1 //===--- CGCall.cpp - Encapsulate calling convention details --------------===//
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7 //===----------------------------------------------------------------------===//
9 // These classes wrap the information about a call or function
10 // definition used to handle ABI compliancy.
12 //===----------------------------------------------------------------------===//
18 #include "CGCleanup.h"
19 #include "CodeGenFunction.h"
20 #include "CodeGenModule.h"
21 #include "TargetInfo.h"
22 #include "clang/AST/Decl.h"
23 #include "clang/AST/DeclCXX.h"
24 #include "clang/AST/DeclObjC.h"
25 #include "clang/Basic/CodeGenOptions.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 "llvm/ADT/StringExtras.h"
31 #include "llvm/Transforms/Utils/Local.h"
32 #include "llvm/Analysis/ValueTracking.h"
33 #include "llvm/IR/Attributes.h"
34 #include "llvm/IR/CallingConv.h"
35 #include "llvm/IR/DataLayout.h"
36 #include "llvm/IR/InlineAsm.h"
37 #include "llvm/IR/IntrinsicInst.h"
38 #include "llvm/IR/Intrinsics.h"
39 using namespace clang;
40 using namespace CodeGen;
44 unsigned CodeGenTypes::ClangCallConvToLLVMCallConv(CallingConv CC) {
46 default: return llvm::CallingConv::C;
47 case CC_X86StdCall: return llvm::CallingConv::X86_StdCall;
48 case CC_X86FastCall: return llvm::CallingConv::X86_FastCall;
49 case CC_X86RegCall: return llvm::CallingConv::X86_RegCall;
50 case CC_X86ThisCall: return llvm::CallingConv::X86_ThisCall;
51 case CC_Win64: return llvm::CallingConv::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_AArch64VectorCall: return llvm::CallingConv::AArch64_VectorCall;
61 case CC_SpirFunction: return llvm::CallingConv::SPIR_FUNC;
62 case CC_OpenCLKernel: return CGM.getTargetCodeGenInfo().getOpenCLKernelCallingConv();
63 case CC_PreserveMost: return llvm::CallingConv::PreserveMost;
64 case CC_PreserveAll: return llvm::CallingConv::PreserveAll;
65 case CC_Swift: return llvm::CallingConv::Swift;
69 /// Derives the 'this' type for codegen purposes, i.e. ignoring method CVR
70 /// qualification. Either or both of RD and MD may be null. A null RD indicates
71 /// that there is no meaningful 'this' type, and a null MD can occur when
72 /// calling a method pointer.
73 CanQualType CodeGenTypes::DeriveThisType(const CXXRecordDecl *RD,
74 const CXXMethodDecl *MD) {
77 RecTy = Context.getTagDeclType(RD)->getCanonicalTypeInternal();
79 RecTy = Context.VoidTy;
82 RecTy = Context.getAddrSpaceQualType(RecTy, MD->getMethodQualifiers().getAddressSpace());
83 return Context.getPointerType(CanQualType::CreateUnsafe(RecTy));
86 /// Returns the canonical formal type of the given C++ method.
87 static CanQual<FunctionProtoType> GetFormalType(const CXXMethodDecl *MD) {
88 return MD->getType()->getCanonicalTypeUnqualified()
89 .getAs<FunctionProtoType>();
92 /// Returns the "extra-canonicalized" return type, which discards
93 /// qualifiers on the return type. Codegen doesn't care about them,
94 /// and it makes ABI code a little easier to be able to assume that
95 /// all parameter and return types are top-level unqualified.
96 static CanQualType GetReturnType(QualType RetTy) {
97 return RetTy->getCanonicalTypeUnqualified().getUnqualifiedType();
100 /// Arrange the argument and result information for a value of the given
101 /// unprototyped freestanding function type.
102 const CGFunctionInfo &
103 CodeGenTypes::arrangeFreeFunctionType(CanQual<FunctionNoProtoType> FTNP) {
104 // When translating an unprototyped function type, always use a
106 return arrangeLLVMFunctionInfo(FTNP->getReturnType().getUnqualifiedType(),
107 /*instanceMethod=*/false,
108 /*chainCall=*/false, None,
109 FTNP->getExtInfo(), {}, RequiredArgs(0));
112 static void addExtParameterInfosForCall(
113 llvm::SmallVectorImpl<FunctionProtoType::ExtParameterInfo> ¶mInfos,
114 const FunctionProtoType *proto,
116 unsigned totalArgs) {
117 assert(proto->hasExtParameterInfos());
118 assert(paramInfos.size() <= prefixArgs);
119 assert(proto->getNumParams() + prefixArgs <= totalArgs);
121 paramInfos.reserve(totalArgs);
123 // Add default infos for any prefix args that don't already have infos.
124 paramInfos.resize(prefixArgs);
126 // Add infos for the prototype.
127 for (const auto &ParamInfo : proto->getExtParameterInfos()) {
128 paramInfos.push_back(ParamInfo);
129 // pass_object_size params have no parameter info.
130 if (ParamInfo.hasPassObjectSize())
131 paramInfos.emplace_back();
134 assert(paramInfos.size() <= totalArgs &&
135 "Did we forget to insert pass_object_size args?");
136 // Add default infos for the variadic and/or suffix arguments.
137 paramInfos.resize(totalArgs);
140 /// Adds the formal parameters in FPT to the given prefix. If any parameter in
141 /// FPT has pass_object_size attrs, then we'll add parameters for those, too.
142 static void appendParameterTypes(const CodeGenTypes &CGT,
143 SmallVectorImpl<CanQualType> &prefix,
144 SmallVectorImpl<FunctionProtoType::ExtParameterInfo> ¶mInfos,
145 CanQual<FunctionProtoType> FPT) {
146 // Fast path: don't touch param info if we don't need to.
147 if (!FPT->hasExtParameterInfos()) {
148 assert(paramInfos.empty() &&
149 "We have paramInfos, but the prototype doesn't?");
150 prefix.append(FPT->param_type_begin(), FPT->param_type_end());
154 unsigned PrefixSize = prefix.size();
155 // In the vast majority of cases, we'll have precisely FPT->getNumParams()
156 // parameters; the only thing that can change this is the presence of
157 // pass_object_size. So, we preallocate for the common case.
158 prefix.reserve(prefix.size() + FPT->getNumParams());
160 auto ExtInfos = FPT->getExtParameterInfos();
161 assert(ExtInfos.size() == FPT->getNumParams());
162 for (unsigned I = 0, E = FPT->getNumParams(); I != E; ++I) {
163 prefix.push_back(FPT->getParamType(I));
164 if (ExtInfos[I].hasPassObjectSize())
165 prefix.push_back(CGT.getContext().getSizeType());
168 addExtParameterInfosForCall(paramInfos, FPT.getTypePtr(), PrefixSize,
172 /// Arrange the LLVM function layout for a value of the given function
173 /// type, on top of any implicit parameters already stored.
174 static const CGFunctionInfo &
175 arrangeLLVMFunctionInfo(CodeGenTypes &CGT, bool instanceMethod,
176 SmallVectorImpl<CanQualType> &prefix,
177 CanQual<FunctionProtoType> FTP) {
178 SmallVector<FunctionProtoType::ExtParameterInfo, 16> paramInfos;
179 RequiredArgs Required = RequiredArgs::forPrototypePlus(FTP, prefix.size());
181 appendParameterTypes(CGT, prefix, paramInfos, FTP);
182 CanQualType resultType = FTP->getReturnType().getUnqualifiedType();
184 return CGT.arrangeLLVMFunctionInfo(resultType, instanceMethod,
185 /*chainCall=*/false, prefix,
186 FTP->getExtInfo(), paramInfos,
190 /// Arrange the argument and result information for a value of the
191 /// given freestanding function type.
192 const CGFunctionInfo &
193 CodeGenTypes::arrangeFreeFunctionType(CanQual<FunctionProtoType> FTP) {
194 SmallVector<CanQualType, 16> argTypes;
195 return ::arrangeLLVMFunctionInfo(*this, /*instanceMethod=*/false, argTypes,
199 static CallingConv getCallingConventionForDecl(const Decl *D, bool IsWindows) {
200 // Set the appropriate calling convention for the Function.
201 if (D->hasAttr<StdCallAttr>())
202 return CC_X86StdCall;
204 if (D->hasAttr<FastCallAttr>())
205 return CC_X86FastCall;
207 if (D->hasAttr<RegCallAttr>())
208 return CC_X86RegCall;
210 if (D->hasAttr<ThisCallAttr>())
211 return CC_X86ThisCall;
213 if (D->hasAttr<VectorCallAttr>())
214 return CC_X86VectorCall;
216 if (D->hasAttr<PascalAttr>())
219 if (PcsAttr *PCS = D->getAttr<PcsAttr>())
220 return (PCS->getPCS() == PcsAttr::AAPCS ? CC_AAPCS : CC_AAPCS_VFP);
222 if (D->hasAttr<AArch64VectorPcsAttr>())
223 return CC_AArch64VectorCall;
225 if (D->hasAttr<IntelOclBiccAttr>())
226 return CC_IntelOclBicc;
228 if (D->hasAttr<MSABIAttr>())
229 return IsWindows ? CC_C : CC_Win64;
231 if (D->hasAttr<SysVABIAttr>())
232 return IsWindows ? CC_X86_64SysV : CC_C;
234 if (D->hasAttr<PreserveMostAttr>())
235 return CC_PreserveMost;
237 if (D->hasAttr<PreserveAllAttr>())
238 return CC_PreserveAll;
243 /// Arrange the argument and result information for a call to an
244 /// unknown C++ non-static member function of the given abstract type.
245 /// (A null RD means we don't have any meaningful "this" argument type,
246 /// so fall back to a generic pointer type).
247 /// The member function must be an ordinary function, i.e. not a
248 /// constructor or destructor.
249 const CGFunctionInfo &
250 CodeGenTypes::arrangeCXXMethodType(const CXXRecordDecl *RD,
251 const FunctionProtoType *FTP,
252 const CXXMethodDecl *MD) {
253 SmallVector<CanQualType, 16> argTypes;
255 // Add the 'this' pointer.
256 argTypes.push_back(DeriveThisType(RD, MD));
258 return ::arrangeLLVMFunctionInfo(
259 *this, true, argTypes,
260 FTP->getCanonicalTypeUnqualified().getAs<FunctionProtoType>());
263 /// Set calling convention for CUDA/HIP kernel.
264 static void setCUDAKernelCallingConvention(CanQualType &FTy, CodeGenModule &CGM,
265 const FunctionDecl *FD) {
266 if (FD->hasAttr<CUDAGlobalAttr>()) {
267 const FunctionType *FT = FTy->getAs<FunctionType>();
268 CGM.getTargetCodeGenInfo().setCUDAKernelCallingConvention(FT);
269 FTy = FT->getCanonicalTypeUnqualified();
273 /// Arrange the argument and result information for a declaration or
274 /// definition of the given C++ non-static member function. The
275 /// member function must be an ordinary function, i.e. not a
276 /// constructor or destructor.
277 const CGFunctionInfo &
278 CodeGenTypes::arrangeCXXMethodDeclaration(const CXXMethodDecl *MD) {
279 assert(!isa<CXXConstructorDecl>(MD) && "wrong method for constructors!");
280 assert(!isa<CXXDestructorDecl>(MD) && "wrong method for destructors!");
282 CanQualType FT = GetFormalType(MD).getAs<Type>();
283 setCUDAKernelCallingConvention(FT, CGM, MD);
284 auto prototype = FT.getAs<FunctionProtoType>();
286 if (MD->isInstance()) {
287 // The abstract case is perfectly fine.
288 const CXXRecordDecl *ThisType = TheCXXABI.getThisArgumentTypeForMethod(MD);
289 return arrangeCXXMethodType(ThisType, prototype.getTypePtr(), MD);
292 return arrangeFreeFunctionType(prototype);
295 bool CodeGenTypes::inheritingCtorHasParams(
296 const InheritedConstructor &Inherited, CXXCtorType Type) {
297 // Parameters are unnecessary if we're constructing a base class subobject
298 // and the inherited constructor lives in a virtual base.
299 return Type == Ctor_Complete ||
300 !Inherited.getShadowDecl()->constructsVirtualBase() ||
301 !Target.getCXXABI().hasConstructorVariants();
304 const CGFunctionInfo &
305 CodeGenTypes::arrangeCXXStructorDeclaration(GlobalDecl GD) {
306 auto *MD = cast<CXXMethodDecl>(GD.getDecl());
308 SmallVector<CanQualType, 16> argTypes;
309 SmallVector<FunctionProtoType::ExtParameterInfo, 16> paramInfos;
310 argTypes.push_back(DeriveThisType(MD->getParent(), MD));
312 bool PassParams = true;
314 if (auto *CD = dyn_cast<CXXConstructorDecl>(MD)) {
315 // A base class inheriting constructor doesn't get forwarded arguments
316 // needed to construct a virtual base (or base class thereof).
317 if (auto Inherited = CD->getInheritedConstructor())
318 PassParams = inheritingCtorHasParams(Inherited, GD.getCtorType());
321 CanQual<FunctionProtoType> FTP = GetFormalType(MD);
323 // Add the formal parameters.
325 appendParameterTypes(*this, argTypes, paramInfos, FTP);
327 CGCXXABI::AddedStructorArgs AddedArgs =
328 TheCXXABI.buildStructorSignature(GD, argTypes);
329 if (!paramInfos.empty()) {
330 // Note: prefix implies after the first param.
331 if (AddedArgs.Prefix)
332 paramInfos.insert(paramInfos.begin() + 1, AddedArgs.Prefix,
333 FunctionProtoType::ExtParameterInfo{});
334 if (AddedArgs.Suffix)
335 paramInfos.append(AddedArgs.Suffix,
336 FunctionProtoType::ExtParameterInfo{});
339 RequiredArgs required =
340 (PassParams && MD->isVariadic() ? RequiredArgs(argTypes.size())
341 : RequiredArgs::All);
343 FunctionType::ExtInfo extInfo = FTP->getExtInfo();
344 CanQualType resultType = TheCXXABI.HasThisReturn(GD)
346 : TheCXXABI.hasMostDerivedReturn(GD)
347 ? CGM.getContext().VoidPtrTy
349 return arrangeLLVMFunctionInfo(resultType, /*instanceMethod=*/true,
350 /*chainCall=*/false, argTypes, extInfo,
351 paramInfos, required);
354 static SmallVector<CanQualType, 16>
355 getArgTypesForCall(ASTContext &ctx, const CallArgList &args) {
356 SmallVector<CanQualType, 16> argTypes;
357 for (auto &arg : args)
358 argTypes.push_back(ctx.getCanonicalParamType(arg.Ty));
362 static SmallVector<CanQualType, 16>
363 getArgTypesForDeclaration(ASTContext &ctx, const FunctionArgList &args) {
364 SmallVector<CanQualType, 16> argTypes;
365 for (auto &arg : args)
366 argTypes.push_back(ctx.getCanonicalParamType(arg->getType()));
370 static llvm::SmallVector<FunctionProtoType::ExtParameterInfo, 16>
371 getExtParameterInfosForCall(const FunctionProtoType *proto,
372 unsigned prefixArgs, unsigned totalArgs) {
373 llvm::SmallVector<FunctionProtoType::ExtParameterInfo, 16> result;
374 if (proto->hasExtParameterInfos()) {
375 addExtParameterInfosForCall(result, proto, prefixArgs, totalArgs);
380 /// Arrange a call to a C++ method, passing the given arguments.
382 /// ExtraPrefixArgs is the number of ABI-specific args passed after the `this`
384 /// ExtraSuffixArgs is the number of ABI-specific args passed at the end of
386 /// PassProtoArgs indicates whether `args` has args for the parameters in the
387 /// given CXXConstructorDecl.
388 const CGFunctionInfo &
389 CodeGenTypes::arrangeCXXConstructorCall(const CallArgList &args,
390 const CXXConstructorDecl *D,
391 CXXCtorType CtorKind,
392 unsigned ExtraPrefixArgs,
393 unsigned ExtraSuffixArgs,
394 bool PassProtoArgs) {
396 SmallVector<CanQualType, 16> ArgTypes;
397 for (const auto &Arg : args)
398 ArgTypes.push_back(Context.getCanonicalParamType(Arg.Ty));
400 // +1 for implicit this, which should always be args[0].
401 unsigned TotalPrefixArgs = 1 + ExtraPrefixArgs;
403 CanQual<FunctionProtoType> FPT = GetFormalType(D);
404 RequiredArgs Required = PassProtoArgs
405 ? RequiredArgs::forPrototypePlus(
406 FPT, TotalPrefixArgs + ExtraSuffixArgs)
409 GlobalDecl GD(D, CtorKind);
410 CanQualType ResultType = TheCXXABI.HasThisReturn(GD)
412 : TheCXXABI.hasMostDerivedReturn(GD)
413 ? CGM.getContext().VoidPtrTy
416 FunctionType::ExtInfo Info = FPT->getExtInfo();
417 llvm::SmallVector<FunctionProtoType::ExtParameterInfo, 16> ParamInfos;
418 // If the prototype args are elided, we should only have ABI-specific args,
419 // which never have param info.
420 if (PassProtoArgs && FPT->hasExtParameterInfos()) {
421 // ABI-specific suffix arguments are treated the same as variadic arguments.
422 addExtParameterInfosForCall(ParamInfos, FPT.getTypePtr(), TotalPrefixArgs,
425 return arrangeLLVMFunctionInfo(ResultType, /*instanceMethod=*/true,
426 /*chainCall=*/false, ArgTypes, Info,
427 ParamInfos, Required);
430 /// Arrange the argument and result information for the declaration or
431 /// definition of the given function.
432 const CGFunctionInfo &
433 CodeGenTypes::arrangeFunctionDeclaration(const FunctionDecl *FD) {
434 if (const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(FD))
435 if (MD->isInstance())
436 return arrangeCXXMethodDeclaration(MD);
438 CanQualType FTy = FD->getType()->getCanonicalTypeUnqualified();
440 assert(isa<FunctionType>(FTy));
441 setCUDAKernelCallingConvention(FTy, CGM, FD);
443 // When declaring a function without a prototype, always use a
444 // non-variadic type.
445 if (CanQual<FunctionNoProtoType> noProto = FTy.getAs<FunctionNoProtoType>()) {
446 return arrangeLLVMFunctionInfo(
447 noProto->getReturnType(), /*instanceMethod=*/false,
448 /*chainCall=*/false, None, noProto->getExtInfo(), {},RequiredArgs::All);
451 return arrangeFreeFunctionType(FTy.castAs<FunctionProtoType>());
454 /// Arrange the argument and result information for the declaration or
455 /// definition of an Objective-C method.
456 const CGFunctionInfo &
457 CodeGenTypes::arrangeObjCMethodDeclaration(const ObjCMethodDecl *MD) {
458 // It happens that this is the same as a call with no optional
459 // arguments, except also using the formal 'self' type.
460 return arrangeObjCMessageSendSignature(MD, MD->getSelfDecl()->getType());
463 /// Arrange the argument and result information for the function type
464 /// through which to perform a send to the given Objective-C method,
465 /// using the given receiver type. The receiver type is not always
466 /// the 'self' type of the method or even an Objective-C pointer type.
467 /// This is *not* the right method for actually performing such a
468 /// message send, due to the possibility of optional arguments.
469 const CGFunctionInfo &
470 CodeGenTypes::arrangeObjCMessageSendSignature(const ObjCMethodDecl *MD,
471 QualType receiverType) {
472 SmallVector<CanQualType, 16> argTys;
473 SmallVector<FunctionProtoType::ExtParameterInfo, 4> extParamInfos(2);
474 argTys.push_back(Context.getCanonicalParamType(receiverType));
475 argTys.push_back(Context.getCanonicalParamType(Context.getObjCSelType()));
477 for (const auto *I : MD->parameters()) {
478 argTys.push_back(Context.getCanonicalParamType(I->getType()));
479 auto extParamInfo = FunctionProtoType::ExtParameterInfo().withIsNoEscape(
480 I->hasAttr<NoEscapeAttr>());
481 extParamInfos.push_back(extParamInfo);
484 FunctionType::ExtInfo einfo;
485 bool IsWindows = getContext().getTargetInfo().getTriple().isOSWindows();
486 einfo = einfo.withCallingConv(getCallingConventionForDecl(MD, IsWindows));
488 if (getContext().getLangOpts().ObjCAutoRefCount &&
489 MD->hasAttr<NSReturnsRetainedAttr>())
490 einfo = einfo.withProducesResult(true);
492 RequiredArgs required =
493 (MD->isVariadic() ? RequiredArgs(argTys.size()) : RequiredArgs::All);
495 return arrangeLLVMFunctionInfo(
496 GetReturnType(MD->getReturnType()), /*instanceMethod=*/false,
497 /*chainCall=*/false, argTys, einfo, extParamInfos, required);
500 const CGFunctionInfo &
501 CodeGenTypes::arrangeUnprototypedObjCMessageSend(QualType returnType,
502 const CallArgList &args) {
503 auto argTypes = getArgTypesForCall(Context, args);
504 FunctionType::ExtInfo einfo;
506 return arrangeLLVMFunctionInfo(
507 GetReturnType(returnType), /*instanceMethod=*/false,
508 /*chainCall=*/false, argTypes, einfo, {}, RequiredArgs::All);
511 const CGFunctionInfo &
512 CodeGenTypes::arrangeGlobalDeclaration(GlobalDecl GD) {
513 // FIXME: Do we need to handle ObjCMethodDecl?
514 const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl());
516 if (isa<CXXConstructorDecl>(GD.getDecl()) ||
517 isa<CXXDestructorDecl>(GD.getDecl()))
518 return arrangeCXXStructorDeclaration(GD);
520 return arrangeFunctionDeclaration(FD);
523 /// Arrange a thunk that takes 'this' as the first parameter followed by
524 /// varargs. Return a void pointer, regardless of the actual return type.
525 /// The body of the thunk will end in a musttail call to a function of the
526 /// correct type, and the caller will bitcast the function to the correct
528 const CGFunctionInfo &
529 CodeGenTypes::arrangeUnprototypedMustTailThunk(const CXXMethodDecl *MD) {
530 assert(MD->isVirtual() && "only methods have thunks");
531 CanQual<FunctionProtoType> FTP = GetFormalType(MD);
532 CanQualType ArgTys[] = {DeriveThisType(MD->getParent(), MD)};
533 return arrangeLLVMFunctionInfo(Context.VoidTy, /*instanceMethod=*/false,
534 /*chainCall=*/false, ArgTys,
535 FTP->getExtInfo(), {}, RequiredArgs(1));
538 const CGFunctionInfo &
539 CodeGenTypes::arrangeMSCtorClosure(const CXXConstructorDecl *CD,
541 assert(CT == Ctor_CopyingClosure || CT == Ctor_DefaultClosure);
543 CanQual<FunctionProtoType> FTP = GetFormalType(CD);
544 SmallVector<CanQualType, 2> ArgTys;
545 const CXXRecordDecl *RD = CD->getParent();
546 ArgTys.push_back(DeriveThisType(RD, CD));
547 if (CT == Ctor_CopyingClosure)
548 ArgTys.push_back(*FTP->param_type_begin());
549 if (RD->getNumVBases() > 0)
550 ArgTys.push_back(Context.IntTy);
551 CallingConv CC = Context.getDefaultCallingConvention(
552 /*IsVariadic=*/false, /*IsCXXMethod=*/true);
553 return arrangeLLVMFunctionInfo(Context.VoidTy, /*instanceMethod=*/true,
554 /*chainCall=*/false, ArgTys,
555 FunctionType::ExtInfo(CC), {},
559 /// Arrange a call as unto a free function, except possibly with an
560 /// additional number of formal parameters considered required.
561 static const CGFunctionInfo &
562 arrangeFreeFunctionLikeCall(CodeGenTypes &CGT,
564 const CallArgList &args,
565 const FunctionType *fnType,
566 unsigned numExtraRequiredArgs,
568 assert(args.size() >= numExtraRequiredArgs);
570 llvm::SmallVector<FunctionProtoType::ExtParameterInfo, 16> paramInfos;
572 // In most cases, there are no optional arguments.
573 RequiredArgs required = RequiredArgs::All;
575 // If we have a variadic prototype, the required arguments are the
576 // extra prefix plus the arguments in the prototype.
577 if (const FunctionProtoType *proto = dyn_cast<FunctionProtoType>(fnType)) {
578 if (proto->isVariadic())
579 required = RequiredArgs::forPrototypePlus(proto, numExtraRequiredArgs);
581 if (proto->hasExtParameterInfos())
582 addExtParameterInfosForCall(paramInfos, proto, numExtraRequiredArgs,
585 // If we don't have a prototype at all, but we're supposed to
586 // explicitly use the variadic convention for unprototyped calls,
587 // treat all of the arguments as required but preserve the nominal
588 // possibility of variadics.
589 } else if (CGM.getTargetCodeGenInfo()
590 .isNoProtoCallVariadic(args,
591 cast<FunctionNoProtoType>(fnType))) {
592 required = RequiredArgs(args.size());
596 SmallVector<CanQualType, 16> argTypes;
597 for (const auto &arg : args)
598 argTypes.push_back(CGT.getContext().getCanonicalParamType(arg.Ty));
599 return CGT.arrangeLLVMFunctionInfo(GetReturnType(fnType->getReturnType()),
600 /*instanceMethod=*/false, chainCall,
601 argTypes, fnType->getExtInfo(), paramInfos,
605 /// Figure out the rules for calling a function with the given formal
606 /// type using the given arguments. The arguments are necessary
607 /// because the function might be unprototyped, in which case it's
608 /// target-dependent in crazy ways.
609 const CGFunctionInfo &
610 CodeGenTypes::arrangeFreeFunctionCall(const CallArgList &args,
611 const FunctionType *fnType,
613 return arrangeFreeFunctionLikeCall(*this, CGM, args, fnType,
614 chainCall ? 1 : 0, chainCall);
617 /// A block function is essentially a free function with an
618 /// extra implicit argument.
619 const CGFunctionInfo &
620 CodeGenTypes::arrangeBlockFunctionCall(const CallArgList &args,
621 const FunctionType *fnType) {
622 return arrangeFreeFunctionLikeCall(*this, CGM, args, fnType, 1,
623 /*chainCall=*/false);
626 const CGFunctionInfo &
627 CodeGenTypes::arrangeBlockFunctionDeclaration(const FunctionProtoType *proto,
628 const FunctionArgList ¶ms) {
629 auto paramInfos = getExtParameterInfosForCall(proto, 1, params.size());
630 auto argTypes = getArgTypesForDeclaration(Context, params);
632 return arrangeLLVMFunctionInfo(GetReturnType(proto->getReturnType()),
633 /*instanceMethod*/ false, /*chainCall*/ false,
634 argTypes, proto->getExtInfo(), paramInfos,
635 RequiredArgs::forPrototypePlus(proto, 1));
638 const CGFunctionInfo &
639 CodeGenTypes::arrangeBuiltinFunctionCall(QualType resultType,
640 const CallArgList &args) {
642 SmallVector<CanQualType, 16> argTypes;
643 for (const auto &Arg : args)
644 argTypes.push_back(Context.getCanonicalParamType(Arg.Ty));
645 return arrangeLLVMFunctionInfo(
646 GetReturnType(resultType), /*instanceMethod=*/false,
647 /*chainCall=*/false, argTypes, FunctionType::ExtInfo(),
648 /*paramInfos=*/ {}, RequiredArgs::All);
651 const CGFunctionInfo &
652 CodeGenTypes::arrangeBuiltinFunctionDeclaration(QualType resultType,
653 const FunctionArgList &args) {
654 auto argTypes = getArgTypesForDeclaration(Context, args);
656 return arrangeLLVMFunctionInfo(
657 GetReturnType(resultType), /*instanceMethod=*/false, /*chainCall=*/false,
658 argTypes, FunctionType::ExtInfo(), {}, RequiredArgs::All);
661 const CGFunctionInfo &
662 CodeGenTypes::arrangeBuiltinFunctionDeclaration(CanQualType resultType,
663 ArrayRef<CanQualType> argTypes) {
664 return arrangeLLVMFunctionInfo(
665 resultType, /*instanceMethod=*/false, /*chainCall=*/false,
666 argTypes, FunctionType::ExtInfo(), {}, RequiredArgs::All);
669 /// Arrange a call to a C++ method, passing the given arguments.
671 /// numPrefixArgs is the number of ABI-specific prefix arguments we have. It
672 /// does not count `this`.
673 const CGFunctionInfo &
674 CodeGenTypes::arrangeCXXMethodCall(const CallArgList &args,
675 const FunctionProtoType *proto,
676 RequiredArgs required,
677 unsigned numPrefixArgs) {
678 assert(numPrefixArgs + 1 <= args.size() &&
679 "Emitting a call with less args than the required prefix?");
680 // Add one to account for `this`. It's a bit awkward here, but we don't count
681 // `this` in similar places elsewhere.
683 getExtParameterInfosForCall(proto, numPrefixArgs + 1, args.size());
686 auto argTypes = getArgTypesForCall(Context, args);
688 FunctionType::ExtInfo info = proto->getExtInfo();
689 return arrangeLLVMFunctionInfo(
690 GetReturnType(proto->getReturnType()), /*instanceMethod=*/true,
691 /*chainCall=*/false, argTypes, info, paramInfos, required);
694 const CGFunctionInfo &CodeGenTypes::arrangeNullaryFunction() {
695 return arrangeLLVMFunctionInfo(
696 getContext().VoidTy, /*instanceMethod=*/false, /*chainCall=*/false,
697 None, FunctionType::ExtInfo(), {}, RequiredArgs::All);
700 const CGFunctionInfo &
701 CodeGenTypes::arrangeCall(const CGFunctionInfo &signature,
702 const CallArgList &args) {
703 assert(signature.arg_size() <= args.size());
704 if (signature.arg_size() == args.size())
707 SmallVector<FunctionProtoType::ExtParameterInfo, 16> paramInfos;
708 auto sigParamInfos = signature.getExtParameterInfos();
709 if (!sigParamInfos.empty()) {
710 paramInfos.append(sigParamInfos.begin(), sigParamInfos.end());
711 paramInfos.resize(args.size());
714 auto argTypes = getArgTypesForCall(Context, args);
716 assert(signature.getRequiredArgs().allowsOptionalArgs());
717 return arrangeLLVMFunctionInfo(signature.getReturnType(),
718 signature.isInstanceMethod(),
719 signature.isChainCall(),
721 signature.getExtInfo(),
723 signature.getRequiredArgs());
728 void computeSPIRKernelABIInfo(CodeGenModule &CGM, CGFunctionInfo &FI);
732 /// Arrange the argument and result information for an abstract value
733 /// of a given function type. This is the method which all of the
734 /// above functions ultimately defer to.
735 const CGFunctionInfo &
736 CodeGenTypes::arrangeLLVMFunctionInfo(CanQualType resultType,
739 ArrayRef<CanQualType> argTypes,
740 FunctionType::ExtInfo info,
741 ArrayRef<FunctionProtoType::ExtParameterInfo> paramInfos,
742 RequiredArgs required) {
743 assert(llvm::all_of(argTypes,
744 [](CanQualType T) { return T.isCanonicalAsParam(); }));
746 // Lookup or create unique function info.
747 llvm::FoldingSetNodeID ID;
748 CGFunctionInfo::Profile(ID, instanceMethod, chainCall, info, paramInfos,
749 required, resultType, argTypes);
751 void *insertPos = nullptr;
752 CGFunctionInfo *FI = FunctionInfos.FindNodeOrInsertPos(ID, insertPos);
756 unsigned CC = ClangCallConvToLLVMCallConv(info.getCC());
758 // Construct the function info. We co-allocate the ArgInfos.
759 FI = CGFunctionInfo::create(CC, instanceMethod, chainCall, info,
760 paramInfos, resultType, argTypes, required);
761 FunctionInfos.InsertNode(FI, insertPos);
763 bool inserted = FunctionsBeingProcessed.insert(FI).second;
765 assert(inserted && "Recursively being processed?");
767 // Compute ABI information.
768 if (CC == llvm::CallingConv::SPIR_KERNEL) {
769 // Force target independent argument handling for the host visible
771 computeSPIRKernelABIInfo(CGM, *FI);
772 } else if (info.getCC() == CC_Swift) {
773 swiftcall::computeABIInfo(CGM, *FI);
775 getABIInfo().computeInfo(*FI);
778 // Loop over all of the computed argument and return value info. If any of
779 // them are direct or extend without a specified coerce type, specify the
781 ABIArgInfo &retInfo = FI->getReturnInfo();
782 if (retInfo.canHaveCoerceToType() && retInfo.getCoerceToType() == nullptr)
783 retInfo.setCoerceToType(ConvertType(FI->getReturnType()));
785 for (auto &I : FI->arguments())
786 if (I.info.canHaveCoerceToType() && I.info.getCoerceToType() == nullptr)
787 I.info.setCoerceToType(ConvertType(I.type));
789 bool erased = FunctionsBeingProcessed.erase(FI); (void)erased;
790 assert(erased && "Not in set?");
795 CGFunctionInfo *CGFunctionInfo::create(unsigned llvmCC,
798 const FunctionType::ExtInfo &info,
799 ArrayRef<ExtParameterInfo> paramInfos,
800 CanQualType resultType,
801 ArrayRef<CanQualType> argTypes,
802 RequiredArgs required) {
803 assert(paramInfos.empty() || paramInfos.size() == argTypes.size());
804 assert(!required.allowsOptionalArgs() ||
805 required.getNumRequiredArgs() <= argTypes.size());
808 operator new(totalSizeToAlloc<ArgInfo, ExtParameterInfo>(
809 argTypes.size() + 1, paramInfos.size()));
811 CGFunctionInfo *FI = new(buffer) CGFunctionInfo();
812 FI->CallingConvention = llvmCC;
813 FI->EffectiveCallingConvention = llvmCC;
814 FI->ASTCallingConvention = info.getCC();
815 FI->InstanceMethod = instanceMethod;
816 FI->ChainCall = chainCall;
817 FI->NoReturn = info.getNoReturn();
818 FI->ReturnsRetained = info.getProducesResult();
819 FI->NoCallerSavedRegs = info.getNoCallerSavedRegs();
820 FI->NoCfCheck = info.getNoCfCheck();
821 FI->Required = required;
822 FI->HasRegParm = info.getHasRegParm();
823 FI->RegParm = info.getRegParm();
824 FI->ArgStruct = nullptr;
825 FI->ArgStructAlign = 0;
826 FI->NumArgs = argTypes.size();
827 FI->HasExtParameterInfos = !paramInfos.empty();
828 FI->getArgsBuffer()[0].type = resultType;
829 for (unsigned i = 0, e = argTypes.size(); i != e; ++i)
830 FI->getArgsBuffer()[i + 1].type = argTypes[i];
831 for (unsigned i = 0, e = paramInfos.size(); i != e; ++i)
832 FI->getExtParameterInfosBuffer()[i] = paramInfos[i];
839 // ABIArgInfo::Expand implementation.
841 // Specifies the way QualType passed as ABIArgInfo::Expand is expanded.
842 struct TypeExpansion {
843 enum TypeExpansionKind {
844 // Elements of constant arrays are expanded recursively.
846 // Record fields are expanded recursively (but if record is a union, only
847 // the field with the largest size is expanded).
849 // For complex types, real and imaginary parts are expanded recursively.
851 // All other types are not expandable.
855 const TypeExpansionKind Kind;
857 TypeExpansion(TypeExpansionKind K) : Kind(K) {}
858 virtual ~TypeExpansion() {}
861 struct ConstantArrayExpansion : TypeExpansion {
865 ConstantArrayExpansion(QualType EltTy, uint64_t NumElts)
866 : TypeExpansion(TEK_ConstantArray), EltTy(EltTy), NumElts(NumElts) {}
867 static bool classof(const TypeExpansion *TE) {
868 return TE->Kind == TEK_ConstantArray;
872 struct RecordExpansion : TypeExpansion {
873 SmallVector<const CXXBaseSpecifier *, 1> Bases;
875 SmallVector<const FieldDecl *, 1> Fields;
877 RecordExpansion(SmallVector<const CXXBaseSpecifier *, 1> &&Bases,
878 SmallVector<const FieldDecl *, 1> &&Fields)
879 : TypeExpansion(TEK_Record), Bases(std::move(Bases)),
880 Fields(std::move(Fields)) {}
881 static bool classof(const TypeExpansion *TE) {
882 return TE->Kind == TEK_Record;
886 struct ComplexExpansion : TypeExpansion {
889 ComplexExpansion(QualType EltTy) : TypeExpansion(TEK_Complex), EltTy(EltTy) {}
890 static bool classof(const TypeExpansion *TE) {
891 return TE->Kind == TEK_Complex;
895 struct NoExpansion : TypeExpansion {
896 NoExpansion() : TypeExpansion(TEK_None) {}
897 static bool classof(const TypeExpansion *TE) {
898 return TE->Kind == TEK_None;
903 static std::unique_ptr<TypeExpansion>
904 getTypeExpansion(QualType Ty, const ASTContext &Context) {
905 if (const ConstantArrayType *AT = Context.getAsConstantArrayType(Ty)) {
906 return llvm::make_unique<ConstantArrayExpansion>(
907 AT->getElementType(), AT->getSize().getZExtValue());
909 if (const RecordType *RT = Ty->getAs<RecordType>()) {
910 SmallVector<const CXXBaseSpecifier *, 1> Bases;
911 SmallVector<const FieldDecl *, 1> Fields;
912 const RecordDecl *RD = RT->getDecl();
913 assert(!RD->hasFlexibleArrayMember() &&
914 "Cannot expand structure with flexible array.");
916 // Unions can be here only in degenerative cases - all the fields are same
917 // after flattening. Thus we have to use the "largest" field.
918 const FieldDecl *LargestFD = nullptr;
919 CharUnits UnionSize = CharUnits::Zero();
921 for (const auto *FD : RD->fields()) {
922 if (FD->isZeroLengthBitField(Context))
924 assert(!FD->isBitField() &&
925 "Cannot expand structure with bit-field members.");
926 CharUnits FieldSize = Context.getTypeSizeInChars(FD->getType());
927 if (UnionSize < FieldSize) {
928 UnionSize = FieldSize;
933 Fields.push_back(LargestFD);
935 if (const auto *CXXRD = dyn_cast<CXXRecordDecl>(RD)) {
936 assert(!CXXRD->isDynamicClass() &&
937 "cannot expand vtable pointers in dynamic classes");
938 for (const CXXBaseSpecifier &BS : CXXRD->bases())
939 Bases.push_back(&BS);
942 for (const auto *FD : RD->fields()) {
943 if (FD->isZeroLengthBitField(Context))
945 assert(!FD->isBitField() &&
946 "Cannot expand structure with bit-field members.");
947 Fields.push_back(FD);
950 return llvm::make_unique<RecordExpansion>(std::move(Bases),
953 if (const ComplexType *CT = Ty->getAs<ComplexType>()) {
954 return llvm::make_unique<ComplexExpansion>(CT->getElementType());
956 return llvm::make_unique<NoExpansion>();
959 static int getExpansionSize(QualType Ty, const ASTContext &Context) {
960 auto Exp = getTypeExpansion(Ty, Context);
961 if (auto CAExp = dyn_cast<ConstantArrayExpansion>(Exp.get())) {
962 return CAExp->NumElts * getExpansionSize(CAExp->EltTy, Context);
964 if (auto RExp = dyn_cast<RecordExpansion>(Exp.get())) {
966 for (auto BS : RExp->Bases)
967 Res += getExpansionSize(BS->getType(), Context);
968 for (auto FD : RExp->Fields)
969 Res += getExpansionSize(FD->getType(), Context);
972 if (isa<ComplexExpansion>(Exp.get()))
974 assert(isa<NoExpansion>(Exp.get()));
979 CodeGenTypes::getExpandedTypes(QualType Ty,
980 SmallVectorImpl<llvm::Type *>::iterator &TI) {
981 auto Exp = getTypeExpansion(Ty, Context);
982 if (auto CAExp = dyn_cast<ConstantArrayExpansion>(Exp.get())) {
983 for (int i = 0, n = CAExp->NumElts; i < n; i++) {
984 getExpandedTypes(CAExp->EltTy, TI);
986 } else if (auto RExp = dyn_cast<RecordExpansion>(Exp.get())) {
987 for (auto BS : RExp->Bases)
988 getExpandedTypes(BS->getType(), TI);
989 for (auto FD : RExp->Fields)
990 getExpandedTypes(FD->getType(), TI);
991 } else if (auto CExp = dyn_cast<ComplexExpansion>(Exp.get())) {
992 llvm::Type *EltTy = ConvertType(CExp->EltTy);
996 assert(isa<NoExpansion>(Exp.get()));
997 *TI++ = ConvertType(Ty);
1001 static void forConstantArrayExpansion(CodeGenFunction &CGF,
1002 ConstantArrayExpansion *CAE,
1004 llvm::function_ref<void(Address)> Fn) {
1005 CharUnits EltSize = CGF.getContext().getTypeSizeInChars(CAE->EltTy);
1006 CharUnits EltAlign =
1007 BaseAddr.getAlignment().alignmentOfArrayElement(EltSize);
1009 for (int i = 0, n = CAE->NumElts; i < n; i++) {
1010 llvm::Value *EltAddr =
1011 CGF.Builder.CreateConstGEP2_32(nullptr, BaseAddr.getPointer(), 0, i);
1012 Fn(Address(EltAddr, EltAlign));
1016 void CodeGenFunction::ExpandTypeFromArgs(
1017 QualType Ty, LValue LV, SmallVectorImpl<llvm::Value *>::iterator &AI) {
1018 assert(LV.isSimple() &&
1019 "Unexpected non-simple lvalue during struct expansion.");
1021 auto Exp = getTypeExpansion(Ty, getContext());
1022 if (auto CAExp = dyn_cast<ConstantArrayExpansion>(Exp.get())) {
1023 forConstantArrayExpansion(*this, CAExp, LV.getAddress(),
1024 [&](Address EltAddr) {
1025 LValue LV = MakeAddrLValue(EltAddr, CAExp->EltTy);
1026 ExpandTypeFromArgs(CAExp->EltTy, LV, AI);
1028 } else if (auto RExp = dyn_cast<RecordExpansion>(Exp.get())) {
1029 Address This = LV.getAddress();
1030 for (const CXXBaseSpecifier *BS : RExp->Bases) {
1031 // Perform a single step derived-to-base conversion.
1033 GetAddressOfBaseClass(This, Ty->getAsCXXRecordDecl(), &BS, &BS + 1,
1034 /*NullCheckValue=*/false, SourceLocation());
1035 LValue SubLV = MakeAddrLValue(Base, BS->getType());
1037 // Recurse onto bases.
1038 ExpandTypeFromArgs(BS->getType(), SubLV, AI);
1040 for (auto FD : RExp->Fields) {
1041 // FIXME: What are the right qualifiers here?
1042 LValue SubLV = EmitLValueForFieldInitialization(LV, FD);
1043 ExpandTypeFromArgs(FD->getType(), SubLV, AI);
1045 } else if (isa<ComplexExpansion>(Exp.get())) {
1046 auto realValue = *AI++;
1047 auto imagValue = *AI++;
1048 EmitStoreOfComplex(ComplexPairTy(realValue, imagValue), LV, /*init*/ true);
1050 assert(isa<NoExpansion>(Exp.get()));
1051 EmitStoreThroughLValue(RValue::get(*AI++), LV);
1055 void CodeGenFunction::ExpandTypeToArgs(
1056 QualType Ty, CallArg Arg, llvm::FunctionType *IRFuncTy,
1057 SmallVectorImpl<llvm::Value *> &IRCallArgs, unsigned &IRCallArgPos) {
1058 auto Exp = getTypeExpansion(Ty, getContext());
1059 if (auto CAExp = dyn_cast<ConstantArrayExpansion>(Exp.get())) {
1060 Address Addr = Arg.hasLValue() ? Arg.getKnownLValue().getAddress()
1061 : Arg.getKnownRValue().getAggregateAddress();
1062 forConstantArrayExpansion(
1063 *this, CAExp, Addr, [&](Address EltAddr) {
1064 CallArg EltArg = CallArg(
1065 convertTempToRValue(EltAddr, CAExp->EltTy, SourceLocation()),
1067 ExpandTypeToArgs(CAExp->EltTy, EltArg, IRFuncTy, IRCallArgs,
1070 } else if (auto RExp = dyn_cast<RecordExpansion>(Exp.get())) {
1071 Address This = Arg.hasLValue() ? Arg.getKnownLValue().getAddress()
1072 : Arg.getKnownRValue().getAggregateAddress();
1073 for (const CXXBaseSpecifier *BS : RExp->Bases) {
1074 // Perform a single step derived-to-base conversion.
1076 GetAddressOfBaseClass(This, Ty->getAsCXXRecordDecl(), &BS, &BS + 1,
1077 /*NullCheckValue=*/false, SourceLocation());
1078 CallArg BaseArg = CallArg(RValue::getAggregate(Base), BS->getType());
1080 // Recurse onto bases.
1081 ExpandTypeToArgs(BS->getType(), BaseArg, IRFuncTy, IRCallArgs,
1085 LValue LV = MakeAddrLValue(This, Ty);
1086 for (auto FD : RExp->Fields) {
1088 CallArg(EmitRValueForField(LV, FD, SourceLocation()), FD->getType());
1089 ExpandTypeToArgs(FD->getType(), FldArg, IRFuncTy, IRCallArgs,
1092 } else if (isa<ComplexExpansion>(Exp.get())) {
1093 ComplexPairTy CV = Arg.getKnownRValue().getComplexVal();
1094 IRCallArgs[IRCallArgPos++] = CV.first;
1095 IRCallArgs[IRCallArgPos++] = CV.second;
1097 assert(isa<NoExpansion>(Exp.get()));
1098 auto RV = Arg.getKnownRValue();
1099 assert(RV.isScalar() &&
1100 "Unexpected non-scalar rvalue during struct expansion.");
1102 // Insert a bitcast as needed.
1103 llvm::Value *V = RV.getScalarVal();
1104 if (IRCallArgPos < IRFuncTy->getNumParams() &&
1105 V->getType() != IRFuncTy->getParamType(IRCallArgPos))
1106 V = Builder.CreateBitCast(V, IRFuncTy->getParamType(IRCallArgPos));
1108 IRCallArgs[IRCallArgPos++] = V;
1112 /// Create a temporary allocation for the purposes of coercion.
1113 static Address CreateTempAllocaForCoercion(CodeGenFunction &CGF, llvm::Type *Ty,
1114 CharUnits MinAlign) {
1115 // Don't use an alignment that's worse than what LLVM would prefer.
1116 auto PrefAlign = CGF.CGM.getDataLayout().getPrefTypeAlignment(Ty);
1117 CharUnits Align = std::max(MinAlign, CharUnits::fromQuantity(PrefAlign));
1119 return CGF.CreateTempAlloca(Ty, Align);
1122 /// EnterStructPointerForCoercedAccess - Given a struct pointer that we are
1123 /// accessing some number of bytes out of it, try to gep into the struct to get
1124 /// at its inner goodness. Dive as deep as possible without entering an element
1125 /// with an in-memory size smaller than DstSize.
1127 EnterStructPointerForCoercedAccess(Address SrcPtr,
1128 llvm::StructType *SrcSTy,
1129 uint64_t DstSize, CodeGenFunction &CGF) {
1130 // We can't dive into a zero-element struct.
1131 if (SrcSTy->getNumElements() == 0) return SrcPtr;
1133 llvm::Type *FirstElt = SrcSTy->getElementType(0);
1135 // If the first elt is at least as large as what we're looking for, or if the
1136 // first element is the same size as the whole struct, we can enter it. The
1137 // comparison must be made on the store size and not the alloca size. Using
1138 // the alloca size may overstate the size of the load.
1139 uint64_t FirstEltSize =
1140 CGF.CGM.getDataLayout().getTypeStoreSize(FirstElt);
1141 if (FirstEltSize < DstSize &&
1142 FirstEltSize < CGF.CGM.getDataLayout().getTypeStoreSize(SrcSTy))
1145 // GEP into the first element.
1146 SrcPtr = CGF.Builder.CreateStructGEP(SrcPtr, 0, "coerce.dive");
1148 // If the first element is a struct, recurse.
1149 llvm::Type *SrcTy = SrcPtr.getElementType();
1150 if (llvm::StructType *SrcSTy = dyn_cast<llvm::StructType>(SrcTy))
1151 return EnterStructPointerForCoercedAccess(SrcPtr, SrcSTy, DstSize, CGF);
1156 /// CoerceIntOrPtrToIntOrPtr - Convert a value Val to the specific Ty where both
1157 /// are either integers or pointers. This does a truncation of the value if it
1158 /// is too large or a zero extension if it is too small.
1160 /// This behaves as if the value were coerced through memory, so on big-endian
1161 /// targets the high bits are preserved in a truncation, while little-endian
1162 /// targets preserve the low bits.
1163 static llvm::Value *CoerceIntOrPtrToIntOrPtr(llvm::Value *Val,
1165 CodeGenFunction &CGF) {
1166 if (Val->getType() == Ty)
1169 if (isa<llvm::PointerType>(Val->getType())) {
1170 // If this is Pointer->Pointer avoid conversion to and from int.
1171 if (isa<llvm::PointerType>(Ty))
1172 return CGF.Builder.CreateBitCast(Val, Ty, "coerce.val");
1174 // Convert the pointer to an integer so we can play with its width.
1175 Val = CGF.Builder.CreatePtrToInt(Val, CGF.IntPtrTy, "coerce.val.pi");
1178 llvm::Type *DestIntTy = Ty;
1179 if (isa<llvm::PointerType>(DestIntTy))
1180 DestIntTy = CGF.IntPtrTy;
1182 if (Val->getType() != DestIntTy) {
1183 const llvm::DataLayout &DL = CGF.CGM.getDataLayout();
1184 if (DL.isBigEndian()) {
1185 // Preserve the high bits on big-endian targets.
1186 // That is what memory coercion does.
1187 uint64_t SrcSize = DL.getTypeSizeInBits(Val->getType());
1188 uint64_t DstSize = DL.getTypeSizeInBits(DestIntTy);
1190 if (SrcSize > DstSize) {
1191 Val = CGF.Builder.CreateLShr(Val, SrcSize - DstSize, "coerce.highbits");
1192 Val = CGF.Builder.CreateTrunc(Val, DestIntTy, "coerce.val.ii");
1194 Val = CGF.Builder.CreateZExt(Val, DestIntTy, "coerce.val.ii");
1195 Val = CGF.Builder.CreateShl(Val, DstSize - SrcSize, "coerce.highbits");
1198 // Little-endian targets preserve the low bits. No shifts required.
1199 Val = CGF.Builder.CreateIntCast(Val, DestIntTy, false, "coerce.val.ii");
1203 if (isa<llvm::PointerType>(Ty))
1204 Val = CGF.Builder.CreateIntToPtr(Val, Ty, "coerce.val.ip");
1210 /// CreateCoercedLoad - Create a load from \arg SrcPtr interpreted as
1211 /// a pointer to an object of type \arg Ty, known to be aligned to
1212 /// \arg SrcAlign bytes.
1214 /// This safely handles the case when the src type is smaller than the
1215 /// destination type; in this situation the values of bits which not
1216 /// present in the src are undefined.
1217 static llvm::Value *CreateCoercedLoad(Address Src, llvm::Type *Ty,
1218 CodeGenFunction &CGF) {
1219 llvm::Type *SrcTy = Src.getElementType();
1221 // If SrcTy and Ty are the same, just do a load.
1223 return CGF.Builder.CreateLoad(Src);
1225 uint64_t DstSize = CGF.CGM.getDataLayout().getTypeAllocSize(Ty);
1227 if (llvm::StructType *SrcSTy = dyn_cast<llvm::StructType>(SrcTy)) {
1228 Src = EnterStructPointerForCoercedAccess(Src, SrcSTy, DstSize, CGF);
1229 SrcTy = Src.getType()->getElementType();
1232 uint64_t SrcSize = CGF.CGM.getDataLayout().getTypeAllocSize(SrcTy);
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>(Ty) || isa<llvm::PointerType>(Ty)) &&
1237 (isa<llvm::IntegerType>(SrcTy) || isa<llvm::PointerType>(SrcTy))) {
1238 llvm::Value *Load = CGF.Builder.CreateLoad(Src);
1239 return CoerceIntOrPtrToIntOrPtr(Load, Ty, CGF);
1242 // If load is legal, just bitcast the src pointer.
1243 if (SrcSize >= DstSize) {
1244 // Generally SrcSize is never greater than DstSize, since this means we are
1245 // losing bits. However, this can happen in cases where the structure has
1246 // additional padding, for example due to a user specified alignment.
1248 // FIXME: Assert that we aren't truncating non-padding bits when have access
1249 // to that information.
1250 Src = CGF.Builder.CreateBitCast(Src,
1251 Ty->getPointerTo(Src.getAddressSpace()));
1252 return CGF.Builder.CreateLoad(Src);
1255 // Otherwise do coercion through memory. This is stupid, but simple.
1256 Address Tmp = CreateTempAllocaForCoercion(CGF, Ty, Src.getAlignment());
1257 Address Casted = CGF.Builder.CreateElementBitCast(Tmp,CGF.Int8Ty);
1258 Address SrcCasted = CGF.Builder.CreateElementBitCast(Src,CGF.Int8Ty);
1259 CGF.Builder.CreateMemCpy(Casted, SrcCasted,
1260 llvm::ConstantInt::get(CGF.IntPtrTy, SrcSize),
1262 return CGF.Builder.CreateLoad(Tmp);
1265 // Function to store a first-class aggregate into memory. We prefer to
1266 // store the elements rather than the aggregate to be more friendly to
1268 // FIXME: Do we need to recurse here?
1269 static void BuildAggStore(CodeGenFunction &CGF, llvm::Value *Val,
1270 Address Dest, bool DestIsVolatile) {
1271 // Prefer scalar stores to first-class aggregate stores.
1272 if (llvm::StructType *STy =
1273 dyn_cast<llvm::StructType>(Val->getType())) {
1274 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1275 Address EltPtr = CGF.Builder.CreateStructGEP(Dest, i);
1276 llvm::Value *Elt = CGF.Builder.CreateExtractValue(Val, i);
1277 CGF.Builder.CreateStore(Elt, EltPtr, DestIsVolatile);
1280 CGF.Builder.CreateStore(Val, Dest, DestIsVolatile);
1284 /// CreateCoercedStore - Create a store to \arg DstPtr from \arg Src,
1285 /// where the source and destination may have different types. The
1286 /// destination is known to be aligned to \arg DstAlign bytes.
1288 /// This safely handles the case when the src type is larger than the
1289 /// destination type; the upper bits of the src will be lost.
1290 static void CreateCoercedStore(llvm::Value *Src,
1293 CodeGenFunction &CGF) {
1294 llvm::Type *SrcTy = Src->getType();
1295 llvm::Type *DstTy = Dst.getType()->getElementType();
1296 if (SrcTy == DstTy) {
1297 CGF.Builder.CreateStore(Src, Dst, DstIsVolatile);
1301 uint64_t SrcSize = CGF.CGM.getDataLayout().getTypeAllocSize(SrcTy);
1303 if (llvm::StructType *DstSTy = dyn_cast<llvm::StructType>(DstTy)) {
1304 Dst = EnterStructPointerForCoercedAccess(Dst, DstSTy, SrcSize, CGF);
1305 DstTy = Dst.getType()->getElementType();
1308 // If the source and destination are integer or pointer types, just do an
1309 // extension or truncation to the desired type.
1310 if ((isa<llvm::IntegerType>(SrcTy) || isa<llvm::PointerType>(SrcTy)) &&
1311 (isa<llvm::IntegerType>(DstTy) || isa<llvm::PointerType>(DstTy))) {
1312 Src = CoerceIntOrPtrToIntOrPtr(Src, DstTy, CGF);
1313 CGF.Builder.CreateStore(Src, Dst, DstIsVolatile);
1317 uint64_t DstSize = CGF.CGM.getDataLayout().getTypeAllocSize(DstTy);
1319 // If store is legal, just bitcast the src pointer.
1320 if (SrcSize <= DstSize) {
1321 Dst = CGF.Builder.CreateElementBitCast(Dst, SrcTy);
1322 BuildAggStore(CGF, Src, Dst, DstIsVolatile);
1324 // Otherwise do coercion through memory. This is stupid, but
1327 // Generally SrcSize is never greater than DstSize, since this means we are
1328 // losing bits. However, this can happen in cases where the structure has
1329 // additional padding, for example due to a user specified alignment.
1331 // FIXME: Assert that we aren't truncating non-padding bits when have access
1332 // to that information.
1333 Address Tmp = CreateTempAllocaForCoercion(CGF, SrcTy, Dst.getAlignment());
1334 CGF.Builder.CreateStore(Src, Tmp);
1335 Address Casted = CGF.Builder.CreateElementBitCast(Tmp,CGF.Int8Ty);
1336 Address DstCasted = CGF.Builder.CreateElementBitCast(Dst,CGF.Int8Ty);
1337 CGF.Builder.CreateMemCpy(DstCasted, Casted,
1338 llvm::ConstantInt::get(CGF.IntPtrTy, DstSize),
1343 static Address emitAddressAtOffset(CodeGenFunction &CGF, Address addr,
1344 const ABIArgInfo &info) {
1345 if (unsigned offset = info.getDirectOffset()) {
1346 addr = CGF.Builder.CreateElementBitCast(addr, CGF.Int8Ty);
1347 addr = CGF.Builder.CreateConstInBoundsByteGEP(addr,
1348 CharUnits::fromQuantity(offset));
1349 addr = CGF.Builder.CreateElementBitCast(addr, info.getCoerceToType());
1356 /// Encapsulates information about the way function arguments from
1357 /// CGFunctionInfo should be passed to actual LLVM IR function.
1358 class ClangToLLVMArgMapping {
1359 static const unsigned InvalidIndex = ~0U;
1360 unsigned InallocaArgNo;
1362 unsigned TotalIRArgs;
1364 /// Arguments of LLVM IR function corresponding to single Clang argument.
1366 unsigned PaddingArgIndex;
1367 // Argument is expanded to IR arguments at positions
1368 // [FirstArgIndex, FirstArgIndex + NumberOfArgs).
1369 unsigned FirstArgIndex;
1370 unsigned NumberOfArgs;
1373 : PaddingArgIndex(InvalidIndex), FirstArgIndex(InvalidIndex),
1377 SmallVector<IRArgs, 8> ArgInfo;
1380 ClangToLLVMArgMapping(const ASTContext &Context, const CGFunctionInfo &FI,
1381 bool OnlyRequiredArgs = false)
1382 : InallocaArgNo(InvalidIndex), SRetArgNo(InvalidIndex), TotalIRArgs(0),
1383 ArgInfo(OnlyRequiredArgs ? FI.getNumRequiredArgs() : FI.arg_size()) {
1384 construct(Context, FI, OnlyRequiredArgs);
1387 bool hasInallocaArg() const { return InallocaArgNo != InvalidIndex; }
1388 unsigned getInallocaArgNo() const {
1389 assert(hasInallocaArg());
1390 return InallocaArgNo;
1393 bool hasSRetArg() const { return SRetArgNo != InvalidIndex; }
1394 unsigned getSRetArgNo() const {
1395 assert(hasSRetArg());
1399 unsigned totalIRArgs() const { return TotalIRArgs; }
1401 bool hasPaddingArg(unsigned ArgNo) const {
1402 assert(ArgNo < ArgInfo.size());
1403 return ArgInfo[ArgNo].PaddingArgIndex != InvalidIndex;
1405 unsigned getPaddingArgNo(unsigned ArgNo) const {
1406 assert(hasPaddingArg(ArgNo));
1407 return ArgInfo[ArgNo].PaddingArgIndex;
1410 /// Returns index of first IR argument corresponding to ArgNo, and their
1412 std::pair<unsigned, unsigned> getIRArgs(unsigned ArgNo) const {
1413 assert(ArgNo < ArgInfo.size());
1414 return std::make_pair(ArgInfo[ArgNo].FirstArgIndex,
1415 ArgInfo[ArgNo].NumberOfArgs);
1419 void construct(const ASTContext &Context, const CGFunctionInfo &FI,
1420 bool OnlyRequiredArgs);
1423 void ClangToLLVMArgMapping::construct(const ASTContext &Context,
1424 const CGFunctionInfo &FI,
1425 bool OnlyRequiredArgs) {
1426 unsigned IRArgNo = 0;
1427 bool SwapThisWithSRet = false;
1428 const ABIArgInfo &RetAI = FI.getReturnInfo();
1430 if (RetAI.getKind() == ABIArgInfo::Indirect) {
1431 SwapThisWithSRet = RetAI.isSRetAfterThis();
1432 SRetArgNo = SwapThisWithSRet ? 1 : IRArgNo++;
1436 unsigned NumArgs = OnlyRequiredArgs ? FI.getNumRequiredArgs() : FI.arg_size();
1437 for (CGFunctionInfo::const_arg_iterator I = FI.arg_begin(); ArgNo < NumArgs;
1439 assert(I != FI.arg_end());
1440 QualType ArgType = I->type;
1441 const ABIArgInfo &AI = I->info;
1442 // Collect data about IR arguments corresponding to Clang argument ArgNo.
1443 auto &IRArgs = ArgInfo[ArgNo];
1445 if (AI.getPaddingType())
1446 IRArgs.PaddingArgIndex = IRArgNo++;
1448 switch (AI.getKind()) {
1449 case ABIArgInfo::Extend:
1450 case ABIArgInfo::Direct: {
1451 // FIXME: handle sseregparm someday...
1452 llvm::StructType *STy = dyn_cast<llvm::StructType>(AI.getCoerceToType());
1453 if (AI.isDirect() && AI.getCanBeFlattened() && STy) {
1454 IRArgs.NumberOfArgs = STy->getNumElements();
1456 IRArgs.NumberOfArgs = 1;
1460 case ABIArgInfo::Indirect:
1461 IRArgs.NumberOfArgs = 1;
1463 case ABIArgInfo::Ignore:
1464 case ABIArgInfo::InAlloca:
1465 // ignore and inalloca doesn't have matching LLVM parameters.
1466 IRArgs.NumberOfArgs = 0;
1468 case ABIArgInfo::CoerceAndExpand:
1469 IRArgs.NumberOfArgs = AI.getCoerceAndExpandTypeSequence().size();
1471 case ABIArgInfo::Expand:
1472 IRArgs.NumberOfArgs = getExpansionSize(ArgType, Context);
1476 if (IRArgs.NumberOfArgs > 0) {
1477 IRArgs.FirstArgIndex = IRArgNo;
1478 IRArgNo += IRArgs.NumberOfArgs;
1481 // Skip over the sret parameter when it comes second. We already handled it
1483 if (IRArgNo == 1 && SwapThisWithSRet)
1486 assert(ArgNo == ArgInfo.size());
1488 if (FI.usesInAlloca())
1489 InallocaArgNo = IRArgNo++;
1491 TotalIRArgs = IRArgNo;
1497 bool CodeGenModule::ReturnTypeUsesSRet(const CGFunctionInfo &FI) {
1498 const auto &RI = FI.getReturnInfo();
1499 return RI.isIndirect() || (RI.isInAlloca() && RI.getInAllocaSRet());
1502 bool CodeGenModule::ReturnSlotInterferesWithArgs(const CGFunctionInfo &FI) {
1503 return ReturnTypeUsesSRet(FI) &&
1504 getTargetCodeGenInfo().doesReturnSlotInterfereWithArgs();
1507 bool CodeGenModule::ReturnTypeUsesFPRet(QualType ResultType) {
1508 if (const BuiltinType *BT = ResultType->getAs<BuiltinType>()) {
1509 switch (BT->getKind()) {
1512 case BuiltinType::Float:
1513 return getTarget().useObjCFPRetForRealType(TargetInfo::Float);
1514 case BuiltinType::Double:
1515 return getTarget().useObjCFPRetForRealType(TargetInfo::Double);
1516 case BuiltinType::LongDouble:
1517 return getTarget().useObjCFPRetForRealType(TargetInfo::LongDouble);
1524 bool CodeGenModule::ReturnTypeUsesFP2Ret(QualType ResultType) {
1525 if (const ComplexType *CT = ResultType->getAs<ComplexType>()) {
1526 if (const BuiltinType *BT = CT->getElementType()->getAs<BuiltinType>()) {
1527 if (BT->getKind() == BuiltinType::LongDouble)
1528 return getTarget().useObjCFP2RetForComplexLongDouble();
1535 llvm::FunctionType *CodeGenTypes::GetFunctionType(GlobalDecl GD) {
1536 const CGFunctionInfo &FI = arrangeGlobalDeclaration(GD);
1537 return GetFunctionType(FI);
1540 llvm::FunctionType *
1541 CodeGenTypes::GetFunctionType(const CGFunctionInfo &FI) {
1543 bool Inserted = FunctionsBeingProcessed.insert(&FI).second;
1545 assert(Inserted && "Recursively being processed?");
1547 llvm::Type *resultType = nullptr;
1548 const ABIArgInfo &retAI = FI.getReturnInfo();
1549 switch (retAI.getKind()) {
1550 case ABIArgInfo::Expand:
1551 llvm_unreachable("Invalid ABI kind for return argument");
1553 case ABIArgInfo::Extend:
1554 case ABIArgInfo::Direct:
1555 resultType = retAI.getCoerceToType();
1558 case ABIArgInfo::InAlloca:
1559 if (retAI.getInAllocaSRet()) {
1560 // sret things on win32 aren't void, they return the sret pointer.
1561 QualType ret = FI.getReturnType();
1562 llvm::Type *ty = ConvertType(ret);
1563 unsigned addressSpace = Context.getTargetAddressSpace(ret);
1564 resultType = llvm::PointerType::get(ty, addressSpace);
1566 resultType = llvm::Type::getVoidTy(getLLVMContext());
1570 case ABIArgInfo::Indirect:
1571 case ABIArgInfo::Ignore:
1572 resultType = llvm::Type::getVoidTy(getLLVMContext());
1575 case ABIArgInfo::CoerceAndExpand:
1576 resultType = retAI.getUnpaddedCoerceAndExpandType();
1580 ClangToLLVMArgMapping IRFunctionArgs(getContext(), FI, true);
1581 SmallVector<llvm::Type*, 8> ArgTypes(IRFunctionArgs.totalIRArgs());
1583 // Add type for sret argument.
1584 if (IRFunctionArgs.hasSRetArg()) {
1585 QualType Ret = FI.getReturnType();
1586 llvm::Type *Ty = ConvertType(Ret);
1587 unsigned AddressSpace = Context.getTargetAddressSpace(Ret);
1588 ArgTypes[IRFunctionArgs.getSRetArgNo()] =
1589 llvm::PointerType::get(Ty, AddressSpace);
1592 // Add type for inalloca argument.
1593 if (IRFunctionArgs.hasInallocaArg()) {
1594 auto ArgStruct = FI.getArgStruct();
1596 ArgTypes[IRFunctionArgs.getInallocaArgNo()] = ArgStruct->getPointerTo();
1599 // Add in all of the required arguments.
1601 CGFunctionInfo::const_arg_iterator it = FI.arg_begin(),
1602 ie = it + FI.getNumRequiredArgs();
1603 for (; it != ie; ++it, ++ArgNo) {
1604 const ABIArgInfo &ArgInfo = it->info;
1606 // Insert a padding type to ensure proper alignment.
1607 if (IRFunctionArgs.hasPaddingArg(ArgNo))
1608 ArgTypes[IRFunctionArgs.getPaddingArgNo(ArgNo)] =
1609 ArgInfo.getPaddingType();
1611 unsigned FirstIRArg, NumIRArgs;
1612 std::tie(FirstIRArg, NumIRArgs) = IRFunctionArgs.getIRArgs(ArgNo);
1614 switch (ArgInfo.getKind()) {
1615 case ABIArgInfo::Ignore:
1616 case ABIArgInfo::InAlloca:
1617 assert(NumIRArgs == 0);
1620 case ABIArgInfo::Indirect: {
1621 assert(NumIRArgs == 1);
1622 // indirect arguments are always on the stack, which is alloca addr space.
1623 llvm::Type *LTy = ConvertTypeForMem(it->type);
1624 ArgTypes[FirstIRArg] = LTy->getPointerTo(
1625 CGM.getDataLayout().getAllocaAddrSpace());
1629 case ABIArgInfo::Extend:
1630 case ABIArgInfo::Direct: {
1631 // Fast-isel and the optimizer generally like scalar values better than
1632 // FCAs, so we flatten them if this is safe to do for this argument.
1633 llvm::Type *argType = ArgInfo.getCoerceToType();
1634 llvm::StructType *st = dyn_cast<llvm::StructType>(argType);
1635 if (st && ArgInfo.isDirect() && ArgInfo.getCanBeFlattened()) {
1636 assert(NumIRArgs == st->getNumElements());
1637 for (unsigned i = 0, e = st->getNumElements(); i != e; ++i)
1638 ArgTypes[FirstIRArg + i] = st->getElementType(i);
1640 assert(NumIRArgs == 1);
1641 ArgTypes[FirstIRArg] = argType;
1646 case ABIArgInfo::CoerceAndExpand: {
1647 auto ArgTypesIter = ArgTypes.begin() + FirstIRArg;
1648 for (auto EltTy : ArgInfo.getCoerceAndExpandTypeSequence()) {
1649 *ArgTypesIter++ = EltTy;
1651 assert(ArgTypesIter == ArgTypes.begin() + FirstIRArg + NumIRArgs);
1655 case ABIArgInfo::Expand:
1656 auto ArgTypesIter = ArgTypes.begin() + FirstIRArg;
1657 getExpandedTypes(it->type, ArgTypesIter);
1658 assert(ArgTypesIter == ArgTypes.begin() + FirstIRArg + NumIRArgs);
1663 bool Erased = FunctionsBeingProcessed.erase(&FI); (void)Erased;
1664 assert(Erased && "Not in set?");
1666 return llvm::FunctionType::get(resultType, ArgTypes, FI.isVariadic());
1669 llvm::Type *CodeGenTypes::GetFunctionTypeForVTable(GlobalDecl GD) {
1670 const CXXMethodDecl *MD = cast<CXXMethodDecl>(GD.getDecl());
1671 const FunctionProtoType *FPT = MD->getType()->getAs<FunctionProtoType>();
1673 if (!isFuncTypeConvertible(FPT))
1674 return llvm::StructType::get(getLLVMContext());
1676 return GetFunctionType(GD);
1679 static void AddAttributesFromFunctionProtoType(ASTContext &Ctx,
1680 llvm::AttrBuilder &FuncAttrs,
1681 const FunctionProtoType *FPT) {
1685 if (!isUnresolvedExceptionSpec(FPT->getExceptionSpecType()) &&
1687 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1690 void CodeGenModule::ConstructDefaultFnAttrList(StringRef Name, bool HasOptnone,
1691 bool AttrOnCallSite,
1692 llvm::AttrBuilder &FuncAttrs) {
1693 // OptimizeNoneAttr takes precedence over -Os or -Oz. No warning needed.
1695 if (CodeGenOpts.OptimizeSize)
1696 FuncAttrs.addAttribute(llvm::Attribute::OptimizeForSize);
1697 if (CodeGenOpts.OptimizeSize == 2)
1698 FuncAttrs.addAttribute(llvm::Attribute::MinSize);
1701 if (CodeGenOpts.DisableRedZone)
1702 FuncAttrs.addAttribute(llvm::Attribute::NoRedZone);
1703 if (CodeGenOpts.IndirectTlsSegRefs)
1704 FuncAttrs.addAttribute("indirect-tls-seg-refs");
1705 if (CodeGenOpts.NoImplicitFloat)
1706 FuncAttrs.addAttribute(llvm::Attribute::NoImplicitFloat);
1708 if (AttrOnCallSite) {
1709 // Attributes that should go on the call site only.
1710 if (!CodeGenOpts.SimplifyLibCalls ||
1711 CodeGenOpts.isNoBuiltinFunc(Name.data()))
1712 FuncAttrs.addAttribute(llvm::Attribute::NoBuiltin);
1713 if (!CodeGenOpts.TrapFuncName.empty())
1714 FuncAttrs.addAttribute("trap-func-name", CodeGenOpts.TrapFuncName);
1716 // Attributes that should go on the function, but not the call site.
1717 if (!CodeGenOpts.DisableFPElim) {
1718 FuncAttrs.addAttribute("no-frame-pointer-elim", "false");
1719 } else if (CodeGenOpts.OmitLeafFramePointer) {
1720 FuncAttrs.addAttribute("no-frame-pointer-elim", "false");
1721 FuncAttrs.addAttribute("no-frame-pointer-elim-non-leaf");
1723 FuncAttrs.addAttribute("no-frame-pointer-elim", "true");
1724 FuncAttrs.addAttribute("no-frame-pointer-elim-non-leaf");
1727 FuncAttrs.addAttribute("less-precise-fpmad",
1728 llvm::toStringRef(CodeGenOpts.LessPreciseFPMAD));
1730 if (CodeGenOpts.NullPointerIsValid)
1731 FuncAttrs.addAttribute("null-pointer-is-valid", "true");
1732 if (!CodeGenOpts.FPDenormalMode.empty())
1733 FuncAttrs.addAttribute("denormal-fp-math", CodeGenOpts.FPDenormalMode);
1735 FuncAttrs.addAttribute("no-trapping-math",
1736 llvm::toStringRef(CodeGenOpts.NoTrappingMath));
1738 // Strict (compliant) code is the default, so only add this attribute to
1739 // indicate that we are trying to workaround a problem case.
1740 if (!CodeGenOpts.StrictFloatCastOverflow)
1741 FuncAttrs.addAttribute("strict-float-cast-overflow", "false");
1743 // TODO: Are these all needed?
1744 // unsafe/inf/nan/nsz are handled by instruction-level FastMathFlags.
1745 FuncAttrs.addAttribute("no-infs-fp-math",
1746 llvm::toStringRef(CodeGenOpts.NoInfsFPMath));
1747 FuncAttrs.addAttribute("no-nans-fp-math",
1748 llvm::toStringRef(CodeGenOpts.NoNaNsFPMath));
1749 FuncAttrs.addAttribute("unsafe-fp-math",
1750 llvm::toStringRef(CodeGenOpts.UnsafeFPMath));
1751 FuncAttrs.addAttribute("use-soft-float",
1752 llvm::toStringRef(CodeGenOpts.SoftFloat));
1753 FuncAttrs.addAttribute("stack-protector-buffer-size",
1754 llvm::utostr(CodeGenOpts.SSPBufferSize));
1755 FuncAttrs.addAttribute("no-signed-zeros-fp-math",
1756 llvm::toStringRef(CodeGenOpts.NoSignedZeros));
1757 FuncAttrs.addAttribute(
1758 "correctly-rounded-divide-sqrt-fp-math",
1759 llvm::toStringRef(CodeGenOpts.CorrectlyRoundedDivSqrt));
1761 if (getLangOpts().OpenCL)
1762 FuncAttrs.addAttribute("denorms-are-zero",
1763 llvm::toStringRef(CodeGenOpts.FlushDenorm));
1765 // TODO: Reciprocal estimate codegen options should apply to instructions?
1766 const std::vector<std::string> &Recips = CodeGenOpts.Reciprocals;
1767 if (!Recips.empty())
1768 FuncAttrs.addAttribute("reciprocal-estimates",
1769 llvm::join(Recips, ","));
1771 if (!CodeGenOpts.PreferVectorWidth.empty() &&
1772 CodeGenOpts.PreferVectorWidth != "none")
1773 FuncAttrs.addAttribute("prefer-vector-width",
1774 CodeGenOpts.PreferVectorWidth);
1776 if (CodeGenOpts.StackRealignment)
1777 FuncAttrs.addAttribute("stackrealign");
1778 if (CodeGenOpts.Backchain)
1779 FuncAttrs.addAttribute("backchain");
1781 if (CodeGenOpts.SpeculativeLoadHardening)
1782 FuncAttrs.addAttribute(llvm::Attribute::SpeculativeLoadHardening);
1785 if (getLangOpts().assumeFunctionsAreConvergent()) {
1786 // Conservatively, mark all functions and calls in CUDA and OpenCL as
1787 // convergent (meaning, they may call an intrinsically convergent op, such
1788 // as __syncthreads() / barrier(), and so can't have certain optimizations
1789 // applied around them). LLVM will remove this attribute where it safely
1791 FuncAttrs.addAttribute(llvm::Attribute::Convergent);
1794 if (getLangOpts().CUDA && getLangOpts().CUDAIsDevice) {
1795 // Exceptions aren't supported in CUDA device code.
1796 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1798 // Respect -fcuda-flush-denormals-to-zero.
1799 if (CodeGenOpts.FlushDenorm)
1800 FuncAttrs.addAttribute("nvptx-f32ftz", "true");
1803 for (StringRef Attr : CodeGenOpts.DefaultFunctionAttrs) {
1804 StringRef Var, Value;
1805 std::tie(Var, Value) = Attr.split('=');
1806 FuncAttrs.addAttribute(Var, Value);
1810 void CodeGenModule::AddDefaultFnAttrs(llvm::Function &F) {
1811 llvm::AttrBuilder FuncAttrs;
1812 ConstructDefaultFnAttrList(F.getName(), F.hasOptNone(),
1813 /* AttrOnCallSite = */ false, FuncAttrs);
1814 F.addAttributes(llvm::AttributeList::FunctionIndex, FuncAttrs);
1817 void CodeGenModule::ConstructAttributeList(
1818 StringRef Name, const CGFunctionInfo &FI, CGCalleeInfo CalleeInfo,
1819 llvm::AttributeList &AttrList, unsigned &CallingConv, bool AttrOnCallSite) {
1820 llvm::AttrBuilder FuncAttrs;
1821 llvm::AttrBuilder RetAttrs;
1823 CallingConv = FI.getEffectiveCallingConvention();
1824 if (FI.isNoReturn())
1825 FuncAttrs.addAttribute(llvm::Attribute::NoReturn);
1827 // If we have information about the function prototype, we can learn
1828 // attributes from there.
1829 AddAttributesFromFunctionProtoType(getContext(), FuncAttrs,
1830 CalleeInfo.getCalleeFunctionProtoType());
1832 const Decl *TargetDecl = CalleeInfo.getCalleeDecl().getDecl();
1834 bool HasOptnone = false;
1835 // FIXME: handle sseregparm someday...
1837 if (TargetDecl->hasAttr<ReturnsTwiceAttr>())
1838 FuncAttrs.addAttribute(llvm::Attribute::ReturnsTwice);
1839 if (TargetDecl->hasAttr<NoThrowAttr>())
1840 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1841 if (TargetDecl->hasAttr<NoReturnAttr>())
1842 FuncAttrs.addAttribute(llvm::Attribute::NoReturn);
1843 if (TargetDecl->hasAttr<ColdAttr>())
1844 FuncAttrs.addAttribute(llvm::Attribute::Cold);
1845 if (TargetDecl->hasAttr<NoDuplicateAttr>())
1846 FuncAttrs.addAttribute(llvm::Attribute::NoDuplicate);
1847 if (TargetDecl->hasAttr<ConvergentAttr>())
1848 FuncAttrs.addAttribute(llvm::Attribute::Convergent);
1850 if (const FunctionDecl *Fn = dyn_cast<FunctionDecl>(TargetDecl)) {
1851 AddAttributesFromFunctionProtoType(
1852 getContext(), FuncAttrs, Fn->getType()->getAs<FunctionProtoType>());
1853 // Don't use [[noreturn]] or _Noreturn for a call to a virtual function.
1854 // These attributes are not inherited by overloads.
1855 const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(Fn);
1856 if (Fn->isNoReturn() && !(AttrOnCallSite && MD && MD->isVirtual()))
1857 FuncAttrs.addAttribute(llvm::Attribute::NoReturn);
1860 // 'const', 'pure' and 'noalias' attributed functions are also nounwind.
1861 if (TargetDecl->hasAttr<ConstAttr>()) {
1862 FuncAttrs.addAttribute(llvm::Attribute::ReadNone);
1863 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1864 } else if (TargetDecl->hasAttr<PureAttr>()) {
1865 FuncAttrs.addAttribute(llvm::Attribute::ReadOnly);
1866 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1867 } else if (TargetDecl->hasAttr<NoAliasAttr>()) {
1868 FuncAttrs.addAttribute(llvm::Attribute::ArgMemOnly);
1869 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1871 if (TargetDecl->hasAttr<RestrictAttr>())
1872 RetAttrs.addAttribute(llvm::Attribute::NoAlias);
1873 if (TargetDecl->hasAttr<ReturnsNonNullAttr>() &&
1874 !CodeGenOpts.NullPointerIsValid)
1875 RetAttrs.addAttribute(llvm::Attribute::NonNull);
1876 if (TargetDecl->hasAttr<AnyX86NoCallerSavedRegistersAttr>())
1877 FuncAttrs.addAttribute("no_caller_saved_registers");
1878 if (TargetDecl->hasAttr<AnyX86NoCfCheckAttr>())
1879 FuncAttrs.addAttribute(llvm::Attribute::NoCfCheck);
1881 HasOptnone = TargetDecl->hasAttr<OptimizeNoneAttr>();
1882 if (auto *AllocSize = TargetDecl->getAttr<AllocSizeAttr>()) {
1883 Optional<unsigned> NumElemsParam;
1884 if (AllocSize->getNumElemsParam().isValid())
1885 NumElemsParam = AllocSize->getNumElemsParam().getLLVMIndex();
1886 FuncAttrs.addAllocSizeAttr(AllocSize->getElemSizeParam().getLLVMIndex(),
1891 ConstructDefaultFnAttrList(Name, HasOptnone, AttrOnCallSite, FuncAttrs);
1893 // This must run after constructing the default function attribute list
1894 // to ensure that the speculative load hardening attribute is removed
1895 // in the case where the -mspeculative-load-hardening flag was passed.
1897 if (TargetDecl->hasAttr<NoSpeculativeLoadHardeningAttr>())
1898 FuncAttrs.removeAttribute(llvm::Attribute::SpeculativeLoadHardening);
1899 if (TargetDecl->hasAttr<SpeculativeLoadHardeningAttr>())
1900 FuncAttrs.addAttribute(llvm::Attribute::SpeculativeLoadHardening);
1903 if (CodeGenOpts.EnableSegmentedStacks &&
1904 !(TargetDecl && TargetDecl->hasAttr<NoSplitStackAttr>()))
1905 FuncAttrs.addAttribute("split-stack");
1907 // Add NonLazyBind attribute to function declarations when -fno-plt
1909 if (TargetDecl && CodeGenOpts.NoPLT) {
1910 if (auto *Fn = dyn_cast<FunctionDecl>(TargetDecl)) {
1911 if (!Fn->isDefined() && !AttrOnCallSite) {
1912 FuncAttrs.addAttribute(llvm::Attribute::NonLazyBind);
1917 if (TargetDecl && TargetDecl->hasAttr<OpenCLKernelAttr>()) {
1918 if (getLangOpts().OpenCLVersion <= 120) {
1919 // OpenCL v1.2 Work groups are always uniform
1920 FuncAttrs.addAttribute("uniform-work-group-size", "true");
1922 // OpenCL v2.0 Work groups may be whether uniform or not.
1923 // '-cl-uniform-work-group-size' compile option gets a hint
1924 // to the compiler that the global work-size be a multiple of
1925 // the work-group size specified to clEnqueueNDRangeKernel
1926 // (i.e. work groups are uniform).
1927 FuncAttrs.addAttribute("uniform-work-group-size",
1928 llvm::toStringRef(CodeGenOpts.UniformWGSize));
1932 if (!AttrOnCallSite) {
1933 bool DisableTailCalls = false;
1935 if (CodeGenOpts.DisableTailCalls)
1936 DisableTailCalls = true;
1937 else if (TargetDecl) {
1938 if (TargetDecl->hasAttr<DisableTailCallsAttr>() ||
1939 TargetDecl->hasAttr<AnyX86InterruptAttr>())
1940 DisableTailCalls = true;
1941 else if (CodeGenOpts.NoEscapingBlockTailCalls) {
1942 if (const auto *BD = dyn_cast<BlockDecl>(TargetDecl))
1943 if (!BD->doesNotEscape())
1944 DisableTailCalls = true;
1948 FuncAttrs.addAttribute("disable-tail-calls",
1949 llvm::toStringRef(DisableTailCalls));
1950 GetCPUAndFeaturesAttributes(CalleeInfo.getCalleeDecl(), FuncAttrs);
1953 ClangToLLVMArgMapping IRFunctionArgs(getContext(), FI);
1955 QualType RetTy = FI.getReturnType();
1956 const ABIArgInfo &RetAI = FI.getReturnInfo();
1957 switch (RetAI.getKind()) {
1958 case ABIArgInfo::Extend:
1959 if (RetAI.isSignExt())
1960 RetAttrs.addAttribute(llvm::Attribute::SExt);
1962 RetAttrs.addAttribute(llvm::Attribute::ZExt);
1964 case ABIArgInfo::Direct:
1965 if (RetAI.getInReg())
1966 RetAttrs.addAttribute(llvm::Attribute::InReg);
1968 case ABIArgInfo::Ignore:
1971 case ABIArgInfo::InAlloca:
1972 case ABIArgInfo::Indirect: {
1973 // inalloca and sret disable readnone and readonly
1974 FuncAttrs.removeAttribute(llvm::Attribute::ReadOnly)
1975 .removeAttribute(llvm::Attribute::ReadNone);
1979 case ABIArgInfo::CoerceAndExpand:
1982 case ABIArgInfo::Expand:
1983 llvm_unreachable("Invalid ABI kind for return argument");
1986 if (const auto *RefTy = RetTy->getAs<ReferenceType>()) {
1987 QualType PTy = RefTy->getPointeeType();
1988 if (!PTy->isIncompleteType() && PTy->isConstantSizeType())
1989 RetAttrs.addDereferenceableAttr(getContext().getTypeSizeInChars(PTy)
1991 else if (getContext().getTargetAddressSpace(PTy) == 0 &&
1992 !CodeGenOpts.NullPointerIsValid)
1993 RetAttrs.addAttribute(llvm::Attribute::NonNull);
1996 bool hasUsedSRet = false;
1997 SmallVector<llvm::AttributeSet, 4> ArgAttrs(IRFunctionArgs.totalIRArgs());
1999 // Attach attributes to sret.
2000 if (IRFunctionArgs.hasSRetArg()) {
2001 llvm::AttrBuilder SRETAttrs;
2002 SRETAttrs.addAttribute(llvm::Attribute::StructRet);
2004 if (RetAI.getInReg())
2005 SRETAttrs.addAttribute(llvm::Attribute::InReg);
2006 ArgAttrs[IRFunctionArgs.getSRetArgNo()] =
2007 llvm::AttributeSet::get(getLLVMContext(), SRETAttrs);
2010 // Attach attributes to inalloca argument.
2011 if (IRFunctionArgs.hasInallocaArg()) {
2012 llvm::AttrBuilder Attrs;
2013 Attrs.addAttribute(llvm::Attribute::InAlloca);
2014 ArgAttrs[IRFunctionArgs.getInallocaArgNo()] =
2015 llvm::AttributeSet::get(getLLVMContext(), Attrs);
2019 for (CGFunctionInfo::const_arg_iterator I = FI.arg_begin(),
2021 I != E; ++I, ++ArgNo) {
2022 QualType ParamType = I->type;
2023 const ABIArgInfo &AI = I->info;
2024 llvm::AttrBuilder Attrs;
2026 // Add attribute for padding argument, if necessary.
2027 if (IRFunctionArgs.hasPaddingArg(ArgNo)) {
2028 if (AI.getPaddingInReg()) {
2029 ArgAttrs[IRFunctionArgs.getPaddingArgNo(ArgNo)] =
2030 llvm::AttributeSet::get(
2032 llvm::AttrBuilder().addAttribute(llvm::Attribute::InReg));
2036 // 'restrict' -> 'noalias' is done in EmitFunctionProlog when we
2037 // have the corresponding parameter variable. It doesn't make
2038 // sense to do it here because parameters are so messed up.
2039 switch (AI.getKind()) {
2040 case ABIArgInfo::Extend:
2042 Attrs.addAttribute(llvm::Attribute::SExt);
2044 Attrs.addAttribute(llvm::Attribute::ZExt);
2046 case ABIArgInfo::Direct:
2047 if (ArgNo == 0 && FI.isChainCall())
2048 Attrs.addAttribute(llvm::Attribute::Nest);
2049 else if (AI.getInReg())
2050 Attrs.addAttribute(llvm::Attribute::InReg);
2053 case ABIArgInfo::Indirect: {
2055 Attrs.addAttribute(llvm::Attribute::InReg);
2057 if (AI.getIndirectByVal())
2058 Attrs.addByValAttr(getTypes().ConvertTypeForMem(ParamType));
2060 CharUnits Align = AI.getIndirectAlign();
2062 // In a byval argument, it is important that the required
2063 // alignment of the type is honored, as LLVM might be creating a
2064 // *new* stack object, and needs to know what alignment to give
2065 // it. (Sometimes it can deduce a sensible alignment on its own,
2066 // but not if clang decides it must emit a packed struct, or the
2067 // user specifies increased alignment requirements.)
2069 // This is different from indirect *not* byval, where the object
2070 // exists already, and the align attribute is purely
2072 assert(!Align.isZero());
2074 // For now, only add this when we have a byval argument.
2075 // TODO: be less lazy about updating test cases.
2076 if (AI.getIndirectByVal())
2077 Attrs.addAlignmentAttr(Align.getQuantity());
2079 // byval disables readnone and readonly.
2080 FuncAttrs.removeAttribute(llvm::Attribute::ReadOnly)
2081 .removeAttribute(llvm::Attribute::ReadNone);
2084 case ABIArgInfo::Ignore:
2085 case ABIArgInfo::Expand:
2086 case ABIArgInfo::CoerceAndExpand:
2089 case ABIArgInfo::InAlloca:
2090 // inalloca disables readnone and readonly.
2091 FuncAttrs.removeAttribute(llvm::Attribute::ReadOnly)
2092 .removeAttribute(llvm::Attribute::ReadNone);
2096 if (const auto *RefTy = ParamType->getAs<ReferenceType>()) {
2097 QualType PTy = RefTy->getPointeeType();
2098 if (!PTy->isIncompleteType() && PTy->isConstantSizeType())
2099 Attrs.addDereferenceableAttr(getContext().getTypeSizeInChars(PTy)
2101 else if (getContext().getTargetAddressSpace(PTy) == 0 &&
2102 !CodeGenOpts.NullPointerIsValid)
2103 Attrs.addAttribute(llvm::Attribute::NonNull);
2106 switch (FI.getExtParameterInfo(ArgNo).getABI()) {
2107 case ParameterABI::Ordinary:
2110 case ParameterABI::SwiftIndirectResult: {
2111 // Add 'sret' if we haven't already used it for something, but
2112 // only if the result is void.
2113 if (!hasUsedSRet && RetTy->isVoidType()) {
2114 Attrs.addAttribute(llvm::Attribute::StructRet);
2118 // Add 'noalias' in either case.
2119 Attrs.addAttribute(llvm::Attribute::NoAlias);
2121 // Add 'dereferenceable' and 'alignment'.
2122 auto PTy = ParamType->getPointeeType();
2123 if (!PTy->isIncompleteType() && PTy->isConstantSizeType()) {
2124 auto info = getContext().getTypeInfoInChars(PTy);
2125 Attrs.addDereferenceableAttr(info.first.getQuantity());
2126 Attrs.addAttribute(llvm::Attribute::getWithAlignment(getLLVMContext(),
2127 info.second.getQuantity()));
2132 case ParameterABI::SwiftErrorResult:
2133 Attrs.addAttribute(llvm::Attribute::SwiftError);
2136 case ParameterABI::SwiftContext:
2137 Attrs.addAttribute(llvm::Attribute::SwiftSelf);
2141 if (FI.getExtParameterInfo(ArgNo).isNoEscape())
2142 Attrs.addAttribute(llvm::Attribute::NoCapture);
2144 if (Attrs.hasAttributes()) {
2145 unsigned FirstIRArg, NumIRArgs;
2146 std::tie(FirstIRArg, NumIRArgs) = IRFunctionArgs.getIRArgs(ArgNo);
2147 for (unsigned i = 0; i < NumIRArgs; i++)
2148 ArgAttrs[FirstIRArg + i] =
2149 llvm::AttributeSet::get(getLLVMContext(), Attrs);
2152 assert(ArgNo == FI.arg_size());
2154 AttrList = llvm::AttributeList::get(
2155 getLLVMContext(), llvm::AttributeSet::get(getLLVMContext(), FuncAttrs),
2156 llvm::AttributeSet::get(getLLVMContext(), RetAttrs), ArgAttrs);
2159 /// An argument came in as a promoted argument; demote it back to its
2161 static llvm::Value *emitArgumentDemotion(CodeGenFunction &CGF,
2163 llvm::Value *value) {
2164 llvm::Type *varType = CGF.ConvertType(var->getType());
2166 // This can happen with promotions that actually don't change the
2167 // underlying type, like the enum promotions.
2168 if (value->getType() == varType) return value;
2170 assert((varType->isIntegerTy() || varType->isFloatingPointTy())
2171 && "unexpected promotion type");
2173 if (isa<llvm::IntegerType>(varType))
2174 return CGF.Builder.CreateTrunc(value, varType, "arg.unpromote");
2176 return CGF.Builder.CreateFPCast(value, varType, "arg.unpromote");
2179 /// Returns the attribute (either parameter attribute, or function
2180 /// attribute), which declares argument ArgNo to be non-null.
2181 static const NonNullAttr *getNonNullAttr(const Decl *FD, const ParmVarDecl *PVD,
2182 QualType ArgType, unsigned ArgNo) {
2183 // FIXME: __attribute__((nonnull)) can also be applied to:
2184 // - references to pointers, where the pointee is known to be
2185 // nonnull (apparently a Clang extension)
2186 // - transparent unions containing pointers
2187 // In the former case, LLVM IR cannot represent the constraint. In
2188 // the latter case, we have no guarantee that the transparent union
2189 // is in fact passed as a pointer.
2190 if (!ArgType->isAnyPointerType() && !ArgType->isBlockPointerType())
2192 // First, check attribute on parameter itself.
2194 if (auto ParmNNAttr = PVD->getAttr<NonNullAttr>())
2197 // Check function attributes.
2200 for (const auto *NNAttr : FD->specific_attrs<NonNullAttr>()) {
2201 if (NNAttr->isNonNull(ArgNo))
2208 struct CopyBackSwiftError final : EHScopeStack::Cleanup {
2211 CopyBackSwiftError(Address temp, Address arg) : Temp(temp), Arg(arg) {}
2212 void Emit(CodeGenFunction &CGF, Flags flags) override {
2213 llvm::Value *errorValue = CGF.Builder.CreateLoad(Temp);
2214 CGF.Builder.CreateStore(errorValue, Arg);
2219 void CodeGenFunction::EmitFunctionProlog(const CGFunctionInfo &FI,
2221 const FunctionArgList &Args) {
2222 if (CurCodeDecl && CurCodeDecl->hasAttr<NakedAttr>())
2223 // Naked functions don't have prologues.
2226 // If this is an implicit-return-zero function, go ahead and
2227 // initialize the return value. TODO: it might be nice to have
2228 // a more general mechanism for this that didn't require synthesized
2229 // return statements.
2230 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(CurCodeDecl)) {
2231 if (FD->hasImplicitReturnZero()) {
2232 QualType RetTy = FD->getReturnType().getUnqualifiedType();
2233 llvm::Type* LLVMTy = CGM.getTypes().ConvertType(RetTy);
2234 llvm::Constant* Zero = llvm::Constant::getNullValue(LLVMTy);
2235 Builder.CreateStore(Zero, ReturnValue);
2239 // FIXME: We no longer need the types from FunctionArgList; lift up and
2242 ClangToLLVMArgMapping IRFunctionArgs(CGM.getContext(), FI);
2243 // Flattened function arguments.
2244 SmallVector<llvm::Value *, 16> FnArgs;
2245 FnArgs.reserve(IRFunctionArgs.totalIRArgs());
2246 for (auto &Arg : Fn->args()) {
2247 FnArgs.push_back(&Arg);
2249 assert(FnArgs.size() == IRFunctionArgs.totalIRArgs());
2251 // If we're using inalloca, all the memory arguments are GEPs off of the last
2252 // parameter, which is a pointer to the complete memory area.
2253 Address ArgStruct = Address::invalid();
2254 if (IRFunctionArgs.hasInallocaArg()) {
2255 ArgStruct = Address(FnArgs[IRFunctionArgs.getInallocaArgNo()],
2256 FI.getArgStructAlignment());
2258 assert(ArgStruct.getType() == FI.getArgStruct()->getPointerTo());
2261 // Name the struct return parameter.
2262 if (IRFunctionArgs.hasSRetArg()) {
2263 auto AI = cast<llvm::Argument>(FnArgs[IRFunctionArgs.getSRetArgNo()]);
2264 AI->setName("agg.result");
2265 AI->addAttr(llvm::Attribute::NoAlias);
2268 // Track if we received the parameter as a pointer (indirect, byval, or
2269 // inalloca). If already have a pointer, EmitParmDecl doesn't need to copy it
2270 // into a local alloca for us.
2271 SmallVector<ParamValue, 16> ArgVals;
2272 ArgVals.reserve(Args.size());
2274 // Create a pointer value for every parameter declaration. This usually
2275 // entails copying one or more LLVM IR arguments into an alloca. Don't push
2276 // any cleanups or do anything that might unwind. We do that separately, so
2277 // we can push the cleanups in the correct order for the ABI.
2278 assert(FI.arg_size() == Args.size() &&
2279 "Mismatch between function signature & arguments.");
2281 CGFunctionInfo::const_arg_iterator info_it = FI.arg_begin();
2282 for (FunctionArgList::const_iterator i = Args.begin(), e = Args.end();
2283 i != e; ++i, ++info_it, ++ArgNo) {
2284 const VarDecl *Arg = *i;
2285 const ABIArgInfo &ArgI = info_it->info;
2288 isa<ParmVarDecl>(Arg) && cast<ParmVarDecl>(Arg)->isKNRPromoted();
2289 // We are converting from ABIArgInfo type to VarDecl type directly, unless
2290 // the parameter is promoted. In this case we convert to
2291 // CGFunctionInfo::ArgInfo type with subsequent argument demotion.
2292 QualType Ty = isPromoted ? info_it->type : Arg->getType();
2293 assert(hasScalarEvaluationKind(Ty) ==
2294 hasScalarEvaluationKind(Arg->getType()));
2296 unsigned FirstIRArg, NumIRArgs;
2297 std::tie(FirstIRArg, NumIRArgs) = IRFunctionArgs.getIRArgs(ArgNo);
2299 switch (ArgI.getKind()) {
2300 case ABIArgInfo::InAlloca: {
2301 assert(NumIRArgs == 0);
2302 auto FieldIndex = ArgI.getInAllocaFieldIndex();
2304 Builder.CreateStructGEP(ArgStruct, FieldIndex, Arg->getName());
2305 ArgVals.push_back(ParamValue::forIndirect(V));
2309 case ABIArgInfo::Indirect: {
2310 assert(NumIRArgs == 1);
2311 Address ParamAddr = Address(FnArgs[FirstIRArg], ArgI.getIndirectAlign());
2313 if (!hasScalarEvaluationKind(Ty)) {
2314 // Aggregates and complex variables are accessed by reference. All we
2315 // need to do is realign the value, if requested.
2316 Address V = ParamAddr;
2317 if (ArgI.getIndirectRealign()) {
2318 Address AlignedTemp = CreateMemTemp(Ty, "coerce");
2320 // Copy from the incoming argument pointer to the temporary with the
2321 // appropriate alignment.
2323 // FIXME: We should have a common utility for generating an aggregate
2325 CharUnits Size = getContext().getTypeSizeInChars(Ty);
2326 auto SizeVal = llvm::ConstantInt::get(IntPtrTy, Size.getQuantity());
2327 Address Dst = Builder.CreateBitCast(AlignedTemp, Int8PtrTy);
2328 Address Src = Builder.CreateBitCast(ParamAddr, Int8PtrTy);
2329 Builder.CreateMemCpy(Dst, Src, SizeVal, false);
2332 ArgVals.push_back(ParamValue::forIndirect(V));
2334 // Load scalar value from indirect argument.
2336 EmitLoadOfScalar(ParamAddr, false, Ty, Arg->getBeginLoc());
2339 V = emitArgumentDemotion(*this, Arg, V);
2340 ArgVals.push_back(ParamValue::forDirect(V));
2345 case ABIArgInfo::Extend:
2346 case ABIArgInfo::Direct: {
2348 // If we have the trivial case, handle it with no muss and fuss.
2349 if (!isa<llvm::StructType>(ArgI.getCoerceToType()) &&
2350 ArgI.getCoerceToType() == ConvertType(Ty) &&
2351 ArgI.getDirectOffset() == 0) {
2352 assert(NumIRArgs == 1);
2353 llvm::Value *V = FnArgs[FirstIRArg];
2354 auto AI = cast<llvm::Argument>(V);
2356 if (const ParmVarDecl *PVD = dyn_cast<ParmVarDecl>(Arg)) {
2357 if (getNonNullAttr(CurCodeDecl, PVD, PVD->getType(),
2358 PVD->getFunctionScopeIndex()) &&
2359 !CGM.getCodeGenOpts().NullPointerIsValid)
2360 AI->addAttr(llvm::Attribute::NonNull);
2362 QualType OTy = PVD->getOriginalType();
2363 if (const auto *ArrTy =
2364 getContext().getAsConstantArrayType(OTy)) {
2365 // A C99 array parameter declaration with the static keyword also
2366 // indicates dereferenceability, and if the size is constant we can
2367 // use the dereferenceable attribute (which requires the size in
2369 if (ArrTy->getSizeModifier() == ArrayType::Static) {
2370 QualType ETy = ArrTy->getElementType();
2371 uint64_t ArrSize = ArrTy->getSize().getZExtValue();
2372 if (!ETy->isIncompleteType() && ETy->isConstantSizeType() &&
2374 llvm::AttrBuilder Attrs;
2375 Attrs.addDereferenceableAttr(
2376 getContext().getTypeSizeInChars(ETy).getQuantity()*ArrSize);
2377 AI->addAttrs(Attrs);
2378 } else if (getContext().getTargetAddressSpace(ETy) == 0 &&
2379 !CGM.getCodeGenOpts().NullPointerIsValid) {
2380 AI->addAttr(llvm::Attribute::NonNull);
2383 } else if (const auto *ArrTy =
2384 getContext().getAsVariableArrayType(OTy)) {
2385 // For C99 VLAs with the static keyword, we don't know the size so
2386 // we can't use the dereferenceable attribute, but in addrspace(0)
2387 // we know that it must be nonnull.
2388 if (ArrTy->getSizeModifier() == VariableArrayType::Static &&
2389 !getContext().getTargetAddressSpace(ArrTy->getElementType()) &&
2390 !CGM.getCodeGenOpts().NullPointerIsValid)
2391 AI->addAttr(llvm::Attribute::NonNull);
2394 const auto *AVAttr = PVD->getAttr<AlignValueAttr>();
2396 if (const auto *TOTy = dyn_cast<TypedefType>(OTy))
2397 AVAttr = TOTy->getDecl()->getAttr<AlignValueAttr>();
2398 if (AVAttr && !SanOpts.has(SanitizerKind::Alignment)) {
2399 // If alignment-assumption sanitizer is enabled, we do *not* add
2400 // alignment attribute here, but emit normal alignment assumption,
2401 // so the UBSAN check could function.
2402 llvm::Value *AlignmentValue =
2403 EmitScalarExpr(AVAttr->getAlignment());
2404 llvm::ConstantInt *AlignmentCI =
2405 cast<llvm::ConstantInt>(AlignmentValue);
2406 unsigned Alignment = std::min((unsigned)AlignmentCI->getZExtValue(),
2407 +llvm::Value::MaximumAlignment);
2408 AI->addAttrs(llvm::AttrBuilder().addAlignmentAttr(Alignment));
2412 if (Arg->getType().isRestrictQualified())
2413 AI->addAttr(llvm::Attribute::NoAlias);
2415 // LLVM expects swifterror parameters to be used in very restricted
2416 // ways. Copy the value into a less-restricted temporary.
2417 if (FI.getExtParameterInfo(ArgNo).getABI()
2418 == ParameterABI::SwiftErrorResult) {
2419 QualType pointeeTy = Ty->getPointeeType();
2420 assert(pointeeTy->isPointerType());
2422 CreateMemTemp(pointeeTy, getPointerAlign(), "swifterror.temp");
2423 Address arg = Address(V, getContext().getTypeAlignInChars(pointeeTy));
2424 llvm::Value *incomingErrorValue = Builder.CreateLoad(arg);
2425 Builder.CreateStore(incomingErrorValue, temp);
2426 V = temp.getPointer();
2428 // Push a cleanup to copy the value back at the end of the function.
2429 // The convention does not guarantee that the value will be written
2430 // back if the function exits with an unwind exception.
2431 EHStack.pushCleanup<CopyBackSwiftError>(NormalCleanup, temp, arg);
2434 // Ensure the argument is the correct type.
2435 if (V->getType() != ArgI.getCoerceToType())
2436 V = Builder.CreateBitCast(V, ArgI.getCoerceToType());
2439 V = emitArgumentDemotion(*this, Arg, V);
2441 // Because of merging of function types from multiple decls it is
2442 // possible for the type of an argument to not match the corresponding
2443 // type in the function type. Since we are codegening the callee
2444 // in here, add a cast to the argument type.
2445 llvm::Type *LTy = ConvertType(Arg->getType());
2446 if (V->getType() != LTy)
2447 V = Builder.CreateBitCast(V, LTy);
2449 ArgVals.push_back(ParamValue::forDirect(V));
2453 Address Alloca = CreateMemTemp(Ty, getContext().getDeclAlign(Arg),
2456 // Pointer to store into.
2457 Address Ptr = emitAddressAtOffset(*this, Alloca, ArgI);
2459 // Fast-isel and the optimizer generally like scalar values better than
2460 // FCAs, so we flatten them if this is safe to do for this argument.
2461 llvm::StructType *STy = dyn_cast<llvm::StructType>(ArgI.getCoerceToType());
2462 if (ArgI.isDirect() && ArgI.getCanBeFlattened() && STy &&
2463 STy->getNumElements() > 1) {
2464 uint64_t SrcSize = CGM.getDataLayout().getTypeAllocSize(STy);
2465 llvm::Type *DstTy = Ptr.getElementType();
2466 uint64_t DstSize = CGM.getDataLayout().getTypeAllocSize(DstTy);
2468 Address AddrToStoreInto = Address::invalid();
2469 if (SrcSize <= DstSize) {
2470 AddrToStoreInto = Builder.CreateElementBitCast(Ptr, STy);
2473 CreateTempAlloca(STy, Alloca.getAlignment(), "coerce");
2476 assert(STy->getNumElements() == NumIRArgs);
2477 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
2478 auto AI = FnArgs[FirstIRArg + i];
2479 AI->setName(Arg->getName() + ".coerce" + Twine(i));
2480 Address EltPtr = Builder.CreateStructGEP(AddrToStoreInto, i);
2481 Builder.CreateStore(AI, EltPtr);
2484 if (SrcSize > DstSize) {
2485 Builder.CreateMemCpy(Ptr, AddrToStoreInto, DstSize);
2489 // Simple case, just do a coerced store of the argument into the alloca.
2490 assert(NumIRArgs == 1);
2491 auto AI = FnArgs[FirstIRArg];
2492 AI->setName(Arg->getName() + ".coerce");
2493 CreateCoercedStore(AI, Ptr, /*DstIsVolatile=*/false, *this);
2496 // Match to what EmitParmDecl is expecting for this type.
2497 if (CodeGenFunction::hasScalarEvaluationKind(Ty)) {
2499 EmitLoadOfScalar(Alloca, false, Ty, Arg->getBeginLoc());
2501 V = emitArgumentDemotion(*this, Arg, V);
2502 ArgVals.push_back(ParamValue::forDirect(V));
2504 ArgVals.push_back(ParamValue::forIndirect(Alloca));
2509 case ABIArgInfo::CoerceAndExpand: {
2510 // Reconstruct into a temporary.
2511 Address alloca = CreateMemTemp(Ty, getContext().getDeclAlign(Arg));
2512 ArgVals.push_back(ParamValue::forIndirect(alloca));
2514 auto coercionType = ArgI.getCoerceAndExpandType();
2515 alloca = Builder.CreateElementBitCast(alloca, coercionType);
2517 unsigned argIndex = FirstIRArg;
2518 for (unsigned i = 0, e = coercionType->getNumElements(); i != e; ++i) {
2519 llvm::Type *eltType = coercionType->getElementType(i);
2520 if (ABIArgInfo::isPaddingForCoerceAndExpand(eltType))
2523 auto eltAddr = Builder.CreateStructGEP(alloca, i);
2524 auto elt = FnArgs[argIndex++];
2525 Builder.CreateStore(elt, eltAddr);
2527 assert(argIndex == FirstIRArg + NumIRArgs);
2531 case ABIArgInfo::Expand: {
2532 // If this structure was expanded into multiple arguments then
2533 // we need to create a temporary and reconstruct it from the
2535 Address Alloca = CreateMemTemp(Ty, getContext().getDeclAlign(Arg));
2536 LValue LV = MakeAddrLValue(Alloca, Ty);
2537 ArgVals.push_back(ParamValue::forIndirect(Alloca));
2539 auto FnArgIter = FnArgs.begin() + FirstIRArg;
2540 ExpandTypeFromArgs(Ty, LV, FnArgIter);
2541 assert(FnArgIter == FnArgs.begin() + FirstIRArg + NumIRArgs);
2542 for (unsigned i = 0, e = NumIRArgs; i != e; ++i) {
2543 auto AI = FnArgs[FirstIRArg + i];
2544 AI->setName(Arg->getName() + "." + Twine(i));
2549 case ABIArgInfo::Ignore:
2550 assert(NumIRArgs == 0);
2551 // Initialize the local variable appropriately.
2552 if (!hasScalarEvaluationKind(Ty)) {
2553 ArgVals.push_back(ParamValue::forIndirect(CreateMemTemp(Ty)));
2555 llvm::Value *U = llvm::UndefValue::get(ConvertType(Arg->getType()));
2556 ArgVals.push_back(ParamValue::forDirect(U));
2562 if (getTarget().getCXXABI().areArgsDestroyedLeftToRightInCallee()) {
2563 for (int I = Args.size() - 1; I >= 0; --I)
2564 EmitParmDecl(*Args[I], ArgVals[I], I + 1);
2566 for (unsigned I = 0, E = Args.size(); I != E; ++I)
2567 EmitParmDecl(*Args[I], ArgVals[I], I + 1);
2571 static void eraseUnusedBitCasts(llvm::Instruction *insn) {
2572 while (insn->use_empty()) {
2573 llvm::BitCastInst *bitcast = dyn_cast<llvm::BitCastInst>(insn);
2574 if (!bitcast) return;
2576 // This is "safe" because we would have used a ConstantExpr otherwise.
2577 insn = cast<llvm::Instruction>(bitcast->getOperand(0));
2578 bitcast->eraseFromParent();
2582 /// Try to emit a fused autorelease of a return result.
2583 static llvm::Value *tryEmitFusedAutoreleaseOfResult(CodeGenFunction &CGF,
2584 llvm::Value *result) {
2585 // We must be immediately followed the cast.
2586 llvm::BasicBlock *BB = CGF.Builder.GetInsertBlock();
2587 if (BB->empty()) return nullptr;
2588 if (&BB->back() != result) return nullptr;
2590 llvm::Type *resultType = result->getType();
2592 // result is in a BasicBlock and is therefore an Instruction.
2593 llvm::Instruction *generator = cast<llvm::Instruction>(result);
2595 SmallVector<llvm::Instruction *, 4> InstsToKill;
2598 // %generator = bitcast %type1* %generator2 to %type2*
2599 while (llvm::BitCastInst *bitcast = dyn_cast<llvm::BitCastInst>(generator)) {
2600 // We would have emitted this as a constant if the operand weren't
2602 generator = cast<llvm::Instruction>(bitcast->getOperand(0));
2604 // Require the generator to be immediately followed by the cast.
2605 if (generator->getNextNode() != bitcast)
2608 InstsToKill.push_back(bitcast);
2612 // %generator = call i8* @objc_retain(i8* %originalResult)
2614 // %generator = call i8* @objc_retainAutoreleasedReturnValue(i8* %originalResult)
2615 llvm::CallInst *call = dyn_cast<llvm::CallInst>(generator);
2616 if (!call) return nullptr;
2618 bool doRetainAutorelease;
2620 if (call->getCalledValue() == CGF.CGM.getObjCEntrypoints().objc_retain) {
2621 doRetainAutorelease = true;
2622 } else if (call->getCalledValue() == CGF.CGM.getObjCEntrypoints()
2623 .objc_retainAutoreleasedReturnValue) {
2624 doRetainAutorelease = false;
2626 // If we emitted an assembly marker for this call (and the
2627 // ARCEntrypoints field should have been set if so), go looking
2628 // for that call. If we can't find it, we can't do this
2629 // optimization. But it should always be the immediately previous
2630 // instruction, unless we needed bitcasts around the call.
2631 if (CGF.CGM.getObjCEntrypoints().retainAutoreleasedReturnValueMarker) {
2632 llvm::Instruction *prev = call->getPrevNode();
2634 if (isa<llvm::BitCastInst>(prev)) {
2635 prev = prev->getPrevNode();
2638 assert(isa<llvm::CallInst>(prev));
2639 assert(cast<llvm::CallInst>(prev)->getCalledValue() ==
2640 CGF.CGM.getObjCEntrypoints().retainAutoreleasedReturnValueMarker);
2641 InstsToKill.push_back(prev);
2647 result = call->getArgOperand(0);
2648 InstsToKill.push_back(call);
2650 // Keep killing bitcasts, for sanity. Note that we no longer care
2651 // about precise ordering as long as there's exactly one use.
2652 while (llvm::BitCastInst *bitcast = dyn_cast<llvm::BitCastInst>(result)) {
2653 if (!bitcast->hasOneUse()) break;
2654 InstsToKill.push_back(bitcast);
2655 result = bitcast->getOperand(0);
2658 // Delete all the unnecessary instructions, from latest to earliest.
2659 for (auto *I : InstsToKill)
2660 I->eraseFromParent();
2662 // Do the fused retain/autorelease if we were asked to.
2663 if (doRetainAutorelease)
2664 result = CGF.EmitARCRetainAutoreleaseReturnValue(result);
2666 // Cast back to the result type.
2667 return CGF.Builder.CreateBitCast(result, resultType);
2670 /// If this is a +1 of the value of an immutable 'self', remove it.
2671 static llvm::Value *tryRemoveRetainOfSelf(CodeGenFunction &CGF,
2672 llvm::Value *result) {
2673 // This is only applicable to a method with an immutable 'self'.
2674 const ObjCMethodDecl *method =
2675 dyn_cast_or_null<ObjCMethodDecl>(CGF.CurCodeDecl);
2676 if (!method) return nullptr;
2677 const VarDecl *self = method->getSelfDecl();
2678 if (!self->getType().isConstQualified()) return nullptr;
2680 // Look for a retain call.
2681 llvm::CallInst *retainCall =
2682 dyn_cast<llvm::CallInst>(result->stripPointerCasts());
2684 retainCall->getCalledValue() != CGF.CGM.getObjCEntrypoints().objc_retain)
2687 // Look for an ordinary load of 'self'.
2688 llvm::Value *retainedValue = retainCall->getArgOperand(0);
2689 llvm::LoadInst *load =
2690 dyn_cast<llvm::LoadInst>(retainedValue->stripPointerCasts());
2691 if (!load || load->isAtomic() || load->isVolatile() ||
2692 load->getPointerOperand() != CGF.GetAddrOfLocalVar(self).getPointer())
2695 // Okay! Burn it all down. This relies for correctness on the
2696 // assumption that the retain is emitted as part of the return and
2697 // that thereafter everything is used "linearly".
2698 llvm::Type *resultType = result->getType();
2699 eraseUnusedBitCasts(cast<llvm::Instruction>(result));
2700 assert(retainCall->use_empty());
2701 retainCall->eraseFromParent();
2702 eraseUnusedBitCasts(cast<llvm::Instruction>(retainedValue));
2704 return CGF.Builder.CreateBitCast(load, resultType);
2707 /// Emit an ARC autorelease of the result of a function.
2709 /// \return the value to actually return from the function
2710 static llvm::Value *emitAutoreleaseOfResult(CodeGenFunction &CGF,
2711 llvm::Value *result) {
2712 // If we're returning 'self', kill the initial retain. This is a
2713 // heuristic attempt to "encourage correctness" in the really unfortunate
2714 // case where we have a return of self during a dealloc and we desperately
2715 // need to avoid the possible autorelease.
2716 if (llvm::Value *self = tryRemoveRetainOfSelf(CGF, result))
2719 // At -O0, try to emit a fused retain/autorelease.
2720 if (CGF.shouldUseFusedARCCalls())
2721 if (llvm::Value *fused = tryEmitFusedAutoreleaseOfResult(CGF, result))
2724 return CGF.EmitARCAutoreleaseReturnValue(result);
2727 /// Heuristically search for a dominating store to the return-value slot.
2728 static llvm::StoreInst *findDominatingStoreToReturnValue(CodeGenFunction &CGF) {
2729 // Check if a User is a store which pointerOperand is the ReturnValue.
2730 // We are looking for stores to the ReturnValue, not for stores of the
2731 // ReturnValue to some other location.
2732 auto GetStoreIfValid = [&CGF](llvm::User *U) -> llvm::StoreInst * {
2733 auto *SI = dyn_cast<llvm::StoreInst>(U);
2734 if (!SI || SI->getPointerOperand() != CGF.ReturnValue.getPointer())
2736 // These aren't actually possible for non-coerced returns, and we
2737 // only care about non-coerced returns on this code path.
2738 assert(!SI->isAtomic() && !SI->isVolatile());
2741 // If there are multiple uses of the return-value slot, just check
2742 // for something immediately preceding the IP. Sometimes this can
2743 // happen with how we generate implicit-returns; it can also happen
2744 // with noreturn cleanups.
2745 if (!CGF.ReturnValue.getPointer()->hasOneUse()) {
2746 llvm::BasicBlock *IP = CGF.Builder.GetInsertBlock();
2747 if (IP->empty()) return nullptr;
2748 llvm::Instruction *I = &IP->back();
2750 // Skip lifetime markers
2751 for (llvm::BasicBlock::reverse_iterator II = IP->rbegin(),
2754 if (llvm::IntrinsicInst *Intrinsic =
2755 dyn_cast<llvm::IntrinsicInst>(&*II)) {
2756 if (Intrinsic->getIntrinsicID() == llvm::Intrinsic::lifetime_end) {
2757 const llvm::Value *CastAddr = Intrinsic->getArgOperand(1);
2761 if (isa<llvm::BitCastInst>(&*II) && (CastAddr == &*II))
2769 return GetStoreIfValid(I);
2772 llvm::StoreInst *store =
2773 GetStoreIfValid(CGF.ReturnValue.getPointer()->user_back());
2774 if (!store) return nullptr;
2776 // Now do a first-and-dirty dominance check: just walk up the
2777 // single-predecessors chain from the current insertion point.
2778 llvm::BasicBlock *StoreBB = store->getParent();
2779 llvm::BasicBlock *IP = CGF.Builder.GetInsertBlock();
2780 while (IP != StoreBB) {
2781 if (!(IP = IP->getSinglePredecessor()))
2785 // Okay, the store's basic block dominates the insertion point; we
2786 // can do our thing.
2790 void CodeGenFunction::EmitFunctionEpilog(const CGFunctionInfo &FI,
2792 SourceLocation EndLoc) {
2793 if (FI.isNoReturn()) {
2794 // Noreturn functions don't return.
2795 EmitUnreachable(EndLoc);
2799 if (CurCodeDecl && CurCodeDecl->hasAttr<NakedAttr>()) {
2800 // Naked functions don't have epilogues.
2801 Builder.CreateUnreachable();
2805 // Functions with no result always return void.
2806 if (!ReturnValue.isValid()) {
2807 Builder.CreateRetVoid();
2811 llvm::DebugLoc RetDbgLoc;
2812 llvm::Value *RV = nullptr;
2813 QualType RetTy = FI.getReturnType();
2814 const ABIArgInfo &RetAI = FI.getReturnInfo();
2816 switch (RetAI.getKind()) {
2817 case ABIArgInfo::InAlloca:
2818 // Aggregrates get evaluated directly into the destination. Sometimes we
2819 // need to return the sret value in a register, though.
2820 assert(hasAggregateEvaluationKind(RetTy));
2821 if (RetAI.getInAllocaSRet()) {
2822 llvm::Function::arg_iterator EI = CurFn->arg_end();
2824 llvm::Value *ArgStruct = &*EI;
2825 llvm::Value *SRet = Builder.CreateStructGEP(
2826 nullptr, ArgStruct, RetAI.getInAllocaFieldIndex());
2827 RV = Builder.CreateAlignedLoad(SRet, getPointerAlign(), "sret");
2831 case ABIArgInfo::Indirect: {
2832 auto AI = CurFn->arg_begin();
2833 if (RetAI.isSRetAfterThis())
2835 switch (getEvaluationKind(RetTy)) {
2838 EmitLoadOfComplex(MakeAddrLValue(ReturnValue, RetTy), EndLoc);
2839 EmitStoreOfComplex(RT, MakeNaturalAlignAddrLValue(&*AI, RetTy),
2844 // Do nothing; aggregrates get evaluated directly into the destination.
2847 EmitStoreOfScalar(Builder.CreateLoad(ReturnValue),
2848 MakeNaturalAlignAddrLValue(&*AI, RetTy),
2855 case ABIArgInfo::Extend:
2856 case ABIArgInfo::Direct:
2857 if (RetAI.getCoerceToType() == ConvertType(RetTy) &&
2858 RetAI.getDirectOffset() == 0) {
2859 // The internal return value temp always will have pointer-to-return-type
2860 // type, just do a load.
2862 // If there is a dominating store to ReturnValue, we can elide
2863 // the load, zap the store, and usually zap the alloca.
2864 if (llvm::StoreInst *SI =
2865 findDominatingStoreToReturnValue(*this)) {
2866 // Reuse the debug location from the store unless there is
2867 // cleanup code to be emitted between the store and return
2869 if (EmitRetDbgLoc && !AutoreleaseResult)
2870 RetDbgLoc = SI->getDebugLoc();
2871 // Get the stored value and nuke the now-dead store.
2872 RV = SI->getValueOperand();
2873 SI->eraseFromParent();
2875 // Otherwise, we have to do a simple load.
2877 RV = Builder.CreateLoad(ReturnValue);
2880 // If the value is offset in memory, apply the offset now.
2881 Address V = emitAddressAtOffset(*this, ReturnValue, RetAI);
2883 RV = CreateCoercedLoad(V, RetAI.getCoerceToType(), *this);
2886 // In ARC, end functions that return a retainable type with a call
2887 // to objc_autoreleaseReturnValue.
2888 if (AutoreleaseResult) {
2890 // Type::isObjCRetainabletype has to be called on a QualType that hasn't
2891 // been stripped of the typedefs, so we cannot use RetTy here. Get the
2892 // original return type of FunctionDecl, CurCodeDecl, and BlockDecl from
2893 // CurCodeDecl or BlockInfo.
2896 if (auto *FD = dyn_cast<FunctionDecl>(CurCodeDecl))
2897 RT = FD->getReturnType();
2898 else if (auto *MD = dyn_cast<ObjCMethodDecl>(CurCodeDecl))
2899 RT = MD->getReturnType();
2900 else if (isa<BlockDecl>(CurCodeDecl))
2901 RT = BlockInfo->BlockExpression->getFunctionType()->getReturnType();
2903 llvm_unreachable("Unexpected function/method type");
2905 assert(getLangOpts().ObjCAutoRefCount &&
2906 !FI.isReturnsRetained() &&
2907 RT->isObjCRetainableType());
2909 RV = emitAutoreleaseOfResult(*this, RV);
2914 case ABIArgInfo::Ignore:
2917 case ABIArgInfo::CoerceAndExpand: {
2918 auto coercionType = RetAI.getCoerceAndExpandType();
2920 // Load all of the coerced elements out into results.
2921 llvm::SmallVector<llvm::Value*, 4> results;
2922 Address addr = Builder.CreateElementBitCast(ReturnValue, coercionType);
2923 for (unsigned i = 0, e = coercionType->getNumElements(); i != e; ++i) {
2924 auto coercedEltType = coercionType->getElementType(i);
2925 if (ABIArgInfo::isPaddingForCoerceAndExpand(coercedEltType))
2928 auto eltAddr = Builder.CreateStructGEP(addr, i);
2929 auto elt = Builder.CreateLoad(eltAddr);
2930 results.push_back(elt);
2933 // If we have one result, it's the single direct result type.
2934 if (results.size() == 1) {
2937 // Otherwise, we need to make a first-class aggregate.
2939 // Construct a return type that lacks padding elements.
2940 llvm::Type *returnType = RetAI.getUnpaddedCoerceAndExpandType();
2942 RV = llvm::UndefValue::get(returnType);
2943 for (unsigned i = 0, e = results.size(); i != e; ++i) {
2944 RV = Builder.CreateInsertValue(RV, results[i], i);
2950 case ABIArgInfo::Expand:
2951 llvm_unreachable("Invalid ABI kind for return argument");
2954 llvm::Instruction *Ret;
2956 EmitReturnValueCheck(RV);
2957 Ret = Builder.CreateRet(RV);
2959 Ret = Builder.CreateRetVoid();
2963 Ret->setDebugLoc(std::move(RetDbgLoc));
2966 void CodeGenFunction::EmitReturnValueCheck(llvm::Value *RV) {
2967 // A current decl may not be available when emitting vtable thunks.
2971 ReturnsNonNullAttr *RetNNAttr = nullptr;
2972 if (SanOpts.has(SanitizerKind::ReturnsNonnullAttribute))
2973 RetNNAttr = CurCodeDecl->getAttr<ReturnsNonNullAttr>();
2975 if (!RetNNAttr && !requiresReturnValueNullabilityCheck())
2978 // Prefer the returns_nonnull attribute if it's present.
2979 SourceLocation AttrLoc;
2980 SanitizerMask CheckKind;
2981 SanitizerHandler Handler;
2983 assert(!requiresReturnValueNullabilityCheck() &&
2984 "Cannot check nullability and the nonnull attribute");
2985 AttrLoc = RetNNAttr->getLocation();
2986 CheckKind = SanitizerKind::ReturnsNonnullAttribute;
2987 Handler = SanitizerHandler::NonnullReturn;
2989 if (auto *DD = dyn_cast<DeclaratorDecl>(CurCodeDecl))
2990 if (auto *TSI = DD->getTypeSourceInfo())
2991 if (auto FTL = TSI->getTypeLoc().castAs<FunctionTypeLoc>())
2992 AttrLoc = FTL.getReturnLoc().findNullabilityLoc();
2993 CheckKind = SanitizerKind::NullabilityReturn;
2994 Handler = SanitizerHandler::NullabilityReturn;
2997 SanitizerScope SanScope(this);
2999 // Make sure the "return" source location is valid. If we're checking a
3000 // nullability annotation, make sure the preconditions for the check are met.
3001 llvm::BasicBlock *Check = createBasicBlock("nullcheck");
3002 llvm::BasicBlock *NoCheck = createBasicBlock("no.nullcheck");
3003 llvm::Value *SLocPtr = Builder.CreateLoad(ReturnLocation, "return.sloc.load");
3004 llvm::Value *CanNullCheck = Builder.CreateIsNotNull(SLocPtr);
3005 if (requiresReturnValueNullabilityCheck())
3007 Builder.CreateAnd(CanNullCheck, RetValNullabilityPrecondition);
3008 Builder.CreateCondBr(CanNullCheck, Check, NoCheck);
3011 // Now do the null check.
3012 llvm::Value *Cond = Builder.CreateIsNotNull(RV);
3013 llvm::Constant *StaticData[] = {EmitCheckSourceLocation(AttrLoc)};
3014 llvm::Value *DynamicData[] = {SLocPtr};
3015 EmitCheck(std::make_pair(Cond, CheckKind), Handler, StaticData, DynamicData);
3020 // The return location should not be used after the check has been emitted.
3021 ReturnLocation = Address::invalid();
3025 static bool isInAllocaArgument(CGCXXABI &ABI, QualType type) {
3026 const CXXRecordDecl *RD = type->getAsCXXRecordDecl();
3027 return RD && ABI.getRecordArgABI(RD) == CGCXXABI::RAA_DirectInMemory;
3030 static AggValueSlot createPlaceholderSlot(CodeGenFunction &CGF,
3032 // FIXME: Generate IR in one pass, rather than going back and fixing up these
3034 llvm::Type *IRTy = CGF.ConvertTypeForMem(Ty);
3035 llvm::Type *IRPtrTy = IRTy->getPointerTo();
3036 llvm::Value *Placeholder = llvm::UndefValue::get(IRPtrTy->getPointerTo());
3038 // FIXME: When we generate this IR in one pass, we shouldn't need
3039 // this win32-specific alignment hack.
3040 CharUnits Align = CharUnits::fromQuantity(4);
3041 Placeholder = CGF.Builder.CreateAlignedLoad(IRPtrTy, Placeholder, Align);
3043 return AggValueSlot::forAddr(Address(Placeholder, Align),
3045 AggValueSlot::IsNotDestructed,
3046 AggValueSlot::DoesNotNeedGCBarriers,
3047 AggValueSlot::IsNotAliased,
3048 AggValueSlot::DoesNotOverlap);
3051 void CodeGenFunction::EmitDelegateCallArg(CallArgList &args,
3052 const VarDecl *param,
3053 SourceLocation loc) {
3054 // StartFunction converted the ABI-lowered parameter(s) into a
3055 // local alloca. We need to turn that into an r-value suitable
3057 Address local = GetAddrOfLocalVar(param);
3059 QualType type = param->getType();
3061 if (isInAllocaArgument(CGM.getCXXABI(), type)) {
3062 CGM.ErrorUnsupported(param, "forwarded non-trivially copyable parameter");
3065 // GetAddrOfLocalVar returns a pointer-to-pointer for references,
3066 // but the argument needs to be the original pointer.
3067 if (type->isReferenceType()) {
3068 args.add(RValue::get(Builder.CreateLoad(local)), type);
3070 // In ARC, move out of consumed arguments so that the release cleanup
3071 // entered by StartFunction doesn't cause an over-release. This isn't
3072 // optimal -O0 code generation, but it should get cleaned up when
3073 // optimization is enabled. This also assumes that delegate calls are
3074 // performed exactly once for a set of arguments, but that should be safe.
3075 } else if (getLangOpts().ObjCAutoRefCount &&
3076 param->hasAttr<NSConsumedAttr>() &&
3077 type->isObjCRetainableType()) {
3078 llvm::Value *ptr = Builder.CreateLoad(local);
3080 llvm::ConstantPointerNull::get(cast<llvm::PointerType>(ptr->getType()));
3081 Builder.CreateStore(null, local);
3082 args.add(RValue::get(ptr), type);
3084 // For the most part, we just need to load the alloca, except that
3085 // aggregate r-values are actually pointers to temporaries.
3087 args.add(convertTempToRValue(local, type, loc), type);
3090 // Deactivate the cleanup for the callee-destructed param that was pushed.
3091 if (hasAggregateEvaluationKind(type) && !CurFuncIsThunk &&
3092 type->getAs<RecordType>()->getDecl()->isParamDestroyedInCallee() &&
3093 type.isDestructedType()) {
3094 EHScopeStack::stable_iterator cleanup =
3095 CalleeDestructedParamCleanups.lookup(cast<ParmVarDecl>(param));
3096 assert(cleanup.isValid() &&
3097 "cleanup for callee-destructed param not recorded");
3098 // This unreachable is a temporary marker which will be removed later.
3099 llvm::Instruction *isActive = Builder.CreateUnreachable();
3100 args.addArgCleanupDeactivation(cleanup, isActive);
3104 static bool isProvablyNull(llvm::Value *addr) {
3105 return isa<llvm::ConstantPointerNull>(addr);
3108 /// Emit the actual writing-back of a writeback.
3109 static void emitWriteback(CodeGenFunction &CGF,
3110 const CallArgList::Writeback &writeback) {
3111 const LValue &srcLV = writeback.Source;
3112 Address srcAddr = srcLV.getAddress();
3113 assert(!isProvablyNull(srcAddr.getPointer()) &&
3114 "shouldn't have writeback for provably null argument");
3116 llvm::BasicBlock *contBB = nullptr;
3118 // If the argument wasn't provably non-null, we need to null check
3119 // before doing the store.
3120 bool provablyNonNull = llvm::isKnownNonZero(srcAddr.getPointer(),
3121 CGF.CGM.getDataLayout());
3122 if (!provablyNonNull) {
3123 llvm::BasicBlock *writebackBB = CGF.createBasicBlock("icr.writeback");
3124 contBB = CGF.createBasicBlock("icr.done");
3126 llvm::Value *isNull =
3127 CGF.Builder.CreateIsNull(srcAddr.getPointer(), "icr.isnull");
3128 CGF.Builder.CreateCondBr(isNull, contBB, writebackBB);
3129 CGF.EmitBlock(writebackBB);
3132 // Load the value to writeback.
3133 llvm::Value *value = CGF.Builder.CreateLoad(writeback.Temporary);
3135 // Cast it back, in case we're writing an id to a Foo* or something.
3136 value = CGF.Builder.CreateBitCast(value, srcAddr.getElementType(),
3137 "icr.writeback-cast");
3139 // Perform the writeback.
3141 // If we have a "to use" value, it's something we need to emit a use
3142 // of. This has to be carefully threaded in: if it's done after the
3143 // release it's potentially undefined behavior (and the optimizer
3144 // will ignore it), and if it happens before the retain then the
3145 // optimizer could move the release there.
3146 if (writeback.ToUse) {
3147 assert(srcLV.getObjCLifetime() == Qualifiers::OCL_Strong);
3149 // Retain the new value. No need to block-copy here: the block's
3150 // being passed up the stack.
3151 value = CGF.EmitARCRetainNonBlock(value);
3153 // Emit the intrinsic use here.
3154 CGF.EmitARCIntrinsicUse(writeback.ToUse);
3156 // Load the old value (primitively).
3157 llvm::Value *oldValue = CGF.EmitLoadOfScalar(srcLV, SourceLocation());
3159 // Put the new value in place (primitively).
3160 CGF.EmitStoreOfScalar(value, srcLV, /*init*/ false);
3162 // Release the old value.
3163 CGF.EmitARCRelease(oldValue, srcLV.isARCPreciseLifetime());
3165 // Otherwise, we can just do a normal lvalue store.
3167 CGF.EmitStoreThroughLValue(RValue::get(value), srcLV);
3170 // Jump to the continuation block.
3171 if (!provablyNonNull)
3172 CGF.EmitBlock(contBB);
3175 static void emitWritebacks(CodeGenFunction &CGF,
3176 const CallArgList &args) {
3177 for (const auto &I : args.writebacks())
3178 emitWriteback(CGF, I);
3181 static void deactivateArgCleanupsBeforeCall(CodeGenFunction &CGF,
3182 const CallArgList &CallArgs) {
3183 ArrayRef<CallArgList::CallArgCleanup> Cleanups =
3184 CallArgs.getCleanupsToDeactivate();
3185 // Iterate in reverse to increase the likelihood of popping the cleanup.
3186 for (const auto &I : llvm::reverse(Cleanups)) {
3187 CGF.DeactivateCleanupBlock(I.Cleanup, I.IsActiveIP);
3188 I.IsActiveIP->eraseFromParent();
3192 static const Expr *maybeGetUnaryAddrOfOperand(const Expr *E) {
3193 if (const UnaryOperator *uop = dyn_cast<UnaryOperator>(E->IgnoreParens()))
3194 if (uop->getOpcode() == UO_AddrOf)
3195 return uop->getSubExpr();
3199 /// Emit an argument that's being passed call-by-writeback. That is,
3200 /// we are passing the address of an __autoreleased temporary; it
3201 /// might be copy-initialized with the current value of the given
3202 /// address, but it will definitely be copied out of after the call.
3203 static void emitWritebackArg(CodeGenFunction &CGF, CallArgList &args,
3204 const ObjCIndirectCopyRestoreExpr *CRE) {
3207 // Make an optimistic effort to emit the address as an l-value.
3208 // This can fail if the argument expression is more complicated.
3209 if (const Expr *lvExpr = maybeGetUnaryAddrOfOperand(CRE->getSubExpr())) {
3210 srcLV = CGF.EmitLValue(lvExpr);
3212 // Otherwise, just emit it as a scalar.
3214 Address srcAddr = CGF.EmitPointerWithAlignment(CRE->getSubExpr());
3216 QualType srcAddrType =
3217 CRE->getSubExpr()->getType()->castAs<PointerType>()->getPointeeType();
3218 srcLV = CGF.MakeAddrLValue(srcAddr, srcAddrType);
3220 Address srcAddr = srcLV.getAddress();
3222 // The dest and src types don't necessarily match in LLVM terms
3223 // because of the crazy ObjC compatibility rules.
3225 llvm::PointerType *destType =
3226 cast<llvm::PointerType>(CGF.ConvertType(CRE->getType()));
3228 // If the address is a constant null, just pass the appropriate null.
3229 if (isProvablyNull(srcAddr.getPointer())) {
3230 args.add(RValue::get(llvm::ConstantPointerNull::get(destType)),
3235 // Create the temporary.
3236 Address temp = CGF.CreateTempAlloca(destType->getElementType(),
3237 CGF.getPointerAlign(),
3239 // Loading an l-value can introduce a cleanup if the l-value is __weak,
3240 // and that cleanup will be conditional if we can't prove that the l-value
3241 // isn't null, so we need to register a dominating point so that the cleanups
3242 // system will make valid IR.
3243 CodeGenFunction::ConditionalEvaluation condEval(CGF);
3245 // Zero-initialize it if we're not doing a copy-initialization.
3246 bool shouldCopy = CRE->shouldCopy();
3249 llvm::ConstantPointerNull::get(
3250 cast<llvm::PointerType>(destType->getElementType()));
3251 CGF.Builder.CreateStore(null, temp);
3254 llvm::BasicBlock *contBB = nullptr;
3255 llvm::BasicBlock *originBB = nullptr;
3257 // If the address is *not* known to be non-null, we need to switch.
3258 llvm::Value *finalArgument;
3260 bool provablyNonNull = llvm::isKnownNonZero(srcAddr.getPointer(),
3261 CGF.CGM.getDataLayout());
3262 if (provablyNonNull) {
3263 finalArgument = temp.getPointer();
3265 llvm::Value *isNull =
3266 CGF.Builder.CreateIsNull(srcAddr.getPointer(), "icr.isnull");
3268 finalArgument = CGF.Builder.CreateSelect(isNull,
3269 llvm::ConstantPointerNull::get(destType),
3270 temp.getPointer(), "icr.argument");
3272 // If we need to copy, then the load has to be conditional, which
3273 // means we need control flow.
3275 originBB = CGF.Builder.GetInsertBlock();
3276 contBB = CGF.createBasicBlock("icr.cont");
3277 llvm::BasicBlock *copyBB = CGF.createBasicBlock("icr.copy");
3278 CGF.Builder.CreateCondBr(isNull, contBB, copyBB);
3279 CGF.EmitBlock(copyBB);
3280 condEval.begin(CGF);
3284 llvm::Value *valueToUse = nullptr;
3286 // Perform a copy if necessary.
3288 RValue srcRV = CGF.EmitLoadOfLValue(srcLV, SourceLocation());
3289 assert(srcRV.isScalar());
3291 llvm::Value *src = srcRV.getScalarVal();
3292 src = CGF.Builder.CreateBitCast(src, destType->getElementType(),
3295 // Use an ordinary store, not a store-to-lvalue.
3296 CGF.Builder.CreateStore(src, temp);
3298 // If optimization is enabled, and the value was held in a
3299 // __strong variable, we need to tell the optimizer that this
3300 // value has to stay alive until we're doing the store back.
3301 // This is because the temporary is effectively unretained,
3302 // and so otherwise we can violate the high-level semantics.
3303 if (CGF.CGM.getCodeGenOpts().OptimizationLevel != 0 &&
3304 srcLV.getObjCLifetime() == Qualifiers::OCL_Strong) {
3309 // Finish the control flow if we needed it.
3310 if (shouldCopy && !provablyNonNull) {
3311 llvm::BasicBlock *copyBB = CGF.Builder.GetInsertBlock();
3312 CGF.EmitBlock(contBB);
3314 // Make a phi for the value to intrinsically use.
3316 llvm::PHINode *phiToUse = CGF.Builder.CreatePHI(valueToUse->getType(), 2,
3318 phiToUse->addIncoming(valueToUse, copyBB);
3319 phiToUse->addIncoming(llvm::UndefValue::get(valueToUse->getType()),
3321 valueToUse = phiToUse;
3327 args.addWriteback(srcLV, temp, valueToUse);
3328 args.add(RValue::get(finalArgument), CRE->getType());
3331 void CallArgList::allocateArgumentMemory(CodeGenFunction &CGF) {
3335 llvm::Function *F = CGF.CGM.getIntrinsic(llvm::Intrinsic::stacksave);
3336 StackBase = CGF.Builder.CreateCall(F, {}, "inalloca.save");
3339 void CallArgList::freeArgumentMemory(CodeGenFunction &CGF) const {
3341 // Restore the stack after the call.
3342 llvm::Function *F = CGF.CGM.getIntrinsic(llvm::Intrinsic::stackrestore);
3343 CGF.Builder.CreateCall(F, StackBase);
3347 void CodeGenFunction::EmitNonNullArgCheck(RValue RV, QualType ArgType,
3348 SourceLocation ArgLoc,
3351 if (!AC.getDecl() || !(SanOpts.has(SanitizerKind::NonnullAttribute) ||
3352 SanOpts.has(SanitizerKind::NullabilityArg)))
3355 // The param decl may be missing in a variadic function.
3356 auto PVD = ParmNum < AC.getNumParams() ? AC.getParamDecl(ParmNum) : nullptr;
3357 unsigned ArgNo = PVD ? PVD->getFunctionScopeIndex() : ParmNum;
3359 // Prefer the nonnull attribute if it's present.
3360 const NonNullAttr *NNAttr = nullptr;
3361 if (SanOpts.has(SanitizerKind::NonnullAttribute))
3362 NNAttr = getNonNullAttr(AC.getDecl(), PVD, ArgType, ArgNo);
3364 bool CanCheckNullability = false;
3365 if (SanOpts.has(SanitizerKind::NullabilityArg) && !NNAttr && PVD) {
3366 auto Nullability = PVD->getType()->getNullability(getContext());
3367 CanCheckNullability = Nullability &&
3368 *Nullability == NullabilityKind::NonNull &&
3369 PVD->getTypeSourceInfo();
3372 if (!NNAttr && !CanCheckNullability)
3375 SourceLocation AttrLoc;
3376 SanitizerMask CheckKind;
3377 SanitizerHandler Handler;
3379 AttrLoc = NNAttr->getLocation();
3380 CheckKind = SanitizerKind::NonnullAttribute;
3381 Handler = SanitizerHandler::NonnullArg;
3383 AttrLoc = PVD->getTypeSourceInfo()->getTypeLoc().findNullabilityLoc();
3384 CheckKind = SanitizerKind::NullabilityArg;
3385 Handler = SanitizerHandler::NullabilityArg;
3388 SanitizerScope SanScope(this);
3389 assert(RV.isScalar());
3390 llvm::Value *V = RV.getScalarVal();
3392 Builder.CreateICmpNE(V, llvm::Constant::getNullValue(V->getType()));
3393 llvm::Constant *StaticData[] = {
3394 EmitCheckSourceLocation(ArgLoc), EmitCheckSourceLocation(AttrLoc),
3395 llvm::ConstantInt::get(Int32Ty, ArgNo + 1),
3397 EmitCheck(std::make_pair(Cond, CheckKind), Handler, StaticData, None);
3400 void CodeGenFunction::EmitCallArgs(
3401 CallArgList &Args, ArrayRef<QualType> ArgTypes,
3402 llvm::iterator_range<CallExpr::const_arg_iterator> ArgRange,
3403 AbstractCallee AC, unsigned ParamsToSkip, EvaluationOrder Order) {
3404 assert((int)ArgTypes.size() == (ArgRange.end() - ArgRange.begin()));
3406 // We *have* to evaluate arguments from right to left in the MS C++ ABI,
3407 // because arguments are destroyed left to right in the callee. As a special
3408 // case, there are certain language constructs that require left-to-right
3409 // evaluation, and in those cases we consider the evaluation order requirement
3410 // to trump the "destruction order is reverse construction order" guarantee.
3412 CGM.getTarget().getCXXABI().areArgsDestroyedLeftToRightInCallee()
3413 ? Order == EvaluationOrder::ForceLeftToRight
3414 : Order != EvaluationOrder::ForceRightToLeft;
3416 auto MaybeEmitImplicitObjectSize = [&](unsigned I, const Expr *Arg,
3417 RValue EmittedArg) {
3418 if (!AC.hasFunctionDecl() || I >= AC.getNumParams())
3420 auto *PS = AC.getParamDecl(I)->getAttr<PassObjectSizeAttr>();
3424 const auto &Context = getContext();
3425 auto SizeTy = Context.getSizeType();
3426 auto T = Builder.getIntNTy(Context.getTypeSize(SizeTy));
3427 assert(EmittedArg.getScalarVal() && "We emitted nothing for the arg?");
3428 llvm::Value *V = evaluateOrEmitBuiltinObjectSize(Arg, PS->getType(), T,
3429 EmittedArg.getScalarVal(),
3431 Args.add(RValue::get(V), SizeTy);
3432 // If we're emitting args in reverse, be sure to do so with
3433 // pass_object_size, as well.
3435 std::swap(Args.back(), *(&Args.back() - 1));
3438 // Insert a stack save if we're going to need any inalloca args.
3439 bool HasInAllocaArgs = false;
3440 if (CGM.getTarget().getCXXABI().isMicrosoft()) {
3441 for (ArrayRef<QualType>::iterator I = ArgTypes.begin(), E = ArgTypes.end();
3442 I != E && !HasInAllocaArgs; ++I)
3443 HasInAllocaArgs = isInAllocaArgument(CGM.getCXXABI(), *I);
3444 if (HasInAllocaArgs) {
3445 assert(getTarget().getTriple().getArch() == llvm::Triple::x86);
3446 Args.allocateArgumentMemory(*this);
3450 // Evaluate each argument in the appropriate order.
3451 size_t CallArgsStart = Args.size();
3452 for (unsigned I = 0, E = ArgTypes.size(); I != E; ++I) {
3453 unsigned Idx = LeftToRight ? I : E - I - 1;
3454 CallExpr::const_arg_iterator Arg = ArgRange.begin() + Idx;
3455 unsigned InitialArgSize = Args.size();
3456 // If *Arg is an ObjCIndirectCopyRestoreExpr, check that either the types of
3457 // the argument and parameter match or the objc method is parameterized.
3458 assert((!isa<ObjCIndirectCopyRestoreExpr>(*Arg) ||
3459 getContext().hasSameUnqualifiedType((*Arg)->getType(),
3461 (isa<ObjCMethodDecl>(AC.getDecl()) &&
3462 isObjCMethodWithTypeParams(cast<ObjCMethodDecl>(AC.getDecl())))) &&
3463 "Argument and parameter types don't match");
3464 EmitCallArg(Args, *Arg, ArgTypes[Idx]);
3465 // In particular, we depend on it being the last arg in Args, and the
3466 // objectsize bits depend on there only being one arg if !LeftToRight.
3467 assert(InitialArgSize + 1 == Args.size() &&
3468 "The code below depends on only adding one arg per EmitCallArg");
3469 (void)InitialArgSize;
3470 // Since pointer argument are never emitted as LValue, it is safe to emit
3471 // non-null argument check for r-value only.
3472 if (!Args.back().hasLValue()) {
3473 RValue RVArg = Args.back().getKnownRValue();
3474 EmitNonNullArgCheck(RVArg, ArgTypes[Idx], (*Arg)->getExprLoc(), AC,
3475 ParamsToSkip + Idx);
3476 // @llvm.objectsize should never have side-effects and shouldn't need
3477 // destruction/cleanups, so we can safely "emit" it after its arg,
3478 // regardless of right-to-leftness
3479 MaybeEmitImplicitObjectSize(Idx, *Arg, RVArg);
3484 // Un-reverse the arguments we just evaluated so they match up with the LLVM
3486 std::reverse(Args.begin() + CallArgsStart, Args.end());
3492 struct DestroyUnpassedArg final : EHScopeStack::Cleanup {
3493 DestroyUnpassedArg(Address Addr, QualType Ty)
3494 : Addr(Addr), Ty(Ty) {}
3499 void Emit(CodeGenFunction &CGF, Flags flags) override {
3500 QualType::DestructionKind DtorKind = Ty.isDestructedType();
3501 if (DtorKind == QualType::DK_cxx_destructor) {
3502 const CXXDestructorDecl *Dtor = Ty->getAsCXXRecordDecl()->getDestructor();
3503 assert(!Dtor->isTrivial());
3504 CGF.EmitCXXDestructorCall(Dtor, Dtor_Complete, /*for vbase*/ false,
3505 /*Delegating=*/false, Addr, Ty);
3507 CGF.callCStructDestructor(CGF.MakeAddrLValue(Addr, Ty));
3512 struct DisableDebugLocationUpdates {
3513 CodeGenFunction &CGF;
3514 bool disabledDebugInfo;
3515 DisableDebugLocationUpdates(CodeGenFunction &CGF, const Expr *E) : CGF(CGF) {
3516 if ((disabledDebugInfo = isa<CXXDefaultArgExpr>(E) && CGF.getDebugInfo()))
3517 CGF.disableDebugInfo();
3519 ~DisableDebugLocationUpdates() {
3520 if (disabledDebugInfo)
3521 CGF.enableDebugInfo();
3525 } // end anonymous namespace
3527 RValue CallArg::getRValue(CodeGenFunction &CGF) const {
3530 LValue Copy = CGF.MakeAddrLValue(CGF.CreateMemTemp(Ty), Ty);
3531 CGF.EmitAggregateCopy(Copy, LV, Ty, AggValueSlot::DoesNotOverlap,
3534 return RValue::getAggregate(Copy.getAddress());
3537 void CallArg::copyInto(CodeGenFunction &CGF, Address Addr) const {
3538 LValue Dst = CGF.MakeAddrLValue(Addr, Ty);
3539 if (!HasLV && RV.isScalar())
3540 CGF.EmitStoreOfScalar(RV.getScalarVal(), Dst, /*isInit=*/true);
3541 else if (!HasLV && RV.isComplex())
3542 CGF.EmitStoreOfComplex(RV.getComplexVal(), Dst, /*init=*/true);
3544 auto Addr = HasLV ? LV.getAddress() : RV.getAggregateAddress();
3545 LValue SrcLV = CGF.MakeAddrLValue(Addr, Ty);
3546 // We assume that call args are never copied into subobjects.
3547 CGF.EmitAggregateCopy(Dst, SrcLV, Ty, AggValueSlot::DoesNotOverlap,
3548 HasLV ? LV.isVolatileQualified()
3549 : RV.isVolatileQualified());
3554 void CodeGenFunction::EmitCallArg(CallArgList &args, const Expr *E,
3556 DisableDebugLocationUpdates Dis(*this, E);
3557 if (const ObjCIndirectCopyRestoreExpr *CRE
3558 = dyn_cast<ObjCIndirectCopyRestoreExpr>(E)) {
3559 assert(getLangOpts().ObjCAutoRefCount);
3560 return emitWritebackArg(*this, args, CRE);
3563 assert(type->isReferenceType() == E->isGLValue() &&
3564 "reference binding to unmaterialized r-value!");
3566 if (E->isGLValue()) {
3567 assert(E->getObjectKind() == OK_Ordinary);
3568 return args.add(EmitReferenceBindingToExpr(E), type);
3571 bool HasAggregateEvalKind = hasAggregateEvaluationKind(type);
3573 // In the Microsoft C++ ABI, aggregate arguments are destructed by the callee.
3574 // However, we still have to push an EH-only cleanup in case we unwind before
3575 // we make it to the call.
3576 if (HasAggregateEvalKind &&
3577 type->getAs<RecordType>()->getDecl()->isParamDestroyedInCallee()) {
3578 // If we're using inalloca, use the argument memory. Otherwise, use a
3581 if (args.isUsingInAlloca())
3582 Slot = createPlaceholderSlot(*this, type);
3584 Slot = CreateAggTemp(type, "agg.tmp");
3586 bool DestroyedInCallee = true, NeedsEHCleanup = true;
3587 if (const auto *RD = type->getAsCXXRecordDecl())
3588 DestroyedInCallee = RD->hasNonTrivialDestructor();
3590 NeedsEHCleanup = needsEHCleanup(type.isDestructedType());
3592 if (DestroyedInCallee)
3593 Slot.setExternallyDestructed();
3595 EmitAggExpr(E, Slot);
3596 RValue RV = Slot.asRValue();
3599 if (DestroyedInCallee && NeedsEHCleanup) {
3600 // Create a no-op GEP between the placeholder and the cleanup so we can
3601 // RAUW it successfully. It also serves as a marker of the first
3602 // instruction where the cleanup is active.
3603 pushFullExprCleanup<DestroyUnpassedArg>(EHCleanup, Slot.getAddress(),
3605 // This unreachable is a temporary marker which will be removed later.
3606 llvm::Instruction *IsActive = Builder.CreateUnreachable();
3607 args.addArgCleanupDeactivation(EHStack.getInnermostEHScope(), IsActive);
3612 if (HasAggregateEvalKind && isa<ImplicitCastExpr>(E) &&
3613 cast<CastExpr>(E)->getCastKind() == CK_LValueToRValue) {
3614 LValue L = EmitLValue(cast<CastExpr>(E)->getSubExpr());
3615 assert(L.isSimple());
3616 args.addUncopiedAggregate(L, type);
3620 args.add(EmitAnyExprToTemp(E), type);
3623 QualType CodeGenFunction::getVarArgType(const Expr *Arg) {
3624 // System headers on Windows define NULL to 0 instead of 0LL on Win64. MSVC
3625 // implicitly widens null pointer constants that are arguments to varargs
3626 // functions to pointer-sized ints.
3627 if (!getTarget().getTriple().isOSWindows())
3628 return Arg->getType();
3630 if (Arg->getType()->isIntegerType() &&
3631 getContext().getTypeSize(Arg->getType()) <
3632 getContext().getTargetInfo().getPointerWidth(0) &&
3633 Arg->isNullPointerConstant(getContext(),
3634 Expr::NPC_ValueDependentIsNotNull)) {
3635 return getContext().getIntPtrType();
3638 return Arg->getType();
3641 // In ObjC ARC mode with no ObjC ARC exception safety, tell the ARC
3642 // optimizer it can aggressively ignore unwind edges.
3644 CodeGenFunction::AddObjCARCExceptionMetadata(llvm::Instruction *Inst) {
3645 if (CGM.getCodeGenOpts().OptimizationLevel != 0 &&
3646 !CGM.getCodeGenOpts().ObjCAutoRefCountExceptions)
3647 Inst->setMetadata("clang.arc.no_objc_arc_exceptions",
3648 CGM.getNoObjCARCExceptionsMetadata());
3651 /// Emits a call to the given no-arguments nounwind runtime function.
3653 CodeGenFunction::EmitNounwindRuntimeCall(llvm::FunctionCallee callee,
3654 const llvm::Twine &name) {
3655 return EmitNounwindRuntimeCall(callee, None, name);
3658 /// Emits a call to the given nounwind runtime function.
3660 CodeGenFunction::EmitNounwindRuntimeCall(llvm::FunctionCallee callee,
3661 ArrayRef<llvm::Value *> args,
3662 const llvm::Twine &name) {
3663 llvm::CallInst *call = EmitRuntimeCall(callee, args, name);
3664 call->setDoesNotThrow();
3668 /// Emits a simple call (never an invoke) to the given no-arguments
3669 /// runtime function.
3670 llvm::CallInst *CodeGenFunction::EmitRuntimeCall(llvm::FunctionCallee callee,
3671 const llvm::Twine &name) {
3672 return EmitRuntimeCall(callee, None, name);
3675 // Calls which may throw must have operand bundles indicating which funclet
3676 // they are nested within.
3677 SmallVector<llvm::OperandBundleDef, 1>
3678 CodeGenFunction::getBundlesForFunclet(llvm::Value *Callee) {
3679 SmallVector<llvm::OperandBundleDef, 1> BundleList;
3680 // There is no need for a funclet operand bundle if we aren't inside a
3682 if (!CurrentFuncletPad)
3685 // Skip intrinsics which cannot throw.
3686 auto *CalleeFn = dyn_cast<llvm::Function>(Callee->stripPointerCasts());
3687 if (CalleeFn && CalleeFn->isIntrinsic() && CalleeFn->doesNotThrow())
3690 BundleList.emplace_back("funclet", CurrentFuncletPad);
3694 /// Emits a simple call (never an invoke) to the given runtime function.
3695 llvm::CallInst *CodeGenFunction::EmitRuntimeCall(llvm::FunctionCallee callee,
3696 ArrayRef<llvm::Value *> args,
3697 const llvm::Twine &name) {
3698 llvm::CallInst *call = Builder.CreateCall(
3699 callee, args, getBundlesForFunclet(callee.getCallee()), name);
3700 call->setCallingConv(getRuntimeCC());
3704 /// Emits a call or invoke to the given noreturn runtime function.
3705 void CodeGenFunction::EmitNoreturnRuntimeCallOrInvoke(
3706 llvm::FunctionCallee callee, ArrayRef<llvm::Value *> args) {
3707 SmallVector<llvm::OperandBundleDef, 1> BundleList =
3708 getBundlesForFunclet(callee.getCallee());
3710 if (getInvokeDest()) {
3711 llvm::InvokeInst *invoke =
3712 Builder.CreateInvoke(callee,
3713 getUnreachableBlock(),
3717 invoke->setDoesNotReturn();
3718 invoke->setCallingConv(getRuntimeCC());
3720 llvm::CallInst *call = Builder.CreateCall(callee, args, BundleList);
3721 call->setDoesNotReturn();
3722 call->setCallingConv(getRuntimeCC());
3723 Builder.CreateUnreachable();
3727 /// Emits a call or invoke instruction to the given nullary runtime function.
3729 CodeGenFunction::EmitRuntimeCallOrInvoke(llvm::FunctionCallee callee,
3730 const Twine &name) {
3731 return EmitRuntimeCallOrInvoke(callee, None, name);
3734 /// Emits a call or invoke instruction to the given runtime function.
3736 CodeGenFunction::EmitRuntimeCallOrInvoke(llvm::FunctionCallee callee,
3737 ArrayRef<llvm::Value *> args,
3738 const Twine &name) {
3739 llvm::CallBase *call = EmitCallOrInvoke(callee, args, name);
3740 call->setCallingConv(getRuntimeCC());
3744 /// Emits a call or invoke instruction to the given function, depending
3745 /// on the current state of the EH stack.
3746 llvm::CallBase *CodeGenFunction::EmitCallOrInvoke(llvm::FunctionCallee Callee,
3747 ArrayRef<llvm::Value *> Args,
3748 const Twine &Name) {
3749 llvm::BasicBlock *InvokeDest = getInvokeDest();
3750 SmallVector<llvm::OperandBundleDef, 1> BundleList =
3751 getBundlesForFunclet(Callee.getCallee());
3753 llvm::CallBase *Inst;
3755 Inst = Builder.CreateCall(Callee, Args, BundleList, Name);
3757 llvm::BasicBlock *ContBB = createBasicBlock("invoke.cont");
3758 Inst = Builder.CreateInvoke(Callee, ContBB, InvokeDest, Args, BundleList,
3763 // In ObjC ARC mode with no ObjC ARC exception safety, tell the ARC
3764 // optimizer it can aggressively ignore unwind edges.
3765 if (CGM.getLangOpts().ObjCAutoRefCount)
3766 AddObjCARCExceptionMetadata(Inst);
3771 void CodeGenFunction::deferPlaceholderReplacement(llvm::Instruction *Old,
3773 DeferredReplacements.push_back(std::make_pair(Old, New));
3776 RValue CodeGenFunction::EmitCall(const CGFunctionInfo &CallInfo,
3777 const CGCallee &Callee,
3778 ReturnValueSlot ReturnValue,
3779 const CallArgList &CallArgs,
3780 llvm::CallBase **callOrInvoke,
3781 SourceLocation Loc) {
3782 // FIXME: We no longer need the types from CallArgs; lift up and simplify.
3784 assert(Callee.isOrdinary() || Callee.isVirtual());
3786 // Handle struct-return functions by passing a pointer to the
3787 // location that we would like to return into.
3788 QualType RetTy = CallInfo.getReturnType();
3789 const ABIArgInfo &RetAI = CallInfo.getReturnInfo();
3791 llvm::FunctionType *IRFuncTy = getTypes().GetFunctionType(CallInfo);
3793 const Decl *TargetDecl = Callee.getAbstractInfo().getCalleeDecl().getDecl();
3794 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(TargetDecl))
3795 // We can only guarantee that a function is called from the correct
3796 // context/function based on the appropriate target attributes,
3797 // so only check in the case where we have both always_inline and target
3798 // since otherwise we could be making a conditional call after a check for
3799 // the proper cpu features (and it won't cause code generation issues due to
3800 // function based code generation).
3801 if (TargetDecl->hasAttr<AlwaysInlineAttr>() &&
3802 TargetDecl->hasAttr<TargetAttr>())
3803 checkTargetFeatures(Loc, FD);
3806 if (!(CallInfo.isVariadic() && CallInfo.getArgStruct())) {
3807 // For an inalloca varargs function, we don't expect CallInfo to match the
3808 // function pointer's type, because the inalloca struct a will have extra
3809 // fields in it for the varargs parameters. Code later in this function
3810 // bitcasts the function pointer to the type derived from CallInfo.
3812 // In other cases, we assert that the types match up (until pointers stop
3813 // having pointee types).
3814 llvm::Type *TypeFromVal;
3815 if (Callee.isVirtual())
3816 TypeFromVal = Callee.getVirtualFunctionType();
3819 Callee.getFunctionPointer()->getType()->getPointerElementType();
3820 assert(IRFuncTy == TypeFromVal);
3824 // 1. Set up the arguments.
3826 // If we're using inalloca, insert the allocation after the stack save.
3827 // FIXME: Do this earlier rather than hacking it in here!
3828 Address ArgMemory = Address::invalid();
3829 if (llvm::StructType *ArgStruct = CallInfo.getArgStruct()) {
3830 const llvm::DataLayout &DL = CGM.getDataLayout();
3831 llvm::Instruction *IP = CallArgs.getStackBase();
3832 llvm::AllocaInst *AI;
3834 IP = IP->getNextNode();
3835 AI = new llvm::AllocaInst(ArgStruct, DL.getAllocaAddrSpace(),
3838 AI = CreateTempAlloca(ArgStruct, "argmem");
3840 auto Align = CallInfo.getArgStructAlignment();
3841 AI->setAlignment(Align.getQuantity());
3842 AI->setUsedWithInAlloca(true);
3843 assert(AI->isUsedWithInAlloca() && !AI->isStaticAlloca());
3844 ArgMemory = Address(AI, Align);
3847 ClangToLLVMArgMapping IRFunctionArgs(CGM.getContext(), CallInfo);
3848 SmallVector<llvm::Value *, 16> IRCallArgs(IRFunctionArgs.totalIRArgs());
3850 // If the call returns a temporary with struct return, create a temporary
3851 // alloca to hold the result, unless one is given to us.
3852 Address SRetPtr = Address::invalid();
3853 Address SRetAlloca = Address::invalid();
3854 llvm::Value *UnusedReturnSizePtr = nullptr;
3855 if (RetAI.isIndirect() || RetAI.isInAlloca() || RetAI.isCoerceAndExpand()) {
3856 if (!ReturnValue.isNull()) {
3857 SRetPtr = ReturnValue.getValue();
3859 SRetPtr = CreateMemTemp(RetTy, "tmp", &SRetAlloca);
3860 if (HaveInsertPoint() && ReturnValue.isUnused()) {
3862 CGM.getDataLayout().getTypeAllocSize(ConvertTypeForMem(RetTy));
3863 UnusedReturnSizePtr = EmitLifetimeStart(size, SRetAlloca.getPointer());
3866 if (IRFunctionArgs.hasSRetArg()) {
3867 IRCallArgs[IRFunctionArgs.getSRetArgNo()] = SRetPtr.getPointer();
3868 } else if (RetAI.isInAlloca()) {
3870 Builder.CreateStructGEP(ArgMemory, RetAI.getInAllocaFieldIndex());
3871 Builder.CreateStore(SRetPtr.getPointer(), Addr);
3875 Address swiftErrorTemp = Address::invalid();
3876 Address swiftErrorArg = Address::invalid();
3878 // Translate all of the arguments as necessary to match the IR lowering.
3879 assert(CallInfo.arg_size() == CallArgs.size() &&
3880 "Mismatch between function signature & arguments.");
3882 CGFunctionInfo::const_arg_iterator info_it = CallInfo.arg_begin();
3883 for (CallArgList::const_iterator I = CallArgs.begin(), E = CallArgs.end();
3884 I != E; ++I, ++info_it, ++ArgNo) {
3885 const ABIArgInfo &ArgInfo = info_it->info;
3887 // Insert a padding argument to ensure proper alignment.
3888 if (IRFunctionArgs.hasPaddingArg(ArgNo))
3889 IRCallArgs[IRFunctionArgs.getPaddingArgNo(ArgNo)] =
3890 llvm::UndefValue::get(ArgInfo.getPaddingType());
3892 unsigned FirstIRArg, NumIRArgs;
3893 std::tie(FirstIRArg, NumIRArgs) = IRFunctionArgs.getIRArgs(ArgNo);
3895 switch (ArgInfo.getKind()) {
3896 case ABIArgInfo::InAlloca: {
3897 assert(NumIRArgs == 0);
3898 assert(getTarget().getTriple().getArch() == llvm::Triple::x86);
3899 if (I->isAggregate()) {
3900 // Replace the placeholder with the appropriate argument slot GEP.
3901 Address Addr = I->hasLValue()
3902 ? I->getKnownLValue().getAddress()
3903 : I->getKnownRValue().getAggregateAddress();
3904 llvm::Instruction *Placeholder =
3905 cast<llvm::Instruction>(Addr.getPointer());
3906 CGBuilderTy::InsertPoint IP = Builder.saveIP();
3907 Builder.SetInsertPoint(Placeholder);
3909 Builder.CreateStructGEP(ArgMemory, ArgInfo.getInAllocaFieldIndex());
3910 Builder.restoreIP(IP);
3911 deferPlaceholderReplacement(Placeholder, Addr.getPointer());
3913 // Store the RValue into the argument struct.
3915 Builder.CreateStructGEP(ArgMemory, ArgInfo.getInAllocaFieldIndex());
3916 unsigned AS = Addr.getType()->getPointerAddressSpace();
3917 llvm::Type *MemType = ConvertTypeForMem(I->Ty)->getPointerTo(AS);
3918 // There are some cases where a trivial bitcast is not avoidable. The
3919 // definition of a type later in a translation unit may change it's type
3920 // from {}* to (%struct.foo*)*.
3921 if (Addr.getType() != MemType)
3922 Addr = Builder.CreateBitCast(Addr, MemType);
3923 I->copyInto(*this, Addr);
3928 case ABIArgInfo::Indirect: {
3929 assert(NumIRArgs == 1);
3930 if (!I->isAggregate()) {
3931 // Make a temporary alloca to pass the argument.
3932 Address Addr = CreateMemTempWithoutCast(
3933 I->Ty, ArgInfo.getIndirectAlign(), "indirect-arg-temp");
3934 IRCallArgs[FirstIRArg] = Addr.getPointer();
3936 I->copyInto(*this, Addr);
3938 // We want to avoid creating an unnecessary temporary+copy here;
3939 // however, we need one in three cases:
3940 // 1. If the argument is not byval, and we are required to copy the
3941 // source. (This case doesn't occur on any common architecture.)
3942 // 2. If the argument is byval, RV is not sufficiently aligned, and
3943 // we cannot force it to be sufficiently aligned.
3944 // 3. If the argument is byval, but RV is not located in default
3945 // or alloca address space.
3946 Address Addr = I->hasLValue()
3947 ? I->getKnownLValue().getAddress()
3948 : I->getKnownRValue().getAggregateAddress();
3949 llvm::Value *V = Addr.getPointer();
3950 CharUnits Align = ArgInfo.getIndirectAlign();
3951 const llvm::DataLayout *TD = &CGM.getDataLayout();
3953 assert((FirstIRArg >= IRFuncTy->getNumParams() ||
3954 IRFuncTy->getParamType(FirstIRArg)->getPointerAddressSpace() ==
3955 TD->getAllocaAddrSpace()) &&
3956 "indirect argument must be in alloca address space");
3958 bool NeedCopy = false;
3960 if (Addr.getAlignment() < Align &&
3961 llvm::getOrEnforceKnownAlignment(V, Align.getQuantity(), *TD) <
3962 Align.getQuantity()) {
3964 } else if (I->hasLValue()) {
3965 auto LV = I->getKnownLValue();
3966 auto AS = LV.getAddressSpace();
3968 if ((!ArgInfo.getIndirectByVal() &&
3969 (LV.getAlignment() >=
3970 getContext().getTypeAlignInChars(I->Ty)))) {
3973 if (!getLangOpts().OpenCL) {
3974 if ((ArgInfo.getIndirectByVal() &&
3975 (AS != LangAS::Default &&
3976 AS != CGM.getASTAllocaAddressSpace()))) {
3980 // For OpenCL even if RV is located in default or alloca address space
3981 // we don't want to perform address space cast for it.
3982 else if ((ArgInfo.getIndirectByVal() &&
3983 Addr.getType()->getAddressSpace() != IRFuncTy->
3984 getParamType(FirstIRArg)->getPointerAddressSpace())) {
3990 // Create an aligned temporary, and copy to it.
3991 Address AI = CreateMemTempWithoutCast(
3992 I->Ty, ArgInfo.getIndirectAlign(), "byval-temp");
3993 IRCallArgs[FirstIRArg] = AI.getPointer();
3994 I->copyInto(*this, AI);
3996 // Skip the extra memcpy call.
3997 auto *T = V->getType()->getPointerElementType()->getPointerTo(
3998 CGM.getDataLayout().getAllocaAddrSpace());
3999 IRCallArgs[FirstIRArg] = getTargetHooks().performAddrSpaceCast(
4000 *this, V, LangAS::Default, CGM.getASTAllocaAddressSpace(), T,
4007 case ABIArgInfo::Ignore:
4008 assert(NumIRArgs == 0);
4011 case ABIArgInfo::Extend:
4012 case ABIArgInfo::Direct: {
4013 if (!isa<llvm::StructType>(ArgInfo.getCoerceToType()) &&
4014 ArgInfo.getCoerceToType() == ConvertType(info_it->type) &&
4015 ArgInfo.getDirectOffset() == 0) {
4016 assert(NumIRArgs == 1);
4018 if (!I->isAggregate())
4019 V = I->getKnownRValue().getScalarVal();
4021 V = Builder.CreateLoad(
4022 I->hasLValue() ? I->getKnownLValue().getAddress()
4023 : I->getKnownRValue().getAggregateAddress());
4025 // Implement swifterror by copying into a new swifterror argument.
4026 // We'll write back in the normal path out of the call.
4027 if (CallInfo.getExtParameterInfo(ArgNo).getABI()
4028 == ParameterABI::SwiftErrorResult) {
4029 assert(!swiftErrorTemp.isValid() && "multiple swifterror args");
4031 QualType pointeeTy = I->Ty->getPointeeType();
4033 Address(V, getContext().getTypeAlignInChars(pointeeTy));
4036 CreateMemTemp(pointeeTy, getPointerAlign(), "swifterror.temp");
4037 V = swiftErrorTemp.getPointer();
4038 cast<llvm::AllocaInst>(V)->setSwiftError(true);
4040 llvm::Value *errorValue = Builder.CreateLoad(swiftErrorArg);
4041 Builder.CreateStore(errorValue, swiftErrorTemp);
4044 // We might have to widen integers, but we should never truncate.
4045 if (ArgInfo.getCoerceToType() != V->getType() &&
4046 V->getType()->isIntegerTy())
4047 V = Builder.CreateZExt(V, ArgInfo.getCoerceToType());
4049 // If the argument doesn't match, perform a bitcast to coerce it. This
4050 // can happen due to trivial type mismatches.
4051 if (FirstIRArg < IRFuncTy->getNumParams() &&
4052 V->getType() != IRFuncTy->getParamType(FirstIRArg))
4053 V = Builder.CreateBitCast(V, IRFuncTy->getParamType(FirstIRArg));
4055 IRCallArgs[FirstIRArg] = V;
4059 // FIXME: Avoid the conversion through memory if possible.
4060 Address Src = Address::invalid();
4061 if (!I->isAggregate()) {
4062 Src = CreateMemTemp(I->Ty, "coerce");
4063 I->copyInto(*this, Src);
4065 Src = I->hasLValue() ? I->getKnownLValue().getAddress()
4066 : I->getKnownRValue().getAggregateAddress();
4069 // If the value is offset in memory, apply the offset now.
4070 Src = emitAddressAtOffset(*this, Src, ArgInfo);
4072 // Fast-isel and the optimizer generally like scalar values better than
4073 // FCAs, so we flatten them if this is safe to do for this argument.
4074 llvm::StructType *STy =
4075 dyn_cast<llvm::StructType>(ArgInfo.getCoerceToType());
4076 if (STy && ArgInfo.isDirect() && ArgInfo.getCanBeFlattened()) {
4077 llvm::Type *SrcTy = Src.getType()->getElementType();
4078 uint64_t SrcSize = CGM.getDataLayout().getTypeAllocSize(SrcTy);
4079 uint64_t DstSize = CGM.getDataLayout().getTypeAllocSize(STy);
4081 // If the source type is smaller than the destination type of the
4082 // coerce-to logic, copy the source value into a temp alloca the size
4083 // of the destination type to allow loading all of it. The bits past
4084 // the source value are left undef.
4085 if (SrcSize < DstSize) {
4087 = CreateTempAlloca(STy, Src.getAlignment(),
4088 Src.getName() + ".coerce");
4089 Builder.CreateMemCpy(TempAlloca, Src, SrcSize);
4092 Src = Builder.CreateBitCast(Src,
4093 STy->getPointerTo(Src.getAddressSpace()));
4096 assert(NumIRArgs == STy->getNumElements());
4097 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
4098 Address EltPtr = Builder.CreateStructGEP(Src, i);
4099 llvm::Value *LI = Builder.CreateLoad(EltPtr);
4100 IRCallArgs[FirstIRArg + i] = LI;
4103 // In the simple case, just pass the coerced loaded value.
4104 assert(NumIRArgs == 1);
4105 IRCallArgs[FirstIRArg] =
4106 CreateCoercedLoad(Src, ArgInfo.getCoerceToType(), *this);
4112 case ABIArgInfo::CoerceAndExpand: {
4113 auto coercionType = ArgInfo.getCoerceAndExpandType();
4114 auto layout = CGM.getDataLayout().getStructLayout(coercionType);
4116 llvm::Value *tempSize = nullptr;
4117 Address addr = Address::invalid();
4118 Address AllocaAddr = Address::invalid();
4119 if (I->isAggregate()) {
4120 addr = I->hasLValue() ? I->getKnownLValue().getAddress()
4121 : I->getKnownRValue().getAggregateAddress();
4124 RValue RV = I->getKnownRValue();
4125 assert(RV.isScalar()); // complex should always just be direct
4127 llvm::Type *scalarType = RV.getScalarVal()->getType();
4128 auto scalarSize = CGM.getDataLayout().getTypeAllocSize(scalarType);
4129 auto scalarAlign = CGM.getDataLayout().getPrefTypeAlignment(scalarType);
4131 // Materialize to a temporary.
4132 addr = CreateTempAlloca(RV.getScalarVal()->getType(),
4133 CharUnits::fromQuantity(std::max(
4134 layout->getAlignment(), scalarAlign)),
4136 /*ArraySize=*/nullptr, &AllocaAddr);
4137 tempSize = EmitLifetimeStart(scalarSize, AllocaAddr.getPointer());
4139 Builder.CreateStore(RV.getScalarVal(), addr);
4142 addr = Builder.CreateElementBitCast(addr, coercionType);
4144 unsigned IRArgPos = FirstIRArg;
4145 for (unsigned i = 0, e = coercionType->getNumElements(); i != e; ++i) {
4146 llvm::Type *eltType = coercionType->getElementType(i);
4147 if (ABIArgInfo::isPaddingForCoerceAndExpand(eltType)) continue;
4148 Address eltAddr = Builder.CreateStructGEP(addr, i);
4149 llvm::Value *elt = Builder.CreateLoad(eltAddr);
4150 IRCallArgs[IRArgPos++] = elt;
4152 assert(IRArgPos == FirstIRArg + NumIRArgs);
4155 EmitLifetimeEnd(tempSize, AllocaAddr.getPointer());
4161 case ABIArgInfo::Expand:
4162 unsigned IRArgPos = FirstIRArg;
4163 ExpandTypeToArgs(I->Ty, *I, IRFuncTy, IRCallArgs, IRArgPos);
4164 assert(IRArgPos == FirstIRArg + NumIRArgs);
4169 const CGCallee &ConcreteCallee = Callee.prepareConcreteCallee(*this);
4170 llvm::Value *CalleePtr = ConcreteCallee.getFunctionPointer();
4172 // If we're using inalloca, set up that argument.
4173 if (ArgMemory.isValid()) {
4174 llvm::Value *Arg = ArgMemory.getPointer();
4175 if (CallInfo.isVariadic()) {
4176 // When passing non-POD arguments by value to variadic functions, we will
4177 // end up with a variadic prototype and an inalloca call site. In such
4178 // cases, we can't do any parameter mismatch checks. Give up and bitcast
4180 unsigned CalleeAS = CalleePtr->getType()->getPointerAddressSpace();
4182 Builder.CreateBitCast(CalleePtr, IRFuncTy->getPointerTo(CalleeAS));
4184 llvm::Type *LastParamTy =
4185 IRFuncTy->getParamType(IRFuncTy->getNumParams() - 1);
4186 if (Arg->getType() != LastParamTy) {
4188 // Assert that these structs have equivalent element types.
4189 llvm::StructType *FullTy = CallInfo.getArgStruct();
4190 llvm::StructType *DeclaredTy = cast<llvm::StructType>(
4191 cast<llvm::PointerType>(LastParamTy)->getElementType());
4192 assert(DeclaredTy->getNumElements() == FullTy->getNumElements());
4193 for (llvm::StructType::element_iterator DI = DeclaredTy->element_begin(),
4194 DE = DeclaredTy->element_end(),
4195 FI = FullTy->element_begin();
4196 DI != DE; ++DI, ++FI)
4199 Arg = Builder.CreateBitCast(Arg, LastParamTy);
4202 assert(IRFunctionArgs.hasInallocaArg());
4203 IRCallArgs[IRFunctionArgs.getInallocaArgNo()] = Arg;
4206 // 2. Prepare the function pointer.
4208 // If the callee is a bitcast of a non-variadic function to have a
4209 // variadic function pointer type, check to see if we can remove the
4210 // bitcast. This comes up with unprototyped functions.
4212 // This makes the IR nicer, but more importantly it ensures that we
4213 // can inline the function at -O0 if it is marked always_inline.
4214 auto simplifyVariadicCallee = [](llvm::FunctionType *CalleeFT,
4215 llvm::Value *Ptr) -> llvm::Function * {
4216 if (!CalleeFT->isVarArg())
4219 // Get underlying value if it's a bitcast
4220 if (llvm::ConstantExpr *CE = dyn_cast<llvm::ConstantExpr>(Ptr)) {
4221 if (CE->getOpcode() == llvm::Instruction::BitCast)
4222 Ptr = CE->getOperand(0);
4225 llvm::Function *OrigFn = dyn_cast<llvm::Function>(Ptr);
4229 llvm::FunctionType *OrigFT = OrigFn->getFunctionType();
4231 // If the original type is variadic, or if any of the component types
4232 // disagree, we cannot remove the cast.
4233 if (OrigFT->isVarArg() ||
4234 OrigFT->getNumParams() != CalleeFT->getNumParams() ||
4235 OrigFT->getReturnType() != CalleeFT->getReturnType())
4238 for (unsigned i = 0, e = OrigFT->getNumParams(); i != e; ++i)
4239 if (OrigFT->getParamType(i) != CalleeFT->getParamType(i))
4245 if (llvm::Function *OrigFn = simplifyVariadicCallee(IRFuncTy, CalleePtr)) {
4247 IRFuncTy = OrigFn->getFunctionType();
4250 // 3. Perform the actual call.
4252 // Deactivate any cleanups that we're supposed to do immediately before
4254 if (!CallArgs.getCleanupsToDeactivate().empty())
4255 deactivateArgCleanupsBeforeCall(*this, CallArgs);
4257 // Assert that the arguments we computed match up. The IR verifier
4258 // will catch this, but this is a common enough source of problems
4259 // during IRGen changes that it's way better for debugging to catch
4260 // it ourselves here.
4262 assert(IRCallArgs.size() == IRFuncTy->getNumParams() || IRFuncTy->isVarArg());
4263 for (unsigned i = 0; i < IRCallArgs.size(); ++i) {
4264 // Inalloca argument can have different type.
4265 if (IRFunctionArgs.hasInallocaArg() &&
4266 i == IRFunctionArgs.getInallocaArgNo())
4268 if (i < IRFuncTy->getNumParams())
4269 assert(IRCallArgs[i]->getType() == IRFuncTy->getParamType(i));
4273 // Update the largest vector width if any arguments have vector types.
4274 for (unsigned i = 0; i < IRCallArgs.size(); ++i) {
4275 if (auto *VT = dyn_cast<llvm::VectorType>(IRCallArgs[i]->getType()))
4276 LargestVectorWidth = std::max(LargestVectorWidth,
4277 VT->getPrimitiveSizeInBits());
4280 // Compute the calling convention and attributes.
4281 unsigned CallingConv;
4282 llvm::AttributeList Attrs;
4283 CGM.ConstructAttributeList(CalleePtr->getName(), CallInfo,
4284 Callee.getAbstractInfo(), Attrs, CallingConv,
4285 /*AttrOnCallSite=*/true);
4287 // Apply some call-site-specific attributes.
4288 // TODO: work this into building the attribute set.
4290 // Apply always_inline to all calls within flatten functions.
4291 // FIXME: should this really take priority over __try, below?
4292 if (CurCodeDecl && CurCodeDecl->hasAttr<FlattenAttr>() &&
4293 !(TargetDecl && TargetDecl->hasAttr<NoInlineAttr>())) {
4295 Attrs.addAttribute(getLLVMContext(), llvm::AttributeList::FunctionIndex,
4296 llvm::Attribute::AlwaysInline);
4299 // Disable inlining inside SEH __try blocks.
4300 if (isSEHTryScope()) {
4302 Attrs.addAttribute(getLLVMContext(), llvm::AttributeList::FunctionIndex,
4303 llvm::Attribute::NoInline);
4306 // Decide whether to use a call or an invoke.
4308 if (currentFunctionUsesSEHTry()) {
4309 // SEH cares about asynchronous exceptions, so everything can "throw."
4310 CannotThrow = false;
4311 } else if (isCleanupPadScope() &&
4312 EHPersonality::get(*this).isMSVCXXPersonality()) {
4313 // The MSVC++ personality will implicitly terminate the program if an
4314 // exception is thrown during a cleanup outside of a try/catch.
4315 // We don't need to model anything in IR to get this behavior.
4318 // Otherwise, nounwind call sites will never throw.
4319 CannotThrow = Attrs.hasAttribute(llvm::AttributeList::FunctionIndex,
4320 llvm::Attribute::NoUnwind);
4323 // If we made a temporary, be sure to clean up after ourselves. Note that we
4324 // can't depend on being inside of an ExprWithCleanups, so we need to manually
4325 // pop this cleanup later on. Being eager about this is OK, since this
4326 // temporary is 'invisible' outside of the callee.
4327 if (UnusedReturnSizePtr)
4328 pushFullExprCleanup<CallLifetimeEnd>(NormalEHLifetimeMarker, SRetAlloca,
4329 UnusedReturnSizePtr);
4331 llvm::BasicBlock *InvokeDest = CannotThrow ? nullptr : getInvokeDest();
4333 SmallVector<llvm::OperandBundleDef, 1> BundleList =
4334 getBundlesForFunclet(CalleePtr);
4336 // Emit the actual call/invoke instruction.
4339 CI = Builder.CreateCall(IRFuncTy, CalleePtr, IRCallArgs, BundleList);
4341 llvm::BasicBlock *Cont = createBasicBlock("invoke.cont");
4342 CI = Builder.CreateInvoke(IRFuncTy, CalleePtr, Cont, InvokeDest, IRCallArgs,
4349 // Apply the attributes and calling convention.
4350 CI->setAttributes(Attrs);
4351 CI->setCallingConv(static_cast<llvm::CallingConv::ID>(CallingConv));
4353 // Apply various metadata.
4355 if (!CI->getType()->isVoidTy())
4356 CI->setName("call");
4358 // Update largest vector width from the return type.
4359 if (auto *VT = dyn_cast<llvm::VectorType>(CI->getType()))
4360 LargestVectorWidth = std::max(LargestVectorWidth,
4361 VT->getPrimitiveSizeInBits());
4363 // Insert instrumentation or attach profile metadata at indirect call sites.
4364 // For more details, see the comment before the definition of
4365 // IPVK_IndirectCallTarget in InstrProfData.inc.
4366 if (!CI->getCalledFunction())
4367 PGO.valueProfile(Builder, llvm::IPVK_IndirectCallTarget,
4370 // In ObjC ARC mode with no ObjC ARC exception safety, tell the ARC
4371 // optimizer it can aggressively ignore unwind edges.
4372 if (CGM.getLangOpts().ObjCAutoRefCount)
4373 AddObjCARCExceptionMetadata(CI);
4375 // Suppress tail calls if requested.
4376 if (llvm::CallInst *Call = dyn_cast<llvm::CallInst>(CI)) {
4377 if (TargetDecl && TargetDecl->hasAttr<NotTailCalledAttr>())
4378 Call->setTailCallKind(llvm::CallInst::TCK_NoTail);
4381 // Add metadata for calls to MSAllocator functions
4382 if (getDebugInfo() && TargetDecl &&
4383 TargetDecl->hasAttr<MSAllocatorAttr>())
4384 getDebugInfo()->addHeapAllocSiteMetadata(CI, RetTy, Loc);
4386 // 4. Finish the call.
4388 // If the call doesn't return, finish the basic block and clear the
4389 // insertion point; this allows the rest of IRGen to discard
4390 // unreachable code.
4391 if (CI->doesNotReturn()) {
4392 if (UnusedReturnSizePtr)
4395 // Strip away the noreturn attribute to better diagnose unreachable UB.
4396 if (SanOpts.has(SanitizerKind::Unreachable)) {
4397 // Also remove from function since CallBase::hasFnAttr additionally checks
4398 // attributes of the called function.
4399 if (auto *F = CI->getCalledFunction())
4400 F->removeFnAttr(llvm::Attribute::NoReturn);
4401 CI->removeAttribute(llvm::AttributeList::FunctionIndex,
4402 llvm::Attribute::NoReturn);
4404 // Avoid incompatibility with ASan which relies on the `noreturn`
4405 // attribute to insert handler calls.
4406 if (SanOpts.hasOneOf(SanitizerKind::Address |
4407 SanitizerKind::KernelAddress)) {
4408 SanitizerScope SanScope(this);
4409 llvm::IRBuilder<>::InsertPointGuard IPGuard(Builder);
4410 Builder.SetInsertPoint(CI);
4411 auto *FnType = llvm::FunctionType::get(CGM.VoidTy, /*isVarArg=*/false);
4412 llvm::FunctionCallee Fn =
4413 CGM.CreateRuntimeFunction(FnType, "__asan_handle_no_return");
4414 EmitNounwindRuntimeCall(Fn);
4418 EmitUnreachable(Loc);
4419 Builder.ClearInsertionPoint();
4421 // FIXME: For now, emit a dummy basic block because expr emitters in
4422 // generally are not ready to handle emitting expressions at unreachable
4424 EnsureInsertPoint();
4426 // Return a reasonable RValue.
4427 return GetUndefRValue(RetTy);
4430 // Perform the swifterror writeback.
4431 if (swiftErrorTemp.isValid()) {
4432 llvm::Value *errorResult = Builder.CreateLoad(swiftErrorTemp);
4433 Builder.CreateStore(errorResult, swiftErrorArg);
4436 // Emit any call-associated writebacks immediately. Arguably this
4437 // should happen after any return-value munging.
4438 if (CallArgs.hasWritebacks())
4439 emitWritebacks(*this, CallArgs);
4441 // The stack cleanup for inalloca arguments has to run out of the normal
4442 // lexical order, so deactivate it and run it manually here.
4443 CallArgs.freeArgumentMemory(*this);
4445 // Extract the return value.
4447 switch (RetAI.getKind()) {
4448 case ABIArgInfo::CoerceAndExpand: {
4449 auto coercionType = RetAI.getCoerceAndExpandType();
4451 Address addr = SRetPtr;
4452 addr = Builder.CreateElementBitCast(addr, coercionType);
4454 assert(CI->getType() == RetAI.getUnpaddedCoerceAndExpandType());
4455 bool requiresExtract = isa<llvm::StructType>(CI->getType());
4457 unsigned unpaddedIndex = 0;
4458 for (unsigned i = 0, e = coercionType->getNumElements(); i != e; ++i) {
4459 llvm::Type *eltType = coercionType->getElementType(i);
4460 if (ABIArgInfo::isPaddingForCoerceAndExpand(eltType)) continue;
4461 Address eltAddr = Builder.CreateStructGEP(addr, i);
4462 llvm::Value *elt = CI;
4463 if (requiresExtract)
4464 elt = Builder.CreateExtractValue(elt, unpaddedIndex++);
4466 assert(unpaddedIndex == 0);
4467 Builder.CreateStore(elt, eltAddr);
4473 case ABIArgInfo::InAlloca:
4474 case ABIArgInfo::Indirect: {
4475 RValue ret = convertTempToRValue(SRetPtr, RetTy, SourceLocation());
4476 if (UnusedReturnSizePtr)
4481 case ABIArgInfo::Ignore:
4482 // If we are ignoring an argument that had a result, make sure to
4483 // construct the appropriate return value for our caller.
4484 return GetUndefRValue(RetTy);
4486 case ABIArgInfo::Extend:
4487 case ABIArgInfo::Direct: {
4488 llvm::Type *RetIRTy = ConvertType(RetTy);
4489 if (RetAI.getCoerceToType() == RetIRTy && RetAI.getDirectOffset() == 0) {
4490 switch (getEvaluationKind(RetTy)) {
4492 llvm::Value *Real = Builder.CreateExtractValue(CI, 0);
4493 llvm::Value *Imag = Builder.CreateExtractValue(CI, 1);
4494 return RValue::getComplex(std::make_pair(Real, Imag));
4496 case TEK_Aggregate: {
4497 Address DestPtr = ReturnValue.getValue();
4498 bool DestIsVolatile = ReturnValue.isVolatile();
4500 if (!DestPtr.isValid()) {
4501 DestPtr = CreateMemTemp(RetTy, "agg.tmp");
4502 DestIsVolatile = false;
4504 BuildAggStore(*this, CI, DestPtr, DestIsVolatile);
4505 return RValue::getAggregate(DestPtr);
4508 // If the argument doesn't match, perform a bitcast to coerce it. This
4509 // can happen due to trivial type mismatches.
4510 llvm::Value *V = CI;
4511 if (V->getType() != RetIRTy)
4512 V = Builder.CreateBitCast(V, RetIRTy);
4513 return RValue::get(V);
4516 llvm_unreachable("bad evaluation kind");
4519 Address DestPtr = ReturnValue.getValue();
4520 bool DestIsVolatile = ReturnValue.isVolatile();
4522 if (!DestPtr.isValid()) {
4523 DestPtr = CreateMemTemp(RetTy, "coerce");
4524 DestIsVolatile = false;
4527 // If the value is offset in memory, apply the offset now.
4528 Address StorePtr = emitAddressAtOffset(*this, DestPtr, RetAI);
4529 CreateCoercedStore(CI, StorePtr, DestIsVolatile, *this);
4531 return convertTempToRValue(DestPtr, RetTy, SourceLocation());
4534 case ABIArgInfo::Expand:
4535 llvm_unreachable("Invalid ABI kind for return argument");
4538 llvm_unreachable("Unhandled ABIArgInfo::Kind");
4541 // Emit the assume_aligned check on the return value.
4542 if (Ret.isScalar() && TargetDecl) {
4543 if (const auto *AA = TargetDecl->getAttr<AssumeAlignedAttr>()) {
4544 llvm::Value *OffsetValue = nullptr;
4545 if (const auto *Offset = AA->getOffset())
4546 OffsetValue = EmitScalarExpr(Offset);
4548 llvm::Value *Alignment = EmitScalarExpr(AA->getAlignment());
4549 llvm::ConstantInt *AlignmentCI = cast<llvm::ConstantInt>(Alignment);
4550 EmitAlignmentAssumption(Ret.getScalarVal(), RetTy, Loc, AA->getLocation(),
4551 AlignmentCI->getZExtValue(), OffsetValue);
4552 } else if (const auto *AA = TargetDecl->getAttr<AllocAlignAttr>()) {
4553 llvm::Value *AlignmentVal = CallArgs[AA->getParamIndex().getLLVMIndex()]
4556 EmitAlignmentAssumption(Ret.getScalarVal(), RetTy, Loc, AA->getLocation(),
4564 CGCallee CGCallee::prepareConcreteCallee(CodeGenFunction &CGF) const {
4566 const CallExpr *CE = getVirtualCallExpr();
4567 return CGF.CGM.getCXXABI().getVirtualFunctionPointer(
4568 CGF, getVirtualMethodDecl(), getThisAddress(), getVirtualFunctionType(),
4569 CE ? CE->getBeginLoc() : SourceLocation());
4575 /* VarArg handling */
4577 Address CodeGenFunction::EmitVAArg(VAArgExpr *VE, Address &VAListAddr) {
4578 VAListAddr = VE->isMicrosoftABI()
4579 ? EmitMSVAListRef(VE->getSubExpr())
4580 : EmitVAListRef(VE->getSubExpr());
4581 QualType Ty = VE->getType();
4582 if (VE->isMicrosoftABI())
4583 return CGM.getTypes().getABIInfo().EmitMSVAArg(*this, VAListAddr, Ty);
4584 return CGM.getTypes().getABIInfo().EmitVAArg(*this, VAListAddr, Ty);