1 //===--- CGCall.cpp - Encapsulate calling convention details ----*- C++ -*-===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // These classes wrap the information about a call or function
11 // definition used to handle ABI compliancy.
12 //
13 //===----------------------------------------------------------------------===//
14 
15 #include "CGCall.h"
16 #include "ABIInfo.h"
17 #include "CGCXXABI.h"
18 #include "CodeGenFunction.h"
19 #include "CodeGenModule.h"
20 #include "TargetInfo.h"
21 #include "clang/AST/Decl.h"
22 #include "clang/AST/DeclCXX.h"
23 #include "clang/AST/DeclObjC.h"
24 #include "clang/Basic/TargetInfo.h"
25 #include "clang/Frontend/CodeGenOptions.h"
26 #include "llvm/IR/Attributes.h"
27 #include "llvm/IR/DataLayout.h"
28 #include "llvm/IR/InlineAsm.h"
29 #include "llvm/Support/CallSite.h"
30 #include "llvm/Transforms/Utils/Local.h"
31 using namespace clang;
32 using namespace CodeGen;
33 
34 /***/
35 
36 static unsigned ClangCallConvToLLVMCallConv(CallingConv CC) {
37   switch (CC) {
38   default: return llvm::CallingConv::C;
39   case CC_X86StdCall: return llvm::CallingConv::X86_StdCall;
40   case CC_X86FastCall: return llvm::CallingConv::X86_FastCall;
41   case CC_X86ThisCall: return llvm::CallingConv::X86_ThisCall;
42   case CC_AAPCS: return llvm::CallingConv::ARM_AAPCS;
43   case CC_AAPCS_VFP: return llvm::CallingConv::ARM_AAPCS_VFP;
44   case CC_IntelOclBicc: return llvm::CallingConv::Intel_OCL_BI;
45   // TODO: add support for CC_X86Pascal to llvm
46   }
47 }
48 
49 /// Derives the 'this' type for codegen purposes, i.e. ignoring method
50 /// qualification.
51 /// FIXME: address space qualification?
52 static CanQualType GetThisType(ASTContext &Context, const CXXRecordDecl *RD) {
53   QualType RecTy = Context.getTagDeclType(RD)->getCanonicalTypeInternal();
54   return Context.getPointerType(CanQualType::CreateUnsafe(RecTy));
55 }
56 
57 /// Returns the canonical formal type of the given C++ method.
58 static CanQual<FunctionProtoType> GetFormalType(const CXXMethodDecl *MD) {
59   return MD->getType()->getCanonicalTypeUnqualified()
60            .getAs<FunctionProtoType>();
61 }
62 
63 /// Returns the "extra-canonicalized" return type, which discards
64 /// qualifiers on the return type.  Codegen doesn't care about them,
65 /// and it makes ABI code a little easier to be able to assume that
66 /// all parameter and return types are top-level unqualified.
67 static CanQualType GetReturnType(QualType RetTy) {
68   return RetTy->getCanonicalTypeUnqualified().getUnqualifiedType();
69 }
70 
71 /// Arrange the argument and result information for a value of the given
72 /// unprototyped freestanding function type.
73 const CGFunctionInfo &
74 CodeGenTypes::arrangeFreeFunctionType(CanQual<FunctionNoProtoType> FTNP) {
75   // When translating an unprototyped function type, always use a
76   // variadic type.
77   return arrangeLLVMFunctionInfo(FTNP->getResultType().getUnqualifiedType(),
78                                  ArrayRef<CanQualType>(),
79                                  FTNP->getExtInfo(),
80                                  RequiredArgs(0));
81 }
82 
83 /// Arrange the LLVM function layout for a value of the given function
84 /// type, on top of any implicit parameters already stored.  Use the
85 /// given ExtInfo instead of the ExtInfo from the function type.
86 static const CGFunctionInfo &arrangeLLVMFunctionInfo(CodeGenTypes &CGT,
87                                        SmallVectorImpl<CanQualType> &prefix,
88                                              CanQual<FunctionProtoType> FTP,
89                                               FunctionType::ExtInfo extInfo) {
90   RequiredArgs required = RequiredArgs::forPrototypePlus(FTP, prefix.size());
91   // FIXME: Kill copy.
92   for (unsigned i = 0, e = FTP->getNumArgs(); i != e; ++i)
93     prefix.push_back(FTP->getArgType(i));
94   CanQualType resultType = FTP->getResultType().getUnqualifiedType();
95   return CGT.arrangeLLVMFunctionInfo(resultType, prefix, extInfo, required);
96 }
97 
98 /// Arrange the argument and result information for a free function (i.e.
99 /// not a C++ or ObjC instance method) of the given type.
100 static const CGFunctionInfo &arrangeFreeFunctionType(CodeGenTypes &CGT,
101                                       SmallVectorImpl<CanQualType> &prefix,
102                                             CanQual<FunctionProtoType> FTP) {
103   return arrangeLLVMFunctionInfo(CGT, prefix, FTP, FTP->getExtInfo());
104 }
105 
106 /// Given the formal ext-info of a C++ instance method, adjust it
107 /// according to the C++ ABI in effect.
108 static void adjustCXXMethodInfo(CodeGenTypes &CGT,
109                                 FunctionType::ExtInfo &extInfo,
110                                 bool isVariadic) {
111   if (extInfo.getCC() == CC_Default) {
112     CallingConv CC = CGT.getContext().getDefaultCXXMethodCallConv(isVariadic);
113     extInfo = extInfo.withCallingConv(CC);
114   }
115 }
116 
117 /// Arrange the argument and result information for a free function (i.e.
118 /// not a C++ or ObjC instance method) of the given type.
119 static const CGFunctionInfo &arrangeCXXMethodType(CodeGenTypes &CGT,
120                                       SmallVectorImpl<CanQualType> &prefix,
121                                             CanQual<FunctionProtoType> FTP) {
122   FunctionType::ExtInfo extInfo = FTP->getExtInfo();
123   adjustCXXMethodInfo(CGT, extInfo, FTP->isVariadic());
124   return arrangeLLVMFunctionInfo(CGT, prefix, FTP, extInfo);
125 }
126 
127 /// Arrange the argument and result information for a value of the
128 /// given freestanding function type.
129 const CGFunctionInfo &
130 CodeGenTypes::arrangeFreeFunctionType(CanQual<FunctionProtoType> FTP) {
131   SmallVector<CanQualType, 16> argTypes;
132   return ::arrangeFreeFunctionType(*this, argTypes, FTP);
133 }
134 
135 static CallingConv getCallingConventionForDecl(const Decl *D) {
136   // Set the appropriate calling convention for the Function.
137   if (D->hasAttr<StdCallAttr>())
138     return CC_X86StdCall;
139 
140   if (D->hasAttr<FastCallAttr>())
141     return CC_X86FastCall;
142 
143   if (D->hasAttr<ThisCallAttr>())
144     return CC_X86ThisCall;
145 
146   if (D->hasAttr<PascalAttr>())
147     return CC_X86Pascal;
148 
149   if (PcsAttr *PCS = D->getAttr<PcsAttr>())
150     return (PCS->getPCS() == PcsAttr::AAPCS ? CC_AAPCS : CC_AAPCS_VFP);
151 
152   if (D->hasAttr<PnaclCallAttr>())
153     return CC_PnaclCall;
154 
155   if (D->hasAttr<IntelOclBiccAttr>())
156     return CC_IntelOclBicc;
157 
158   return CC_C;
159 }
160 
161 /// Arrange the argument and result information for a call to an
162 /// unknown C++ non-static member function of the given abstract type.
163 /// The member function must be an ordinary function, i.e. not a
164 /// constructor or destructor.
165 const CGFunctionInfo &
166 CodeGenTypes::arrangeCXXMethodType(const CXXRecordDecl *RD,
167                                    const FunctionProtoType *FTP) {
168   SmallVector<CanQualType, 16> argTypes;
169 
170   // Add the 'this' pointer.
171   argTypes.push_back(GetThisType(Context, RD));
172 
173   return ::arrangeCXXMethodType(*this, argTypes,
174               FTP->getCanonicalTypeUnqualified().getAs<FunctionProtoType>());
175 }
176 
177 /// Arrange the argument and result information for a declaration or
178 /// definition of the given C++ non-static member function.  The
179 /// member function must be an ordinary function, i.e. not a
180 /// constructor or destructor.
181 const CGFunctionInfo &
182 CodeGenTypes::arrangeCXXMethodDeclaration(const CXXMethodDecl *MD) {
183   assert(!isa<CXXConstructorDecl>(MD) && "wrong method for contructors!");
184   assert(!isa<CXXDestructorDecl>(MD) && "wrong method for destructors!");
185 
186   CanQual<FunctionProtoType> prototype = GetFormalType(MD);
187 
188   if (MD->isInstance()) {
189     // The abstract case is perfectly fine.
190     return arrangeCXXMethodType(MD->getParent(), prototype.getTypePtr());
191   }
192 
193   return arrangeFreeFunctionType(prototype);
194 }
195 
196 /// Arrange the argument and result information for a declaration
197 /// or definition to the given constructor variant.
198 const CGFunctionInfo &
199 CodeGenTypes::arrangeCXXConstructorDeclaration(const CXXConstructorDecl *D,
200                                                CXXCtorType ctorKind) {
201   SmallVector<CanQualType, 16> argTypes;
202   argTypes.push_back(GetThisType(Context, D->getParent()));
203   CanQualType resultType = Context.VoidTy;
204 
205   TheCXXABI.BuildConstructorSignature(D, ctorKind, resultType, argTypes);
206 
207   CanQual<FunctionProtoType> FTP = GetFormalType(D);
208 
209   RequiredArgs required = RequiredArgs::forPrototypePlus(FTP, argTypes.size());
210 
211   // Add the formal parameters.
212   for (unsigned i = 0, e = FTP->getNumArgs(); i != e; ++i)
213     argTypes.push_back(FTP->getArgType(i));
214 
215   FunctionType::ExtInfo extInfo = FTP->getExtInfo();
216   adjustCXXMethodInfo(*this, extInfo, FTP->isVariadic());
217   return arrangeLLVMFunctionInfo(resultType, argTypes, extInfo, required);
218 }
219 
220 /// Arrange the argument and result information for a declaration,
221 /// definition, or call to the given destructor variant.  It so
222 /// happens that all three cases produce the same information.
223 const CGFunctionInfo &
224 CodeGenTypes::arrangeCXXDestructor(const CXXDestructorDecl *D,
225                                    CXXDtorType dtorKind) {
226   SmallVector<CanQualType, 2> argTypes;
227   argTypes.push_back(GetThisType(Context, D->getParent()));
228   CanQualType resultType = Context.VoidTy;
229 
230   TheCXXABI.BuildDestructorSignature(D, dtorKind, resultType, argTypes);
231 
232   CanQual<FunctionProtoType> FTP = GetFormalType(D);
233   assert(FTP->getNumArgs() == 0 && "dtor with formal parameters");
234   assert(FTP->isVariadic() == 0 && "dtor with formal parameters");
235 
236   FunctionType::ExtInfo extInfo = FTP->getExtInfo();
237   adjustCXXMethodInfo(*this, extInfo, false);
238   return arrangeLLVMFunctionInfo(resultType, argTypes, extInfo,
239                                  RequiredArgs::All);
240 }
241 
242 /// Arrange the argument and result information for the declaration or
243 /// definition of the given function.
244 const CGFunctionInfo &
245 CodeGenTypes::arrangeFunctionDeclaration(const FunctionDecl *FD) {
246   if (const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(FD))
247     if (MD->isInstance())
248       return arrangeCXXMethodDeclaration(MD);
249 
250   CanQualType FTy = FD->getType()->getCanonicalTypeUnqualified();
251 
252   assert(isa<FunctionType>(FTy));
253 
254   // When declaring a function without a prototype, always use a
255   // non-variadic type.
256   if (isa<FunctionNoProtoType>(FTy)) {
257     CanQual<FunctionNoProtoType> noProto = FTy.getAs<FunctionNoProtoType>();
258     return arrangeLLVMFunctionInfo(noProto->getResultType(),
259                                    ArrayRef<CanQualType>(),
260                                    noProto->getExtInfo(),
261                                    RequiredArgs::All);
262   }
263 
264   assert(isa<FunctionProtoType>(FTy));
265   return arrangeFreeFunctionType(FTy.getAs<FunctionProtoType>());
266 }
267 
268 /// Arrange the argument and result information for the declaration or
269 /// definition of an Objective-C method.
270 const CGFunctionInfo &
271 CodeGenTypes::arrangeObjCMethodDeclaration(const ObjCMethodDecl *MD) {
272   // It happens that this is the same as a call with no optional
273   // arguments, except also using the formal 'self' type.
274   return arrangeObjCMessageSendSignature(MD, MD->getSelfDecl()->getType());
275 }
276 
277 /// Arrange the argument and result information for the function type
278 /// through which to perform a send to the given Objective-C method,
279 /// using the given receiver type.  The receiver type is not always
280 /// the 'self' type of the method or even an Objective-C pointer type.
281 /// This is *not* the right method for actually performing such a
282 /// message send, due to the possibility of optional arguments.
283 const CGFunctionInfo &
284 CodeGenTypes::arrangeObjCMessageSendSignature(const ObjCMethodDecl *MD,
285                                               QualType receiverType) {
286   SmallVector<CanQualType, 16> argTys;
287   argTys.push_back(Context.getCanonicalParamType(receiverType));
288   argTys.push_back(Context.getCanonicalParamType(Context.getObjCSelType()));
289   // FIXME: Kill copy?
290   for (ObjCMethodDecl::param_const_iterator i = MD->param_begin(),
291          e = MD->param_end(); i != e; ++i) {
292     argTys.push_back(Context.getCanonicalParamType((*i)->getType()));
293   }
294 
295   FunctionType::ExtInfo einfo;
296   einfo = einfo.withCallingConv(getCallingConventionForDecl(MD));
297 
298   if (getContext().getLangOpts().ObjCAutoRefCount &&
299       MD->hasAttr<NSReturnsRetainedAttr>())
300     einfo = einfo.withProducesResult(true);
301 
302   RequiredArgs required =
303     (MD->isVariadic() ? RequiredArgs(argTys.size()) : RequiredArgs::All);
304 
305   return arrangeLLVMFunctionInfo(GetReturnType(MD->getResultType()), argTys,
306                                  einfo, required);
307 }
308 
309 const CGFunctionInfo &
310 CodeGenTypes::arrangeGlobalDeclaration(GlobalDecl GD) {
311   // FIXME: Do we need to handle ObjCMethodDecl?
312   const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl());
313 
314   if (const CXXConstructorDecl *CD = dyn_cast<CXXConstructorDecl>(FD))
315     return arrangeCXXConstructorDeclaration(CD, GD.getCtorType());
316 
317   if (const CXXDestructorDecl *DD = dyn_cast<CXXDestructorDecl>(FD))
318     return arrangeCXXDestructor(DD, GD.getDtorType());
319 
320   return arrangeFunctionDeclaration(FD);
321 }
322 
323 /// Arrange a call as unto a free function, except possibly with an
324 /// additional number of formal parameters considered required.
325 static const CGFunctionInfo &
326 arrangeFreeFunctionLikeCall(CodeGenTypes &CGT,
327                             const CallArgList &args,
328                             const FunctionType *fnType,
329                             unsigned numExtraRequiredArgs) {
330   assert(args.size() >= numExtraRequiredArgs);
331 
332   // In most cases, there are no optional arguments.
333   RequiredArgs required = RequiredArgs::All;
334 
335   // If we have a variadic prototype, the required arguments are the
336   // extra prefix plus the arguments in the prototype.
337   if (const FunctionProtoType *proto = dyn_cast<FunctionProtoType>(fnType)) {
338     if (proto->isVariadic())
339       required = RequiredArgs(proto->getNumArgs() + numExtraRequiredArgs);
340 
341   // If we don't have a prototype at all, but we're supposed to
342   // explicitly use the variadic convention for unprototyped calls,
343   // treat all of the arguments as required but preserve the nominal
344   // possibility of variadics.
345   } else if (CGT.CGM.getTargetCodeGenInfo()
346                .isNoProtoCallVariadic(args, cast<FunctionNoProtoType>(fnType))) {
347     required = RequiredArgs(args.size());
348   }
349 
350   return CGT.arrangeFreeFunctionCall(fnType->getResultType(), args,
351                                      fnType->getExtInfo(), required);
352 }
353 
354 /// Figure out the rules for calling a function with the given formal
355 /// type using the given arguments.  The arguments are necessary
356 /// because the function might be unprototyped, in which case it's
357 /// target-dependent in crazy ways.
358 const CGFunctionInfo &
359 CodeGenTypes::arrangeFreeFunctionCall(const CallArgList &args,
360                                       const FunctionType *fnType) {
361   return arrangeFreeFunctionLikeCall(*this, args, fnType, 0);
362 }
363 
364 /// A block function call is essentially a free-function call with an
365 /// extra implicit argument.
366 const CGFunctionInfo &
367 CodeGenTypes::arrangeBlockFunctionCall(const CallArgList &args,
368                                        const FunctionType *fnType) {
369   return arrangeFreeFunctionLikeCall(*this, args, fnType, 1);
370 }
371 
372 const CGFunctionInfo &
373 CodeGenTypes::arrangeFreeFunctionCall(QualType resultType,
374                                       const CallArgList &args,
375                                       FunctionType::ExtInfo info,
376                                       RequiredArgs required) {
377   // FIXME: Kill copy.
378   SmallVector<CanQualType, 16> argTypes;
379   for (CallArgList::const_iterator i = args.begin(), e = args.end();
380        i != e; ++i)
381     argTypes.push_back(Context.getCanonicalParamType(i->Ty));
382   return arrangeLLVMFunctionInfo(GetReturnType(resultType), argTypes, info,
383                                  required);
384 }
385 
386 /// Arrange a call to a C++ method, passing the given arguments.
387 const CGFunctionInfo &
388 CodeGenTypes::arrangeCXXMethodCall(const CallArgList &args,
389                                    const FunctionProtoType *FPT,
390                                    RequiredArgs required) {
391   // FIXME: Kill copy.
392   SmallVector<CanQualType, 16> argTypes;
393   for (CallArgList::const_iterator i = args.begin(), e = args.end();
394        i != e; ++i)
395     argTypes.push_back(Context.getCanonicalParamType(i->Ty));
396 
397   FunctionType::ExtInfo info = FPT->getExtInfo();
398   adjustCXXMethodInfo(*this, info, FPT->isVariadic());
399   return arrangeLLVMFunctionInfo(GetReturnType(FPT->getResultType()),
400                                  argTypes, info, required);
401 }
402 
403 const CGFunctionInfo &
404 CodeGenTypes::arrangeFunctionDeclaration(QualType resultType,
405                                          const FunctionArgList &args,
406                                          const FunctionType::ExtInfo &info,
407                                          bool isVariadic) {
408   // FIXME: Kill copy.
409   SmallVector<CanQualType, 16> argTypes;
410   for (FunctionArgList::const_iterator i = args.begin(), e = args.end();
411        i != e; ++i)
412     argTypes.push_back(Context.getCanonicalParamType((*i)->getType()));
413 
414   RequiredArgs required =
415     (isVariadic ? RequiredArgs(args.size()) : RequiredArgs::All);
416   return arrangeLLVMFunctionInfo(GetReturnType(resultType), argTypes, info,
417                                  required);
418 }
419 
420 const CGFunctionInfo &CodeGenTypes::arrangeNullaryFunction() {
421   return arrangeLLVMFunctionInfo(getContext().VoidTy, ArrayRef<CanQualType>(),
422                                  FunctionType::ExtInfo(), RequiredArgs::All);
423 }
424 
425 /// Arrange the argument and result information for an abstract value
426 /// of a given function type.  This is the method which all of the
427 /// above functions ultimately defer to.
428 const CGFunctionInfo &
429 CodeGenTypes::arrangeLLVMFunctionInfo(CanQualType resultType,
430                                       ArrayRef<CanQualType> argTypes,
431                                       FunctionType::ExtInfo info,
432                                       RequiredArgs required) {
433 #ifndef NDEBUG
434   for (ArrayRef<CanQualType>::const_iterator
435          I = argTypes.begin(), E = argTypes.end(); I != E; ++I)
436     assert(I->isCanonicalAsParam());
437 #endif
438 
439   unsigned CC = ClangCallConvToLLVMCallConv(info.getCC());
440 
441   // Lookup or create unique function info.
442   llvm::FoldingSetNodeID ID;
443   CGFunctionInfo::Profile(ID, info, required, resultType, argTypes);
444 
445   void *insertPos = 0;
446   CGFunctionInfo *FI = FunctionInfos.FindNodeOrInsertPos(ID, insertPos);
447   if (FI)
448     return *FI;
449 
450   // Construct the function info.  We co-allocate the ArgInfos.
451   FI = CGFunctionInfo::create(CC, info, resultType, argTypes, required);
452   FunctionInfos.InsertNode(FI, insertPos);
453 
454   bool inserted = FunctionsBeingProcessed.insert(FI); (void)inserted;
455   assert(inserted && "Recursively being processed?");
456 
457   // Compute ABI information.
458   getABIInfo().computeInfo(*FI);
459 
460   // Loop over all of the computed argument and return value info.  If any of
461   // them are direct or extend without a specified coerce type, specify the
462   // default now.
463   ABIArgInfo &retInfo = FI->getReturnInfo();
464   if (retInfo.canHaveCoerceToType() && retInfo.getCoerceToType() == 0)
465     retInfo.setCoerceToType(ConvertType(FI->getReturnType()));
466 
467   for (CGFunctionInfo::arg_iterator I = FI->arg_begin(), E = FI->arg_end();
468        I != E; ++I)
469     if (I->info.canHaveCoerceToType() && I->info.getCoerceToType() == 0)
470       I->info.setCoerceToType(ConvertType(I->type));
471 
472   bool erased = FunctionsBeingProcessed.erase(FI); (void)erased;
473   assert(erased && "Not in set?");
474 
475   return *FI;
476 }
477 
478 CGFunctionInfo *CGFunctionInfo::create(unsigned llvmCC,
479                                        const FunctionType::ExtInfo &info,
480                                        CanQualType resultType,
481                                        ArrayRef<CanQualType> argTypes,
482                                        RequiredArgs required) {
483   void *buffer = operator new(sizeof(CGFunctionInfo) +
484                               sizeof(ArgInfo) * (argTypes.size() + 1));
485   CGFunctionInfo *FI = new(buffer) CGFunctionInfo();
486   FI->CallingConvention = llvmCC;
487   FI->EffectiveCallingConvention = llvmCC;
488   FI->ASTCallingConvention = info.getCC();
489   FI->NoReturn = info.getNoReturn();
490   FI->ReturnsRetained = info.getProducesResult();
491   FI->Required = required;
492   FI->HasRegParm = info.getHasRegParm();
493   FI->RegParm = info.getRegParm();
494   FI->NumArgs = argTypes.size();
495   FI->getArgsBuffer()[0].type = resultType;
496   for (unsigned i = 0, e = argTypes.size(); i != e; ++i)
497     FI->getArgsBuffer()[i + 1].type = argTypes[i];
498   return FI;
499 }
500 
501 /***/
502 
503 void CodeGenTypes::GetExpandedTypes(QualType type,
504                      SmallVectorImpl<llvm::Type*> &expandedTypes) {
505   if (const ConstantArrayType *AT = Context.getAsConstantArrayType(type)) {
506     uint64_t NumElts = AT->getSize().getZExtValue();
507     for (uint64_t Elt = 0; Elt < NumElts; ++Elt)
508       GetExpandedTypes(AT->getElementType(), expandedTypes);
509   } else if (const RecordType *RT = type->getAs<RecordType>()) {
510     const RecordDecl *RD = RT->getDecl();
511     assert(!RD->hasFlexibleArrayMember() &&
512            "Cannot expand structure with flexible array.");
513     if (RD->isUnion()) {
514       // Unions can be here only in degenerative cases - all the fields are same
515       // after flattening. Thus we have to use the "largest" field.
516       const FieldDecl *LargestFD = 0;
517       CharUnits UnionSize = CharUnits::Zero();
518 
519       for (RecordDecl::field_iterator i = RD->field_begin(), e = RD->field_end();
520            i != e; ++i) {
521         const FieldDecl *FD = *i;
522         assert(!FD->isBitField() &&
523                "Cannot expand structure with bit-field members.");
524         CharUnits FieldSize = getContext().getTypeSizeInChars(FD->getType());
525         if (UnionSize < FieldSize) {
526           UnionSize = FieldSize;
527           LargestFD = FD;
528         }
529       }
530       if (LargestFD)
531         GetExpandedTypes(LargestFD->getType(), expandedTypes);
532     } else {
533       for (RecordDecl::field_iterator i = RD->field_begin(), e = RD->field_end();
534            i != e; ++i) {
535         assert(!i->isBitField() &&
536                "Cannot expand structure with bit-field members.");
537         GetExpandedTypes(i->getType(), expandedTypes);
538       }
539     }
540   } else if (const ComplexType *CT = type->getAs<ComplexType>()) {
541     llvm::Type *EltTy = ConvertType(CT->getElementType());
542     expandedTypes.push_back(EltTy);
543     expandedTypes.push_back(EltTy);
544   } else
545     expandedTypes.push_back(ConvertType(type));
546 }
547 
548 llvm::Function::arg_iterator
549 CodeGenFunction::ExpandTypeFromArgs(QualType Ty, LValue LV,
550                                     llvm::Function::arg_iterator AI) {
551   assert(LV.isSimple() &&
552          "Unexpected non-simple lvalue during struct expansion.");
553 
554   if (const ConstantArrayType *AT = getContext().getAsConstantArrayType(Ty)) {
555     unsigned NumElts = AT->getSize().getZExtValue();
556     QualType EltTy = AT->getElementType();
557     for (unsigned Elt = 0; Elt < NumElts; ++Elt) {
558       llvm::Value *EltAddr = Builder.CreateConstGEP2_32(LV.getAddress(), 0, Elt);
559       LValue LV = MakeAddrLValue(EltAddr, EltTy);
560       AI = ExpandTypeFromArgs(EltTy, LV, AI);
561     }
562   } else if (const RecordType *RT = Ty->getAs<RecordType>()) {
563     RecordDecl *RD = RT->getDecl();
564     if (RD->isUnion()) {
565       // Unions can be here only in degenerative cases - all the fields are same
566       // after flattening. Thus we have to use the "largest" field.
567       const FieldDecl *LargestFD = 0;
568       CharUnits UnionSize = CharUnits::Zero();
569 
570       for (RecordDecl::field_iterator i = RD->field_begin(), e = RD->field_end();
571            i != e; ++i) {
572         const FieldDecl *FD = *i;
573         assert(!FD->isBitField() &&
574                "Cannot expand structure with bit-field members.");
575         CharUnits FieldSize = getContext().getTypeSizeInChars(FD->getType());
576         if (UnionSize < FieldSize) {
577           UnionSize = FieldSize;
578           LargestFD = FD;
579         }
580       }
581       if (LargestFD) {
582         // FIXME: What are the right qualifiers here?
583         LValue SubLV = EmitLValueForField(LV, LargestFD);
584         AI = ExpandTypeFromArgs(LargestFD->getType(), SubLV, AI);
585       }
586     } else {
587       for (RecordDecl::field_iterator i = RD->field_begin(), e = RD->field_end();
588            i != e; ++i) {
589         FieldDecl *FD = *i;
590         QualType FT = FD->getType();
591 
592         // FIXME: What are the right qualifiers here?
593         LValue SubLV = EmitLValueForField(LV, FD);
594         AI = ExpandTypeFromArgs(FT, SubLV, AI);
595       }
596     }
597   } else if (const ComplexType *CT = Ty->getAs<ComplexType>()) {
598     QualType EltTy = CT->getElementType();
599     llvm::Value *RealAddr = Builder.CreateStructGEP(LV.getAddress(), 0, "real");
600     EmitStoreThroughLValue(RValue::get(AI++), MakeAddrLValue(RealAddr, EltTy));
601     llvm::Value *ImagAddr = Builder.CreateStructGEP(LV.getAddress(), 1, "imag");
602     EmitStoreThroughLValue(RValue::get(AI++), MakeAddrLValue(ImagAddr, EltTy));
603   } else {
604     EmitStoreThroughLValue(RValue::get(AI), LV);
605     ++AI;
606   }
607 
608   return AI;
609 }
610 
611 /// EnterStructPointerForCoercedAccess - Given a struct pointer that we are
612 /// accessing some number of bytes out of it, try to gep into the struct to get
613 /// at its inner goodness.  Dive as deep as possible without entering an element
614 /// with an in-memory size smaller than DstSize.
615 static llvm::Value *
616 EnterStructPointerForCoercedAccess(llvm::Value *SrcPtr,
617                                    llvm::StructType *SrcSTy,
618                                    uint64_t DstSize, CodeGenFunction &CGF) {
619   // We can't dive into a zero-element struct.
620   if (SrcSTy->getNumElements() == 0) return SrcPtr;
621 
622   llvm::Type *FirstElt = SrcSTy->getElementType(0);
623 
624   // If the first elt is at least as large as what we're looking for, or if the
625   // first element is the same size as the whole struct, we can enter it.
626   uint64_t FirstEltSize =
627     CGF.CGM.getDataLayout().getTypeAllocSize(FirstElt);
628   if (FirstEltSize < DstSize &&
629       FirstEltSize < CGF.CGM.getDataLayout().getTypeAllocSize(SrcSTy))
630     return SrcPtr;
631 
632   // GEP into the first element.
633   SrcPtr = CGF.Builder.CreateConstGEP2_32(SrcPtr, 0, 0, "coerce.dive");
634 
635   // If the first element is a struct, recurse.
636   llvm::Type *SrcTy =
637     cast<llvm::PointerType>(SrcPtr->getType())->getElementType();
638   if (llvm::StructType *SrcSTy = dyn_cast<llvm::StructType>(SrcTy))
639     return EnterStructPointerForCoercedAccess(SrcPtr, SrcSTy, DstSize, CGF);
640 
641   return SrcPtr;
642 }
643 
644 /// CoerceIntOrPtrToIntOrPtr - Convert a value Val to the specific Ty where both
645 /// are either integers or pointers.  This does a truncation of the value if it
646 /// is too large or a zero extension if it is too small.
647 static llvm::Value *CoerceIntOrPtrToIntOrPtr(llvm::Value *Val,
648                                              llvm::Type *Ty,
649                                              CodeGenFunction &CGF) {
650   if (Val->getType() == Ty)
651     return Val;
652 
653   if (isa<llvm::PointerType>(Val->getType())) {
654     // If this is Pointer->Pointer avoid conversion to and from int.
655     if (isa<llvm::PointerType>(Ty))
656       return CGF.Builder.CreateBitCast(Val, Ty, "coerce.val");
657 
658     // Convert the pointer to an integer so we can play with its width.
659     Val = CGF.Builder.CreatePtrToInt(Val, CGF.IntPtrTy, "coerce.val.pi");
660   }
661 
662   llvm::Type *DestIntTy = Ty;
663   if (isa<llvm::PointerType>(DestIntTy))
664     DestIntTy = CGF.IntPtrTy;
665 
666   if (Val->getType() != DestIntTy)
667     Val = CGF.Builder.CreateIntCast(Val, DestIntTy, false, "coerce.val.ii");
668 
669   if (isa<llvm::PointerType>(Ty))
670     Val = CGF.Builder.CreateIntToPtr(Val, Ty, "coerce.val.ip");
671   return Val;
672 }
673 
674 
675 
676 /// CreateCoercedLoad - Create a load from \arg SrcPtr interpreted as
677 /// a pointer to an object of type \arg Ty.
678 ///
679 /// This safely handles the case when the src type is smaller than the
680 /// destination type; in this situation the values of bits which not
681 /// present in the src are undefined.
682 static llvm::Value *CreateCoercedLoad(llvm::Value *SrcPtr,
683                                       llvm::Type *Ty,
684                                       CodeGenFunction &CGF) {
685   llvm::Type *SrcTy =
686     cast<llvm::PointerType>(SrcPtr->getType())->getElementType();
687 
688   // If SrcTy and Ty are the same, just do a load.
689   if (SrcTy == Ty)
690     return CGF.Builder.CreateLoad(SrcPtr);
691 
692   uint64_t DstSize = CGF.CGM.getDataLayout().getTypeAllocSize(Ty);
693 
694   if (llvm::StructType *SrcSTy = dyn_cast<llvm::StructType>(SrcTy)) {
695     SrcPtr = EnterStructPointerForCoercedAccess(SrcPtr, SrcSTy, DstSize, CGF);
696     SrcTy = cast<llvm::PointerType>(SrcPtr->getType())->getElementType();
697   }
698 
699   uint64_t SrcSize = CGF.CGM.getDataLayout().getTypeAllocSize(SrcTy);
700 
701   // If the source and destination are integer or pointer types, just do an
702   // extension or truncation to the desired type.
703   if ((isa<llvm::IntegerType>(Ty) || isa<llvm::PointerType>(Ty)) &&
704       (isa<llvm::IntegerType>(SrcTy) || isa<llvm::PointerType>(SrcTy))) {
705     llvm::LoadInst *Load = CGF.Builder.CreateLoad(SrcPtr);
706     return CoerceIntOrPtrToIntOrPtr(Load, Ty, CGF);
707   }
708 
709   // If load is legal, just bitcast the src pointer.
710   if (SrcSize >= DstSize) {
711     // Generally SrcSize is never greater than DstSize, since this means we are
712     // losing bits. However, this can happen in cases where the structure has
713     // additional padding, for example due to a user specified alignment.
714     //
715     // FIXME: Assert that we aren't truncating non-padding bits when have access
716     // to that information.
717     llvm::Value *Casted =
718       CGF.Builder.CreateBitCast(SrcPtr, llvm::PointerType::getUnqual(Ty));
719     llvm::LoadInst *Load = CGF.Builder.CreateLoad(Casted);
720     // FIXME: Use better alignment / avoid requiring aligned load.
721     Load->setAlignment(1);
722     return Load;
723   }
724 
725   // Otherwise do coercion through memory. This is stupid, but
726   // simple.
727   llvm::Value *Tmp = CGF.CreateTempAlloca(Ty);
728   llvm::Type *I8PtrTy = CGF.Builder.getInt8PtrTy();
729   llvm::Value *Casted = CGF.Builder.CreateBitCast(Tmp, I8PtrTy);
730   llvm::Value *SrcCasted = CGF.Builder.CreateBitCast(SrcPtr, I8PtrTy);
731   // FIXME: Use better alignment.
732   CGF.Builder.CreateMemCpy(Casted, SrcCasted,
733       llvm::ConstantInt::get(CGF.IntPtrTy, SrcSize),
734       1, false);
735   return CGF.Builder.CreateLoad(Tmp);
736 }
737 
738 // Function to store a first-class aggregate into memory.  We prefer to
739 // store the elements rather than the aggregate to be more friendly to
740 // fast-isel.
741 // FIXME: Do we need to recurse here?
742 static void BuildAggStore(CodeGenFunction &CGF, llvm::Value *Val,
743                           llvm::Value *DestPtr, bool DestIsVolatile,
744                           bool LowAlignment) {
745   // Prefer scalar stores to first-class aggregate stores.
746   if (llvm::StructType *STy =
747         dyn_cast<llvm::StructType>(Val->getType())) {
748     for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
749       llvm::Value *EltPtr = CGF.Builder.CreateConstGEP2_32(DestPtr, 0, i);
750       llvm::Value *Elt = CGF.Builder.CreateExtractValue(Val, i);
751       llvm::StoreInst *SI = CGF.Builder.CreateStore(Elt, EltPtr,
752                                                     DestIsVolatile);
753       if (LowAlignment)
754         SI->setAlignment(1);
755     }
756   } else {
757     llvm::StoreInst *SI = CGF.Builder.CreateStore(Val, DestPtr, DestIsVolatile);
758     if (LowAlignment)
759       SI->setAlignment(1);
760   }
761 }
762 
763 /// CreateCoercedStore - Create a store to \arg DstPtr from \arg Src,
764 /// where the source and destination may have different types.
765 ///
766 /// This safely handles the case when the src type is larger than the
767 /// destination type; the upper bits of the src will be lost.
768 static void CreateCoercedStore(llvm::Value *Src,
769                                llvm::Value *DstPtr,
770                                bool DstIsVolatile,
771                                CodeGenFunction &CGF) {
772   llvm::Type *SrcTy = Src->getType();
773   llvm::Type *DstTy =
774     cast<llvm::PointerType>(DstPtr->getType())->getElementType();
775   if (SrcTy == DstTy) {
776     CGF.Builder.CreateStore(Src, DstPtr, DstIsVolatile);
777     return;
778   }
779 
780   uint64_t SrcSize = CGF.CGM.getDataLayout().getTypeAllocSize(SrcTy);
781 
782   if (llvm::StructType *DstSTy = dyn_cast<llvm::StructType>(DstTy)) {
783     DstPtr = EnterStructPointerForCoercedAccess(DstPtr, DstSTy, SrcSize, CGF);
784     DstTy = cast<llvm::PointerType>(DstPtr->getType())->getElementType();
785   }
786 
787   // If the source and destination are integer or pointer types, just do an
788   // extension or truncation to the desired type.
789   if ((isa<llvm::IntegerType>(SrcTy) || isa<llvm::PointerType>(SrcTy)) &&
790       (isa<llvm::IntegerType>(DstTy) || isa<llvm::PointerType>(DstTy))) {
791     Src = CoerceIntOrPtrToIntOrPtr(Src, DstTy, CGF);
792     CGF.Builder.CreateStore(Src, DstPtr, DstIsVolatile);
793     return;
794   }
795 
796   uint64_t DstSize = CGF.CGM.getDataLayout().getTypeAllocSize(DstTy);
797 
798   // If store is legal, just bitcast the src pointer.
799   if (SrcSize <= DstSize) {
800     llvm::Value *Casted =
801       CGF.Builder.CreateBitCast(DstPtr, llvm::PointerType::getUnqual(SrcTy));
802     // FIXME: Use better alignment / avoid requiring aligned store.
803     BuildAggStore(CGF, Src, Casted, DstIsVolatile, true);
804   } else {
805     // Otherwise do coercion through memory. This is stupid, but
806     // simple.
807 
808     // Generally SrcSize is never greater than DstSize, since this means we are
809     // losing bits. However, this can happen in cases where the structure has
810     // additional padding, for example due to a user specified alignment.
811     //
812     // FIXME: Assert that we aren't truncating non-padding bits when have access
813     // to that information.
814     llvm::Value *Tmp = CGF.CreateTempAlloca(SrcTy);
815     CGF.Builder.CreateStore(Src, Tmp);
816     llvm::Type *I8PtrTy = CGF.Builder.getInt8PtrTy();
817     llvm::Value *Casted = CGF.Builder.CreateBitCast(Tmp, I8PtrTy);
818     llvm::Value *DstCasted = CGF.Builder.CreateBitCast(DstPtr, I8PtrTy);
819     // FIXME: Use better alignment.
820     CGF.Builder.CreateMemCpy(DstCasted, Casted,
821         llvm::ConstantInt::get(CGF.IntPtrTy, DstSize),
822         1, false);
823   }
824 }
825 
826 /***/
827 
828 bool CodeGenModule::ReturnTypeUsesSRet(const CGFunctionInfo &FI) {
829   return FI.getReturnInfo().isIndirect();
830 }
831 
832 bool CodeGenModule::ReturnTypeUsesFPRet(QualType ResultType) {
833   if (const BuiltinType *BT = ResultType->getAs<BuiltinType>()) {
834     switch (BT->getKind()) {
835     default:
836       return false;
837     case BuiltinType::Float:
838       return getContext().getTargetInfo().useObjCFPRetForRealType(TargetInfo::Float);
839     case BuiltinType::Double:
840       return getContext().getTargetInfo().useObjCFPRetForRealType(TargetInfo::Double);
841     case BuiltinType::LongDouble:
842       return getContext().getTargetInfo().useObjCFPRetForRealType(
843         TargetInfo::LongDouble);
844     }
845   }
846 
847   return false;
848 }
849 
850 bool CodeGenModule::ReturnTypeUsesFP2Ret(QualType ResultType) {
851   if (const ComplexType *CT = ResultType->getAs<ComplexType>()) {
852     if (const BuiltinType *BT = CT->getElementType()->getAs<BuiltinType>()) {
853       if (BT->getKind() == BuiltinType::LongDouble)
854         return getContext().getTargetInfo().useObjCFP2RetForComplexLongDouble();
855     }
856   }
857 
858   return false;
859 }
860 
861 llvm::FunctionType *CodeGenTypes::GetFunctionType(GlobalDecl GD) {
862   const CGFunctionInfo &FI = arrangeGlobalDeclaration(GD);
863   return GetFunctionType(FI);
864 }
865 
866 llvm::FunctionType *
867 CodeGenTypes::GetFunctionType(const CGFunctionInfo &FI) {
868 
869   bool Inserted = FunctionsBeingProcessed.insert(&FI); (void)Inserted;
870   assert(Inserted && "Recursively being processed?");
871 
872   SmallVector<llvm::Type*, 8> argTypes;
873   llvm::Type *resultType = 0;
874 
875   const ABIArgInfo &retAI = FI.getReturnInfo();
876   switch (retAI.getKind()) {
877   case ABIArgInfo::Expand:
878     llvm_unreachable("Invalid ABI kind for return argument");
879 
880   case ABIArgInfo::Extend:
881   case ABIArgInfo::Direct:
882     resultType = retAI.getCoerceToType();
883     break;
884 
885   case ABIArgInfo::Indirect: {
886     assert(!retAI.getIndirectAlign() && "Align unused on indirect return.");
887     resultType = llvm::Type::getVoidTy(getLLVMContext());
888 
889     QualType ret = FI.getReturnType();
890     llvm::Type *ty = ConvertType(ret);
891     unsigned addressSpace = Context.getTargetAddressSpace(ret);
892     argTypes.push_back(llvm::PointerType::get(ty, addressSpace));
893     break;
894   }
895 
896   case ABIArgInfo::Ignore:
897     resultType = llvm::Type::getVoidTy(getLLVMContext());
898     break;
899   }
900 
901   // Add in all of the required arguments.
902   CGFunctionInfo::const_arg_iterator it = FI.arg_begin(), ie;
903   if (FI.isVariadic()) {
904     ie = it + FI.getRequiredArgs().getNumRequiredArgs();
905   } else {
906     ie = FI.arg_end();
907   }
908   for (; it != ie; ++it) {
909     const ABIArgInfo &argAI = it->info;
910 
911     // Insert a padding type to ensure proper alignment.
912     if (llvm::Type *PaddingType = argAI.getPaddingType())
913       argTypes.push_back(PaddingType);
914 
915     switch (argAI.getKind()) {
916     case ABIArgInfo::Ignore:
917       break;
918 
919     case ABIArgInfo::Indirect: {
920       // indirect arguments are always on the stack, which is addr space #0.
921       llvm::Type *LTy = ConvertTypeForMem(it->type);
922       argTypes.push_back(LTy->getPointerTo());
923       break;
924     }
925 
926     case ABIArgInfo::Extend:
927     case ABIArgInfo::Direct: {
928       // If the coerce-to type is a first class aggregate, flatten it.  Either
929       // way is semantically identical, but fast-isel and the optimizer
930       // generally likes scalar values better than FCAs.
931       llvm::Type *argType = argAI.getCoerceToType();
932       if (llvm::StructType *st = dyn_cast<llvm::StructType>(argType)) {
933         for (unsigned i = 0, e = st->getNumElements(); i != e; ++i)
934           argTypes.push_back(st->getElementType(i));
935       } else {
936         argTypes.push_back(argType);
937       }
938       break;
939     }
940 
941     case ABIArgInfo::Expand:
942       GetExpandedTypes(it->type, argTypes);
943       break;
944     }
945   }
946 
947   bool Erased = FunctionsBeingProcessed.erase(&FI); (void)Erased;
948   assert(Erased && "Not in set?");
949 
950   return llvm::FunctionType::get(resultType, argTypes, FI.isVariadic());
951 }
952 
953 llvm::Type *CodeGenTypes::GetFunctionTypeForVTable(GlobalDecl GD) {
954   const CXXMethodDecl *MD = cast<CXXMethodDecl>(GD.getDecl());
955   const FunctionProtoType *FPT = MD->getType()->getAs<FunctionProtoType>();
956 
957   if (!isFuncTypeConvertible(FPT))
958     return llvm::StructType::get(getLLVMContext());
959 
960   const CGFunctionInfo *Info;
961   if (isa<CXXDestructorDecl>(MD))
962     Info = &arrangeCXXDestructor(cast<CXXDestructorDecl>(MD), GD.getDtorType());
963   else
964     Info = &arrangeCXXMethodDeclaration(MD);
965   return GetFunctionType(*Info);
966 }
967 
968 void CodeGenModule::ConstructAttributeList(const CGFunctionInfo &FI,
969                                            const Decl *TargetDecl,
970                                            AttributeListType &PAL,
971                                            unsigned &CallingConv) {
972   llvm::AttrBuilder FuncAttrs;
973   llvm::AttrBuilder RetAttrs;
974 
975   CallingConv = FI.getEffectiveCallingConvention();
976 
977   if (FI.isNoReturn())
978     FuncAttrs.addAttribute(llvm::Attribute::NoReturn);
979 
980   // FIXME: handle sseregparm someday...
981   if (TargetDecl) {
982     if (TargetDecl->hasAttr<ReturnsTwiceAttr>())
983       FuncAttrs.addAttribute(llvm::Attribute::ReturnsTwice);
984     if (TargetDecl->hasAttr<NoThrowAttr>())
985       FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
986     if (TargetDecl->hasAttr<NoReturnAttr>())
987       FuncAttrs.addAttribute(llvm::Attribute::NoReturn);
988 
989     if (const FunctionDecl *Fn = dyn_cast<FunctionDecl>(TargetDecl)) {
990       const FunctionProtoType *FPT = Fn->getType()->getAs<FunctionProtoType>();
991       if (FPT && FPT->isNothrow(getContext()))
992         FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
993       if (Fn->isNoReturn())
994         FuncAttrs.addAttribute(llvm::Attribute::NoReturn);
995     }
996 
997     // 'const' and 'pure' attribute functions are also nounwind.
998     if (TargetDecl->hasAttr<ConstAttr>()) {
999       FuncAttrs.addAttribute(llvm::Attribute::ReadNone);
1000       FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1001     } else if (TargetDecl->hasAttr<PureAttr>()) {
1002       FuncAttrs.addAttribute(llvm::Attribute::ReadOnly);
1003       FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1004     }
1005     if (TargetDecl->hasAttr<MallocAttr>())
1006       RetAttrs.addAttribute(llvm::Attribute::NoAlias);
1007   }
1008 
1009   if (CodeGenOpts.OptimizeSize)
1010     FuncAttrs.addAttribute(llvm::Attribute::OptimizeForSize);
1011   if (CodeGenOpts.OptimizeSize == 2)
1012     FuncAttrs.addAttribute(llvm::Attribute::MinSize);
1013   if (CodeGenOpts.DisableRedZone)
1014     FuncAttrs.addAttribute(llvm::Attribute::NoRedZone);
1015   if (CodeGenOpts.NoImplicitFloat)
1016     FuncAttrs.addAttribute(llvm::Attribute::NoImplicitFloat);
1017 
1018   QualType RetTy = FI.getReturnType();
1019   unsigned Index = 1;
1020   const ABIArgInfo &RetAI = FI.getReturnInfo();
1021   switch (RetAI.getKind()) {
1022   case ABIArgInfo::Extend:
1023    if (RetTy->hasSignedIntegerRepresentation())
1024      RetAttrs.addAttribute(llvm::Attribute::SExt);
1025    else if (RetTy->hasUnsignedIntegerRepresentation())
1026      RetAttrs.addAttribute(llvm::Attribute::ZExt);
1027     break;
1028   case ABIArgInfo::Direct:
1029   case ABIArgInfo::Ignore:
1030     break;
1031 
1032   case ABIArgInfo::Indirect: {
1033     llvm::AttrBuilder SRETAttrs;
1034     SRETAttrs.addAttribute(llvm::Attribute::StructRet);
1035     if (RetAI.getInReg())
1036       SRETAttrs.addAttribute(llvm::Attribute::InReg);
1037     PAL.push_back(llvm::
1038                   AttributeSet::get(getLLVMContext(), Index, SRETAttrs));
1039 
1040     ++Index;
1041     // sret disables readnone and readonly
1042     FuncAttrs.removeAttribute(llvm::Attribute::ReadOnly)
1043       .removeAttribute(llvm::Attribute::ReadNone);
1044     break;
1045   }
1046 
1047   case ABIArgInfo::Expand:
1048     llvm_unreachable("Invalid ABI kind for return argument");
1049   }
1050 
1051   if (RetAttrs.hasAttributes())
1052     PAL.push_back(llvm::
1053                   AttributeSet::get(getLLVMContext(),
1054                                     llvm::AttributeSet::ReturnIndex,
1055                                     RetAttrs));
1056 
1057   for (CGFunctionInfo::const_arg_iterator it = FI.arg_begin(),
1058          ie = FI.arg_end(); it != ie; ++it) {
1059     QualType ParamType = it->type;
1060     const ABIArgInfo &AI = it->info;
1061     llvm::AttrBuilder Attrs;
1062 
1063     if (AI.getPaddingType()) {
1064       if (AI.getPaddingInReg())
1065         PAL.push_back(llvm::AttributeSet::get(getLLVMContext(), Index,
1066                                               llvm::Attribute::InReg));
1067       // Increment Index if there is padding.
1068       ++Index;
1069     }
1070 
1071     // 'restrict' -> 'noalias' is done in EmitFunctionProlog when we
1072     // have the corresponding parameter variable.  It doesn't make
1073     // sense to do it here because parameters are so messed up.
1074     switch (AI.getKind()) {
1075     case ABIArgInfo::Extend:
1076       if (ParamType->isSignedIntegerOrEnumerationType())
1077         Attrs.addAttribute(llvm::Attribute::SExt);
1078       else if (ParamType->isUnsignedIntegerOrEnumerationType())
1079         Attrs.addAttribute(llvm::Attribute::ZExt);
1080       // FALL THROUGH
1081     case ABIArgInfo::Direct:
1082       if (AI.getInReg())
1083         Attrs.addAttribute(llvm::Attribute::InReg);
1084 
1085       // FIXME: handle sseregparm someday...
1086 
1087       if (llvm::StructType *STy =
1088           dyn_cast<llvm::StructType>(AI.getCoerceToType())) {
1089         unsigned Extra = STy->getNumElements()-1;  // 1 will be added below.
1090         if (Attrs.hasAttributes())
1091           for (unsigned I = 0; I < Extra; ++I)
1092             PAL.push_back(llvm::AttributeSet::get(getLLVMContext(), Index + I,
1093                                                   Attrs));
1094         Index += Extra;
1095       }
1096       break;
1097 
1098     case ABIArgInfo::Indirect:
1099       if (AI.getInReg())
1100         Attrs.addAttribute(llvm::Attribute::InReg);
1101 
1102       if (AI.getIndirectByVal())
1103         Attrs.addAttribute(llvm::Attribute::ByVal);
1104 
1105       Attrs.addAlignmentAttr(AI.getIndirectAlign());
1106 
1107       // byval disables readnone and readonly.
1108       FuncAttrs.removeAttribute(llvm::Attribute::ReadOnly)
1109         .removeAttribute(llvm::Attribute::ReadNone);
1110       break;
1111 
1112     case ABIArgInfo::Ignore:
1113       // Skip increment, no matching LLVM parameter.
1114       continue;
1115 
1116     case ABIArgInfo::Expand: {
1117       SmallVector<llvm::Type*, 8> types;
1118       // FIXME: This is rather inefficient. Do we ever actually need to do
1119       // anything here? The result should be just reconstructed on the other
1120       // side, so extension should be a non-issue.
1121       getTypes().GetExpandedTypes(ParamType, types);
1122       Index += types.size();
1123       continue;
1124     }
1125     }
1126 
1127     if (Attrs.hasAttributes())
1128       PAL.push_back(llvm::AttributeSet::get(getLLVMContext(), Index, Attrs));
1129     ++Index;
1130   }
1131   if (FuncAttrs.hasAttributes())
1132     PAL.push_back(llvm::
1133                   AttributeSet::get(getLLVMContext(),
1134                                     llvm::AttributeSet::FunctionIndex,
1135                                     FuncAttrs));
1136 }
1137 
1138 /// An argument came in as a promoted argument; demote it back to its
1139 /// declared type.
1140 static llvm::Value *emitArgumentDemotion(CodeGenFunction &CGF,
1141                                          const VarDecl *var,
1142                                          llvm::Value *value) {
1143   llvm::Type *varType = CGF.ConvertType(var->getType());
1144 
1145   // This can happen with promotions that actually don't change the
1146   // underlying type, like the enum promotions.
1147   if (value->getType() == varType) return value;
1148 
1149   assert((varType->isIntegerTy() || varType->isFloatingPointTy())
1150          && "unexpected promotion type");
1151 
1152   if (isa<llvm::IntegerType>(varType))
1153     return CGF.Builder.CreateTrunc(value, varType, "arg.unpromote");
1154 
1155   return CGF.Builder.CreateFPCast(value, varType, "arg.unpromote");
1156 }
1157 
1158 void CodeGenFunction::EmitFunctionProlog(const CGFunctionInfo &FI,
1159                                          llvm::Function *Fn,
1160                                          const FunctionArgList &Args) {
1161   // If this is an implicit-return-zero function, go ahead and
1162   // initialize the return value.  TODO: it might be nice to have
1163   // a more general mechanism for this that didn't require synthesized
1164   // return statements.
1165   if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(CurFuncDecl)) {
1166     if (FD->hasImplicitReturnZero()) {
1167       QualType RetTy = FD->getResultType().getUnqualifiedType();
1168       llvm::Type* LLVMTy = CGM.getTypes().ConvertType(RetTy);
1169       llvm::Constant* Zero = llvm::Constant::getNullValue(LLVMTy);
1170       Builder.CreateStore(Zero, ReturnValue);
1171     }
1172   }
1173 
1174   // FIXME: We no longer need the types from FunctionArgList; lift up and
1175   // simplify.
1176 
1177   // Emit allocs for param decls.  Give the LLVM Argument nodes names.
1178   llvm::Function::arg_iterator AI = Fn->arg_begin();
1179 
1180   // Name the struct return argument.
1181   if (CGM.ReturnTypeUsesSRet(FI)) {
1182     AI->setName("agg.result");
1183     AI->addAttr(llvm::AttributeSet::get(getLLVMContext(),
1184                                         AI->getArgNo() + 1,
1185                                         llvm::Attribute::NoAlias));
1186     ++AI;
1187   }
1188 
1189   assert(FI.arg_size() == Args.size() &&
1190          "Mismatch between function signature & arguments.");
1191   unsigned ArgNo = 1;
1192   CGFunctionInfo::const_arg_iterator info_it = FI.arg_begin();
1193   for (FunctionArgList::const_iterator i = Args.begin(), e = Args.end();
1194        i != e; ++i, ++info_it, ++ArgNo) {
1195     const VarDecl *Arg = *i;
1196     QualType Ty = info_it->type;
1197     const ABIArgInfo &ArgI = info_it->info;
1198 
1199     bool isPromoted =
1200       isa<ParmVarDecl>(Arg) && cast<ParmVarDecl>(Arg)->isKNRPromoted();
1201 
1202     // Skip the dummy padding argument.
1203     if (ArgI.getPaddingType())
1204       ++AI;
1205 
1206     switch (ArgI.getKind()) {
1207     case ABIArgInfo::Indirect: {
1208       llvm::Value *V = AI;
1209 
1210       if (hasAggregateLLVMType(Ty)) {
1211         // Aggregates and complex variables are accessed by reference.  All we
1212         // need to do is realign the value, if requested
1213         if (ArgI.getIndirectRealign()) {
1214           llvm::Value *AlignedTemp = CreateMemTemp(Ty, "coerce");
1215 
1216           // Copy from the incoming argument pointer to the temporary with the
1217           // appropriate alignment.
1218           //
1219           // FIXME: We should have a common utility for generating an aggregate
1220           // copy.
1221           llvm::Type *I8PtrTy = Builder.getInt8PtrTy();
1222           CharUnits Size = getContext().getTypeSizeInChars(Ty);
1223           llvm::Value *Dst = Builder.CreateBitCast(AlignedTemp, I8PtrTy);
1224           llvm::Value *Src = Builder.CreateBitCast(V, I8PtrTy);
1225           Builder.CreateMemCpy(Dst,
1226                                Src,
1227                                llvm::ConstantInt::get(IntPtrTy,
1228                                                       Size.getQuantity()),
1229                                ArgI.getIndirectAlign(),
1230                                false);
1231           V = AlignedTemp;
1232         }
1233       } else {
1234         // Load scalar value from indirect argument.
1235         CharUnits Alignment = getContext().getTypeAlignInChars(Ty);
1236         V = EmitLoadOfScalar(V, false, Alignment.getQuantity(), Ty);
1237 
1238         if (isPromoted)
1239           V = emitArgumentDemotion(*this, Arg, V);
1240       }
1241       EmitParmDecl(*Arg, V, ArgNo);
1242       break;
1243     }
1244 
1245     case ABIArgInfo::Extend:
1246     case ABIArgInfo::Direct: {
1247 
1248       // If we have the trivial case, handle it with no muss and fuss.
1249       if (!isa<llvm::StructType>(ArgI.getCoerceToType()) &&
1250           ArgI.getCoerceToType() == ConvertType(Ty) &&
1251           ArgI.getDirectOffset() == 0) {
1252         assert(AI != Fn->arg_end() && "Argument mismatch!");
1253         llvm::Value *V = AI;
1254 
1255         if (Arg->getType().isRestrictQualified())
1256           AI->addAttr(llvm::AttributeSet::get(getLLVMContext(),
1257                                               AI->getArgNo() + 1,
1258                                               llvm::Attribute::NoAlias));
1259 
1260         // Ensure the argument is the correct type.
1261         if (V->getType() != ArgI.getCoerceToType())
1262           V = Builder.CreateBitCast(V, ArgI.getCoerceToType());
1263 
1264         if (isPromoted)
1265           V = emitArgumentDemotion(*this, Arg, V);
1266 
1267         // Because of merging of function types from multiple decls it is
1268         // possible for the type of an argument to not match the corresponding
1269         // type in the function type. Since we are codegening the callee
1270         // in here, add a cast to the argument type.
1271         llvm::Type *LTy = ConvertType(Arg->getType());
1272         if (V->getType() != LTy)
1273           V = Builder.CreateBitCast(V, LTy);
1274 
1275         EmitParmDecl(*Arg, V, ArgNo);
1276         break;
1277       }
1278 
1279       llvm::AllocaInst *Alloca = CreateMemTemp(Ty, Arg->getName());
1280 
1281       // The alignment we need to use is the max of the requested alignment for
1282       // the argument plus the alignment required by our access code below.
1283       unsigned AlignmentToUse =
1284         CGM.getDataLayout().getABITypeAlignment(ArgI.getCoerceToType());
1285       AlignmentToUse = std::max(AlignmentToUse,
1286                         (unsigned)getContext().getDeclAlign(Arg).getQuantity());
1287 
1288       Alloca->setAlignment(AlignmentToUse);
1289       llvm::Value *V = Alloca;
1290       llvm::Value *Ptr = V;    // Pointer to store into.
1291 
1292       // If the value is offset in memory, apply the offset now.
1293       if (unsigned Offs = ArgI.getDirectOffset()) {
1294         Ptr = Builder.CreateBitCast(Ptr, Builder.getInt8PtrTy());
1295         Ptr = Builder.CreateConstGEP1_32(Ptr, Offs);
1296         Ptr = Builder.CreateBitCast(Ptr,
1297                           llvm::PointerType::getUnqual(ArgI.getCoerceToType()));
1298       }
1299 
1300       // If the coerce-to type is a first class aggregate, we flatten it and
1301       // pass the elements. Either way is semantically identical, but fast-isel
1302       // and the optimizer generally likes scalar values better than FCAs.
1303       llvm::StructType *STy = dyn_cast<llvm::StructType>(ArgI.getCoerceToType());
1304       if (STy && STy->getNumElements() > 1) {
1305         uint64_t SrcSize = CGM.getDataLayout().getTypeAllocSize(STy);
1306         llvm::Type *DstTy =
1307           cast<llvm::PointerType>(Ptr->getType())->getElementType();
1308         uint64_t DstSize = CGM.getDataLayout().getTypeAllocSize(DstTy);
1309 
1310         if (SrcSize <= DstSize) {
1311           Ptr = Builder.CreateBitCast(Ptr, llvm::PointerType::getUnqual(STy));
1312 
1313           for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1314             assert(AI != Fn->arg_end() && "Argument mismatch!");
1315             AI->setName(Arg->getName() + ".coerce" + Twine(i));
1316             llvm::Value *EltPtr = Builder.CreateConstGEP2_32(Ptr, 0, i);
1317             Builder.CreateStore(AI++, EltPtr);
1318           }
1319         } else {
1320           llvm::AllocaInst *TempAlloca =
1321             CreateTempAlloca(ArgI.getCoerceToType(), "coerce");
1322           TempAlloca->setAlignment(AlignmentToUse);
1323           llvm::Value *TempV = TempAlloca;
1324 
1325           for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1326             assert(AI != Fn->arg_end() && "Argument mismatch!");
1327             AI->setName(Arg->getName() + ".coerce" + Twine(i));
1328             llvm::Value *EltPtr = Builder.CreateConstGEP2_32(TempV, 0, i);
1329             Builder.CreateStore(AI++, EltPtr);
1330           }
1331 
1332           Builder.CreateMemCpy(Ptr, TempV, DstSize, AlignmentToUse);
1333         }
1334       } else {
1335         // Simple case, just do a coerced store of the argument into the alloca.
1336         assert(AI != Fn->arg_end() && "Argument mismatch!");
1337         AI->setName(Arg->getName() + ".coerce");
1338         CreateCoercedStore(AI++, Ptr, /*DestIsVolatile=*/false, *this);
1339       }
1340 
1341 
1342       // Match to what EmitParmDecl is expecting for this type.
1343       if (!CodeGenFunction::hasAggregateLLVMType(Ty)) {
1344         V = EmitLoadOfScalar(V, false, AlignmentToUse, Ty);
1345         if (isPromoted)
1346           V = emitArgumentDemotion(*this, Arg, V);
1347       }
1348       EmitParmDecl(*Arg, V, ArgNo);
1349       continue;  // Skip ++AI increment, already done.
1350     }
1351 
1352     case ABIArgInfo::Expand: {
1353       // If this structure was expanded into multiple arguments then
1354       // we need to create a temporary and reconstruct it from the
1355       // arguments.
1356       llvm::AllocaInst *Alloca = CreateMemTemp(Ty);
1357       CharUnits Align = getContext().getDeclAlign(Arg);
1358       Alloca->setAlignment(Align.getQuantity());
1359       LValue LV = MakeAddrLValue(Alloca, Ty, Align);
1360       llvm::Function::arg_iterator End = ExpandTypeFromArgs(Ty, LV, AI);
1361       EmitParmDecl(*Arg, Alloca, ArgNo);
1362 
1363       // Name the arguments used in expansion and increment AI.
1364       unsigned Index = 0;
1365       for (; AI != End; ++AI, ++Index)
1366         AI->setName(Arg->getName() + "." + Twine(Index));
1367       continue;
1368     }
1369 
1370     case ABIArgInfo::Ignore:
1371       // Initialize the local variable appropriately.
1372       if (hasAggregateLLVMType(Ty))
1373         EmitParmDecl(*Arg, CreateMemTemp(Ty), ArgNo);
1374       else
1375         EmitParmDecl(*Arg, llvm::UndefValue::get(ConvertType(Arg->getType())),
1376                      ArgNo);
1377 
1378       // Skip increment, no matching LLVM parameter.
1379       continue;
1380     }
1381 
1382     ++AI;
1383   }
1384   assert(AI == Fn->arg_end() && "Argument mismatch!");
1385 }
1386 
1387 static void eraseUnusedBitCasts(llvm::Instruction *insn) {
1388   while (insn->use_empty()) {
1389     llvm::BitCastInst *bitcast = dyn_cast<llvm::BitCastInst>(insn);
1390     if (!bitcast) return;
1391 
1392     // This is "safe" because we would have used a ConstantExpr otherwise.
1393     insn = cast<llvm::Instruction>(bitcast->getOperand(0));
1394     bitcast->eraseFromParent();
1395   }
1396 }
1397 
1398 /// Try to emit a fused autorelease of a return result.
1399 static llvm::Value *tryEmitFusedAutoreleaseOfResult(CodeGenFunction &CGF,
1400                                                     llvm::Value *result) {
1401   // We must be immediately followed the cast.
1402   llvm::BasicBlock *BB = CGF.Builder.GetInsertBlock();
1403   if (BB->empty()) return 0;
1404   if (&BB->back() != result) return 0;
1405 
1406   llvm::Type *resultType = result->getType();
1407 
1408   // result is in a BasicBlock and is therefore an Instruction.
1409   llvm::Instruction *generator = cast<llvm::Instruction>(result);
1410 
1411   SmallVector<llvm::Instruction*,4> insnsToKill;
1412 
1413   // Look for:
1414   //  %generator = bitcast %type1* %generator2 to %type2*
1415   while (llvm::BitCastInst *bitcast = dyn_cast<llvm::BitCastInst>(generator)) {
1416     // We would have emitted this as a constant if the operand weren't
1417     // an Instruction.
1418     generator = cast<llvm::Instruction>(bitcast->getOperand(0));
1419 
1420     // Require the generator to be immediately followed by the cast.
1421     if (generator->getNextNode() != bitcast)
1422       return 0;
1423 
1424     insnsToKill.push_back(bitcast);
1425   }
1426 
1427   // Look for:
1428   //   %generator = call i8* @objc_retain(i8* %originalResult)
1429   // or
1430   //   %generator = call i8* @objc_retainAutoreleasedReturnValue(i8* %originalResult)
1431   llvm::CallInst *call = dyn_cast<llvm::CallInst>(generator);
1432   if (!call) return 0;
1433 
1434   bool doRetainAutorelease;
1435 
1436   if (call->getCalledValue() == CGF.CGM.getARCEntrypoints().objc_retain) {
1437     doRetainAutorelease = true;
1438   } else if (call->getCalledValue() == CGF.CGM.getARCEntrypoints()
1439                                           .objc_retainAutoreleasedReturnValue) {
1440     doRetainAutorelease = false;
1441 
1442     // If we emitted an assembly marker for this call (and the
1443     // ARCEntrypoints field should have been set if so), go looking
1444     // for that call.  If we can't find it, we can't do this
1445     // optimization.  But it should always be the immediately previous
1446     // instruction, unless we needed bitcasts around the call.
1447     if (CGF.CGM.getARCEntrypoints().retainAutoreleasedReturnValueMarker) {
1448       llvm::Instruction *prev = call->getPrevNode();
1449       assert(prev);
1450       if (isa<llvm::BitCastInst>(prev)) {
1451         prev = prev->getPrevNode();
1452         assert(prev);
1453       }
1454       assert(isa<llvm::CallInst>(prev));
1455       assert(cast<llvm::CallInst>(prev)->getCalledValue() ==
1456                CGF.CGM.getARCEntrypoints().retainAutoreleasedReturnValueMarker);
1457       insnsToKill.push_back(prev);
1458     }
1459   } else {
1460     return 0;
1461   }
1462 
1463   result = call->getArgOperand(0);
1464   insnsToKill.push_back(call);
1465 
1466   // Keep killing bitcasts, for sanity.  Note that we no longer care
1467   // about precise ordering as long as there's exactly one use.
1468   while (llvm::BitCastInst *bitcast = dyn_cast<llvm::BitCastInst>(result)) {
1469     if (!bitcast->hasOneUse()) break;
1470     insnsToKill.push_back(bitcast);
1471     result = bitcast->getOperand(0);
1472   }
1473 
1474   // Delete all the unnecessary instructions, from latest to earliest.
1475   for (SmallVectorImpl<llvm::Instruction*>::iterator
1476          i = insnsToKill.begin(), e = insnsToKill.end(); i != e; ++i)
1477     (*i)->eraseFromParent();
1478 
1479   // Do the fused retain/autorelease if we were asked to.
1480   if (doRetainAutorelease)
1481     result = CGF.EmitARCRetainAutoreleaseReturnValue(result);
1482 
1483   // Cast back to the result type.
1484   return CGF.Builder.CreateBitCast(result, resultType);
1485 }
1486 
1487 /// If this is a +1 of the value of an immutable 'self', remove it.
1488 static llvm::Value *tryRemoveRetainOfSelf(CodeGenFunction &CGF,
1489                                           llvm::Value *result) {
1490   // This is only applicable to a method with an immutable 'self'.
1491   const ObjCMethodDecl *method =
1492     dyn_cast_or_null<ObjCMethodDecl>(CGF.CurCodeDecl);
1493   if (!method) return 0;
1494   const VarDecl *self = method->getSelfDecl();
1495   if (!self->getType().isConstQualified()) return 0;
1496 
1497   // Look for a retain call.
1498   llvm::CallInst *retainCall =
1499     dyn_cast<llvm::CallInst>(result->stripPointerCasts());
1500   if (!retainCall ||
1501       retainCall->getCalledValue() != CGF.CGM.getARCEntrypoints().objc_retain)
1502     return 0;
1503 
1504   // Look for an ordinary load of 'self'.
1505   llvm::Value *retainedValue = retainCall->getArgOperand(0);
1506   llvm::LoadInst *load =
1507     dyn_cast<llvm::LoadInst>(retainedValue->stripPointerCasts());
1508   if (!load || load->isAtomic() || load->isVolatile() ||
1509       load->getPointerOperand() != CGF.GetAddrOfLocalVar(self))
1510     return 0;
1511 
1512   // Okay!  Burn it all down.  This relies for correctness on the
1513   // assumption that the retain is emitted as part of the return and
1514   // that thereafter everything is used "linearly".
1515   llvm::Type *resultType = result->getType();
1516   eraseUnusedBitCasts(cast<llvm::Instruction>(result));
1517   assert(retainCall->use_empty());
1518   retainCall->eraseFromParent();
1519   eraseUnusedBitCasts(cast<llvm::Instruction>(retainedValue));
1520 
1521   return CGF.Builder.CreateBitCast(load, resultType);
1522 }
1523 
1524 /// Emit an ARC autorelease of the result of a function.
1525 ///
1526 /// \return the value to actually return from the function
1527 static llvm::Value *emitAutoreleaseOfResult(CodeGenFunction &CGF,
1528                                             llvm::Value *result) {
1529   // If we're returning 'self', kill the initial retain.  This is a
1530   // heuristic attempt to "encourage correctness" in the really unfortunate
1531   // case where we have a return of self during a dealloc and we desperately
1532   // need to avoid the possible autorelease.
1533   if (llvm::Value *self = tryRemoveRetainOfSelf(CGF, result))
1534     return self;
1535 
1536   // At -O0, try to emit a fused retain/autorelease.
1537   if (CGF.shouldUseFusedARCCalls())
1538     if (llvm::Value *fused = tryEmitFusedAutoreleaseOfResult(CGF, result))
1539       return fused;
1540 
1541   return CGF.EmitARCAutoreleaseReturnValue(result);
1542 }
1543 
1544 /// Heuristically search for a dominating store to the return-value slot.
1545 static llvm::StoreInst *findDominatingStoreToReturnValue(CodeGenFunction &CGF) {
1546   // If there are multiple uses of the return-value slot, just check
1547   // for something immediately preceding the IP.  Sometimes this can
1548   // happen with how we generate implicit-returns; it can also happen
1549   // with noreturn cleanups.
1550   if (!CGF.ReturnValue->hasOneUse()) {
1551     llvm::BasicBlock *IP = CGF.Builder.GetInsertBlock();
1552     if (IP->empty()) return 0;
1553     llvm::StoreInst *store = dyn_cast<llvm::StoreInst>(&IP->back());
1554     if (!store) return 0;
1555     if (store->getPointerOperand() != CGF.ReturnValue) return 0;
1556     assert(!store->isAtomic() && !store->isVolatile()); // see below
1557     return store;
1558   }
1559 
1560   llvm::StoreInst *store =
1561     dyn_cast<llvm::StoreInst>(CGF.ReturnValue->use_back());
1562   if (!store) return 0;
1563 
1564   // These aren't actually possible for non-coerced returns, and we
1565   // only care about non-coerced returns on this code path.
1566   assert(!store->isAtomic() && !store->isVolatile());
1567 
1568   // Now do a first-and-dirty dominance check: just walk up the
1569   // single-predecessors chain from the current insertion point.
1570   llvm::BasicBlock *StoreBB = store->getParent();
1571   llvm::BasicBlock *IP = CGF.Builder.GetInsertBlock();
1572   while (IP != StoreBB) {
1573     if (!(IP = IP->getSinglePredecessor()))
1574       return 0;
1575   }
1576 
1577   // Okay, the store's basic block dominates the insertion point; we
1578   // can do our thing.
1579   return store;
1580 }
1581 
1582 void CodeGenFunction::EmitFunctionEpilog(const CGFunctionInfo &FI) {
1583   // Functions with no result always return void.
1584   if (ReturnValue == 0) {
1585     Builder.CreateRetVoid();
1586     return;
1587   }
1588 
1589   llvm::DebugLoc RetDbgLoc;
1590   llvm::Value *RV = 0;
1591   QualType RetTy = FI.getReturnType();
1592   const ABIArgInfo &RetAI = FI.getReturnInfo();
1593 
1594   switch (RetAI.getKind()) {
1595   case ABIArgInfo::Indirect: {
1596     unsigned Alignment = getContext().getTypeAlignInChars(RetTy).getQuantity();
1597     if (RetTy->isAnyComplexType()) {
1598       ComplexPairTy RT = LoadComplexFromAddr(ReturnValue, false);
1599       StoreComplexToAddr(RT, CurFn->arg_begin(), false);
1600     } else if (CodeGenFunction::hasAggregateLLVMType(RetTy)) {
1601       // Do nothing; aggregrates get evaluated directly into the destination.
1602     } else {
1603       EmitStoreOfScalar(Builder.CreateLoad(ReturnValue), CurFn->arg_begin(),
1604                         false, Alignment, RetTy);
1605     }
1606     break;
1607   }
1608 
1609   case ABIArgInfo::Extend:
1610   case ABIArgInfo::Direct:
1611     if (RetAI.getCoerceToType() == ConvertType(RetTy) &&
1612         RetAI.getDirectOffset() == 0) {
1613       // The internal return value temp always will have pointer-to-return-type
1614       // type, just do a load.
1615 
1616       // If there is a dominating store to ReturnValue, we can elide
1617       // the load, zap the store, and usually zap the alloca.
1618       if (llvm::StoreInst *SI = findDominatingStoreToReturnValue(*this)) {
1619         // Get the stored value and nuke the now-dead store.
1620         RetDbgLoc = SI->getDebugLoc();
1621         RV = SI->getValueOperand();
1622         SI->eraseFromParent();
1623 
1624         // If that was the only use of the return value, nuke it as well now.
1625         if (ReturnValue->use_empty() && isa<llvm::AllocaInst>(ReturnValue)) {
1626           cast<llvm::AllocaInst>(ReturnValue)->eraseFromParent();
1627           ReturnValue = 0;
1628         }
1629 
1630       // Otherwise, we have to do a simple load.
1631       } else {
1632         RV = Builder.CreateLoad(ReturnValue);
1633       }
1634     } else {
1635       llvm::Value *V = ReturnValue;
1636       // If the value is offset in memory, apply the offset now.
1637       if (unsigned Offs = RetAI.getDirectOffset()) {
1638         V = Builder.CreateBitCast(V, Builder.getInt8PtrTy());
1639         V = Builder.CreateConstGEP1_32(V, Offs);
1640         V = Builder.CreateBitCast(V,
1641                          llvm::PointerType::getUnqual(RetAI.getCoerceToType()));
1642       }
1643 
1644       RV = CreateCoercedLoad(V, RetAI.getCoerceToType(), *this);
1645     }
1646 
1647     // In ARC, end functions that return a retainable type with a call
1648     // to objc_autoreleaseReturnValue.
1649     if (AutoreleaseResult) {
1650       assert(getLangOpts().ObjCAutoRefCount &&
1651              !FI.isReturnsRetained() &&
1652              RetTy->isObjCRetainableType());
1653       RV = emitAutoreleaseOfResult(*this, RV);
1654     }
1655 
1656     break;
1657 
1658   case ABIArgInfo::Ignore:
1659     break;
1660 
1661   case ABIArgInfo::Expand:
1662     llvm_unreachable("Invalid ABI kind for return argument");
1663   }
1664 
1665   llvm::Instruction *Ret = RV ? Builder.CreateRet(RV) : Builder.CreateRetVoid();
1666   if (!RetDbgLoc.isUnknown())
1667     Ret->setDebugLoc(RetDbgLoc);
1668 }
1669 
1670 void CodeGenFunction::EmitDelegateCallArg(CallArgList &args,
1671                                           const VarDecl *param) {
1672   // StartFunction converted the ABI-lowered parameter(s) into a
1673   // local alloca.  We need to turn that into an r-value suitable
1674   // for EmitCall.
1675   llvm::Value *local = GetAddrOfLocalVar(param);
1676 
1677   QualType type = param->getType();
1678 
1679   // For the most part, we just need to load the alloca, except:
1680   // 1) aggregate r-values are actually pointers to temporaries, and
1681   // 2) references to aggregates are pointers directly to the aggregate.
1682   // I don't know why references to non-aggregates are different here.
1683   if (const ReferenceType *ref = type->getAs<ReferenceType>()) {
1684     if (hasAggregateLLVMType(ref->getPointeeType()))
1685       return args.add(RValue::getAggregate(local), type);
1686 
1687     // Locals which are references to scalars are represented
1688     // with allocas holding the pointer.
1689     return args.add(RValue::get(Builder.CreateLoad(local)), type);
1690   }
1691 
1692   if (type->isAnyComplexType()) {
1693     ComplexPairTy complex = LoadComplexFromAddr(local, /*volatile*/ false);
1694     return args.add(RValue::getComplex(complex), type);
1695   }
1696 
1697   if (hasAggregateLLVMType(type))
1698     return args.add(RValue::getAggregate(local), type);
1699 
1700   unsigned alignment = getContext().getDeclAlign(param).getQuantity();
1701   llvm::Value *value = EmitLoadOfScalar(local, false, alignment, type);
1702   return args.add(RValue::get(value), type);
1703 }
1704 
1705 static bool isProvablyNull(llvm::Value *addr) {
1706   return isa<llvm::ConstantPointerNull>(addr);
1707 }
1708 
1709 static bool isProvablyNonNull(llvm::Value *addr) {
1710   return isa<llvm::AllocaInst>(addr);
1711 }
1712 
1713 /// Emit the actual writing-back of a writeback.
1714 static void emitWriteback(CodeGenFunction &CGF,
1715                           const CallArgList::Writeback &writeback) {
1716   llvm::Value *srcAddr = writeback.Address;
1717   assert(!isProvablyNull(srcAddr) &&
1718          "shouldn't have writeback for provably null argument");
1719 
1720   llvm::BasicBlock *contBB = 0;
1721 
1722   // If the argument wasn't provably non-null, we need to null check
1723   // before doing the store.
1724   bool provablyNonNull = isProvablyNonNull(srcAddr);
1725   if (!provablyNonNull) {
1726     llvm::BasicBlock *writebackBB = CGF.createBasicBlock("icr.writeback");
1727     contBB = CGF.createBasicBlock("icr.done");
1728 
1729     llvm::Value *isNull = CGF.Builder.CreateIsNull(srcAddr, "icr.isnull");
1730     CGF.Builder.CreateCondBr(isNull, contBB, writebackBB);
1731     CGF.EmitBlock(writebackBB);
1732   }
1733 
1734   // Load the value to writeback.
1735   llvm::Value *value = CGF.Builder.CreateLoad(writeback.Temporary);
1736 
1737   // Cast it back, in case we're writing an id to a Foo* or something.
1738   value = CGF.Builder.CreateBitCast(value,
1739                cast<llvm::PointerType>(srcAddr->getType())->getElementType(),
1740                             "icr.writeback-cast");
1741 
1742   // Perform the writeback.
1743   QualType srcAddrType = writeback.AddressType;
1744   CGF.EmitStoreThroughLValue(RValue::get(value),
1745                              CGF.MakeAddrLValue(srcAddr, srcAddrType));
1746 
1747   // Jump to the continuation block.
1748   if (!provablyNonNull)
1749     CGF.EmitBlock(contBB);
1750 }
1751 
1752 static void emitWritebacks(CodeGenFunction &CGF,
1753                            const CallArgList &args) {
1754   for (CallArgList::writeback_iterator
1755          i = args.writeback_begin(), e = args.writeback_end(); i != e; ++i)
1756     emitWriteback(CGF, *i);
1757 }
1758 
1759 /// Emit an argument that's being passed call-by-writeback.  That is,
1760 /// we are passing the address of
1761 static void emitWritebackArg(CodeGenFunction &CGF, CallArgList &args,
1762                              const ObjCIndirectCopyRestoreExpr *CRE) {
1763   llvm::Value *srcAddr = CGF.EmitScalarExpr(CRE->getSubExpr());
1764 
1765   // The dest and src types don't necessarily match in LLVM terms
1766   // because of the crazy ObjC compatibility rules.
1767 
1768   llvm::PointerType *destType =
1769     cast<llvm::PointerType>(CGF.ConvertType(CRE->getType()));
1770 
1771   // If the address is a constant null, just pass the appropriate null.
1772   if (isProvablyNull(srcAddr)) {
1773     args.add(RValue::get(llvm::ConstantPointerNull::get(destType)),
1774              CRE->getType());
1775     return;
1776   }
1777 
1778   QualType srcAddrType =
1779     CRE->getSubExpr()->getType()->castAs<PointerType>()->getPointeeType();
1780 
1781   // Create the temporary.
1782   llvm::Value *temp = CGF.CreateTempAlloca(destType->getElementType(),
1783                                            "icr.temp");
1784   // Loading an l-value can introduce a cleanup if the l-value is __weak,
1785   // and that cleanup will be conditional if we can't prove that the l-value
1786   // isn't null, so we need to register a dominating point so that the cleanups
1787   // system will make valid IR.
1788   CodeGenFunction::ConditionalEvaluation condEval(CGF);
1789 
1790   // Zero-initialize it if we're not doing a copy-initialization.
1791   bool shouldCopy = CRE->shouldCopy();
1792   if (!shouldCopy) {
1793     llvm::Value *null =
1794       llvm::ConstantPointerNull::get(
1795         cast<llvm::PointerType>(destType->getElementType()));
1796     CGF.Builder.CreateStore(null, temp);
1797   }
1798 
1799   llvm::BasicBlock *contBB = 0;
1800 
1801   // If the address is *not* known to be non-null, we need to switch.
1802   llvm::Value *finalArgument;
1803 
1804   bool provablyNonNull = isProvablyNonNull(srcAddr);
1805   if (provablyNonNull) {
1806     finalArgument = temp;
1807   } else {
1808     llvm::Value *isNull = CGF.Builder.CreateIsNull(srcAddr, "icr.isnull");
1809 
1810     finalArgument = CGF.Builder.CreateSelect(isNull,
1811                                    llvm::ConstantPointerNull::get(destType),
1812                                              temp, "icr.argument");
1813 
1814     // If we need to copy, then the load has to be conditional, which
1815     // means we need control flow.
1816     if (shouldCopy) {
1817       contBB = CGF.createBasicBlock("icr.cont");
1818       llvm::BasicBlock *copyBB = CGF.createBasicBlock("icr.copy");
1819       CGF.Builder.CreateCondBr(isNull, contBB, copyBB);
1820       CGF.EmitBlock(copyBB);
1821       condEval.begin(CGF);
1822     }
1823   }
1824 
1825   // Perform a copy if necessary.
1826   if (shouldCopy) {
1827     LValue srcLV = CGF.MakeAddrLValue(srcAddr, srcAddrType);
1828     RValue srcRV = CGF.EmitLoadOfLValue(srcLV);
1829     assert(srcRV.isScalar());
1830 
1831     llvm::Value *src = srcRV.getScalarVal();
1832     src = CGF.Builder.CreateBitCast(src, destType->getElementType(),
1833                                     "icr.cast");
1834 
1835     // Use an ordinary store, not a store-to-lvalue.
1836     CGF.Builder.CreateStore(src, temp);
1837   }
1838 
1839   // Finish the control flow if we needed it.
1840   if (shouldCopy && !provablyNonNull) {
1841     CGF.EmitBlock(contBB);
1842     condEval.end(CGF);
1843   }
1844 
1845   args.addWriteback(srcAddr, srcAddrType, temp);
1846   args.add(RValue::get(finalArgument), CRE->getType());
1847 }
1848 
1849 void CodeGenFunction::EmitCallArg(CallArgList &args, const Expr *E,
1850                                   QualType type) {
1851   if (const ObjCIndirectCopyRestoreExpr *CRE
1852         = dyn_cast<ObjCIndirectCopyRestoreExpr>(E)) {
1853     assert(getLangOpts().ObjCAutoRefCount);
1854     assert(getContext().hasSameType(E->getType(), type));
1855     return emitWritebackArg(*this, args, CRE);
1856   }
1857 
1858   assert(type->isReferenceType() == E->isGLValue() &&
1859          "reference binding to unmaterialized r-value!");
1860 
1861   if (E->isGLValue()) {
1862     assert(E->getObjectKind() == OK_Ordinary);
1863     return args.add(EmitReferenceBindingToExpr(E, /*InitializedDecl=*/0),
1864                     type);
1865   }
1866 
1867   if (hasAggregateLLVMType(type) && !E->getType()->isAnyComplexType() &&
1868       isa<ImplicitCastExpr>(E) &&
1869       cast<CastExpr>(E)->getCastKind() == CK_LValueToRValue) {
1870     LValue L = EmitLValue(cast<CastExpr>(E)->getSubExpr());
1871     assert(L.isSimple());
1872     args.add(L.asAggregateRValue(), type, /*NeedsCopy*/true);
1873     return;
1874   }
1875 
1876   args.add(EmitAnyExprToTemp(E), type);
1877 }
1878 
1879 // In ObjC ARC mode with no ObjC ARC exception safety, tell the ARC
1880 // optimizer it can aggressively ignore unwind edges.
1881 void
1882 CodeGenFunction::AddObjCARCExceptionMetadata(llvm::Instruction *Inst) {
1883   if (CGM.getCodeGenOpts().OptimizationLevel != 0 &&
1884       !CGM.getCodeGenOpts().ObjCAutoRefCountExceptions)
1885     Inst->setMetadata("clang.arc.no_objc_arc_exceptions",
1886                       CGM.getNoObjCARCExceptionsMetadata());
1887 }
1888 
1889 /// Emits a call or invoke instruction to the given function, depending
1890 /// on the current state of the EH stack.
1891 llvm::CallSite
1892 CodeGenFunction::EmitCallOrInvoke(llvm::Value *Callee,
1893                                   ArrayRef<llvm::Value *> Args,
1894                                   const Twine &Name) {
1895   llvm::BasicBlock *InvokeDest = getInvokeDest();
1896 
1897   llvm::Instruction *Inst;
1898   if (!InvokeDest)
1899     Inst = Builder.CreateCall(Callee, Args, Name);
1900   else {
1901     llvm::BasicBlock *ContBB = createBasicBlock("invoke.cont");
1902     Inst = Builder.CreateInvoke(Callee, ContBB, InvokeDest, Args, Name);
1903     EmitBlock(ContBB);
1904   }
1905 
1906   // In ObjC ARC mode with no ObjC ARC exception safety, tell the ARC
1907   // optimizer it can aggressively ignore unwind edges.
1908   if (CGM.getLangOpts().ObjCAutoRefCount)
1909     AddObjCARCExceptionMetadata(Inst);
1910 
1911   return Inst;
1912 }
1913 
1914 llvm::CallSite
1915 CodeGenFunction::EmitCallOrInvoke(llvm::Value *Callee,
1916                                   const Twine &Name) {
1917   return EmitCallOrInvoke(Callee, ArrayRef<llvm::Value *>(), Name);
1918 }
1919 
1920 static void checkArgMatches(llvm::Value *Elt, unsigned &ArgNo,
1921                             llvm::FunctionType *FTy) {
1922   if (ArgNo < FTy->getNumParams())
1923     assert(Elt->getType() == FTy->getParamType(ArgNo));
1924   else
1925     assert(FTy->isVarArg());
1926   ++ArgNo;
1927 }
1928 
1929 void CodeGenFunction::ExpandTypeToArgs(QualType Ty, RValue RV,
1930                                        SmallVector<llvm::Value*,16> &Args,
1931                                        llvm::FunctionType *IRFuncTy) {
1932   if (const ConstantArrayType *AT = getContext().getAsConstantArrayType(Ty)) {
1933     unsigned NumElts = AT->getSize().getZExtValue();
1934     QualType EltTy = AT->getElementType();
1935     llvm::Value *Addr = RV.getAggregateAddr();
1936     for (unsigned Elt = 0; Elt < NumElts; ++Elt) {
1937       llvm::Value *EltAddr = Builder.CreateConstGEP2_32(Addr, 0, Elt);
1938       LValue LV = MakeAddrLValue(EltAddr, EltTy);
1939       RValue EltRV;
1940       if (EltTy->isAnyComplexType())
1941         // FIXME: Volatile?
1942         EltRV = RValue::getComplex(LoadComplexFromAddr(LV.getAddress(), false));
1943       else if (CodeGenFunction::hasAggregateLLVMType(EltTy))
1944         EltRV = LV.asAggregateRValue();
1945       else
1946         EltRV = EmitLoadOfLValue(LV);
1947       ExpandTypeToArgs(EltTy, EltRV, Args, IRFuncTy);
1948     }
1949   } else if (const RecordType *RT = Ty->getAs<RecordType>()) {
1950     RecordDecl *RD = RT->getDecl();
1951     assert(RV.isAggregate() && "Unexpected rvalue during struct expansion");
1952     LValue LV = MakeAddrLValue(RV.getAggregateAddr(), Ty);
1953 
1954     if (RD->isUnion()) {
1955       const FieldDecl *LargestFD = 0;
1956       CharUnits UnionSize = CharUnits::Zero();
1957 
1958       for (RecordDecl::field_iterator i = RD->field_begin(), e = RD->field_end();
1959            i != e; ++i) {
1960         const FieldDecl *FD = *i;
1961         assert(!FD->isBitField() &&
1962                "Cannot expand structure with bit-field members.");
1963         CharUnits FieldSize = getContext().getTypeSizeInChars(FD->getType());
1964         if (UnionSize < FieldSize) {
1965           UnionSize = FieldSize;
1966           LargestFD = FD;
1967         }
1968       }
1969       if (LargestFD) {
1970         RValue FldRV = EmitRValueForField(LV, LargestFD);
1971         ExpandTypeToArgs(LargestFD->getType(), FldRV, Args, IRFuncTy);
1972       }
1973     } else {
1974       for (RecordDecl::field_iterator i = RD->field_begin(), e = RD->field_end();
1975            i != e; ++i) {
1976         FieldDecl *FD = *i;
1977 
1978         RValue FldRV = EmitRValueForField(LV, FD);
1979         ExpandTypeToArgs(FD->getType(), FldRV, Args, IRFuncTy);
1980       }
1981     }
1982   } else if (Ty->isAnyComplexType()) {
1983     ComplexPairTy CV = RV.getComplexVal();
1984     Args.push_back(CV.first);
1985     Args.push_back(CV.second);
1986   } else {
1987     assert(RV.isScalar() &&
1988            "Unexpected non-scalar rvalue during struct expansion.");
1989 
1990     // Insert a bitcast as needed.
1991     llvm::Value *V = RV.getScalarVal();
1992     if (Args.size() < IRFuncTy->getNumParams() &&
1993         V->getType() != IRFuncTy->getParamType(Args.size()))
1994       V = Builder.CreateBitCast(V, IRFuncTy->getParamType(Args.size()));
1995 
1996     Args.push_back(V);
1997   }
1998 }
1999 
2000 
2001 RValue CodeGenFunction::EmitCall(const CGFunctionInfo &CallInfo,
2002                                  llvm::Value *Callee,
2003                                  ReturnValueSlot ReturnValue,
2004                                  const CallArgList &CallArgs,
2005                                  const Decl *TargetDecl,
2006                                  llvm::Instruction **callOrInvoke) {
2007   // FIXME: We no longer need the types from CallArgs; lift up and simplify.
2008   SmallVector<llvm::Value*, 16> Args;
2009 
2010   // Handle struct-return functions by passing a pointer to the
2011   // location that we would like to return into.
2012   QualType RetTy = CallInfo.getReturnType();
2013   const ABIArgInfo &RetAI = CallInfo.getReturnInfo();
2014 
2015   // IRArgNo - Keep track of the argument number in the callee we're looking at.
2016   unsigned IRArgNo = 0;
2017   llvm::FunctionType *IRFuncTy =
2018     cast<llvm::FunctionType>(
2019                   cast<llvm::PointerType>(Callee->getType())->getElementType());
2020 
2021   // If the call returns a temporary with struct return, create a temporary
2022   // alloca to hold the result, unless one is given to us.
2023   if (CGM.ReturnTypeUsesSRet(CallInfo)) {
2024     llvm::Value *Value = ReturnValue.getValue();
2025     if (!Value)
2026       Value = CreateMemTemp(RetTy);
2027     Args.push_back(Value);
2028     checkArgMatches(Value, IRArgNo, IRFuncTy);
2029   }
2030 
2031   assert(CallInfo.arg_size() == CallArgs.size() &&
2032          "Mismatch between function signature & arguments.");
2033   CGFunctionInfo::const_arg_iterator info_it = CallInfo.arg_begin();
2034   for (CallArgList::const_iterator I = CallArgs.begin(), E = CallArgs.end();
2035        I != E; ++I, ++info_it) {
2036     const ABIArgInfo &ArgInfo = info_it->info;
2037     RValue RV = I->RV;
2038 
2039     unsigned TypeAlign =
2040       getContext().getTypeAlignInChars(I->Ty).getQuantity();
2041 
2042     // Insert a padding argument to ensure proper alignment.
2043     if (llvm::Type *PaddingType = ArgInfo.getPaddingType()) {
2044       Args.push_back(llvm::UndefValue::get(PaddingType));
2045       ++IRArgNo;
2046     }
2047 
2048     switch (ArgInfo.getKind()) {
2049     case ABIArgInfo::Indirect: {
2050       if (RV.isScalar() || RV.isComplex()) {
2051         // Make a temporary alloca to pass the argument.
2052         llvm::AllocaInst *AI = CreateMemTemp(I->Ty);
2053         if (ArgInfo.getIndirectAlign() > AI->getAlignment())
2054           AI->setAlignment(ArgInfo.getIndirectAlign());
2055         Args.push_back(AI);
2056 
2057         if (RV.isScalar())
2058           EmitStoreOfScalar(RV.getScalarVal(), Args.back(), false,
2059                             TypeAlign, I->Ty);
2060         else
2061           StoreComplexToAddr(RV.getComplexVal(), Args.back(), false);
2062 
2063         // Validate argument match.
2064         checkArgMatches(AI, IRArgNo, IRFuncTy);
2065       } else {
2066         // We want to avoid creating an unnecessary temporary+copy here;
2067         // however, we need one in two cases:
2068         // 1. If the argument is not byval, and we are required to copy the
2069         //    source.  (This case doesn't occur on any common architecture.)
2070         // 2. If the argument is byval, RV is not sufficiently aligned, and
2071         //    we cannot force it to be sufficiently aligned.
2072         llvm::Value *Addr = RV.getAggregateAddr();
2073         unsigned Align = ArgInfo.getIndirectAlign();
2074         const llvm::DataLayout *TD = &CGM.getDataLayout();
2075         if ((!ArgInfo.getIndirectByVal() && I->NeedsCopy) ||
2076             (ArgInfo.getIndirectByVal() && TypeAlign < Align &&
2077              llvm::getOrEnforceKnownAlignment(Addr, Align, TD) < Align)) {
2078           // Create an aligned temporary, and copy to it.
2079           llvm::AllocaInst *AI = CreateMemTemp(I->Ty);
2080           if (Align > AI->getAlignment())
2081             AI->setAlignment(Align);
2082           Args.push_back(AI);
2083           EmitAggregateCopy(AI, Addr, I->Ty, RV.isVolatileQualified());
2084 
2085           // Validate argument match.
2086           checkArgMatches(AI, IRArgNo, IRFuncTy);
2087         } else {
2088           // Skip the extra memcpy call.
2089           Args.push_back(Addr);
2090 
2091           // Validate argument match.
2092           checkArgMatches(Addr, IRArgNo, IRFuncTy);
2093         }
2094       }
2095       break;
2096     }
2097 
2098     case ABIArgInfo::Ignore:
2099       break;
2100 
2101     case ABIArgInfo::Extend:
2102     case ABIArgInfo::Direct: {
2103       if (!isa<llvm::StructType>(ArgInfo.getCoerceToType()) &&
2104           ArgInfo.getCoerceToType() == ConvertType(info_it->type) &&
2105           ArgInfo.getDirectOffset() == 0) {
2106         llvm::Value *V;
2107         if (RV.isScalar())
2108           V = RV.getScalarVal();
2109         else
2110           V = Builder.CreateLoad(RV.getAggregateAddr());
2111 
2112         // If the argument doesn't match, perform a bitcast to coerce it.  This
2113         // can happen due to trivial type mismatches.
2114         if (IRArgNo < IRFuncTy->getNumParams() &&
2115             V->getType() != IRFuncTy->getParamType(IRArgNo))
2116           V = Builder.CreateBitCast(V, IRFuncTy->getParamType(IRArgNo));
2117         Args.push_back(V);
2118 
2119         checkArgMatches(V, IRArgNo, IRFuncTy);
2120         break;
2121       }
2122 
2123       // FIXME: Avoid the conversion through memory if possible.
2124       llvm::Value *SrcPtr;
2125       if (RV.isScalar()) {
2126         SrcPtr = CreateMemTemp(I->Ty, "coerce");
2127         EmitStoreOfScalar(RV.getScalarVal(), SrcPtr, false, TypeAlign, I->Ty);
2128       } else if (RV.isComplex()) {
2129         SrcPtr = CreateMemTemp(I->Ty, "coerce");
2130         StoreComplexToAddr(RV.getComplexVal(), SrcPtr, false);
2131       } else
2132         SrcPtr = RV.getAggregateAddr();
2133 
2134       // If the value is offset in memory, apply the offset now.
2135       if (unsigned Offs = ArgInfo.getDirectOffset()) {
2136         SrcPtr = Builder.CreateBitCast(SrcPtr, Builder.getInt8PtrTy());
2137         SrcPtr = Builder.CreateConstGEP1_32(SrcPtr, Offs);
2138         SrcPtr = Builder.CreateBitCast(SrcPtr,
2139                        llvm::PointerType::getUnqual(ArgInfo.getCoerceToType()));
2140 
2141       }
2142 
2143       // If the coerce-to type is a first class aggregate, we flatten it and
2144       // pass the elements. Either way is semantically identical, but fast-isel
2145       // and the optimizer generally likes scalar values better than FCAs.
2146       if (llvm::StructType *STy =
2147             dyn_cast<llvm::StructType>(ArgInfo.getCoerceToType())) {
2148         llvm::Type *SrcTy =
2149           cast<llvm::PointerType>(SrcPtr->getType())->getElementType();
2150         uint64_t SrcSize = CGM.getDataLayout().getTypeAllocSize(SrcTy);
2151         uint64_t DstSize = CGM.getDataLayout().getTypeAllocSize(STy);
2152 
2153         // If the source type is smaller than the destination type of the
2154         // coerce-to logic, copy the source value into a temp alloca the size
2155         // of the destination type to allow loading all of it. The bits past
2156         // the source value are left undef.
2157         if (SrcSize < DstSize) {
2158           llvm::AllocaInst *TempAlloca
2159             = CreateTempAlloca(STy, SrcPtr->getName() + ".coerce");
2160           Builder.CreateMemCpy(TempAlloca, SrcPtr, SrcSize, 0);
2161           SrcPtr = TempAlloca;
2162         } else {
2163           SrcPtr = Builder.CreateBitCast(SrcPtr,
2164                                          llvm::PointerType::getUnqual(STy));
2165         }
2166 
2167         for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
2168           llvm::Value *EltPtr = Builder.CreateConstGEP2_32(SrcPtr, 0, i);
2169           llvm::LoadInst *LI = Builder.CreateLoad(EltPtr);
2170           // We don't know what we're loading from.
2171           LI->setAlignment(1);
2172           Args.push_back(LI);
2173 
2174           // Validate argument match.
2175           checkArgMatches(LI, IRArgNo, IRFuncTy);
2176         }
2177       } else {
2178         // In the simple case, just pass the coerced loaded value.
2179         Args.push_back(CreateCoercedLoad(SrcPtr, ArgInfo.getCoerceToType(),
2180                                          *this));
2181 
2182         // Validate argument match.
2183         checkArgMatches(Args.back(), IRArgNo, IRFuncTy);
2184       }
2185 
2186       break;
2187     }
2188 
2189     case ABIArgInfo::Expand:
2190       ExpandTypeToArgs(I->Ty, RV, Args, IRFuncTy);
2191       IRArgNo = Args.size();
2192       break;
2193     }
2194   }
2195 
2196   // If the callee is a bitcast of a function to a varargs pointer to function
2197   // type, check to see if we can remove the bitcast.  This handles some cases
2198   // with unprototyped functions.
2199   if (llvm::ConstantExpr *CE = dyn_cast<llvm::ConstantExpr>(Callee))
2200     if (llvm::Function *CalleeF = dyn_cast<llvm::Function>(CE->getOperand(0))) {
2201       llvm::PointerType *CurPT=cast<llvm::PointerType>(Callee->getType());
2202       llvm::FunctionType *CurFT =
2203         cast<llvm::FunctionType>(CurPT->getElementType());
2204       llvm::FunctionType *ActualFT = CalleeF->getFunctionType();
2205 
2206       if (CE->getOpcode() == llvm::Instruction::BitCast &&
2207           ActualFT->getReturnType() == CurFT->getReturnType() &&
2208           ActualFT->getNumParams() == CurFT->getNumParams() &&
2209           ActualFT->getNumParams() == Args.size() &&
2210           (CurFT->isVarArg() || !ActualFT->isVarArg())) {
2211         bool ArgsMatch = true;
2212         for (unsigned i = 0, e = ActualFT->getNumParams(); i != e; ++i)
2213           if (ActualFT->getParamType(i) != CurFT->getParamType(i)) {
2214             ArgsMatch = false;
2215             break;
2216           }
2217 
2218         // Strip the cast if we can get away with it.  This is a nice cleanup,
2219         // but also allows us to inline the function at -O0 if it is marked
2220         // always_inline.
2221         if (ArgsMatch)
2222           Callee = CalleeF;
2223       }
2224     }
2225 
2226   unsigned CallingConv;
2227   CodeGen::AttributeListType AttributeList;
2228   CGM.ConstructAttributeList(CallInfo, TargetDecl, AttributeList, CallingConv);
2229   llvm::AttributeSet Attrs = llvm::AttributeSet::get(getLLVMContext(),
2230                                                    AttributeList);
2231 
2232   llvm::BasicBlock *InvokeDest = 0;
2233   if (!Attrs.hasAttribute(llvm::AttributeSet::FunctionIndex,
2234                           llvm::Attribute::NoUnwind))
2235     InvokeDest = getInvokeDest();
2236 
2237   llvm::CallSite CS;
2238   if (!InvokeDest) {
2239     CS = Builder.CreateCall(Callee, Args);
2240   } else {
2241     llvm::BasicBlock *Cont = createBasicBlock("invoke.cont");
2242     CS = Builder.CreateInvoke(Callee, Cont, InvokeDest, Args);
2243     EmitBlock(Cont);
2244   }
2245   if (callOrInvoke)
2246     *callOrInvoke = CS.getInstruction();
2247 
2248   CS.setAttributes(Attrs);
2249   CS.setCallingConv(static_cast<llvm::CallingConv::ID>(CallingConv));
2250 
2251   // In ObjC ARC mode with no ObjC ARC exception safety, tell the ARC
2252   // optimizer it can aggressively ignore unwind edges.
2253   if (CGM.getLangOpts().ObjCAutoRefCount)
2254     AddObjCARCExceptionMetadata(CS.getInstruction());
2255 
2256   // If the call doesn't return, finish the basic block and clear the
2257   // insertion point; this allows the rest of IRgen to discard
2258   // unreachable code.
2259   if (CS.doesNotReturn()) {
2260     Builder.CreateUnreachable();
2261     Builder.ClearInsertionPoint();
2262 
2263     // FIXME: For now, emit a dummy basic block because expr emitters in
2264     // generally are not ready to handle emitting expressions at unreachable
2265     // points.
2266     EnsureInsertPoint();
2267 
2268     // Return a reasonable RValue.
2269     return GetUndefRValue(RetTy);
2270   }
2271 
2272   llvm::Instruction *CI = CS.getInstruction();
2273   if (Builder.isNamePreserving() && !CI->getType()->isVoidTy())
2274     CI->setName("call");
2275 
2276   // Emit any writebacks immediately.  Arguably this should happen
2277   // after any return-value munging.
2278   if (CallArgs.hasWritebacks())
2279     emitWritebacks(*this, CallArgs);
2280 
2281   switch (RetAI.getKind()) {
2282   case ABIArgInfo::Indirect: {
2283     unsigned Alignment = getContext().getTypeAlignInChars(RetTy).getQuantity();
2284     if (RetTy->isAnyComplexType())
2285       return RValue::getComplex(LoadComplexFromAddr(Args[0], false));
2286     if (CodeGenFunction::hasAggregateLLVMType(RetTy))
2287       return RValue::getAggregate(Args[0]);
2288     return RValue::get(EmitLoadOfScalar(Args[0], false, Alignment, RetTy));
2289   }
2290 
2291   case ABIArgInfo::Ignore:
2292     // If we are ignoring an argument that had a result, make sure to
2293     // construct the appropriate return value for our caller.
2294     return GetUndefRValue(RetTy);
2295 
2296   case ABIArgInfo::Extend:
2297   case ABIArgInfo::Direct: {
2298     llvm::Type *RetIRTy = ConvertType(RetTy);
2299     if (RetAI.getCoerceToType() == RetIRTy && RetAI.getDirectOffset() == 0) {
2300       if (RetTy->isAnyComplexType()) {
2301         llvm::Value *Real = Builder.CreateExtractValue(CI, 0);
2302         llvm::Value *Imag = Builder.CreateExtractValue(CI, 1);
2303         return RValue::getComplex(std::make_pair(Real, Imag));
2304       }
2305       if (CodeGenFunction::hasAggregateLLVMType(RetTy)) {
2306         llvm::Value *DestPtr = ReturnValue.getValue();
2307         bool DestIsVolatile = ReturnValue.isVolatile();
2308 
2309         if (!DestPtr) {
2310           DestPtr = CreateMemTemp(RetTy, "agg.tmp");
2311           DestIsVolatile = false;
2312         }
2313         BuildAggStore(*this, CI, DestPtr, DestIsVolatile, false);
2314         return RValue::getAggregate(DestPtr);
2315       }
2316 
2317       // If the argument doesn't match, perform a bitcast to coerce it.  This
2318       // can happen due to trivial type mismatches.
2319       llvm::Value *V = CI;
2320       if (V->getType() != RetIRTy)
2321         V = Builder.CreateBitCast(V, RetIRTy);
2322       return RValue::get(V);
2323     }
2324 
2325     llvm::Value *DestPtr = ReturnValue.getValue();
2326     bool DestIsVolatile = ReturnValue.isVolatile();
2327 
2328     if (!DestPtr) {
2329       DestPtr = CreateMemTemp(RetTy, "coerce");
2330       DestIsVolatile = false;
2331     }
2332 
2333     // If the value is offset in memory, apply the offset now.
2334     llvm::Value *StorePtr = DestPtr;
2335     if (unsigned Offs = RetAI.getDirectOffset()) {
2336       StorePtr = Builder.CreateBitCast(StorePtr, Builder.getInt8PtrTy());
2337       StorePtr = Builder.CreateConstGEP1_32(StorePtr, Offs);
2338       StorePtr = Builder.CreateBitCast(StorePtr,
2339                          llvm::PointerType::getUnqual(RetAI.getCoerceToType()));
2340     }
2341     CreateCoercedStore(CI, StorePtr, DestIsVolatile, *this);
2342 
2343     unsigned Alignment = getContext().getTypeAlignInChars(RetTy).getQuantity();
2344     if (RetTy->isAnyComplexType())
2345       return RValue::getComplex(LoadComplexFromAddr(DestPtr, false));
2346     if (CodeGenFunction::hasAggregateLLVMType(RetTy))
2347       return RValue::getAggregate(DestPtr);
2348     return RValue::get(EmitLoadOfScalar(DestPtr, false, Alignment, RetTy));
2349   }
2350 
2351   case ABIArgInfo::Expand:
2352     llvm_unreachable("Invalid ABI kind for return argument");
2353   }
2354 
2355   llvm_unreachable("Unhandled ABIArgInfo::Kind");
2356 }
2357 
2358 /* VarArg handling */
2359 
2360 llvm::Value *CodeGenFunction::EmitVAArg(llvm::Value *VAListAddr, QualType Ty) {
2361   return CGM.getTypes().getABIInfo().EmitVAArg(VAListAddr, Ty, *this);
2362 }
2363