1 //===--- CGCall.cpp - Encapsulate calling convention details --------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // These classes wrap the information about a call or function
11 // definition used to handle ABI compliancy.
13 //===----------------------------------------------------------------------===//
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/CodeGen/CGFunctionInfo.h"
26 #include "clang/Frontend/CodeGenOptions.h"
27 #include "llvm/ADT/StringExtras.h"
28 #include "llvm/IR/Attributes.h"
29 #include "llvm/IR/CallSite.h"
30 #include "llvm/IR/DataLayout.h"
31 #include "llvm/IR/InlineAsm.h"
32 #include "llvm/IR/Intrinsics.h"
33 #include "llvm/IR/IntrinsicInst.h"
34 #include "llvm/Transforms/Utils/Local.h"
36 using namespace clang;
37 using namespace CodeGen;
41 static unsigned ClangCallConvToLLVMCallConv(CallingConv CC) {
43 default: return llvm::CallingConv::C;
44 case CC_X86StdCall: return llvm::CallingConv::X86_StdCall;
45 case CC_X86FastCall: return llvm::CallingConv::X86_FastCall;
46 case CC_X86ThisCall: return llvm::CallingConv::X86_ThisCall;
47 case CC_X86_64Win64: return llvm::CallingConv::X86_64_Win64;
48 case CC_X86_64SysV: return llvm::CallingConv::X86_64_SysV;
49 case CC_AAPCS: return llvm::CallingConv::ARM_AAPCS;
50 case CC_AAPCS_VFP: return llvm::CallingConv::ARM_AAPCS_VFP;
51 case CC_IntelOclBicc: return llvm::CallingConv::Intel_OCL_BI;
52 // TODO: Add support for __pascal to LLVM.
53 case CC_X86Pascal: return llvm::CallingConv::C;
54 // TODO: Add support for __vectorcall to LLVM.
55 case CC_X86VectorCall: return llvm::CallingConv::X86_VectorCall;
56 case CC_SpirFunction: return llvm::CallingConv::SPIR_FUNC;
57 case CC_SpirKernel: return llvm::CallingConv::SPIR_KERNEL;
61 /// Derives the 'this' type for codegen purposes, i.e. ignoring method
63 /// FIXME: address space qualification?
64 static CanQualType GetThisType(ASTContext &Context, const CXXRecordDecl *RD) {
65 QualType RecTy = Context.getTagDeclType(RD)->getCanonicalTypeInternal();
66 return Context.getPointerType(CanQualType::CreateUnsafe(RecTy));
69 /// Returns the canonical formal type of the given C++ method.
70 static CanQual<FunctionProtoType> GetFormalType(const CXXMethodDecl *MD) {
71 return MD->getType()->getCanonicalTypeUnqualified()
72 .getAs<FunctionProtoType>();
75 /// Returns the "extra-canonicalized" return type, which discards
76 /// qualifiers on the return type. Codegen doesn't care about them,
77 /// and it makes ABI code a little easier to be able to assume that
78 /// all parameter and return types are top-level unqualified.
79 static CanQualType GetReturnType(QualType RetTy) {
80 return RetTy->getCanonicalTypeUnqualified().getUnqualifiedType();
83 /// Arrange the argument and result information for a value of the given
84 /// unprototyped freestanding function type.
85 const CGFunctionInfo &
86 CodeGenTypes::arrangeFreeFunctionType(CanQual<FunctionNoProtoType> FTNP) {
87 // When translating an unprototyped function type, always use a
89 return arrangeLLVMFunctionInfo(FTNP->getReturnType().getUnqualifiedType(),
90 /*instanceMethod=*/false,
91 /*chainCall=*/false, None,
92 FTNP->getExtInfo(), RequiredArgs(0));
95 /// Arrange the LLVM function layout for a value of the given function
96 /// type, on top of any implicit parameters already stored.
97 static const CGFunctionInfo &
98 arrangeLLVMFunctionInfo(CodeGenTypes &CGT, bool instanceMethod,
99 SmallVectorImpl<CanQualType> &prefix,
100 CanQual<FunctionProtoType> FTP) {
101 RequiredArgs required = RequiredArgs::forPrototypePlus(FTP, prefix.size());
103 prefix.append(FTP->param_type_begin(), FTP->param_type_end());
104 CanQualType resultType = FTP->getReturnType().getUnqualifiedType();
105 return CGT.arrangeLLVMFunctionInfo(resultType, instanceMethod,
106 /*chainCall=*/false, prefix,
107 FTP->getExtInfo(), required);
110 /// Arrange the argument and result information for a value of the
111 /// given freestanding function type.
112 const CGFunctionInfo &
113 CodeGenTypes::arrangeFreeFunctionType(CanQual<FunctionProtoType> FTP) {
114 SmallVector<CanQualType, 16> argTypes;
115 return ::arrangeLLVMFunctionInfo(*this, /*instanceMethod=*/false, argTypes,
119 static CallingConv getCallingConventionForDecl(const Decl *D, bool IsWindows) {
120 // Set the appropriate calling convention for the Function.
121 if (D->hasAttr<StdCallAttr>())
122 return CC_X86StdCall;
124 if (D->hasAttr<FastCallAttr>())
125 return CC_X86FastCall;
127 if (D->hasAttr<ThisCallAttr>())
128 return CC_X86ThisCall;
130 if (D->hasAttr<VectorCallAttr>())
131 return CC_X86VectorCall;
133 if (D->hasAttr<PascalAttr>())
136 if (PcsAttr *PCS = D->getAttr<PcsAttr>())
137 return (PCS->getPCS() == PcsAttr::AAPCS ? CC_AAPCS : CC_AAPCS_VFP);
139 if (D->hasAttr<IntelOclBiccAttr>())
140 return CC_IntelOclBicc;
142 if (D->hasAttr<MSABIAttr>())
143 return IsWindows ? CC_C : CC_X86_64Win64;
145 if (D->hasAttr<SysVABIAttr>())
146 return IsWindows ? CC_X86_64SysV : CC_C;
151 /// Arrange the argument and result information for a call to an
152 /// unknown C++ non-static member function of the given abstract type.
153 /// (Zero value of RD means we don't have any meaningful "this" argument type,
154 /// so fall back to a generic pointer type).
155 /// The member function must be an ordinary function, i.e. not a
156 /// constructor or destructor.
157 const CGFunctionInfo &
158 CodeGenTypes::arrangeCXXMethodType(const CXXRecordDecl *RD,
159 const FunctionProtoType *FTP) {
160 SmallVector<CanQualType, 16> argTypes;
162 // Add the 'this' pointer.
164 argTypes.push_back(GetThisType(Context, RD));
166 argTypes.push_back(Context.VoidPtrTy);
168 return ::arrangeLLVMFunctionInfo(
169 *this, true, argTypes,
170 FTP->getCanonicalTypeUnqualified().getAs<FunctionProtoType>());
173 /// Arrange the argument and result information for a declaration or
174 /// definition of the given C++ non-static member function. The
175 /// member function must be an ordinary function, i.e. not a
176 /// constructor or destructor.
177 const CGFunctionInfo &
178 CodeGenTypes::arrangeCXXMethodDeclaration(const CXXMethodDecl *MD) {
179 assert(!isa<CXXConstructorDecl>(MD) && "wrong method for constructors!");
180 assert(!isa<CXXDestructorDecl>(MD) && "wrong method for destructors!");
182 CanQual<FunctionProtoType> prototype = GetFormalType(MD);
184 if (MD->isInstance()) {
185 // The abstract case is perfectly fine.
186 const CXXRecordDecl *ThisType = TheCXXABI.getThisArgumentTypeForMethod(MD);
187 return arrangeCXXMethodType(ThisType, prototype.getTypePtr());
190 return arrangeFreeFunctionType(prototype);
193 const CGFunctionInfo &
194 CodeGenTypes::arrangeCXXStructorDeclaration(const CXXMethodDecl *MD,
197 SmallVector<CanQualType, 16> argTypes;
198 argTypes.push_back(GetThisType(Context, MD->getParent()));
201 if (auto *CD = dyn_cast<CXXConstructorDecl>(MD)) {
202 GD = GlobalDecl(CD, toCXXCtorType(Type));
204 auto *DD = dyn_cast<CXXDestructorDecl>(MD);
205 GD = GlobalDecl(DD, toCXXDtorType(Type));
208 CanQual<FunctionProtoType> FTP = GetFormalType(MD);
210 // Add the formal parameters.
211 argTypes.append(FTP->param_type_begin(), FTP->param_type_end());
213 TheCXXABI.buildStructorSignature(MD, Type, argTypes);
215 RequiredArgs required =
216 (MD->isVariadic() ? RequiredArgs(argTypes.size()) : RequiredArgs::All);
218 FunctionType::ExtInfo extInfo = FTP->getExtInfo();
219 CanQualType resultType = TheCXXABI.HasThisReturn(GD)
221 : TheCXXABI.hasMostDerivedReturn(GD)
222 ? CGM.getContext().VoidPtrTy
224 return arrangeLLVMFunctionInfo(resultType, /*instanceMethod=*/true,
225 /*chainCall=*/false, argTypes, extInfo,
229 /// Arrange a call to a C++ method, passing the given arguments.
230 const CGFunctionInfo &
231 CodeGenTypes::arrangeCXXConstructorCall(const CallArgList &args,
232 const CXXConstructorDecl *D,
233 CXXCtorType CtorKind,
234 unsigned ExtraArgs) {
236 SmallVector<CanQualType, 16> ArgTypes;
237 for (const auto &Arg : args)
238 ArgTypes.push_back(Context.getCanonicalParamType(Arg.Ty));
240 CanQual<FunctionProtoType> FPT = GetFormalType(D);
241 RequiredArgs Required = RequiredArgs::forPrototypePlus(FPT, 1 + ExtraArgs);
242 GlobalDecl GD(D, CtorKind);
243 CanQualType ResultType = TheCXXABI.HasThisReturn(GD)
245 : TheCXXABI.hasMostDerivedReturn(GD)
246 ? CGM.getContext().VoidPtrTy
249 FunctionType::ExtInfo Info = FPT->getExtInfo();
250 return arrangeLLVMFunctionInfo(ResultType, /*instanceMethod=*/true,
251 /*chainCall=*/false, ArgTypes, Info,
255 /// Arrange the argument and result information for the declaration or
256 /// definition of the given function.
257 const CGFunctionInfo &
258 CodeGenTypes::arrangeFunctionDeclaration(const FunctionDecl *FD) {
259 if (const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(FD))
260 if (MD->isInstance())
261 return arrangeCXXMethodDeclaration(MD);
263 CanQualType FTy = FD->getType()->getCanonicalTypeUnqualified();
265 assert(isa<FunctionType>(FTy));
267 // When declaring a function without a prototype, always use a
268 // non-variadic type.
269 if (isa<FunctionNoProtoType>(FTy)) {
270 CanQual<FunctionNoProtoType> noProto = FTy.getAs<FunctionNoProtoType>();
271 return arrangeLLVMFunctionInfo(
272 noProto->getReturnType(), /*instanceMethod=*/false,
273 /*chainCall=*/false, None, noProto->getExtInfo(), RequiredArgs::All);
276 assert(isa<FunctionProtoType>(FTy));
277 return arrangeFreeFunctionType(FTy.getAs<FunctionProtoType>());
280 /// Arrange the argument and result information for the declaration or
281 /// definition of an Objective-C method.
282 const CGFunctionInfo &
283 CodeGenTypes::arrangeObjCMethodDeclaration(const ObjCMethodDecl *MD) {
284 // It happens that this is the same as a call with no optional
285 // arguments, except also using the formal 'self' type.
286 return arrangeObjCMessageSendSignature(MD, MD->getSelfDecl()->getType());
289 /// Arrange the argument and result information for the function type
290 /// through which to perform a send to the given Objective-C method,
291 /// using the given receiver type. The receiver type is not always
292 /// the 'self' type of the method or even an Objective-C pointer type.
293 /// This is *not* the right method for actually performing such a
294 /// message send, due to the possibility of optional arguments.
295 const CGFunctionInfo &
296 CodeGenTypes::arrangeObjCMessageSendSignature(const ObjCMethodDecl *MD,
297 QualType receiverType) {
298 SmallVector<CanQualType, 16> argTys;
299 argTys.push_back(Context.getCanonicalParamType(receiverType));
300 argTys.push_back(Context.getCanonicalParamType(Context.getObjCSelType()));
302 for (const auto *I : MD->params()) {
303 argTys.push_back(Context.getCanonicalParamType(I->getType()));
306 FunctionType::ExtInfo einfo;
307 bool IsWindows = getContext().getTargetInfo().getTriple().isOSWindows();
308 einfo = einfo.withCallingConv(getCallingConventionForDecl(MD, IsWindows));
310 if (getContext().getLangOpts().ObjCAutoRefCount &&
311 MD->hasAttr<NSReturnsRetainedAttr>())
312 einfo = einfo.withProducesResult(true);
314 RequiredArgs required =
315 (MD->isVariadic() ? RequiredArgs(argTys.size()) : RequiredArgs::All);
317 return arrangeLLVMFunctionInfo(
318 GetReturnType(MD->getReturnType()), /*instanceMethod=*/false,
319 /*chainCall=*/false, argTys, einfo, required);
322 const CGFunctionInfo &
323 CodeGenTypes::arrangeGlobalDeclaration(GlobalDecl GD) {
324 // FIXME: Do we need to handle ObjCMethodDecl?
325 const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl());
327 if (const CXXConstructorDecl *CD = dyn_cast<CXXConstructorDecl>(FD))
328 return arrangeCXXStructorDeclaration(CD, getFromCtorType(GD.getCtorType()));
330 if (const CXXDestructorDecl *DD = dyn_cast<CXXDestructorDecl>(FD))
331 return arrangeCXXStructorDeclaration(DD, getFromDtorType(GD.getDtorType()));
333 return arrangeFunctionDeclaration(FD);
336 /// Arrange a thunk that takes 'this' as the first parameter followed by
337 /// varargs. Return a void pointer, regardless of the actual return type.
338 /// The body of the thunk will end in a musttail call to a function of the
339 /// correct type, and the caller will bitcast the function to the correct
341 const CGFunctionInfo &
342 CodeGenTypes::arrangeMSMemberPointerThunk(const CXXMethodDecl *MD) {
343 assert(MD->isVirtual() && "only virtual memptrs have thunks");
344 CanQual<FunctionProtoType> FTP = GetFormalType(MD);
345 CanQualType ArgTys[] = { GetThisType(Context, MD->getParent()) };
346 return arrangeLLVMFunctionInfo(Context.VoidTy, /*instanceMethod=*/false,
347 /*chainCall=*/false, ArgTys,
348 FTP->getExtInfo(), RequiredArgs(1));
351 const CGFunctionInfo &
352 CodeGenTypes::arrangeMSCtorClosure(const CXXConstructorDecl *CD,
354 assert(CT == Ctor_CopyingClosure || CT == Ctor_DefaultClosure);
356 CanQual<FunctionProtoType> FTP = GetFormalType(CD);
357 SmallVector<CanQualType, 2> ArgTys;
358 const CXXRecordDecl *RD = CD->getParent();
359 ArgTys.push_back(GetThisType(Context, RD));
360 if (CT == Ctor_CopyingClosure)
361 ArgTys.push_back(*FTP->param_type_begin());
362 if (RD->getNumVBases() > 0)
363 ArgTys.push_back(Context.IntTy);
364 CallingConv CC = Context.getDefaultCallingConvention(
365 /*IsVariadic=*/false, /*IsCXXMethod=*/true);
366 return arrangeLLVMFunctionInfo(Context.VoidTy, /*instanceMethod=*/true,
367 /*chainCall=*/false, ArgTys,
368 FunctionType::ExtInfo(CC), RequiredArgs::All);
371 /// Arrange a call as unto a free function, except possibly with an
372 /// additional number of formal parameters considered required.
373 static const CGFunctionInfo &
374 arrangeFreeFunctionLikeCall(CodeGenTypes &CGT,
376 const CallArgList &args,
377 const FunctionType *fnType,
378 unsigned numExtraRequiredArgs,
380 assert(args.size() >= numExtraRequiredArgs);
382 // In most cases, there are no optional arguments.
383 RequiredArgs required = RequiredArgs::All;
385 // If we have a variadic prototype, the required arguments are the
386 // extra prefix plus the arguments in the prototype.
387 if (const FunctionProtoType *proto = dyn_cast<FunctionProtoType>(fnType)) {
388 if (proto->isVariadic())
389 required = RequiredArgs(proto->getNumParams() + numExtraRequiredArgs);
391 // If we don't have a prototype at all, but we're supposed to
392 // explicitly use the variadic convention for unprototyped calls,
393 // treat all of the arguments as required but preserve the nominal
394 // possibility of variadics.
395 } else if (CGM.getTargetCodeGenInfo()
396 .isNoProtoCallVariadic(args,
397 cast<FunctionNoProtoType>(fnType))) {
398 required = RequiredArgs(args.size());
402 SmallVector<CanQualType, 16> argTypes;
403 for (const auto &arg : args)
404 argTypes.push_back(CGT.getContext().getCanonicalParamType(arg.Ty));
405 return CGT.arrangeLLVMFunctionInfo(GetReturnType(fnType->getReturnType()),
406 /*instanceMethod=*/false, chainCall,
407 argTypes, fnType->getExtInfo(), required);
410 /// Figure out the rules for calling a function with the given formal
411 /// type using the given arguments. The arguments are necessary
412 /// because the function might be unprototyped, in which case it's
413 /// target-dependent in crazy ways.
414 const CGFunctionInfo &
415 CodeGenTypes::arrangeFreeFunctionCall(const CallArgList &args,
416 const FunctionType *fnType,
418 return arrangeFreeFunctionLikeCall(*this, CGM, args, fnType,
419 chainCall ? 1 : 0, chainCall);
422 /// A block function call is essentially a free-function call with an
423 /// extra implicit argument.
424 const CGFunctionInfo &
425 CodeGenTypes::arrangeBlockFunctionCall(const CallArgList &args,
426 const FunctionType *fnType) {
427 return arrangeFreeFunctionLikeCall(*this, CGM, args, fnType, 1,
428 /*chainCall=*/false);
431 const CGFunctionInfo &
432 CodeGenTypes::arrangeFreeFunctionCall(QualType resultType,
433 const CallArgList &args,
434 FunctionType::ExtInfo info,
435 RequiredArgs required) {
437 SmallVector<CanQualType, 16> argTypes;
438 for (const auto &Arg : args)
439 argTypes.push_back(Context.getCanonicalParamType(Arg.Ty));
440 return arrangeLLVMFunctionInfo(
441 GetReturnType(resultType), /*instanceMethod=*/false,
442 /*chainCall=*/false, argTypes, info, required);
445 /// Arrange a call to a C++ method, passing the given arguments.
446 const CGFunctionInfo &
447 CodeGenTypes::arrangeCXXMethodCall(const CallArgList &args,
448 const FunctionProtoType *FPT,
449 RequiredArgs required) {
451 SmallVector<CanQualType, 16> argTypes;
452 for (const auto &Arg : args)
453 argTypes.push_back(Context.getCanonicalParamType(Arg.Ty));
455 FunctionType::ExtInfo info = FPT->getExtInfo();
456 return arrangeLLVMFunctionInfo(
457 GetReturnType(FPT->getReturnType()), /*instanceMethod=*/true,
458 /*chainCall=*/false, argTypes, info, required);
461 const CGFunctionInfo &CodeGenTypes::arrangeFreeFunctionDeclaration(
462 QualType resultType, const FunctionArgList &args,
463 const FunctionType::ExtInfo &info, bool isVariadic) {
465 SmallVector<CanQualType, 16> argTypes;
466 for (auto Arg : args)
467 argTypes.push_back(Context.getCanonicalParamType(Arg->getType()));
469 RequiredArgs required =
470 (isVariadic ? RequiredArgs(args.size()) : RequiredArgs::All);
471 return arrangeLLVMFunctionInfo(
472 GetReturnType(resultType), /*instanceMethod=*/false,
473 /*chainCall=*/false, argTypes, info, required);
476 const CGFunctionInfo &CodeGenTypes::arrangeNullaryFunction() {
477 return arrangeLLVMFunctionInfo(
478 getContext().VoidTy, /*instanceMethod=*/false, /*chainCall=*/false,
479 None, FunctionType::ExtInfo(), RequiredArgs::All);
482 /// Arrange the argument and result information for an abstract value
483 /// of a given function type. This is the method which all of the
484 /// above functions ultimately defer to.
485 const CGFunctionInfo &
486 CodeGenTypes::arrangeLLVMFunctionInfo(CanQualType resultType,
489 ArrayRef<CanQualType> argTypes,
490 FunctionType::ExtInfo info,
491 RequiredArgs required) {
492 assert(std::all_of(argTypes.begin(), argTypes.end(),
493 std::mem_fun_ref(&CanQualType::isCanonicalAsParam)));
495 unsigned CC = ClangCallConvToLLVMCallConv(info.getCC());
497 // Lookup or create unique function info.
498 llvm::FoldingSetNodeID ID;
499 CGFunctionInfo::Profile(ID, instanceMethod, chainCall, info, required,
500 resultType, argTypes);
502 void *insertPos = nullptr;
503 CGFunctionInfo *FI = FunctionInfos.FindNodeOrInsertPos(ID, insertPos);
507 // Construct the function info. We co-allocate the ArgInfos.
508 FI = CGFunctionInfo::create(CC, instanceMethod, chainCall, info,
509 resultType, argTypes, required);
510 FunctionInfos.InsertNode(FI, insertPos);
512 bool inserted = FunctionsBeingProcessed.insert(FI).second;
514 assert(inserted && "Recursively being processed?");
516 // Compute ABI information.
517 getABIInfo().computeInfo(*FI);
519 // Loop over all of the computed argument and return value info. If any of
520 // them are direct or extend without a specified coerce type, specify the
522 ABIArgInfo &retInfo = FI->getReturnInfo();
523 if (retInfo.canHaveCoerceToType() && retInfo.getCoerceToType() == nullptr)
524 retInfo.setCoerceToType(ConvertType(FI->getReturnType()));
526 for (auto &I : FI->arguments())
527 if (I.info.canHaveCoerceToType() && I.info.getCoerceToType() == nullptr)
528 I.info.setCoerceToType(ConvertType(I.type));
530 bool erased = FunctionsBeingProcessed.erase(FI); (void)erased;
531 assert(erased && "Not in set?");
536 CGFunctionInfo *CGFunctionInfo::create(unsigned llvmCC,
539 const FunctionType::ExtInfo &info,
540 CanQualType resultType,
541 ArrayRef<CanQualType> argTypes,
542 RequiredArgs required) {
543 void *buffer = operator new(sizeof(CGFunctionInfo) +
544 sizeof(ArgInfo) * (argTypes.size() + 1));
545 CGFunctionInfo *FI = new(buffer) CGFunctionInfo();
546 FI->CallingConvention = llvmCC;
547 FI->EffectiveCallingConvention = llvmCC;
548 FI->ASTCallingConvention = info.getCC();
549 FI->InstanceMethod = instanceMethod;
550 FI->ChainCall = chainCall;
551 FI->NoReturn = info.getNoReturn();
552 FI->ReturnsRetained = info.getProducesResult();
553 FI->Required = required;
554 FI->HasRegParm = info.getHasRegParm();
555 FI->RegParm = info.getRegParm();
556 FI->ArgStruct = nullptr;
557 FI->NumArgs = argTypes.size();
558 FI->getArgsBuffer()[0].type = resultType;
559 for (unsigned i = 0, e = argTypes.size(); i != e; ++i)
560 FI->getArgsBuffer()[i + 1].type = argTypes[i];
567 // ABIArgInfo::Expand implementation.
569 // Specifies the way QualType passed as ABIArgInfo::Expand is expanded.
570 struct TypeExpansion {
571 enum TypeExpansionKind {
572 // Elements of constant arrays are expanded recursively.
574 // Record fields are expanded recursively (but if record is a union, only
575 // the field with the largest size is expanded).
577 // For complex types, real and imaginary parts are expanded recursively.
579 // All other types are not expandable.
583 const TypeExpansionKind Kind;
585 TypeExpansion(TypeExpansionKind K) : Kind(K) {}
586 virtual ~TypeExpansion() {}
589 struct ConstantArrayExpansion : TypeExpansion {
593 ConstantArrayExpansion(QualType EltTy, uint64_t NumElts)
594 : TypeExpansion(TEK_ConstantArray), EltTy(EltTy), NumElts(NumElts) {}
595 static bool classof(const TypeExpansion *TE) {
596 return TE->Kind == TEK_ConstantArray;
600 struct RecordExpansion : TypeExpansion {
601 SmallVector<const CXXBaseSpecifier *, 1> Bases;
603 SmallVector<const FieldDecl *, 1> Fields;
605 RecordExpansion(SmallVector<const CXXBaseSpecifier *, 1> &&Bases,
606 SmallVector<const FieldDecl *, 1> &&Fields)
607 : TypeExpansion(TEK_Record), Bases(Bases), Fields(Fields) {}
608 static bool classof(const TypeExpansion *TE) {
609 return TE->Kind == TEK_Record;
613 struct ComplexExpansion : TypeExpansion {
616 ComplexExpansion(QualType EltTy) : TypeExpansion(TEK_Complex), EltTy(EltTy) {}
617 static bool classof(const TypeExpansion *TE) {
618 return TE->Kind == TEK_Complex;
622 struct NoExpansion : TypeExpansion {
623 NoExpansion() : TypeExpansion(TEK_None) {}
624 static bool classof(const TypeExpansion *TE) {
625 return TE->Kind == TEK_None;
630 static std::unique_ptr<TypeExpansion>
631 getTypeExpansion(QualType Ty, const ASTContext &Context) {
632 if (const ConstantArrayType *AT = Context.getAsConstantArrayType(Ty)) {
633 return llvm::make_unique<ConstantArrayExpansion>(
634 AT->getElementType(), AT->getSize().getZExtValue());
636 if (const RecordType *RT = Ty->getAs<RecordType>()) {
637 SmallVector<const CXXBaseSpecifier *, 1> Bases;
638 SmallVector<const FieldDecl *, 1> Fields;
639 const RecordDecl *RD = RT->getDecl();
640 assert(!RD->hasFlexibleArrayMember() &&
641 "Cannot expand structure with flexible array.");
643 // Unions can be here only in degenerative cases - all the fields are same
644 // after flattening. Thus we have to use the "largest" field.
645 const FieldDecl *LargestFD = nullptr;
646 CharUnits UnionSize = CharUnits::Zero();
648 for (const auto *FD : RD->fields()) {
649 // Skip zero length bitfields.
650 if (FD->isBitField() && FD->getBitWidthValue(Context) == 0)
652 assert(!FD->isBitField() &&
653 "Cannot expand structure with bit-field members.");
654 CharUnits FieldSize = Context.getTypeSizeInChars(FD->getType());
655 if (UnionSize < FieldSize) {
656 UnionSize = FieldSize;
661 Fields.push_back(LargestFD);
663 if (const auto *CXXRD = dyn_cast<CXXRecordDecl>(RD)) {
664 assert(!CXXRD->isDynamicClass() &&
665 "cannot expand vtable pointers in dynamic classes");
666 for (const CXXBaseSpecifier &BS : CXXRD->bases())
667 Bases.push_back(&BS);
670 for (const auto *FD : RD->fields()) {
671 // Skip zero length bitfields.
672 if (FD->isBitField() && FD->getBitWidthValue(Context) == 0)
674 assert(!FD->isBitField() &&
675 "Cannot expand structure with bit-field members.");
676 Fields.push_back(FD);
679 return llvm::make_unique<RecordExpansion>(std::move(Bases),
682 if (const ComplexType *CT = Ty->getAs<ComplexType>()) {
683 return llvm::make_unique<ComplexExpansion>(CT->getElementType());
685 return llvm::make_unique<NoExpansion>();
688 static int getExpansionSize(QualType Ty, const ASTContext &Context) {
689 auto Exp = getTypeExpansion(Ty, Context);
690 if (auto CAExp = dyn_cast<ConstantArrayExpansion>(Exp.get())) {
691 return CAExp->NumElts * getExpansionSize(CAExp->EltTy, Context);
693 if (auto RExp = dyn_cast<RecordExpansion>(Exp.get())) {
695 for (auto BS : RExp->Bases)
696 Res += getExpansionSize(BS->getType(), Context);
697 for (auto FD : RExp->Fields)
698 Res += getExpansionSize(FD->getType(), Context);
701 if (isa<ComplexExpansion>(Exp.get()))
703 assert(isa<NoExpansion>(Exp.get()));
708 CodeGenTypes::getExpandedTypes(QualType Ty,
709 SmallVectorImpl<llvm::Type *>::iterator &TI) {
710 auto Exp = getTypeExpansion(Ty, Context);
711 if (auto CAExp = dyn_cast<ConstantArrayExpansion>(Exp.get())) {
712 for (int i = 0, n = CAExp->NumElts; i < n; i++) {
713 getExpandedTypes(CAExp->EltTy, TI);
715 } else if (auto RExp = dyn_cast<RecordExpansion>(Exp.get())) {
716 for (auto BS : RExp->Bases)
717 getExpandedTypes(BS->getType(), TI);
718 for (auto FD : RExp->Fields)
719 getExpandedTypes(FD->getType(), TI);
720 } else if (auto CExp = dyn_cast<ComplexExpansion>(Exp.get())) {
721 llvm::Type *EltTy = ConvertType(CExp->EltTy);
725 assert(isa<NoExpansion>(Exp.get()));
726 *TI++ = ConvertType(Ty);
730 void CodeGenFunction::ExpandTypeFromArgs(
731 QualType Ty, LValue LV, SmallVectorImpl<llvm::Argument *>::iterator &AI) {
732 assert(LV.isSimple() &&
733 "Unexpected non-simple lvalue during struct expansion.");
735 auto Exp = getTypeExpansion(Ty, getContext());
736 if (auto CAExp = dyn_cast<ConstantArrayExpansion>(Exp.get())) {
737 for (int i = 0, n = CAExp->NumElts; i < n; i++) {
738 llvm::Value *EltAddr =
739 Builder.CreateConstGEP2_32(nullptr, LV.getAddress(), 0, i);
740 LValue LV = MakeAddrLValue(EltAddr, CAExp->EltTy);
741 ExpandTypeFromArgs(CAExp->EltTy, LV, AI);
743 } else if (auto RExp = dyn_cast<RecordExpansion>(Exp.get())) {
744 llvm::Value *This = LV.getAddress();
745 for (const CXXBaseSpecifier *BS : RExp->Bases) {
746 // Perform a single step derived-to-base conversion.
748 GetAddressOfBaseClass(This, Ty->getAsCXXRecordDecl(), &BS, &BS + 1,
749 /*NullCheckValue=*/false, SourceLocation());
750 LValue SubLV = MakeAddrLValue(Base, BS->getType());
752 // Recurse onto bases.
753 ExpandTypeFromArgs(BS->getType(), SubLV, AI);
755 for (auto FD : RExp->Fields) {
756 // FIXME: What are the right qualifiers here?
757 LValue SubLV = EmitLValueForField(LV, FD);
758 ExpandTypeFromArgs(FD->getType(), SubLV, AI);
760 } else if (auto CExp = dyn_cast<ComplexExpansion>(Exp.get())) {
761 llvm::Value *RealAddr =
762 Builder.CreateStructGEP(nullptr, LV.getAddress(), 0, "real");
763 EmitStoreThroughLValue(RValue::get(*AI++),
764 MakeAddrLValue(RealAddr, CExp->EltTy));
765 llvm::Value *ImagAddr =
766 Builder.CreateStructGEP(nullptr, LV.getAddress(), 1, "imag");
767 EmitStoreThroughLValue(RValue::get(*AI++),
768 MakeAddrLValue(ImagAddr, CExp->EltTy));
770 assert(isa<NoExpansion>(Exp.get()));
771 EmitStoreThroughLValue(RValue::get(*AI++), LV);
775 void CodeGenFunction::ExpandTypeToArgs(
776 QualType Ty, RValue RV, llvm::FunctionType *IRFuncTy,
777 SmallVectorImpl<llvm::Value *> &IRCallArgs, unsigned &IRCallArgPos) {
778 auto Exp = getTypeExpansion(Ty, getContext());
779 if (auto CAExp = dyn_cast<ConstantArrayExpansion>(Exp.get())) {
780 llvm::Value *Addr = RV.getAggregateAddr();
781 for (int i = 0, n = CAExp->NumElts; i < n; i++) {
782 llvm::Value *EltAddr = Builder.CreateConstGEP2_32(nullptr, Addr, 0, i);
784 convertTempToRValue(EltAddr, CAExp->EltTy, SourceLocation());
785 ExpandTypeToArgs(CAExp->EltTy, EltRV, IRFuncTy, IRCallArgs, IRCallArgPos);
787 } else if (auto RExp = dyn_cast<RecordExpansion>(Exp.get())) {
788 llvm::Value *This = RV.getAggregateAddr();
789 for (const CXXBaseSpecifier *BS : RExp->Bases) {
790 // Perform a single step derived-to-base conversion.
792 GetAddressOfBaseClass(This, Ty->getAsCXXRecordDecl(), &BS, &BS + 1,
793 /*NullCheckValue=*/false, SourceLocation());
794 RValue BaseRV = RValue::getAggregate(Base);
796 // Recurse onto bases.
797 ExpandTypeToArgs(BS->getType(), BaseRV, IRFuncTy, IRCallArgs,
801 LValue LV = MakeAddrLValue(This, Ty);
802 for (auto FD : RExp->Fields) {
803 RValue FldRV = EmitRValueForField(LV, FD, SourceLocation());
804 ExpandTypeToArgs(FD->getType(), FldRV, IRFuncTy, IRCallArgs,
807 } else if (isa<ComplexExpansion>(Exp.get())) {
808 ComplexPairTy CV = RV.getComplexVal();
809 IRCallArgs[IRCallArgPos++] = CV.first;
810 IRCallArgs[IRCallArgPos++] = CV.second;
812 assert(isa<NoExpansion>(Exp.get()));
813 assert(RV.isScalar() &&
814 "Unexpected non-scalar rvalue during struct expansion.");
816 // Insert a bitcast as needed.
817 llvm::Value *V = RV.getScalarVal();
818 if (IRCallArgPos < IRFuncTy->getNumParams() &&
819 V->getType() != IRFuncTy->getParamType(IRCallArgPos))
820 V = Builder.CreateBitCast(V, IRFuncTy->getParamType(IRCallArgPos));
822 IRCallArgs[IRCallArgPos++] = V;
826 /// EnterStructPointerForCoercedAccess - Given a struct pointer that we are
827 /// accessing some number of bytes out of it, try to gep into the struct to get
828 /// at its inner goodness. Dive as deep as possible without entering an element
829 /// with an in-memory size smaller than DstSize.
831 EnterStructPointerForCoercedAccess(llvm::Value *SrcPtr,
832 llvm::StructType *SrcSTy,
833 uint64_t DstSize, CodeGenFunction &CGF) {
834 // We can't dive into a zero-element struct.
835 if (SrcSTy->getNumElements() == 0) return SrcPtr;
837 llvm::Type *FirstElt = SrcSTy->getElementType(0);
839 // If the first elt is at least as large as what we're looking for, or if the
840 // first element is the same size as the whole struct, we can enter it. The
841 // comparison must be made on the store size and not the alloca size. Using
842 // the alloca size may overstate the size of the load.
843 uint64_t FirstEltSize =
844 CGF.CGM.getDataLayout().getTypeStoreSize(FirstElt);
845 if (FirstEltSize < DstSize &&
846 FirstEltSize < CGF.CGM.getDataLayout().getTypeStoreSize(SrcSTy))
849 // GEP into the first element.
850 SrcPtr = CGF.Builder.CreateConstGEP2_32(SrcSTy, SrcPtr, 0, 0, "coerce.dive");
852 // If the first element is a struct, recurse.
854 cast<llvm::PointerType>(SrcPtr->getType())->getElementType();
855 if (llvm::StructType *SrcSTy = dyn_cast<llvm::StructType>(SrcTy))
856 return EnterStructPointerForCoercedAccess(SrcPtr, SrcSTy, DstSize, CGF);
861 /// CoerceIntOrPtrToIntOrPtr - Convert a value Val to the specific Ty where both
862 /// are either integers or pointers. This does a truncation of the value if it
863 /// is too large or a zero extension if it is too small.
865 /// This behaves as if the value were coerced through memory, so on big-endian
866 /// targets the high bits are preserved in a truncation, while little-endian
867 /// targets preserve the low bits.
868 static llvm::Value *CoerceIntOrPtrToIntOrPtr(llvm::Value *Val,
870 CodeGenFunction &CGF) {
871 if (Val->getType() == Ty)
874 if (isa<llvm::PointerType>(Val->getType())) {
875 // If this is Pointer->Pointer avoid conversion to and from int.
876 if (isa<llvm::PointerType>(Ty))
877 return CGF.Builder.CreateBitCast(Val, Ty, "coerce.val");
879 // Convert the pointer to an integer so we can play with its width.
880 Val = CGF.Builder.CreatePtrToInt(Val, CGF.IntPtrTy, "coerce.val.pi");
883 llvm::Type *DestIntTy = Ty;
884 if (isa<llvm::PointerType>(DestIntTy))
885 DestIntTy = CGF.IntPtrTy;
887 if (Val->getType() != DestIntTy) {
888 const llvm::DataLayout &DL = CGF.CGM.getDataLayout();
889 if (DL.isBigEndian()) {
890 // Preserve the high bits on big-endian targets.
891 // That is what memory coercion does.
892 uint64_t SrcSize = DL.getTypeSizeInBits(Val->getType());
893 uint64_t DstSize = DL.getTypeSizeInBits(DestIntTy);
895 if (SrcSize > DstSize) {
896 Val = CGF.Builder.CreateLShr(Val, SrcSize - DstSize, "coerce.highbits");
897 Val = CGF.Builder.CreateTrunc(Val, DestIntTy, "coerce.val.ii");
899 Val = CGF.Builder.CreateZExt(Val, DestIntTy, "coerce.val.ii");
900 Val = CGF.Builder.CreateShl(Val, DstSize - SrcSize, "coerce.highbits");
903 // Little-endian targets preserve the low bits. No shifts required.
904 Val = CGF.Builder.CreateIntCast(Val, DestIntTy, false, "coerce.val.ii");
908 if (isa<llvm::PointerType>(Ty))
909 Val = CGF.Builder.CreateIntToPtr(Val, Ty, "coerce.val.ip");
915 /// CreateCoercedLoad - Create a load from \arg SrcPtr interpreted as
916 /// a pointer to an object of type \arg Ty.
918 /// This safely handles the case when the src type is smaller than the
919 /// destination type; in this situation the values of bits which not
920 /// present in the src are undefined.
921 static llvm::Value *CreateCoercedLoad(llvm::Value *SrcPtr,
923 CodeGenFunction &CGF) {
925 cast<llvm::PointerType>(SrcPtr->getType())->getElementType();
927 // If SrcTy and Ty are the same, just do a load.
929 return CGF.Builder.CreateLoad(SrcPtr);
931 uint64_t DstSize = CGF.CGM.getDataLayout().getTypeAllocSize(Ty);
933 if (llvm::StructType *SrcSTy = dyn_cast<llvm::StructType>(SrcTy)) {
934 SrcPtr = EnterStructPointerForCoercedAccess(SrcPtr, SrcSTy, DstSize, CGF);
935 SrcTy = cast<llvm::PointerType>(SrcPtr->getType())->getElementType();
938 uint64_t SrcSize = CGF.CGM.getDataLayout().getTypeAllocSize(SrcTy);
940 // If the source and destination are integer or pointer types, just do an
941 // extension or truncation to the desired type.
942 if ((isa<llvm::IntegerType>(Ty) || isa<llvm::PointerType>(Ty)) &&
943 (isa<llvm::IntegerType>(SrcTy) || isa<llvm::PointerType>(SrcTy))) {
944 llvm::LoadInst *Load = CGF.Builder.CreateLoad(SrcPtr);
945 return CoerceIntOrPtrToIntOrPtr(Load, Ty, CGF);
948 // If load is legal, just bitcast the src pointer.
949 if (SrcSize >= DstSize) {
950 // Generally SrcSize is never greater than DstSize, since this means we are
951 // losing bits. However, this can happen in cases where the structure has
952 // additional padding, for example due to a user specified alignment.
954 // FIXME: Assert that we aren't truncating non-padding bits when have access
955 // to that information.
956 llvm::Value *Casted =
957 CGF.Builder.CreateBitCast(SrcPtr, llvm::PointerType::getUnqual(Ty));
958 llvm::LoadInst *Load = CGF.Builder.CreateLoad(Casted);
959 // FIXME: Use better alignment / avoid requiring aligned load.
960 Load->setAlignment(1);
964 // Otherwise do coercion through memory. This is stupid, but
966 llvm::Value *Tmp = CGF.CreateTempAlloca(Ty);
967 llvm::Type *I8PtrTy = CGF.Builder.getInt8PtrTy();
968 llvm::Value *Casted = CGF.Builder.CreateBitCast(Tmp, I8PtrTy);
969 llvm::Value *SrcCasted = CGF.Builder.CreateBitCast(SrcPtr, I8PtrTy);
970 // FIXME: Use better alignment.
971 CGF.Builder.CreateMemCpy(Casted, SrcCasted,
972 llvm::ConstantInt::get(CGF.IntPtrTy, SrcSize),
974 return CGF.Builder.CreateLoad(Tmp);
977 // Function to store a first-class aggregate into memory. We prefer to
978 // store the elements rather than the aggregate to be more friendly to
980 // FIXME: Do we need to recurse here?
981 static void BuildAggStore(CodeGenFunction &CGF, llvm::Value *Val,
982 llvm::Value *DestPtr, bool DestIsVolatile,
984 // Prefer scalar stores to first-class aggregate stores.
985 if (llvm::StructType *STy =
986 dyn_cast<llvm::StructType>(Val->getType())) {
987 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
988 llvm::Value *EltPtr = CGF.Builder.CreateConstGEP2_32(STy, DestPtr, 0, i);
989 llvm::Value *Elt = CGF.Builder.CreateExtractValue(Val, i);
990 llvm::StoreInst *SI = CGF.Builder.CreateStore(Elt, EltPtr,
996 llvm::StoreInst *SI = CGF.Builder.CreateStore(Val, DestPtr, DestIsVolatile);
1002 /// CreateCoercedStore - Create a store to \arg DstPtr from \arg Src,
1003 /// where the source and destination may have different types.
1005 /// This safely handles the case when the src type is larger than the
1006 /// destination type; the upper bits of the src will be lost.
1007 static void CreateCoercedStore(llvm::Value *Src,
1008 llvm::Value *DstPtr,
1010 CodeGenFunction &CGF) {
1011 llvm::Type *SrcTy = Src->getType();
1013 cast<llvm::PointerType>(DstPtr->getType())->getElementType();
1014 if (SrcTy == DstTy) {
1015 CGF.Builder.CreateStore(Src, DstPtr, DstIsVolatile);
1019 uint64_t SrcSize = CGF.CGM.getDataLayout().getTypeAllocSize(SrcTy);
1021 if (llvm::StructType *DstSTy = dyn_cast<llvm::StructType>(DstTy)) {
1022 DstPtr = EnterStructPointerForCoercedAccess(DstPtr, DstSTy, SrcSize, CGF);
1023 DstTy = cast<llvm::PointerType>(DstPtr->getType())->getElementType();
1026 // If the source and destination are integer or pointer types, just do an
1027 // extension or truncation to the desired type.
1028 if ((isa<llvm::IntegerType>(SrcTy) || isa<llvm::PointerType>(SrcTy)) &&
1029 (isa<llvm::IntegerType>(DstTy) || isa<llvm::PointerType>(DstTy))) {
1030 Src = CoerceIntOrPtrToIntOrPtr(Src, DstTy, CGF);
1031 CGF.Builder.CreateStore(Src, DstPtr, DstIsVolatile);
1035 uint64_t DstSize = CGF.CGM.getDataLayout().getTypeAllocSize(DstTy);
1037 // If store is legal, just bitcast the src pointer.
1038 if (SrcSize <= DstSize) {
1039 llvm::Value *Casted =
1040 CGF.Builder.CreateBitCast(DstPtr, llvm::PointerType::getUnqual(SrcTy));
1041 // FIXME: Use better alignment / avoid requiring aligned store.
1042 BuildAggStore(CGF, Src, Casted, DstIsVolatile, true);
1044 // Otherwise do coercion through memory. This is stupid, but
1047 // Generally SrcSize is never greater than DstSize, since this means we are
1048 // losing bits. However, this can happen in cases where the structure has
1049 // additional padding, for example due to a user specified alignment.
1051 // FIXME: Assert that we aren't truncating non-padding bits when have access
1052 // to that information.
1053 llvm::Value *Tmp = CGF.CreateTempAlloca(SrcTy);
1054 CGF.Builder.CreateStore(Src, Tmp);
1055 llvm::Type *I8PtrTy = CGF.Builder.getInt8PtrTy();
1056 llvm::Value *Casted = CGF.Builder.CreateBitCast(Tmp, I8PtrTy);
1057 llvm::Value *DstCasted = CGF.Builder.CreateBitCast(DstPtr, I8PtrTy);
1058 // FIXME: Use better alignment.
1059 CGF.Builder.CreateMemCpy(DstCasted, Casted,
1060 llvm::ConstantInt::get(CGF.IntPtrTy, DstSize),
1067 /// Encapsulates information about the way function arguments from
1068 /// CGFunctionInfo should be passed to actual LLVM IR function.
1069 class ClangToLLVMArgMapping {
1070 static const unsigned InvalidIndex = ~0U;
1071 unsigned InallocaArgNo;
1073 unsigned TotalIRArgs;
1075 /// Arguments of LLVM IR function corresponding to single Clang argument.
1077 unsigned PaddingArgIndex;
1078 // Argument is expanded to IR arguments at positions
1079 // [FirstArgIndex, FirstArgIndex + NumberOfArgs).
1080 unsigned FirstArgIndex;
1081 unsigned NumberOfArgs;
1084 : PaddingArgIndex(InvalidIndex), FirstArgIndex(InvalidIndex),
1088 SmallVector<IRArgs, 8> ArgInfo;
1091 ClangToLLVMArgMapping(const ASTContext &Context, const CGFunctionInfo &FI,
1092 bool OnlyRequiredArgs = false)
1093 : InallocaArgNo(InvalidIndex), SRetArgNo(InvalidIndex), TotalIRArgs(0),
1094 ArgInfo(OnlyRequiredArgs ? FI.getNumRequiredArgs() : FI.arg_size()) {
1095 construct(Context, FI, OnlyRequiredArgs);
1098 bool hasInallocaArg() const { return InallocaArgNo != InvalidIndex; }
1099 unsigned getInallocaArgNo() const {
1100 assert(hasInallocaArg());
1101 return InallocaArgNo;
1104 bool hasSRetArg() const { return SRetArgNo != InvalidIndex; }
1105 unsigned getSRetArgNo() const {
1106 assert(hasSRetArg());
1110 unsigned totalIRArgs() const { return TotalIRArgs; }
1112 bool hasPaddingArg(unsigned ArgNo) const {
1113 assert(ArgNo < ArgInfo.size());
1114 return ArgInfo[ArgNo].PaddingArgIndex != InvalidIndex;
1116 unsigned getPaddingArgNo(unsigned ArgNo) const {
1117 assert(hasPaddingArg(ArgNo));
1118 return ArgInfo[ArgNo].PaddingArgIndex;
1121 /// Returns index of first IR argument corresponding to ArgNo, and their
1123 std::pair<unsigned, unsigned> getIRArgs(unsigned ArgNo) const {
1124 assert(ArgNo < ArgInfo.size());
1125 return std::make_pair(ArgInfo[ArgNo].FirstArgIndex,
1126 ArgInfo[ArgNo].NumberOfArgs);
1130 void construct(const ASTContext &Context, const CGFunctionInfo &FI,
1131 bool OnlyRequiredArgs);
1134 void ClangToLLVMArgMapping::construct(const ASTContext &Context,
1135 const CGFunctionInfo &FI,
1136 bool OnlyRequiredArgs) {
1137 unsigned IRArgNo = 0;
1138 bool SwapThisWithSRet = false;
1139 const ABIArgInfo &RetAI = FI.getReturnInfo();
1141 if (RetAI.getKind() == ABIArgInfo::Indirect) {
1142 SwapThisWithSRet = RetAI.isSRetAfterThis();
1143 SRetArgNo = SwapThisWithSRet ? 1 : IRArgNo++;
1147 unsigned NumArgs = OnlyRequiredArgs ? FI.getNumRequiredArgs() : FI.arg_size();
1148 for (CGFunctionInfo::const_arg_iterator I = FI.arg_begin(); ArgNo < NumArgs;
1150 assert(I != FI.arg_end());
1151 QualType ArgType = I->type;
1152 const ABIArgInfo &AI = I->info;
1153 // Collect data about IR arguments corresponding to Clang argument ArgNo.
1154 auto &IRArgs = ArgInfo[ArgNo];
1156 if (AI.getPaddingType())
1157 IRArgs.PaddingArgIndex = IRArgNo++;
1159 switch (AI.getKind()) {
1160 case ABIArgInfo::Extend:
1161 case ABIArgInfo::Direct: {
1162 // FIXME: handle sseregparm someday...
1163 llvm::StructType *STy = dyn_cast<llvm::StructType>(AI.getCoerceToType());
1164 if (AI.isDirect() && AI.getCanBeFlattened() && STy) {
1165 IRArgs.NumberOfArgs = STy->getNumElements();
1167 IRArgs.NumberOfArgs = 1;
1171 case ABIArgInfo::Indirect:
1172 IRArgs.NumberOfArgs = 1;
1174 case ABIArgInfo::Ignore:
1175 case ABIArgInfo::InAlloca:
1176 // ignore and inalloca doesn't have matching LLVM parameters.
1177 IRArgs.NumberOfArgs = 0;
1179 case ABIArgInfo::Expand: {
1180 IRArgs.NumberOfArgs = getExpansionSize(ArgType, Context);
1185 if (IRArgs.NumberOfArgs > 0) {
1186 IRArgs.FirstArgIndex = IRArgNo;
1187 IRArgNo += IRArgs.NumberOfArgs;
1190 // Skip over the sret parameter when it comes second. We already handled it
1192 if (IRArgNo == 1 && SwapThisWithSRet)
1195 assert(ArgNo == ArgInfo.size());
1197 if (FI.usesInAlloca())
1198 InallocaArgNo = IRArgNo++;
1200 TotalIRArgs = IRArgNo;
1206 bool CodeGenModule::ReturnTypeUsesSRet(const CGFunctionInfo &FI) {
1207 return FI.getReturnInfo().isIndirect();
1210 bool CodeGenModule::ReturnSlotInterferesWithArgs(const CGFunctionInfo &FI) {
1211 return ReturnTypeUsesSRet(FI) &&
1212 getTargetCodeGenInfo().doesReturnSlotInterfereWithArgs();
1215 bool CodeGenModule::ReturnTypeUsesFPRet(QualType ResultType) {
1216 if (const BuiltinType *BT = ResultType->getAs<BuiltinType>()) {
1217 switch (BT->getKind()) {
1220 case BuiltinType::Float:
1221 return getTarget().useObjCFPRetForRealType(TargetInfo::Float);
1222 case BuiltinType::Double:
1223 return getTarget().useObjCFPRetForRealType(TargetInfo::Double);
1224 case BuiltinType::LongDouble:
1225 return getTarget().useObjCFPRetForRealType(TargetInfo::LongDouble);
1232 bool CodeGenModule::ReturnTypeUsesFP2Ret(QualType ResultType) {
1233 if (const ComplexType *CT = ResultType->getAs<ComplexType>()) {
1234 if (const BuiltinType *BT = CT->getElementType()->getAs<BuiltinType>()) {
1235 if (BT->getKind() == BuiltinType::LongDouble)
1236 return getTarget().useObjCFP2RetForComplexLongDouble();
1243 llvm::FunctionType *CodeGenTypes::GetFunctionType(GlobalDecl GD) {
1244 const CGFunctionInfo &FI = arrangeGlobalDeclaration(GD);
1245 return GetFunctionType(FI);
1248 llvm::FunctionType *
1249 CodeGenTypes::GetFunctionType(const CGFunctionInfo &FI) {
1251 bool Inserted = FunctionsBeingProcessed.insert(&FI).second;
1253 assert(Inserted && "Recursively being processed?");
1255 llvm::Type *resultType = nullptr;
1256 const ABIArgInfo &retAI = FI.getReturnInfo();
1257 switch (retAI.getKind()) {
1258 case ABIArgInfo::Expand:
1259 llvm_unreachable("Invalid ABI kind for return argument");
1261 case ABIArgInfo::Extend:
1262 case ABIArgInfo::Direct:
1263 resultType = retAI.getCoerceToType();
1266 case ABIArgInfo::InAlloca:
1267 if (retAI.getInAllocaSRet()) {
1268 // sret things on win32 aren't void, they return the sret pointer.
1269 QualType ret = FI.getReturnType();
1270 llvm::Type *ty = ConvertType(ret);
1271 unsigned addressSpace = Context.getTargetAddressSpace(ret);
1272 resultType = llvm::PointerType::get(ty, addressSpace);
1274 resultType = llvm::Type::getVoidTy(getLLVMContext());
1278 case ABIArgInfo::Indirect: {
1279 assert(!retAI.getIndirectAlign() && "Align unused on indirect return.");
1280 resultType = llvm::Type::getVoidTy(getLLVMContext());
1284 case ABIArgInfo::Ignore:
1285 resultType = llvm::Type::getVoidTy(getLLVMContext());
1289 ClangToLLVMArgMapping IRFunctionArgs(getContext(), FI, true);
1290 SmallVector<llvm::Type*, 8> ArgTypes(IRFunctionArgs.totalIRArgs());
1292 // Add type for sret argument.
1293 if (IRFunctionArgs.hasSRetArg()) {
1294 QualType Ret = FI.getReturnType();
1295 llvm::Type *Ty = ConvertType(Ret);
1296 unsigned AddressSpace = Context.getTargetAddressSpace(Ret);
1297 ArgTypes[IRFunctionArgs.getSRetArgNo()] =
1298 llvm::PointerType::get(Ty, AddressSpace);
1301 // Add type for inalloca argument.
1302 if (IRFunctionArgs.hasInallocaArg()) {
1303 auto ArgStruct = FI.getArgStruct();
1305 ArgTypes[IRFunctionArgs.getInallocaArgNo()] = ArgStruct->getPointerTo();
1308 // Add in all of the required arguments.
1310 CGFunctionInfo::const_arg_iterator it = FI.arg_begin(),
1311 ie = it + FI.getNumRequiredArgs();
1312 for (; it != ie; ++it, ++ArgNo) {
1313 const ABIArgInfo &ArgInfo = it->info;
1315 // Insert a padding type to ensure proper alignment.
1316 if (IRFunctionArgs.hasPaddingArg(ArgNo))
1317 ArgTypes[IRFunctionArgs.getPaddingArgNo(ArgNo)] =
1318 ArgInfo.getPaddingType();
1320 unsigned FirstIRArg, NumIRArgs;
1321 std::tie(FirstIRArg, NumIRArgs) = IRFunctionArgs.getIRArgs(ArgNo);
1323 switch (ArgInfo.getKind()) {
1324 case ABIArgInfo::Ignore:
1325 case ABIArgInfo::InAlloca:
1326 assert(NumIRArgs == 0);
1329 case ABIArgInfo::Indirect: {
1330 assert(NumIRArgs == 1);
1331 // indirect arguments are always on the stack, which is addr space #0.
1332 llvm::Type *LTy = ConvertTypeForMem(it->type);
1333 ArgTypes[FirstIRArg] = LTy->getPointerTo();
1337 case ABIArgInfo::Extend:
1338 case ABIArgInfo::Direct: {
1339 // Fast-isel and the optimizer generally like scalar values better than
1340 // FCAs, so we flatten them if this is safe to do for this argument.
1341 llvm::Type *argType = ArgInfo.getCoerceToType();
1342 llvm::StructType *st = dyn_cast<llvm::StructType>(argType);
1343 if (st && ArgInfo.isDirect() && ArgInfo.getCanBeFlattened()) {
1344 assert(NumIRArgs == st->getNumElements());
1345 for (unsigned i = 0, e = st->getNumElements(); i != e; ++i)
1346 ArgTypes[FirstIRArg + i] = st->getElementType(i);
1348 assert(NumIRArgs == 1);
1349 ArgTypes[FirstIRArg] = argType;
1354 case ABIArgInfo::Expand:
1355 auto ArgTypesIter = ArgTypes.begin() + FirstIRArg;
1356 getExpandedTypes(it->type, ArgTypesIter);
1357 assert(ArgTypesIter == ArgTypes.begin() + FirstIRArg + NumIRArgs);
1362 bool Erased = FunctionsBeingProcessed.erase(&FI); (void)Erased;
1363 assert(Erased && "Not in set?");
1365 return llvm::FunctionType::get(resultType, ArgTypes, FI.isVariadic());
1368 llvm::Type *CodeGenTypes::GetFunctionTypeForVTable(GlobalDecl GD) {
1369 const CXXMethodDecl *MD = cast<CXXMethodDecl>(GD.getDecl());
1370 const FunctionProtoType *FPT = MD->getType()->getAs<FunctionProtoType>();
1372 if (!isFuncTypeConvertible(FPT))
1373 return llvm::StructType::get(getLLVMContext());
1375 const CGFunctionInfo *Info;
1376 if (isa<CXXDestructorDecl>(MD))
1378 &arrangeCXXStructorDeclaration(MD, getFromDtorType(GD.getDtorType()));
1380 Info = &arrangeCXXMethodDeclaration(MD);
1381 return GetFunctionType(*Info);
1384 void CodeGenModule::ConstructAttributeList(const CGFunctionInfo &FI,
1385 const Decl *TargetDecl,
1386 AttributeListType &PAL,
1387 unsigned &CallingConv,
1388 bool AttrOnCallSite) {
1389 llvm::AttrBuilder FuncAttrs;
1390 llvm::AttrBuilder RetAttrs;
1391 bool HasOptnone = false;
1393 CallingConv = FI.getEffectiveCallingConvention();
1395 if (FI.isNoReturn())
1396 FuncAttrs.addAttribute(llvm::Attribute::NoReturn);
1398 // FIXME: handle sseregparm someday...
1400 if (TargetDecl->hasAttr<ReturnsTwiceAttr>())
1401 FuncAttrs.addAttribute(llvm::Attribute::ReturnsTwice);
1402 if (TargetDecl->hasAttr<NoThrowAttr>())
1403 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1404 if (TargetDecl->hasAttr<NoReturnAttr>())
1405 FuncAttrs.addAttribute(llvm::Attribute::NoReturn);
1406 if (TargetDecl->hasAttr<NoDuplicateAttr>())
1407 FuncAttrs.addAttribute(llvm::Attribute::NoDuplicate);
1409 if (const FunctionDecl *Fn = dyn_cast<FunctionDecl>(TargetDecl)) {
1410 const FunctionProtoType *FPT = Fn->getType()->getAs<FunctionProtoType>();
1411 if (FPT && FPT->isNothrow(getContext()))
1412 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1413 // Don't use [[noreturn]] or _Noreturn for a call to a virtual function.
1414 // These attributes are not inherited by overloads.
1415 const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(Fn);
1416 if (Fn->isNoReturn() && !(AttrOnCallSite && MD && MD->isVirtual()))
1417 FuncAttrs.addAttribute(llvm::Attribute::NoReturn);
1420 // 'const' and 'pure' attribute functions are also nounwind.
1421 if (TargetDecl->hasAttr<ConstAttr>()) {
1422 FuncAttrs.addAttribute(llvm::Attribute::ReadNone);
1423 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1424 } else if (TargetDecl->hasAttr<PureAttr>()) {
1425 FuncAttrs.addAttribute(llvm::Attribute::ReadOnly);
1426 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1428 if (TargetDecl->hasAttr<RestrictAttr>())
1429 RetAttrs.addAttribute(llvm::Attribute::NoAlias);
1430 if (TargetDecl->hasAttr<ReturnsNonNullAttr>())
1431 RetAttrs.addAttribute(llvm::Attribute::NonNull);
1433 HasOptnone = TargetDecl->hasAttr<OptimizeNoneAttr>();
1436 // OptimizeNoneAttr takes precedence over -Os or -Oz. No warning needed.
1438 if (CodeGenOpts.OptimizeSize)
1439 FuncAttrs.addAttribute(llvm::Attribute::OptimizeForSize);
1440 if (CodeGenOpts.OptimizeSize == 2)
1441 FuncAttrs.addAttribute(llvm::Attribute::MinSize);
1444 if (CodeGenOpts.DisableRedZone)
1445 FuncAttrs.addAttribute(llvm::Attribute::NoRedZone);
1446 if (CodeGenOpts.NoImplicitFloat)
1447 FuncAttrs.addAttribute(llvm::Attribute::NoImplicitFloat);
1448 if (CodeGenOpts.EnableSegmentedStacks &&
1449 !(TargetDecl && TargetDecl->hasAttr<NoSplitStackAttr>()))
1450 FuncAttrs.addAttribute("split-stack");
1452 if (AttrOnCallSite) {
1453 // Attributes that should go on the call site only.
1454 if (!CodeGenOpts.SimplifyLibCalls)
1455 FuncAttrs.addAttribute(llvm::Attribute::NoBuiltin);
1457 // Attributes that should go on the function, but not the call site.
1458 if (!CodeGenOpts.DisableFPElim) {
1459 FuncAttrs.addAttribute("no-frame-pointer-elim", "false");
1460 } else if (CodeGenOpts.OmitLeafFramePointer) {
1461 FuncAttrs.addAttribute("no-frame-pointer-elim", "false");
1462 FuncAttrs.addAttribute("no-frame-pointer-elim-non-leaf");
1464 FuncAttrs.addAttribute("no-frame-pointer-elim", "true");
1465 FuncAttrs.addAttribute("no-frame-pointer-elim-non-leaf");
1468 FuncAttrs.addAttribute("less-precise-fpmad",
1469 llvm::toStringRef(CodeGenOpts.LessPreciseFPMAD));
1470 FuncAttrs.addAttribute("no-infs-fp-math",
1471 llvm::toStringRef(CodeGenOpts.NoInfsFPMath));
1472 FuncAttrs.addAttribute("no-nans-fp-math",
1473 llvm::toStringRef(CodeGenOpts.NoNaNsFPMath));
1474 FuncAttrs.addAttribute("unsafe-fp-math",
1475 llvm::toStringRef(CodeGenOpts.UnsafeFPMath));
1476 FuncAttrs.addAttribute("use-soft-float",
1477 llvm::toStringRef(CodeGenOpts.SoftFloat));
1478 FuncAttrs.addAttribute("stack-protector-buffer-size",
1479 llvm::utostr(CodeGenOpts.SSPBufferSize));
1481 if (!CodeGenOpts.StackRealignment)
1482 FuncAttrs.addAttribute("no-realign-stack");
1484 // Add target-cpu and target-features work if they differ from the defaults.
1485 std::string &CPU = getTarget().getTargetOpts().CPU;
1487 FuncAttrs.addAttribute("target-cpu", CPU);
1489 // TODO: Features gets us the features on the command line including
1490 // feature dependencies. For canonicalization purposes we might want to
1491 // avoid putting features in the target-features set if we know it'll be one
1492 // of the default features in the backend, e.g. corei7-avx and +avx or figure
1493 // out non-explicit dependencies.
1494 std::vector<std::string> &Features = getTarget().getTargetOpts().Features;
1495 if (!Features.empty()) {
1496 std::stringstream S;
1497 std::copy(Features.begin(), Features.end(),
1498 std::ostream_iterator<std::string>(S, ","));
1499 // The drop_back gets rid of the trailing space.
1500 FuncAttrs.addAttribute("target-features",
1501 StringRef(S.str()).drop_back(1));
1505 ClangToLLVMArgMapping IRFunctionArgs(getContext(), FI);
1507 QualType RetTy = FI.getReturnType();
1508 const ABIArgInfo &RetAI = FI.getReturnInfo();
1509 switch (RetAI.getKind()) {
1510 case ABIArgInfo::Extend:
1511 if (RetTy->hasSignedIntegerRepresentation())
1512 RetAttrs.addAttribute(llvm::Attribute::SExt);
1513 else if (RetTy->hasUnsignedIntegerRepresentation())
1514 RetAttrs.addAttribute(llvm::Attribute::ZExt);
1516 case ABIArgInfo::Direct:
1517 if (RetAI.getInReg())
1518 RetAttrs.addAttribute(llvm::Attribute::InReg);
1520 case ABIArgInfo::Ignore:
1523 case ABIArgInfo::InAlloca:
1524 case ABIArgInfo::Indirect: {
1525 // inalloca and sret disable readnone and readonly
1526 FuncAttrs.removeAttribute(llvm::Attribute::ReadOnly)
1527 .removeAttribute(llvm::Attribute::ReadNone);
1531 case ABIArgInfo::Expand:
1532 llvm_unreachable("Invalid ABI kind for return argument");
1535 if (const auto *RefTy = RetTy->getAs<ReferenceType>()) {
1536 QualType PTy = RefTy->getPointeeType();
1537 if (!PTy->isIncompleteType() && PTy->isConstantSizeType())
1538 RetAttrs.addDereferenceableAttr(getContext().getTypeSizeInChars(PTy)
1540 else if (getContext().getTargetAddressSpace(PTy) == 0)
1541 RetAttrs.addAttribute(llvm::Attribute::NonNull);
1544 // Attach return attributes.
1545 if (RetAttrs.hasAttributes()) {
1546 PAL.push_back(llvm::AttributeSet::get(
1547 getLLVMContext(), llvm::AttributeSet::ReturnIndex, RetAttrs));
1550 // Attach attributes to sret.
1551 if (IRFunctionArgs.hasSRetArg()) {
1552 llvm::AttrBuilder SRETAttrs;
1553 SRETAttrs.addAttribute(llvm::Attribute::StructRet);
1554 if (RetAI.getInReg())
1555 SRETAttrs.addAttribute(llvm::Attribute::InReg);
1556 PAL.push_back(llvm::AttributeSet::get(
1557 getLLVMContext(), IRFunctionArgs.getSRetArgNo() + 1, SRETAttrs));
1560 // Attach attributes to inalloca argument.
1561 if (IRFunctionArgs.hasInallocaArg()) {
1562 llvm::AttrBuilder Attrs;
1563 Attrs.addAttribute(llvm::Attribute::InAlloca);
1564 PAL.push_back(llvm::AttributeSet::get(
1565 getLLVMContext(), IRFunctionArgs.getInallocaArgNo() + 1, Attrs));
1569 for (CGFunctionInfo::const_arg_iterator I = FI.arg_begin(),
1571 I != E; ++I, ++ArgNo) {
1572 QualType ParamType = I->type;
1573 const ABIArgInfo &AI = I->info;
1574 llvm::AttrBuilder Attrs;
1576 // Add attribute for padding argument, if necessary.
1577 if (IRFunctionArgs.hasPaddingArg(ArgNo)) {
1578 if (AI.getPaddingInReg())
1579 PAL.push_back(llvm::AttributeSet::get(
1580 getLLVMContext(), IRFunctionArgs.getPaddingArgNo(ArgNo) + 1,
1581 llvm::Attribute::InReg));
1584 // 'restrict' -> 'noalias' is done in EmitFunctionProlog when we
1585 // have the corresponding parameter variable. It doesn't make
1586 // sense to do it here because parameters are so messed up.
1587 switch (AI.getKind()) {
1588 case ABIArgInfo::Extend:
1589 if (ParamType->isSignedIntegerOrEnumerationType())
1590 Attrs.addAttribute(llvm::Attribute::SExt);
1591 else if (ParamType->isUnsignedIntegerOrEnumerationType()) {
1592 if (getTypes().getABIInfo().shouldSignExtUnsignedType(ParamType))
1593 Attrs.addAttribute(llvm::Attribute::SExt);
1595 Attrs.addAttribute(llvm::Attribute::ZExt);
1598 case ABIArgInfo::Direct:
1599 if (ArgNo == 0 && FI.isChainCall())
1600 Attrs.addAttribute(llvm::Attribute::Nest);
1601 else if (AI.getInReg())
1602 Attrs.addAttribute(llvm::Attribute::InReg);
1605 case ABIArgInfo::Indirect:
1607 Attrs.addAttribute(llvm::Attribute::InReg);
1609 if (AI.getIndirectByVal())
1610 Attrs.addAttribute(llvm::Attribute::ByVal);
1612 Attrs.addAlignmentAttr(AI.getIndirectAlign());
1614 // byval disables readnone and readonly.
1615 FuncAttrs.removeAttribute(llvm::Attribute::ReadOnly)
1616 .removeAttribute(llvm::Attribute::ReadNone);
1619 case ABIArgInfo::Ignore:
1620 case ABIArgInfo::Expand:
1623 case ABIArgInfo::InAlloca:
1624 // inalloca disables readnone and readonly.
1625 FuncAttrs.removeAttribute(llvm::Attribute::ReadOnly)
1626 .removeAttribute(llvm::Attribute::ReadNone);
1630 if (const auto *RefTy = ParamType->getAs<ReferenceType>()) {
1631 QualType PTy = RefTy->getPointeeType();
1632 if (!PTy->isIncompleteType() && PTy->isConstantSizeType())
1633 Attrs.addDereferenceableAttr(getContext().getTypeSizeInChars(PTy)
1635 else if (getContext().getTargetAddressSpace(PTy) == 0)
1636 Attrs.addAttribute(llvm::Attribute::NonNull);
1639 if (Attrs.hasAttributes()) {
1640 unsigned FirstIRArg, NumIRArgs;
1641 std::tie(FirstIRArg, NumIRArgs) = IRFunctionArgs.getIRArgs(ArgNo);
1642 for (unsigned i = 0; i < NumIRArgs; i++)
1643 PAL.push_back(llvm::AttributeSet::get(getLLVMContext(),
1644 FirstIRArg + i + 1, Attrs));
1647 assert(ArgNo == FI.arg_size());
1649 if (FuncAttrs.hasAttributes())
1650 PAL.push_back(llvm::
1651 AttributeSet::get(getLLVMContext(),
1652 llvm::AttributeSet::FunctionIndex,
1656 /// An argument came in as a promoted argument; demote it back to its
1658 static llvm::Value *emitArgumentDemotion(CodeGenFunction &CGF,
1660 llvm::Value *value) {
1661 llvm::Type *varType = CGF.ConvertType(var->getType());
1663 // This can happen with promotions that actually don't change the
1664 // underlying type, like the enum promotions.
1665 if (value->getType() == varType) return value;
1667 assert((varType->isIntegerTy() || varType->isFloatingPointTy())
1668 && "unexpected promotion type");
1670 if (isa<llvm::IntegerType>(varType))
1671 return CGF.Builder.CreateTrunc(value, varType, "arg.unpromote");
1673 return CGF.Builder.CreateFPCast(value, varType, "arg.unpromote");
1676 /// Returns the attribute (either parameter attribute, or function
1677 /// attribute), which declares argument ArgNo to be non-null.
1678 static const NonNullAttr *getNonNullAttr(const Decl *FD, const ParmVarDecl *PVD,
1679 QualType ArgType, unsigned ArgNo) {
1680 // FIXME: __attribute__((nonnull)) can also be applied to:
1681 // - references to pointers, where the pointee is known to be
1682 // nonnull (apparently a Clang extension)
1683 // - transparent unions containing pointers
1684 // In the former case, LLVM IR cannot represent the constraint. In
1685 // the latter case, we have no guarantee that the transparent union
1686 // is in fact passed as a pointer.
1687 if (!ArgType->isAnyPointerType() && !ArgType->isBlockPointerType())
1689 // First, check attribute on parameter itself.
1691 if (auto ParmNNAttr = PVD->getAttr<NonNullAttr>())
1694 // Check function attributes.
1697 for (const auto *NNAttr : FD->specific_attrs<NonNullAttr>()) {
1698 if (NNAttr->isNonNull(ArgNo))
1704 void CodeGenFunction::EmitFunctionProlog(const CGFunctionInfo &FI,
1706 const FunctionArgList &Args) {
1707 if (CurCodeDecl && CurCodeDecl->hasAttr<NakedAttr>())
1708 // Naked functions don't have prologues.
1711 // If this is an implicit-return-zero function, go ahead and
1712 // initialize the return value. TODO: it might be nice to have
1713 // a more general mechanism for this that didn't require synthesized
1714 // return statements.
1715 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(CurCodeDecl)) {
1716 if (FD->hasImplicitReturnZero()) {
1717 QualType RetTy = FD->getReturnType().getUnqualifiedType();
1718 llvm::Type* LLVMTy = CGM.getTypes().ConvertType(RetTy);
1719 llvm::Constant* Zero = llvm::Constant::getNullValue(LLVMTy);
1720 Builder.CreateStore(Zero, ReturnValue);
1724 // FIXME: We no longer need the types from FunctionArgList; lift up and
1727 ClangToLLVMArgMapping IRFunctionArgs(CGM.getContext(), FI);
1728 // Flattened function arguments.
1729 SmallVector<llvm::Argument *, 16> FnArgs;
1730 FnArgs.reserve(IRFunctionArgs.totalIRArgs());
1731 for (auto &Arg : Fn->args()) {
1732 FnArgs.push_back(&Arg);
1734 assert(FnArgs.size() == IRFunctionArgs.totalIRArgs());
1736 // If we're using inalloca, all the memory arguments are GEPs off of the last
1737 // parameter, which is a pointer to the complete memory area.
1738 llvm::Value *ArgStruct = nullptr;
1739 if (IRFunctionArgs.hasInallocaArg()) {
1740 ArgStruct = FnArgs[IRFunctionArgs.getInallocaArgNo()];
1741 assert(ArgStruct->getType() == FI.getArgStruct()->getPointerTo());
1744 // Name the struct return parameter.
1745 if (IRFunctionArgs.hasSRetArg()) {
1746 auto AI = FnArgs[IRFunctionArgs.getSRetArgNo()];
1747 AI->setName("agg.result");
1748 AI->addAttr(llvm::AttributeSet::get(getLLVMContext(), AI->getArgNo() + 1,
1749 llvm::Attribute::NoAlias));
1752 // Track if we received the parameter as a pointer (indirect, byval, or
1753 // inalloca). If already have a pointer, EmitParmDecl doesn't need to copy it
1754 // into a local alloca for us.
1755 enum ValOrPointer { HaveValue = 0, HavePointer = 1 };
1756 typedef llvm::PointerIntPair<llvm::Value *, 1> ValueAndIsPtr;
1757 SmallVector<ValueAndIsPtr, 16> ArgVals;
1758 ArgVals.reserve(Args.size());
1760 // Create a pointer value for every parameter declaration. This usually
1761 // entails copying one or more LLVM IR arguments into an alloca. Don't push
1762 // any cleanups or do anything that might unwind. We do that separately, so
1763 // we can push the cleanups in the correct order for the ABI.
1764 assert(FI.arg_size() == Args.size() &&
1765 "Mismatch between function signature & arguments.");
1767 CGFunctionInfo::const_arg_iterator info_it = FI.arg_begin();
1768 for (FunctionArgList::const_iterator i = Args.begin(), e = Args.end();
1769 i != e; ++i, ++info_it, ++ArgNo) {
1770 const VarDecl *Arg = *i;
1771 QualType Ty = info_it->type;
1772 const ABIArgInfo &ArgI = info_it->info;
1775 isa<ParmVarDecl>(Arg) && cast<ParmVarDecl>(Arg)->isKNRPromoted();
1777 unsigned FirstIRArg, NumIRArgs;
1778 std::tie(FirstIRArg, NumIRArgs) = IRFunctionArgs.getIRArgs(ArgNo);
1780 switch (ArgI.getKind()) {
1781 case ABIArgInfo::InAlloca: {
1782 assert(NumIRArgs == 0);
1784 Builder.CreateStructGEP(FI.getArgStruct(), ArgStruct,
1785 ArgI.getInAllocaFieldIndex(), Arg->getName());
1786 ArgVals.push_back(ValueAndIsPtr(V, HavePointer));
1790 case ABIArgInfo::Indirect: {
1791 assert(NumIRArgs == 1);
1792 llvm::Value *V = FnArgs[FirstIRArg];
1794 if (!hasScalarEvaluationKind(Ty)) {
1795 // Aggregates and complex variables are accessed by reference. All we
1796 // need to do is realign the value, if requested
1797 if (ArgI.getIndirectRealign()) {
1798 llvm::Value *AlignedTemp = CreateMemTemp(Ty, "coerce");
1800 // Copy from the incoming argument pointer to the temporary with the
1801 // appropriate alignment.
1803 // FIXME: We should have a common utility for generating an aggregate
1805 llvm::Type *I8PtrTy = Builder.getInt8PtrTy();
1806 CharUnits Size = getContext().getTypeSizeInChars(Ty);
1807 llvm::Value *Dst = Builder.CreateBitCast(AlignedTemp, I8PtrTy);
1808 llvm::Value *Src = Builder.CreateBitCast(V, I8PtrTy);
1809 Builder.CreateMemCpy(Dst,
1811 llvm::ConstantInt::get(IntPtrTy,
1812 Size.getQuantity()),
1813 ArgI.getIndirectAlign(),
1817 ArgVals.push_back(ValueAndIsPtr(V, HavePointer));
1819 // Load scalar value from indirect argument.
1820 V = EmitLoadOfScalar(V, false, ArgI.getIndirectAlign(), Ty,
1821 Arg->getLocStart());
1824 V = emitArgumentDemotion(*this, Arg, V);
1825 ArgVals.push_back(ValueAndIsPtr(V, HaveValue));
1830 case ABIArgInfo::Extend:
1831 case ABIArgInfo::Direct: {
1833 // If we have the trivial case, handle it with no muss and fuss.
1834 if (!isa<llvm::StructType>(ArgI.getCoerceToType()) &&
1835 ArgI.getCoerceToType() == ConvertType(Ty) &&
1836 ArgI.getDirectOffset() == 0) {
1837 assert(NumIRArgs == 1);
1838 auto AI = FnArgs[FirstIRArg];
1839 llvm::Value *V = AI;
1841 if (const ParmVarDecl *PVD = dyn_cast<ParmVarDecl>(Arg)) {
1842 if (getNonNullAttr(CurCodeDecl, PVD, PVD->getType(),
1843 PVD->getFunctionScopeIndex()))
1844 AI->addAttr(llvm::AttributeSet::get(getLLVMContext(),
1846 llvm::Attribute::NonNull));
1848 QualType OTy = PVD->getOriginalType();
1849 if (const auto *ArrTy =
1850 getContext().getAsConstantArrayType(OTy)) {
1851 // A C99 array parameter declaration with the static keyword also
1852 // indicates dereferenceability, and if the size is constant we can
1853 // use the dereferenceable attribute (which requires the size in
1855 if (ArrTy->getSizeModifier() == ArrayType::Static) {
1856 QualType ETy = ArrTy->getElementType();
1857 uint64_t ArrSize = ArrTy->getSize().getZExtValue();
1858 if (!ETy->isIncompleteType() && ETy->isConstantSizeType() &&
1860 llvm::AttrBuilder Attrs;
1861 Attrs.addDereferenceableAttr(
1862 getContext().getTypeSizeInChars(ETy).getQuantity()*ArrSize);
1863 AI->addAttr(llvm::AttributeSet::get(getLLVMContext(),
1864 AI->getArgNo() + 1, Attrs));
1865 } else if (getContext().getTargetAddressSpace(ETy) == 0) {
1866 AI->addAttr(llvm::AttributeSet::get(getLLVMContext(),
1868 llvm::Attribute::NonNull));
1871 } else if (const auto *ArrTy =
1872 getContext().getAsVariableArrayType(OTy)) {
1873 // For C99 VLAs with the static keyword, we don't know the size so
1874 // we can't use the dereferenceable attribute, but in addrspace(0)
1875 // we know that it must be nonnull.
1876 if (ArrTy->getSizeModifier() == VariableArrayType::Static &&
1877 !getContext().getTargetAddressSpace(ArrTy->getElementType()))
1878 AI->addAttr(llvm::AttributeSet::get(getLLVMContext(),
1880 llvm::Attribute::NonNull));
1883 const auto *AVAttr = PVD->getAttr<AlignValueAttr>();
1885 if (const auto *TOTy = dyn_cast<TypedefType>(OTy))
1886 AVAttr = TOTy->getDecl()->getAttr<AlignValueAttr>();
1888 llvm::Value *AlignmentValue =
1889 EmitScalarExpr(AVAttr->getAlignment());
1890 llvm::ConstantInt *AlignmentCI =
1891 cast<llvm::ConstantInt>(AlignmentValue);
1892 unsigned Alignment =
1893 std::min((unsigned) AlignmentCI->getZExtValue(),
1894 +llvm::Value::MaximumAlignment);
1896 llvm::AttrBuilder Attrs;
1897 Attrs.addAlignmentAttr(Alignment);
1898 AI->addAttr(llvm::AttributeSet::get(getLLVMContext(),
1899 AI->getArgNo() + 1, Attrs));
1903 if (Arg->getType().isRestrictQualified())
1904 AI->addAttr(llvm::AttributeSet::get(getLLVMContext(),
1906 llvm::Attribute::NoAlias));
1908 // Ensure the argument is the correct type.
1909 if (V->getType() != ArgI.getCoerceToType())
1910 V = Builder.CreateBitCast(V, ArgI.getCoerceToType());
1913 V = emitArgumentDemotion(*this, Arg, V);
1915 if (const CXXMethodDecl *MD =
1916 dyn_cast_or_null<CXXMethodDecl>(CurCodeDecl)) {
1917 if (MD->isVirtual() && Arg == CXXABIThisDecl)
1918 V = CGM.getCXXABI().
1919 adjustThisParameterInVirtualFunctionPrologue(*this, CurGD, V);
1922 // Because of merging of function types from multiple decls it is
1923 // possible for the type of an argument to not match the corresponding
1924 // type in the function type. Since we are codegening the callee
1925 // in here, add a cast to the argument type.
1926 llvm::Type *LTy = ConvertType(Arg->getType());
1927 if (V->getType() != LTy)
1928 V = Builder.CreateBitCast(V, LTy);
1930 ArgVals.push_back(ValueAndIsPtr(V, HaveValue));
1934 llvm::AllocaInst *Alloca = CreateMemTemp(Ty, Arg->getName());
1936 // The alignment we need to use is the max of the requested alignment for
1937 // the argument plus the alignment required by our access code below.
1938 unsigned AlignmentToUse =
1939 CGM.getDataLayout().getABITypeAlignment(ArgI.getCoerceToType());
1940 AlignmentToUse = std::max(AlignmentToUse,
1941 (unsigned)getContext().getDeclAlign(Arg).getQuantity());
1943 Alloca->setAlignment(AlignmentToUse);
1944 llvm::Value *V = Alloca;
1945 llvm::Value *Ptr = V; // Pointer to store into.
1947 // If the value is offset in memory, apply the offset now.
1948 if (unsigned Offs = ArgI.getDirectOffset()) {
1949 Ptr = Builder.CreateBitCast(Ptr, Builder.getInt8PtrTy());
1950 Ptr = Builder.CreateConstGEP1_32(Builder.getInt8Ty(), Ptr, Offs);
1951 Ptr = Builder.CreateBitCast(Ptr,
1952 llvm::PointerType::getUnqual(ArgI.getCoerceToType()));
1955 // Fast-isel and the optimizer generally like scalar values better than
1956 // FCAs, so we flatten them if this is safe to do for this argument.
1957 llvm::StructType *STy = dyn_cast<llvm::StructType>(ArgI.getCoerceToType());
1958 if (ArgI.isDirect() && ArgI.getCanBeFlattened() && STy &&
1959 STy->getNumElements() > 1) {
1960 uint64_t SrcSize = CGM.getDataLayout().getTypeAllocSize(STy);
1962 cast<llvm::PointerType>(Ptr->getType())->getElementType();
1963 uint64_t DstSize = CGM.getDataLayout().getTypeAllocSize(DstTy);
1965 if (SrcSize <= DstSize) {
1966 Ptr = Builder.CreateBitCast(Ptr, llvm::PointerType::getUnqual(STy));
1968 assert(STy->getNumElements() == NumIRArgs);
1969 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1970 auto AI = FnArgs[FirstIRArg + i];
1971 AI->setName(Arg->getName() + ".coerce" + Twine(i));
1972 llvm::Value *EltPtr = Builder.CreateConstGEP2_32(STy, Ptr, 0, i);
1973 Builder.CreateStore(AI, EltPtr);
1976 llvm::AllocaInst *TempAlloca =
1977 CreateTempAlloca(ArgI.getCoerceToType(), "coerce");
1978 TempAlloca->setAlignment(AlignmentToUse);
1979 llvm::Value *TempV = TempAlloca;
1981 assert(STy->getNumElements() == NumIRArgs);
1982 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1983 auto AI = FnArgs[FirstIRArg + i];
1984 AI->setName(Arg->getName() + ".coerce" + Twine(i));
1985 llvm::Value *EltPtr =
1986 Builder.CreateConstGEP2_32(ArgI.getCoerceToType(), TempV, 0, i);
1987 Builder.CreateStore(AI, EltPtr);
1990 Builder.CreateMemCpy(Ptr, TempV, DstSize, AlignmentToUse);
1993 // Simple case, just do a coerced store of the argument into the alloca.
1994 assert(NumIRArgs == 1);
1995 auto AI = FnArgs[FirstIRArg];
1996 AI->setName(Arg->getName() + ".coerce");
1997 CreateCoercedStore(AI, Ptr, /*DestIsVolatile=*/false, *this);
2001 // Match to what EmitParmDecl is expecting for this type.
2002 if (CodeGenFunction::hasScalarEvaluationKind(Ty)) {
2003 V = EmitLoadOfScalar(V, false, AlignmentToUse, Ty, Arg->getLocStart());
2005 V = emitArgumentDemotion(*this, Arg, V);
2006 ArgVals.push_back(ValueAndIsPtr(V, HaveValue));
2008 ArgVals.push_back(ValueAndIsPtr(V, HavePointer));
2013 case ABIArgInfo::Expand: {
2014 // If this structure was expanded into multiple arguments then
2015 // we need to create a temporary and reconstruct it from the
2017 llvm::AllocaInst *Alloca = CreateMemTemp(Ty);
2018 CharUnits Align = getContext().getDeclAlign(Arg);
2019 Alloca->setAlignment(Align.getQuantity());
2020 LValue LV = MakeAddrLValue(Alloca, Ty, Align);
2021 ArgVals.push_back(ValueAndIsPtr(Alloca, HavePointer));
2023 auto FnArgIter = FnArgs.begin() + FirstIRArg;
2024 ExpandTypeFromArgs(Ty, LV, FnArgIter);
2025 assert(FnArgIter == FnArgs.begin() + FirstIRArg + NumIRArgs);
2026 for (unsigned i = 0, e = NumIRArgs; i != e; ++i) {
2027 auto AI = FnArgs[FirstIRArg + i];
2028 AI->setName(Arg->getName() + "." + Twine(i));
2033 case ABIArgInfo::Ignore:
2034 assert(NumIRArgs == 0);
2035 // Initialize the local variable appropriately.
2036 if (!hasScalarEvaluationKind(Ty)) {
2037 ArgVals.push_back(ValueAndIsPtr(CreateMemTemp(Ty), HavePointer));
2039 llvm::Value *U = llvm::UndefValue::get(ConvertType(Arg->getType()));
2040 ArgVals.push_back(ValueAndIsPtr(U, HaveValue));
2046 if (getTarget().getCXXABI().areArgsDestroyedLeftToRightInCallee()) {
2047 for (int I = Args.size() - 1; I >= 0; --I)
2048 EmitParmDecl(*Args[I], ArgVals[I].getPointer(), ArgVals[I].getInt(),
2051 for (unsigned I = 0, E = Args.size(); I != E; ++I)
2052 EmitParmDecl(*Args[I], ArgVals[I].getPointer(), ArgVals[I].getInt(),
2057 static void eraseUnusedBitCasts(llvm::Instruction *insn) {
2058 while (insn->use_empty()) {
2059 llvm::BitCastInst *bitcast = dyn_cast<llvm::BitCastInst>(insn);
2060 if (!bitcast) return;
2062 // This is "safe" because we would have used a ConstantExpr otherwise.
2063 insn = cast<llvm::Instruction>(bitcast->getOperand(0));
2064 bitcast->eraseFromParent();
2068 /// Try to emit a fused autorelease of a return result.
2069 static llvm::Value *tryEmitFusedAutoreleaseOfResult(CodeGenFunction &CGF,
2070 llvm::Value *result) {
2071 // We must be immediately followed the cast.
2072 llvm::BasicBlock *BB = CGF.Builder.GetInsertBlock();
2073 if (BB->empty()) return nullptr;
2074 if (&BB->back() != result) return nullptr;
2076 llvm::Type *resultType = result->getType();
2078 // result is in a BasicBlock and is therefore an Instruction.
2079 llvm::Instruction *generator = cast<llvm::Instruction>(result);
2081 SmallVector<llvm::Instruction*,4> insnsToKill;
2084 // %generator = bitcast %type1* %generator2 to %type2*
2085 while (llvm::BitCastInst *bitcast = dyn_cast<llvm::BitCastInst>(generator)) {
2086 // We would have emitted this as a constant if the operand weren't
2088 generator = cast<llvm::Instruction>(bitcast->getOperand(0));
2090 // Require the generator to be immediately followed by the cast.
2091 if (generator->getNextNode() != bitcast)
2094 insnsToKill.push_back(bitcast);
2098 // %generator = call i8* @objc_retain(i8* %originalResult)
2100 // %generator = call i8* @objc_retainAutoreleasedReturnValue(i8* %originalResult)
2101 llvm::CallInst *call = dyn_cast<llvm::CallInst>(generator);
2102 if (!call) return nullptr;
2104 bool doRetainAutorelease;
2106 if (call->getCalledValue() == CGF.CGM.getARCEntrypoints().objc_retain) {
2107 doRetainAutorelease = true;
2108 } else if (call->getCalledValue() == CGF.CGM.getARCEntrypoints()
2109 .objc_retainAutoreleasedReturnValue) {
2110 doRetainAutorelease = false;
2112 // If we emitted an assembly marker for this call (and the
2113 // ARCEntrypoints field should have been set if so), go looking
2114 // for that call. If we can't find it, we can't do this
2115 // optimization. But it should always be the immediately previous
2116 // instruction, unless we needed bitcasts around the call.
2117 if (CGF.CGM.getARCEntrypoints().retainAutoreleasedReturnValueMarker) {
2118 llvm::Instruction *prev = call->getPrevNode();
2120 if (isa<llvm::BitCastInst>(prev)) {
2121 prev = prev->getPrevNode();
2124 assert(isa<llvm::CallInst>(prev));
2125 assert(cast<llvm::CallInst>(prev)->getCalledValue() ==
2126 CGF.CGM.getARCEntrypoints().retainAutoreleasedReturnValueMarker);
2127 insnsToKill.push_back(prev);
2133 result = call->getArgOperand(0);
2134 insnsToKill.push_back(call);
2136 // Keep killing bitcasts, for sanity. Note that we no longer care
2137 // about precise ordering as long as there's exactly one use.
2138 while (llvm::BitCastInst *bitcast = dyn_cast<llvm::BitCastInst>(result)) {
2139 if (!bitcast->hasOneUse()) break;
2140 insnsToKill.push_back(bitcast);
2141 result = bitcast->getOperand(0);
2144 // Delete all the unnecessary instructions, from latest to earliest.
2145 for (SmallVectorImpl<llvm::Instruction*>::iterator
2146 i = insnsToKill.begin(), e = insnsToKill.end(); i != e; ++i)
2147 (*i)->eraseFromParent();
2149 // Do the fused retain/autorelease if we were asked to.
2150 if (doRetainAutorelease)
2151 result = CGF.EmitARCRetainAutoreleaseReturnValue(result);
2153 // Cast back to the result type.
2154 return CGF.Builder.CreateBitCast(result, resultType);
2157 /// If this is a +1 of the value of an immutable 'self', remove it.
2158 static llvm::Value *tryRemoveRetainOfSelf(CodeGenFunction &CGF,
2159 llvm::Value *result) {
2160 // This is only applicable to a method with an immutable 'self'.
2161 const ObjCMethodDecl *method =
2162 dyn_cast_or_null<ObjCMethodDecl>(CGF.CurCodeDecl);
2163 if (!method) return nullptr;
2164 const VarDecl *self = method->getSelfDecl();
2165 if (!self->getType().isConstQualified()) return nullptr;
2167 // Look for a retain call.
2168 llvm::CallInst *retainCall =
2169 dyn_cast<llvm::CallInst>(result->stripPointerCasts());
2171 retainCall->getCalledValue() != CGF.CGM.getARCEntrypoints().objc_retain)
2174 // Look for an ordinary load of 'self'.
2175 llvm::Value *retainedValue = retainCall->getArgOperand(0);
2176 llvm::LoadInst *load =
2177 dyn_cast<llvm::LoadInst>(retainedValue->stripPointerCasts());
2178 if (!load || load->isAtomic() || load->isVolatile() ||
2179 load->getPointerOperand() != CGF.GetAddrOfLocalVar(self))
2182 // Okay! Burn it all down. This relies for correctness on the
2183 // assumption that the retain is emitted as part of the return and
2184 // that thereafter everything is used "linearly".
2185 llvm::Type *resultType = result->getType();
2186 eraseUnusedBitCasts(cast<llvm::Instruction>(result));
2187 assert(retainCall->use_empty());
2188 retainCall->eraseFromParent();
2189 eraseUnusedBitCasts(cast<llvm::Instruction>(retainedValue));
2191 return CGF.Builder.CreateBitCast(load, resultType);
2194 /// Emit an ARC autorelease of the result of a function.
2196 /// \return the value to actually return from the function
2197 static llvm::Value *emitAutoreleaseOfResult(CodeGenFunction &CGF,
2198 llvm::Value *result) {
2199 // If we're returning 'self', kill the initial retain. This is a
2200 // heuristic attempt to "encourage correctness" in the really unfortunate
2201 // case where we have a return of self during a dealloc and we desperately
2202 // need to avoid the possible autorelease.
2203 if (llvm::Value *self = tryRemoveRetainOfSelf(CGF, result))
2206 // At -O0, try to emit a fused retain/autorelease.
2207 if (CGF.shouldUseFusedARCCalls())
2208 if (llvm::Value *fused = tryEmitFusedAutoreleaseOfResult(CGF, result))
2211 return CGF.EmitARCAutoreleaseReturnValue(result);
2214 /// Heuristically search for a dominating store to the return-value slot.
2215 static llvm::StoreInst *findDominatingStoreToReturnValue(CodeGenFunction &CGF) {
2216 // If there are multiple uses of the return-value slot, just check
2217 // for something immediately preceding the IP. Sometimes this can
2218 // happen with how we generate implicit-returns; it can also happen
2219 // with noreturn cleanups.
2220 if (!CGF.ReturnValue->hasOneUse()) {
2221 llvm::BasicBlock *IP = CGF.Builder.GetInsertBlock();
2222 if (IP->empty()) return nullptr;
2223 llvm::Instruction *I = &IP->back();
2225 // Skip lifetime markers
2226 for (llvm::BasicBlock::reverse_iterator II = IP->rbegin(),
2229 if (llvm::IntrinsicInst *Intrinsic =
2230 dyn_cast<llvm::IntrinsicInst>(&*II)) {
2231 if (Intrinsic->getIntrinsicID() == llvm::Intrinsic::lifetime_end) {
2232 const llvm::Value *CastAddr = Intrinsic->getArgOperand(1);
2234 if (isa<llvm::BitCastInst>(&*II)) {
2235 if (CastAddr == &*II) {
2245 llvm::StoreInst *store = dyn_cast<llvm::StoreInst>(I);
2246 if (!store) return nullptr;
2247 if (store->getPointerOperand() != CGF.ReturnValue) return nullptr;
2248 assert(!store->isAtomic() && !store->isVolatile()); // see below
2252 llvm::StoreInst *store =
2253 dyn_cast<llvm::StoreInst>(CGF.ReturnValue->user_back());
2254 if (!store) return nullptr;
2256 // These aren't actually possible for non-coerced returns, and we
2257 // only care about non-coerced returns on this code path.
2258 assert(!store->isAtomic() && !store->isVolatile());
2260 // Now do a first-and-dirty dominance check: just walk up the
2261 // single-predecessors chain from the current insertion point.
2262 llvm::BasicBlock *StoreBB = store->getParent();
2263 llvm::BasicBlock *IP = CGF.Builder.GetInsertBlock();
2264 while (IP != StoreBB) {
2265 if (!(IP = IP->getSinglePredecessor()))
2269 // Okay, the store's basic block dominates the insertion point; we
2270 // can do our thing.
2274 void CodeGenFunction::EmitFunctionEpilog(const CGFunctionInfo &FI,
2276 SourceLocation EndLoc) {
2277 if (CurCodeDecl && CurCodeDecl->hasAttr<NakedAttr>()) {
2278 // Naked functions don't have epilogues.
2279 Builder.CreateUnreachable();
2283 // Functions with no result always return void.
2285 Builder.CreateRetVoid();
2289 llvm::DebugLoc RetDbgLoc;
2290 llvm::Value *RV = nullptr;
2291 QualType RetTy = FI.getReturnType();
2292 const ABIArgInfo &RetAI = FI.getReturnInfo();
2294 switch (RetAI.getKind()) {
2295 case ABIArgInfo::InAlloca:
2296 // Aggregrates get evaluated directly into the destination. Sometimes we
2297 // need to return the sret value in a register, though.
2298 assert(hasAggregateEvaluationKind(RetTy));
2299 if (RetAI.getInAllocaSRet()) {
2300 llvm::Function::arg_iterator EI = CurFn->arg_end();
2302 llvm::Value *ArgStruct = EI;
2303 llvm::Value *SRet = Builder.CreateStructGEP(
2304 nullptr, ArgStruct, RetAI.getInAllocaFieldIndex());
2305 RV = Builder.CreateLoad(SRet, "sret");
2309 case ABIArgInfo::Indirect: {
2310 auto AI = CurFn->arg_begin();
2311 if (RetAI.isSRetAfterThis())
2313 switch (getEvaluationKind(RetTy)) {
2316 EmitLoadOfComplex(MakeNaturalAlignAddrLValue(ReturnValue, RetTy),
2318 EmitStoreOfComplex(RT, MakeNaturalAlignAddrLValue(AI, RetTy),
2323 // Do nothing; aggregrates get evaluated directly into the destination.
2326 EmitStoreOfScalar(Builder.CreateLoad(ReturnValue),
2327 MakeNaturalAlignAddrLValue(AI, RetTy),
2334 case ABIArgInfo::Extend:
2335 case ABIArgInfo::Direct:
2336 if (RetAI.getCoerceToType() == ConvertType(RetTy) &&
2337 RetAI.getDirectOffset() == 0) {
2338 // The internal return value temp always will have pointer-to-return-type
2339 // type, just do a load.
2341 // If there is a dominating store to ReturnValue, we can elide
2342 // the load, zap the store, and usually zap the alloca.
2343 if (llvm::StoreInst *SI =
2344 findDominatingStoreToReturnValue(*this)) {
2345 // Reuse the debug location from the store unless there is
2346 // cleanup code to be emitted between the store and return
2348 if (EmitRetDbgLoc && !AutoreleaseResult)
2349 RetDbgLoc = SI->getDebugLoc();
2350 // Get the stored value and nuke the now-dead store.
2351 RV = SI->getValueOperand();
2352 SI->eraseFromParent();
2354 // If that was the only use of the return value, nuke it as well now.
2355 if (ReturnValue->use_empty() && isa<llvm::AllocaInst>(ReturnValue)) {
2356 cast<llvm::AllocaInst>(ReturnValue)->eraseFromParent();
2357 ReturnValue = nullptr;
2360 // Otherwise, we have to do a simple load.
2362 RV = Builder.CreateLoad(ReturnValue);
2365 llvm::Value *V = ReturnValue;
2366 // If the value is offset in memory, apply the offset now.
2367 if (unsigned Offs = RetAI.getDirectOffset()) {
2368 V = Builder.CreateBitCast(V, Builder.getInt8PtrTy());
2369 V = Builder.CreateConstGEP1_32(Builder.getInt8Ty(), V, Offs);
2370 V = Builder.CreateBitCast(V,
2371 llvm::PointerType::getUnqual(RetAI.getCoerceToType()));
2374 RV = CreateCoercedLoad(V, RetAI.getCoerceToType(), *this);
2377 // In ARC, end functions that return a retainable type with a call
2378 // to objc_autoreleaseReturnValue.
2379 if (AutoreleaseResult) {
2380 assert(getLangOpts().ObjCAutoRefCount &&
2381 !FI.isReturnsRetained() &&
2382 RetTy->isObjCRetainableType());
2383 RV = emitAutoreleaseOfResult(*this, RV);
2388 case ABIArgInfo::Ignore:
2391 case ABIArgInfo::Expand:
2392 llvm_unreachable("Invalid ABI kind for return argument");
2395 llvm::Instruction *Ret;
2397 if (SanOpts.has(SanitizerKind::ReturnsNonnullAttribute)) {
2398 if (auto RetNNAttr = CurGD.getDecl()->getAttr<ReturnsNonNullAttr>()) {
2399 SanitizerScope SanScope(this);
2400 llvm::Value *Cond = Builder.CreateICmpNE(
2401 RV, llvm::Constant::getNullValue(RV->getType()));
2402 llvm::Constant *StaticData[] = {
2403 EmitCheckSourceLocation(EndLoc),
2404 EmitCheckSourceLocation(RetNNAttr->getLocation()),
2406 EmitCheck(std::make_pair(Cond, SanitizerKind::ReturnsNonnullAttribute),
2407 "nonnull_return", StaticData, None);
2410 Ret = Builder.CreateRet(RV);
2412 Ret = Builder.CreateRetVoid();
2416 Ret->setDebugLoc(std::move(RetDbgLoc));
2419 static bool isInAllocaArgument(CGCXXABI &ABI, QualType type) {
2420 const CXXRecordDecl *RD = type->getAsCXXRecordDecl();
2421 return RD && ABI.getRecordArgABI(RD) == CGCXXABI::RAA_DirectInMemory;
2424 static AggValueSlot createPlaceholderSlot(CodeGenFunction &CGF, QualType Ty) {
2425 // FIXME: Generate IR in one pass, rather than going back and fixing up these
2427 llvm::Type *IRTy = CGF.ConvertTypeForMem(Ty);
2428 llvm::Value *Placeholder =
2429 llvm::UndefValue::get(IRTy->getPointerTo()->getPointerTo());
2430 Placeholder = CGF.Builder.CreateLoad(Placeholder);
2431 return AggValueSlot::forAddr(Placeholder, CharUnits::Zero(),
2433 AggValueSlot::IsNotDestructed,
2434 AggValueSlot::DoesNotNeedGCBarriers,
2435 AggValueSlot::IsNotAliased);
2438 void CodeGenFunction::EmitDelegateCallArg(CallArgList &args,
2439 const VarDecl *param,
2440 SourceLocation loc) {
2441 // StartFunction converted the ABI-lowered parameter(s) into a
2442 // local alloca. We need to turn that into an r-value suitable
2444 llvm::Value *local = GetAddrOfLocalVar(param);
2446 QualType type = param->getType();
2448 // For the most part, we just need to load the alloca, except:
2449 // 1) aggregate r-values are actually pointers to temporaries, and
2450 // 2) references to non-scalars are pointers directly to the aggregate.
2451 // I don't know why references to scalars are different here.
2452 if (const ReferenceType *ref = type->getAs<ReferenceType>()) {
2453 if (!hasScalarEvaluationKind(ref->getPointeeType()))
2454 return args.add(RValue::getAggregate(local), type);
2456 // Locals which are references to scalars are represented
2457 // with allocas holding the pointer.
2458 return args.add(RValue::get(Builder.CreateLoad(local)), type);
2461 assert(!isInAllocaArgument(CGM.getCXXABI(), type) &&
2462 "cannot emit delegate call arguments for inalloca arguments!");
2464 args.add(convertTempToRValue(local, type, loc), type);
2467 static bool isProvablyNull(llvm::Value *addr) {
2468 return isa<llvm::ConstantPointerNull>(addr);
2471 static bool isProvablyNonNull(llvm::Value *addr) {
2472 return isa<llvm::AllocaInst>(addr);
2475 /// Emit the actual writing-back of a writeback.
2476 static void emitWriteback(CodeGenFunction &CGF,
2477 const CallArgList::Writeback &writeback) {
2478 const LValue &srcLV = writeback.Source;
2479 llvm::Value *srcAddr = srcLV.getAddress();
2480 assert(!isProvablyNull(srcAddr) &&
2481 "shouldn't have writeback for provably null argument");
2483 llvm::BasicBlock *contBB = nullptr;
2485 // If the argument wasn't provably non-null, we need to null check
2486 // before doing the store.
2487 bool provablyNonNull = isProvablyNonNull(srcAddr);
2488 if (!provablyNonNull) {
2489 llvm::BasicBlock *writebackBB = CGF.createBasicBlock("icr.writeback");
2490 contBB = CGF.createBasicBlock("icr.done");
2492 llvm::Value *isNull = CGF.Builder.CreateIsNull(srcAddr, "icr.isnull");
2493 CGF.Builder.CreateCondBr(isNull, contBB, writebackBB);
2494 CGF.EmitBlock(writebackBB);
2497 // Load the value to writeback.
2498 llvm::Value *value = CGF.Builder.CreateLoad(writeback.Temporary);
2500 // Cast it back, in case we're writing an id to a Foo* or something.
2501 value = CGF.Builder.CreateBitCast(value,
2502 cast<llvm::PointerType>(srcAddr->getType())->getElementType(),
2503 "icr.writeback-cast");
2505 // Perform the writeback.
2507 // If we have a "to use" value, it's something we need to emit a use
2508 // of. This has to be carefully threaded in: if it's done after the
2509 // release it's potentially undefined behavior (and the optimizer
2510 // will ignore it), and if it happens before the retain then the
2511 // optimizer could move the release there.
2512 if (writeback.ToUse) {
2513 assert(srcLV.getObjCLifetime() == Qualifiers::OCL_Strong);
2515 // Retain the new value. No need to block-copy here: the block's
2516 // being passed up the stack.
2517 value = CGF.EmitARCRetainNonBlock(value);
2519 // Emit the intrinsic use here.
2520 CGF.EmitARCIntrinsicUse(writeback.ToUse);
2522 // Load the old value (primitively).
2523 llvm::Value *oldValue = CGF.EmitLoadOfScalar(srcLV, SourceLocation());
2525 // Put the new value in place (primitively).
2526 CGF.EmitStoreOfScalar(value, srcLV, /*init*/ false);
2528 // Release the old value.
2529 CGF.EmitARCRelease(oldValue, srcLV.isARCPreciseLifetime());
2531 // Otherwise, we can just do a normal lvalue store.
2533 CGF.EmitStoreThroughLValue(RValue::get(value), srcLV);
2536 // Jump to the continuation block.
2537 if (!provablyNonNull)
2538 CGF.EmitBlock(contBB);
2541 static void emitWritebacks(CodeGenFunction &CGF,
2542 const CallArgList &args) {
2543 for (const auto &I : args.writebacks())
2544 emitWriteback(CGF, I);
2547 static void deactivateArgCleanupsBeforeCall(CodeGenFunction &CGF,
2548 const CallArgList &CallArgs) {
2549 assert(CGF.getTarget().getCXXABI().areArgsDestroyedLeftToRightInCallee());
2550 ArrayRef<CallArgList::CallArgCleanup> Cleanups =
2551 CallArgs.getCleanupsToDeactivate();
2552 // Iterate in reverse to increase the likelihood of popping the cleanup.
2553 for (ArrayRef<CallArgList::CallArgCleanup>::reverse_iterator
2554 I = Cleanups.rbegin(), E = Cleanups.rend(); I != E; ++I) {
2555 CGF.DeactivateCleanupBlock(I->Cleanup, I->IsActiveIP);
2556 I->IsActiveIP->eraseFromParent();
2560 static const Expr *maybeGetUnaryAddrOfOperand(const Expr *E) {
2561 if (const UnaryOperator *uop = dyn_cast<UnaryOperator>(E->IgnoreParens()))
2562 if (uop->getOpcode() == UO_AddrOf)
2563 return uop->getSubExpr();
2567 /// Emit an argument that's being passed call-by-writeback. That is,
2568 /// we are passing the address of
2569 static void emitWritebackArg(CodeGenFunction &CGF, CallArgList &args,
2570 const ObjCIndirectCopyRestoreExpr *CRE) {
2573 // Make an optimistic effort to emit the address as an l-value.
2574 // This can fail if the the argument expression is more complicated.
2575 if (const Expr *lvExpr = maybeGetUnaryAddrOfOperand(CRE->getSubExpr())) {
2576 srcLV = CGF.EmitLValue(lvExpr);
2578 // Otherwise, just emit it as a scalar.
2580 llvm::Value *srcAddr = CGF.EmitScalarExpr(CRE->getSubExpr());
2582 QualType srcAddrType =
2583 CRE->getSubExpr()->getType()->castAs<PointerType>()->getPointeeType();
2584 srcLV = CGF.MakeNaturalAlignAddrLValue(srcAddr, srcAddrType);
2586 llvm::Value *srcAddr = srcLV.getAddress();
2588 // The dest and src types don't necessarily match in LLVM terms
2589 // because of the crazy ObjC compatibility rules.
2591 llvm::PointerType *destType =
2592 cast<llvm::PointerType>(CGF.ConvertType(CRE->getType()));
2594 // If the address is a constant null, just pass the appropriate null.
2595 if (isProvablyNull(srcAddr)) {
2596 args.add(RValue::get(llvm::ConstantPointerNull::get(destType)),
2601 // Create the temporary.
2602 llvm::Value *temp = CGF.CreateTempAlloca(destType->getElementType(),
2604 // Loading an l-value can introduce a cleanup if the l-value is __weak,
2605 // and that cleanup will be conditional if we can't prove that the l-value
2606 // isn't null, so we need to register a dominating point so that the cleanups
2607 // system will make valid IR.
2608 CodeGenFunction::ConditionalEvaluation condEval(CGF);
2610 // Zero-initialize it if we're not doing a copy-initialization.
2611 bool shouldCopy = CRE->shouldCopy();
2614 llvm::ConstantPointerNull::get(
2615 cast<llvm::PointerType>(destType->getElementType()));
2616 CGF.Builder.CreateStore(null, temp);
2619 llvm::BasicBlock *contBB = nullptr;
2620 llvm::BasicBlock *originBB = nullptr;
2622 // If the address is *not* known to be non-null, we need to switch.
2623 llvm::Value *finalArgument;
2625 bool provablyNonNull = isProvablyNonNull(srcAddr);
2626 if (provablyNonNull) {
2627 finalArgument = temp;
2629 llvm::Value *isNull = CGF.Builder.CreateIsNull(srcAddr, "icr.isnull");
2631 finalArgument = CGF.Builder.CreateSelect(isNull,
2632 llvm::ConstantPointerNull::get(destType),
2633 temp, "icr.argument");
2635 // If we need to copy, then the load has to be conditional, which
2636 // means we need control flow.
2638 originBB = CGF.Builder.GetInsertBlock();
2639 contBB = CGF.createBasicBlock("icr.cont");
2640 llvm::BasicBlock *copyBB = CGF.createBasicBlock("icr.copy");
2641 CGF.Builder.CreateCondBr(isNull, contBB, copyBB);
2642 CGF.EmitBlock(copyBB);
2643 condEval.begin(CGF);
2647 llvm::Value *valueToUse = nullptr;
2649 // Perform a copy if necessary.
2651 RValue srcRV = CGF.EmitLoadOfLValue(srcLV, SourceLocation());
2652 assert(srcRV.isScalar());
2654 llvm::Value *src = srcRV.getScalarVal();
2655 src = CGF.Builder.CreateBitCast(src, destType->getElementType(),
2658 // Use an ordinary store, not a store-to-lvalue.
2659 CGF.Builder.CreateStore(src, temp);
2661 // If optimization is enabled, and the value was held in a
2662 // __strong variable, we need to tell the optimizer that this
2663 // value has to stay alive until we're doing the store back.
2664 // This is because the temporary is effectively unretained,
2665 // and so otherwise we can violate the high-level semantics.
2666 if (CGF.CGM.getCodeGenOpts().OptimizationLevel != 0 &&
2667 srcLV.getObjCLifetime() == Qualifiers::OCL_Strong) {
2672 // Finish the control flow if we needed it.
2673 if (shouldCopy && !provablyNonNull) {
2674 llvm::BasicBlock *copyBB = CGF.Builder.GetInsertBlock();
2675 CGF.EmitBlock(contBB);
2677 // Make a phi for the value to intrinsically use.
2679 llvm::PHINode *phiToUse = CGF.Builder.CreatePHI(valueToUse->getType(), 2,
2681 phiToUse->addIncoming(valueToUse, copyBB);
2682 phiToUse->addIncoming(llvm::UndefValue::get(valueToUse->getType()),
2684 valueToUse = phiToUse;
2690 args.addWriteback(srcLV, temp, valueToUse);
2691 args.add(RValue::get(finalArgument), CRE->getType());
2694 void CallArgList::allocateArgumentMemory(CodeGenFunction &CGF) {
2695 assert(!StackBase && !StackCleanup.isValid());
2698 llvm::Function *F = CGF.CGM.getIntrinsic(llvm::Intrinsic::stacksave);
2699 StackBase = CGF.Builder.CreateCall(F, {}, "inalloca.save");
2701 // Control gets really tied up in landing pads, so we have to spill the
2702 // stacksave to an alloca to avoid violating SSA form.
2703 // TODO: This is dead if we never emit the cleanup. We should create the
2704 // alloca and store lazily on the first cleanup emission.
2705 StackBaseMem = CGF.CreateTempAlloca(CGF.Int8PtrTy, "inalloca.spmem");
2706 CGF.Builder.CreateStore(StackBase, StackBaseMem);
2707 CGF.pushStackRestore(EHCleanup, StackBaseMem);
2708 StackCleanup = CGF.EHStack.getInnermostEHScope();
2709 assert(StackCleanup.isValid());
2712 void CallArgList::freeArgumentMemory(CodeGenFunction &CGF) const {
2714 CGF.DeactivateCleanupBlock(StackCleanup, StackBase);
2715 llvm::Value *F = CGF.CGM.getIntrinsic(llvm::Intrinsic::stackrestore);
2716 // We could load StackBase from StackBaseMem, but in the non-exceptional
2717 // case we can skip it.
2718 CGF.Builder.CreateCall(F, StackBase);
2722 static void emitNonNullArgCheck(CodeGenFunction &CGF, RValue RV,
2723 QualType ArgType, SourceLocation ArgLoc,
2724 const FunctionDecl *FD, unsigned ParmNum) {
2725 if (!CGF.SanOpts.has(SanitizerKind::NonnullAttribute) || !FD)
2727 auto PVD = ParmNum < FD->getNumParams() ? FD->getParamDecl(ParmNum) : nullptr;
2728 unsigned ArgNo = PVD ? PVD->getFunctionScopeIndex() : ParmNum;
2729 auto NNAttr = getNonNullAttr(FD, PVD, ArgType, ArgNo);
2732 CodeGenFunction::SanitizerScope SanScope(&CGF);
2733 assert(RV.isScalar());
2734 llvm::Value *V = RV.getScalarVal();
2736 CGF.Builder.CreateICmpNE(V, llvm::Constant::getNullValue(V->getType()));
2737 llvm::Constant *StaticData[] = {
2738 CGF.EmitCheckSourceLocation(ArgLoc),
2739 CGF.EmitCheckSourceLocation(NNAttr->getLocation()),
2740 llvm::ConstantInt::get(CGF.Int32Ty, ArgNo + 1),
2742 CGF.EmitCheck(std::make_pair(Cond, SanitizerKind::NonnullAttribute),
2743 "nonnull_arg", StaticData, None);
2746 void CodeGenFunction::EmitCallArgs(CallArgList &Args,
2747 ArrayRef<QualType> ArgTypes,
2748 CallExpr::const_arg_iterator ArgBeg,
2749 CallExpr::const_arg_iterator ArgEnd,
2750 const FunctionDecl *CalleeDecl,
2751 unsigned ParamsToSkip) {
2752 // We *have* to evaluate arguments from right to left in the MS C++ ABI,
2753 // because arguments are destroyed left to right in the callee.
2754 if (CGM.getTarget().getCXXABI().areArgsDestroyedLeftToRightInCallee()) {
2755 // Insert a stack save if we're going to need any inalloca args.
2756 bool HasInAllocaArgs = false;
2757 for (ArrayRef<QualType>::iterator I = ArgTypes.begin(), E = ArgTypes.end();
2758 I != E && !HasInAllocaArgs; ++I)
2759 HasInAllocaArgs = isInAllocaArgument(CGM.getCXXABI(), *I);
2760 if (HasInAllocaArgs) {
2761 assert(getTarget().getTriple().getArch() == llvm::Triple::x86);
2762 Args.allocateArgumentMemory(*this);
2765 // Evaluate each argument.
2766 size_t CallArgsStart = Args.size();
2767 for (int I = ArgTypes.size() - 1; I >= 0; --I) {
2768 CallExpr::const_arg_iterator Arg = ArgBeg + I;
2769 EmitCallArg(Args, *Arg, ArgTypes[I]);
2770 emitNonNullArgCheck(*this, Args.back().RV, ArgTypes[I], Arg->getExprLoc(),
2771 CalleeDecl, ParamsToSkip + I);
2774 // Un-reverse the arguments we just evaluated so they match up with the LLVM
2776 std::reverse(Args.begin() + CallArgsStart, Args.end());
2780 for (unsigned I = 0, E = ArgTypes.size(); I != E; ++I) {
2781 CallExpr::const_arg_iterator Arg = ArgBeg + I;
2782 assert(Arg != ArgEnd);
2783 EmitCallArg(Args, *Arg, ArgTypes[I]);
2784 emitNonNullArgCheck(*this, Args.back().RV, ArgTypes[I], Arg->getExprLoc(),
2785 CalleeDecl, ParamsToSkip + I);
2791 struct DestroyUnpassedArg : EHScopeStack::Cleanup {
2792 DestroyUnpassedArg(llvm::Value *Addr, QualType Ty)
2793 : Addr(Addr), Ty(Ty) {}
2798 void Emit(CodeGenFunction &CGF, Flags flags) override {
2799 const CXXDestructorDecl *Dtor = Ty->getAsCXXRecordDecl()->getDestructor();
2800 assert(!Dtor->isTrivial());
2801 CGF.EmitCXXDestructorCall(Dtor, Dtor_Complete, /*for vbase*/ false,
2802 /*Delegating=*/false, Addr);
2808 struct DisableDebugLocationUpdates {
2809 CodeGenFunction &CGF;
2810 bool disabledDebugInfo;
2811 DisableDebugLocationUpdates(CodeGenFunction &CGF, const Expr *E) : CGF(CGF) {
2812 if ((disabledDebugInfo = isa<CXXDefaultArgExpr>(E) && CGF.getDebugInfo()))
2813 CGF.disableDebugInfo();
2815 ~DisableDebugLocationUpdates() {
2816 if (disabledDebugInfo)
2817 CGF.enableDebugInfo();
2821 void CodeGenFunction::EmitCallArg(CallArgList &args, const Expr *E,
2823 DisableDebugLocationUpdates Dis(*this, E);
2824 if (const ObjCIndirectCopyRestoreExpr *CRE
2825 = dyn_cast<ObjCIndirectCopyRestoreExpr>(E)) {
2826 assert(getLangOpts().ObjCAutoRefCount);
2827 assert(getContext().hasSameType(E->getType(), type));
2828 return emitWritebackArg(*this, args, CRE);
2831 assert(type->isReferenceType() == E->isGLValue() &&
2832 "reference binding to unmaterialized r-value!");
2834 if (E->isGLValue()) {
2835 assert(E->getObjectKind() == OK_Ordinary);
2836 return args.add(EmitReferenceBindingToExpr(E), type);
2839 bool HasAggregateEvalKind = hasAggregateEvaluationKind(type);
2841 // In the Microsoft C++ ABI, aggregate arguments are destructed by the callee.
2842 // However, we still have to push an EH-only cleanup in case we unwind before
2843 // we make it to the call.
2844 if (HasAggregateEvalKind &&
2845 CGM.getTarget().getCXXABI().areArgsDestroyedLeftToRightInCallee()) {
2846 // If we're using inalloca, use the argument memory. Otherwise, use a
2849 if (args.isUsingInAlloca())
2850 Slot = createPlaceholderSlot(*this, type);
2852 Slot = CreateAggTemp(type, "agg.tmp");
2854 const CXXRecordDecl *RD = type->getAsCXXRecordDecl();
2855 bool DestroyedInCallee =
2856 RD && RD->hasNonTrivialDestructor() &&
2857 CGM.getCXXABI().getRecordArgABI(RD) != CGCXXABI::RAA_Default;
2858 if (DestroyedInCallee)
2859 Slot.setExternallyDestructed();
2861 EmitAggExpr(E, Slot);
2862 RValue RV = Slot.asRValue();
2865 if (DestroyedInCallee) {
2866 // Create a no-op GEP between the placeholder and the cleanup so we can
2867 // RAUW it successfully. It also serves as a marker of the first
2868 // instruction where the cleanup is active.
2869 pushFullExprCleanup<DestroyUnpassedArg>(EHCleanup, Slot.getAddr(), type);
2870 // This unreachable is a temporary marker which will be removed later.
2871 llvm::Instruction *IsActive = Builder.CreateUnreachable();
2872 args.addArgCleanupDeactivation(EHStack.getInnermostEHScope(), IsActive);
2877 if (HasAggregateEvalKind && isa<ImplicitCastExpr>(E) &&
2878 cast<CastExpr>(E)->getCastKind() == CK_LValueToRValue) {
2879 LValue L = EmitLValue(cast<CastExpr>(E)->getSubExpr());
2880 assert(L.isSimple());
2881 if (L.getAlignment() >= getContext().getTypeAlignInChars(type)) {
2882 args.add(L.asAggregateRValue(), type, /*NeedsCopy*/true);
2884 // We can't represent a misaligned lvalue in the CallArgList, so copy
2885 // to an aligned temporary now.
2886 llvm::Value *tmp = CreateMemTemp(type);
2887 EmitAggregateCopy(tmp, L.getAddress(), type, L.isVolatile(),
2889 args.add(RValue::getAggregate(tmp), type);
2894 args.add(EmitAnyExprToTemp(E), type);
2897 QualType CodeGenFunction::getVarArgType(const Expr *Arg) {
2898 // System headers on Windows define NULL to 0 instead of 0LL on Win64. MSVC
2899 // implicitly widens null pointer constants that are arguments to varargs
2900 // functions to pointer-sized ints.
2901 if (!getTarget().getTriple().isOSWindows())
2902 return Arg->getType();
2904 if (Arg->getType()->isIntegerType() &&
2905 getContext().getTypeSize(Arg->getType()) <
2906 getContext().getTargetInfo().getPointerWidth(0) &&
2907 Arg->isNullPointerConstant(getContext(),
2908 Expr::NPC_ValueDependentIsNotNull)) {
2909 return getContext().getIntPtrType();
2912 return Arg->getType();
2915 // In ObjC ARC mode with no ObjC ARC exception safety, tell the ARC
2916 // optimizer it can aggressively ignore unwind edges.
2918 CodeGenFunction::AddObjCARCExceptionMetadata(llvm::Instruction *Inst) {
2919 if (CGM.getCodeGenOpts().OptimizationLevel != 0 &&
2920 !CGM.getCodeGenOpts().ObjCAutoRefCountExceptions)
2921 Inst->setMetadata("clang.arc.no_objc_arc_exceptions",
2922 CGM.getNoObjCARCExceptionsMetadata());
2925 /// Emits a call to the given no-arguments nounwind runtime function.
2927 CodeGenFunction::EmitNounwindRuntimeCall(llvm::Value *callee,
2928 const llvm::Twine &name) {
2929 return EmitNounwindRuntimeCall(callee, None, name);
2932 /// Emits a call to the given nounwind runtime function.
2934 CodeGenFunction::EmitNounwindRuntimeCall(llvm::Value *callee,
2935 ArrayRef<llvm::Value*> args,
2936 const llvm::Twine &name) {
2937 llvm::CallInst *call = EmitRuntimeCall(callee, args, name);
2938 call->setDoesNotThrow();
2942 /// Emits a simple call (never an invoke) to the given no-arguments
2943 /// runtime function.
2945 CodeGenFunction::EmitRuntimeCall(llvm::Value *callee,
2946 const llvm::Twine &name) {
2947 return EmitRuntimeCall(callee, None, name);
2950 /// Emits a simple call (never an invoke) to the given runtime
2953 CodeGenFunction::EmitRuntimeCall(llvm::Value *callee,
2954 ArrayRef<llvm::Value*> args,
2955 const llvm::Twine &name) {
2956 llvm::CallInst *call = Builder.CreateCall(callee, args, name);
2957 call->setCallingConv(getRuntimeCC());
2961 /// Emits a call or invoke to the given noreturn runtime function.
2962 void CodeGenFunction::EmitNoreturnRuntimeCallOrInvoke(llvm::Value *callee,
2963 ArrayRef<llvm::Value*> args) {
2964 if (getInvokeDest()) {
2965 llvm::InvokeInst *invoke =
2966 Builder.CreateInvoke(callee,
2967 getUnreachableBlock(),
2970 invoke->setDoesNotReturn();
2971 invoke->setCallingConv(getRuntimeCC());
2973 llvm::CallInst *call = Builder.CreateCall(callee, args);
2974 call->setDoesNotReturn();
2975 call->setCallingConv(getRuntimeCC());
2976 Builder.CreateUnreachable();
2980 /// Emits a call or invoke instruction to the given nullary runtime
2983 CodeGenFunction::EmitRuntimeCallOrInvoke(llvm::Value *callee,
2984 const Twine &name) {
2985 return EmitRuntimeCallOrInvoke(callee, None, name);
2988 /// Emits a call or invoke instruction to the given runtime function.
2990 CodeGenFunction::EmitRuntimeCallOrInvoke(llvm::Value *callee,
2991 ArrayRef<llvm::Value*> args,
2992 const Twine &name) {
2993 llvm::CallSite callSite = EmitCallOrInvoke(callee, args, name);
2994 callSite.setCallingConv(getRuntimeCC());
2999 CodeGenFunction::EmitCallOrInvoke(llvm::Value *Callee,
3000 const Twine &Name) {
3001 return EmitCallOrInvoke(Callee, None, Name);
3004 /// Emits a call or invoke instruction to the given function, depending
3005 /// on the current state of the EH stack.
3007 CodeGenFunction::EmitCallOrInvoke(llvm::Value *Callee,
3008 ArrayRef<llvm::Value *> Args,
3009 const Twine &Name) {
3010 llvm::BasicBlock *InvokeDest = getInvokeDest();
3012 llvm::Instruction *Inst;
3014 Inst = Builder.CreateCall(Callee, Args, Name);
3016 llvm::BasicBlock *ContBB = createBasicBlock("invoke.cont");
3017 Inst = Builder.CreateInvoke(Callee, ContBB, InvokeDest, Args, Name);
3021 // In ObjC ARC mode with no ObjC ARC exception safety, tell the ARC
3022 // optimizer it can aggressively ignore unwind edges.
3023 if (CGM.getLangOpts().ObjCAutoRefCount)
3024 AddObjCARCExceptionMetadata(Inst);
3026 return llvm::CallSite(Inst);
3029 /// \brief Store a non-aggregate value to an address to initialize it. For
3030 /// initialization, a non-atomic store will be used.
3031 static void EmitInitStoreOfNonAggregate(CodeGenFunction &CGF, RValue Src,
3034 CGF.EmitStoreOfScalar(Src.getScalarVal(), Dst, /*init=*/true);
3036 CGF.EmitStoreOfComplex(Src.getComplexVal(), Dst, /*init=*/true);
3039 void CodeGenFunction::deferPlaceholderReplacement(llvm::Instruction *Old,
3041 DeferredReplacements.push_back(std::make_pair(Old, New));
3044 RValue CodeGenFunction::EmitCall(const CGFunctionInfo &CallInfo,
3045 llvm::Value *Callee,
3046 ReturnValueSlot ReturnValue,
3047 const CallArgList &CallArgs,
3048 const Decl *TargetDecl,
3049 llvm::Instruction **callOrInvoke) {
3050 // FIXME: We no longer need the types from CallArgs; lift up and simplify.
3052 // Handle struct-return functions by passing a pointer to the
3053 // location that we would like to return into.
3054 QualType RetTy = CallInfo.getReturnType();
3055 const ABIArgInfo &RetAI = CallInfo.getReturnInfo();
3057 llvm::FunctionType *IRFuncTy =
3058 cast<llvm::FunctionType>(
3059 cast<llvm::PointerType>(Callee->getType())->getElementType());
3061 // If we're using inalloca, insert the allocation after the stack save.
3062 // FIXME: Do this earlier rather than hacking it in here!
3063 llvm::AllocaInst *ArgMemory = nullptr;
3064 if (llvm::StructType *ArgStruct = CallInfo.getArgStruct()) {
3065 llvm::Instruction *IP = CallArgs.getStackBase();
3066 llvm::AllocaInst *AI;
3068 IP = IP->getNextNode();
3069 AI = new llvm::AllocaInst(ArgStruct, "argmem", IP);
3071 AI = CreateTempAlloca(ArgStruct, "argmem");
3073 AI->setUsedWithInAlloca(true);
3074 assert(AI->isUsedWithInAlloca() && !AI->isStaticAlloca());
3078 ClangToLLVMArgMapping IRFunctionArgs(CGM.getContext(), CallInfo);
3079 SmallVector<llvm::Value *, 16> IRCallArgs(IRFunctionArgs.totalIRArgs());
3081 // If the call returns a temporary with struct return, create a temporary
3082 // alloca to hold the result, unless one is given to us.
3083 llvm::Value *SRetPtr = nullptr;
3084 if (RetAI.isIndirect() || RetAI.isInAlloca()) {
3085 SRetPtr = ReturnValue.getValue();
3087 SRetPtr = CreateMemTemp(RetTy);
3088 if (IRFunctionArgs.hasSRetArg()) {
3089 IRCallArgs[IRFunctionArgs.getSRetArgNo()] = SRetPtr;
3092 Builder.CreateStructGEP(ArgMemory->getAllocatedType(), ArgMemory,
3093 RetAI.getInAllocaFieldIndex());
3094 Builder.CreateStore(SRetPtr, Addr);
3098 assert(CallInfo.arg_size() == CallArgs.size() &&
3099 "Mismatch between function signature & arguments.");
3101 CGFunctionInfo::const_arg_iterator info_it = CallInfo.arg_begin();
3102 for (CallArgList::const_iterator I = CallArgs.begin(), E = CallArgs.end();
3103 I != E; ++I, ++info_it, ++ArgNo) {
3104 const ABIArgInfo &ArgInfo = info_it->info;
3107 CharUnits TypeAlign = getContext().getTypeAlignInChars(I->Ty);
3109 // Insert a padding argument to ensure proper alignment.
3110 if (IRFunctionArgs.hasPaddingArg(ArgNo))
3111 IRCallArgs[IRFunctionArgs.getPaddingArgNo(ArgNo)] =
3112 llvm::UndefValue::get(ArgInfo.getPaddingType());
3114 unsigned FirstIRArg, NumIRArgs;
3115 std::tie(FirstIRArg, NumIRArgs) = IRFunctionArgs.getIRArgs(ArgNo);
3117 switch (ArgInfo.getKind()) {
3118 case ABIArgInfo::InAlloca: {
3119 assert(NumIRArgs == 0);
3120 assert(getTarget().getTriple().getArch() == llvm::Triple::x86);
3121 if (RV.isAggregate()) {
3122 // Replace the placeholder with the appropriate argument slot GEP.
3123 llvm::Instruction *Placeholder =
3124 cast<llvm::Instruction>(RV.getAggregateAddr());
3125 CGBuilderTy::InsertPoint IP = Builder.saveIP();
3126 Builder.SetInsertPoint(Placeholder);
3128 Builder.CreateStructGEP(ArgMemory->getAllocatedType(), ArgMemory,
3129 ArgInfo.getInAllocaFieldIndex());
3130 Builder.restoreIP(IP);
3131 deferPlaceholderReplacement(Placeholder, Addr);
3133 // Store the RValue into the argument struct.
3135 Builder.CreateStructGEP(ArgMemory->getAllocatedType(), ArgMemory,
3136 ArgInfo.getInAllocaFieldIndex());
3137 unsigned AS = Addr->getType()->getPointerAddressSpace();
3138 llvm::Type *MemType = ConvertTypeForMem(I->Ty)->getPointerTo(AS);
3139 // There are some cases where a trivial bitcast is not avoidable. The
3140 // definition of a type later in a translation unit may change it's type
3141 // from {}* to (%struct.foo*)*.
3142 if (Addr->getType() != MemType)
3143 Addr = Builder.CreateBitCast(Addr, MemType);
3144 LValue argLV = MakeAddrLValue(Addr, I->Ty, TypeAlign);
3145 EmitInitStoreOfNonAggregate(*this, RV, argLV);
3150 case ABIArgInfo::Indirect: {
3151 assert(NumIRArgs == 1);
3152 if (RV.isScalar() || RV.isComplex()) {
3153 // Make a temporary alloca to pass the argument.
3154 llvm::AllocaInst *AI = CreateMemTemp(I->Ty);
3155 if (ArgInfo.getIndirectAlign() > AI->getAlignment())
3156 AI->setAlignment(ArgInfo.getIndirectAlign());
3157 IRCallArgs[FirstIRArg] = AI;
3159 LValue argLV = MakeAddrLValue(AI, I->Ty, TypeAlign);
3160 EmitInitStoreOfNonAggregate(*this, RV, argLV);
3162 // We want to avoid creating an unnecessary temporary+copy here;
3163 // however, we need one in three cases:
3164 // 1. If the argument is not byval, and we are required to copy the
3165 // source. (This case doesn't occur on any common architecture.)
3166 // 2. If the argument is byval, RV is not sufficiently aligned, and
3167 // we cannot force it to be sufficiently aligned.
3168 // 3. If the argument is byval, but RV is located in an address space
3169 // different than that of the argument (0).
3170 llvm::Value *Addr = RV.getAggregateAddr();
3171 unsigned Align = ArgInfo.getIndirectAlign();
3172 const llvm::DataLayout *TD = &CGM.getDataLayout();
3173 const unsigned RVAddrSpace = Addr->getType()->getPointerAddressSpace();
3174 const unsigned ArgAddrSpace =
3175 (FirstIRArg < IRFuncTy->getNumParams()
3176 ? IRFuncTy->getParamType(FirstIRArg)->getPointerAddressSpace()
3178 if ((!ArgInfo.getIndirectByVal() && I->NeedsCopy) ||
3179 (ArgInfo.getIndirectByVal() && TypeAlign.getQuantity() < Align &&
3180 llvm::getOrEnforceKnownAlignment(Addr, Align, *TD) < Align) ||
3181 (ArgInfo.getIndirectByVal() && (RVAddrSpace != ArgAddrSpace))) {
3182 // Create an aligned temporary, and copy to it.
3183 llvm::AllocaInst *AI = CreateMemTemp(I->Ty);
3184 if (Align > AI->getAlignment())
3185 AI->setAlignment(Align);
3186 IRCallArgs[FirstIRArg] = AI;
3187 EmitAggregateCopy(AI, Addr, I->Ty, RV.isVolatileQualified());
3189 // Skip the extra memcpy call.
3190 IRCallArgs[FirstIRArg] = Addr;
3196 case ABIArgInfo::Ignore:
3197 assert(NumIRArgs == 0);
3200 case ABIArgInfo::Extend:
3201 case ABIArgInfo::Direct: {
3202 if (!isa<llvm::StructType>(ArgInfo.getCoerceToType()) &&
3203 ArgInfo.getCoerceToType() == ConvertType(info_it->type) &&
3204 ArgInfo.getDirectOffset() == 0) {
3205 assert(NumIRArgs == 1);
3208 V = RV.getScalarVal();
3210 V = Builder.CreateLoad(RV.getAggregateAddr());
3212 // We might have to widen integers, but we should never truncate.
3213 if (ArgInfo.getCoerceToType() != V->getType() &&
3214 V->getType()->isIntegerTy())
3215 V = Builder.CreateZExt(V, ArgInfo.getCoerceToType());
3217 // If the argument doesn't match, perform a bitcast to coerce it. This
3218 // can happen due to trivial type mismatches.
3219 if (FirstIRArg < IRFuncTy->getNumParams() &&
3220 V->getType() != IRFuncTy->getParamType(FirstIRArg))
3221 V = Builder.CreateBitCast(V, IRFuncTy->getParamType(FirstIRArg));
3222 IRCallArgs[FirstIRArg] = V;
3226 // FIXME: Avoid the conversion through memory if possible.
3227 llvm::Value *SrcPtr;
3228 if (RV.isScalar() || RV.isComplex()) {
3229 SrcPtr = CreateMemTemp(I->Ty, "coerce");
3230 LValue SrcLV = MakeAddrLValue(SrcPtr, I->Ty, TypeAlign);
3231 EmitInitStoreOfNonAggregate(*this, RV, SrcLV);
3233 SrcPtr = RV.getAggregateAddr();
3235 // If the value is offset in memory, apply the offset now.
3236 if (unsigned Offs = ArgInfo.getDirectOffset()) {
3237 SrcPtr = Builder.CreateBitCast(SrcPtr, Builder.getInt8PtrTy());
3238 SrcPtr = Builder.CreateConstGEP1_32(Builder.getInt8Ty(), SrcPtr, Offs);
3239 SrcPtr = Builder.CreateBitCast(SrcPtr,
3240 llvm::PointerType::getUnqual(ArgInfo.getCoerceToType()));
3244 // Fast-isel and the optimizer generally like scalar values better than
3245 // FCAs, so we flatten them if this is safe to do for this argument.
3246 llvm::StructType *STy =
3247 dyn_cast<llvm::StructType>(ArgInfo.getCoerceToType());
3248 if (STy && ArgInfo.isDirect() && ArgInfo.getCanBeFlattened()) {
3250 cast<llvm::PointerType>(SrcPtr->getType())->getElementType();
3251 uint64_t SrcSize = CGM.getDataLayout().getTypeAllocSize(SrcTy);
3252 uint64_t DstSize = CGM.getDataLayout().getTypeAllocSize(STy);
3254 // If the source type is smaller than the destination type of the
3255 // coerce-to logic, copy the source value into a temp alloca the size
3256 // of the destination type to allow loading all of it. The bits past
3257 // the source value are left undef.
3258 if (SrcSize < DstSize) {
3259 llvm::AllocaInst *TempAlloca
3260 = CreateTempAlloca(STy, SrcPtr->getName() + ".coerce");
3261 Builder.CreateMemCpy(TempAlloca, SrcPtr, SrcSize, 0);
3262 SrcPtr = TempAlloca;
3264 SrcPtr = Builder.CreateBitCast(SrcPtr,
3265 llvm::PointerType::getUnqual(STy));
3268 assert(NumIRArgs == STy->getNumElements());
3269 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
3270 llvm::Value *EltPtr = Builder.CreateConstGEP2_32(STy, SrcPtr, 0, i);
3271 llvm::LoadInst *LI = Builder.CreateLoad(EltPtr);
3272 // We don't know what we're loading from.
3273 LI->setAlignment(1);
3274 IRCallArgs[FirstIRArg + i] = LI;
3277 // In the simple case, just pass the coerced loaded value.
3278 assert(NumIRArgs == 1);
3279 IRCallArgs[FirstIRArg] =
3280 CreateCoercedLoad(SrcPtr, ArgInfo.getCoerceToType(), *this);
3286 case ABIArgInfo::Expand:
3287 unsigned IRArgPos = FirstIRArg;
3288 ExpandTypeToArgs(I->Ty, RV, IRFuncTy, IRCallArgs, IRArgPos);
3289 assert(IRArgPos == FirstIRArg + NumIRArgs);
3295 llvm::Value *Arg = ArgMemory;
3296 if (CallInfo.isVariadic()) {
3297 // When passing non-POD arguments by value to variadic functions, we will
3298 // end up with a variadic prototype and an inalloca call site. In such
3299 // cases, we can't do any parameter mismatch checks. Give up and bitcast
3302 cast<llvm::PointerType>(Callee->getType())->getAddressSpace();
3303 Callee = Builder.CreateBitCast(
3304 Callee, getTypes().GetFunctionType(CallInfo)->getPointerTo(CalleeAS));
3306 llvm::Type *LastParamTy =
3307 IRFuncTy->getParamType(IRFuncTy->getNumParams() - 1);
3308 if (Arg->getType() != LastParamTy) {
3310 // Assert that these structs have equivalent element types.
3311 llvm::StructType *FullTy = CallInfo.getArgStruct();
3312 llvm::StructType *DeclaredTy = cast<llvm::StructType>(
3313 cast<llvm::PointerType>(LastParamTy)->getElementType());
3314 assert(DeclaredTy->getNumElements() == FullTy->getNumElements());
3315 for (llvm::StructType::element_iterator DI = DeclaredTy->element_begin(),
3316 DE = DeclaredTy->element_end(),
3317 FI = FullTy->element_begin();
3318 DI != DE; ++DI, ++FI)
3321 Arg = Builder.CreateBitCast(Arg, LastParamTy);
3324 assert(IRFunctionArgs.hasInallocaArg());
3325 IRCallArgs[IRFunctionArgs.getInallocaArgNo()] = Arg;
3328 if (!CallArgs.getCleanupsToDeactivate().empty())
3329 deactivateArgCleanupsBeforeCall(*this, CallArgs);
3331 // If the callee is a bitcast of a function to a varargs pointer to function
3332 // type, check to see if we can remove the bitcast. This handles some cases
3333 // with unprototyped functions.
3334 if (llvm::ConstantExpr *CE = dyn_cast<llvm::ConstantExpr>(Callee))
3335 if (llvm::Function *CalleeF = dyn_cast<llvm::Function>(CE->getOperand(0))) {
3336 llvm::PointerType *CurPT=cast<llvm::PointerType>(Callee->getType());
3337 llvm::FunctionType *CurFT =
3338 cast<llvm::FunctionType>(CurPT->getElementType());
3339 llvm::FunctionType *ActualFT = CalleeF->getFunctionType();
3341 if (CE->getOpcode() == llvm::Instruction::BitCast &&
3342 ActualFT->getReturnType() == CurFT->getReturnType() &&
3343 ActualFT->getNumParams() == CurFT->getNumParams() &&
3344 ActualFT->getNumParams() == IRCallArgs.size() &&
3345 (CurFT->isVarArg() || !ActualFT->isVarArg())) {
3346 bool ArgsMatch = true;
3347 for (unsigned i = 0, e = ActualFT->getNumParams(); i != e; ++i)
3348 if (ActualFT->getParamType(i) != CurFT->getParamType(i)) {
3353 // Strip the cast if we can get away with it. This is a nice cleanup,
3354 // but also allows us to inline the function at -O0 if it is marked
3361 assert(IRCallArgs.size() == IRFuncTy->getNumParams() || IRFuncTy->isVarArg());
3362 for (unsigned i = 0; i < IRCallArgs.size(); ++i) {
3363 // Inalloca argument can have different type.
3364 if (IRFunctionArgs.hasInallocaArg() &&
3365 i == IRFunctionArgs.getInallocaArgNo())
3367 if (i < IRFuncTy->getNumParams())
3368 assert(IRCallArgs[i]->getType() == IRFuncTy->getParamType(i));
3371 unsigned CallingConv;
3372 CodeGen::AttributeListType AttributeList;
3373 CGM.ConstructAttributeList(CallInfo, TargetDecl, AttributeList,
3375 llvm::AttributeSet Attrs = llvm::AttributeSet::get(getLLVMContext(),
3378 llvm::BasicBlock *InvokeDest = nullptr;
3379 if (!Attrs.hasAttribute(llvm::AttributeSet::FunctionIndex,
3380 llvm::Attribute::NoUnwind) ||
3381 currentFunctionUsesSEHTry())
3382 InvokeDest = getInvokeDest();
3386 CS = Builder.CreateCall(Callee, IRCallArgs);
3388 llvm::BasicBlock *Cont = createBasicBlock("invoke.cont");
3389 CS = Builder.CreateInvoke(Callee, Cont, InvokeDest, IRCallArgs);
3393 *callOrInvoke = CS.getInstruction();
3395 if (CurCodeDecl && CurCodeDecl->hasAttr<FlattenAttr>() &&
3396 !CS.hasFnAttr(llvm::Attribute::NoInline))
3398 Attrs.addAttribute(getLLVMContext(), llvm::AttributeSet::FunctionIndex,
3399 llvm::Attribute::AlwaysInline);
3401 // Disable inlining inside SEH __try blocks.
3402 if (isSEHTryScope())
3404 Attrs.addAttribute(getLLVMContext(), llvm::AttributeSet::FunctionIndex,
3405 llvm::Attribute::NoInline);
3407 CS.setAttributes(Attrs);
3408 CS.setCallingConv(static_cast<llvm::CallingConv::ID>(CallingConv));
3410 // In ObjC ARC mode with no ObjC ARC exception safety, tell the ARC
3411 // optimizer it can aggressively ignore unwind edges.
3412 if (CGM.getLangOpts().ObjCAutoRefCount)
3413 AddObjCARCExceptionMetadata(CS.getInstruction());
3415 // If the call doesn't return, finish the basic block and clear the
3416 // insertion point; this allows the rest of IRgen to discard
3417 // unreachable code.
3418 if (CS.doesNotReturn()) {
3419 Builder.CreateUnreachable();
3420 Builder.ClearInsertionPoint();
3422 // FIXME: For now, emit a dummy basic block because expr emitters in
3423 // generally are not ready to handle emitting expressions at unreachable
3425 EnsureInsertPoint();
3427 // Return a reasonable RValue.
3428 return GetUndefRValue(RetTy);
3431 llvm::Instruction *CI = CS.getInstruction();
3432 if (Builder.isNamePreserving() && !CI->getType()->isVoidTy())
3433 CI->setName("call");
3435 // Emit any writebacks immediately. Arguably this should happen
3436 // after any return-value munging.
3437 if (CallArgs.hasWritebacks())
3438 emitWritebacks(*this, CallArgs);
3440 // The stack cleanup for inalloca arguments has to run out of the normal
3441 // lexical order, so deactivate it and run it manually here.
3442 CallArgs.freeArgumentMemory(*this);
3445 switch (RetAI.getKind()) {
3446 case ABIArgInfo::InAlloca:
3447 case ABIArgInfo::Indirect:
3448 return convertTempToRValue(SRetPtr, RetTy, SourceLocation());
3450 case ABIArgInfo::Ignore:
3451 // If we are ignoring an argument that had a result, make sure to
3452 // construct the appropriate return value for our caller.
3453 return GetUndefRValue(RetTy);
3455 case ABIArgInfo::Extend:
3456 case ABIArgInfo::Direct: {
3457 llvm::Type *RetIRTy = ConvertType(RetTy);
3458 if (RetAI.getCoerceToType() == RetIRTy && RetAI.getDirectOffset() == 0) {
3459 switch (getEvaluationKind(RetTy)) {
3461 llvm::Value *Real = Builder.CreateExtractValue(CI, 0);
3462 llvm::Value *Imag = Builder.CreateExtractValue(CI, 1);
3463 return RValue::getComplex(std::make_pair(Real, Imag));
3465 case TEK_Aggregate: {
3466 llvm::Value *DestPtr = ReturnValue.getValue();
3467 bool DestIsVolatile = ReturnValue.isVolatile();
3470 DestPtr = CreateMemTemp(RetTy, "agg.tmp");
3471 DestIsVolatile = false;
3473 BuildAggStore(*this, CI, DestPtr, DestIsVolatile, false);
3474 return RValue::getAggregate(DestPtr);
3477 // If the argument doesn't match, perform a bitcast to coerce it. This
3478 // can happen due to trivial type mismatches.
3479 llvm::Value *V = CI;
3480 if (V->getType() != RetIRTy)
3481 V = Builder.CreateBitCast(V, RetIRTy);
3482 return RValue::get(V);
3485 llvm_unreachable("bad evaluation kind");
3488 llvm::Value *DestPtr = ReturnValue.getValue();
3489 bool DestIsVolatile = ReturnValue.isVolatile();
3492 DestPtr = CreateMemTemp(RetTy, "coerce");
3493 DestIsVolatile = false;
3496 // If the value is offset in memory, apply the offset now.
3497 llvm::Value *StorePtr = DestPtr;
3498 if (unsigned Offs = RetAI.getDirectOffset()) {
3499 StorePtr = Builder.CreateBitCast(StorePtr, Builder.getInt8PtrTy());
3501 Builder.CreateConstGEP1_32(Builder.getInt8Ty(), StorePtr, Offs);
3502 StorePtr = Builder.CreateBitCast(StorePtr,
3503 llvm::PointerType::getUnqual(RetAI.getCoerceToType()));
3505 CreateCoercedStore(CI, StorePtr, DestIsVolatile, *this);
3507 return convertTempToRValue(DestPtr, RetTy, SourceLocation());
3510 case ABIArgInfo::Expand:
3511 llvm_unreachable("Invalid ABI kind for return argument");
3514 llvm_unreachable("Unhandled ABIArgInfo::Kind");
3517 if (Ret.isScalar() && TargetDecl) {
3518 if (const auto *AA = TargetDecl->getAttr<AssumeAlignedAttr>()) {
3519 llvm::Value *OffsetValue = nullptr;
3520 if (const auto *Offset = AA->getOffset())
3521 OffsetValue = EmitScalarExpr(Offset);
3523 llvm::Value *Alignment = EmitScalarExpr(AA->getAlignment());
3524 llvm::ConstantInt *AlignmentCI = cast<llvm::ConstantInt>(Alignment);
3525 EmitAlignmentAssumption(Ret.getScalarVal(), AlignmentCI->getZExtValue(),
3533 /* VarArg handling */
3535 llvm::Value *CodeGenFunction::EmitVAArg(llvm::Value *VAListAddr, QualType Ty) {
3536 return CGM.getTypes().getABIInfo().EmitVAArg(VAListAddr, Ty, *this);