_missing_format_string)
1074      << Fn->getSourceRange();
1075    return;
1076  }
1077
1078  const Expr *OrigFormatExpr = TheCall->getArg(format_idx)->IgnoreParenCasts();
1079
1080  // CHECK: format string is not a string literal.
1081  //
1082  // Dynamically generated format strings are difficult to
1083  // automatically vet at compile time.  Requiring that format strings
1084  // are string literals: (1) permits the checking of format strings by
1085  // the compiler and thereby (2) can practically remove the source of
1086  // many format string exploits.
1087
1088  // Format string can be either ObjC string (e.g. @"%d") or
1089  // C string (e.g. "%d")
1090  // ObjC string uses the same format specifiers as C string, so we can use
1091  // the same format string checking logic for both ObjC and C strings.
1092  if (SemaCheckStringLiteral(OrigFormatExpr, TheCall, HasVAListArg, format_idx,
1093                             firstDataArg, isPrintf))
1094    return;  // Literal format string found, check done!
1095
1096  // If there are no arguments specified, warn with -Wformat-security, otherwise
1097  // warn only with -Wformat-nonliteral.
1098  if (TheCall->getNumArgs() == format_idx+1)
1099    Diag(TheCall->getArg(format_idx)->getLocStart(),
1100         diag::warn_format_nonliteral_noargs)
1101      << OrigFormatExpr->getSourceRange();
1102  else
1103    Diag(TheCall->getArg(format_idx)->getLocStart(),
1104         diag::warn_format_nonliteral)
1105           << OrigFormatExpr->getSourceRange();
1106}
1107
1108namespace {
1109class CheckFormatHandler : public analyze_format_string::FormatStringHandler {
1110protected:
1111  Sema &S;
1112  const StringLiteral *FExpr;
1113  const Expr *OrigFormatExpr;
1114  const unsigned FirstDataArg;
1115  const unsigned NumDataArgs;
1116  const bool IsObjCLiteral;
1117  const char *Beg; // Start of format string.
1118  const bool HasVAListArg;
1119  const CallExpr *TheCall;
1120  unsigned FormatIdx;
1121  llvm::BitVector CoveredArgs;
1122  bool usesPositionalArgs;
1123  bool atFirstArg;
1124public:
1125  CheckFormatHandler(Sema &s, const StringLiteral *fexpr,
1126                     const Expr *origFormatExpr, unsigned firstDataArg,
1127                     unsigned numDataArgs, bool isObjCLiteral,
1128                     const char *beg, bool hasVAListArg,
1129                     const CallExpr *theCall, unsigned formatIdx)
1130    : S(s), FExpr(fexpr), OrigFormatExpr(origFormatExpr),
1131      FirstDataArg(firstDataArg),
1132      NumDataArgs(numDataArgs),
1133      IsObjCLiteral(isObjCLiteral), Beg(beg),
1134      HasVAListArg(hasVAListArg),
1135      TheCall(theCall), FormatIdx(formatIdx),
1136      usesPositionalArgs(false), atFirstArg(true) {
1137        CoveredArgs.resize(numDataArgs);
1138        CoveredArgs.reset();
1139      }
1140
1141  void DoneProcessing();
1142
1143  void HandleIncompleteSpecifier(const char *startSpecifier,
1144                                 unsigned specifierLen);
1145
1146  virtual void HandleInvalidPosition(const char *startSpecifier,
1147                                     unsigned specifierLen,
1148                                     analyze_format_string::PositionContext p);
1149
1150  virtual void HandleZeroPosition(const char *startPos, unsigned posLen);
1151
1152  void HandleNullChar(const char *nullCharacter);
1153
1154protected:
1155  bool HandleInvalidConversionSpecifier(unsigned argIndex, SourceLocation Loc,
1156                                        const char *startSpec,
1157                                        unsigned specifierLen,
1158                                        const char *csStart, unsigned csLen);
1159
1160  SourceRange getFormatStringRange();
1161  CharSourceRange getSpecifierRange(const char *startSpecifier,
1162                                    unsigned specifierLen);
1163  SourceLocation getLocationOfByte(const char *x);
1164
1165  const Expr *getDataArg(unsigned i) const;
1166
1167  bool CheckNumArgs(const analyze_format_string::FormatSpecifier &FS,
1168                    const analyze_format_string::ConversionSpecifier &CS,
1169                    const char *startSpecifier, unsigned specifierLen,
1170                    unsigned argIndex);
1171};
1172}
1173
1174SourceRange CheckFormatHandler::getFormatStringRange() {
1175  return OrigFormatExpr->getSourceRange();
1176}
1177
1178CharSourceRange CheckFormatHandler::
1179getSpecifierRange(const char *startSpecifier, unsigned specifierLen) {
1180  SourceLocation Start = getLocationOfByte(startSpecifier);
1181  SourceLocation End   = getLocationOfByte(startSpecifier + specifierLen - 1);
1182
1183  // Advance the end SourceLocation by one due to half-open ranges.
1184  End = End.getFileLocWithOffset(1);
1185
1186  return CharSourceRange::getCharRange(Start, End);
1187}
1188
1189SourceLocation CheckFormatHandler::getLocationOfByte(const char *x) {
1190  return S.getLocationOfStringLiteralByte(FExpr, x - Beg);
1191}
1192
1193void CheckFormatHandler::HandleIncompleteSpecifier(const char *startSpecifier,
1194                                                   unsigned specifierLen){
1195  SourceLocation Loc = getLocationOfByte(startSpecifier);
1196  S.Diag(Loc, diag::warn_printf_incomplete_specifier)
1197    << getSpecifierRange(startSpecifier, specifierLen);
1198}
1199
1200void
1201CheckFormatHandler::HandleInvalidPosition(const char *startPos, unsigned posLen,
1202                                     analyze_format_string::PositionContext p) {
1203  SourceLocation Loc = getLocationOfByte(startPos);
1204  S.Diag(Loc, diag::warn_format_invalid_positional_specifier)
1205    << (unsigned) p << getSpecifierRange(startPos, posLen);
1206}
1207
1208void CheckFormatHandler::HandleZeroPosition(const char *startPos,
1209                                            unsigned posLen) {
1210  SourceLocation Loc = getLocationOfByte(startPos);
1211  S.Diag(Loc, diag::warn_format_zero_positional_specifier)
1212    << getSpecifierRange(startPos, posLen);
1213}
1214
1215void CheckFormatHandler::HandleNullChar(const char *nullCharacter) {
1216  if (!IsObjCLiteral) {
1217    // The presence of a null character is likely an error.
1218    S.Diag(getLocationOfByte(nullCharacter),
1219           diag::warn_printf_format_string_contains_null_char)
1220      << getFormatStringRange();
1221  }
1222}
1223
1224const Expr *CheckFormatHandler::getDataArg(unsigned i) const {
1225  return TheCall->getArg(FirstDataArg + i);
1226}
1227
1228void CheckFormatHandler::DoneProcessing() {
1229    // Does the number of data arguments exceed the number of
1230    // format conversions in the format string?
1231  if (!HasVAListArg) {
1232      // Find any arguments that weren't covered.
1233    CoveredArgs.flip();
1234    signed notCoveredArg = CoveredArgs.find_first();
1235    if (notCoveredArg >= 0) {
1236      assert((unsigned)notCoveredArg < NumDataArgs);
1237      S.Diag(getDataArg((unsigned) notCoveredArg)->getLocStart(),
1238             diag::warn_printf_data_arg_not_used)
1239      << getFormatStringRange();
1240    }
1241  }
1242}
1243
1244bool
1245CheckFormatHandler::HandleInvalidConversionSpecifier(unsigned argIndex,
1246                                                     SourceLocation Loc,
1247                                                     const char *startSpec,
1248                                                     unsigned specifierLen,
1249                                                     const char *csStart,
1250                                                     unsigned csLen) {
1251
1252  bool keepGoing = true;
1253  if (argIndex < NumDataArgs) {
1254    // Consider the argument coverered, even though the specifier doesn't
1255    // make sense.
1256    CoveredArgs.set(argIndex);
1257  }
1258  else {
1259    // If argIndex exceeds the number of data arguments we
1260    // don't issue a warning because that is just a cascade of warnings (and
1261    // they may have intended '%%' anyway). We don't want to continue processing
1262    // the format string after this point, however, as we will like just get
1263    // gibberish when trying to match arguments.
1264    keepGoing = false;
1265  }
1266
1267  S.Diag(Loc, diag::warn_format_invalid_conversion)
1268    << llvm::StringRef(csStart, csLen)
1269    << getSpecifierRange(startSpec, specifierLen);
1270
1271  return keepGoing;
1272}
1273
1274bool
1275CheckFormatHandler::CheckNumArgs(
1276  const analyze_format_string::FormatSpecifier &FS,
1277  const analyze_format_string::ConversionSpecifier &CS,
1278  const char *startSpecifier, unsigned specifierLen, unsigned argIndex) {
1279
1280  if (argIndex >= NumDataArgs) {
1281    if (FS.usesPositionalArg())  {
1282      S.Diag(getLocationOfByte(CS.getStart()),
1283             diag::warn_printf_positional_arg_exceeds_data_args)
1284      << (argIndex+1) << NumDataArgs
1285      << getSpecifierRange(startSpecifier, specifierLen);
1286    }
1287    else {
1288      S.Diag(getLocationOfByte(CS.getStart()),
1289             diag::warn_printf_insufficient_data_args)
1290      << getSpecifierRange(startSpecifier, specifierLen);
1291    }
1292
1293    return false;
1294  }
1295  return true;
1296}
1297
1298//===--- CHECK: Printf format string checking ------------------------------===//
1299
1300namespace {
1301class CheckPrintfHandler : public CheckFormatHandler {
1302public:
1303  CheckPrintfHandler(Sema &s, const StringLiteral *fexpr,
1304                     const Expr *origFormatExpr, unsigned firstDataArg,
1305                     unsigned numDataArgs, bool isObjCLiteral,
1306                     const char *beg, bool hasVAListArg,
1307                     const CallExpr *theCall, unsigned formatIdx)
1308  : CheckFormatHandler(s, fexpr, origFormatExpr, firstDataArg,
1309                       numDataArgs, isObjCLiteral, beg, hasVAListArg,
1310                       theCall, formatIdx) {}
1311
1312
1313  bool HandleInvalidPrintfConversionSpecifier(
1314                                      const analyze_printf::PrintfSpecifier &FS,
1315                                      const char *startSpecifier,
1316                                      unsigned specifierLen);
1317
1318  bool HandlePrintfSpecifier(const analyze_printf::PrintfSpecifier &FS,
1319                             const char *startSpecifier,
1320                             unsigned specifierLen);
1321
1322  bool HandleAmount(const analyze_format_string::OptionalAmount &Amt, unsigned k,
1323                    const char *startSpecifier, unsigned specifierLen);
1324  void HandleInvalidAmount(const analyze_printf::PrintfSpecifier &FS,
1325                           const analyze_printf::OptionalAmount &Amt,
1326                           unsigned type,
1327                           const char *startSpecifier, unsigned specifierLen);
1328  void HandleFlag(const analyze_printf::PrintfSpecifier &FS,
1329                  const analyze_printf::OptionalFlag &flag,
1330                  const char *startSpecifier, unsigned specifierLen);
1331  void HandleIgnoredFlag(const analyze_printf::PrintfSpecifier &FS,
1332                         const analyze_printf::OptionalFlag &ignoredFlag,
1333                         const analyze_printf::OptionalFlag &flag,
1334                         const char *startSpecifier, unsigned specifierLen);
1335};
1336}
1337
1338bool CheckPrintfHandler::HandleInvalidPrintfConversionSpecifier(
1339                                      const analyze_printf::PrintfSpecifier &FS,
1340                                      const char *startSpecifier,
1341                                      unsigned specifierLen) {
1342  const analyze_printf::PrintfConversionSpecifier &CS =
1343    FS.getConversionSpecifier();
1344
1345  return HandleInvalidConversionSpecifier(FS.getArgIndex(),
1346                                          getLocationOfByte(CS.getStart()),
1347                                          startSpecifier, specifierLen,
1348                                          CS.getStart(), CS.getLength());
1349}
1350
1351bool CheckPrintfHandler::HandleAmount(
1352                               const analyze_format_string::OptionalAmount &Amt,
1353                               unsigned k, const char *startSpecifier,
1354                               unsigned specifierLen) {
1355
1356  if (Amt.hasDataArgument()) {
1357    if (!HasVAListArg) {
1358      unsigned argIndex = Amt.getArgIndex();
1359      if (argIndex >= NumDataArgs) {
1360        S.Diag(getLocationOfByte(Amt.getStart()),
1361               diag::warn_printf_asterisk_missing_arg)
1362          << k << getSpecifierRange(startSpecifier, specifierLen);
1363        // Don't do any more checking.  We will just emit
1364        // spurious errors.
1365        return false;
1366      }
1367
1368      // Type check the data argument.  It should be an 'int'.
1369      // Although not in conformance with C99, we also allow the argument to be
1370      // an 'unsigned int' as that is a reasonably safe case.  GCC also
1371      // doesn't emit a warning for that case.
1372      CoveredArgs.set(argIndex);
1373      const Expr *Arg = getDataArg(argIndex);
1374      QualType T = Arg->getType();
1375
1376      const analyze_printf::ArgTypeResult &ATR = Amt.getArgType(S.Context);
1377      assert(ATR.isValid());
1378
1379      if (!ATR.matchesType(S.Context, T)) {
1380        S.Diag(getLocationOfByte(Amt.getStart()),
1381               diag::warn_printf_asterisk_wrong_type)
1382          << k
1383          << ATR.getRepresentativeType(S.Context) << T
1384          << getSpecifierRange(startSpecifier, specifierLen)
1385          << Arg->getSourceRange();
1386        // Don't do any more checking.  We will just emit
1387        // spurious errors.
1388        return false;
1389      }
1390    }
1391  }
1392  return true;
1393}
1394
1395void CheckPrintfHandler::HandleInvalidAmount(
1396                                      const analyze_printf::PrintfSpecifier &FS,
1397                                      const analyze_printf::OptionalAmount &Amt,
1398                                      unsigned type,
1399                                      const char *startSpecifier,
1400                                      unsigned specifierLen) {
1401  const analyze_printf::PrintfConversionSpecifier &CS =
1402    FS.getConversionSpecifier();
1403  switch (Amt.getHowSpecified()) {
1404  case analyze_printf::OptionalAmount::Constant:
1405    S.Diag(getLocationOfByte(Amt.getStart()),
1406        diag::warn_printf_nonsensical_optional_amount)
1407      << type
1408      << CS.toString()
1409      << getSpecifierRange(startSpecifier, specifierLen)
1410      << FixItHint::CreateRemoval(getSpecifierRange(Amt.getStart(),
1411          Amt.getConstantLength()));
1412    break;
1413
1414  default:
1415    S.Diag(getLocationOfByte(Amt.getStart()),
1416        diag::warn_printf_nonsensical_optional_amount)
1417      << type
1418      << CS.toString()
1419      << getSpecifierRange(startSpecifier, specifierLen);
1420    break;
1421  }
1422}
1423
1424void CheckPrintfHandler::HandleFlag(const analyze_printf::PrintfSpecifier &FS,
1425                                    const analyze_printf::OptionalFlag &flag,
1426                                    const char *startSpecifier,
1427                                    unsigned specifierLen) {
1428  // Warn about pointless flag with a fixit removal.
1429  const analyze_printf::PrintfConversionSpecifier &CS =
1430    FS.getConversionSpecifier();
1431  S.Diag(getLocationOfByte(flag.getPosition()),
1432      diag::warn_printf_nonsensical_flag)
1433    << flag.toString() << CS.toString()
1434    << getSpecifierRange(startSpecifier, specifierLen)
1435    << FixItHint::CreateRemoval(getSpecifierRange(flag.getPosition(), 1));
1436}
1437
1438void CheckPrintfHandler::HandleIgnoredFlag(
1439                                const analyze_printf::PrintfSpecifier &FS,
1440                                const analyze_printf::OptionalFlag &ignoredFlag,
1441                                const analyze_printf::OptionalFlag &flag,
1442                                const char *startSpecifier,
1443                                unsigned specifierLen) {
1444  // Warn about ignored flag with a fixit removal.
1445  S.Diag(getLocationOfByte(ignoredFlag.getPosition()),
1446      diag::warn_printf_ignored_flag)
1447    << ignoredFlag.toString() << flag.toString()
1448    << getSpecifierRange(startSpecifier, specifierLen)
1449    << FixItHint::CreateRemoval(getSpecifierRange(
1450        ignoredFlag.getPosition(), 1));
1451}
1452
1453bool
1454CheckPrintfHandler::HandlePrintfSpecifier(const analyze_printf::PrintfSpecifier
1455                                            &FS,
1456                                          const char *startSpecifier,
1457                                          unsigned specifierLen) {
1458
1459  using namespace analyze_format_string;
1460  using namespace analyze_printf;
1461  const PrintfConversionSpecifier &CS = FS.getConversionSpecifier();
1462
1463  if (FS.consumesDataArgument()) {
1464    if (atFirstArg) {
1465        atFirstArg = false;
1466        usesPositionalArgs = FS.usesPositionalArg();
1467    }
1468    else if (usesPositionalArgs != FS.usesPositionalArg()) {
1469      // Cannot mix-and-match positional and non-positional arguments.
1470      S.Diag(getLocationOfByte(CS.getStart()),
1471             diag::warn_format_mix_positional_nonpositional_args)
1472        << getSpecifierRange(startSpecifier, specifierLen);
1473      return false;
1474    }
1475  }
1476
1477  // First check if the field width, precision, and conversion specifier
1478  // have matching data arguments.
1479  if (!HandleAmount(FS.getFieldWidth(), /* field width */ 0,
1480                    startSpecifier, specifierLen)) {
1481    return false;
1482  }
1483
1484  if (!HandleAmount(FS.getPrecision(), /* precision */ 1,
1485                    startSpecifier, specifierLen)) {
1486    return false;
1487  }
1488
1489  if (!CS.consumesDataArgument()) {
1490    // FIXME: Technically specifying a precision or field width here
1491    // makes no sense.  Worth issuing a warning at some point.
1492    return true;
1493  }
1494
1495  // Consume the argument.
1496  unsigned argIndex = FS.getArgIndex();
1497  if (argIndex < NumDataArgs) {
1498    // The check to see if the argIndex is valid will come later.
1499    // We set the bit here because we may exit early from this
1500    // function if we encounter some other error.
1501    CoveredArgs.set(argIndex);
1502  }
1503
1504  // Check for using an Objective-C specific conversion specifier
1505  // in a non-ObjC literal.
1506  if (!IsObjCLiteral && CS.isObjCArg()) {
1507    return HandleInvalidPrintfConversionSpecifier(FS, startSpecifier,
1508                                                  specifierLen);
1509  }
1510
1511  // Check for invalid use of field width
1512  if (!FS.hasValidFieldWidth()) {
1513    HandleInvalidAmount(FS, FS.getFieldWidth(), /* field width */ 0,
1514        startSpecifier, specifierLen);
1515  }
1516
1517  // Check for invalid use of precision
1518  if (!FS.hasValidPrecision()) {
1519    HandleInvalidAmount(FS, FS.getPrecision(), /* precision */ 1,
1520        startSpecifier, specifierLen);
1521  }
1522
1523  // Check each flag does not conflict with any other component.
1524  if (!FS.hasValidThousandsGroupingPrefix())
1525    HandleFlag(FS, FS.hasThousandsGrouping(), startSpecifier, specifierLen);
1526  if (!FS.hasValidLeadingZeros())
1527    HandleFlag(FS, FS.hasLeadingZeros(), startSpecifier, specifierLen);
1528  if (!FS.hasValidPlusPrefix())
1529    HandleFlag(FS, FS.hasPlusPrefix(), startSpecifier, specifierLen);
1530  if (!FS.hasValidSpacePrefix())
1531    HandleFlag(FS, FS.hasSpacePrefix(), startSpecifier, specifierLen);
1532  if (!FS.hasValidAlternativeForm())
1533    HandleFlag(FS, FS.hasAlternativeForm(), startSpecifier, specifierLen);
1534  if (!FS.hasValidLeftJustified())
1535    HandleFlag(FS, FS.isLeftJustified(), startSpecifier, specifierLen);
1536
1537  // Check that flags are not ignored by another flag
1538  if (FS.hasSpacePrefix() && FS.hasPlusPrefix()) // ' ' ignored by '+'
1539    HandleIgnoredFlag(FS, FS.hasSpacePrefix(), FS.hasPlusPrefix(),
1540        startSpecifier, specifierLen);
1541  if (FS.hasLeadingZeros() && FS.isLeftJustified()) // '0' ignored by '-'
1542    HandleIgnoredFlag(FS, FS.hasLeadingZeros(), FS.isLeftJustified(),
1543            startSpecifier, specifierLen);
1544
1545  // Check the length modifier is valid with the given conversion specifier.
1546  const LengthModifier &LM = FS.getLengthModifier();
1547  if (!FS.hasValidLengthModifier())
1548    S.Diag(getLocationOfByte(LM.getStart()),
1549        diag::warn_format_nonsensical_length)
1550      << LM.toString() << CS.toString()
1551      << getSpecifierRange(startSpecifier, specifierLen)
1552      << FixItHint::CreateRemoval(getSpecifierRange(LM.getStart(),
1553          LM.getLength()));
1554
1555  // Are we using '%n'?
1556  if (CS.getKind() == ConversionSpecifier::nArg) {
1557    // Issue a warning about this being a possible security issue.
1558    S.Diag(getLocationOfByte(CS.getStart()), diag::warn_printf_write_back)
1559      << getSpecifierRange(startSpecifier, specifierLen);
1560    // Continue checking the other format specifiers.
1561    return true;
1562  }
1563
1564  // The remaining checks depend on the data arguments.
1565  if (HasVAListArg)
1566    return true;
1567
1568  if (!CheckNumArgs(FS, CS, startSpecifier, specifierLen, argIndex))
1569    return false;
1570
1571  // Now type check the data expression that matches the
1572  // format specifier.
1573  const Expr *Ex = getDataArg(argIndex);
1574  const analyze_printf::ArgTypeResult &ATR = FS.getArgType(S.Context);
1575  if (ATR.isValid() && !ATR.matchesType(S.Context, Ex->getType())) {
1576    // Check if we didn't match because of an implicit cast from a 'char'
1577    // or 'short' to an 'int'.  This is done because printf is a varargs
1578    // function.
1579    if (const ImplicitCastExpr *ICE = dyn_cast<ImplicitCastExpr>(Ex))
1580      if (ICE->getType() == S.Context.IntTy) {
1581        // All further checking is done on the subexpression.
1582        Ex = ICE->getSubExpr();
1583        if (ATR.matchesType(S.Context, Ex->getType()))
1584          return true;
1585      }
1586
1587    // We may be able to offer a FixItHint if it is a supported type.
1588    PrintfSpecifier fixedFS = FS;
1589    bool success = fixedFS.fixType(Ex->getType());
1590
1591    if (success) {
1592      // Get the fix string from the fixed format specifier
1593      llvm::SmallString<128> buf;
1594      llvm::raw_svector_ostream os(buf);
1595      fixedFS.toString(os);
1596
1597      // FIXME: getRepresentativeType() perhaps should return a string
1598      // instead of a QualType to better handle when the representative
1599      // type is 'wint_t' (which is defined in the system headers).
1600      S.Diag(getLocationOfByte(CS.getStart()),
1601          diag::warn_printf_conversion_argument_type_mismatch)
1602        << ATR.getRepresentativeType(S.Context) << Ex->getType()
1603        << getSpecifierRange(startSpecifier, specifierLen)
1604        << Ex->getSourceRange()
1605        << FixItHint::CreateReplacement(
1606            getSpecifierRange(startSpecifier, specifierLen),
1607            os.str());
1608    }
1609    else {
1610      S.Diag(getLocationOfByte(CS.getStart()),
1611             diag::warn_printf_conversion_argument_type_mismatch)
1612        << ATR.getRepresentativeType(S.Context) << Ex->getType()
1613        << getSpecifierRange(startSpecifier, specifierLen)
1614        << Ex->getSourceRange();
1615    }
1616  }
1617
1618  return true;
1619}
1620
1621//===--- CHECK: Scanf format string checking ------------------------------===//
1622
1623namespace {
1624class CheckScanfHandler : public CheckFormatHandler {
1625public:
1626  CheckScanfHandler(Sema &s, const StringLiteral *fexpr,
1627                    const Expr *origFormatExpr, unsigned firstDataArg,
1628                    unsigned numDataArgs, bool isObjCLiteral,
1629                    const char *beg, bool hasVAListArg,
1630                    const CallExpr *theCall, unsigned formatIdx)
1631  : CheckFormatHandler(s, fexpr, origFormatExpr, firstDataArg,
1632                       numDataArgs, isObjCLiteral, beg, hasVAListArg,
1633                       theCall, formatIdx) {}
1634
1635  bool HandleScanfSpecifier(const analyze_scanf::ScanfSpecifier &FS,
1636                            const char *startSpecifier,
1637                            unsigned specifierLen);
1638
1639  bool HandleInvalidScanfConversionSpecifier(
1640          const analyze_scanf::ScanfSpecifier &FS,
1641          const char *startSpecifier,
1642          unsigned specifierLen);
1643
1644  void HandleIncompleteScanList(const char *start, const char *end);
1645};
1646}
1647
1648void CheckScanfHandler::HandleIncompleteScanList(const char *start,
1649                                                 const char *end) {
1650  S.Diag(getLocationOfByte(end), diag::warn_scanf_scanlist_incomplete)
1651    << getSpecifierRange(start, end - start);
1652}
1653
1654bool CheckScanfHandler::HandleInvalidScanfConversionSpecifier(
1655                                        const analyze_scanf::ScanfSpecifier &FS,
1656                                        const char *startSpecifier,
1657                                        unsigned specifierLen) {
1658
1659  const analyze_scanf::ScanfConversionSpecifier &CS =
1660    FS.getConversionSpecifier();
1661
1662  return HandleInvalidConversionSpecifier(FS.getArgIndex(),
1663                                          getLocationOfByte(CS.getStart()),
1664                                          startSpecifier, specifierLen,
1665                                          CS.getStart(), CS.getLength());
1666}
1667
1668bool CheckScanfHandler::HandleScanfSpecifier(
1669                                       const analyze_scanf::ScanfSpecifier &FS,
1670                                       const char *startSpecifier,
1671                                       unsigned specifierLen) {
1672
1673  using namespace analyze_scanf;
1674  using namespace analyze_format_string;
1675
1676  const ScanfConversionSpecifier &CS = FS.getConversionSpecifier();
1677
1678  // Handle case where '%' and '*' don't consume an argument.  These shouldn't
1679  // be used to decide if we are using positional arguments consistently.
1680  if (FS.consumesDataArgument()) {
1681    if (atFirstArg) {
1682      atFirstArg = false;
1683      usesPositionalArgs = FS.usesPositionalArg();
1684    }
1685    else if (usesPositionalArgs != FS.usesPositionalArg()) {
1686      // Cannot mix-and-match positional and non-positional arguments.
1687      S.Diag(getLocationOfByte(CS.getStart()),
1688             diag::warn_format_mix_positional_nonpositional_args)
1689        << getSpecifierRange(startSpecifier, specifierLen);
1690      return false;
1691    }
1692  }
1693
1694  // Check if the field with is non-zero.
1695  const OptionalAmount &Amt = FS.getFieldWidth();
1696  if (Amt.getHowSpecified() == OptionalAmount::Constant) {
1697    if (Amt.getConstantAmount() == 0) {
1698      const CharSourceRange &R = getSpecifierRange(Amt.getStart(),
1699                                                   Amt.getConstantLength());
1700      S.Diag(getLocationOfByte(Amt.getStart()),
1701             diag::warn_scanf_nonzero_width)
1702        << R << FixItHint::CreateRemoval(R);
1703    }
1704  }
1705
1706  if (!FS.consumesDataArgument()) {
1707    // FIXME: Technically specifying a precision or field width here
1708    // makes no sense.  Worth issuing a warning at some point.
1709    return true;
1710  }
1711
1712  // Consume the argument.
1713  unsigned argIndex = FS.getArgIndex();
1714  if (argIndex < NumDataArgs) {
1715      // The check to see if the argIndex is valid will come later.
1716      // We set the bit here because we may exit early from this
1717      // function if we encounter some other error.
1718    CoveredArgs.set(argIndex);
1719  }
1720
1721  // Check the length modifier is valid with the given conversion specifier.
1722  const LengthModifier &LM = FS.getLengthModifier();
1723  if (!FS.hasValidLengthModifier()) {
1724    S.Diag(getLocationOfByte(LM.getStart()),
1725           diag::warn_format_nonsensical_length)
1726      << LM.toString() << CS.toString()
1727      << getSpecifierRange(startSpecifier, specifierLen)
1728      << FixItHint::CreateRemoval(getSpecifierRange(LM.getStart(),
1729                                                    LM.getLength()));
1730  }
1731
1732  // The remaining checks depend on the data arguments.
1733  if (HasVAListArg)
1734    return true;
1735
1736  if (!CheckNumArgs(FS, CS, startSpecifier, specifierLen, argIndex))
1737    return false;
1738
1739  // FIXME: Check that the argument type matches the format specifier.
1740
1741  return true;
1742}
1743
1744void Sema::CheckFormatString(const StringLiteral *FExpr,
1745                             const Expr *OrigFormatExpr,
1746                             const CallExpr *TheCall, bool HasVAListArg,
1747                             unsigned format_idx, unsigned firstDataArg,
1748                             bool isPrintf) {
1749
1750  // CHECK: is the format string a wide literal?
1751  if (FExpr->isWide()) {
1752    Diag(FExpr->getLocStart(),
1753         diag::warn_format_string_is_wide_literal)
1754    << OrigFormatExpr->getSourceRange();
1755    return;
1756  }
1757
1758  // Str - The format string.  NOTE: this is NOT null-terminated!
1759  llvm::StringRef StrRef = FExpr->getString();
1760  const char *Str = StrRef.data();
1761  unsigned StrLen = StrRef.size();
1762
1763  // CHECK: empty format string?
1764  if (StrLen == 0) {
1765    Diag(FExpr->getLocStart(), diag::warn_empty_format_string)
1766    << OrigFormatExpr->getSourceRange();
1767    return;
1768  }
1769
1770  if (isPrintf) {
1771    CheckPrintfHandler H(*this, FExpr, OrigFormatExpr, firstDataArg,
1772                         TheCall->getNumArgs() - firstDataArg,
1773                         isa<ObjCStringLiteral>(OrigFormatExpr), Str,
1774                         HasVAListArg, TheCall, format_idx);
1775
1776    if (!analyze_format_string::ParsePrintfString(H, Str, Str + StrLen))
1777      H.DoneProcessing();
1778  }
1779  else {
1780    CheckScanfHandler H(*this, FExpr, OrigFormatExpr, firstDataArg,
1781                        TheCall->getNumArgs() - firstDataArg,
1782                        isa<ObjCStringLiteral>(OrigFormatExpr), Str,
1783                        HasVAListArg, TheCall, format_idx);
1784
1785    if (!analyze_format_string::ParseScanfString(H, Str, Str + StrLen))
1786      H.DoneProcessing();
1787  }
1788}
1789
1790//===--- CHECK: Return Address of Stack Variable --------------------------===//
1791
1792static Expr *EvalVal(Expr *E, llvm::SmallVectorImpl<DeclRefExpr *> &refVars);
1793static Expr *EvalAddr(Expr* E, llvm::SmallVectorImpl<DeclRefExpr *> &refVars);
1794
1795/// CheckReturnStackAddr - Check if a return statement returns the address
1796///   of a stack variable.
1797void
1798Sema::CheckReturnStackAddr(Expr *RetValExp, QualType lhsType,
1799                           SourceLocation ReturnLoc) {
1800
1801  Expr *stackE = 0;
1802  llvm::SmallVector<DeclRefExpr *, 8> refVars;
1803
1804  // Perform checking for returned stack addresses, local blocks,
1805  // label addresses or references to temporaries.
1806  if (lhsType->isPointerType() || lhsType->isBlockPointerType()) {
1807    stackE = EvalAddr(RetValExp, refVars);
1808  } else if (lhsType->isReferenceType()) {
1809    stackE = EvalVal(RetValExp, refVars);
1810  }
1811
1812  if (stackE == 0)
1813    return; // Nothing suspicious was found.
1814
1815  SourceLocation diagLoc;
1816  SourceRange diagRange;
1817  if (refVars.empty()) {
1818    diagLoc = stackE->getLocStart();
1819    diagRange = stackE->getSourceRange();
1820  } else {
1821    // We followed through a reference variable. 'stackE' contains the
1822    // problematic expression but we will warn at the return statement pointing
1823    // at the reference variable. We will later display the "trail" of
1824    // reference variables using notes.
1825    diagLoc = refVars[0]->getLocStart();
1826    diagRange = refVars[0]->getSourceRange();
1827  }
1828
1829  if (DeclRefExpr *DR = dyn_cast<DeclRefExpr>(stackE)) { //address of local var.
1830    Diag(diagLoc, lhsType->isReferenceType() ? diag::warn_ret_stack_ref
1831                                             : diag::warn_ret_stack_addr)
1832     << DR->getDecl()->getDeclName() << diagRange;
1833  } else if (isa<BlockExpr>(stackE)) { // local block.
1834    Diag(diagLoc, diag::err_ret_local_block) << diagRange;
1835  } else if (isa<AddrLabelExpr>(stackE)) { // address of label.
1836    Diag(diagLoc, diag::warn_ret_addr_label) << diagRange;
1837  } else { // local temporary.
1838    Diag(diagLoc, lhsType->isReferenceType() ? diag::warn_ret_local_temp_ref
1839                                             : diag::warn_ret_local_temp_addr)
1840     << diagRange;
1841  }
1842
1843  // Display the "trail" of reference variables that we followed until we
1844  // found the problematic expression using notes.
1845  for (unsigned i = 0, e = refVars.size(); i != e; ++i) {
1846    VarDecl *VD = cast<VarDecl>(refVars[i]->getDecl());
1847    // If this var binds to another reference var, show the range of the next
1848    // var, otherwise the var binds to the problematic expression, in which case
1849    // show the range of the expression.
1850    SourceRange range = (i < e-1) ? refVars[i+1]->getSourceRange()
1851                                  : stackE->getSourceRange();
1852    Diag(VD->getLocation(), diag::note_ref_var_local_bind)
1853      << VD->getDeclName() << range;
1854  }
1855}
1856
1857/// EvalAddr - EvalAddr and EvalVal are mutually recursive functions that
1858///  check if the expression in a return statement evaluates to an address
1859///  to a location on the stack, a local block, an address of a label, or a
1860///  reference to local temporary. The recursion is used to traverse the
1861///  AST of the return expression, with recursion backtracking when we
1862///  encounter a subexpression that (1) clearly does not lead to one of the
1863///  above problematic expressions (2) is something we cannot determine leads to
1864///  a problematic expression based on such local checking.
1865///
1866///  Both EvalAddr and EvalVal follow through reference variables to evaluate
1867///  the expression that they point to. Such variables are added to the
1868///  'refVars' vector so that we know what the reference variable "trail" was.
1869///
1870///  EvalAddr processes expressions that are pointers that are used as
1871///  references (and not L-values).  EvalVal handles all other values.
1872///  At the base case of the recursion is a check for the above problematic
1873///  expressions.
1874///
1875///  This implementation handles:
1876///
1877///   * pointer-to-pointer casts
1878///   * implicit conversions from array references to pointers
1879///   * taking the address of fields
1880///   * arbitrary interplay between "&" and "*" operators
1881///   * pointer arithmetic from an address of a stack variable
1882///   * taking the address of an array element where the array is on the stack
1883static Expr *EvalAddr(Expr *E, llvm::SmallVectorImpl<DeclRefExpr *> &refVars) {
1884  if (E->isTypeDependent())
1885      return NULL;
1886
1887  // We should only be called for evaluating pointer expressions.
1888  assert((E->getType()->isAnyPointerType() ||
1889          E->getType()->isBlockPointerType() ||
1890          E->getType()->isObjCQualifiedIdType()) &&
1891         "EvalAddr only works on pointers");
1892
1893  // Our "symbolic interpreter" is just a dispatch off the currently
1894  // viewed AST node.  We then recursively traverse the AST by calling
1895  // EvalAddr and EvalVal appropriately.
1896  switch (E->getStmtClass()) {
1897  case Stmt::ParenExprClass:
1898    // Ignore parentheses.
1899    return EvalAddr(cast<ParenExpr>(E)->getSubExpr(), refVars);
1900
1901  case Stmt::DeclRefExprClass: {
1902    DeclRefExpr *DR = cast<DeclRefExpr>(E);
1903
1904    if (VarDecl *V = dyn_cast<VarDecl>(DR->getDecl()))
1905      // If this is a reference variable, follow through to the expression that
1906      // it points to.
1907      if (V->hasLocalStorage() &&
1908          V->getType()->isReferenceType() && V->hasInit()) {
1909        // Add the reference variable to the "trail".
1910        refVars.push_back(DR);
1911        return EvalAddr(V->getInit(), refVars);
1912      }
1913
1914    return NULL;
1915  }
1916
1917  case Stmt::UnaryOperatorClass: {
1918    // The only unary operator that make sense to handle here
1919    // is AddrOf.  All others don't make sense as pointers.
1920    UnaryOperator *U = cast<UnaryOperator>(E);
1921
1922    if (U->getOpcode() == UO_AddrOf)
1923      return EvalVal(U->getSubExpr(), refVars);
1924    else
1925      return NULL;
1926  }
1927
1928  case Stmt::BinaryOperatorClass: {
1929    // Handle pointer arithmetic.  All other binary operators are not valid
1930    // in this context.
1931    BinaryOperator *B = cast<BinaryOperator>(E);
1932    BinaryOperatorKind op = B->getOpcode();
1933
1934    if (op != BO_Add && op != BO_Sub)
1935      return NULL;
1936
1937    Expr *Base = B->getLHS();
1938
1939    // Determine which argument is the real pointer base.  It could be
1940    // the RHS argument instead of the LHS.
1941    if (!Base->getType()->isPointerType()) Base = B->getRHS();
1942
1943    assert (Base->getType()->isPointerType());
1944    return EvalAddr(Base, refVars);
1945  }
1946
1947  // For conditional operators we need to see if either the LHS or RHS are
1948  // valid DeclRefExpr*s.  If one of them is valid, we return it.
1949  case Stmt::ConditionalOperatorClass: {
1950    ConditionalOperator *C = cast<ConditionalOperator>(E);
1951
1952    // Handle the GNU extension for missing LHS.
1953    if (Expr *lhsExpr = C->getLHS()) {
1954    // In C++, we can have a throw-expression, which has 'void' type.
1955      if (!lhsExpr->getType()->isVoidType())
1956        if (Expr* LHS = EvalAddr(lhsExpr, refVars))
1957          return LHS;
1958    }
1959
1960    // In C++, we can have a throw-expression, which has 'void' type.
1961    if (C->getRHS()->getType()->isVoidType())
1962      return NULL;
1963
1964    return EvalAddr(C->getRHS(), refVars);
1965  }
1966
1967  case Stmt::BlockExprClass:
1968    if (cast<BlockExpr>(E)->getBlockDecl()->hasCaptures())
1969      return E; // local block.
1970    return NULL;
1971
1972  case Stmt::AddrLabelExprClass:
1973    return E; // address of label.
1974
1975  // For casts, we need to handle conversions from arrays to
1976  // pointer values, and pointer-to-pointer conversions.
1977  case Stmt::ImplicitCastExprClass:
1978  case Stmt::CStyleCastExprClass:
1979  case Stmt::CXXFunctionalCastExprClass: {
1980    Expr* SubExpr = cast<CastExpr>(E)->getSubExpr();
1981    QualType T = SubExpr->getType();
1982
1983    if (SubExpr->getType()->isPointerType() ||
1984        SubExpr->getType()->isBlockPointerType() ||
1985        SubExpr->getType()->isObjCQualifiedIdType())
1986      return EvalAddr(SubExpr, refVars);
1987    else if (T->isArrayType())
1988      return EvalVal(SubExpr, refVars);
1989    else
1990      return 0;
1991  }
1992
1993  // C++ casts.  For dynamic casts, static casts, and const casts, we
1994  // are always converting from a pointer-to-pointer, so we just blow
1995  // through the cast.  In the case the dynamic cast doesn't fail (and
1996  // return NULL), we take the conservative route and report cases
1997  // where we return the address of a stack variable.  For Reinterpre
1998  // FIXME: The comment about is wrong; we're not always converting
1999  // from pointer to pointer. I'm guessing that this code should also
2000  // handle references to objects.
2001  case Stmt::CXXStaticCastExprClass:
2002  case Stmt::CXXDynamicCastExprClass:
2003  case Stmt::CXXConstCastExprClass:
2004  case Stmt::CXXReinterpretCastExprClass: {
2005      Expr *S = cast<CXXNamedCastExpr>(E)->getSubExpr();
2006      if (S->getType()->isPointerType() || S->getType()->isBlockPointerType())
2007        return EvalAddr(S, refVars);
2008      else
2009        return NULL;
2010  }
2011
2012  // Everything else: we simply don't reason about them.
2013  default:
2014    return NULL;
2015  }
2016}
2017
2018
2019///  EvalVal - This function is complements EvalAddr in the mutual recursion.
2020///   See the comments for EvalAddr for more details.
2021static Expr *EvalVal(Expr *E, llvm::SmallVectorImpl<DeclRefExpr *> &refVars) {
2022do {
2023  // We should only be called for evaluating non-pointer expressions, or
2024  // expressions with a pointer type that are not used as references but instead
2025  // are l-values (e.g., DeclRefExpr with a pointer type).
2026
2027  // Our "symbolic interpreter" is just a dispatch off the currently
2028  // viewed AST node.  We then recursively traverse the AST by calling
2029  // EvalAddr and EvalVal appropriately.
2030  switch (E->getStmtClass()) {
2031  case Stmt::ImplicitCastExprClass: {
2032    ImplicitCastExpr *IE = cast<ImplicitCastExpr>(E);
2033    if (IE->getValueKind() == VK_LValue) {
2034      E = IE->getSubExpr();
2035      continue;
2036    }
2037    return NULL;
2038  }
2039
2040  case Stmt::DeclRefExprClass: {
2041    // When we hit a DeclRefExpr we are looking at code that refers to a
2042    // variable's name. If it's not a reference variable we check if it has
2043    // local storage within the function, and if so, return the expression.
2044    DeclRefExpr *DR = cast<DeclRefExpr>(E);
2045
2046    if (VarDecl *V = dyn_cast<VarDecl>(DR->getDecl()))
2047      if (V->hasLocalStorage()) {
2048        if (!V->getType()->isReferenceType())
2049          return DR;
2050
2051        // Reference variable, follow through to the expression that
2052        // it points to.
2053        if (V->hasInit()) {
2054          // Add the reference variable to the "trail".
2055          refVars.push_back(DR);
2056          return EvalVal(V->getInit(), refVars);
2057        }
2058      }
2059
2060    return NULL;
2061  }
2062
2063  case Stmt::ParenExprClass: {
2064    // Ignore parentheses.
2065    E = cast<ParenExpr>(E)->getSubExpr();
2066    continue;
2067  }
2068
2069  case Stmt::UnaryOperatorClass: {
2070    // The only unary operator that make sense to handle here
2071    // is Deref.  All others don't resolve to a "name."  This includes
2072    // handling all sorts of rvalues passed to a unary operator.
2073    UnaryOperator *U = cast<UnaryOperator>(E);
2074
2075    if (U->getOpcode() == UO_Deref)
2076      return EvalAddr(U->getSubExpr(), refVars);
2077
2078    return NULL;
2079  }
2080
2081  case Stmt::ArraySubscriptExprClass: {
2082    // Array subscripts are potential references to data on the stack.  We
2083    // retrieve the DeclRefExpr* for the array variable if it indeed
2084    // has local storage.
2085    return EvalAddr(cast<ArraySubscriptExpr>(E)->getBase(), refVars);
2086  }
2087
2088  case Stmt::ConditionalOperatorClass: {
2089    // For conditional operators we need to see if either the LHS or RHS are
2090    // non-NULL Expr's.  If one is non-NULL, we return it.
2091    ConditionalOperator *C = cast<ConditionalOperator>(E);
2092
2093    // Handle the GNU extension for missing LHS.
2094    if (Expr *lhsExpr = C->getLHS())
2095      if (Expr *LHS = EvalVal(lhsExpr, refVars))
2096        return LHS;
2097
2098    return EvalVal(C->getRHS(), refVars);
2099  }
2100
2101  // Accesses to members are potential references to data on the stack.
2102  case Stmt::MemberExprClass: {
2103    MemberExpr *M = cast<MemberExpr>(E);
2104
2105    // Check for indirect access.  We only want direct field accesses.
2106    if (M->isArrow())
2107      return NULL;
2108
2109    // Check whether the member type is itself a reference, in which case
2110    // we're not going to refer to the member, but to what the member refers to.
2111    if (M->getMemberDecl()->getType()->isReferenceType())
2112      return NULL;
2113
2114    return EvalVal(M->getBase(), refVars);
2115  }
2116
2117  default:
2118    // Check that we don't return or take the address of a reference to a
2119    // temporary. This is only useful in C++.
2120    if (!E->isTypeDependent() && E->isRValue())
2121      return E;
2122
2123    // Everything else: we simply don't reason about them.
2124    return NULL;
2125  }
2126} while (true);
2127}
2128
2129//===--- CHECK: Floating-Point comparisons (-Wfloat-equal) ---------------===//
2130
2131/// Check for comparisons of floating point operands using != and ==.
2132/// Issue a warning if these are no self-comparisons, as they are not likely
2133/// to do what the programmer intended.
2134void Sema::CheckFloatComparison(SourceLocation loc, Expr* lex, Expr *rex) {
2135  bool EmitWarning = true;
2136
2137  Expr* LeftExprSansParen = lex->IgnoreParenImpCasts();
2138  Expr* RightExprSansParen = rex->IgnoreParenImpCasts();
2139
2140  // Special case: check for x == x (which is OK).
2141  // Do not emit warnings for such cases.
2142  if (DeclRefExpr* DRL = dyn_cast<DeclRefExpr>(LeftExprSansParen))
2143    if (DeclRefExpr* DRR = dyn_cast<DeclRefExpr>(RightExprSansParen))
2144      if (DRL->getDecl() == DRR->getDecl())
2145        EmitWarning = false;
2146
2147
2148  // Special case: check for comparisons against literals that can be exactly
2149  //  represented by APFloat.  In such cases, do not emit a warning.  This
2150  //  is a heuristic: often comparison against such literals are used to
2151  //  detect if a value in a variable has not changed.  This clearly can
2152  //  lead to false negatives.
2153  if (EmitWarning) {
2154    if (FloatingLiteral* FLL = dyn_cast<FloatingLiteral>(LeftExprSansParen)) {
2155      if (FLL->isExact())
2156        EmitWarning = false;
2157    } else
2158      if (FloatingLiteral* FLR = dyn_cast<FloatingLiteral>(RightExprSansParen)){
2159        if (FLR->isExact())
2160          EmitWarning = false;
2161    }
2162  }
2163
2164  // Check for comparisons with builtin types.
2165  if (EmitWarning)
2166    if (CallExpr* CL = dyn_cast<CallExpr>(LeftExprSansParen))
2167      if (CL->isBuiltinCall(Context))
2168        EmitWarning = false;
2169
2170  if (EmitWarning)
2171    if (CallExpr* CR = dyn_cast<CallExpr>(RightExprSansParen))
2172      if (CR->isBuiltinCall(Context))
2173        EmitWarning = false;
2174
2175  // Emit the diagnostic.
2176  if (EmitWarning)
2177    Diag(loc, diag::warn_floatingpoint_eq)
2178      << lex->getSourceRange() << rex->getSourceRange();
2179}
2180
2181//===--- CHECK: Integer mixed-sign comparisons (-Wsign-compare) --------===//
2182//===--- CHECK: Lossy implicit conversions (-Wconversion) --------------===//
2183
2184namespace {
2185
2186/// Structure recording the 'active' range of an integer-valued
2187/// expression.
2188struct IntRange {
2189  /// The number of bits active in the int.
2190  unsigned Width;
2191
2192  /// True if the int is known not to have negative values.
2193  bool NonNegative;
2194
2195  IntRange(unsigned Width, bool NonNegative)
2196    : Width(Width), NonNegative(NonNegative)
2197  {}
2198
2199  /// Returns the range of the bool type.
2200  static IntRange forBoolType() {
2201    return IntRange(1, true);
2202  }
2203
2204  /// Returns the range of an opaque value of the given integral type.
2205  static IntRange forValueOfType(ASTContext &C, QualType T) {
2206    return forValueOfCanonicalType(C,
2207                          T->getCanonicalTypeInternal().getTypePtr());
2208  }
2209
2210  /// Returns the range of an opaque value of a canonical integral type.
2211  static IntRange forValueOfCanonicalType(ASTContext &C, const Type *T) {
2212    assert(T->isCanonicalUnqualified());
2213
2214    if (const VectorType *VT = dyn_cast<VectorType>(T))
2215      T = VT->getElementType().getTypePtr();
2216    if (const ComplexType *CT = dyn_cast<ComplexType>(T))
2217      T = CT->getElementType().getTypePtr();
2218
2219    // For enum types, use the known bit width of the enumerators.
2220    if (const EnumType *ET = dyn_cast<EnumType>(T)) {
2221      EnumDecl *Enum = ET->getDecl();
2222      if (!Enum->isDefinition())
2223        return IntRange(C.getIntWidth(QualType(T, 0)), false);
2224
2225      unsigned NumPositive = Enum->getNumPositiveBits();
2226      unsigned NumNegative = Enum->getNumNegativeBits();
2227
2228      return IntRange(std::max(NumPositive, NumNegative), NumNegative == 0);
2229    }
2230
2231    const BuiltinType *BT = cast<BuiltinType>(T);
2232    assert(BT->isInteger());
2233
2234    return IntRange(C.getIntWidth(QualType(T, 0)), BT->isUnsignedInteger());
2235  }
2236
2237  /// Returns the "target" range of a canonical integral type, i.e.
2238  /// the range of values expressible in the type.
2239  ///
2240  /// This matches forValueOfCanonicalType except that enums have the
2241  /// full range of their type, not the range of their enumerators.
2242  static IntRange forTargetOfCanonicalType(ASTContext &C, const Type *T) {
2243    assert(T->isCanonicalUnqualified());
2244
2245    if (const VectorType *VT = dyn_cast<VectorType>(T))
2246      T = VT->getElementType().getTypePtr();
2247    if (const ComplexType *CT = dyn_cast<ComplexType>(T))
2248      T = CT->getElementType().getTypePtr();
2249    if (const EnumType *ET = dyn_cast<EnumType>(T))
2250      T = ET->getDecl()->getIntegerType().getTypePtr();
2251
2252    const BuiltinType *BT = cast<BuiltinType>(T);
2253    assert(BT->isInteger());
2254
2255    return IntRange(C.getIntWidth(QualType(T, 0)), BT->isUnsignedInteger());
2256  }
2257
2258  /// Returns the supremum of two ranges: i.e. their conservative merge.
2259  static IntRange join(IntRange L, IntRange R) {
2260    return IntRange(std::max(L.Width, R.Width),
2261                    L.NonNegative && R.NonNegative);
2262  }
2263
2264  /// Returns the infinum of two ranges: i.e. their aggressive merge.
2265  static IntRange meet(IntRange L, IntRange R) {
2266    return IntRange(std::min(L.Width, R.Width),
2267                    L.NonNegative || R.NonNegative);
2268  }
2269};
2270
2271IntRange GetValueRange(ASTContext &C, llvm::APSInt &value, unsigned MaxWidth) {
2272  if (value.isSigned() && value.isNegative())
2273    return IntRange(value.getMinSignedBits(), false);
2274
2275  if (value.getBitWidth() > MaxWidth)
2276    value = value.trunc(MaxWidth);
2277
2278  // isNonNegative() just checks the sign bit without considering
2279  // signedness.
2280  return IntRange(value.getActiveBits(), true);
2281}
2282
2283IntRange GetValueRange(ASTContext &C, APValue &result, QualType Ty,
2284                       unsigned MaxWidth) {
2285  if (result.isInt())
2286    return GetValueRange(C, result.getInt(), MaxWidth);
2287
2288  if (result.isVector()) {
2289    IntRange R = GetValueRange(C, result.getVectorElt(0), Ty, MaxWidth);
2290    for (unsigned i = 1, e = result.getVectorLength(); i != e; ++i) {
2291      IntRange El = GetValueRange(C, result.getVectorElt(i), Ty, MaxWidth);
2292      R = IntRange::join(R, El);
2293    }
2294    return R;
2295  }
2296
2297  if (result.isComplexInt()) {
2298    IntRange R = GetValueRange(C, result.getComplexIntReal(), MaxWidth);
2299    IntRange I = GetValueRange(C, result.getComplexIntImag(), MaxWidth);
2300    return IntRange::join(R, I);
2301  }
2302
2303  // This can happen with lossless casts to intptr_t of "based" lvalues.
2304  // Assume it might use arbitrary bits.
2305  // FIXME: The only reason we need to pass the type in here is to get
2306  // the sign right on this one case.  It would be nice if APValue
2307  // preserved this.
2308  assert(result.isLValue());
2309  return IntRange(MaxWidth, Ty->isUnsignedIntegerType());
2310}
2311
2312/// Pseudo-evaluate the given integer expression, estimating the
2313/// range of values it might take.
2314///
2315/// \param MaxWidth - the width to which the value will be truncated
2316IntRange GetExprRange(ASTContext &C, Expr *E, unsigned MaxWidth) {
2317  E = E->IgnoreParens();
2318
2319  // Try a full evaluation first.
2320  Expr::EvalResult result;
2321  if (E->Evaluate(result, C))
2322    return GetValueRange(C, result.Val, E->getType(), MaxWidth);
2323
2324  // I think we only want to look through implicit casts here; if the
2325  // user has an explicit widening cast, we should treat the value as
2326  // being of the new, wider type.
2327  if (ImplicitCastExpr *CE = dyn_cast<ImplicitCastExpr>(E)) {
2328    if (CE->getCastKind() == CK_NoOp)
2329      return GetExprRange(C, CE->getSubExpr(), MaxWidth);
2330
2331    IntRange OutputTypeRange = IntRange::forValueOfType(C, CE->getType());
2332
2333    bool isIntegerCast = (CE->getCastKind() == CK_IntegralCast);
2334
2335    // Assume that non-integer casts can span the full range of the type.
2336    if (!isIntegerCast)
2337      return OutputTypeRange;
2338
2339    IntRange SubRange
2340      = GetExprRange(C, CE->getSubExpr(),
2341                     std::min(MaxWidth, OutputTypeRange.Width));
2342
2343    // Bail out if the subexpr's range is as wide as the cast type.
2344    if (SubRange.Width >= OutputTypeRange.Width)
2345      return OutputTypeRange;
2346
2347    // Otherwise, we take the smaller width, and we're non-negative if
2348    // either the output type or the subexpr is.
2349    return IntRange(SubRange.Width,
2350                    SubRange.NonNegative || OutputTypeRange.NonNegative);
2351  }
2352
2353  if (ConditionalOperator *CO = dyn_cast<ConditionalOperator>(E)) {
2354    // If we can fold the condition, just take that operand.
2355    bool CondResult;
2356    if (CO->getCond()->EvaluateAsBooleanCondition(CondResult, C))
2357      return GetExprRange(C, CondResult ? CO->getTrueExpr()
2358                                        : CO->getFalseExpr(),
2359                          MaxWidth);
2360
2361    // Otherwise, conservatively merge.
2362    IntRange L = GetExprRange(C, CO->getTrueExpr(), MaxWidth);
2363    IntRange R = GetExprRange(C, CO->getFalseExpr(), MaxWidth);
2364    return IntRange::join(L, R);
2365  }
2366
2367  if (BinaryOperator *BO = dyn_cast<BinaryOperator>(E)) {
2368    switch (BO->getOpcode()) {
2369
2370    // Boolean-valued operations are single-bit and positive.
2371    case BO_LAnd:
2372    case BO_LOr:
2373    case BO_LT:
2374    case BO_GT:
2375    case BO_LE:
2376    case BO_GE:
2377    case BO_EQ:
2378    case BO_NE:
2379      return IntRange::forBoolType();
2380
2381    // The type of these compound assignments is the type of the LHS,
2382    // so the RHS is not necessarily an integer.
2383    case BO_MulAssign:
2384    case BO_DivAssign:
2385    case BO_RemAssign:
2386    case BO_AddAssign:
2387    case BO_SubAssign:
2388      return IntRange::forValueOfType(C, E->getType());
2389
2390    // Operations with opaque sources are black-listed.
2391    case BO_PtrMemD:
2392    case BO_PtrMemI:
2393      return IntRange::forValueOfType(C, E->getType());
2394
2395    // Bitwise-and uses the *infinum* of the two source ranges.
2396    case BO_And:
2397    case BO_AndAssign:
2398      return IntRange::meet(GetExprRange(C, BO->getLHS(), MaxWidth),
2399                            GetExprRange(C, BO->getRHS(), MaxWidth));
2400
2401    // Left shift gets black-listed based on a judgement call.
2402    case BO_Shl:
2403      // ...except that we want to treat '1 << (blah)' as logically
2404      // positive.  It's an important idiom.
2405      if (IntegerLiteral *I
2406            = dyn_cast<IntegerLiteral>(BO->getLHS()->IgnoreParenCasts())) {
2407        if (I->getValue() == 1) {
2408          IntRange R = IntRange::forValueOfType(C, E->getType());
2409          return IntRange(R.Width, /*NonNegative*/ true);
2410        }
2411      }
2412      // fallthrough
2413
2414    case BO_ShlAssign:
2415      return IntRange::forValueOfType(C, E->getType());
2416
2417    // Right shift by a constant can narrow its left argument.
2418    case BO_Shr:
2419    case BO_ShrAssign: {
2420      IntRange L = GetExprRange(C, BO->getLHS(), MaxWidth);
2421
2422      // If the shift amount is a positive constant, drop the width by
2423      // that much.
2424      llvm::APSInt shift;
2425      if (BO->getRHS()->isIntegerConstantExpr(shift, C) &&
2426          shift.isNonNegative()) {
2427        unsigned zext = shift.getZExtValue();
2428        if (zext >= L.Width)
2429          L.Width = (L.NonNegative ? 0 : 1);
2430        else
2431          L.Width -= zext;
2432      }
2433
2434      return L;
2435    }
2436
2437    // Comma acts as its right operand.
2438    case BO_Comma:
2439      return GetExprRange(C, BO->getRHS(), MaxWidth);
2440
2441    // Black-list pointer subtractions.
2442    case BO_Sub:
2443      if (BO->getLHS()->getType()->isPointerType())
2444        return IntRange::forValueOfType(C, E->getType());
2445      // fallthrough
2446
2447    default:
2448      break;
2449    }
2450
2451    // Treat every other operator as if it were closed on the
2452    // narrowest type that encompasses both operands.
2453    IntRange L = GetExprRange(C, BO->getLHS(), MaxWidth);
2454    IntRange R = GetExprRange(C, BO->getRHS(), MaxWidth);
2455    return IntRange::join(L, R);
2456  }
2457
2458  if (UnaryOperator *UO = dyn_cast<UnaryOperator>(E)) {
2459    switch (UO->getOpcode()) {
2460    // Boolean-valued operations are white-listed.
2461    case UO_LNot:
2462      return IntRange::forBoolType();
2463
2464    // Operations with opaque sources are black-listed.
2465    case UO_Deref:
2466    case UO_AddrOf: // should be impossible
2467      return IntRange::forValueOfType(C, E->getType());
2468
2469    default:
2470      return GetExprRange(C, UO->getSubExpr(), MaxWidth);
2471    }
2472  }
2473
2474  if (dyn_cast<OffsetOfExpr>(E)) {
2475    IntRange::forValueOfType(C, E->getType());
2476  }
2477
2478  FieldDecl *BitField = E->getBitField();
2479  if (BitField) {
2480    llvm::APSInt BitWidthAP = BitField->getBitWidth()->EvaluateAsInt(C);
2481    unsigned BitWidth = BitWidthAP.getZExtValue();
2482
2483    return IntRange(BitWidth, BitField->getType()->isUnsignedIntegerType());
2484  }
2485
2486  return IntRange::forValueOfType(C, E->getType());
2487}
2488
2489IntRange GetExprRange(ASTContext &C, Expr *E) {
2490  return GetExprRange(C, E, C.getIntWidth(E->getType()));
2491}
2492
2493/// Checks whether the given value, which currently has the given
2494/// source semantics, has the same value when coerced through the
2495/// target semantics.
2496bool IsSameFloatAfterCast(const llvm::APFloat &value,
2497                          const llvm::fltSemantics &Src,
2498                          const llvm::fltSemantics &Tgt) {
2499  llvm::APFloat truncated = value;
2500
2501  bool ignored;
2502  truncated.convert(Src, llvm::APFloat::rmNearestTiesToEven, &ignored);
2503  truncated.convert(Tgt, llvm::APFloat::rmNearestTiesToEven, &ignored);
2504
2505  return truncated.bitwiseIsEqual(value);
2506}
2507
2508/// Checks whether the given value, which currently has the given
2509/// source semantics, has the same value when coerced through the
2510/// target semantics.
2511///
2512/// The value might be a vector of floats (or a complex number).
2513bool IsSameFloatAfterCast(const APValue &value,
2514                          const llvm::fltSemantics &Src,
2515                          const llvm::fltSemantics &Tgt) {
2516  if (value.isFloat())
2517    return IsSameFloatAfterCast(value.getFloat(), Src, Tgt);
2518
2519  if (value.isVector()) {
2520    for (unsigned i = 0, e = value.getVectorLength(); i != e; ++i)
2521      if (!IsSameFloatAfterCast(value.getVectorElt(i), Src, Tgt))
2522        return false;
2523    return true;
2524  }
2525
2526  assert(value.isComplexFloat());
2527  return (IsSameFloatAfterCast(value.getComplexFloatReal(), Src, Tgt) &&
2528          IsSameFloatAfterCast(value.getComplexFloatImag(), Src, Tgt));
2529}
2530
2531void AnalyzeImplicitConversions(Sema &S, Expr *E, SourceLocation CC);
2532
2533static bool IsZero(Sema &S, Expr *E) {
2534  // Suppress cases where we are comparing against an enum constant.
2535  if (const DeclRefExpr *DR =
2536      dyn_cast<DeclRefExpr>(E->IgnoreParenImpCasts()))
2537    if (isa<EnumConstantDecl>(DR->getDecl()))
2538      return false;
2539
2540  // Suppress cases where the '0' value is expanded from a macro.
2541  if (E->getLocStart().isMacroID())
2542    return false;
2543
2544  llvm::APSInt Value;
2545  return E->isIntegerConstantExpr(Value, S.Context) && Value == 0;
2546}
2547
2548static bool HasEnumType(Expr *E) {
2549  // Strip off implicit integral promotions.
2550  while (ImplicitCastExpr *ICE = dyn_cast<ImplicitCastExpr>(E)) {
2551    if (ICE->getCastKind() != CK_IntegralCast &&
2552        ICE->getCastKind() != CK_NoOp)
2553      break;
2554    E = ICE->getSubExpr();
2555  }
2556
2557  return E->getType()->isEnumeralType();
2558}
2559
2560void CheckTrivialUnsignedComparison(Sema &S, BinaryOperator *E) {
2561  BinaryOperatorKind op = E->getOpcode();
2562  if (E->isValueDependent())
2563    return;
2564
2565  if (op == BO_LT && IsZero(S, E->getRHS())) {
2566    S.Diag(E->getOperatorLoc(), diag::warn_lunsigned_always_true_comparison)
2567      << "< 0" << "false" << HasEnumType(E->getLHS())
2568      << E->getLHS()->getSourceRange() << E->getRHS()->getSourceRange();
2569  } else if (op == BO_GE && IsZero(S, E->getRHS())) {
2570    S.Diag(E->getOperatorLoc(), diag::warn_lunsigned_always_true_comparison)
2571      << ">= 0" << "true" << HasEnumType(E->getLHS())
2572      << E->getLHS()->getSourceRange() << E->getRHS()->getSourceRange();
2573  } else if (op == BO_GT && IsZero(S, E->getLHS())) {
2574    S.Diag(E->getOperatorLoc(), diag::warn_runsigned_always_true_comparison)
2575      << "0 >" << "false" << HasEnumType(E->getRHS())
2576      << E->getLHS()->getSourceRange() << E->getRHS()->getSourceRange();
2577  } else if (op == BO_LE && IsZero(S, E->getLHS())) {
2578    S.Diag(E->getOperatorLoc(), diag::warn_runsigned_always_true_comparison)
2579      << "0 <=" << "true" << HasEnumType(E->getRHS())
2580      << E->getLHS()->getSourceRange() << E->getRHS()->getSourceRange();
2581  }
2582}
2583
2584/// Analyze the operands of the given comparison.  Implements the
2585/// fallback case from AnalyzeComparison.
2586void AnalyzeImpConvsInComparison(Sema &S, BinaryOperator *E) {
2587  AnalyzeImplicitConversions(S, E->getLHS(), E->getOperatorLoc());
2588  AnalyzeImplicitConversions(S, E->getRHS(), E->getOperatorLoc());
2589}
2590
2591/// \brief Implements -Wsign-compare.
2592///
2593/// \param lex the left-hand expression
2594/// \param rex the right-hand expression
2595/// \param OpLoc the location of the joining operator
2596/// \param BinOpc binary opcode or 0
2597void AnalyzeComparison(Sema &S, BinaryOperator *E) {
2598  // The type the comparison is being performed in.
2599  QualType T = E->getLHS()->getType();
2600  assert(S.Context.hasSameUnqualifiedType(T, E->getRHS()->getType())
2601         && "comparison with mismatched types");
2602
2603  // We don't do anything special if this isn't an unsigned integral
2604  // comparison:  we're only interested in integral comparisons, and
2605  // signed comparisons only happen in cases we don't care to warn about.
2606  //
2607  // We also don't care about value-dependent expressions or expressions
2608  // whose result is a constant.
2609  if (!T->hasUnsignedIntegerRepresentation()
2610      || E->isValueDependent() || E->isIntegerConstantExpr(S.Context))
2611    return AnalyzeImpConvsInComparison(S, E);
2612
2613  Expr *lex = E->getLHS()->IgnoreParenImpCasts();
2614  Expr *rex = E->getRHS()->IgnoreParenImpCasts();
2615
2616  // Check to see if one of the (unmodified) operands is of different
2617  // signedness.
2618  Expr *signedOperand, *unsignedOperand;
2619  if (lex->getType()->hasSignedIntegerRepresentation()) {
2620    assert(!rex->getType()->hasSignedIntegerRepresentation() &&
2621           "unsigned comparison between two signed integer expressions?");
2622    signedOperand = lex;
2623    unsignedOperand = rex;
2624  } else if (rex->getType()->hasSignedIntegerRepresentation()) {
2625    signedOperand = rex;
2626    unsignedOperand = lex;
2627  } else {
2628    CheckTrivialUnsignedComparison(S, E);
2629    return AnalyzeImpConvsInComparison(S, E);
2630  }
2631
2632  // Otherwise, calculate the effective range of the signed operand.
2633  IntRange signedRange = GetExprRange(S.Context, signedOperand);
2634
2635  // Go ahead and analyze implicit conversions in the operands.  Note
2636  // that we skip the implicit conversions on both sides.
2637  AnalyzeImplicitConversions(S, lex, E->getOperatorLoc());
2638  AnalyzeImplicitConversions(S, rex, E->getOperatorLoc());
2639
2640  // If the signed range is non-negative, -Wsign-compare won't fire,
2641  // but we should still check for comparisons which are always true
2642  // or false.
2643  if (signedRange.NonNegative)
2644    return CheckTrivialUnsignedComparison(S, E);
2645
2646  // For (in)equality comparisons, if the unsigned operand is a
2647  // constant which cannot collide with a overflowed signed operand,
2648  // then reinterpreting the signed operand as unsigned will not
2649  // change the result of the comparison.
2650  if (E->isEqualityOp()) {
2651    unsigned comparisonWidth = S.Context.getIntWidth(T);
2652    IntRange unsignedRange = GetExprRange(S.Context, unsignedOperand);
2653
2654    // We should never be unable to prove that the unsigned operand is
2655    // non-negative.
2656    assert(unsignedRange.NonNegative && "unsigned range includes negative?");
2657
2658    if (unsignedRange.Width < comparisonWidth)
2659      return;
2660  }
2661
2662  S.Diag(E->getOperatorLoc(), diag::warn_mixed_sign_comparison)
2663    << lex->getType() << rex->getType()
2664    << lex->getSourceRange() << rex->getSourceRange();
2665}
2666
2667/// Analyzes an attempt to assign the given value to a bitfield.
2668///
2669/// Returns true if there was something fishy about the attempt.
2670bool AnalyzeBitFieldAssignment(Sema &S, FieldDecl *Bitfield, Expr *Init,
2671                               SourceLocation InitLoc) {
2672  assert(Bitfield->isBitField());
2673  if (Bitfield->isInvalidDecl())
2674    return false;
2675
2676  // White-list bool bitfields.
2677  if (Bitfield->getType()->isBooleanType())
2678    return false;
2679
2680  // Ignore value- or type-dependent expressions.
2681  if (Bitfield->getBitWidth()->isValueDependent() ||
2682      Bitfield->getBitWidth()->isTypeDependent() ||
2683      Init->isValueDependent() ||
2684      Init->isTypeDependent())
2685    return false;
2686
2687  Expr *OriginalInit = Init->IgnoreParenImpCasts();
2688
2689  llvm::APSInt Width(32);
2690  Expr::EvalResult InitValue;
2691  if (!Bitfield->getBitWidth()->isIntegerConstantExpr(Width, S.Context) ||
2692      !OriginalInit->Evaluate(InitValue, S.Context) ||
2693      !InitValue.Val.isInt())
2694    return false;
2695
2696  const llvm::APSInt &Value = InitValue.Val.getInt();
2697  unsigned OriginalWidth = Value.getBitWidth();
2698  unsigned FieldWidth = Width.getZExtValue();
2699
2700  if (OriginalWidth <= FieldWidth)
2701    return false;
2702
2703  llvm::APSInt TruncatedValue = Value.trunc(FieldWidth);
2704
2705  // It's fairly common to write values into signed bitfields
2706  // that, if sign-extended, would end up becoming a different
2707  // value.  We don't want to warn about that.
2708  if (Value.isSigned() && Value.isNegative())
2709    TruncatedValue = TruncatedValue.sext(OriginalWidth);
2710  else
2711    TruncatedValue = TruncatedValue.zext(OriginalWidth);
2712
2713  if (Value == TruncatedValue)
2714    return false;
2715
2716  std::string PrettyValue = Value.toString(10);
2717  std::string PrettyTrunc = TruncatedValue.toString(10);
2718
2719  S.Diag(InitLoc, diag::warn_impcast_bitfield_precision_constant)
2720    << PrettyValue << PrettyTrunc << OriginalInit->getType()
2721    << Init->getSourceRange();
2722
2723  return true;
2724}
2725
2726/// Analyze the given simple or compound assignment for warning-worthy
2727/// operations.
2728void AnalyzeAssignment(Sema &S, BinaryOperator *E) {
2729  // Just recurse on the LHS.
2730  AnalyzeImplicitConversions(S, E->getLHS(), E->getOperatorLoc());
2731
2732  // We want to recurse on the RHS as normal unless we're assigning to
2733  // a bitfield.
2734  if (FieldDecl *Bitfield = E->getLHS()->getBitField()) {
2735    if (AnalyzeBitFieldAssignment(S, Bitfield, E->getRHS(),
2736                                  E->getOperatorLoc())) {
2737      // Recurse, ignoring any implicit conversions on the RHS.
2738      return AnalyzeImplicitConversions(S, E->getRHS()->IgnoreParenImpCasts(),
2739                                        E->getOperatorLoc());
2740    }
2741  }
2742
2743  AnalyzeImplicitConversions(S, E->getRHS(), E->getOperatorLoc());
2744}
2745
2746/// Diagnose an implicit cast;  purely a helper for CheckImplicitConversion.
2747void DiagnoseImpCast(Sema &S, Expr *E, QualType T, SourceLocation CContext,
2748                     unsigned diag) {
2749  S.Diag(E->getExprLoc(), diag)
2750    << E->getType() << T << E->getSourceRange() << SourceRange(CContext);
2751}
2752
2753/// Diagnose an implicit cast;  purely a helper for CheckImplicitConversion.
2754void DiagnoseImpCast(Sema &S, Expr *E, QualType SourceType, QualType T,
2755                     SourceLocation CContext, unsigned diag) {
2756  S.Diag(E->getExprLoc(), diag)
2757    << SourceType << T << E->getSourceRange() << SourceRange(CContext);
2758}
2759
2760std::string PrettyPrintInRange(const llvm::APSInt &Value, IntRange Range) {
2761  if (!Range.Width) return "0";
2762
2763  llvm::APSInt ValueInRange = Value;
2764  ValueInRange.setIsSigned(!Range.NonNegative);
2765  ValueInRange = ValueInRange.trunc(Range.Width);
2766  return ValueInRange.toString(10);
2767}
2768
2769static bool isFromSystemMacro(Sema &S, SourceLocation loc) {
2770  SourceManager &smgr = S.Context.getSourceManager();
2771  return loc.isMacroID() && smgr.isInSystemHeader(smgr.getSpellingLoc(loc));
2772}
2773
2774void CheckImplicitConversion(Sema &S, Expr *E, QualType T,
2775                             SourceLocation CC, bool *ICContext = 0) {
2776  if (E->isTypeDependent() || E->isValueDependent()) return;
2777
2778  const Type *Source = S.Context.getCanonicalType(E->getType()).getTypePtr();
2779  const Type *Target = S.Context.getCanonicalType(T).getTypePtr();
2780  if (Source == Target) return;
2781  if (Target->isDependentType()) return;
2782
2783  // If the conversion context location is invalid don't complain.
2784  // We also don't want to emit a warning if the issue occurs from the
2785  // instantiation of a system macro.  The problem is that 'getSpellingLoc()'
2786  // is slow, so we delay this check as long as possible.  Once we detect
2787  // we are in that scenario, we just return.
2788  if (CC.isInvalid())
2789    return;
2790
2791  // Never diagnose implicit casts to bool.
2792  if (Target->isSpecificBuiltinType(BuiltinType::Bool))
2793    return;
2794
2795  // Strip vector types.
2796  if (isa<VectorType>(Source)) {
2797    if (!isa<VectorType>(Target)) {
2798      if (isFromSystemMacro(S, CC))
2799        return;
2800      return DiagnoseImpCast(S, E, T, CC, diag::warn_impcast_vector_scalar);
2801    }
2802
2803    Source = cast<VectorType>(Source)->getElementType().getTypePtr();
2804    Target = cast<VectorType>(Target)->getElementType().getTypePtr();
2805  }
2806
2807  // Strip complex types.
2808  if (isa<ComplexType>(Source)) {
2809    if (!isa<ComplexType>(Target)) {
2810      if (isFromSystemMacro(S, CC))
2811        return;
2812
2813      return DiagnoseImpCast(S, E, T, CC, diag::warn_impcast_complex_scalar);
2814    }
2815
2816    Source = cast<ComplexType>(Source)->getElementType().getTypePtr();
2817    Target = cast<ComplexType>(Target)->getElementType().getTypePtr();
2818  }
2819
2820  const BuiltinType *SourceBT = dyn_cast<BuiltinType>(Source);
2821  const BuiltinType *TargetBT = dyn_cast<BuiltinType>(Target);
2822
2823  // If the source is floating point...
2824  if (SourceBT && SourceBT->isFloatingPoint()) {
2825    // ...and the target is floating point...
2826    if (TargetBT && TargetBT->isFloatingPoint()) {
2827      // ...then warn if we're dropping FP rank.
2828
2829      // Builtin FP kinds are ordered by increasing FP rank.
2830      if (SourceBT->getKind() > TargetBT->getKind()) {
2831        // Don't warn about float constants that are precisely
2832        // representable in the target type.
2833        Expr::EvalResult result;
2834        if (E->Evaluate(result, S.Context)) {
2835          // Value might be a float, a float vector, or a float complex.
2836          if (IsSameFloatAfterCast(result.Val,
2837                   S.Context.getFloatTypeSemantics(QualType(TargetBT, 0)),
2838                   S.Context.getFloatTypeSemantics(QualType(SourceBT, 0))))
2839            return;
2840        }
2841
2842        if (isFromSystemMacro(S, CC))
2843          return;
2844
2845        DiagnoseImpCast(S, E, T, CC, diag::warn_impcast_float_precision);
2846      }
2847      return;
2848    }
2849
2850    // If the target is integral, always warn.
2851    if ((TargetBT && TargetBT->isInteger())) {
2852      if (isFromSystemMacro(S, CC))
2853        return;
2854
2855      Expr *InnerE = E->IgnoreParenImpCasts();
2856      if (FloatingLiteral *LiteralExpr = dyn_cast<FloatingLiteral>(InnerE)) {
2857        DiagnoseImpCast(S, LiteralExpr, T, CC,
2858                        diag::warn_impcast_literal_float_to_integer);
2859      } else {
2860        DiagnoseImpCast(S, E, T, CC, diag::warn_impcast_float_integer);
2861      }
2862    }
2863
2864    return;
2865  }
2866
2867  if (!Source->isIntegerType() || !Target->isIntegerType())
2868    return;
2869
2870  IntRange SourceRange = GetExprRange(S.Context, E);
2871  IntRange TargetRange = IntRange::forTargetOfCanonicalType(S.Context, Target);
2872
2873  if (SourceRange.Width > TargetRange.Width) {
2874    // If the source is a constant, use a default-on diagnostic.
2875    // TODO: this should happen for bitfield stores, too.
2876    llvm::APSInt Value(32);
2877    if (E->isIntegerConstantExpr(Value, S.Context)) {
2878      if (isFromSystemMacro(S, CC))
2879        return;
2880
2881      std::string PrettySourceValue = Value.toString(10);
2882      std::string PrettyTargetValue = PrettyPrintInRange(Value, TargetRange);
2883
2884      S.Diag(E->getExprLoc(), diag::warn_impcast_integer_precision_constant)
2885        << PrettySourceValue << PrettyTargetValue
2886        << E->getType() << T << E->getSourceRange() << clang::SourceRange(CC);
2887      return;
2888    }
2889
2890    // People want to build with -Wshorten-64-to-32 and not -Wconversion
2891    // and by god we'll let them.
2892
2893    if (isFromSystemMacro(S, CC))
2894      return;
2895
2896    if (SourceRange.Width == 64 && TargetRange.Width == 32)
2897      return DiagnoseImpCast(S, E, T, CC, diag::warn_impcast_integer_64_32);
2898    return DiagnoseImpCast(S, E, T, CC, diag::warn_impcast_integer_precision);
2899  }
2900
2901  if ((TargetRange.NonNegative && !SourceRange.NonNegative) ||
2902      (!TargetRange.NonNegative && SourceRange.NonNegative &&
2903       SourceRange.Width == TargetRange.Width)) {
2904
2905    if (isFromSystemMacro(S, CC))
2906      return;
2907
2908    unsigned DiagID = diag::warn_impcast_integer_sign;
2909
2910    // Traditionally, gcc has warned about this under -Wsign-compare.
2911    // We also want to warn about it in -Wconversion.
2912    // So if -Wconversion is off, use a completely identical diagnostic
2913    // in the sign-compare group.
2914    // The conditional-checking code will
2915    if (ICContext) {
2916      DiagID = diag::warn_impcast_integer_sign_conditional;
2917      *ICContext = true;
2918    }
2919
2920    return DiagnoseImpCast(S, E, T, CC, DiagID);
2921  }
2922
2923  // Diagnose conversions between different enumeration types.
2924  // In C, we pretend that the type of an EnumConstantDecl is its enumeration
2925  // type, to give us better diagnostics.
2926  QualType SourceType = E->getType();
2927  if (!S.getLangOptions().CPlusPlus) {
2928    if (DeclRefExpr *DRE = dyn_cast<DeclRefExpr>(E))
2929      if (EnumConstantDecl *ECD = dyn_cast<EnumConstantDecl>(DRE->getDecl())) {
2930        EnumDecl *Enum = cast<EnumDecl>(ECD->getDeclContext());
2931        SourceType = S.Context.getTypeDeclType(Enum);
2932        Source = S.Context.getCanonicalType(SourceType).getTypePtr();
2933      }
2934  }
2935
2936  if (const EnumType *SourceEnum = Source->getAs<EnumType>())
2937    if (const EnumType *TargetEnum = Target->getAs<EnumType>())
2938      if ((SourceEnum->getDecl()->getIdentifier() ||
2939           SourceEnum->getDecl()->getTypedefForAnonDecl()) &&
2940          (TargetEnum->getDecl()->getIdentifier() ||
2941           TargetEnum->getDecl()->getTypedefForAnonDecl()) &&
2942          SourceEnum != TargetEnum) {
2943        if (isFromSystemMacro(S, CC))
2944          return;
2945
2946        return DiagnoseImpCast(S, E, SourceType, T, CC,
2947                               diag::warn_impcast_different_enum_types);
2948      }
2949
2950  return;
2951}
2952
2953void CheckConditionalOperator(Sema &S, ConditionalOperator *E, QualType T);
2954
2955void CheckConditionalOperand(Sema &S, Expr *E, QualType T,
2956                             SourceLocation CC, bool &ICContext) {
2957  E = E->IgnoreParenImpCasts();
2958
2959  if (isa<ConditionalOperator>(E))
2960    return CheckConditionalOperator(S, cast<ConditionalOperator>(E), T);
2961
2962  AnalyzeImplicitConversions(S, E, CC);
2963  if (E->getType() != T)
2964    return CheckImplicitConversion(S, E, T, CC, &ICContext);
2965  return;
2966}
2967
2968void CheckConditionalOperator(Sema &S, ConditionalOperator *E, QualType T) {
2969  SourceLocation CC = E->getQuestionLoc();
2970
2971  AnalyzeImplicitConversions(S, E->getCond(), CC);
2972
2973  bool Suspicious = false;
2974  CheckConditionalOperand(S, E->getTrueExpr(), T, CC, Suspicious);
2975  CheckConditionalOperand(S, E->getFalseExpr(), T, CC, Suspicious);
2976
2977  // If -Wconversion would have warned about either of the candidates
2978  // for a signedness conversion to the context type...
2979  if (!Suspicious) return;
2980
2981  // ...but it's currently ignored...
2982  if (S.Diags.getDiagnosticLevel(diag::warn_impcast_integer_sign_conditional,
2983                                 CC))
2984    return;
2985
2986  // ...and -Wsign-compare isn't...
2987  if (!S.Diags.getDiagnosticLevel(diag::warn_mixed_sign_conditional, CC))
2988    return;
2989
2990  // ...then check whether it would have warned about either of the
2991  // candidates for a signedness conversion to the condition type.
2992  if (E->getType() != T) {
2993    Suspicious = false;
2994    CheckImplicitConversion(S, E->getTrueExpr()->IgnoreParenImpCasts(),
2995                            E->getType(), CC, &Suspicious);
2996    if (!Suspicious)
2997      CheckImplicitConversion(S, E->getFalseExpr()->IgnoreParenImpCasts(),
2998                              E->getType(), CC, &Suspicious);
2999    if (!Suspicious)
3000      return;
3001  }
3002
3003  // If so, emit a diagnostic under -Wsign-compare.
3004  Expr *lex = E->getTrueExpr()->IgnoreParenImpCasts();
3005  Expr *rex = E->getFalseExpr()->IgnoreParenImpCasts();
3006  S.Diag(E->getQuestionLoc(), diag::warn_mixed_sign_conditional)
3007    << lex->getType() << rex->getType()
3008    << lex->getSourceRange() << rex->getSourceRange();
3009}
3010
3011/// AnalyzeImplicitConversions - Find and report any interesting
3012/// implicit conversions in the given expression.  There are a couple
3013/// of competing diagnostics here, -Wconversion and -Wsign-compare.
3014void AnalyzeImplicitConversions(Sema &S, Expr *OrigE, SourceLocation CC) {
3015  QualType T = OrigE->getType();
3016  Expr *E = OrigE->IgnoreParenImpCasts();
3017
3018  // For conditional operators, we analyze the arguments as if they
3019  // were being fed directly into the output.
3020  if (isa<ConditionalOperator>(E)) {
3021    ConditionalOperator *CO = cast<ConditionalOperator>(E);
3022    CheckConditionalOperator(S, CO, T);
3023    return;
3024  }
3025
3026  // Go ahead and check any implicit conversions we might have skipped.
3027  // The non-canonical typecheck is just an optimization;
3028  // CheckImplicitConversion will filter out dead implicit conversions.
3029  if (E->getType() != T)
3030    CheckImplicitConversion(S, E, T, CC);
3031
3032  // Now continue drilling into this expression.
3033
3034  // Skip past explicit casts.
3035  if (isa<ExplicitCastExpr>(E)) {
3036    E = cast<ExplicitCastExpr>(E)->getSubExpr()->IgnoreParenImpCasts();
3037    return AnalyzeImplicitConversions(S, E, CC);
3038  }
3039
3040  if (BinaryOperator *BO = dyn_cast<BinaryOperator>(E)) {
3041    // Do a somewhat different check with comparison operators.
3042    if (BO->isComparisonOp())
3043      return AnalyzeComparison(S, BO);
3044
3045    // And with assignments and compound assignments.
3046    if (BO->isAssignmentOp())
3047      return AnalyzeAssignment(S, BO);
3048  }
3049
3050  // These break the otherwise-useful invariant below.  Fortunately,
3051  // we don't really need to recurse into them, because any internal
3052  // expressions should have been analyzed already when they were
3053  // built into statements.
3054  if (isa<StmtExpr>(E)) return;
3055
3056  // Don't descend into unevaluated contexts.
3057  if (isa<UnaryExprOrTypeTraitExpr>(E)) return;
3058
3059  // Now just recurse over the expression's children.
3060  CC = E->getExprLoc();
3061  for (Stmt::child_range I = E->children(); I; ++I)
3062    AnalyzeImplicitConversions(S, cast<Expr>(*I), CC);
3063}
3064
3065} // end anonymous namespace
3066
3067/// Diagnoses "dangerous" implicit conversions within the given
3068/// expression (which is a full expression).  Implements -Wconversion
3069/// and -Wsign-compare.
3070///
3071/// \param CC the "context" location of the implicit conversion, i.e.
3072///   the most location of the syntactic entity requiring the implicit
3073///   conversion
3074void Sema::CheckImplicitConversions(Expr *E, SourceLocation CC) {
3075  // Don't diagnose in unevaluated contexts.
3076  if (ExprEvalContexts.back().Context == Sema::Unevaluated)
3077    return;
3078
3079  // Don't diagnose for value- or type-dependent expressions.
3080  if (E->isTypeDependent() || E->isValueDependent())
3081    return;
3082
3083  // This is not the right CC for (e.g.) a variable initialization.
3084  AnalyzeImplicitConversions(*this, E, CC);
3085}
3086
3087void Sema::CheckBitFieldInitialization(SourceLocation InitLoc,
3088                                       FieldDecl *BitField,
3089                                       Expr *Init) {
3090  (void) AnalyzeBitFieldAssignment(*this, BitField, Init, InitLoc);
3091}
3092
3093/// CheckParmsForFunctionDef - Check that the parameters of the given
3094/// function are appropriate for the definition of a function. This
3095/// takes care of any checks that cannot be performed on the
3096/// declaration itself, e.g., that the types of each of the function
3097/// parameters are complete.
3098bool Sema::CheckParmsForFunctionDef(ParmVarDecl **P, ParmVarDecl **PEnd,
3099                                    bool CheckParameterNames) {
3100  bool HasInvalidParm = false;
3101  for (; P != PEnd; ++P) {
3102    ParmVarDecl *Param = *P;
3103
3104    // C99 6.7.5.3p4: the parameters in a parameter type list in a
3105    // function declarator that is part of a function definition of
3106    // that function shall not have incomplete type.
3107    //
3108    // This is also C++ [dcl.fct]p6.
3109    if (!Param->isInvalidDecl() &&
3110        RequireCompleteType(Param->getLocation(), Param->getType(),
3111                               diag::err_typecheck_decl_incomplete_type)) {
3112      Param->setInvalidDecl();
3113      HasInvalidParm = true;
3114    }
3115
3116    // C99 6.9.1p5: If the declarator includes a parameter type list, the
3117    // declaration of each parameter shall include an identifier.
3118    if (CheckParameterNames &&
3119        Param->getIdentifier() == 0 &&
3120        !Param->isImplicit() &&
3121        !getLangOptions().CPlusPlus)
3122      Diag(Param->getLocation(), diag::err_parameter_name_omitted);
3123
3124    // C99 6.7.5.3p12:
3125    //   If the function declarator is not part of a definition of that
3126    //   function, parameters may have incomplete type and may use the [*]
3127    //   notation in their sequences of declarator specifiers to specify
3128    //   variable length array types.
3129    QualType PType = Param->getOriginalType();
3130    if (const ArrayType *AT = Context.getAsArrayType(PType)) {
3131      if (AT->getSizeModifier() == ArrayType::Star) {
3132        // FIXME: This diagnosic should point the the '[*]' if source-location
3133        // information is added for it.
3134        Diag(Param->getLocation(), diag::err_array_star_in_function_definition);
3135      }
3136    }
3137  }
3138
3139  return HasInvalidParm;
3140}
3141
3142/// CheckCastAlign - Implements -Wcast-align, which warns when a
3143/// pointer cast increases the alignment requirements.
3144void Sema::CheckCastAlign(Expr *Op, QualType T, SourceRange TRange) {
3145  // This is actually a lot of work to potentially be doing on every
3146  // cast; don't do it if we're ignoring -Wcast_align (as is the default).
3147  if (getDiagnostics().getDiagnosticLevel(diag::warn_cast_align,
3148                                          TRange.getBegin())
3149        == Diagnostic::Ignored)
3150    return;
3151
3152  // Ignore dependent types.
3153  if (T->isDependentType() || Op->getType()->isDependentType())
3154    return;
3155
3156  // Require that the destination be a pointer type.
3157  const PointerType *DestPtr = T->getAs<PointerType>();
3158  if (!DestPtr) return;
3159
3160  // If the destination has alignment 1, we're done.
3161  QualType DestPointee = DestPtr->getPointeeType();
3162  if (DestPointee->isIncompleteType()) return;
3163  CharUnits DestAlign = Context.getTypeAlignInChars(DestPointee);
3164  if (DestAlign.isOne()) return;
3165
3166  // Require that the source be a pointer type.
3167  const PointerType *SrcPtr = Op->getType()->getAs<PointerType>();
3168  if (!SrcPtr) return;
3169  QualType SrcPointee = SrcPtr->getPointeeType();
3170
3171  // Whitelist casts from cv void*.  We already implicitly
3172  // whitelisted casts to cv void*, since they have alignment 1.
3173  // Also whitelist casts involving incomplete types, which implicitly
3174  // includes 'void'.
3175  if (SrcPointee->isIncompleteType()) return;
3176
3177  CharUnits SrcAlign = Context.getTypeAlignInChars(SrcPointee);
3178  if (SrcAlign >= DestAlign) return;
3179
3180  Diag(TRange.getBegin(), diag::warn_cast_align)
3181    << Op->getType() << T
3182    << static_cast<unsigned>(SrcAlign.getQuantity())
3183    << static_cast<unsigned>(DestAlign.getQuantity())
3184    << TRange << Op->getSourceRange();
3185}
3186
3187static void CheckArrayAccess_Check(Sema &S,
3188                                   const clang::ArraySubscriptExpr *E) {
3189  const Expr *BaseExpr = E->getBase()->IgnoreParenImpCasts();
3190  const ConstantArrayType *ArrayTy =
3191    S.Context.getAsConstantArrayType(BaseExpr->getType());
3192  if (!ArrayTy)
3193    return;
3194
3195  const Expr *IndexExpr = E->getIdx();
3196  if (IndexExpr->isValueDependent())
3197    return;
3198  llvm::APSInt index;
3199  if (!IndexExpr->isIntegerConstantExpr(index, S.Context))
3200    return;
3201
3202  if (index.isUnsigned() || !index.isNegative()) {
3203    llvm::APInt size = ArrayTy->getSize();
3204    if (!size.isStrictlyPositive())
3205      return;
3206    if (size.getBitWidth() > index.getBitWidth())
3207      index = index.sext(size.getBitWidth());
3208    else if (size.getBitWidth() < index.getBitWidth())
3209      size = size.sext(index.getBitWidth());
3210
3211    if (index.slt(size))
3212      return;
3213
3214    S.DiagRuntimeBehavior(E->getBase()->getLocStart(), BaseExpr,
3215                          S.PDiag(diag::warn_array_index_exceeds_bounds)
3216                            << index.toString(10, true)
3217                            << size.toString(10, true)
3218                            << IndexExpr->getSourceRange());
3219  } else {
3220    S.DiagRuntimeBehavior(E->getBase()->getLocStart(), BaseExpr,
3221                          S.PDiag(diag::warn_array_index_precedes_bounds)
3222                            << index.toString(10, true)
3223                            << IndexExpr->getSourceRange());
3224  }
3225
3226  const NamedDecl *ND = NULL;
3227  if (const DeclRefExpr *DRE = dyn_cast<DeclRefExpr>(BaseExpr))
3228    ND = dyn_cast<NamedDecl>(DRE->getDecl());
3229  if (const MemberExpr *ME = dyn_cast<MemberExpr>(BaseExpr))
3230    ND = dyn_cast<NamedDecl>(ME->getMemberDecl());
3231  if (ND)
3232    S.DiagRuntimeBehavior(ND->getLocStart(), BaseExpr,
3233                          S.PDiag(diag::note_array_index_out_of_bounds)
3234                            << ND->getDeclName());
3235}
3236
3237void Sema::CheckArrayAccess(const Expr *expr) {
3238  while (true)
3239    switch (expr->getStmtClass()) {
3240      case Stmt::ParenExprClass:
3241        expr = cast<ParenExpr>(expr)->getSubExpr();
3242        continue;
3243      case Stmt::ArraySubscriptExprClass:
3244        CheckArrayAccess_Check(*this, cast<ArraySubscriptExpr>(expr));
3245        return;
3246      case Stmt::ConditionalOperatorClass: {
3247        const ConditionalOperator *cond = cast<ConditionalOperator>(expr);
3248        if (const Expr *lhs = cond->getLHS())
3249          CheckArrayAccess(lhs);
3250        if (const Expr *rhs = cond->getRHS())
3251          CheckArrayAccess(rhs);
3252        return;
3253      }
3254      default:
3255        return;
3256    }
3257}
3258