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 //===----------------------------------------------------------------------===//
19 #include "CGCleanup.h"
20 #include "CodeGenFunction.h"
21 #include "CodeGenModule.h"
22 #include "TargetInfo.h"
23 #include "clang/AST/Decl.h"
24 #include "clang/AST/DeclCXX.h"
25 #include "clang/AST/DeclObjC.h"
26 #include "clang/Basic/TargetBuiltins.h"
27 #include "clang/Basic/TargetInfo.h"
28 #include "clang/CodeGen/CGFunctionInfo.h"
29 #include "clang/CodeGen/SwiftCallingConv.h"
30 #include "clang/Frontend/CodeGenOptions.h"
31 #include "llvm/ADT/StringExtras.h"
32 #include "llvm/IR/Attributes.h"
33 #include "llvm/IR/CallingConv.h"
34 #include "llvm/IR/CallSite.h"
35 #include "llvm/IR/DataLayout.h"
36 #include "llvm/IR/InlineAsm.h"
37 #include "llvm/IR/Intrinsics.h"
38 #include "llvm/IR/IntrinsicInst.h"
39 #include "llvm/Transforms/Utils/Local.h"
40 using namespace clang;
41 using namespace CodeGen;
45 unsigned CodeGenTypes::ClangCallConvToLLVMCallConv(CallingConv CC) {
47 default: return llvm::CallingConv::C;
48 case CC_X86StdCall: return llvm::CallingConv::X86_StdCall;
49 case CC_X86FastCall: return llvm::CallingConv::X86_FastCall;
50 case CC_X86ThisCall: return llvm::CallingConv::X86_ThisCall;
51 case CC_X86_64Win64: return llvm::CallingConv::X86_64_Win64;
52 case CC_X86_64SysV: return llvm::CallingConv::X86_64_SysV;
53 case CC_AAPCS: return llvm::CallingConv::ARM_AAPCS;
54 case CC_AAPCS_VFP: return llvm::CallingConv::ARM_AAPCS_VFP;
55 case CC_IntelOclBicc: return llvm::CallingConv::Intel_OCL_BI;
56 // TODO: Add support for __pascal to LLVM.
57 case CC_X86Pascal: return llvm::CallingConv::C;
58 // TODO: Add support for __vectorcall to LLVM.
59 case CC_X86VectorCall: return llvm::CallingConv::X86_VectorCall;
60 case CC_SpirFunction: return llvm::CallingConv::SPIR_FUNC;
61 case CC_OpenCLKernel: return CGM.getTargetCodeGenInfo().getOpenCLKernelCallingConv();
62 case CC_PreserveMost: return llvm::CallingConv::PreserveMost;
63 case CC_PreserveAll: return llvm::CallingConv::PreserveAll;
64 case CC_Swift: return llvm::CallingConv::Swift;
68 /// Derives the 'this' type for codegen purposes, i.e. ignoring method
70 /// FIXME: address space qualification?
71 static CanQualType GetThisType(ASTContext &Context, const CXXRecordDecl *RD) {
72 QualType RecTy = Context.getTagDeclType(RD)->getCanonicalTypeInternal();
73 return Context.getPointerType(CanQualType::CreateUnsafe(RecTy));
76 /// Returns the canonical formal type of the given C++ method.
77 static CanQual<FunctionProtoType> GetFormalType(const CXXMethodDecl *MD) {
78 return MD->getType()->getCanonicalTypeUnqualified()
79 .getAs<FunctionProtoType>();
82 /// Returns the "extra-canonicalized" return type, which discards
83 /// qualifiers on the return type. Codegen doesn't care about them,
84 /// and it makes ABI code a little easier to be able to assume that
85 /// all parameter and return types are top-level unqualified.
86 static CanQualType GetReturnType(QualType RetTy) {
87 return RetTy->getCanonicalTypeUnqualified().getUnqualifiedType();
90 /// Arrange the argument and result information for a value of the given
91 /// unprototyped freestanding function type.
92 const CGFunctionInfo &
93 CodeGenTypes::arrangeFreeFunctionType(CanQual<FunctionNoProtoType> FTNP) {
94 // When translating an unprototyped function type, always use a
96 return arrangeLLVMFunctionInfo(FTNP->getReturnType().getUnqualifiedType(),
97 /*instanceMethod=*/false,
98 /*chainCall=*/false, None,
99 FTNP->getExtInfo(), {}, RequiredArgs(0));
102 /// Adds the formal paramaters in FPT to the given prefix. If any parameter in
103 /// FPT has pass_object_size attrs, then we'll add parameters for those, too.
104 static void appendParameterTypes(const CodeGenTypes &CGT,
105 SmallVectorImpl<CanQualType> &prefix,
106 SmallVectorImpl<FunctionProtoType::ExtParameterInfo> ¶mInfos,
107 CanQual<FunctionProtoType> FPT,
108 const FunctionDecl *FD) {
109 // Fill out paramInfos.
110 if (FPT->hasExtParameterInfos() || !paramInfos.empty()) {
111 assert(paramInfos.size() <= prefix.size());
112 auto protoParamInfos = FPT->getExtParameterInfos();
113 paramInfos.reserve(prefix.size() + protoParamInfos.size());
114 paramInfos.resize(prefix.size());
115 paramInfos.append(protoParamInfos.begin(), protoParamInfos.end());
118 // Fast path: unknown target.
120 prefix.append(FPT->param_type_begin(), FPT->param_type_end());
124 // In the vast majority cases, we'll have precisely FPT->getNumParams()
125 // parameters; the only thing that can change this is the presence of
126 // pass_object_size. So, we preallocate for the common case.
127 prefix.reserve(prefix.size() + FPT->getNumParams());
129 assert(FD->getNumParams() == FPT->getNumParams());
130 for (unsigned I = 0, E = FPT->getNumParams(); I != E; ++I) {
131 prefix.push_back(FPT->getParamType(I));
132 if (FD->getParamDecl(I)->hasAttr<PassObjectSizeAttr>())
133 prefix.push_back(CGT.getContext().getSizeType());
137 /// Arrange the LLVM function layout for a value of the given function
138 /// type, on top of any implicit parameters already stored.
139 static const CGFunctionInfo &
140 arrangeLLVMFunctionInfo(CodeGenTypes &CGT, bool instanceMethod,
141 SmallVectorImpl<CanQualType> &prefix,
142 CanQual<FunctionProtoType> FTP,
143 const FunctionDecl *FD) {
144 SmallVector<FunctionProtoType::ExtParameterInfo, 16> paramInfos;
145 RequiredArgs Required =
146 RequiredArgs::forPrototypePlus(FTP, prefix.size(), FD);
148 appendParameterTypes(CGT, prefix, paramInfos, FTP, FD);
149 CanQualType resultType = FTP->getReturnType().getUnqualifiedType();
151 return CGT.arrangeLLVMFunctionInfo(resultType, instanceMethod,
152 /*chainCall=*/false, prefix,
153 FTP->getExtInfo(), paramInfos,
157 /// Arrange the argument and result information for a value of the
158 /// given freestanding function type.
159 const CGFunctionInfo &
160 CodeGenTypes::arrangeFreeFunctionType(CanQual<FunctionProtoType> FTP,
161 const FunctionDecl *FD) {
162 SmallVector<CanQualType, 16> argTypes;
163 return ::arrangeLLVMFunctionInfo(*this, /*instanceMethod=*/false, argTypes,
167 static CallingConv getCallingConventionForDecl(const Decl *D, bool IsWindows) {
168 // Set the appropriate calling convention for the Function.
169 if (D->hasAttr<StdCallAttr>())
170 return CC_X86StdCall;
172 if (D->hasAttr<FastCallAttr>())
173 return CC_X86FastCall;
175 if (D->hasAttr<ThisCallAttr>())
176 return CC_X86ThisCall;
178 if (D->hasAttr<VectorCallAttr>())
179 return CC_X86VectorCall;
181 if (D->hasAttr<PascalAttr>())
184 if (PcsAttr *PCS = D->getAttr<PcsAttr>())
185 return (PCS->getPCS() == PcsAttr::AAPCS ? CC_AAPCS : CC_AAPCS_VFP);
187 if (D->hasAttr<IntelOclBiccAttr>())
188 return CC_IntelOclBicc;
190 if (D->hasAttr<MSABIAttr>())
191 return IsWindows ? CC_C : CC_X86_64Win64;
193 if (D->hasAttr<SysVABIAttr>())
194 return IsWindows ? CC_X86_64SysV : CC_C;
196 if (D->hasAttr<PreserveMostAttr>())
197 return CC_PreserveMost;
199 if (D->hasAttr<PreserveAllAttr>())
200 return CC_PreserveAll;
205 /// Arrange the argument and result information for a call to an
206 /// unknown C++ non-static member function of the given abstract type.
207 /// (Zero value of RD means we don't have any meaningful "this" argument type,
208 /// so fall back to a generic pointer type).
209 /// The member function must be an ordinary function, i.e. not a
210 /// constructor or destructor.
211 const CGFunctionInfo &
212 CodeGenTypes::arrangeCXXMethodType(const CXXRecordDecl *RD,
213 const FunctionProtoType *FTP,
214 const CXXMethodDecl *MD) {
215 SmallVector<CanQualType, 16> argTypes;
217 // Add the 'this' pointer.
219 argTypes.push_back(GetThisType(Context, RD));
221 argTypes.push_back(Context.VoidPtrTy);
223 return ::arrangeLLVMFunctionInfo(
224 *this, true, argTypes,
225 FTP->getCanonicalTypeUnqualified().getAs<FunctionProtoType>(), MD);
228 /// Arrange the argument and result information for a declaration or
229 /// definition of the given C++ non-static member function. The
230 /// member function must be an ordinary function, i.e. not a
231 /// constructor or destructor.
232 const CGFunctionInfo &
233 CodeGenTypes::arrangeCXXMethodDeclaration(const CXXMethodDecl *MD) {
234 assert(!isa<CXXConstructorDecl>(MD) && "wrong method for constructors!");
235 assert(!isa<CXXDestructorDecl>(MD) && "wrong method for destructors!");
237 CanQual<FunctionProtoType> prototype = GetFormalType(MD);
239 if (MD->isInstance()) {
240 // The abstract case is perfectly fine.
241 const CXXRecordDecl *ThisType = TheCXXABI.getThisArgumentTypeForMethod(MD);
242 return arrangeCXXMethodType(ThisType, prototype.getTypePtr(), MD);
245 return arrangeFreeFunctionType(prototype, MD);
248 bool CodeGenTypes::inheritingCtorHasParams(
249 const InheritedConstructor &Inherited, CXXCtorType Type) {
250 // Parameters are unnecessary if we're constructing a base class subobject
251 // and the inherited constructor lives in a virtual base.
252 return Type == Ctor_Complete ||
253 !Inherited.getShadowDecl()->constructsVirtualBase() ||
254 !Target.getCXXABI().hasConstructorVariants();
257 const CGFunctionInfo &
258 CodeGenTypes::arrangeCXXStructorDeclaration(const CXXMethodDecl *MD,
261 SmallVector<CanQualType, 16> argTypes;
262 SmallVector<FunctionProtoType::ExtParameterInfo, 16> paramInfos;
263 argTypes.push_back(GetThisType(Context, MD->getParent()));
265 bool PassParams = true;
268 if (auto *CD = dyn_cast<CXXConstructorDecl>(MD)) {
269 GD = GlobalDecl(CD, toCXXCtorType(Type));
271 // A base class inheriting constructor doesn't get forwarded arguments
272 // needed to construct a virtual base (or base class thereof).
273 if (auto Inherited = CD->getInheritedConstructor())
274 PassParams = inheritingCtorHasParams(Inherited, toCXXCtorType(Type));
276 auto *DD = dyn_cast<CXXDestructorDecl>(MD);
277 GD = GlobalDecl(DD, toCXXDtorType(Type));
280 CanQual<FunctionProtoType> FTP = GetFormalType(MD);
282 // Add the formal parameters.
284 appendParameterTypes(*this, argTypes, paramInfos, FTP, MD);
286 TheCXXABI.buildStructorSignature(MD, Type, argTypes);
288 RequiredArgs required =
289 (PassParams && MD->isVariadic() ? RequiredArgs(argTypes.size())
290 : RequiredArgs::All);
292 FunctionType::ExtInfo extInfo = FTP->getExtInfo();
293 CanQualType resultType = TheCXXABI.HasThisReturn(GD)
295 : TheCXXABI.hasMostDerivedReturn(GD)
296 ? CGM.getContext().VoidPtrTy
298 return arrangeLLVMFunctionInfo(resultType, /*instanceMethod=*/true,
299 /*chainCall=*/false, argTypes, extInfo,
300 paramInfos, required);
303 static SmallVector<CanQualType, 16>
304 getArgTypesForCall(ASTContext &ctx, const CallArgList &args) {
305 SmallVector<CanQualType, 16> argTypes;
306 for (auto &arg : args)
307 argTypes.push_back(ctx.getCanonicalParamType(arg.Ty));
311 static SmallVector<CanQualType, 16>
312 getArgTypesForDeclaration(ASTContext &ctx, const FunctionArgList &args) {
313 SmallVector<CanQualType, 16> argTypes;
314 for (auto &arg : args)
315 argTypes.push_back(ctx.getCanonicalParamType(arg->getType()));
319 static void addExtParameterInfosForCall(
320 llvm::SmallVectorImpl<FunctionProtoType::ExtParameterInfo> ¶mInfos,
321 const FunctionProtoType *proto,
323 unsigned totalArgs) {
324 assert(proto->hasExtParameterInfos());
325 assert(paramInfos.size() <= prefixArgs);
326 assert(proto->getNumParams() + prefixArgs <= totalArgs);
328 // Add default infos for any prefix args that don't already have infos.
329 paramInfos.resize(prefixArgs);
331 // Add infos for the prototype.
332 auto protoInfos = proto->getExtParameterInfos();
333 paramInfos.append(protoInfos.begin(), protoInfos.end());
335 // Add default infos for the variadic arguments.
336 paramInfos.resize(totalArgs);
339 static llvm::SmallVector<FunctionProtoType::ExtParameterInfo, 16>
340 getExtParameterInfosForCall(const FunctionProtoType *proto,
341 unsigned prefixArgs, unsigned totalArgs) {
342 llvm::SmallVector<FunctionProtoType::ExtParameterInfo, 16> result;
343 if (proto->hasExtParameterInfos()) {
344 addExtParameterInfosForCall(result, proto, prefixArgs, totalArgs);
349 /// Arrange a call to a C++ method, passing the given arguments.
350 const CGFunctionInfo &
351 CodeGenTypes::arrangeCXXConstructorCall(const CallArgList &args,
352 const CXXConstructorDecl *D,
353 CXXCtorType CtorKind,
354 unsigned ExtraArgs) {
356 SmallVector<CanQualType, 16> ArgTypes;
357 for (const auto &Arg : args)
358 ArgTypes.push_back(Context.getCanonicalParamType(Arg.Ty));
360 CanQual<FunctionProtoType> FPT = GetFormalType(D);
361 RequiredArgs Required = RequiredArgs::forPrototypePlus(FPT, 1 + ExtraArgs, D);
362 GlobalDecl GD(D, CtorKind);
363 CanQualType ResultType = TheCXXABI.HasThisReturn(GD)
365 : TheCXXABI.hasMostDerivedReturn(GD)
366 ? CGM.getContext().VoidPtrTy
369 FunctionType::ExtInfo Info = FPT->getExtInfo();
370 auto ParamInfos = getExtParameterInfosForCall(FPT.getTypePtr(), 1 + ExtraArgs,
372 return arrangeLLVMFunctionInfo(ResultType, /*instanceMethod=*/true,
373 /*chainCall=*/false, ArgTypes, Info,
374 ParamInfos, Required);
377 /// Arrange the argument and result information for the declaration or
378 /// definition of the given function.
379 const CGFunctionInfo &
380 CodeGenTypes::arrangeFunctionDeclaration(const FunctionDecl *FD) {
381 if (const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(FD))
382 if (MD->isInstance())
383 return arrangeCXXMethodDeclaration(MD);
385 CanQualType FTy = FD->getType()->getCanonicalTypeUnqualified();
387 assert(isa<FunctionType>(FTy));
389 // When declaring a function without a prototype, always use a
390 // non-variadic type.
391 if (isa<FunctionNoProtoType>(FTy)) {
392 CanQual<FunctionNoProtoType> noProto = FTy.getAs<FunctionNoProtoType>();
393 return arrangeLLVMFunctionInfo(
394 noProto->getReturnType(), /*instanceMethod=*/false,
395 /*chainCall=*/false, None, noProto->getExtInfo(), {},RequiredArgs::All);
398 assert(isa<FunctionProtoType>(FTy));
399 return arrangeFreeFunctionType(FTy.getAs<FunctionProtoType>(), FD);
402 /// Arrange the argument and result information for the declaration or
403 /// definition of an Objective-C method.
404 const CGFunctionInfo &
405 CodeGenTypes::arrangeObjCMethodDeclaration(const ObjCMethodDecl *MD) {
406 // It happens that this is the same as a call with no optional
407 // arguments, except also using the formal 'self' type.
408 return arrangeObjCMessageSendSignature(MD, MD->getSelfDecl()->getType());
411 /// Arrange the argument and result information for the function type
412 /// through which to perform a send to the given Objective-C method,
413 /// using the given receiver type. The receiver type is not always
414 /// the 'self' type of the method or even an Objective-C pointer type.
415 /// This is *not* the right method for actually performing such a
416 /// message send, due to the possibility of optional arguments.
417 const CGFunctionInfo &
418 CodeGenTypes::arrangeObjCMessageSendSignature(const ObjCMethodDecl *MD,
419 QualType receiverType) {
420 SmallVector<CanQualType, 16> argTys;
421 argTys.push_back(Context.getCanonicalParamType(receiverType));
422 argTys.push_back(Context.getCanonicalParamType(Context.getObjCSelType()));
424 for (const auto *I : MD->parameters()) {
425 argTys.push_back(Context.getCanonicalParamType(I->getType()));
428 FunctionType::ExtInfo einfo;
429 bool IsWindows = getContext().getTargetInfo().getTriple().isOSWindows();
430 einfo = einfo.withCallingConv(getCallingConventionForDecl(MD, IsWindows));
432 if (getContext().getLangOpts().ObjCAutoRefCount &&
433 MD->hasAttr<NSReturnsRetainedAttr>())
434 einfo = einfo.withProducesResult(true);
436 RequiredArgs required =
437 (MD->isVariadic() ? RequiredArgs(argTys.size()) : RequiredArgs::All);
439 return arrangeLLVMFunctionInfo(
440 GetReturnType(MD->getReturnType()), /*instanceMethod=*/false,
441 /*chainCall=*/false, argTys, einfo, {}, required);
444 const CGFunctionInfo &
445 CodeGenTypes::arrangeUnprototypedObjCMessageSend(QualType returnType,
446 const CallArgList &args) {
447 auto argTypes = getArgTypesForCall(Context, args);
448 FunctionType::ExtInfo einfo;
450 return arrangeLLVMFunctionInfo(
451 GetReturnType(returnType), /*instanceMethod=*/false,
452 /*chainCall=*/false, argTypes, einfo, {}, RequiredArgs::All);
455 const CGFunctionInfo &
456 CodeGenTypes::arrangeGlobalDeclaration(GlobalDecl GD) {
457 // FIXME: Do we need to handle ObjCMethodDecl?
458 const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl());
460 if (const CXXConstructorDecl *CD = dyn_cast<CXXConstructorDecl>(FD))
461 return arrangeCXXStructorDeclaration(CD, getFromCtorType(GD.getCtorType()));
463 if (const CXXDestructorDecl *DD = dyn_cast<CXXDestructorDecl>(FD))
464 return arrangeCXXStructorDeclaration(DD, getFromDtorType(GD.getDtorType()));
466 return arrangeFunctionDeclaration(FD);
469 /// Arrange a thunk that takes 'this' as the first parameter followed by
470 /// varargs. Return a void pointer, regardless of the actual return type.
471 /// The body of the thunk will end in a musttail call to a function of the
472 /// correct type, and the caller will bitcast the function to the correct
474 const CGFunctionInfo &
475 CodeGenTypes::arrangeMSMemberPointerThunk(const CXXMethodDecl *MD) {
476 assert(MD->isVirtual() && "only virtual memptrs have thunks");
477 CanQual<FunctionProtoType> FTP = GetFormalType(MD);
478 CanQualType ArgTys[] = { GetThisType(Context, MD->getParent()) };
479 return arrangeLLVMFunctionInfo(Context.VoidTy, /*instanceMethod=*/false,
480 /*chainCall=*/false, ArgTys,
481 FTP->getExtInfo(), {}, RequiredArgs(1));
484 const CGFunctionInfo &
485 CodeGenTypes::arrangeMSCtorClosure(const CXXConstructorDecl *CD,
487 assert(CT == Ctor_CopyingClosure || CT == Ctor_DefaultClosure);
489 CanQual<FunctionProtoType> FTP = GetFormalType(CD);
490 SmallVector<CanQualType, 2> ArgTys;
491 const CXXRecordDecl *RD = CD->getParent();
492 ArgTys.push_back(GetThisType(Context, RD));
493 if (CT == Ctor_CopyingClosure)
494 ArgTys.push_back(*FTP->param_type_begin());
495 if (RD->getNumVBases() > 0)
496 ArgTys.push_back(Context.IntTy);
497 CallingConv CC = Context.getDefaultCallingConvention(
498 /*IsVariadic=*/false, /*IsCXXMethod=*/true);
499 return arrangeLLVMFunctionInfo(Context.VoidTy, /*instanceMethod=*/true,
500 /*chainCall=*/false, ArgTys,
501 FunctionType::ExtInfo(CC), {},
505 /// Arrange a call as unto a free function, except possibly with an
506 /// additional number of formal parameters considered required.
507 static const CGFunctionInfo &
508 arrangeFreeFunctionLikeCall(CodeGenTypes &CGT,
510 const CallArgList &args,
511 const FunctionType *fnType,
512 unsigned numExtraRequiredArgs,
514 assert(args.size() >= numExtraRequiredArgs);
516 llvm::SmallVector<FunctionProtoType::ExtParameterInfo, 16> paramInfos;
518 // In most cases, there are no optional arguments.
519 RequiredArgs required = RequiredArgs::All;
521 // If we have a variadic prototype, the required arguments are the
522 // extra prefix plus the arguments in the prototype.
523 if (const FunctionProtoType *proto = dyn_cast<FunctionProtoType>(fnType)) {
524 if (proto->isVariadic())
525 required = RequiredArgs(proto->getNumParams() + numExtraRequiredArgs);
527 if (proto->hasExtParameterInfos())
528 addExtParameterInfosForCall(paramInfos, proto, numExtraRequiredArgs,
531 // If we don't have a prototype at all, but we're supposed to
532 // explicitly use the variadic convention for unprototyped calls,
533 // treat all of the arguments as required but preserve the nominal
534 // possibility of variadics.
535 } else if (CGM.getTargetCodeGenInfo()
536 .isNoProtoCallVariadic(args,
537 cast<FunctionNoProtoType>(fnType))) {
538 required = RequiredArgs(args.size());
542 SmallVector<CanQualType, 16> argTypes;
543 for (const auto &arg : args)
544 argTypes.push_back(CGT.getContext().getCanonicalParamType(arg.Ty));
545 return CGT.arrangeLLVMFunctionInfo(GetReturnType(fnType->getReturnType()),
546 /*instanceMethod=*/false, chainCall,
547 argTypes, fnType->getExtInfo(), paramInfos,
551 /// Figure out the rules for calling a function with the given formal
552 /// type using the given arguments. The arguments are necessary
553 /// because the function might be unprototyped, in which case it's
554 /// target-dependent in crazy ways.
555 const CGFunctionInfo &
556 CodeGenTypes::arrangeFreeFunctionCall(const CallArgList &args,
557 const FunctionType *fnType,
559 return arrangeFreeFunctionLikeCall(*this, CGM, args, fnType,
560 chainCall ? 1 : 0, chainCall);
563 /// A block function is essentially a free function with an
564 /// extra implicit argument.
565 const CGFunctionInfo &
566 CodeGenTypes::arrangeBlockFunctionCall(const CallArgList &args,
567 const FunctionType *fnType) {
568 return arrangeFreeFunctionLikeCall(*this, CGM, args, fnType, 1,
569 /*chainCall=*/false);
572 const CGFunctionInfo &
573 CodeGenTypes::arrangeBlockFunctionDeclaration(const FunctionProtoType *proto,
574 const FunctionArgList ¶ms) {
575 auto paramInfos = getExtParameterInfosForCall(proto, 1, params.size());
576 auto argTypes = getArgTypesForDeclaration(Context, params);
578 return arrangeLLVMFunctionInfo(
579 GetReturnType(proto->getReturnType()),
580 /*instanceMethod*/ false, /*chainCall*/ false, argTypes,
581 proto->getExtInfo(), paramInfos,
582 RequiredArgs::forPrototypePlus(proto, 1, nullptr));
585 const CGFunctionInfo &
586 CodeGenTypes::arrangeBuiltinFunctionCall(QualType resultType,
587 const CallArgList &args) {
589 SmallVector<CanQualType, 16> argTypes;
590 for (const auto &Arg : args)
591 argTypes.push_back(Context.getCanonicalParamType(Arg.Ty));
592 return arrangeLLVMFunctionInfo(
593 GetReturnType(resultType), /*instanceMethod=*/false,
594 /*chainCall=*/false, argTypes, FunctionType::ExtInfo(),
595 /*paramInfos=*/ {}, RequiredArgs::All);
598 const CGFunctionInfo &
599 CodeGenTypes::arrangeBuiltinFunctionDeclaration(QualType resultType,
600 const FunctionArgList &args) {
601 auto argTypes = getArgTypesForDeclaration(Context, args);
603 return arrangeLLVMFunctionInfo(
604 GetReturnType(resultType), /*instanceMethod=*/false, /*chainCall=*/false,
605 argTypes, FunctionType::ExtInfo(), {}, RequiredArgs::All);
608 const CGFunctionInfo &
609 CodeGenTypes::arrangeBuiltinFunctionDeclaration(CanQualType resultType,
610 ArrayRef<CanQualType> argTypes) {
611 return arrangeLLVMFunctionInfo(
612 resultType, /*instanceMethod=*/false, /*chainCall=*/false,
613 argTypes, FunctionType::ExtInfo(), {}, RequiredArgs::All);
616 /// Arrange a call to a C++ method, passing the given arguments.
617 const CGFunctionInfo &
618 CodeGenTypes::arrangeCXXMethodCall(const CallArgList &args,
619 const FunctionProtoType *proto,
620 RequiredArgs required) {
621 unsigned numRequiredArgs =
622 (proto->isVariadic() ? required.getNumRequiredArgs() : args.size());
623 unsigned numPrefixArgs = numRequiredArgs - proto->getNumParams();
625 getExtParameterInfosForCall(proto, numPrefixArgs, args.size());
628 auto argTypes = getArgTypesForCall(Context, args);
630 FunctionType::ExtInfo info = proto->getExtInfo();
631 return arrangeLLVMFunctionInfo(
632 GetReturnType(proto->getReturnType()), /*instanceMethod=*/true,
633 /*chainCall=*/false, argTypes, info, paramInfos, required);
636 const CGFunctionInfo &CodeGenTypes::arrangeNullaryFunction() {
637 return arrangeLLVMFunctionInfo(
638 getContext().VoidTy, /*instanceMethod=*/false, /*chainCall=*/false,
639 None, FunctionType::ExtInfo(), {}, RequiredArgs::All);
642 const CGFunctionInfo &
643 CodeGenTypes::arrangeCall(const CGFunctionInfo &signature,
644 const CallArgList &args) {
645 assert(signature.arg_size() <= args.size());
646 if (signature.arg_size() == args.size())
649 SmallVector<FunctionProtoType::ExtParameterInfo, 16> paramInfos;
650 auto sigParamInfos = signature.getExtParameterInfos();
651 if (!sigParamInfos.empty()) {
652 paramInfos.append(sigParamInfos.begin(), sigParamInfos.end());
653 paramInfos.resize(args.size());
656 auto argTypes = getArgTypesForCall(Context, args);
658 assert(signature.getRequiredArgs().allowsOptionalArgs());
659 return arrangeLLVMFunctionInfo(signature.getReturnType(),
660 signature.isInstanceMethod(),
661 signature.isChainCall(),
663 signature.getExtInfo(),
665 signature.getRequiredArgs());
668 /// Arrange the argument and result information for an abstract value
669 /// of a given function type. This is the method which all of the
670 /// above functions ultimately defer to.
671 const CGFunctionInfo &
672 CodeGenTypes::arrangeLLVMFunctionInfo(CanQualType resultType,
675 ArrayRef<CanQualType> argTypes,
676 FunctionType::ExtInfo info,
677 ArrayRef<FunctionProtoType::ExtParameterInfo> paramInfos,
678 RequiredArgs required) {
679 assert(std::all_of(argTypes.begin(), argTypes.end(),
680 std::mem_fun_ref(&CanQualType::isCanonicalAsParam)));
682 // Lookup or create unique function info.
683 llvm::FoldingSetNodeID ID;
684 CGFunctionInfo::Profile(ID, instanceMethod, chainCall, info, paramInfos,
685 required, resultType, argTypes);
687 void *insertPos = nullptr;
688 CGFunctionInfo *FI = FunctionInfos.FindNodeOrInsertPos(ID, insertPos);
692 unsigned CC = ClangCallConvToLLVMCallConv(info.getCC());
694 // Construct the function info. We co-allocate the ArgInfos.
695 FI = CGFunctionInfo::create(CC, instanceMethod, chainCall, info,
696 paramInfos, resultType, argTypes, required);
697 FunctionInfos.InsertNode(FI, insertPos);
699 bool inserted = FunctionsBeingProcessed.insert(FI).second;
701 assert(inserted && "Recursively being processed?");
703 // Compute ABI information.
704 if (info.getCC() != CC_Swift) {
705 getABIInfo().computeInfo(*FI);
707 swiftcall::computeABIInfo(CGM, *FI);
710 // Loop over all of the computed argument and return value info. If any of
711 // them are direct or extend without a specified coerce type, specify the
713 ABIArgInfo &retInfo = FI->getReturnInfo();
714 if (retInfo.canHaveCoerceToType() && retInfo.getCoerceToType() == nullptr)
715 retInfo.setCoerceToType(ConvertType(FI->getReturnType()));
717 for (auto &I : FI->arguments())
718 if (I.info.canHaveCoerceToType() && I.info.getCoerceToType() == nullptr)
719 I.info.setCoerceToType(ConvertType(I.type));
721 bool erased = FunctionsBeingProcessed.erase(FI); (void)erased;
722 assert(erased && "Not in set?");
727 CGFunctionInfo *CGFunctionInfo::create(unsigned llvmCC,
730 const FunctionType::ExtInfo &info,
731 ArrayRef<ExtParameterInfo> paramInfos,
732 CanQualType resultType,
733 ArrayRef<CanQualType> argTypes,
734 RequiredArgs required) {
735 assert(paramInfos.empty() || paramInfos.size() == argTypes.size());
738 operator new(totalSizeToAlloc<ArgInfo, ExtParameterInfo>(
739 argTypes.size() + 1, paramInfos.size()));
741 CGFunctionInfo *FI = new(buffer) CGFunctionInfo();
742 FI->CallingConvention = llvmCC;
743 FI->EffectiveCallingConvention = llvmCC;
744 FI->ASTCallingConvention = info.getCC();
745 FI->InstanceMethod = instanceMethod;
746 FI->ChainCall = chainCall;
747 FI->NoReturn = info.getNoReturn();
748 FI->ReturnsRetained = info.getProducesResult();
749 FI->Required = required;
750 FI->HasRegParm = info.getHasRegParm();
751 FI->RegParm = info.getRegParm();
752 FI->ArgStruct = nullptr;
753 FI->ArgStructAlign = 0;
754 FI->NumArgs = argTypes.size();
755 FI->HasExtParameterInfos = !paramInfos.empty();
756 FI->getArgsBuffer()[0].type = resultType;
757 for (unsigned i = 0, e = argTypes.size(); i != e; ++i)
758 FI->getArgsBuffer()[i + 1].type = argTypes[i];
759 for (unsigned i = 0, e = paramInfos.size(); i != e; ++i)
760 FI->getExtParameterInfosBuffer()[i] = paramInfos[i];
767 // ABIArgInfo::Expand implementation.
769 // Specifies the way QualType passed as ABIArgInfo::Expand is expanded.
770 struct TypeExpansion {
771 enum TypeExpansionKind {
772 // Elements of constant arrays are expanded recursively.
774 // Record fields are expanded recursively (but if record is a union, only
775 // the field with the largest size is expanded).
777 // For complex types, real and imaginary parts are expanded recursively.
779 // All other types are not expandable.
783 const TypeExpansionKind Kind;
785 TypeExpansion(TypeExpansionKind K) : Kind(K) {}
786 virtual ~TypeExpansion() {}
789 struct ConstantArrayExpansion : TypeExpansion {
793 ConstantArrayExpansion(QualType EltTy, uint64_t NumElts)
794 : TypeExpansion(TEK_ConstantArray), EltTy(EltTy), NumElts(NumElts) {}
795 static bool classof(const TypeExpansion *TE) {
796 return TE->Kind == TEK_ConstantArray;
800 struct RecordExpansion : TypeExpansion {
801 SmallVector<const CXXBaseSpecifier *, 1> Bases;
803 SmallVector<const FieldDecl *, 1> Fields;
805 RecordExpansion(SmallVector<const CXXBaseSpecifier *, 1> &&Bases,
806 SmallVector<const FieldDecl *, 1> &&Fields)
807 : TypeExpansion(TEK_Record), Bases(std::move(Bases)),
808 Fields(std::move(Fields)) {}
809 static bool classof(const TypeExpansion *TE) {
810 return TE->Kind == TEK_Record;
814 struct ComplexExpansion : TypeExpansion {
817 ComplexExpansion(QualType EltTy) : TypeExpansion(TEK_Complex), EltTy(EltTy) {}
818 static bool classof(const TypeExpansion *TE) {
819 return TE->Kind == TEK_Complex;
823 struct NoExpansion : TypeExpansion {
824 NoExpansion() : TypeExpansion(TEK_None) {}
825 static bool classof(const TypeExpansion *TE) {
826 return TE->Kind == TEK_None;
831 static std::unique_ptr<TypeExpansion>
832 getTypeExpansion(QualType Ty, const ASTContext &Context) {
833 if (const ConstantArrayType *AT = Context.getAsConstantArrayType(Ty)) {
834 return llvm::make_unique<ConstantArrayExpansion>(
835 AT->getElementType(), AT->getSize().getZExtValue());
837 if (const RecordType *RT = Ty->getAs<RecordType>()) {
838 SmallVector<const CXXBaseSpecifier *, 1> Bases;
839 SmallVector<const FieldDecl *, 1> Fields;
840 const RecordDecl *RD = RT->getDecl();
841 assert(!RD->hasFlexibleArrayMember() &&
842 "Cannot expand structure with flexible array.");
844 // Unions can be here only in degenerative cases - all the fields are same
845 // after flattening. Thus we have to use the "largest" field.
846 const FieldDecl *LargestFD = nullptr;
847 CharUnits UnionSize = CharUnits::Zero();
849 for (const auto *FD : RD->fields()) {
850 // Skip zero length bitfields.
851 if (FD->isBitField() && FD->getBitWidthValue(Context) == 0)
853 assert(!FD->isBitField() &&
854 "Cannot expand structure with bit-field members.");
855 CharUnits FieldSize = Context.getTypeSizeInChars(FD->getType());
856 if (UnionSize < FieldSize) {
857 UnionSize = FieldSize;
862 Fields.push_back(LargestFD);
864 if (const auto *CXXRD = dyn_cast<CXXRecordDecl>(RD)) {
865 assert(!CXXRD->isDynamicClass() &&
866 "cannot expand vtable pointers in dynamic classes");
867 for (const CXXBaseSpecifier &BS : CXXRD->bases())
868 Bases.push_back(&BS);
871 for (const auto *FD : RD->fields()) {
872 // Skip zero length bitfields.
873 if (FD->isBitField() && FD->getBitWidthValue(Context) == 0)
875 assert(!FD->isBitField() &&
876 "Cannot expand structure with bit-field members.");
877 Fields.push_back(FD);
880 return llvm::make_unique<RecordExpansion>(std::move(Bases),
883 if (const ComplexType *CT = Ty->getAs<ComplexType>()) {
884 return llvm::make_unique<ComplexExpansion>(CT->getElementType());
886 return llvm::make_unique<NoExpansion>();
889 static int getExpansionSize(QualType Ty, const ASTContext &Context) {
890 auto Exp = getTypeExpansion(Ty, Context);
891 if (auto CAExp = dyn_cast<ConstantArrayExpansion>(Exp.get())) {
892 return CAExp->NumElts * getExpansionSize(CAExp->EltTy, Context);
894 if (auto RExp = dyn_cast<RecordExpansion>(Exp.get())) {
896 for (auto BS : RExp->Bases)
897 Res += getExpansionSize(BS->getType(), Context);
898 for (auto FD : RExp->Fields)
899 Res += getExpansionSize(FD->getType(), Context);
902 if (isa<ComplexExpansion>(Exp.get()))
904 assert(isa<NoExpansion>(Exp.get()));
909 CodeGenTypes::getExpandedTypes(QualType Ty,
910 SmallVectorImpl<llvm::Type *>::iterator &TI) {
911 auto Exp = getTypeExpansion(Ty, Context);
912 if (auto CAExp = dyn_cast<ConstantArrayExpansion>(Exp.get())) {
913 for (int i = 0, n = CAExp->NumElts; i < n; i++) {
914 getExpandedTypes(CAExp->EltTy, TI);
916 } else if (auto RExp = dyn_cast<RecordExpansion>(Exp.get())) {
917 for (auto BS : RExp->Bases)
918 getExpandedTypes(BS->getType(), TI);
919 for (auto FD : RExp->Fields)
920 getExpandedTypes(FD->getType(), TI);
921 } else if (auto CExp = dyn_cast<ComplexExpansion>(Exp.get())) {
922 llvm::Type *EltTy = ConvertType(CExp->EltTy);
926 assert(isa<NoExpansion>(Exp.get()));
927 *TI++ = ConvertType(Ty);
931 static void forConstantArrayExpansion(CodeGenFunction &CGF,
932 ConstantArrayExpansion *CAE,
934 llvm::function_ref<void(Address)> Fn) {
935 CharUnits EltSize = CGF.getContext().getTypeSizeInChars(CAE->EltTy);
937 BaseAddr.getAlignment().alignmentOfArrayElement(EltSize);
939 for (int i = 0, n = CAE->NumElts; i < n; i++) {
940 llvm::Value *EltAddr =
941 CGF.Builder.CreateConstGEP2_32(nullptr, BaseAddr.getPointer(), 0, i);
942 Fn(Address(EltAddr, EltAlign));
946 void CodeGenFunction::ExpandTypeFromArgs(
947 QualType Ty, LValue LV, SmallVectorImpl<llvm::Value *>::iterator &AI) {
948 assert(LV.isSimple() &&
949 "Unexpected non-simple lvalue during struct expansion.");
951 auto Exp = getTypeExpansion(Ty, getContext());
952 if (auto CAExp = dyn_cast<ConstantArrayExpansion>(Exp.get())) {
953 forConstantArrayExpansion(*this, CAExp, LV.getAddress(),
954 [&](Address EltAddr) {
955 LValue LV = MakeAddrLValue(EltAddr, CAExp->EltTy);
956 ExpandTypeFromArgs(CAExp->EltTy, LV, AI);
958 } else if (auto RExp = dyn_cast<RecordExpansion>(Exp.get())) {
959 Address This = LV.getAddress();
960 for (const CXXBaseSpecifier *BS : RExp->Bases) {
961 // Perform a single step derived-to-base conversion.
963 GetAddressOfBaseClass(This, Ty->getAsCXXRecordDecl(), &BS, &BS + 1,
964 /*NullCheckValue=*/false, SourceLocation());
965 LValue SubLV = MakeAddrLValue(Base, BS->getType());
967 // Recurse onto bases.
968 ExpandTypeFromArgs(BS->getType(), SubLV, AI);
970 for (auto FD : RExp->Fields) {
971 // FIXME: What are the right qualifiers here?
972 LValue SubLV = EmitLValueForFieldInitialization(LV, FD);
973 ExpandTypeFromArgs(FD->getType(), SubLV, AI);
975 } else if (isa<ComplexExpansion>(Exp.get())) {
976 auto realValue = *AI++;
977 auto imagValue = *AI++;
978 EmitStoreOfComplex(ComplexPairTy(realValue, imagValue), LV, /*init*/ true);
980 assert(isa<NoExpansion>(Exp.get()));
981 EmitStoreThroughLValue(RValue::get(*AI++), LV);
985 void CodeGenFunction::ExpandTypeToArgs(
986 QualType Ty, RValue RV, llvm::FunctionType *IRFuncTy,
987 SmallVectorImpl<llvm::Value *> &IRCallArgs, unsigned &IRCallArgPos) {
988 auto Exp = getTypeExpansion(Ty, getContext());
989 if (auto CAExp = dyn_cast<ConstantArrayExpansion>(Exp.get())) {
990 forConstantArrayExpansion(*this, CAExp, RV.getAggregateAddress(),
991 [&](Address EltAddr) {
993 convertTempToRValue(EltAddr, CAExp->EltTy, SourceLocation());
994 ExpandTypeToArgs(CAExp->EltTy, EltRV, IRFuncTy, IRCallArgs, IRCallArgPos);
996 } else if (auto RExp = dyn_cast<RecordExpansion>(Exp.get())) {
997 Address This = RV.getAggregateAddress();
998 for (const CXXBaseSpecifier *BS : RExp->Bases) {
999 // Perform a single step derived-to-base conversion.
1001 GetAddressOfBaseClass(This, Ty->getAsCXXRecordDecl(), &BS, &BS + 1,
1002 /*NullCheckValue=*/false, SourceLocation());
1003 RValue BaseRV = RValue::getAggregate(Base);
1005 // Recurse onto bases.
1006 ExpandTypeToArgs(BS->getType(), BaseRV, IRFuncTy, IRCallArgs,
1010 LValue LV = MakeAddrLValue(This, Ty);
1011 for (auto FD : RExp->Fields) {
1012 RValue FldRV = EmitRValueForField(LV, FD, SourceLocation());
1013 ExpandTypeToArgs(FD->getType(), FldRV, IRFuncTy, IRCallArgs,
1016 } else if (isa<ComplexExpansion>(Exp.get())) {
1017 ComplexPairTy CV = RV.getComplexVal();
1018 IRCallArgs[IRCallArgPos++] = CV.first;
1019 IRCallArgs[IRCallArgPos++] = CV.second;
1021 assert(isa<NoExpansion>(Exp.get()));
1022 assert(RV.isScalar() &&
1023 "Unexpected non-scalar rvalue during struct expansion.");
1025 // Insert a bitcast as needed.
1026 llvm::Value *V = RV.getScalarVal();
1027 if (IRCallArgPos < IRFuncTy->getNumParams() &&
1028 V->getType() != IRFuncTy->getParamType(IRCallArgPos))
1029 V = Builder.CreateBitCast(V, IRFuncTy->getParamType(IRCallArgPos));
1031 IRCallArgs[IRCallArgPos++] = V;
1035 /// Create a temporary allocation for the purposes of coercion.
1036 static Address CreateTempAllocaForCoercion(CodeGenFunction &CGF, llvm::Type *Ty,
1037 CharUnits MinAlign) {
1038 // Don't use an alignment that's worse than what LLVM would prefer.
1039 auto PrefAlign = CGF.CGM.getDataLayout().getPrefTypeAlignment(Ty);
1040 CharUnits Align = std::max(MinAlign, CharUnits::fromQuantity(PrefAlign));
1042 return CGF.CreateTempAlloca(Ty, Align);
1045 /// EnterStructPointerForCoercedAccess - Given a struct pointer that we are
1046 /// accessing some number of bytes out of it, try to gep into the struct to get
1047 /// at its inner goodness. Dive as deep as possible without entering an element
1048 /// with an in-memory size smaller than DstSize.
1050 EnterStructPointerForCoercedAccess(Address SrcPtr,
1051 llvm::StructType *SrcSTy,
1052 uint64_t DstSize, CodeGenFunction &CGF) {
1053 // We can't dive into a zero-element struct.
1054 if (SrcSTy->getNumElements() == 0) return SrcPtr;
1056 llvm::Type *FirstElt = SrcSTy->getElementType(0);
1058 // If the first elt is at least as large as what we're looking for, or if the
1059 // first element is the same size as the whole struct, we can enter it. The
1060 // comparison must be made on the store size and not the alloca size. Using
1061 // the alloca size may overstate the size of the load.
1062 uint64_t FirstEltSize =
1063 CGF.CGM.getDataLayout().getTypeStoreSize(FirstElt);
1064 if (FirstEltSize < DstSize &&
1065 FirstEltSize < CGF.CGM.getDataLayout().getTypeStoreSize(SrcSTy))
1068 // GEP into the first element.
1069 SrcPtr = CGF.Builder.CreateStructGEP(SrcPtr, 0, CharUnits(), "coerce.dive");
1071 // If the first element is a struct, recurse.
1072 llvm::Type *SrcTy = SrcPtr.getElementType();
1073 if (llvm::StructType *SrcSTy = dyn_cast<llvm::StructType>(SrcTy))
1074 return EnterStructPointerForCoercedAccess(SrcPtr, SrcSTy, DstSize, CGF);
1079 /// CoerceIntOrPtrToIntOrPtr - Convert a value Val to the specific Ty where both
1080 /// are either integers or pointers. This does a truncation of the value if it
1081 /// is too large or a zero extension if it is too small.
1083 /// This behaves as if the value were coerced through memory, so on big-endian
1084 /// targets the high bits are preserved in a truncation, while little-endian
1085 /// targets preserve the low bits.
1086 static llvm::Value *CoerceIntOrPtrToIntOrPtr(llvm::Value *Val,
1088 CodeGenFunction &CGF) {
1089 if (Val->getType() == Ty)
1092 if (isa<llvm::PointerType>(Val->getType())) {
1093 // If this is Pointer->Pointer avoid conversion to and from int.
1094 if (isa<llvm::PointerType>(Ty))
1095 return CGF.Builder.CreateBitCast(Val, Ty, "coerce.val");
1097 // Convert the pointer to an integer so we can play with its width.
1098 Val = CGF.Builder.CreatePtrToInt(Val, CGF.IntPtrTy, "coerce.val.pi");
1101 llvm::Type *DestIntTy = Ty;
1102 if (isa<llvm::PointerType>(DestIntTy))
1103 DestIntTy = CGF.IntPtrTy;
1105 if (Val->getType() != DestIntTy) {
1106 const llvm::DataLayout &DL = CGF.CGM.getDataLayout();
1107 if (DL.isBigEndian()) {
1108 // Preserve the high bits on big-endian targets.
1109 // That is what memory coercion does.
1110 uint64_t SrcSize = DL.getTypeSizeInBits(Val->getType());
1111 uint64_t DstSize = DL.getTypeSizeInBits(DestIntTy);
1113 if (SrcSize > DstSize) {
1114 Val = CGF.Builder.CreateLShr(Val, SrcSize - DstSize, "coerce.highbits");
1115 Val = CGF.Builder.CreateTrunc(Val, DestIntTy, "coerce.val.ii");
1117 Val = CGF.Builder.CreateZExt(Val, DestIntTy, "coerce.val.ii");
1118 Val = CGF.Builder.CreateShl(Val, DstSize - SrcSize, "coerce.highbits");
1121 // Little-endian targets preserve the low bits. No shifts required.
1122 Val = CGF.Builder.CreateIntCast(Val, DestIntTy, false, "coerce.val.ii");
1126 if (isa<llvm::PointerType>(Ty))
1127 Val = CGF.Builder.CreateIntToPtr(Val, Ty, "coerce.val.ip");
1133 /// CreateCoercedLoad - Create a load from \arg SrcPtr interpreted as
1134 /// a pointer to an object of type \arg Ty, known to be aligned to
1135 /// \arg SrcAlign bytes.
1137 /// This safely handles the case when the src type is smaller than the
1138 /// destination type; in this situation the values of bits which not
1139 /// present in the src are undefined.
1140 static llvm::Value *CreateCoercedLoad(Address Src, llvm::Type *Ty,
1141 CodeGenFunction &CGF) {
1142 llvm::Type *SrcTy = Src.getElementType();
1144 // If SrcTy and Ty are the same, just do a load.
1146 return CGF.Builder.CreateLoad(Src);
1148 uint64_t DstSize = CGF.CGM.getDataLayout().getTypeAllocSize(Ty);
1150 if (llvm::StructType *SrcSTy = dyn_cast<llvm::StructType>(SrcTy)) {
1151 Src = EnterStructPointerForCoercedAccess(Src, SrcSTy, DstSize, CGF);
1152 SrcTy = Src.getType()->getElementType();
1155 uint64_t SrcSize = CGF.CGM.getDataLayout().getTypeAllocSize(SrcTy);
1157 // If the source and destination are integer or pointer types, just do an
1158 // extension or truncation to the desired type.
1159 if ((isa<llvm::IntegerType>(Ty) || isa<llvm::PointerType>(Ty)) &&
1160 (isa<llvm::IntegerType>(SrcTy) || isa<llvm::PointerType>(SrcTy))) {
1161 llvm::Value *Load = CGF.Builder.CreateLoad(Src);
1162 return CoerceIntOrPtrToIntOrPtr(Load, Ty, CGF);
1165 // If load is legal, just bitcast the src pointer.
1166 if (SrcSize >= DstSize) {
1167 // Generally SrcSize is never greater than DstSize, since this means we are
1168 // losing bits. However, this can happen in cases where the structure has
1169 // additional padding, for example due to a user specified alignment.
1171 // FIXME: Assert that we aren't truncating non-padding bits when have access
1172 // to that information.
1173 Src = CGF.Builder.CreateBitCast(Src, llvm::PointerType::getUnqual(Ty));
1174 return CGF.Builder.CreateLoad(Src);
1177 // Otherwise do coercion through memory. This is stupid, but simple.
1178 Address Tmp = CreateTempAllocaForCoercion(CGF, Ty, Src.getAlignment());
1179 Address Casted = CGF.Builder.CreateBitCast(Tmp, CGF.Int8PtrTy);
1180 Address SrcCasted = CGF.Builder.CreateBitCast(Src, CGF.Int8PtrTy);
1181 CGF.Builder.CreateMemCpy(Casted, SrcCasted,
1182 llvm::ConstantInt::get(CGF.IntPtrTy, SrcSize),
1184 return CGF.Builder.CreateLoad(Tmp);
1187 // Function to store a first-class aggregate into memory. We prefer to
1188 // store the elements rather than the aggregate to be more friendly to
1190 // FIXME: Do we need to recurse here?
1191 static void BuildAggStore(CodeGenFunction &CGF, llvm::Value *Val,
1192 Address Dest, bool DestIsVolatile) {
1193 // Prefer scalar stores to first-class aggregate stores.
1194 if (llvm::StructType *STy =
1195 dyn_cast<llvm::StructType>(Val->getType())) {
1196 const llvm::StructLayout *Layout =
1197 CGF.CGM.getDataLayout().getStructLayout(STy);
1199 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1200 auto EltOffset = CharUnits::fromQuantity(Layout->getElementOffset(i));
1201 Address EltPtr = CGF.Builder.CreateStructGEP(Dest, i, EltOffset);
1202 llvm::Value *Elt = CGF.Builder.CreateExtractValue(Val, i);
1203 CGF.Builder.CreateStore(Elt, EltPtr, DestIsVolatile);
1206 CGF.Builder.CreateStore(Val, Dest, DestIsVolatile);
1210 /// CreateCoercedStore - Create a store to \arg DstPtr from \arg Src,
1211 /// where the source and destination may have different types. The
1212 /// destination is known to be aligned to \arg DstAlign bytes.
1214 /// This safely handles the case when the src type is larger than the
1215 /// destination type; the upper bits of the src will be lost.
1216 static void CreateCoercedStore(llvm::Value *Src,
1219 CodeGenFunction &CGF) {
1220 llvm::Type *SrcTy = Src->getType();
1221 llvm::Type *DstTy = Dst.getType()->getElementType();
1222 if (SrcTy == DstTy) {
1223 CGF.Builder.CreateStore(Src, Dst, DstIsVolatile);
1227 uint64_t SrcSize = CGF.CGM.getDataLayout().getTypeAllocSize(SrcTy);
1229 if (llvm::StructType *DstSTy = dyn_cast<llvm::StructType>(DstTy)) {
1230 Dst = EnterStructPointerForCoercedAccess(Dst, DstSTy, SrcSize, CGF);
1231 DstTy = Dst.getType()->getElementType();
1234 // If the source and destination are integer or pointer types, just do an
1235 // extension or truncation to the desired type.
1236 if ((isa<llvm::IntegerType>(SrcTy) || isa<llvm::PointerType>(SrcTy)) &&
1237 (isa<llvm::IntegerType>(DstTy) || isa<llvm::PointerType>(DstTy))) {
1238 Src = CoerceIntOrPtrToIntOrPtr(Src, DstTy, CGF);
1239 CGF.Builder.CreateStore(Src, Dst, DstIsVolatile);
1243 uint64_t DstSize = CGF.CGM.getDataLayout().getTypeAllocSize(DstTy);
1245 // If store is legal, just bitcast the src pointer.
1246 if (SrcSize <= DstSize) {
1247 Dst = CGF.Builder.CreateBitCast(Dst, llvm::PointerType::getUnqual(SrcTy));
1248 BuildAggStore(CGF, Src, Dst, DstIsVolatile);
1250 // Otherwise do coercion through memory. This is stupid, but
1253 // Generally SrcSize is never greater than DstSize, since this means we are
1254 // losing bits. However, this can happen in cases where the structure has
1255 // additional padding, for example due to a user specified alignment.
1257 // FIXME: Assert that we aren't truncating non-padding bits when have access
1258 // to that information.
1259 Address Tmp = CreateTempAllocaForCoercion(CGF, SrcTy, Dst.getAlignment());
1260 CGF.Builder.CreateStore(Src, Tmp);
1261 Address Casted = CGF.Builder.CreateBitCast(Tmp, CGF.Int8PtrTy);
1262 Address DstCasted = CGF.Builder.CreateBitCast(Dst, CGF.Int8PtrTy);
1263 CGF.Builder.CreateMemCpy(DstCasted, Casted,
1264 llvm::ConstantInt::get(CGF.IntPtrTy, DstSize),
1269 static Address emitAddressAtOffset(CodeGenFunction &CGF, Address addr,
1270 const ABIArgInfo &info) {
1271 if (unsigned offset = info.getDirectOffset()) {
1272 addr = CGF.Builder.CreateElementBitCast(addr, CGF.Int8Ty);
1273 addr = CGF.Builder.CreateConstInBoundsByteGEP(addr,
1274 CharUnits::fromQuantity(offset));
1275 addr = CGF.Builder.CreateElementBitCast(addr, info.getCoerceToType());
1282 /// Encapsulates information about the way function arguments from
1283 /// CGFunctionInfo should be passed to actual LLVM IR function.
1284 class ClangToLLVMArgMapping {
1285 static const unsigned InvalidIndex = ~0U;
1286 unsigned InallocaArgNo;
1288 unsigned TotalIRArgs;
1290 /// Arguments of LLVM IR function corresponding to single Clang argument.
1292 unsigned PaddingArgIndex;
1293 // Argument is expanded to IR arguments at positions
1294 // [FirstArgIndex, FirstArgIndex + NumberOfArgs).
1295 unsigned FirstArgIndex;
1296 unsigned NumberOfArgs;
1299 : PaddingArgIndex(InvalidIndex), FirstArgIndex(InvalidIndex),
1303 SmallVector<IRArgs, 8> ArgInfo;
1306 ClangToLLVMArgMapping(const ASTContext &Context, const CGFunctionInfo &FI,
1307 bool OnlyRequiredArgs = false)
1308 : InallocaArgNo(InvalidIndex), SRetArgNo(InvalidIndex), TotalIRArgs(0),
1309 ArgInfo(OnlyRequiredArgs ? FI.getNumRequiredArgs() : FI.arg_size()) {
1310 construct(Context, FI, OnlyRequiredArgs);
1313 bool hasInallocaArg() const { return InallocaArgNo != InvalidIndex; }
1314 unsigned getInallocaArgNo() const {
1315 assert(hasInallocaArg());
1316 return InallocaArgNo;
1319 bool hasSRetArg() const { return SRetArgNo != InvalidIndex; }
1320 unsigned getSRetArgNo() const {
1321 assert(hasSRetArg());
1325 unsigned totalIRArgs() const { return TotalIRArgs; }
1327 bool hasPaddingArg(unsigned ArgNo) const {
1328 assert(ArgNo < ArgInfo.size());
1329 return ArgInfo[ArgNo].PaddingArgIndex != InvalidIndex;
1331 unsigned getPaddingArgNo(unsigned ArgNo) const {
1332 assert(hasPaddingArg(ArgNo));
1333 return ArgInfo[ArgNo].PaddingArgIndex;
1336 /// Returns index of first IR argument corresponding to ArgNo, and their
1338 std::pair<unsigned, unsigned> getIRArgs(unsigned ArgNo) const {
1339 assert(ArgNo < ArgInfo.size());
1340 return std::make_pair(ArgInfo[ArgNo].FirstArgIndex,
1341 ArgInfo[ArgNo].NumberOfArgs);
1345 void construct(const ASTContext &Context, const CGFunctionInfo &FI,
1346 bool OnlyRequiredArgs);
1349 void ClangToLLVMArgMapping::construct(const ASTContext &Context,
1350 const CGFunctionInfo &FI,
1351 bool OnlyRequiredArgs) {
1352 unsigned IRArgNo = 0;
1353 bool SwapThisWithSRet = false;
1354 const ABIArgInfo &RetAI = FI.getReturnInfo();
1356 if (RetAI.getKind() == ABIArgInfo::Indirect) {
1357 SwapThisWithSRet = RetAI.isSRetAfterThis();
1358 SRetArgNo = SwapThisWithSRet ? 1 : IRArgNo++;
1362 unsigned NumArgs = OnlyRequiredArgs ? FI.getNumRequiredArgs() : FI.arg_size();
1363 for (CGFunctionInfo::const_arg_iterator I = FI.arg_begin(); ArgNo < NumArgs;
1365 assert(I != FI.arg_end());
1366 QualType ArgType = I->type;
1367 const ABIArgInfo &AI = I->info;
1368 // Collect data about IR arguments corresponding to Clang argument ArgNo.
1369 auto &IRArgs = ArgInfo[ArgNo];
1371 if (AI.getPaddingType())
1372 IRArgs.PaddingArgIndex = IRArgNo++;
1374 switch (AI.getKind()) {
1375 case ABIArgInfo::Extend:
1376 case ABIArgInfo::Direct: {
1377 // FIXME: handle sseregparm someday...
1378 llvm::StructType *STy = dyn_cast<llvm::StructType>(AI.getCoerceToType());
1379 if (AI.isDirect() && AI.getCanBeFlattened() && STy) {
1380 IRArgs.NumberOfArgs = STy->getNumElements();
1382 IRArgs.NumberOfArgs = 1;
1386 case ABIArgInfo::Indirect:
1387 IRArgs.NumberOfArgs = 1;
1389 case ABIArgInfo::Ignore:
1390 case ABIArgInfo::InAlloca:
1391 // ignore and inalloca doesn't have matching LLVM parameters.
1392 IRArgs.NumberOfArgs = 0;
1394 case ABIArgInfo::CoerceAndExpand:
1395 IRArgs.NumberOfArgs = AI.getCoerceAndExpandTypeSequence().size();
1397 case ABIArgInfo::Expand:
1398 IRArgs.NumberOfArgs = getExpansionSize(ArgType, Context);
1402 if (IRArgs.NumberOfArgs > 0) {
1403 IRArgs.FirstArgIndex = IRArgNo;
1404 IRArgNo += IRArgs.NumberOfArgs;
1407 // Skip over the sret parameter when it comes second. We already handled it
1409 if (IRArgNo == 1 && SwapThisWithSRet)
1412 assert(ArgNo == ArgInfo.size());
1414 if (FI.usesInAlloca())
1415 InallocaArgNo = IRArgNo++;
1417 TotalIRArgs = IRArgNo;
1423 bool CodeGenModule::ReturnTypeUsesSRet(const CGFunctionInfo &FI) {
1424 return FI.getReturnInfo().isIndirect();
1427 bool CodeGenModule::ReturnSlotInterferesWithArgs(const CGFunctionInfo &FI) {
1428 return ReturnTypeUsesSRet(FI) &&
1429 getTargetCodeGenInfo().doesReturnSlotInterfereWithArgs();
1432 bool CodeGenModule::ReturnTypeUsesFPRet(QualType ResultType) {
1433 if (const BuiltinType *BT = ResultType->getAs<BuiltinType>()) {
1434 switch (BT->getKind()) {
1437 case BuiltinType::Float:
1438 return getTarget().useObjCFPRetForRealType(TargetInfo::Float);
1439 case BuiltinType::Double:
1440 return getTarget().useObjCFPRetForRealType(TargetInfo::Double);
1441 case BuiltinType::LongDouble:
1442 return getTarget().useObjCFPRetForRealType(TargetInfo::LongDouble);
1449 bool CodeGenModule::ReturnTypeUsesFP2Ret(QualType ResultType) {
1450 if (const ComplexType *CT = ResultType->getAs<ComplexType>()) {
1451 if (const BuiltinType *BT = CT->getElementType()->getAs<BuiltinType>()) {
1452 if (BT->getKind() == BuiltinType::LongDouble)
1453 return getTarget().useObjCFP2RetForComplexLongDouble();
1460 llvm::FunctionType *CodeGenTypes::GetFunctionType(GlobalDecl GD) {
1461 const CGFunctionInfo &FI = arrangeGlobalDeclaration(GD);
1462 return GetFunctionType(FI);
1465 llvm::FunctionType *
1466 CodeGenTypes::GetFunctionType(const CGFunctionInfo &FI) {
1468 bool Inserted = FunctionsBeingProcessed.insert(&FI).second;
1470 assert(Inserted && "Recursively being processed?");
1472 llvm::Type *resultType = nullptr;
1473 const ABIArgInfo &retAI = FI.getReturnInfo();
1474 switch (retAI.getKind()) {
1475 case ABIArgInfo::Expand:
1476 llvm_unreachable("Invalid ABI kind for return argument");
1478 case ABIArgInfo::Extend:
1479 case ABIArgInfo::Direct:
1480 resultType = retAI.getCoerceToType();
1483 case ABIArgInfo::InAlloca:
1484 if (retAI.getInAllocaSRet()) {
1485 // sret things on win32 aren't void, they return the sret pointer.
1486 QualType ret = FI.getReturnType();
1487 llvm::Type *ty = ConvertType(ret);
1488 unsigned addressSpace = Context.getTargetAddressSpace(ret);
1489 resultType = llvm::PointerType::get(ty, addressSpace);
1491 resultType = llvm::Type::getVoidTy(getLLVMContext());
1495 case ABIArgInfo::Indirect:
1496 case ABIArgInfo::Ignore:
1497 resultType = llvm::Type::getVoidTy(getLLVMContext());
1500 case ABIArgInfo::CoerceAndExpand:
1501 resultType = retAI.getUnpaddedCoerceAndExpandType();
1505 ClangToLLVMArgMapping IRFunctionArgs(getContext(), FI, true);
1506 SmallVector<llvm::Type*, 8> ArgTypes(IRFunctionArgs.totalIRArgs());
1508 // Add type for sret argument.
1509 if (IRFunctionArgs.hasSRetArg()) {
1510 QualType Ret = FI.getReturnType();
1511 llvm::Type *Ty = ConvertType(Ret);
1512 unsigned AddressSpace = Context.getTargetAddressSpace(Ret);
1513 ArgTypes[IRFunctionArgs.getSRetArgNo()] =
1514 llvm::PointerType::get(Ty, AddressSpace);
1517 // Add type for inalloca argument.
1518 if (IRFunctionArgs.hasInallocaArg()) {
1519 auto ArgStruct = FI.getArgStruct();
1521 ArgTypes[IRFunctionArgs.getInallocaArgNo()] = ArgStruct->getPointerTo();
1524 // Add in all of the required arguments.
1526 CGFunctionInfo::const_arg_iterator it = FI.arg_begin(),
1527 ie = it + FI.getNumRequiredArgs();
1528 for (; it != ie; ++it, ++ArgNo) {
1529 const ABIArgInfo &ArgInfo = it->info;
1531 // Insert a padding type to ensure proper alignment.
1532 if (IRFunctionArgs.hasPaddingArg(ArgNo))
1533 ArgTypes[IRFunctionArgs.getPaddingArgNo(ArgNo)] =
1534 ArgInfo.getPaddingType();
1536 unsigned FirstIRArg, NumIRArgs;
1537 std::tie(FirstIRArg, NumIRArgs) = IRFunctionArgs.getIRArgs(ArgNo);
1539 switch (ArgInfo.getKind()) {
1540 case ABIArgInfo::Ignore:
1541 case ABIArgInfo::InAlloca:
1542 assert(NumIRArgs == 0);
1545 case ABIArgInfo::Indirect: {
1546 assert(NumIRArgs == 1);
1547 // indirect arguments are always on the stack, which is addr space #0.
1548 llvm::Type *LTy = ConvertTypeForMem(it->type);
1549 ArgTypes[FirstIRArg] = LTy->getPointerTo();
1553 case ABIArgInfo::Extend:
1554 case ABIArgInfo::Direct: {
1555 // Fast-isel and the optimizer generally like scalar values better than
1556 // FCAs, so we flatten them if this is safe to do for this argument.
1557 llvm::Type *argType = ArgInfo.getCoerceToType();
1558 llvm::StructType *st = dyn_cast<llvm::StructType>(argType);
1559 if (st && ArgInfo.isDirect() && ArgInfo.getCanBeFlattened()) {
1560 assert(NumIRArgs == st->getNumElements());
1561 for (unsigned i = 0, e = st->getNumElements(); i != e; ++i)
1562 ArgTypes[FirstIRArg + i] = st->getElementType(i);
1564 assert(NumIRArgs == 1);
1565 ArgTypes[FirstIRArg] = argType;
1570 case ABIArgInfo::CoerceAndExpand: {
1571 auto ArgTypesIter = ArgTypes.begin() + FirstIRArg;
1572 for (auto EltTy : ArgInfo.getCoerceAndExpandTypeSequence()) {
1573 *ArgTypesIter++ = EltTy;
1575 assert(ArgTypesIter == ArgTypes.begin() + FirstIRArg + NumIRArgs);
1579 case ABIArgInfo::Expand:
1580 auto ArgTypesIter = ArgTypes.begin() + FirstIRArg;
1581 getExpandedTypes(it->type, ArgTypesIter);
1582 assert(ArgTypesIter == ArgTypes.begin() + FirstIRArg + NumIRArgs);
1587 bool Erased = FunctionsBeingProcessed.erase(&FI); (void)Erased;
1588 assert(Erased && "Not in set?");
1590 return llvm::FunctionType::get(resultType, ArgTypes, FI.isVariadic());
1593 llvm::Type *CodeGenTypes::GetFunctionTypeForVTable(GlobalDecl GD) {
1594 const CXXMethodDecl *MD = cast<CXXMethodDecl>(GD.getDecl());
1595 const FunctionProtoType *FPT = MD->getType()->getAs<FunctionProtoType>();
1597 if (!isFuncTypeConvertible(FPT))
1598 return llvm::StructType::get(getLLVMContext());
1600 const CGFunctionInfo *Info;
1601 if (isa<CXXDestructorDecl>(MD))
1603 &arrangeCXXStructorDeclaration(MD, getFromDtorType(GD.getDtorType()));
1605 Info = &arrangeCXXMethodDeclaration(MD);
1606 return GetFunctionType(*Info);
1609 static void AddAttributesFromFunctionProtoType(ASTContext &Ctx,
1610 llvm::AttrBuilder &FuncAttrs,
1611 const FunctionProtoType *FPT) {
1615 if (!isUnresolvedExceptionSpec(FPT->getExceptionSpecType()) &&
1616 FPT->isNothrow(Ctx))
1617 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1620 void CodeGenModule::ConstructAttributeList(
1621 StringRef Name, const CGFunctionInfo &FI, CGCalleeInfo CalleeInfo,
1622 AttributeListType &PAL, unsigned &CallingConv, bool AttrOnCallSite) {
1623 llvm::AttrBuilder FuncAttrs;
1624 llvm::AttrBuilder RetAttrs;
1625 bool HasOptnone = false;
1627 CallingConv = FI.getEffectiveCallingConvention();
1629 if (FI.isNoReturn())
1630 FuncAttrs.addAttribute(llvm::Attribute::NoReturn);
1632 // If we have information about the function prototype, we can learn
1633 // attributes form there.
1634 AddAttributesFromFunctionProtoType(getContext(), FuncAttrs,
1635 CalleeInfo.getCalleeFunctionProtoType());
1637 const Decl *TargetDecl = CalleeInfo.getCalleeDecl();
1639 bool HasAnyX86InterruptAttr = false;
1640 // FIXME: handle sseregparm someday...
1642 if (TargetDecl->hasAttr<ReturnsTwiceAttr>())
1643 FuncAttrs.addAttribute(llvm::Attribute::ReturnsTwice);
1644 if (TargetDecl->hasAttr<NoThrowAttr>())
1645 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1646 if (TargetDecl->hasAttr<NoReturnAttr>())
1647 FuncAttrs.addAttribute(llvm::Attribute::NoReturn);
1648 if (TargetDecl->hasAttr<NoDuplicateAttr>())
1649 FuncAttrs.addAttribute(llvm::Attribute::NoDuplicate);
1651 if (const FunctionDecl *Fn = dyn_cast<FunctionDecl>(TargetDecl)) {
1652 AddAttributesFromFunctionProtoType(
1653 getContext(), FuncAttrs, Fn->getType()->getAs<FunctionProtoType>());
1654 // Don't use [[noreturn]] or _Noreturn for a call to a virtual function.
1655 // These attributes are not inherited by overloads.
1656 const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(Fn);
1657 if (Fn->isNoReturn() && !(AttrOnCallSite && MD && MD->isVirtual()))
1658 FuncAttrs.addAttribute(llvm::Attribute::NoReturn);
1661 // 'const', 'pure' and 'noalias' attributed functions are also nounwind.
1662 if (TargetDecl->hasAttr<ConstAttr>()) {
1663 FuncAttrs.addAttribute(llvm::Attribute::ReadNone);
1664 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1665 } else if (TargetDecl->hasAttr<PureAttr>()) {
1666 FuncAttrs.addAttribute(llvm::Attribute::ReadOnly);
1667 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1668 } else if (TargetDecl->hasAttr<NoAliasAttr>()) {
1669 FuncAttrs.addAttribute(llvm::Attribute::ArgMemOnly);
1670 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1672 if (TargetDecl->hasAttr<RestrictAttr>())
1673 RetAttrs.addAttribute(llvm::Attribute::NoAlias);
1674 if (TargetDecl->hasAttr<ReturnsNonNullAttr>())
1675 RetAttrs.addAttribute(llvm::Attribute::NonNull);
1677 HasAnyX86InterruptAttr = TargetDecl->hasAttr<AnyX86InterruptAttr>();
1678 HasOptnone = TargetDecl->hasAttr<OptimizeNoneAttr>();
1681 // OptimizeNoneAttr takes precedence over -Os or -Oz. No warning needed.
1683 if (CodeGenOpts.OptimizeSize)
1684 FuncAttrs.addAttribute(llvm::Attribute::OptimizeForSize);
1685 if (CodeGenOpts.OptimizeSize == 2)
1686 FuncAttrs.addAttribute(llvm::Attribute::MinSize);
1689 if (CodeGenOpts.DisableRedZone)
1690 FuncAttrs.addAttribute(llvm::Attribute::NoRedZone);
1691 if (CodeGenOpts.NoImplicitFloat)
1692 FuncAttrs.addAttribute(llvm::Attribute::NoImplicitFloat);
1693 if (CodeGenOpts.EnableSegmentedStacks &&
1694 !(TargetDecl && TargetDecl->hasAttr<NoSplitStackAttr>()))
1695 FuncAttrs.addAttribute("split-stack");
1697 if (AttrOnCallSite) {
1698 // Attributes that should go on the call site only.
1699 if (!CodeGenOpts.SimplifyLibCalls ||
1700 CodeGenOpts.isNoBuiltinFunc(Name.data()))
1701 FuncAttrs.addAttribute(llvm::Attribute::NoBuiltin);
1702 if (!CodeGenOpts.TrapFuncName.empty())
1703 FuncAttrs.addAttribute("trap-func-name", CodeGenOpts.TrapFuncName);
1705 // Attributes that should go on the function, but not the call site.
1706 if (!CodeGenOpts.DisableFPElim) {
1707 FuncAttrs.addAttribute("no-frame-pointer-elim", "false");
1708 } else if (CodeGenOpts.OmitLeafFramePointer) {
1709 FuncAttrs.addAttribute("no-frame-pointer-elim", "false");
1710 FuncAttrs.addAttribute("no-frame-pointer-elim-non-leaf");
1712 FuncAttrs.addAttribute("no-frame-pointer-elim", "true");
1713 FuncAttrs.addAttribute("no-frame-pointer-elim-non-leaf");
1716 bool DisableTailCalls =
1717 CodeGenOpts.DisableTailCalls || HasAnyX86InterruptAttr ||
1718 (TargetDecl && TargetDecl->hasAttr<DisableTailCallsAttr>());
1719 FuncAttrs.addAttribute(
1720 "disable-tail-calls",
1721 llvm::toStringRef(DisableTailCalls));
1723 FuncAttrs.addAttribute("less-precise-fpmad",
1724 llvm::toStringRef(CodeGenOpts.LessPreciseFPMAD));
1725 FuncAttrs.addAttribute("no-infs-fp-math",
1726 llvm::toStringRef(CodeGenOpts.NoInfsFPMath));
1727 FuncAttrs.addAttribute("no-nans-fp-math",
1728 llvm::toStringRef(CodeGenOpts.NoNaNsFPMath));
1729 FuncAttrs.addAttribute("unsafe-fp-math",
1730 llvm::toStringRef(CodeGenOpts.UnsafeFPMath));
1731 FuncAttrs.addAttribute("use-soft-float",
1732 llvm::toStringRef(CodeGenOpts.SoftFloat));
1733 FuncAttrs.addAttribute("stack-protector-buffer-size",
1734 llvm::utostr(CodeGenOpts.SSPBufferSize));
1735 FuncAttrs.addAttribute("no-signed-zeros-fp-math",
1736 llvm::toStringRef(CodeGenOpts.NoSignedZeros));
1738 if (CodeGenOpts.StackRealignment)
1739 FuncAttrs.addAttribute("stackrealign");
1740 if (CodeGenOpts.Backchain)
1741 FuncAttrs.addAttribute("backchain");
1743 // Add target-cpu and target-features attributes to functions. If
1744 // we have a decl for the function and it has a target attribute then
1745 // parse that and add it to the feature set.
1746 StringRef TargetCPU = getTarget().getTargetOpts().CPU;
1747 const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(TargetDecl);
1748 if (FD && FD->hasAttr<TargetAttr>()) {
1749 llvm::StringMap<bool> FeatureMap;
1750 getFunctionFeatureMap(FeatureMap, FD);
1752 // Produce the canonical string for this set of features.
1753 std::vector<std::string> Features;
1754 for (llvm::StringMap<bool>::const_iterator it = FeatureMap.begin(),
1755 ie = FeatureMap.end();
1757 Features.push_back((it->second ? "+" : "-") + it->first().str());
1759 // Now add the target-cpu and target-features to the function.
1760 // While we populated the feature map above, we still need to
1761 // get and parse the target attribute so we can get the cpu for
1763 const auto *TD = FD->getAttr<TargetAttr>();
1764 TargetAttr::ParsedTargetAttr ParsedAttr = TD->parse();
1765 if (ParsedAttr.second != "")
1766 TargetCPU = ParsedAttr.second;
1767 if (TargetCPU != "")
1768 FuncAttrs.addAttribute("target-cpu", TargetCPU);
1769 if (!Features.empty()) {
1770 std::sort(Features.begin(), Features.end());
1771 FuncAttrs.addAttribute(
1773 llvm::join(Features.begin(), Features.end(), ","));
1776 // Otherwise just add the existing target cpu and target features to the
1778 std::vector<std::string> &Features = getTarget().getTargetOpts().Features;
1779 if (TargetCPU != "")
1780 FuncAttrs.addAttribute("target-cpu", TargetCPU);
1781 if (!Features.empty()) {
1782 std::sort(Features.begin(), Features.end());
1783 FuncAttrs.addAttribute(
1785 llvm::join(Features.begin(), Features.end(), ","));
1790 if (getLangOpts().CUDA && getLangOpts().CUDAIsDevice) {
1791 // Conservatively, mark all functions and calls in CUDA as convergent
1792 // (meaning, they may call an intrinsically convergent op, such as
1793 // __syncthreads(), and so can't have certain optimizations applied around
1794 // them). LLVM will remove this attribute where it safely can.
1795 FuncAttrs.addAttribute(llvm::Attribute::Convergent);
1797 // Respect -fcuda-flush-denormals-to-zero.
1798 if (getLangOpts().CUDADeviceFlushDenormalsToZero)
1799 FuncAttrs.addAttribute("nvptx-f32ftz", "true");
1802 ClangToLLVMArgMapping IRFunctionArgs(getContext(), FI);
1804 QualType RetTy = FI.getReturnType();
1805 const ABIArgInfo &RetAI = FI.getReturnInfo();
1806 switch (RetAI.getKind()) {
1807 case ABIArgInfo::Extend:
1808 if (RetTy->hasSignedIntegerRepresentation())
1809 RetAttrs.addAttribute(llvm::Attribute::SExt);
1810 else if (RetTy->hasUnsignedIntegerRepresentation())
1811 RetAttrs.addAttribute(llvm::Attribute::ZExt);
1813 case ABIArgInfo::Direct:
1814 if (RetAI.getInReg())
1815 RetAttrs.addAttribute(llvm::Attribute::InReg);
1817 case ABIArgInfo::Ignore:
1820 case ABIArgInfo::InAlloca:
1821 case ABIArgInfo::Indirect: {
1822 // inalloca and sret disable readnone and readonly
1823 FuncAttrs.removeAttribute(llvm::Attribute::ReadOnly)
1824 .removeAttribute(llvm::Attribute::ReadNone);
1828 case ABIArgInfo::CoerceAndExpand:
1831 case ABIArgInfo::Expand:
1832 llvm_unreachable("Invalid ABI kind for return argument");
1835 if (const auto *RefTy = RetTy->getAs<ReferenceType>()) {
1836 QualType PTy = RefTy->getPointeeType();
1837 if (!PTy->isIncompleteType() && PTy->isConstantSizeType())
1838 RetAttrs.addDereferenceableAttr(getContext().getTypeSizeInChars(PTy)
1840 else if (getContext().getTargetAddressSpace(PTy) == 0)
1841 RetAttrs.addAttribute(llvm::Attribute::NonNull);
1844 // Attach return attributes.
1845 if (RetAttrs.hasAttributes()) {
1846 PAL.push_back(llvm::AttributeSet::get(
1847 getLLVMContext(), llvm::AttributeSet::ReturnIndex, RetAttrs));
1850 bool hasUsedSRet = false;
1852 // Attach attributes to sret.
1853 if (IRFunctionArgs.hasSRetArg()) {
1854 llvm::AttrBuilder SRETAttrs;
1855 SRETAttrs.addAttribute(llvm::Attribute::StructRet);
1857 if (RetAI.getInReg())
1858 SRETAttrs.addAttribute(llvm::Attribute::InReg);
1859 PAL.push_back(llvm::AttributeSet::get(
1860 getLLVMContext(), IRFunctionArgs.getSRetArgNo() + 1, SRETAttrs));
1863 // Attach attributes to inalloca argument.
1864 if (IRFunctionArgs.hasInallocaArg()) {
1865 llvm::AttrBuilder Attrs;
1866 Attrs.addAttribute(llvm::Attribute::InAlloca);
1867 PAL.push_back(llvm::AttributeSet::get(
1868 getLLVMContext(), IRFunctionArgs.getInallocaArgNo() + 1, Attrs));
1872 for (CGFunctionInfo::const_arg_iterator I = FI.arg_begin(),
1874 I != E; ++I, ++ArgNo) {
1875 QualType ParamType = I->type;
1876 const ABIArgInfo &AI = I->info;
1877 llvm::AttrBuilder Attrs;
1879 // Add attribute for padding argument, if necessary.
1880 if (IRFunctionArgs.hasPaddingArg(ArgNo)) {
1881 if (AI.getPaddingInReg())
1882 PAL.push_back(llvm::AttributeSet::get(
1883 getLLVMContext(), IRFunctionArgs.getPaddingArgNo(ArgNo) + 1,
1884 llvm::Attribute::InReg));
1887 // 'restrict' -> 'noalias' is done in EmitFunctionProlog when we
1888 // have the corresponding parameter variable. It doesn't make
1889 // sense to do it here because parameters are so messed up.
1890 switch (AI.getKind()) {
1891 case ABIArgInfo::Extend:
1892 if (ParamType->isSignedIntegerOrEnumerationType())
1893 Attrs.addAttribute(llvm::Attribute::SExt);
1894 else if (ParamType->isUnsignedIntegerOrEnumerationType()) {
1895 if (getTypes().getABIInfo().shouldSignExtUnsignedType(ParamType))
1896 Attrs.addAttribute(llvm::Attribute::SExt);
1898 Attrs.addAttribute(llvm::Attribute::ZExt);
1901 case ABIArgInfo::Direct:
1902 if (ArgNo == 0 && FI.isChainCall())
1903 Attrs.addAttribute(llvm::Attribute::Nest);
1904 else if (AI.getInReg())
1905 Attrs.addAttribute(llvm::Attribute::InReg);
1908 case ABIArgInfo::Indirect: {
1910 Attrs.addAttribute(llvm::Attribute::InReg);
1912 if (AI.getIndirectByVal())
1913 Attrs.addAttribute(llvm::Attribute::ByVal);
1915 CharUnits Align = AI.getIndirectAlign();
1917 // In a byval argument, it is important that the required
1918 // alignment of the type is honored, as LLVM might be creating a
1919 // *new* stack object, and needs to know what alignment to give
1920 // it. (Sometimes it can deduce a sensible alignment on its own,
1921 // but not if clang decides it must emit a packed struct, or the
1922 // user specifies increased alignment requirements.)
1924 // This is different from indirect *not* byval, where the object
1925 // exists already, and the align attribute is purely
1927 assert(!Align.isZero());
1929 // For now, only add this when we have a byval argument.
1930 // TODO: be less lazy about updating test cases.
1931 if (AI.getIndirectByVal())
1932 Attrs.addAlignmentAttr(Align.getQuantity());
1934 // byval disables readnone and readonly.
1935 FuncAttrs.removeAttribute(llvm::Attribute::ReadOnly)
1936 .removeAttribute(llvm::Attribute::ReadNone);
1939 case ABIArgInfo::Ignore:
1940 case ABIArgInfo::Expand:
1941 case ABIArgInfo::CoerceAndExpand:
1944 case ABIArgInfo::InAlloca:
1945 // inalloca disables readnone and readonly.
1946 FuncAttrs.removeAttribute(llvm::Attribute::ReadOnly)
1947 .removeAttribute(llvm::Attribute::ReadNone);
1951 if (const auto *RefTy = ParamType->getAs<ReferenceType>()) {
1952 QualType PTy = RefTy->getPointeeType();
1953 if (!PTy->isIncompleteType() && PTy->isConstantSizeType())
1954 Attrs.addDereferenceableAttr(getContext().getTypeSizeInChars(PTy)
1956 else if (getContext().getTargetAddressSpace(PTy) == 0)
1957 Attrs.addAttribute(llvm::Attribute::NonNull);
1960 switch (FI.getExtParameterInfo(ArgNo).getABI()) {
1961 case ParameterABI::Ordinary:
1964 case ParameterABI::SwiftIndirectResult: {
1965 // Add 'sret' if we haven't already used it for something, but
1966 // only if the result is void.
1967 if (!hasUsedSRet && RetTy->isVoidType()) {
1968 Attrs.addAttribute(llvm::Attribute::StructRet);
1972 // Add 'noalias' in either case.
1973 Attrs.addAttribute(llvm::Attribute::NoAlias);
1975 // Add 'dereferenceable' and 'alignment'.
1976 auto PTy = ParamType->getPointeeType();
1977 if (!PTy->isIncompleteType() && PTy->isConstantSizeType()) {
1978 auto info = getContext().getTypeInfoInChars(PTy);
1979 Attrs.addDereferenceableAttr(info.first.getQuantity());
1980 Attrs.addAttribute(llvm::Attribute::getWithAlignment(getLLVMContext(),
1981 info.second.getQuantity()));
1986 case ParameterABI::SwiftErrorResult:
1987 Attrs.addAttribute(llvm::Attribute::SwiftError);
1990 case ParameterABI::SwiftContext:
1991 Attrs.addAttribute(llvm::Attribute::SwiftSelf);
1995 if (Attrs.hasAttributes()) {
1996 unsigned FirstIRArg, NumIRArgs;
1997 std::tie(FirstIRArg, NumIRArgs) = IRFunctionArgs.getIRArgs(ArgNo);
1998 for (unsigned i = 0; i < NumIRArgs; i++)
1999 PAL.push_back(llvm::AttributeSet::get(getLLVMContext(),
2000 FirstIRArg + i + 1, Attrs));
2003 assert(ArgNo == FI.arg_size());
2005 if (FuncAttrs.hasAttributes())
2006 PAL.push_back(llvm::
2007 AttributeSet::get(getLLVMContext(),
2008 llvm::AttributeSet::FunctionIndex,
2012 /// An argument came in as a promoted argument; demote it back to its
2014 static llvm::Value *emitArgumentDemotion(CodeGenFunction &CGF,
2016 llvm::Value *value) {
2017 llvm::Type *varType = CGF.ConvertType(var->getType());
2019 // This can happen with promotions that actually don't change the
2020 // underlying type, like the enum promotions.
2021 if (value->getType() == varType) return value;
2023 assert((varType->isIntegerTy() || varType->isFloatingPointTy())
2024 && "unexpected promotion type");
2026 if (isa<llvm::IntegerType>(varType))
2027 return CGF.Builder.CreateTrunc(value, varType, "arg.unpromote");
2029 return CGF.Builder.CreateFPCast(value, varType, "arg.unpromote");
2032 /// Returns the attribute (either parameter attribute, or function
2033 /// attribute), which declares argument ArgNo to be non-null.
2034 static const NonNullAttr *getNonNullAttr(const Decl *FD, const ParmVarDecl *PVD,
2035 QualType ArgType, unsigned ArgNo) {
2036 // FIXME: __attribute__((nonnull)) can also be applied to:
2037 // - references to pointers, where the pointee is known to be
2038 // nonnull (apparently a Clang extension)
2039 // - transparent unions containing pointers
2040 // In the former case, LLVM IR cannot represent the constraint. In
2041 // the latter case, we have no guarantee that the transparent union
2042 // is in fact passed as a pointer.
2043 if (!ArgType->isAnyPointerType() && !ArgType->isBlockPointerType())
2045 // First, check attribute on parameter itself.
2047 if (auto ParmNNAttr = PVD->getAttr<NonNullAttr>())
2050 // Check function attributes.
2053 for (const auto *NNAttr : FD->specific_attrs<NonNullAttr>()) {
2054 if (NNAttr->isNonNull(ArgNo))
2061 struct CopyBackSwiftError final : EHScopeStack::Cleanup {
2064 CopyBackSwiftError(Address temp, Address arg) : Temp(temp), Arg(arg) {}
2065 void Emit(CodeGenFunction &CGF, Flags flags) override {
2066 llvm::Value *errorValue = CGF.Builder.CreateLoad(Temp);
2067 CGF.Builder.CreateStore(errorValue, Arg);
2072 void CodeGenFunction::EmitFunctionProlog(const CGFunctionInfo &FI,
2074 const FunctionArgList &Args) {
2075 if (CurCodeDecl && CurCodeDecl->hasAttr<NakedAttr>())
2076 // Naked functions don't have prologues.
2079 // If this is an implicit-return-zero function, go ahead and
2080 // initialize the return value. TODO: it might be nice to have
2081 // a more general mechanism for this that didn't require synthesized
2082 // return statements.
2083 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(CurCodeDecl)) {
2084 if (FD->hasImplicitReturnZero()) {
2085 QualType RetTy = FD->getReturnType().getUnqualifiedType();
2086 llvm::Type* LLVMTy = CGM.getTypes().ConvertType(RetTy);
2087 llvm::Constant* Zero = llvm::Constant::getNullValue(LLVMTy);
2088 Builder.CreateStore(Zero, ReturnValue);
2092 // FIXME: We no longer need the types from FunctionArgList; lift up and
2095 ClangToLLVMArgMapping IRFunctionArgs(CGM.getContext(), FI);
2096 // Flattened function arguments.
2097 SmallVector<llvm::Value *, 16> FnArgs;
2098 FnArgs.reserve(IRFunctionArgs.totalIRArgs());
2099 for (auto &Arg : Fn->args()) {
2100 FnArgs.push_back(&Arg);
2102 assert(FnArgs.size() == IRFunctionArgs.totalIRArgs());
2104 // If we're using inalloca, all the memory arguments are GEPs off of the last
2105 // parameter, which is a pointer to the complete memory area.
2106 Address ArgStruct = Address::invalid();
2107 const llvm::StructLayout *ArgStructLayout = nullptr;
2108 if (IRFunctionArgs.hasInallocaArg()) {
2109 ArgStructLayout = CGM.getDataLayout().getStructLayout(FI.getArgStruct());
2110 ArgStruct = Address(FnArgs[IRFunctionArgs.getInallocaArgNo()],
2111 FI.getArgStructAlignment());
2113 assert(ArgStruct.getType() == FI.getArgStruct()->getPointerTo());
2116 // Name the struct return parameter.
2117 if (IRFunctionArgs.hasSRetArg()) {
2118 auto AI = cast<llvm::Argument>(FnArgs[IRFunctionArgs.getSRetArgNo()]);
2119 AI->setName("agg.result");
2120 AI->addAttr(llvm::AttributeSet::get(getLLVMContext(), AI->getArgNo() + 1,
2121 llvm::Attribute::NoAlias));
2124 // Track if we received the parameter as a pointer (indirect, byval, or
2125 // inalloca). If already have a pointer, EmitParmDecl doesn't need to copy it
2126 // into a local alloca for us.
2127 SmallVector<ParamValue, 16> ArgVals;
2128 ArgVals.reserve(Args.size());
2130 // Create a pointer value for every parameter declaration. This usually
2131 // entails copying one or more LLVM IR arguments into an alloca. Don't push
2132 // any cleanups or do anything that might unwind. We do that separately, so
2133 // we can push the cleanups in the correct order for the ABI.
2134 assert(FI.arg_size() == Args.size() &&
2135 "Mismatch between function signature & arguments.");
2137 CGFunctionInfo::const_arg_iterator info_it = FI.arg_begin();
2138 for (FunctionArgList::const_iterator i = Args.begin(), e = Args.end();
2139 i != e; ++i, ++info_it, ++ArgNo) {
2140 const VarDecl *Arg = *i;
2141 QualType Ty = info_it->type;
2142 const ABIArgInfo &ArgI = info_it->info;
2145 isa<ParmVarDecl>(Arg) && cast<ParmVarDecl>(Arg)->isKNRPromoted();
2147 unsigned FirstIRArg, NumIRArgs;
2148 std::tie(FirstIRArg, NumIRArgs) = IRFunctionArgs.getIRArgs(ArgNo);
2150 switch (ArgI.getKind()) {
2151 case ABIArgInfo::InAlloca: {
2152 assert(NumIRArgs == 0);
2153 auto FieldIndex = ArgI.getInAllocaFieldIndex();
2154 CharUnits FieldOffset =
2155 CharUnits::fromQuantity(ArgStructLayout->getElementOffset(FieldIndex));
2156 Address V = Builder.CreateStructGEP(ArgStruct, FieldIndex, FieldOffset,
2158 ArgVals.push_back(ParamValue::forIndirect(V));
2162 case ABIArgInfo::Indirect: {
2163 assert(NumIRArgs == 1);
2164 Address ParamAddr = Address(FnArgs[FirstIRArg], ArgI.getIndirectAlign());
2166 if (!hasScalarEvaluationKind(Ty)) {
2167 // Aggregates and complex variables are accessed by reference. All we
2168 // need to do is realign the value, if requested.
2169 Address V = ParamAddr;
2170 if (ArgI.getIndirectRealign()) {
2171 Address AlignedTemp = CreateMemTemp(Ty, "coerce");
2173 // Copy from the incoming argument pointer to the temporary with the
2174 // appropriate alignment.
2176 // FIXME: We should have a common utility for generating an aggregate
2178 CharUnits Size = getContext().getTypeSizeInChars(Ty);
2179 auto SizeVal = llvm::ConstantInt::get(IntPtrTy, Size.getQuantity());
2180 Address Dst = Builder.CreateBitCast(AlignedTemp, Int8PtrTy);
2181 Address Src = Builder.CreateBitCast(ParamAddr, Int8PtrTy);
2182 Builder.CreateMemCpy(Dst, Src, SizeVal, false);
2185 ArgVals.push_back(ParamValue::forIndirect(V));
2187 // Load scalar value from indirect argument.
2189 EmitLoadOfScalar(ParamAddr, false, Ty, Arg->getLocStart());
2192 V = emitArgumentDemotion(*this, Arg, V);
2193 ArgVals.push_back(ParamValue::forDirect(V));
2198 case ABIArgInfo::Extend:
2199 case ABIArgInfo::Direct: {
2201 // If we have the trivial case, handle it with no muss and fuss.
2202 if (!isa<llvm::StructType>(ArgI.getCoerceToType()) &&
2203 ArgI.getCoerceToType() == ConvertType(Ty) &&
2204 ArgI.getDirectOffset() == 0) {
2205 assert(NumIRArgs == 1);
2206 llvm::Value *V = FnArgs[FirstIRArg];
2207 auto AI = cast<llvm::Argument>(V);
2209 if (const ParmVarDecl *PVD = dyn_cast<ParmVarDecl>(Arg)) {
2210 if (getNonNullAttr(CurCodeDecl, PVD, PVD->getType(),
2211 PVD->getFunctionScopeIndex()))
2212 AI->addAttr(llvm::AttributeSet::get(getLLVMContext(),
2214 llvm::Attribute::NonNull));
2216 QualType OTy = PVD->getOriginalType();
2217 if (const auto *ArrTy =
2218 getContext().getAsConstantArrayType(OTy)) {
2219 // A C99 array parameter declaration with the static keyword also
2220 // indicates dereferenceability, and if the size is constant we can
2221 // use the dereferenceable attribute (which requires the size in
2223 if (ArrTy->getSizeModifier() == ArrayType::Static) {
2224 QualType ETy = ArrTy->getElementType();
2225 uint64_t ArrSize = ArrTy->getSize().getZExtValue();
2226 if (!ETy->isIncompleteType() && ETy->isConstantSizeType() &&
2228 llvm::AttrBuilder Attrs;
2229 Attrs.addDereferenceableAttr(
2230 getContext().getTypeSizeInChars(ETy).getQuantity()*ArrSize);
2231 AI->addAttr(llvm::AttributeSet::get(getLLVMContext(),
2232 AI->getArgNo() + 1, Attrs));
2233 } else if (getContext().getTargetAddressSpace(ETy) == 0) {
2234 AI->addAttr(llvm::AttributeSet::get(getLLVMContext(),
2236 llvm::Attribute::NonNull));
2239 } else if (const auto *ArrTy =
2240 getContext().getAsVariableArrayType(OTy)) {
2241 // For C99 VLAs with the static keyword, we don't know the size so
2242 // we can't use the dereferenceable attribute, but in addrspace(0)
2243 // we know that it must be nonnull.
2244 if (ArrTy->getSizeModifier() == VariableArrayType::Static &&
2245 !getContext().getTargetAddressSpace(ArrTy->getElementType()))
2246 AI->addAttr(llvm::AttributeSet::get(getLLVMContext(),
2248 llvm::Attribute::NonNull));
2251 const auto *AVAttr = PVD->getAttr<AlignValueAttr>();
2253 if (const auto *TOTy = dyn_cast<TypedefType>(OTy))
2254 AVAttr = TOTy->getDecl()->getAttr<AlignValueAttr>();
2256 llvm::Value *AlignmentValue =
2257 EmitScalarExpr(AVAttr->getAlignment());
2258 llvm::ConstantInt *AlignmentCI =
2259 cast<llvm::ConstantInt>(AlignmentValue);
2260 unsigned Alignment =
2261 std::min((unsigned) AlignmentCI->getZExtValue(),
2262 +llvm::Value::MaximumAlignment);
2264 llvm::AttrBuilder Attrs;
2265 Attrs.addAlignmentAttr(Alignment);
2266 AI->addAttr(llvm::AttributeSet::get(getLLVMContext(),
2267 AI->getArgNo() + 1, Attrs));
2271 if (Arg->getType().isRestrictQualified())
2272 AI->addAttr(llvm::AttributeSet::get(getLLVMContext(),
2274 llvm::Attribute::NoAlias));
2276 // LLVM expects swifterror parameters to be used in very restricted
2277 // ways. Copy the value into a less-restricted temporary.
2278 if (FI.getExtParameterInfo(ArgNo).getABI()
2279 == ParameterABI::SwiftErrorResult) {
2280 QualType pointeeTy = Ty->getPointeeType();
2281 assert(pointeeTy->isPointerType());
2283 CreateMemTemp(pointeeTy, getPointerAlign(), "swifterror.temp");
2284 Address arg = Address(V, getContext().getTypeAlignInChars(pointeeTy));
2285 llvm::Value *incomingErrorValue = Builder.CreateLoad(arg);
2286 Builder.CreateStore(incomingErrorValue, temp);
2287 V = temp.getPointer();
2289 // Push a cleanup to copy the value back at the end of the function.
2290 // The convention does not guarantee that the value will be written
2291 // back if the function exits with an unwind exception.
2292 EHStack.pushCleanup<CopyBackSwiftError>(NormalCleanup, temp, arg);
2295 // Ensure the argument is the correct type.
2296 if (V->getType() != ArgI.getCoerceToType())
2297 V = Builder.CreateBitCast(V, ArgI.getCoerceToType());
2300 V = emitArgumentDemotion(*this, Arg, V);
2302 if (const CXXMethodDecl *MD =
2303 dyn_cast_or_null<CXXMethodDecl>(CurCodeDecl)) {
2304 if (MD->isVirtual() && Arg == CXXABIThisDecl)
2305 V = CGM.getCXXABI().
2306 adjustThisParameterInVirtualFunctionPrologue(*this, CurGD, V);
2309 // Because of merging of function types from multiple decls it is
2310 // possible for the type of an argument to not match the corresponding
2311 // type in the function type. Since we are codegening the callee
2312 // in here, add a cast to the argument type.
2313 llvm::Type *LTy = ConvertType(Arg->getType());
2314 if (V->getType() != LTy)
2315 V = Builder.CreateBitCast(V, LTy);
2317 ArgVals.push_back(ParamValue::forDirect(V));
2321 Address Alloca = CreateMemTemp(Ty, getContext().getDeclAlign(Arg),
2324 // Pointer to store into.
2325 Address Ptr = emitAddressAtOffset(*this, Alloca, ArgI);
2327 // Fast-isel and the optimizer generally like scalar values better than
2328 // FCAs, so we flatten them if this is safe to do for this argument.
2329 llvm::StructType *STy = dyn_cast<llvm::StructType>(ArgI.getCoerceToType());
2330 if (ArgI.isDirect() && ArgI.getCanBeFlattened() && STy &&
2331 STy->getNumElements() > 1) {
2332 auto SrcLayout = CGM.getDataLayout().getStructLayout(STy);
2333 uint64_t SrcSize = CGM.getDataLayout().getTypeAllocSize(STy);
2334 llvm::Type *DstTy = Ptr.getElementType();
2335 uint64_t DstSize = CGM.getDataLayout().getTypeAllocSize(DstTy);
2337 Address AddrToStoreInto = Address::invalid();
2338 if (SrcSize <= DstSize) {
2340 Builder.CreateBitCast(Ptr, llvm::PointerType::getUnqual(STy));
2343 CreateTempAlloca(STy, Alloca.getAlignment(), "coerce");
2346 assert(STy->getNumElements() == NumIRArgs);
2347 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
2348 auto AI = FnArgs[FirstIRArg + i];
2349 AI->setName(Arg->getName() + ".coerce" + Twine(i));
2350 auto Offset = CharUnits::fromQuantity(SrcLayout->getElementOffset(i));
2352 Builder.CreateStructGEP(AddrToStoreInto, i, Offset);
2353 Builder.CreateStore(AI, EltPtr);
2356 if (SrcSize > DstSize) {
2357 Builder.CreateMemCpy(Ptr, AddrToStoreInto, DstSize);
2361 // Simple case, just do a coerced store of the argument into the alloca.
2362 assert(NumIRArgs == 1);
2363 auto AI = FnArgs[FirstIRArg];
2364 AI->setName(Arg->getName() + ".coerce");
2365 CreateCoercedStore(AI, Ptr, /*DestIsVolatile=*/false, *this);
2368 // Match to what EmitParmDecl is expecting for this type.
2369 if (CodeGenFunction::hasScalarEvaluationKind(Ty)) {
2371 EmitLoadOfScalar(Alloca, false, Ty, Arg->getLocStart());
2373 V = emitArgumentDemotion(*this, Arg, V);
2374 ArgVals.push_back(ParamValue::forDirect(V));
2376 ArgVals.push_back(ParamValue::forIndirect(Alloca));
2381 case ABIArgInfo::CoerceAndExpand: {
2382 // Reconstruct into a temporary.
2383 Address alloca = CreateMemTemp(Ty, getContext().getDeclAlign(Arg));
2384 ArgVals.push_back(ParamValue::forIndirect(alloca));
2386 auto coercionType = ArgI.getCoerceAndExpandType();
2387 alloca = Builder.CreateElementBitCast(alloca, coercionType);
2388 auto layout = CGM.getDataLayout().getStructLayout(coercionType);
2390 unsigned argIndex = FirstIRArg;
2391 for (unsigned i = 0, e = coercionType->getNumElements(); i != e; ++i) {
2392 llvm::Type *eltType = coercionType->getElementType(i);
2393 if (ABIArgInfo::isPaddingForCoerceAndExpand(eltType))
2396 auto eltAddr = Builder.CreateStructGEP(alloca, i, layout);
2397 auto elt = FnArgs[argIndex++];
2398 Builder.CreateStore(elt, eltAddr);
2400 assert(argIndex == FirstIRArg + NumIRArgs);
2404 case ABIArgInfo::Expand: {
2405 // If this structure was expanded into multiple arguments then
2406 // we need to create a temporary and reconstruct it from the
2408 Address Alloca = CreateMemTemp(Ty, getContext().getDeclAlign(Arg));
2409 LValue LV = MakeAddrLValue(Alloca, Ty);
2410 ArgVals.push_back(ParamValue::forIndirect(Alloca));
2412 auto FnArgIter = FnArgs.begin() + FirstIRArg;
2413 ExpandTypeFromArgs(Ty, LV, FnArgIter);
2414 assert(FnArgIter == FnArgs.begin() + FirstIRArg + NumIRArgs);
2415 for (unsigned i = 0, e = NumIRArgs; i != e; ++i) {
2416 auto AI = FnArgs[FirstIRArg + i];
2417 AI->setName(Arg->getName() + "." + Twine(i));
2422 case ABIArgInfo::Ignore:
2423 assert(NumIRArgs == 0);
2424 // Initialize the local variable appropriately.
2425 if (!hasScalarEvaluationKind(Ty)) {
2426 ArgVals.push_back(ParamValue::forIndirect(CreateMemTemp(Ty)));
2428 llvm::Value *U = llvm::UndefValue::get(ConvertType(Arg->getType()));
2429 ArgVals.push_back(ParamValue::forDirect(U));
2435 if (getTarget().getCXXABI().areArgsDestroyedLeftToRightInCallee()) {
2436 for (int I = Args.size() - 1; I >= 0; --I)
2437 EmitParmDecl(*Args[I], ArgVals[I], I + 1);
2439 for (unsigned I = 0, E = Args.size(); I != E; ++I)
2440 EmitParmDecl(*Args[I], ArgVals[I], I + 1);
2444 static void eraseUnusedBitCasts(llvm::Instruction *insn) {
2445 while (insn->use_empty()) {
2446 llvm::BitCastInst *bitcast = dyn_cast<llvm::BitCastInst>(insn);
2447 if (!bitcast) return;
2449 // This is "safe" because we would have used a ConstantExpr otherwise.
2450 insn = cast<llvm::Instruction>(bitcast->getOperand(0));
2451 bitcast->eraseFromParent();
2455 /// Try to emit a fused autorelease of a return result.
2456 static llvm::Value *tryEmitFusedAutoreleaseOfResult(CodeGenFunction &CGF,
2457 llvm::Value *result) {
2458 // We must be immediately followed the cast.
2459 llvm::BasicBlock *BB = CGF.Builder.GetInsertBlock();
2460 if (BB->empty()) return nullptr;
2461 if (&BB->back() != result) return nullptr;
2463 llvm::Type *resultType = result->getType();
2465 // result is in a BasicBlock and is therefore an Instruction.
2466 llvm::Instruction *generator = cast<llvm::Instruction>(result);
2468 SmallVector<llvm::Instruction*,4> insnsToKill;
2471 // %generator = bitcast %type1* %generator2 to %type2*
2472 while (llvm::BitCastInst *bitcast = dyn_cast<llvm::BitCastInst>(generator)) {
2473 // We would have emitted this as a constant if the operand weren't
2475 generator = cast<llvm::Instruction>(bitcast->getOperand(0));
2477 // Require the generator to be immediately followed by the cast.
2478 if (generator->getNextNode() != bitcast)
2481 insnsToKill.push_back(bitcast);
2485 // %generator = call i8* @objc_retain(i8* %originalResult)
2487 // %generator = call i8* @objc_retainAutoreleasedReturnValue(i8* %originalResult)
2488 llvm::CallInst *call = dyn_cast<llvm::CallInst>(generator);
2489 if (!call) return nullptr;
2491 bool doRetainAutorelease;
2493 if (call->getCalledValue() == CGF.CGM.getObjCEntrypoints().objc_retain) {
2494 doRetainAutorelease = true;
2495 } else if (call->getCalledValue() == CGF.CGM.getObjCEntrypoints()
2496 .objc_retainAutoreleasedReturnValue) {
2497 doRetainAutorelease = false;
2499 // If we emitted an assembly marker for this call (and the
2500 // ARCEntrypoints field should have been set if so), go looking
2501 // for that call. If we can't find it, we can't do this
2502 // optimization. But it should always be the immediately previous
2503 // instruction, unless we needed bitcasts around the call.
2504 if (CGF.CGM.getObjCEntrypoints().retainAutoreleasedReturnValueMarker) {
2505 llvm::Instruction *prev = call->getPrevNode();
2507 if (isa<llvm::BitCastInst>(prev)) {
2508 prev = prev->getPrevNode();
2511 assert(isa<llvm::CallInst>(prev));
2512 assert(cast<llvm::CallInst>(prev)->getCalledValue() ==
2513 CGF.CGM.getObjCEntrypoints().retainAutoreleasedReturnValueMarker);
2514 insnsToKill.push_back(prev);
2520 result = call->getArgOperand(0);
2521 insnsToKill.push_back(call);
2523 // Keep killing bitcasts, for sanity. Note that we no longer care
2524 // about precise ordering as long as there's exactly one use.
2525 while (llvm::BitCastInst *bitcast = dyn_cast<llvm::BitCastInst>(result)) {
2526 if (!bitcast->hasOneUse()) break;
2527 insnsToKill.push_back(bitcast);
2528 result = bitcast->getOperand(0);
2531 // Delete all the unnecessary instructions, from latest to earliest.
2532 for (SmallVectorImpl<llvm::Instruction*>::iterator
2533 i = insnsToKill.begin(), e = insnsToKill.end(); i != e; ++i)
2534 (*i)->eraseFromParent();
2536 // Do the fused retain/autorelease if we were asked to.
2537 if (doRetainAutorelease)
2538 result = CGF.EmitARCRetainAutoreleaseReturnValue(result);
2540 // Cast back to the result type.
2541 return CGF.Builder.CreateBitCast(result, resultType);
2544 /// If this is a +1 of the value of an immutable 'self', remove it.
2545 static llvm::Value *tryRemoveRetainOfSelf(CodeGenFunction &CGF,
2546 llvm::Value *result) {
2547 // This is only applicable to a method with an immutable 'self'.
2548 const ObjCMethodDecl *method =
2549 dyn_cast_or_null<ObjCMethodDecl>(CGF.CurCodeDecl);
2550 if (!method) return nullptr;
2551 const VarDecl *self = method->getSelfDecl();
2552 if (!self->getType().isConstQualified()) return nullptr;
2554 // Look for a retain call.
2555 llvm::CallInst *retainCall =
2556 dyn_cast<llvm::CallInst>(result->stripPointerCasts());
2558 retainCall->getCalledValue() != CGF.CGM.getObjCEntrypoints().objc_retain)
2561 // Look for an ordinary load of 'self'.
2562 llvm::Value *retainedValue = retainCall->getArgOperand(0);
2563 llvm::LoadInst *load =
2564 dyn_cast<llvm::LoadInst>(retainedValue->stripPointerCasts());
2565 if (!load || load->isAtomic() || load->isVolatile() ||
2566 load->getPointerOperand() != CGF.GetAddrOfLocalVar(self).getPointer())
2569 // Okay! Burn it all down. This relies for correctness on the
2570 // assumption that the retain is emitted as part of the return and
2571 // that thereafter everything is used "linearly".
2572 llvm::Type *resultType = result->getType();
2573 eraseUnusedBitCasts(cast<llvm::Instruction>(result));
2574 assert(retainCall->use_empty());
2575 retainCall->eraseFromParent();
2576 eraseUnusedBitCasts(cast<llvm::Instruction>(retainedValue));
2578 return CGF.Builder.CreateBitCast(load, resultType);
2581 /// Emit an ARC autorelease of the result of a function.
2583 /// \return the value to actually return from the function
2584 static llvm::Value *emitAutoreleaseOfResult(CodeGenFunction &CGF,
2585 llvm::Value *result) {
2586 // If we're returning 'self', kill the initial retain. This is a
2587 // heuristic attempt to "encourage correctness" in the really unfortunate
2588 // case where we have a return of self during a dealloc and we desperately
2589 // need to avoid the possible autorelease.
2590 if (llvm::Value *self = tryRemoveRetainOfSelf(CGF, result))
2593 // At -O0, try to emit a fused retain/autorelease.
2594 if (CGF.shouldUseFusedARCCalls())
2595 if (llvm::Value *fused = tryEmitFusedAutoreleaseOfResult(CGF, result))
2598 return CGF.EmitARCAutoreleaseReturnValue(result);
2601 /// Heuristically search for a dominating store to the return-value slot.
2602 static llvm::StoreInst *findDominatingStoreToReturnValue(CodeGenFunction &CGF) {
2603 // Check if a User is a store which pointerOperand is the ReturnValue.
2604 // We are looking for stores to the ReturnValue, not for stores of the
2605 // ReturnValue to some other location.
2606 auto GetStoreIfValid = [&CGF](llvm::User *U) -> llvm::StoreInst * {
2607 auto *SI = dyn_cast<llvm::StoreInst>(U);
2608 if (!SI || SI->getPointerOperand() != CGF.ReturnValue.getPointer())
2610 // These aren't actually possible for non-coerced returns, and we
2611 // only care about non-coerced returns on this code path.
2612 assert(!SI->isAtomic() && !SI->isVolatile());
2615 // If there are multiple uses of the return-value slot, just check
2616 // for something immediately preceding the IP. Sometimes this can
2617 // happen with how we generate implicit-returns; it can also happen
2618 // with noreturn cleanups.
2619 if (!CGF.ReturnValue.getPointer()->hasOneUse()) {
2620 llvm::BasicBlock *IP = CGF.Builder.GetInsertBlock();
2621 if (IP->empty()) return nullptr;
2622 llvm::Instruction *I = &IP->back();
2624 // Skip lifetime markers
2625 for (llvm::BasicBlock::reverse_iterator II = IP->rbegin(),
2628 if (llvm::IntrinsicInst *Intrinsic =
2629 dyn_cast<llvm::IntrinsicInst>(&*II)) {
2630 if (Intrinsic->getIntrinsicID() == llvm::Intrinsic::lifetime_end) {
2631 const llvm::Value *CastAddr = Intrinsic->getArgOperand(1);
2635 if (isa<llvm::BitCastInst>(&*II) && (CastAddr == &*II))
2643 return GetStoreIfValid(I);
2646 llvm::StoreInst *store =
2647 GetStoreIfValid(CGF.ReturnValue.getPointer()->user_back());
2648 if (!store) return nullptr;
2650 // Now do a first-and-dirty dominance check: just walk up the
2651 // single-predecessors chain from the current insertion point.
2652 llvm::BasicBlock *StoreBB = store->getParent();
2653 llvm::BasicBlock *IP = CGF.Builder.GetInsertBlock();
2654 while (IP != StoreBB) {
2655 if (!(IP = IP->getSinglePredecessor()))
2659 // Okay, the store's basic block dominates the insertion point; we
2660 // can do our thing.
2664 void CodeGenFunction::EmitFunctionEpilog(const CGFunctionInfo &FI,
2666 SourceLocation EndLoc) {
2667 if (CurCodeDecl && CurCodeDecl->hasAttr<NakedAttr>()) {
2668 // Naked functions don't have epilogues.
2669 Builder.CreateUnreachable();
2673 // Functions with no result always return void.
2674 if (!ReturnValue.isValid()) {
2675 Builder.CreateRetVoid();
2679 llvm::DebugLoc RetDbgLoc;
2680 llvm::Value *RV = nullptr;
2681 QualType RetTy = FI.getReturnType();
2682 const ABIArgInfo &RetAI = FI.getReturnInfo();
2684 switch (RetAI.getKind()) {
2685 case ABIArgInfo::InAlloca:
2686 // Aggregrates get evaluated directly into the destination. Sometimes we
2687 // need to return the sret value in a register, though.
2688 assert(hasAggregateEvaluationKind(RetTy));
2689 if (RetAI.getInAllocaSRet()) {
2690 llvm::Function::arg_iterator EI = CurFn->arg_end();
2692 llvm::Value *ArgStruct = &*EI;
2693 llvm::Value *SRet = Builder.CreateStructGEP(
2694 nullptr, ArgStruct, RetAI.getInAllocaFieldIndex());
2695 RV = Builder.CreateAlignedLoad(SRet, getPointerAlign(), "sret");
2699 case ABIArgInfo::Indirect: {
2700 auto AI = CurFn->arg_begin();
2701 if (RetAI.isSRetAfterThis())
2703 switch (getEvaluationKind(RetTy)) {
2706 EmitLoadOfComplex(MakeAddrLValue(ReturnValue, RetTy), EndLoc);
2707 EmitStoreOfComplex(RT, MakeNaturalAlignAddrLValue(&*AI, RetTy),
2712 // Do nothing; aggregrates get evaluated directly into the destination.
2715 EmitStoreOfScalar(Builder.CreateLoad(ReturnValue),
2716 MakeNaturalAlignAddrLValue(&*AI, RetTy),
2723 case ABIArgInfo::Extend:
2724 case ABIArgInfo::Direct:
2725 if (RetAI.getCoerceToType() == ConvertType(RetTy) &&
2726 RetAI.getDirectOffset() == 0) {
2727 // The internal return value temp always will have pointer-to-return-type
2728 // type, just do a load.
2730 // If there is a dominating store to ReturnValue, we can elide
2731 // the load, zap the store, and usually zap the alloca.
2732 if (llvm::StoreInst *SI =
2733 findDominatingStoreToReturnValue(*this)) {
2734 // Reuse the debug location from the store unless there is
2735 // cleanup code to be emitted between the store and return
2737 if (EmitRetDbgLoc && !AutoreleaseResult)
2738 RetDbgLoc = SI->getDebugLoc();
2739 // Get the stored value and nuke the now-dead store.
2740 RV = SI->getValueOperand();
2741 SI->eraseFromParent();
2743 // If that was the only use of the return value, nuke it as well now.
2744 auto returnValueInst = ReturnValue.getPointer();
2745 if (returnValueInst->use_empty()) {
2746 if (auto alloca = dyn_cast<llvm::AllocaInst>(returnValueInst)) {
2747 alloca->eraseFromParent();
2748 ReturnValue = Address::invalid();
2752 // Otherwise, we have to do a simple load.
2754 RV = Builder.CreateLoad(ReturnValue);
2757 // If the value is offset in memory, apply the offset now.
2758 Address V = emitAddressAtOffset(*this, ReturnValue, RetAI);
2760 RV = CreateCoercedLoad(V, RetAI.getCoerceToType(), *this);
2763 // In ARC, end functions that return a retainable type with a call
2764 // to objc_autoreleaseReturnValue.
2765 if (AutoreleaseResult) {
2767 // Type::isObjCRetainabletype has to be called on a QualType that hasn't
2768 // been stripped of the typedefs, so we cannot use RetTy here. Get the
2769 // original return type of FunctionDecl, CurCodeDecl, and BlockDecl from
2770 // CurCodeDecl or BlockInfo.
2773 if (auto *FD = dyn_cast<FunctionDecl>(CurCodeDecl))
2774 RT = FD->getReturnType();
2775 else if (auto *MD = dyn_cast<ObjCMethodDecl>(CurCodeDecl))
2776 RT = MD->getReturnType();
2777 else if (isa<BlockDecl>(CurCodeDecl))
2778 RT = BlockInfo->BlockExpression->getFunctionType()->getReturnType();
2780 llvm_unreachable("Unexpected function/method type");
2782 assert(getLangOpts().ObjCAutoRefCount &&
2783 !FI.isReturnsRetained() &&
2784 RT->isObjCRetainableType());
2786 RV = emitAutoreleaseOfResult(*this, RV);
2791 case ABIArgInfo::Ignore:
2794 case ABIArgInfo::CoerceAndExpand: {
2795 auto coercionType = RetAI.getCoerceAndExpandType();
2796 auto layout = CGM.getDataLayout().getStructLayout(coercionType);
2798 // Load all of the coerced elements out into results.
2799 llvm::SmallVector<llvm::Value*, 4> results;
2800 Address addr = Builder.CreateElementBitCast(ReturnValue, coercionType);
2801 for (unsigned i = 0, e = coercionType->getNumElements(); i != e; ++i) {
2802 auto coercedEltType = coercionType->getElementType(i);
2803 if (ABIArgInfo::isPaddingForCoerceAndExpand(coercedEltType))
2806 auto eltAddr = Builder.CreateStructGEP(addr, i, layout);
2807 auto elt = Builder.CreateLoad(eltAddr);
2808 results.push_back(elt);
2811 // If we have one result, it's the single direct result type.
2812 if (results.size() == 1) {
2815 // Otherwise, we need to make a first-class aggregate.
2817 // Construct a return type that lacks padding elements.
2818 llvm::Type *returnType = RetAI.getUnpaddedCoerceAndExpandType();
2820 RV = llvm::UndefValue::get(returnType);
2821 for (unsigned i = 0, e = results.size(); i != e; ++i) {
2822 RV = Builder.CreateInsertValue(RV, results[i], i);
2828 case ABIArgInfo::Expand:
2829 llvm_unreachable("Invalid ABI kind for return argument");
2832 llvm::Instruction *Ret;
2834 if (CurCodeDecl && SanOpts.has(SanitizerKind::ReturnsNonnullAttribute)) {
2835 if (auto RetNNAttr = CurCodeDecl->getAttr<ReturnsNonNullAttr>()) {
2836 SanitizerScope SanScope(this);
2837 llvm::Value *Cond = Builder.CreateICmpNE(
2838 RV, llvm::Constant::getNullValue(RV->getType()));
2839 llvm::Constant *StaticData[] = {
2840 EmitCheckSourceLocation(EndLoc),
2841 EmitCheckSourceLocation(RetNNAttr->getLocation()),
2843 EmitCheck(std::make_pair(Cond, SanitizerKind::ReturnsNonnullAttribute),
2844 "nonnull_return", StaticData, None);
2847 Ret = Builder.CreateRet(RV);
2849 Ret = Builder.CreateRetVoid();
2853 Ret->setDebugLoc(std::move(RetDbgLoc));
2856 static bool isInAllocaArgument(CGCXXABI &ABI, QualType type) {
2857 const CXXRecordDecl *RD = type->getAsCXXRecordDecl();
2858 return RD && ABI.getRecordArgABI(RD) == CGCXXABI::RAA_DirectInMemory;
2861 static AggValueSlot createPlaceholderSlot(CodeGenFunction &CGF,
2863 // FIXME: Generate IR in one pass, rather than going back and fixing up these
2865 llvm::Type *IRTy = CGF.ConvertTypeForMem(Ty);
2866 llvm::Value *Placeholder =
2867 llvm::UndefValue::get(IRTy->getPointerTo()->getPointerTo());
2868 Placeholder = CGF.Builder.CreateDefaultAlignedLoad(Placeholder);
2870 // FIXME: When we generate this IR in one pass, we shouldn't need
2871 // this win32-specific alignment hack.
2872 CharUnits Align = CharUnits::fromQuantity(4);
2874 return AggValueSlot::forAddr(Address(Placeholder, Align),
2876 AggValueSlot::IsNotDestructed,
2877 AggValueSlot::DoesNotNeedGCBarriers,
2878 AggValueSlot::IsNotAliased);
2881 void CodeGenFunction::EmitDelegateCallArg(CallArgList &args,
2882 const VarDecl *param,
2883 SourceLocation loc) {
2884 // StartFunction converted the ABI-lowered parameter(s) into a
2885 // local alloca. We need to turn that into an r-value suitable
2887 Address local = GetAddrOfLocalVar(param);
2889 QualType type = param->getType();
2891 assert(!isInAllocaArgument(CGM.getCXXABI(), type) &&
2892 "cannot emit delegate call arguments for inalloca arguments!");
2894 // For the most part, we just need to load the alloca, except that
2895 // aggregate r-values are actually pointers to temporaries.
2896 if (type->isReferenceType())
2897 args.add(RValue::get(Builder.CreateLoad(local)), type);
2899 args.add(convertTempToRValue(local, type, loc), type);
2902 static bool isProvablyNull(llvm::Value *addr) {
2903 return isa<llvm::ConstantPointerNull>(addr);
2906 static bool isProvablyNonNull(llvm::Value *addr) {
2907 return isa<llvm::AllocaInst>(addr);
2910 /// Emit the actual writing-back of a writeback.
2911 static void emitWriteback(CodeGenFunction &CGF,
2912 const CallArgList::Writeback &writeback) {
2913 const LValue &srcLV = writeback.Source;
2914 Address srcAddr = srcLV.getAddress();
2915 assert(!isProvablyNull(srcAddr.getPointer()) &&
2916 "shouldn't have writeback for provably null argument");
2918 llvm::BasicBlock *contBB = nullptr;
2920 // If the argument wasn't provably non-null, we need to null check
2921 // before doing the store.
2922 bool provablyNonNull = isProvablyNonNull(srcAddr.getPointer());
2923 if (!provablyNonNull) {
2924 llvm::BasicBlock *writebackBB = CGF.createBasicBlock("icr.writeback");
2925 contBB = CGF.createBasicBlock("icr.done");
2927 llvm::Value *isNull =
2928 CGF.Builder.CreateIsNull(srcAddr.getPointer(), "icr.isnull");
2929 CGF.Builder.CreateCondBr(isNull, contBB, writebackBB);
2930 CGF.EmitBlock(writebackBB);
2933 // Load the value to writeback.
2934 llvm::Value *value = CGF.Builder.CreateLoad(writeback.Temporary);
2936 // Cast it back, in case we're writing an id to a Foo* or something.
2937 value = CGF.Builder.CreateBitCast(value, srcAddr.getElementType(),
2938 "icr.writeback-cast");
2940 // Perform the writeback.
2942 // If we have a "to use" value, it's something we need to emit a use
2943 // of. This has to be carefully threaded in: if it's done after the
2944 // release it's potentially undefined behavior (and the optimizer
2945 // will ignore it), and if it happens before the retain then the
2946 // optimizer could move the release there.
2947 if (writeback.ToUse) {
2948 assert(srcLV.getObjCLifetime() == Qualifiers::OCL_Strong);
2950 // Retain the new value. No need to block-copy here: the block's
2951 // being passed up the stack.
2952 value = CGF.EmitARCRetainNonBlock(value);
2954 // Emit the intrinsic use here.
2955 CGF.EmitARCIntrinsicUse(writeback.ToUse);
2957 // Load the old value (primitively).
2958 llvm::Value *oldValue = CGF.EmitLoadOfScalar(srcLV, SourceLocation());
2960 // Put the new value in place (primitively).
2961 CGF.EmitStoreOfScalar(value, srcLV, /*init*/ false);
2963 // Release the old value.
2964 CGF.EmitARCRelease(oldValue, srcLV.isARCPreciseLifetime());
2966 // Otherwise, we can just do a normal lvalue store.
2968 CGF.EmitStoreThroughLValue(RValue::get(value), srcLV);
2971 // Jump to the continuation block.
2972 if (!provablyNonNull)
2973 CGF.EmitBlock(contBB);
2976 static void emitWritebacks(CodeGenFunction &CGF,
2977 const CallArgList &args) {
2978 for (const auto &I : args.writebacks())
2979 emitWriteback(CGF, I);
2982 static void deactivateArgCleanupsBeforeCall(CodeGenFunction &CGF,
2983 const CallArgList &CallArgs) {
2984 assert(CGF.getTarget().getCXXABI().areArgsDestroyedLeftToRightInCallee());
2985 ArrayRef<CallArgList::CallArgCleanup> Cleanups =
2986 CallArgs.getCleanupsToDeactivate();
2987 // Iterate in reverse to increase the likelihood of popping the cleanup.
2988 for (const auto &I : llvm::reverse(Cleanups)) {
2989 CGF.DeactivateCleanupBlock(I.Cleanup, I.IsActiveIP);
2990 I.IsActiveIP->eraseFromParent();
2994 static const Expr *maybeGetUnaryAddrOfOperand(const Expr *E) {
2995 if (const UnaryOperator *uop = dyn_cast<UnaryOperator>(E->IgnoreParens()))
2996 if (uop->getOpcode() == UO_AddrOf)
2997 return uop->getSubExpr();
3001 /// Emit an argument that's being passed call-by-writeback. That is,
3002 /// we are passing the address of an __autoreleased temporary; it
3003 /// might be copy-initialized with the current value of the given
3004 /// address, but it will definitely be copied out of after the call.
3005 static void emitWritebackArg(CodeGenFunction &CGF, CallArgList &args,
3006 const ObjCIndirectCopyRestoreExpr *CRE) {
3009 // Make an optimistic effort to emit the address as an l-value.
3010 // This can fail if the argument expression is more complicated.
3011 if (const Expr *lvExpr = maybeGetUnaryAddrOfOperand(CRE->getSubExpr())) {
3012 srcLV = CGF.EmitLValue(lvExpr);
3014 // Otherwise, just emit it as a scalar.
3016 Address srcAddr = CGF.EmitPointerWithAlignment(CRE->getSubExpr());
3018 QualType srcAddrType =
3019 CRE->getSubExpr()->getType()->castAs<PointerType>()->getPointeeType();
3020 srcLV = CGF.MakeAddrLValue(srcAddr, srcAddrType);
3022 Address srcAddr = srcLV.getAddress();
3024 // The dest and src types don't necessarily match in LLVM terms
3025 // because of the crazy ObjC compatibility rules.
3027 llvm::PointerType *destType =
3028 cast<llvm::PointerType>(CGF.ConvertType(CRE->getType()));
3030 // If the address is a constant null, just pass the appropriate null.
3031 if (isProvablyNull(srcAddr.getPointer())) {
3032 args.add(RValue::get(llvm::ConstantPointerNull::get(destType)),
3037 // Create the temporary.
3038 Address temp = CGF.CreateTempAlloca(destType->getElementType(),
3039 CGF.getPointerAlign(),
3041 // Loading an l-value can introduce a cleanup if the l-value is __weak,
3042 // and that cleanup will be conditional if we can't prove that the l-value
3043 // isn't null, so we need to register a dominating point so that the cleanups
3044 // system will make valid IR.
3045 CodeGenFunction::ConditionalEvaluation condEval(CGF);
3047 // Zero-initialize it if we're not doing a copy-initialization.
3048 bool shouldCopy = CRE->shouldCopy();
3051 llvm::ConstantPointerNull::get(
3052 cast<llvm::PointerType>(destType->getElementType()));
3053 CGF.Builder.CreateStore(null, temp);
3056 llvm::BasicBlock *contBB = nullptr;
3057 llvm::BasicBlock *originBB = nullptr;
3059 // If the address is *not* known to be non-null, we need to switch.
3060 llvm::Value *finalArgument;
3062 bool provablyNonNull = isProvablyNonNull(srcAddr.getPointer());
3063 if (provablyNonNull) {
3064 finalArgument = temp.getPointer();
3066 llvm::Value *isNull =
3067 CGF.Builder.CreateIsNull(srcAddr.getPointer(), "icr.isnull");
3069 finalArgument = CGF.Builder.CreateSelect(isNull,
3070 llvm::ConstantPointerNull::get(destType),
3071 temp.getPointer(), "icr.argument");
3073 // If we need to copy, then the load has to be conditional, which
3074 // means we need control flow.
3076 originBB = CGF.Builder.GetInsertBlock();
3077 contBB = CGF.createBasicBlock("icr.cont");
3078 llvm::BasicBlock *copyBB = CGF.createBasicBlock("icr.copy");
3079 CGF.Builder.CreateCondBr(isNull, contBB, copyBB);
3080 CGF.EmitBlock(copyBB);
3081 condEval.begin(CGF);
3085 llvm::Value *valueToUse = nullptr;
3087 // Perform a copy if necessary.
3089 RValue srcRV = CGF.EmitLoadOfLValue(srcLV, SourceLocation());
3090 assert(srcRV.isScalar());
3092 llvm::Value *src = srcRV.getScalarVal();
3093 src = CGF.Builder.CreateBitCast(src, destType->getElementType(),
3096 // Use an ordinary store, not a store-to-lvalue.
3097 CGF.Builder.CreateStore(src, temp);
3099 // If optimization is enabled, and the value was held in a
3100 // __strong variable, we need to tell the optimizer that this
3101 // value has to stay alive until we're doing the store back.
3102 // This is because the temporary is effectively unretained,
3103 // and so otherwise we can violate the high-level semantics.
3104 if (CGF.CGM.getCodeGenOpts().OptimizationLevel != 0 &&
3105 srcLV.getObjCLifetime() == Qualifiers::OCL_Strong) {
3110 // Finish the control flow if we needed it.
3111 if (shouldCopy && !provablyNonNull) {
3112 llvm::BasicBlock *copyBB = CGF.Builder.GetInsertBlock();
3113 CGF.EmitBlock(contBB);
3115 // Make a phi for the value to intrinsically use.
3117 llvm::PHINode *phiToUse = CGF.Builder.CreatePHI(valueToUse->getType(), 2,
3119 phiToUse->addIncoming(valueToUse, copyBB);
3120 phiToUse->addIncoming(llvm::UndefValue::get(valueToUse->getType()),
3122 valueToUse = phiToUse;
3128 args.addWriteback(srcLV, temp, valueToUse);
3129 args.add(RValue::get(finalArgument), CRE->getType());
3132 void CallArgList::allocateArgumentMemory(CodeGenFunction &CGF) {
3133 assert(!StackBase && !StackCleanup.isValid());
3136 llvm::Function *F = CGF.CGM.getIntrinsic(llvm::Intrinsic::stacksave);
3137 StackBase = CGF.Builder.CreateCall(F, {}, "inalloca.save");
3140 void CallArgList::freeArgumentMemory(CodeGenFunction &CGF) const {
3142 // Restore the stack after the call.
3143 llvm::Value *F = CGF.CGM.getIntrinsic(llvm::Intrinsic::stackrestore);
3144 CGF.Builder.CreateCall(F, StackBase);
3148 void CodeGenFunction::EmitNonNullArgCheck(RValue RV, QualType ArgType,
3149 SourceLocation ArgLoc,
3150 const FunctionDecl *FD,
3152 if (!SanOpts.has(SanitizerKind::NonnullAttribute) || !FD)
3154 auto PVD = ParmNum < FD->getNumParams() ? FD->getParamDecl(ParmNum) : nullptr;
3155 unsigned ArgNo = PVD ? PVD->getFunctionScopeIndex() : ParmNum;
3156 auto NNAttr = getNonNullAttr(FD, PVD, ArgType, ArgNo);
3159 SanitizerScope SanScope(this);
3160 assert(RV.isScalar());
3161 llvm::Value *V = RV.getScalarVal();
3163 Builder.CreateICmpNE(V, llvm::Constant::getNullValue(V->getType()));
3164 llvm::Constant *StaticData[] = {
3165 EmitCheckSourceLocation(ArgLoc),
3166 EmitCheckSourceLocation(NNAttr->getLocation()),
3167 llvm::ConstantInt::get(Int32Ty, ArgNo + 1),
3169 EmitCheck(std::make_pair(Cond, SanitizerKind::NonnullAttribute),
3170 "nonnull_arg", StaticData, None);
3173 void CodeGenFunction::EmitCallArgs(
3174 CallArgList &Args, ArrayRef<QualType> ArgTypes,
3175 llvm::iterator_range<CallExpr::const_arg_iterator> ArgRange,
3176 const FunctionDecl *CalleeDecl, unsigned ParamsToSkip) {
3177 assert((int)ArgTypes.size() == (ArgRange.end() - ArgRange.begin()));
3179 auto MaybeEmitImplicitObjectSize = [&](unsigned I, const Expr *Arg) {
3180 if (CalleeDecl == nullptr || I >= CalleeDecl->getNumParams())
3182 auto *PS = CalleeDecl->getParamDecl(I)->getAttr<PassObjectSizeAttr>();
3186 const auto &Context = getContext();
3187 auto SizeTy = Context.getSizeType();
3188 auto T = Builder.getIntNTy(Context.getTypeSize(SizeTy));
3189 llvm::Value *V = evaluateOrEmitBuiltinObjectSize(Arg, PS->getType(), T);
3190 Args.add(RValue::get(V), SizeTy);
3193 // We *have* to evaluate arguments from right to left in the MS C++ ABI,
3194 // because arguments are destroyed left to right in the callee.
3195 if (CGM.getTarget().getCXXABI().areArgsDestroyedLeftToRightInCallee()) {
3196 // Insert a stack save if we're going to need any inalloca args.
3197 bool HasInAllocaArgs = false;
3198 for (ArrayRef<QualType>::iterator I = ArgTypes.begin(), E = ArgTypes.end();
3199 I != E && !HasInAllocaArgs; ++I)
3200 HasInAllocaArgs = isInAllocaArgument(CGM.getCXXABI(), *I);
3201 if (HasInAllocaArgs) {
3202 assert(getTarget().getTriple().getArch() == llvm::Triple::x86);
3203 Args.allocateArgumentMemory(*this);
3206 // Evaluate each argument.
3207 size_t CallArgsStart = Args.size();
3208 for (int I = ArgTypes.size() - 1; I >= 0; --I) {
3209 CallExpr::const_arg_iterator Arg = ArgRange.begin() + I;
3210 MaybeEmitImplicitObjectSize(I, *Arg);
3211 EmitCallArg(Args, *Arg, ArgTypes[I]);
3212 EmitNonNullArgCheck(Args.back().RV, ArgTypes[I], (*Arg)->getExprLoc(),
3213 CalleeDecl, ParamsToSkip + I);
3216 // Un-reverse the arguments we just evaluated so they match up with the LLVM
3218 std::reverse(Args.begin() + CallArgsStart, Args.end());
3222 for (unsigned I = 0, E = ArgTypes.size(); I != E; ++I) {
3223 CallExpr::const_arg_iterator Arg = ArgRange.begin() + I;
3224 assert(Arg != ArgRange.end());
3225 EmitCallArg(Args, *Arg, ArgTypes[I]);
3226 EmitNonNullArgCheck(Args.back().RV, ArgTypes[I], (*Arg)->getExprLoc(),
3227 CalleeDecl, ParamsToSkip + I);
3228 MaybeEmitImplicitObjectSize(I, *Arg);
3234 struct DestroyUnpassedArg final : EHScopeStack::Cleanup {
3235 DestroyUnpassedArg(Address Addr, QualType Ty)
3236 : Addr(Addr), Ty(Ty) {}
3241 void Emit(CodeGenFunction &CGF, Flags flags) override {
3242 const CXXDestructorDecl *Dtor = Ty->getAsCXXRecordDecl()->getDestructor();
3243 assert(!Dtor->isTrivial());
3244 CGF.EmitCXXDestructorCall(Dtor, Dtor_Complete, /*for vbase*/ false,
3245 /*Delegating=*/false, Addr);
3249 struct DisableDebugLocationUpdates {
3250 CodeGenFunction &CGF;
3251 bool disabledDebugInfo;
3252 DisableDebugLocationUpdates(CodeGenFunction &CGF, const Expr *E) : CGF(CGF) {
3253 if ((disabledDebugInfo = isa<CXXDefaultArgExpr>(E) && CGF.getDebugInfo()))
3254 CGF.disableDebugInfo();
3256 ~DisableDebugLocationUpdates() {
3257 if (disabledDebugInfo)
3258 CGF.enableDebugInfo();
3262 } // end anonymous namespace
3264 void CodeGenFunction::EmitCallArg(CallArgList &args, const Expr *E,
3266 DisableDebugLocationUpdates Dis(*this, E);
3267 if (const ObjCIndirectCopyRestoreExpr *CRE
3268 = dyn_cast<ObjCIndirectCopyRestoreExpr>(E)) {
3269 assert(getLangOpts().ObjCAutoRefCount);
3270 assert(getContext().hasSameType(E->getType(), type));
3271 return emitWritebackArg(*this, args, CRE);
3274 assert(type->isReferenceType() == E->isGLValue() &&
3275 "reference binding to unmaterialized r-value!");
3277 if (E->isGLValue()) {
3278 assert(E->getObjectKind() == OK_Ordinary);
3279 return args.add(EmitReferenceBindingToExpr(E), type);
3282 bool HasAggregateEvalKind = hasAggregateEvaluationKind(type);
3284 // In the Microsoft C++ ABI, aggregate arguments are destructed by the callee.
3285 // However, we still have to push an EH-only cleanup in case we unwind before
3286 // we make it to the call.
3287 if (HasAggregateEvalKind &&
3288 CGM.getTarget().getCXXABI().areArgsDestroyedLeftToRightInCallee()) {
3289 // If we're using inalloca, use the argument memory. Otherwise, use a
3292 if (args.isUsingInAlloca())
3293 Slot = createPlaceholderSlot(*this, type);
3295 Slot = CreateAggTemp(type, "agg.tmp");
3297 const CXXRecordDecl *RD = type->getAsCXXRecordDecl();
3298 bool DestroyedInCallee =
3299 RD && RD->hasNonTrivialDestructor() &&
3300 CGM.getCXXABI().getRecordArgABI(RD) != CGCXXABI::RAA_Default;
3301 if (DestroyedInCallee)
3302 Slot.setExternallyDestructed();
3304 EmitAggExpr(E, Slot);
3305 RValue RV = Slot.asRValue();
3308 if (DestroyedInCallee) {
3309 // Create a no-op GEP between the placeholder and the cleanup so we can
3310 // RAUW it successfully. It also serves as a marker of the first
3311 // instruction where the cleanup is active.
3312 pushFullExprCleanup<DestroyUnpassedArg>(EHCleanup, Slot.getAddress(),
3314 // This unreachable is a temporary marker which will be removed later.
3315 llvm::Instruction *IsActive = Builder.CreateUnreachable();
3316 args.addArgCleanupDeactivation(EHStack.getInnermostEHScope(), IsActive);
3321 if (HasAggregateEvalKind && isa<ImplicitCastExpr>(E) &&
3322 cast<CastExpr>(E)->getCastKind() == CK_LValueToRValue) {
3323 LValue L = EmitLValue(cast<CastExpr>(E)->getSubExpr());
3324 assert(L.isSimple());
3325 if (L.getAlignment() >= getContext().getTypeAlignInChars(type)) {
3326 args.add(L.asAggregateRValue(), type, /*NeedsCopy*/true);
3328 // We can't represent a misaligned lvalue in the CallArgList, so copy
3329 // to an aligned temporary now.
3330 Address tmp = CreateMemTemp(type);
3331 EmitAggregateCopy(tmp, L.getAddress(), type, L.isVolatile());
3332 args.add(RValue::getAggregate(tmp), type);
3337 args.add(EmitAnyExprToTemp(E), type);
3340 QualType CodeGenFunction::getVarArgType(const Expr *Arg) {
3341 // System headers on Windows define NULL to 0 instead of 0LL on Win64. MSVC
3342 // implicitly widens null pointer constants that are arguments to varargs
3343 // functions to pointer-sized ints.
3344 if (!getTarget().getTriple().isOSWindows())
3345 return Arg->getType();
3347 if (Arg->getType()->isIntegerType() &&
3348 getContext().getTypeSize(Arg->getType()) <
3349 getContext().getTargetInfo().getPointerWidth(0) &&
3350 Arg->isNullPointerConstant(getContext(),
3351 Expr::NPC_ValueDependentIsNotNull)) {
3352 return getContext().getIntPtrType();
3355 return Arg->getType();
3358 // In ObjC ARC mode with no ObjC ARC exception safety, tell the ARC
3359 // optimizer it can aggressively ignore unwind edges.
3361 CodeGenFunction::AddObjCARCExceptionMetadata(llvm::Instruction *Inst) {
3362 if (CGM.getCodeGenOpts().OptimizationLevel != 0 &&
3363 !CGM.getCodeGenOpts().ObjCAutoRefCountExceptions)
3364 Inst->setMetadata("clang.arc.no_objc_arc_exceptions",
3365 CGM.getNoObjCARCExceptionsMetadata());
3368 /// Emits a call to the given no-arguments nounwind runtime function.
3370 CodeGenFunction::EmitNounwindRuntimeCall(llvm::Value *callee,
3371 const llvm::Twine &name) {
3372 return EmitNounwindRuntimeCall(callee, None, name);
3375 /// Emits a call to the given nounwind runtime function.
3377 CodeGenFunction::EmitNounwindRuntimeCall(llvm::Value *callee,
3378 ArrayRef<llvm::Value*> args,
3379 const llvm::Twine &name) {
3380 llvm::CallInst *call = EmitRuntimeCall(callee, args, name);
3381 call->setDoesNotThrow();
3385 /// Emits a simple call (never an invoke) to the given no-arguments
3386 /// runtime function.
3388 CodeGenFunction::EmitRuntimeCall(llvm::Value *callee,
3389 const llvm::Twine &name) {
3390 return EmitRuntimeCall(callee, None, name);
3393 // Calls which may throw must have operand bundles indicating which funclet
3394 // they are nested within.
3396 getBundlesForFunclet(llvm::Value *Callee, llvm::Instruction *CurrentFuncletPad,
3397 SmallVectorImpl<llvm::OperandBundleDef> &BundleList) {
3398 // There is no need for a funclet operand bundle if we aren't inside a
3400 if (!CurrentFuncletPad)
3403 // Skip intrinsics which cannot throw.
3404 auto *CalleeFn = dyn_cast<llvm::Function>(Callee->stripPointerCasts());
3405 if (CalleeFn && CalleeFn->isIntrinsic() && CalleeFn->doesNotThrow())
3408 BundleList.emplace_back("funclet", CurrentFuncletPad);
3411 /// Emits a simple call (never an invoke) to the given runtime function.
3413 CodeGenFunction::EmitRuntimeCall(llvm::Value *callee,
3414 ArrayRef<llvm::Value*> args,
3415 const llvm::Twine &name) {
3416 SmallVector<llvm::OperandBundleDef, 1> BundleList;
3417 getBundlesForFunclet(callee, CurrentFuncletPad, BundleList);
3419 llvm::CallInst *call = Builder.CreateCall(callee, args, BundleList, name);
3420 call->setCallingConv(getRuntimeCC());
3424 /// Emits a call or invoke to the given noreturn runtime function.
3425 void CodeGenFunction::EmitNoreturnRuntimeCallOrInvoke(llvm::Value *callee,
3426 ArrayRef<llvm::Value*> args) {
3427 SmallVector<llvm::OperandBundleDef, 1> BundleList;
3428 getBundlesForFunclet(callee, CurrentFuncletPad, BundleList);
3430 if (getInvokeDest()) {
3431 llvm::InvokeInst *invoke =
3432 Builder.CreateInvoke(callee,
3433 getUnreachableBlock(),
3437 invoke->setDoesNotReturn();
3438 invoke->setCallingConv(getRuntimeCC());
3440 llvm::CallInst *call = Builder.CreateCall(callee, args, BundleList);
3441 call->setDoesNotReturn();
3442 call->setCallingConv(getRuntimeCC());
3443 Builder.CreateUnreachable();
3447 /// Emits a call or invoke instruction to the given nullary runtime function.
3449 CodeGenFunction::EmitRuntimeCallOrInvoke(llvm::Value *callee,
3450 const Twine &name) {
3451 return EmitRuntimeCallOrInvoke(callee, None, name);
3454 /// Emits a call or invoke instruction to the given runtime function.
3456 CodeGenFunction::EmitRuntimeCallOrInvoke(llvm::Value *callee,
3457 ArrayRef<llvm::Value*> args,
3458 const Twine &name) {
3459 llvm::CallSite callSite = EmitCallOrInvoke(callee, args, name);
3460 callSite.setCallingConv(getRuntimeCC());
3464 /// Emits a call or invoke instruction to the given function, depending
3465 /// on the current state of the EH stack.
3467 CodeGenFunction::EmitCallOrInvoke(llvm::Value *Callee,
3468 ArrayRef<llvm::Value *> Args,
3469 const Twine &Name) {
3470 llvm::BasicBlock *InvokeDest = getInvokeDest();
3471 SmallVector<llvm::OperandBundleDef, 1> BundleList;
3472 getBundlesForFunclet(Callee, CurrentFuncletPad, BundleList);
3474 llvm::Instruction *Inst;
3476 Inst = Builder.CreateCall(Callee, Args, BundleList, Name);
3478 llvm::BasicBlock *ContBB = createBasicBlock("invoke.cont");
3479 Inst = Builder.CreateInvoke(Callee, ContBB, InvokeDest, Args, BundleList,
3484 // In ObjC ARC mode with no ObjC ARC exception safety, tell the ARC
3485 // optimizer it can aggressively ignore unwind edges.
3486 if (CGM.getLangOpts().ObjCAutoRefCount)
3487 AddObjCARCExceptionMetadata(Inst);
3489 return llvm::CallSite(Inst);
3492 /// \brief Store a non-aggregate value to an address to initialize it. For
3493 /// initialization, a non-atomic store will be used.
3494 static void EmitInitStoreOfNonAggregate(CodeGenFunction &CGF, RValue Src,
3497 CGF.EmitStoreOfScalar(Src.getScalarVal(), Dst, /*init=*/true);
3499 CGF.EmitStoreOfComplex(Src.getComplexVal(), Dst, /*init=*/true);
3502 void CodeGenFunction::deferPlaceholderReplacement(llvm::Instruction *Old,
3504 DeferredReplacements.push_back(std::make_pair(Old, New));
3507 RValue CodeGenFunction::EmitCall(const CGFunctionInfo &CallInfo,
3508 llvm::Value *Callee,
3509 ReturnValueSlot ReturnValue,
3510 const CallArgList &CallArgs,
3511 CGCalleeInfo CalleeInfo,
3512 llvm::Instruction **callOrInvoke) {
3513 // FIXME: We no longer need the types from CallArgs; lift up and simplify.
3515 // Handle struct-return functions by passing a pointer to the
3516 // location that we would like to return into.
3517 QualType RetTy = CallInfo.getReturnType();
3518 const ABIArgInfo &RetAI = CallInfo.getReturnInfo();
3520 llvm::FunctionType *IRFuncTy =
3521 cast<llvm::FunctionType>(
3522 cast<llvm::PointerType>(Callee->getType())->getElementType());
3524 // If we're using inalloca, insert the allocation after the stack save.
3525 // FIXME: Do this earlier rather than hacking it in here!
3526 Address ArgMemory = Address::invalid();
3527 const llvm::StructLayout *ArgMemoryLayout = nullptr;
3528 if (llvm::StructType *ArgStruct = CallInfo.getArgStruct()) {
3529 ArgMemoryLayout = CGM.getDataLayout().getStructLayout(ArgStruct);
3530 llvm::Instruction *IP = CallArgs.getStackBase();
3531 llvm::AllocaInst *AI;
3533 IP = IP->getNextNode();
3534 AI = new llvm::AllocaInst(ArgStruct, "argmem", IP);
3536 AI = CreateTempAlloca(ArgStruct, "argmem");
3538 auto Align = CallInfo.getArgStructAlignment();
3539 AI->setAlignment(Align.getQuantity());
3540 AI->setUsedWithInAlloca(true);
3541 assert(AI->isUsedWithInAlloca() && !AI->isStaticAlloca());
3542 ArgMemory = Address(AI, Align);
3545 // Helper function to drill into the inalloca allocation.
3546 auto createInAllocaStructGEP = [&](unsigned FieldIndex) -> Address {
3548 CharUnits::fromQuantity(ArgMemoryLayout->getElementOffset(FieldIndex));
3549 return Builder.CreateStructGEP(ArgMemory, FieldIndex, FieldOffset);
3552 ClangToLLVMArgMapping IRFunctionArgs(CGM.getContext(), CallInfo);
3553 SmallVector<llvm::Value *, 16> IRCallArgs(IRFunctionArgs.totalIRArgs());
3555 // If the call returns a temporary with struct return, create a temporary
3556 // alloca to hold the result, unless one is given to us.
3557 Address SRetPtr = Address::invalid();
3558 size_t UnusedReturnSize = 0;
3559 if (RetAI.isIndirect() || RetAI.isInAlloca() || RetAI.isCoerceAndExpand()) {
3560 if (!ReturnValue.isNull()) {
3561 SRetPtr = ReturnValue.getValue();
3563 SRetPtr = CreateMemTemp(RetTy);
3564 if (HaveInsertPoint() && ReturnValue.isUnused()) {
3566 CGM.getDataLayout().getTypeAllocSize(ConvertTypeForMem(RetTy));
3567 if (EmitLifetimeStart(size, SRetPtr.getPointer()))
3568 UnusedReturnSize = size;
3571 if (IRFunctionArgs.hasSRetArg()) {
3572 IRCallArgs[IRFunctionArgs.getSRetArgNo()] = SRetPtr.getPointer();
3573 } else if (RetAI.isInAlloca()) {
3574 Address Addr = createInAllocaStructGEP(RetAI.getInAllocaFieldIndex());
3575 Builder.CreateStore(SRetPtr.getPointer(), Addr);
3579 Address swiftErrorTemp = Address::invalid();
3580 Address swiftErrorArg = Address::invalid();
3582 assert(CallInfo.arg_size() == CallArgs.size() &&
3583 "Mismatch between function signature & arguments.");
3585 CGFunctionInfo::const_arg_iterator info_it = CallInfo.arg_begin();
3586 for (CallArgList::const_iterator I = CallArgs.begin(), E = CallArgs.end();
3587 I != E; ++I, ++info_it, ++ArgNo) {
3588 const ABIArgInfo &ArgInfo = info_it->info;
3591 // Insert a padding argument to ensure proper alignment.
3592 if (IRFunctionArgs.hasPaddingArg(ArgNo))
3593 IRCallArgs[IRFunctionArgs.getPaddingArgNo(ArgNo)] =
3594 llvm::UndefValue::get(ArgInfo.getPaddingType());
3596 unsigned FirstIRArg, NumIRArgs;
3597 std::tie(FirstIRArg, NumIRArgs) = IRFunctionArgs.getIRArgs(ArgNo);
3599 switch (ArgInfo.getKind()) {
3600 case ABIArgInfo::InAlloca: {
3601 assert(NumIRArgs == 0);
3602 assert(getTarget().getTriple().getArch() == llvm::Triple::x86);
3603 if (RV.isAggregate()) {
3604 // Replace the placeholder with the appropriate argument slot GEP.
3605 llvm::Instruction *Placeholder =
3606 cast<llvm::Instruction>(RV.getAggregatePointer());
3607 CGBuilderTy::InsertPoint IP = Builder.saveIP();
3608 Builder.SetInsertPoint(Placeholder);
3609 Address Addr = createInAllocaStructGEP(ArgInfo.getInAllocaFieldIndex());
3610 Builder.restoreIP(IP);
3611 deferPlaceholderReplacement(Placeholder, Addr.getPointer());
3613 // Store the RValue into the argument struct.
3614 Address Addr = createInAllocaStructGEP(ArgInfo.getInAllocaFieldIndex());
3615 unsigned AS = Addr.getType()->getPointerAddressSpace();
3616 llvm::Type *MemType = ConvertTypeForMem(I->Ty)->getPointerTo(AS);
3617 // There are some cases where a trivial bitcast is not avoidable. The
3618 // definition of a type later in a translation unit may change it's type
3619 // from {}* to (%struct.foo*)*.
3620 if (Addr.getType() != MemType)
3621 Addr = Builder.CreateBitCast(Addr, MemType);
3622 LValue argLV = MakeAddrLValue(Addr, I->Ty);
3623 EmitInitStoreOfNonAggregate(*this, RV, argLV);
3628 case ABIArgInfo::Indirect: {
3629 assert(NumIRArgs == 1);
3630 if (RV.isScalar() || RV.isComplex()) {
3631 // Make a temporary alloca to pass the argument.
3632 Address Addr = CreateMemTemp(I->Ty, ArgInfo.getIndirectAlign());
3633 IRCallArgs[FirstIRArg] = Addr.getPointer();
3635 LValue argLV = MakeAddrLValue(Addr, I->Ty);
3636 EmitInitStoreOfNonAggregate(*this, RV, argLV);
3638 // We want to avoid creating an unnecessary temporary+copy here;
3639 // however, we need one in three cases:
3640 // 1. If the argument is not byval, and we are required to copy the
3641 // source. (This case doesn't occur on any common architecture.)
3642 // 2. If the argument is byval, RV is not sufficiently aligned, and
3643 // we cannot force it to be sufficiently aligned.
3644 // 3. If the argument is byval, but RV is located in an address space
3645 // different than that of the argument (0).
3646 Address Addr = RV.getAggregateAddress();
3647 CharUnits Align = ArgInfo.getIndirectAlign();
3648 const llvm::DataLayout *TD = &CGM.getDataLayout();
3649 const unsigned RVAddrSpace = Addr.getType()->getAddressSpace();
3650 const unsigned ArgAddrSpace =
3651 (FirstIRArg < IRFuncTy->getNumParams()
3652 ? IRFuncTy->getParamType(FirstIRArg)->getPointerAddressSpace()
3654 if ((!ArgInfo.getIndirectByVal() && I->NeedsCopy) ||
3655 (ArgInfo.getIndirectByVal() && Addr.getAlignment() < Align &&
3656 llvm::getOrEnforceKnownAlignment(Addr.getPointer(),
3657 Align.getQuantity(), *TD)
3658 < Align.getQuantity()) ||
3659 (ArgInfo.getIndirectByVal() && (RVAddrSpace != ArgAddrSpace))) {
3660 // Create an aligned temporary, and copy to it.
3661 Address AI = CreateMemTemp(I->Ty, ArgInfo.getIndirectAlign());
3662 IRCallArgs[FirstIRArg] = AI.getPointer();
3663 EmitAggregateCopy(AI, Addr, I->Ty, RV.isVolatileQualified());
3665 // Skip the extra memcpy call.
3666 IRCallArgs[FirstIRArg] = Addr.getPointer();
3672 case ABIArgInfo::Ignore:
3673 assert(NumIRArgs == 0);
3676 case ABIArgInfo::Extend:
3677 case ABIArgInfo::Direct: {
3678 if (!isa<llvm::StructType>(ArgInfo.getCoerceToType()) &&
3679 ArgInfo.getCoerceToType() == ConvertType(info_it->type) &&
3680 ArgInfo.getDirectOffset() == 0) {
3681 assert(NumIRArgs == 1);
3684 V = RV.getScalarVal();
3686 V = Builder.CreateLoad(RV.getAggregateAddress());
3688 // Implement swifterror by copying into a new swifterror argument.
3689 // We'll write back in the normal path out of the call.
3690 if (CallInfo.getExtParameterInfo(ArgNo).getABI()
3691 == ParameterABI::SwiftErrorResult) {
3692 assert(!swiftErrorTemp.isValid() && "multiple swifterror args");
3694 QualType pointeeTy = I->Ty->getPointeeType();
3696 Address(V, getContext().getTypeAlignInChars(pointeeTy));
3699 CreateMemTemp(pointeeTy, getPointerAlign(), "swifterror.temp");
3700 V = swiftErrorTemp.getPointer();
3701 cast<llvm::AllocaInst>(V)->setSwiftError(true);
3703 llvm::Value *errorValue = Builder.CreateLoad(swiftErrorArg);
3704 Builder.CreateStore(errorValue, swiftErrorTemp);
3707 // We might have to widen integers, but we should never truncate.
3708 if (ArgInfo.getCoerceToType() != V->getType() &&
3709 V->getType()->isIntegerTy())
3710 V = Builder.CreateZExt(V, ArgInfo.getCoerceToType());
3712 // If the argument doesn't match, perform a bitcast to coerce it. This
3713 // can happen due to trivial type mismatches.
3714 if (FirstIRArg < IRFuncTy->getNumParams() &&
3715 V->getType() != IRFuncTy->getParamType(FirstIRArg))
3716 V = Builder.CreateBitCast(V, IRFuncTy->getParamType(FirstIRArg));
3718 IRCallArgs[FirstIRArg] = V;
3722 // FIXME: Avoid the conversion through memory if possible.
3723 Address Src = Address::invalid();
3724 if (RV.isScalar() || RV.isComplex()) {
3725 Src = CreateMemTemp(I->Ty, "coerce");
3726 LValue SrcLV = MakeAddrLValue(Src, I->Ty);
3727 EmitInitStoreOfNonAggregate(*this, RV, SrcLV);
3729 Src = RV.getAggregateAddress();
3732 // If the value is offset in memory, apply the offset now.
3733 Src = emitAddressAtOffset(*this, Src, ArgInfo);
3735 // Fast-isel and the optimizer generally like scalar values better than
3736 // FCAs, so we flatten them if this is safe to do for this argument.
3737 llvm::StructType *STy =
3738 dyn_cast<llvm::StructType>(ArgInfo.getCoerceToType());
3739 if (STy && ArgInfo.isDirect() && ArgInfo.getCanBeFlattened()) {
3740 llvm::Type *SrcTy = Src.getType()->getElementType();
3741 uint64_t SrcSize = CGM.getDataLayout().getTypeAllocSize(SrcTy);
3742 uint64_t DstSize = CGM.getDataLayout().getTypeAllocSize(STy);
3744 // If the source type is smaller than the destination type of the
3745 // coerce-to logic, copy the source value into a temp alloca the size
3746 // of the destination type to allow loading all of it. The bits past
3747 // the source value are left undef.
3748 if (SrcSize < DstSize) {
3750 = CreateTempAlloca(STy, Src.getAlignment(),
3751 Src.getName() + ".coerce");
3752 Builder.CreateMemCpy(TempAlloca, Src, SrcSize);
3755 Src = Builder.CreateBitCast(Src, llvm::PointerType::getUnqual(STy));
3758 auto SrcLayout = CGM.getDataLayout().getStructLayout(STy);
3759 assert(NumIRArgs == STy->getNumElements());
3760 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
3761 auto Offset = CharUnits::fromQuantity(SrcLayout->getElementOffset(i));
3762 Address EltPtr = Builder.CreateStructGEP(Src, i, Offset);
3763 llvm::Value *LI = Builder.CreateLoad(EltPtr);
3764 IRCallArgs[FirstIRArg + i] = LI;
3767 // In the simple case, just pass the coerced loaded value.
3768 assert(NumIRArgs == 1);
3769 IRCallArgs[FirstIRArg] =
3770 CreateCoercedLoad(Src, ArgInfo.getCoerceToType(), *this);
3776 case ABIArgInfo::CoerceAndExpand: {
3777 auto coercionType = ArgInfo.getCoerceAndExpandType();
3778 auto layout = CGM.getDataLayout().getStructLayout(coercionType);
3780 llvm::Value *tempSize = nullptr;
3781 Address addr = Address::invalid();
3782 if (RV.isAggregate()) {
3783 addr = RV.getAggregateAddress();
3785 assert(RV.isScalar()); // complex should always just be direct
3787 llvm::Type *scalarType = RV.getScalarVal()->getType();
3788 auto scalarSize = CGM.getDataLayout().getTypeAllocSize(scalarType);
3789 auto scalarAlign = CGM.getDataLayout().getPrefTypeAlignment(scalarType);
3791 tempSize = llvm::ConstantInt::get(CGM.Int64Ty, scalarSize);
3793 // Materialize to a temporary.
3794 addr = CreateTempAlloca(RV.getScalarVal()->getType(),
3795 CharUnits::fromQuantity(std::max(layout->getAlignment(),
3797 EmitLifetimeStart(scalarSize, addr.getPointer());
3799 Builder.CreateStore(RV.getScalarVal(), addr);
3802 addr = Builder.CreateElementBitCast(addr, coercionType);
3804 unsigned IRArgPos = FirstIRArg;
3805 for (unsigned i = 0, e = coercionType->getNumElements(); i != e; ++i) {
3806 llvm::Type *eltType = coercionType->getElementType(i);
3807 if (ABIArgInfo::isPaddingForCoerceAndExpand(eltType)) continue;
3808 Address eltAddr = Builder.CreateStructGEP(addr, i, layout);
3809 llvm::Value *elt = Builder.CreateLoad(eltAddr);
3810 IRCallArgs[IRArgPos++] = elt;
3812 assert(IRArgPos == FirstIRArg + NumIRArgs);
3815 EmitLifetimeEnd(tempSize, addr.getPointer());
3821 case ABIArgInfo::Expand:
3822 unsigned IRArgPos = FirstIRArg;
3823 ExpandTypeToArgs(I->Ty, RV, IRFuncTy, IRCallArgs, IRArgPos);
3824 assert(IRArgPos == FirstIRArg + NumIRArgs);
3829 if (ArgMemory.isValid()) {
3830 llvm::Value *Arg = ArgMemory.getPointer();
3831 if (CallInfo.isVariadic()) {
3832 // When passing non-POD arguments by value to variadic functions, we will
3833 // end up with a variadic prototype and an inalloca call site. In such
3834 // cases, we can't do any parameter mismatch checks. Give up and bitcast
3837 cast<llvm::PointerType>(Callee->getType())->getAddressSpace();
3838 Callee = Builder.CreateBitCast(
3839 Callee, getTypes().GetFunctionType(CallInfo)->getPointerTo(CalleeAS));
3841 llvm::Type *LastParamTy =
3842 IRFuncTy->getParamType(IRFuncTy->getNumParams() - 1);
3843 if (Arg->getType() != LastParamTy) {
3845 // Assert that these structs have equivalent element types.
3846 llvm::StructType *FullTy = CallInfo.getArgStruct();
3847 llvm::StructType *DeclaredTy = cast<llvm::StructType>(
3848 cast<llvm::PointerType>(LastParamTy)->getElementType());
3849 assert(DeclaredTy->getNumElements() == FullTy->getNumElements());
3850 for (llvm::StructType::element_iterator DI = DeclaredTy->element_begin(),
3851 DE = DeclaredTy->element_end(),
3852 FI = FullTy->element_begin();
3853 DI != DE; ++DI, ++FI)
3856 Arg = Builder.CreateBitCast(Arg, LastParamTy);
3859 assert(IRFunctionArgs.hasInallocaArg());
3860 IRCallArgs[IRFunctionArgs.getInallocaArgNo()] = Arg;
3863 if (!CallArgs.getCleanupsToDeactivate().empty())
3864 deactivateArgCleanupsBeforeCall(*this, CallArgs);
3866 // If the callee is a bitcast of a function to a varargs pointer to function
3867 // type, check to see if we can remove the bitcast. This handles some cases
3868 // with unprototyped functions.
3869 if (llvm::ConstantExpr *CE = dyn_cast<llvm::ConstantExpr>(Callee))
3870 if (llvm::Function *CalleeF = dyn_cast<llvm::Function>(CE->getOperand(0))) {
3871 llvm::PointerType *CurPT=cast<llvm::PointerType>(Callee->getType());
3872 llvm::FunctionType *CurFT =
3873 cast<llvm::FunctionType>(CurPT->getElementType());
3874 llvm::FunctionType *ActualFT = CalleeF->getFunctionType();
3876 if (CE->getOpcode() == llvm::Instruction::BitCast &&
3877 ActualFT->getReturnType() == CurFT->getReturnType() &&
3878 ActualFT->getNumParams() == CurFT->getNumParams() &&
3879 ActualFT->getNumParams() == IRCallArgs.size() &&
3880 (CurFT->isVarArg() || !ActualFT->isVarArg())) {
3881 bool ArgsMatch = true;
3882 for (unsigned i = 0, e = ActualFT->getNumParams(); i != e; ++i)
3883 if (ActualFT->getParamType(i) != CurFT->getParamType(i)) {
3888 // Strip the cast if we can get away with it. This is a nice cleanup,
3889 // but also allows us to inline the function at -O0 if it is marked
3896 assert(IRCallArgs.size() == IRFuncTy->getNumParams() || IRFuncTy->isVarArg());
3897 for (unsigned i = 0; i < IRCallArgs.size(); ++i) {
3898 // Inalloca argument can have different type.
3899 if (IRFunctionArgs.hasInallocaArg() &&
3900 i == IRFunctionArgs.getInallocaArgNo())
3902 if (i < IRFuncTy->getNumParams())
3903 assert(IRCallArgs[i]->getType() == IRFuncTy->getParamType(i));
3906 unsigned CallingConv;
3907 CodeGen::AttributeListType AttributeList;
3908 CGM.ConstructAttributeList(Callee->getName(), CallInfo, CalleeInfo,
3909 AttributeList, CallingConv,
3910 /*AttrOnCallSite=*/true);
3911 llvm::AttributeSet Attrs = llvm::AttributeSet::get(getLLVMContext(),
3915 if (currentFunctionUsesSEHTry()) {
3916 // SEH cares about asynchronous exceptions, everything can "throw."
3917 CannotThrow = false;
3918 } else if (isCleanupPadScope() &&
3919 EHPersonality::get(*this).isMSVCXXPersonality()) {
3920 // The MSVC++ personality will implicitly terminate the program if an
3921 // exception is thrown. An unwind edge cannot be reached.
3924 // Otherwise, nowunind callsites will never throw.
3925 CannotThrow = Attrs.hasAttribute(llvm::AttributeSet::FunctionIndex,
3926 llvm::Attribute::NoUnwind);
3928 llvm::BasicBlock *InvokeDest = CannotThrow ? nullptr : getInvokeDest();
3930 SmallVector<llvm::OperandBundleDef, 1> BundleList;
3931 getBundlesForFunclet(Callee, CurrentFuncletPad, BundleList);
3935 CS = Builder.CreateCall(Callee, IRCallArgs, BundleList);
3937 llvm::BasicBlock *Cont = createBasicBlock("invoke.cont");
3938 CS = Builder.CreateInvoke(Callee, Cont, InvokeDest, IRCallArgs,
3943 *callOrInvoke = CS.getInstruction();
3945 if (CurCodeDecl && CurCodeDecl->hasAttr<FlattenAttr>() &&
3946 !CS.hasFnAttr(llvm::Attribute::NoInline))
3948 Attrs.addAttribute(getLLVMContext(), llvm::AttributeSet::FunctionIndex,
3949 llvm::Attribute::AlwaysInline);
3951 // Disable inlining inside SEH __try blocks.
3952 if (isSEHTryScope())
3954 Attrs.addAttribute(getLLVMContext(), llvm::AttributeSet::FunctionIndex,
3955 llvm::Attribute::NoInline);
3957 CS.setAttributes(Attrs);
3958 CS.setCallingConv(static_cast<llvm::CallingConv::ID>(CallingConv));
3960 // Insert instrumentation or attach profile metadata at indirect call sites.
3961 // For more details, see the comment before the definition of
3962 // IPVK_IndirectCallTarget in InstrProfData.inc.
3963 if (!CS.getCalledFunction())
3964 PGO.valueProfile(Builder, llvm::IPVK_IndirectCallTarget,
3965 CS.getInstruction(), Callee);
3967 // In ObjC ARC mode with no ObjC ARC exception safety, tell the ARC
3968 // optimizer it can aggressively ignore unwind edges.
3969 if (CGM.getLangOpts().ObjCAutoRefCount)
3970 AddObjCARCExceptionMetadata(CS.getInstruction());
3972 // If the call doesn't return, finish the basic block and clear the
3973 // insertion point; this allows the rest of IRgen to discard
3974 // unreachable code.
3975 if (CS.doesNotReturn()) {
3976 if (UnusedReturnSize)
3977 EmitLifetimeEnd(llvm::ConstantInt::get(Int64Ty, UnusedReturnSize),
3978 SRetPtr.getPointer());
3980 Builder.CreateUnreachable();
3981 Builder.ClearInsertionPoint();
3983 // FIXME: For now, emit a dummy basic block because expr emitters in
3984 // generally are not ready to handle emitting expressions at unreachable
3986 EnsureInsertPoint();
3988 // Return a reasonable RValue.
3989 return GetUndefRValue(RetTy);
3992 llvm::Instruction *CI = CS.getInstruction();
3993 if (!CI->getType()->isVoidTy())
3994 CI->setName("call");
3996 // Perform the swifterror writeback.
3997 if (swiftErrorTemp.isValid()) {
3998 llvm::Value *errorResult = Builder.CreateLoad(swiftErrorTemp);
3999 Builder.CreateStore(errorResult, swiftErrorArg);
4002 // Emit any writebacks immediately. Arguably this should happen
4003 // after any return-value munging.
4004 if (CallArgs.hasWritebacks())
4005 emitWritebacks(*this, CallArgs);
4007 // The stack cleanup for inalloca arguments has to run out of the normal
4008 // lexical order, so deactivate it and run it manually here.
4009 CallArgs.freeArgumentMemory(*this);
4011 if (llvm::CallInst *Call = dyn_cast<llvm::CallInst>(CI)) {
4012 const Decl *TargetDecl = CalleeInfo.getCalleeDecl();
4013 if (TargetDecl && TargetDecl->hasAttr<NotTailCalledAttr>())
4014 Call->setTailCallKind(llvm::CallInst::TCK_NoTail);
4018 switch (RetAI.getKind()) {
4019 case ABIArgInfo::CoerceAndExpand: {
4020 auto coercionType = RetAI.getCoerceAndExpandType();
4021 auto layout = CGM.getDataLayout().getStructLayout(coercionType);
4023 Address addr = SRetPtr;
4024 addr = Builder.CreateElementBitCast(addr, coercionType);
4026 assert(CI->getType() == RetAI.getUnpaddedCoerceAndExpandType());
4027 bool requiresExtract = isa<llvm::StructType>(CI->getType());
4029 unsigned unpaddedIndex = 0;
4030 for (unsigned i = 0, e = coercionType->getNumElements(); i != e; ++i) {
4031 llvm::Type *eltType = coercionType->getElementType(i);
4032 if (ABIArgInfo::isPaddingForCoerceAndExpand(eltType)) continue;
4033 Address eltAddr = Builder.CreateStructGEP(addr, i, layout);
4034 llvm::Value *elt = CI;
4035 if (requiresExtract)
4036 elt = Builder.CreateExtractValue(elt, unpaddedIndex++);
4038 assert(unpaddedIndex == 0);
4039 Builder.CreateStore(elt, eltAddr);
4044 case ABIArgInfo::InAlloca:
4045 case ABIArgInfo::Indirect: {
4046 RValue ret = convertTempToRValue(SRetPtr, RetTy, SourceLocation());
4047 if (UnusedReturnSize)
4048 EmitLifetimeEnd(llvm::ConstantInt::get(Int64Ty, UnusedReturnSize),
4049 SRetPtr.getPointer());
4053 case ABIArgInfo::Ignore:
4054 // If we are ignoring an argument that had a result, make sure to
4055 // construct the appropriate return value for our caller.
4056 return GetUndefRValue(RetTy);
4058 case ABIArgInfo::Extend:
4059 case ABIArgInfo::Direct: {
4060 llvm::Type *RetIRTy = ConvertType(RetTy);
4061 if (RetAI.getCoerceToType() == RetIRTy && RetAI.getDirectOffset() == 0) {
4062 switch (getEvaluationKind(RetTy)) {
4064 llvm::Value *Real = Builder.CreateExtractValue(CI, 0);
4065 llvm::Value *Imag = Builder.CreateExtractValue(CI, 1);
4066 return RValue::getComplex(std::make_pair(Real, Imag));
4068 case TEK_Aggregate: {
4069 Address DestPtr = ReturnValue.getValue();
4070 bool DestIsVolatile = ReturnValue.isVolatile();
4072 if (!DestPtr.isValid()) {
4073 DestPtr = CreateMemTemp(RetTy, "agg.tmp");
4074 DestIsVolatile = false;
4076 BuildAggStore(*this, CI, DestPtr, DestIsVolatile);
4077 return RValue::getAggregate(DestPtr);
4080 // If the argument doesn't match, perform a bitcast to coerce it. This
4081 // can happen due to trivial type mismatches.
4082 llvm::Value *V = CI;
4083 if (V->getType() != RetIRTy)
4084 V = Builder.CreateBitCast(V, RetIRTy);
4085 return RValue::get(V);
4088 llvm_unreachable("bad evaluation kind");
4091 Address DestPtr = ReturnValue.getValue();
4092 bool DestIsVolatile = ReturnValue.isVolatile();
4094 if (!DestPtr.isValid()) {
4095 DestPtr = CreateMemTemp(RetTy, "coerce");
4096 DestIsVolatile = false;
4099 // If the value is offset in memory, apply the offset now.
4100 Address StorePtr = emitAddressAtOffset(*this, DestPtr, RetAI);
4101 CreateCoercedStore(CI, StorePtr, DestIsVolatile, *this);
4103 return convertTempToRValue(DestPtr, RetTy, SourceLocation());
4106 case ABIArgInfo::Expand:
4107 llvm_unreachable("Invalid ABI kind for return argument");
4110 llvm_unreachable("Unhandled ABIArgInfo::Kind");
4113 const Decl *TargetDecl = CalleeInfo.getCalleeDecl();
4115 if (Ret.isScalar() && TargetDecl) {
4116 if (const auto *AA = TargetDecl->getAttr<AssumeAlignedAttr>()) {
4117 llvm::Value *OffsetValue = nullptr;
4118 if (const auto *Offset = AA->getOffset())
4119 OffsetValue = EmitScalarExpr(Offset);
4121 llvm::Value *Alignment = EmitScalarExpr(AA->getAlignment());
4122 llvm::ConstantInt *AlignmentCI = cast<llvm::ConstantInt>(Alignment);
4123 EmitAlignmentAssumption(Ret.getScalarVal(), AlignmentCI->getZExtValue(),
4131 /* VarArg handling */
4133 Address CodeGenFunction::EmitVAArg(VAArgExpr *VE, Address &VAListAddr) {
4134 VAListAddr = VE->isMicrosoftABI()
4135 ? EmitMSVAListRef(VE->getSubExpr())
4136 : EmitVAListRef(VE->getSubExpr());
4137 QualType Ty = VE->getType();
4138 if (VE->isMicrosoftABI())
4139 return CGM.getTypes().getABIInfo().EmitMSVAArg(*this, VAListAddr, Ty);
4140 return CGM.getTypes().getABIInfo().EmitVAArg(*this, VAListAddr, Ty);