1 //===--- CGCall.cpp - Encapsulate calling convention details --------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // These classes wrap the information about a call or function
11 // definition used to handle ABI compliancy.
13 //===----------------------------------------------------------------------===//
19 #include "CGCleanup.h"
20 #include "CodeGenFunction.h"
21 #include "CodeGenModule.h"
22 #include "TargetInfo.h"
23 #include "clang/AST/Decl.h"
24 #include "clang/AST/DeclCXX.h"
25 #include "clang/AST/DeclObjC.h"
26 #include "clang/Basic/TargetBuiltins.h"
27 #include "clang/Basic/TargetInfo.h"
28 #include "clang/CodeGen/CGFunctionInfo.h"
29 #include "clang/CodeGen/SwiftCallingConv.h"
30 #include "clang/Frontend/CodeGenOptions.h"
31 #include "llvm/ADT/StringExtras.h"
32 #include "llvm/Analysis/ValueTracking.h"
33 #include "llvm/IR/Attributes.h"
34 #include "llvm/IR/CallingConv.h"
35 #include "llvm/IR/CallSite.h"
36 #include "llvm/IR/DataLayout.h"
37 #include "llvm/IR/InlineAsm.h"
38 #include "llvm/IR/Intrinsics.h"
39 #include "llvm/IR/IntrinsicInst.h"
40 #include "llvm/Transforms/Utils/Local.h"
41 using namespace clang;
42 using namespace CodeGen;
46 unsigned CodeGenTypes::ClangCallConvToLLVMCallConv(CallingConv CC) {
48 default: return llvm::CallingConv::C;
49 case CC_X86StdCall: return llvm::CallingConv::X86_StdCall;
50 case CC_X86FastCall: return llvm::CallingConv::X86_FastCall;
51 case CC_X86RegCall: return llvm::CallingConv::X86_RegCall;
52 case CC_X86ThisCall: return llvm::CallingConv::X86_ThisCall;
53 case CC_X86_64Win64: return llvm::CallingConv::X86_64_Win64;
54 case CC_X86_64SysV: return llvm::CallingConv::X86_64_SysV;
55 case CC_AAPCS: return llvm::CallingConv::ARM_AAPCS;
56 case CC_AAPCS_VFP: return llvm::CallingConv::ARM_AAPCS_VFP;
57 case CC_IntelOclBicc: return llvm::CallingConv::Intel_OCL_BI;
58 // TODO: Add support for __pascal to LLVM.
59 case CC_X86Pascal: return llvm::CallingConv::C;
60 // TODO: Add support for __vectorcall to LLVM.
61 case CC_X86VectorCall: return llvm::CallingConv::X86_VectorCall;
62 case CC_SpirFunction: return llvm::CallingConv::SPIR_FUNC;
63 case CC_OpenCLKernel: return CGM.getTargetCodeGenInfo().getOpenCLKernelCallingConv();
64 case CC_PreserveMost: return llvm::CallingConv::PreserveMost;
65 case CC_PreserveAll: return llvm::CallingConv::PreserveAll;
66 case CC_Swift: return llvm::CallingConv::Swift;
70 /// Derives the 'this' type for codegen purposes, i.e. ignoring method
72 /// FIXME: address space qualification?
73 static CanQualType GetThisType(ASTContext &Context, const CXXRecordDecl *RD) {
74 QualType RecTy = Context.getTagDeclType(RD)->getCanonicalTypeInternal();
75 return Context.getPointerType(CanQualType::CreateUnsafe(RecTy));
78 /// Returns the canonical formal type of the given C++ method.
79 static CanQual<FunctionProtoType> GetFormalType(const CXXMethodDecl *MD) {
80 return MD->getType()->getCanonicalTypeUnqualified()
81 .getAs<FunctionProtoType>();
84 /// Returns the "extra-canonicalized" return type, which discards
85 /// qualifiers on the return type. Codegen doesn't care about them,
86 /// and it makes ABI code a little easier to be able to assume that
87 /// all parameter and return types are top-level unqualified.
88 static CanQualType GetReturnType(QualType RetTy) {
89 return RetTy->getCanonicalTypeUnqualified().getUnqualifiedType();
92 /// Arrange the argument and result information for a value of the given
93 /// unprototyped freestanding function type.
94 const CGFunctionInfo &
95 CodeGenTypes::arrangeFreeFunctionType(CanQual<FunctionNoProtoType> FTNP) {
96 // When translating an unprototyped function type, always use a
98 return arrangeLLVMFunctionInfo(FTNP->getReturnType().getUnqualifiedType(),
99 /*instanceMethod=*/false,
100 /*chainCall=*/false, None,
101 FTNP->getExtInfo(), {}, RequiredArgs(0));
104 static void addExtParameterInfosForCall(
105 llvm::SmallVectorImpl<FunctionProtoType::ExtParameterInfo> ¶mInfos,
106 const FunctionProtoType *proto,
108 unsigned totalArgs) {
109 assert(proto->hasExtParameterInfos());
110 assert(paramInfos.size() <= prefixArgs);
111 assert(proto->getNumParams() + prefixArgs <= totalArgs);
113 paramInfos.reserve(totalArgs);
115 // Add default infos for any prefix args that don't already have infos.
116 paramInfos.resize(prefixArgs);
118 // Add infos for the prototype.
119 for (const auto &ParamInfo : proto->getExtParameterInfos()) {
120 paramInfos.push_back(ParamInfo);
121 // pass_object_size params have no parameter info.
122 if (ParamInfo.hasPassObjectSize())
123 paramInfos.emplace_back();
126 assert(paramInfos.size() <= totalArgs &&
127 "Did we forget to insert pass_object_size args?");
128 // Add default infos for the variadic and/or suffix arguments.
129 paramInfos.resize(totalArgs);
132 /// Adds the formal paramaters in FPT to the given prefix. If any parameter in
133 /// FPT has pass_object_size attrs, then we'll add parameters for those, too.
134 static void appendParameterTypes(const CodeGenTypes &CGT,
135 SmallVectorImpl<CanQualType> &prefix,
136 SmallVectorImpl<FunctionProtoType::ExtParameterInfo> ¶mInfos,
137 CanQual<FunctionProtoType> FPT) {
138 // Fast path: don't touch param info if we don't need to.
139 if (!FPT->hasExtParameterInfos()) {
140 assert(paramInfos.empty() &&
141 "We have paramInfos, but the prototype doesn't?");
142 prefix.append(FPT->param_type_begin(), FPT->param_type_end());
146 unsigned PrefixSize = prefix.size();
147 // In the vast majority of cases, we'll have precisely FPT->getNumParams()
148 // parameters; the only thing that can change this is the presence of
149 // pass_object_size. So, we preallocate for the common case.
150 prefix.reserve(prefix.size() + FPT->getNumParams());
152 auto ExtInfos = FPT->getExtParameterInfos();
153 assert(ExtInfos.size() == FPT->getNumParams());
154 for (unsigned I = 0, E = FPT->getNumParams(); I != E; ++I) {
155 prefix.push_back(FPT->getParamType(I));
156 if (ExtInfos[I].hasPassObjectSize())
157 prefix.push_back(CGT.getContext().getSizeType());
160 addExtParameterInfosForCall(paramInfos, FPT.getTypePtr(), PrefixSize,
164 /// Arrange the LLVM function layout for a value of the given function
165 /// type, on top of any implicit parameters already stored.
166 static const CGFunctionInfo &
167 arrangeLLVMFunctionInfo(CodeGenTypes &CGT, bool instanceMethod,
168 SmallVectorImpl<CanQualType> &prefix,
169 CanQual<FunctionProtoType> FTP,
170 const FunctionDecl *FD) {
171 SmallVector<FunctionProtoType::ExtParameterInfo, 16> paramInfos;
172 RequiredArgs Required =
173 RequiredArgs::forPrototypePlus(FTP, prefix.size(), FD);
175 appendParameterTypes(CGT, prefix, paramInfos, FTP);
176 CanQualType resultType = FTP->getReturnType().getUnqualifiedType();
178 return CGT.arrangeLLVMFunctionInfo(resultType, instanceMethod,
179 /*chainCall=*/false, prefix,
180 FTP->getExtInfo(), paramInfos,
184 /// Arrange the argument and result information for a value of the
185 /// given freestanding function type.
186 const CGFunctionInfo &
187 CodeGenTypes::arrangeFreeFunctionType(CanQual<FunctionProtoType> FTP,
188 const FunctionDecl *FD) {
189 SmallVector<CanQualType, 16> argTypes;
190 return ::arrangeLLVMFunctionInfo(*this, /*instanceMethod=*/false, argTypes,
194 static CallingConv getCallingConventionForDecl(const Decl *D, bool IsWindows) {
195 // Set the appropriate calling convention for the Function.
196 if (D->hasAttr<StdCallAttr>())
197 return CC_X86StdCall;
199 if (D->hasAttr<FastCallAttr>())
200 return CC_X86FastCall;
202 if (D->hasAttr<RegCallAttr>())
203 return CC_X86RegCall;
205 if (D->hasAttr<ThisCallAttr>())
206 return CC_X86ThisCall;
208 if (D->hasAttr<VectorCallAttr>())
209 return CC_X86VectorCall;
211 if (D->hasAttr<PascalAttr>())
214 if (PcsAttr *PCS = D->getAttr<PcsAttr>())
215 return (PCS->getPCS() == PcsAttr::AAPCS ? CC_AAPCS : CC_AAPCS_VFP);
217 if (D->hasAttr<IntelOclBiccAttr>())
218 return CC_IntelOclBicc;
220 if (D->hasAttr<MSABIAttr>())
221 return IsWindows ? CC_C : CC_X86_64Win64;
223 if (D->hasAttr<SysVABIAttr>())
224 return IsWindows ? CC_X86_64SysV : CC_C;
226 if (D->hasAttr<PreserveMostAttr>())
227 return CC_PreserveMost;
229 if (D->hasAttr<PreserveAllAttr>())
230 return CC_PreserveAll;
235 /// Arrange the argument and result information for a call to an
236 /// unknown C++ non-static member function of the given abstract type.
237 /// (Zero value of RD means we don't have any meaningful "this" argument type,
238 /// so fall back to a generic pointer type).
239 /// The member function must be an ordinary function, i.e. not a
240 /// constructor or destructor.
241 const CGFunctionInfo &
242 CodeGenTypes::arrangeCXXMethodType(const CXXRecordDecl *RD,
243 const FunctionProtoType *FTP,
244 const CXXMethodDecl *MD) {
245 SmallVector<CanQualType, 16> argTypes;
247 // Add the 'this' pointer.
249 argTypes.push_back(GetThisType(Context, RD));
251 argTypes.push_back(Context.VoidPtrTy);
253 return ::arrangeLLVMFunctionInfo(
254 *this, true, argTypes,
255 FTP->getCanonicalTypeUnqualified().getAs<FunctionProtoType>(), MD);
258 /// Arrange the argument and result information for a declaration or
259 /// definition of the given C++ non-static member function. The
260 /// member function must be an ordinary function, i.e. not a
261 /// constructor or destructor.
262 const CGFunctionInfo &
263 CodeGenTypes::arrangeCXXMethodDeclaration(const CXXMethodDecl *MD) {
264 assert(!isa<CXXConstructorDecl>(MD) && "wrong method for constructors!");
265 assert(!isa<CXXDestructorDecl>(MD) && "wrong method for destructors!");
267 CanQual<FunctionProtoType> prototype = GetFormalType(MD);
269 if (MD->isInstance()) {
270 // The abstract case is perfectly fine.
271 const CXXRecordDecl *ThisType = TheCXXABI.getThisArgumentTypeForMethod(MD);
272 return arrangeCXXMethodType(ThisType, prototype.getTypePtr(), MD);
275 return arrangeFreeFunctionType(prototype, MD);
278 bool CodeGenTypes::inheritingCtorHasParams(
279 const InheritedConstructor &Inherited, CXXCtorType Type) {
280 // Parameters are unnecessary if we're constructing a base class subobject
281 // and the inherited constructor lives in a virtual base.
282 return Type == Ctor_Complete ||
283 !Inherited.getShadowDecl()->constructsVirtualBase() ||
284 !Target.getCXXABI().hasConstructorVariants();
287 const CGFunctionInfo &
288 CodeGenTypes::arrangeCXXStructorDeclaration(const CXXMethodDecl *MD,
291 SmallVector<CanQualType, 16> argTypes;
292 SmallVector<FunctionProtoType::ExtParameterInfo, 16> paramInfos;
293 argTypes.push_back(GetThisType(Context, MD->getParent()));
295 bool PassParams = true;
298 if (auto *CD = dyn_cast<CXXConstructorDecl>(MD)) {
299 GD = GlobalDecl(CD, toCXXCtorType(Type));
301 // A base class inheriting constructor doesn't get forwarded arguments
302 // needed to construct a virtual base (or base class thereof).
303 if (auto Inherited = CD->getInheritedConstructor())
304 PassParams = inheritingCtorHasParams(Inherited, toCXXCtorType(Type));
306 auto *DD = dyn_cast<CXXDestructorDecl>(MD);
307 GD = GlobalDecl(DD, toCXXDtorType(Type));
310 CanQual<FunctionProtoType> FTP = GetFormalType(MD);
312 // Add the formal parameters.
314 appendParameterTypes(*this, argTypes, paramInfos, FTP);
316 CGCXXABI::AddedStructorArgs AddedArgs =
317 TheCXXABI.buildStructorSignature(MD, Type, argTypes);
318 if (!paramInfos.empty()) {
319 // Note: prefix implies after the first param.
320 if (AddedArgs.Prefix)
321 paramInfos.insert(paramInfos.begin() + 1, AddedArgs.Prefix,
322 FunctionProtoType::ExtParameterInfo{});
323 if (AddedArgs.Suffix)
324 paramInfos.append(AddedArgs.Suffix,
325 FunctionProtoType::ExtParameterInfo{});
328 RequiredArgs required =
329 (PassParams && MD->isVariadic() ? RequiredArgs(argTypes.size())
330 : RequiredArgs::All);
332 FunctionType::ExtInfo extInfo = FTP->getExtInfo();
333 CanQualType resultType = TheCXXABI.HasThisReturn(GD)
335 : TheCXXABI.hasMostDerivedReturn(GD)
336 ? CGM.getContext().VoidPtrTy
338 return arrangeLLVMFunctionInfo(resultType, /*instanceMethod=*/true,
339 /*chainCall=*/false, argTypes, extInfo,
340 paramInfos, required);
343 static SmallVector<CanQualType, 16>
344 getArgTypesForCall(ASTContext &ctx, const CallArgList &args) {
345 SmallVector<CanQualType, 16> argTypes;
346 for (auto &arg : args)
347 argTypes.push_back(ctx.getCanonicalParamType(arg.Ty));
351 static SmallVector<CanQualType, 16>
352 getArgTypesForDeclaration(ASTContext &ctx, const FunctionArgList &args) {
353 SmallVector<CanQualType, 16> argTypes;
354 for (auto &arg : args)
355 argTypes.push_back(ctx.getCanonicalParamType(arg->getType()));
359 static llvm::SmallVector<FunctionProtoType::ExtParameterInfo, 16>
360 getExtParameterInfosForCall(const FunctionProtoType *proto,
361 unsigned prefixArgs, unsigned totalArgs) {
362 llvm::SmallVector<FunctionProtoType::ExtParameterInfo, 16> result;
363 if (proto->hasExtParameterInfos()) {
364 addExtParameterInfosForCall(result, proto, prefixArgs, totalArgs);
369 /// Arrange a call to a C++ method, passing the given arguments.
371 /// ExtraPrefixArgs is the number of ABI-specific args passed after the `this`
373 /// ExtraSuffixArgs is the number of ABI-specific args passed at the end of
375 /// PassProtoArgs indicates whether `args` has args for the parameters in the
376 /// given CXXConstructorDecl.
377 const CGFunctionInfo &
378 CodeGenTypes::arrangeCXXConstructorCall(const CallArgList &args,
379 const CXXConstructorDecl *D,
380 CXXCtorType CtorKind,
381 unsigned ExtraPrefixArgs,
382 unsigned ExtraSuffixArgs,
383 bool PassProtoArgs) {
385 SmallVector<CanQualType, 16> ArgTypes;
386 for (const auto &Arg : args)
387 ArgTypes.push_back(Context.getCanonicalParamType(Arg.Ty));
389 // +1 for implicit this, which should always be args[0].
390 unsigned TotalPrefixArgs = 1 + ExtraPrefixArgs;
392 CanQual<FunctionProtoType> FPT = GetFormalType(D);
393 RequiredArgs Required =
394 RequiredArgs::forPrototypePlus(FPT, TotalPrefixArgs + ExtraSuffixArgs, D);
395 GlobalDecl GD(D, CtorKind);
396 CanQualType ResultType = TheCXXABI.HasThisReturn(GD)
398 : TheCXXABI.hasMostDerivedReturn(GD)
399 ? CGM.getContext().VoidPtrTy
402 FunctionType::ExtInfo Info = FPT->getExtInfo();
403 llvm::SmallVector<FunctionProtoType::ExtParameterInfo, 16> ParamInfos;
404 // If the prototype args are elided, we should only have ABI-specific args,
405 // which never have param info.
406 if (PassProtoArgs && FPT->hasExtParameterInfos()) {
407 // ABI-specific suffix arguments are treated the same as variadic arguments.
408 addExtParameterInfosForCall(ParamInfos, FPT.getTypePtr(), TotalPrefixArgs,
411 return arrangeLLVMFunctionInfo(ResultType, /*instanceMethod=*/true,
412 /*chainCall=*/false, ArgTypes, Info,
413 ParamInfos, Required);
416 /// Arrange the argument and result information for the declaration or
417 /// definition of the given function.
418 const CGFunctionInfo &
419 CodeGenTypes::arrangeFunctionDeclaration(const FunctionDecl *FD) {
420 if (const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(FD))
421 if (MD->isInstance())
422 return arrangeCXXMethodDeclaration(MD);
424 CanQualType FTy = FD->getType()->getCanonicalTypeUnqualified();
426 assert(isa<FunctionType>(FTy));
428 // When declaring a function without a prototype, always use a
429 // non-variadic type.
430 if (CanQual<FunctionNoProtoType> noProto = FTy.getAs<FunctionNoProtoType>()) {
431 return arrangeLLVMFunctionInfo(
432 noProto->getReturnType(), /*instanceMethod=*/false,
433 /*chainCall=*/false, None, noProto->getExtInfo(), {},RequiredArgs::All);
436 return arrangeFreeFunctionType(FTy.castAs<FunctionProtoType>(), FD);
439 /// Arrange the argument and result information for the declaration or
440 /// definition of an Objective-C method.
441 const CGFunctionInfo &
442 CodeGenTypes::arrangeObjCMethodDeclaration(const ObjCMethodDecl *MD) {
443 // It happens that this is the same as a call with no optional
444 // arguments, except also using the formal 'self' type.
445 return arrangeObjCMessageSendSignature(MD, MD->getSelfDecl()->getType());
448 /// Arrange the argument and result information for the function type
449 /// through which to perform a send to the given Objective-C method,
450 /// using the given receiver type. The receiver type is not always
451 /// the 'self' type of the method or even an Objective-C pointer type.
452 /// This is *not* the right method for actually performing such a
453 /// message send, due to the possibility of optional arguments.
454 const CGFunctionInfo &
455 CodeGenTypes::arrangeObjCMessageSendSignature(const ObjCMethodDecl *MD,
456 QualType receiverType) {
457 SmallVector<CanQualType, 16> argTys;
458 argTys.push_back(Context.getCanonicalParamType(receiverType));
459 argTys.push_back(Context.getCanonicalParamType(Context.getObjCSelType()));
461 for (const auto *I : MD->parameters()) {
462 argTys.push_back(Context.getCanonicalParamType(I->getType()));
465 FunctionType::ExtInfo einfo;
466 bool IsWindows = getContext().getTargetInfo().getTriple().isOSWindows();
467 einfo = einfo.withCallingConv(getCallingConventionForDecl(MD, IsWindows));
469 if (getContext().getLangOpts().ObjCAutoRefCount &&
470 MD->hasAttr<NSReturnsRetainedAttr>())
471 einfo = einfo.withProducesResult(true);
473 RequiredArgs required =
474 (MD->isVariadic() ? RequiredArgs(argTys.size()) : RequiredArgs::All);
476 return arrangeLLVMFunctionInfo(
477 GetReturnType(MD->getReturnType()), /*instanceMethod=*/false,
478 /*chainCall=*/false, argTys, einfo, {}, required);
481 const CGFunctionInfo &
482 CodeGenTypes::arrangeUnprototypedObjCMessageSend(QualType returnType,
483 const CallArgList &args) {
484 auto argTypes = getArgTypesForCall(Context, args);
485 FunctionType::ExtInfo einfo;
487 return arrangeLLVMFunctionInfo(
488 GetReturnType(returnType), /*instanceMethod=*/false,
489 /*chainCall=*/false, argTypes, einfo, {}, RequiredArgs::All);
492 const CGFunctionInfo &
493 CodeGenTypes::arrangeGlobalDeclaration(GlobalDecl GD) {
494 // FIXME: Do we need to handle ObjCMethodDecl?
495 const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl());
497 if (const CXXConstructorDecl *CD = dyn_cast<CXXConstructorDecl>(FD))
498 return arrangeCXXStructorDeclaration(CD, getFromCtorType(GD.getCtorType()));
500 if (const CXXDestructorDecl *DD = dyn_cast<CXXDestructorDecl>(FD))
501 return arrangeCXXStructorDeclaration(DD, getFromDtorType(GD.getDtorType()));
503 return arrangeFunctionDeclaration(FD);
506 /// Arrange a thunk that takes 'this' as the first parameter followed by
507 /// varargs. Return a void pointer, regardless of the actual return type.
508 /// The body of the thunk will end in a musttail call to a function of the
509 /// correct type, and the caller will bitcast the function to the correct
511 const CGFunctionInfo &
512 CodeGenTypes::arrangeMSMemberPointerThunk(const CXXMethodDecl *MD) {
513 assert(MD->isVirtual() && "only virtual memptrs have thunks");
514 CanQual<FunctionProtoType> FTP = GetFormalType(MD);
515 CanQualType ArgTys[] = { GetThisType(Context, MD->getParent()) };
516 return arrangeLLVMFunctionInfo(Context.VoidTy, /*instanceMethod=*/false,
517 /*chainCall=*/false, ArgTys,
518 FTP->getExtInfo(), {}, RequiredArgs(1));
521 const CGFunctionInfo &
522 CodeGenTypes::arrangeMSCtorClosure(const CXXConstructorDecl *CD,
524 assert(CT == Ctor_CopyingClosure || CT == Ctor_DefaultClosure);
526 CanQual<FunctionProtoType> FTP = GetFormalType(CD);
527 SmallVector<CanQualType, 2> ArgTys;
528 const CXXRecordDecl *RD = CD->getParent();
529 ArgTys.push_back(GetThisType(Context, RD));
530 if (CT == Ctor_CopyingClosure)
531 ArgTys.push_back(*FTP->param_type_begin());
532 if (RD->getNumVBases() > 0)
533 ArgTys.push_back(Context.IntTy);
534 CallingConv CC = Context.getDefaultCallingConvention(
535 /*IsVariadic=*/false, /*IsCXXMethod=*/true);
536 return arrangeLLVMFunctionInfo(Context.VoidTy, /*instanceMethod=*/true,
537 /*chainCall=*/false, ArgTys,
538 FunctionType::ExtInfo(CC), {},
542 /// Arrange a call as unto a free function, except possibly with an
543 /// additional number of formal parameters considered required.
544 static const CGFunctionInfo &
545 arrangeFreeFunctionLikeCall(CodeGenTypes &CGT,
547 const CallArgList &args,
548 const FunctionType *fnType,
549 unsigned numExtraRequiredArgs,
551 assert(args.size() >= numExtraRequiredArgs);
553 llvm::SmallVector<FunctionProtoType::ExtParameterInfo, 16> paramInfos;
555 // In most cases, there are no optional arguments.
556 RequiredArgs required = RequiredArgs::All;
558 // If we have a variadic prototype, the required arguments are the
559 // extra prefix plus the arguments in the prototype.
560 if (const FunctionProtoType *proto = dyn_cast<FunctionProtoType>(fnType)) {
561 if (proto->isVariadic())
562 required = RequiredArgs(proto->getNumParams() + numExtraRequiredArgs);
564 if (proto->hasExtParameterInfos())
565 addExtParameterInfosForCall(paramInfos, proto, numExtraRequiredArgs,
568 // If we don't have a prototype at all, but we're supposed to
569 // explicitly use the variadic convention for unprototyped calls,
570 // treat all of the arguments as required but preserve the nominal
571 // possibility of variadics.
572 } else if (CGM.getTargetCodeGenInfo()
573 .isNoProtoCallVariadic(args,
574 cast<FunctionNoProtoType>(fnType))) {
575 required = RequiredArgs(args.size());
579 SmallVector<CanQualType, 16> argTypes;
580 for (const auto &arg : args)
581 argTypes.push_back(CGT.getContext().getCanonicalParamType(arg.Ty));
582 return CGT.arrangeLLVMFunctionInfo(GetReturnType(fnType->getReturnType()),
583 /*instanceMethod=*/false, chainCall,
584 argTypes, fnType->getExtInfo(), paramInfos,
588 /// Figure out the rules for calling a function with the given formal
589 /// type using the given arguments. The arguments are necessary
590 /// because the function might be unprototyped, in which case it's
591 /// target-dependent in crazy ways.
592 const CGFunctionInfo &
593 CodeGenTypes::arrangeFreeFunctionCall(const CallArgList &args,
594 const FunctionType *fnType,
596 return arrangeFreeFunctionLikeCall(*this, CGM, args, fnType,
597 chainCall ? 1 : 0, chainCall);
600 /// A block function is essentially a free function with an
601 /// extra implicit argument.
602 const CGFunctionInfo &
603 CodeGenTypes::arrangeBlockFunctionCall(const CallArgList &args,
604 const FunctionType *fnType) {
605 return arrangeFreeFunctionLikeCall(*this, CGM, args, fnType, 1,
606 /*chainCall=*/false);
609 const CGFunctionInfo &
610 CodeGenTypes::arrangeBlockFunctionDeclaration(const FunctionProtoType *proto,
611 const FunctionArgList ¶ms) {
612 auto paramInfos = getExtParameterInfosForCall(proto, 1, params.size());
613 auto argTypes = getArgTypesForDeclaration(Context, params);
615 return arrangeLLVMFunctionInfo(
616 GetReturnType(proto->getReturnType()),
617 /*instanceMethod*/ false, /*chainCall*/ false, argTypes,
618 proto->getExtInfo(), paramInfos,
619 RequiredArgs::forPrototypePlus(proto, 1, nullptr));
622 const CGFunctionInfo &
623 CodeGenTypes::arrangeBuiltinFunctionCall(QualType resultType,
624 const CallArgList &args) {
626 SmallVector<CanQualType, 16> argTypes;
627 for (const auto &Arg : args)
628 argTypes.push_back(Context.getCanonicalParamType(Arg.Ty));
629 return arrangeLLVMFunctionInfo(
630 GetReturnType(resultType), /*instanceMethod=*/false,
631 /*chainCall=*/false, argTypes, FunctionType::ExtInfo(),
632 /*paramInfos=*/ {}, RequiredArgs::All);
635 const CGFunctionInfo &
636 CodeGenTypes::arrangeBuiltinFunctionDeclaration(QualType resultType,
637 const FunctionArgList &args) {
638 auto argTypes = getArgTypesForDeclaration(Context, args);
640 return arrangeLLVMFunctionInfo(
641 GetReturnType(resultType), /*instanceMethod=*/false, /*chainCall=*/false,
642 argTypes, FunctionType::ExtInfo(), {}, RequiredArgs::All);
645 const CGFunctionInfo &
646 CodeGenTypes::arrangeBuiltinFunctionDeclaration(CanQualType resultType,
647 ArrayRef<CanQualType> argTypes) {
648 return arrangeLLVMFunctionInfo(
649 resultType, /*instanceMethod=*/false, /*chainCall=*/false,
650 argTypes, FunctionType::ExtInfo(), {}, RequiredArgs::All);
653 /// Arrange a call to a C++ method, passing the given arguments.
655 /// numPrefixArgs is the number of ABI-specific prefix arguments we have. It
656 /// does not count `this`.
657 const CGFunctionInfo &
658 CodeGenTypes::arrangeCXXMethodCall(const CallArgList &args,
659 const FunctionProtoType *proto,
660 RequiredArgs required,
661 unsigned numPrefixArgs) {
662 assert(numPrefixArgs + 1 <= args.size() &&
663 "Emitting a call with less args than the required prefix?");
664 // Add one to account for `this`. It's a bit awkward here, but we don't count
665 // `this` in similar places elsewhere.
667 getExtParameterInfosForCall(proto, numPrefixArgs + 1, args.size());
670 auto argTypes = getArgTypesForCall(Context, args);
672 FunctionType::ExtInfo info = proto->getExtInfo();
673 return arrangeLLVMFunctionInfo(
674 GetReturnType(proto->getReturnType()), /*instanceMethod=*/true,
675 /*chainCall=*/false, argTypes, info, paramInfos, required);
678 const CGFunctionInfo &CodeGenTypes::arrangeNullaryFunction() {
679 return arrangeLLVMFunctionInfo(
680 getContext().VoidTy, /*instanceMethod=*/false, /*chainCall=*/false,
681 None, FunctionType::ExtInfo(), {}, RequiredArgs::All);
684 const CGFunctionInfo &
685 CodeGenTypes::arrangeCall(const CGFunctionInfo &signature,
686 const CallArgList &args) {
687 assert(signature.arg_size() <= args.size());
688 if (signature.arg_size() == args.size())
691 SmallVector<FunctionProtoType::ExtParameterInfo, 16> paramInfos;
692 auto sigParamInfos = signature.getExtParameterInfos();
693 if (!sigParamInfos.empty()) {
694 paramInfos.append(sigParamInfos.begin(), sigParamInfos.end());
695 paramInfos.resize(args.size());
698 auto argTypes = getArgTypesForCall(Context, args);
700 assert(signature.getRequiredArgs().allowsOptionalArgs());
701 return arrangeLLVMFunctionInfo(signature.getReturnType(),
702 signature.isInstanceMethod(),
703 signature.isChainCall(),
705 signature.getExtInfo(),
707 signature.getRequiredArgs());
712 void computeSPIRKernelABIInfo(CodeGenModule &CGM, CGFunctionInfo &FI);
716 /// Arrange the argument and result information for an abstract value
717 /// of a given function type. This is the method which all of the
718 /// above functions ultimately defer to.
719 const CGFunctionInfo &
720 CodeGenTypes::arrangeLLVMFunctionInfo(CanQualType resultType,
723 ArrayRef<CanQualType> argTypes,
724 FunctionType::ExtInfo info,
725 ArrayRef<FunctionProtoType::ExtParameterInfo> paramInfos,
726 RequiredArgs required) {
727 assert(std::all_of(argTypes.begin(), argTypes.end(),
728 [](CanQualType T) { return T.isCanonicalAsParam(); }));
730 // Lookup or create unique function info.
731 llvm::FoldingSetNodeID ID;
732 CGFunctionInfo::Profile(ID, instanceMethod, chainCall, info, paramInfos,
733 required, resultType, argTypes);
735 void *insertPos = nullptr;
736 CGFunctionInfo *FI = FunctionInfos.FindNodeOrInsertPos(ID, insertPos);
740 unsigned CC = ClangCallConvToLLVMCallConv(info.getCC());
742 // Construct the function info. We co-allocate the ArgInfos.
743 FI = CGFunctionInfo::create(CC, instanceMethod, chainCall, info,
744 paramInfos, resultType, argTypes, required);
745 FunctionInfos.InsertNode(FI, insertPos);
747 bool inserted = FunctionsBeingProcessed.insert(FI).second;
749 assert(inserted && "Recursively being processed?");
751 // Compute ABI information.
752 if (CC == llvm::CallingConv::SPIR_KERNEL) {
753 // Force target independent argument handling for the host visible
755 computeSPIRKernelABIInfo(CGM, *FI);
756 } else if (info.getCC() == CC_Swift) {
757 swiftcall::computeABIInfo(CGM, *FI);
759 getABIInfo().computeInfo(*FI);
762 // Loop over all of the computed argument and return value info. If any of
763 // them are direct or extend without a specified coerce type, specify the
765 ABIArgInfo &retInfo = FI->getReturnInfo();
766 if (retInfo.canHaveCoerceToType() && retInfo.getCoerceToType() == nullptr)
767 retInfo.setCoerceToType(ConvertType(FI->getReturnType()));
769 for (auto &I : FI->arguments())
770 if (I.info.canHaveCoerceToType() && I.info.getCoerceToType() == nullptr)
771 I.info.setCoerceToType(ConvertType(I.type));
773 bool erased = FunctionsBeingProcessed.erase(FI); (void)erased;
774 assert(erased && "Not in set?");
779 CGFunctionInfo *CGFunctionInfo::create(unsigned llvmCC,
782 const FunctionType::ExtInfo &info,
783 ArrayRef<ExtParameterInfo> paramInfos,
784 CanQualType resultType,
785 ArrayRef<CanQualType> argTypes,
786 RequiredArgs required) {
787 assert(paramInfos.empty() || paramInfos.size() == argTypes.size());
790 operator new(totalSizeToAlloc<ArgInfo, ExtParameterInfo>(
791 argTypes.size() + 1, paramInfos.size()));
793 CGFunctionInfo *FI = new(buffer) CGFunctionInfo();
794 FI->CallingConvention = llvmCC;
795 FI->EffectiveCallingConvention = llvmCC;
796 FI->ASTCallingConvention = info.getCC();
797 FI->InstanceMethod = instanceMethod;
798 FI->ChainCall = chainCall;
799 FI->NoReturn = info.getNoReturn();
800 FI->ReturnsRetained = info.getProducesResult();
801 FI->NoCallerSavedRegs = info.getNoCallerSavedRegs();
802 FI->Required = required;
803 FI->HasRegParm = info.getHasRegParm();
804 FI->RegParm = info.getRegParm();
805 FI->ArgStruct = nullptr;
806 FI->ArgStructAlign = 0;
807 FI->NumArgs = argTypes.size();
808 FI->HasExtParameterInfos = !paramInfos.empty();
809 FI->getArgsBuffer()[0].type = resultType;
810 for (unsigned i = 0, e = argTypes.size(); i != e; ++i)
811 FI->getArgsBuffer()[i + 1].type = argTypes[i];
812 for (unsigned i = 0, e = paramInfos.size(); i != e; ++i)
813 FI->getExtParameterInfosBuffer()[i] = paramInfos[i];
820 // ABIArgInfo::Expand implementation.
822 // Specifies the way QualType passed as ABIArgInfo::Expand is expanded.
823 struct TypeExpansion {
824 enum TypeExpansionKind {
825 // Elements of constant arrays are expanded recursively.
827 // Record fields are expanded recursively (but if record is a union, only
828 // the field with the largest size is expanded).
830 // For complex types, real and imaginary parts are expanded recursively.
832 // All other types are not expandable.
836 const TypeExpansionKind Kind;
838 TypeExpansion(TypeExpansionKind K) : Kind(K) {}
839 virtual ~TypeExpansion() {}
842 struct ConstantArrayExpansion : TypeExpansion {
846 ConstantArrayExpansion(QualType EltTy, uint64_t NumElts)
847 : TypeExpansion(TEK_ConstantArray), EltTy(EltTy), NumElts(NumElts) {}
848 static bool classof(const TypeExpansion *TE) {
849 return TE->Kind == TEK_ConstantArray;
853 struct RecordExpansion : TypeExpansion {
854 SmallVector<const CXXBaseSpecifier *, 1> Bases;
856 SmallVector<const FieldDecl *, 1> Fields;
858 RecordExpansion(SmallVector<const CXXBaseSpecifier *, 1> &&Bases,
859 SmallVector<const FieldDecl *, 1> &&Fields)
860 : TypeExpansion(TEK_Record), Bases(std::move(Bases)),
861 Fields(std::move(Fields)) {}
862 static bool classof(const TypeExpansion *TE) {
863 return TE->Kind == TEK_Record;
867 struct ComplexExpansion : TypeExpansion {
870 ComplexExpansion(QualType EltTy) : TypeExpansion(TEK_Complex), EltTy(EltTy) {}
871 static bool classof(const TypeExpansion *TE) {
872 return TE->Kind == TEK_Complex;
876 struct NoExpansion : TypeExpansion {
877 NoExpansion() : TypeExpansion(TEK_None) {}
878 static bool classof(const TypeExpansion *TE) {
879 return TE->Kind == TEK_None;
884 static std::unique_ptr<TypeExpansion>
885 getTypeExpansion(QualType Ty, const ASTContext &Context) {
886 if (const ConstantArrayType *AT = Context.getAsConstantArrayType(Ty)) {
887 return llvm::make_unique<ConstantArrayExpansion>(
888 AT->getElementType(), AT->getSize().getZExtValue());
890 if (const RecordType *RT = Ty->getAs<RecordType>()) {
891 SmallVector<const CXXBaseSpecifier *, 1> Bases;
892 SmallVector<const FieldDecl *, 1> Fields;
893 const RecordDecl *RD = RT->getDecl();
894 assert(!RD->hasFlexibleArrayMember() &&
895 "Cannot expand structure with flexible array.");
897 // Unions can be here only in degenerative cases - all the fields are same
898 // after flattening. Thus we have to use the "largest" field.
899 const FieldDecl *LargestFD = nullptr;
900 CharUnits UnionSize = CharUnits::Zero();
902 for (const auto *FD : RD->fields()) {
903 // Skip zero length bitfields.
904 if (FD->isBitField() && FD->getBitWidthValue(Context) == 0)
906 assert(!FD->isBitField() &&
907 "Cannot expand structure with bit-field members.");
908 CharUnits FieldSize = Context.getTypeSizeInChars(FD->getType());
909 if (UnionSize < FieldSize) {
910 UnionSize = FieldSize;
915 Fields.push_back(LargestFD);
917 if (const auto *CXXRD = dyn_cast<CXXRecordDecl>(RD)) {
918 assert(!CXXRD->isDynamicClass() &&
919 "cannot expand vtable pointers in dynamic classes");
920 for (const CXXBaseSpecifier &BS : CXXRD->bases())
921 Bases.push_back(&BS);
924 for (const auto *FD : RD->fields()) {
925 // Skip zero length bitfields.
926 if (FD->isBitField() && FD->getBitWidthValue(Context) == 0)
928 assert(!FD->isBitField() &&
929 "Cannot expand structure with bit-field members.");
930 Fields.push_back(FD);
933 return llvm::make_unique<RecordExpansion>(std::move(Bases),
936 if (const ComplexType *CT = Ty->getAs<ComplexType>()) {
937 return llvm::make_unique<ComplexExpansion>(CT->getElementType());
939 return llvm::make_unique<NoExpansion>();
942 static int getExpansionSize(QualType Ty, const ASTContext &Context) {
943 auto Exp = getTypeExpansion(Ty, Context);
944 if (auto CAExp = dyn_cast<ConstantArrayExpansion>(Exp.get())) {
945 return CAExp->NumElts * getExpansionSize(CAExp->EltTy, Context);
947 if (auto RExp = dyn_cast<RecordExpansion>(Exp.get())) {
949 for (auto BS : RExp->Bases)
950 Res += getExpansionSize(BS->getType(), Context);
951 for (auto FD : RExp->Fields)
952 Res += getExpansionSize(FD->getType(), Context);
955 if (isa<ComplexExpansion>(Exp.get()))
957 assert(isa<NoExpansion>(Exp.get()));
962 CodeGenTypes::getExpandedTypes(QualType Ty,
963 SmallVectorImpl<llvm::Type *>::iterator &TI) {
964 auto Exp = getTypeExpansion(Ty, Context);
965 if (auto CAExp = dyn_cast<ConstantArrayExpansion>(Exp.get())) {
966 for (int i = 0, n = CAExp->NumElts; i < n; i++) {
967 getExpandedTypes(CAExp->EltTy, TI);
969 } else if (auto RExp = dyn_cast<RecordExpansion>(Exp.get())) {
970 for (auto BS : RExp->Bases)
971 getExpandedTypes(BS->getType(), TI);
972 for (auto FD : RExp->Fields)
973 getExpandedTypes(FD->getType(), TI);
974 } else if (auto CExp = dyn_cast<ComplexExpansion>(Exp.get())) {
975 llvm::Type *EltTy = ConvertType(CExp->EltTy);
979 assert(isa<NoExpansion>(Exp.get()));
980 *TI++ = ConvertType(Ty);
984 static void forConstantArrayExpansion(CodeGenFunction &CGF,
985 ConstantArrayExpansion *CAE,
987 llvm::function_ref<void(Address)> Fn) {
988 CharUnits EltSize = CGF.getContext().getTypeSizeInChars(CAE->EltTy);
990 BaseAddr.getAlignment().alignmentOfArrayElement(EltSize);
992 for (int i = 0, n = CAE->NumElts; i < n; i++) {
993 llvm::Value *EltAddr =
994 CGF.Builder.CreateConstGEP2_32(nullptr, BaseAddr.getPointer(), 0, i);
995 Fn(Address(EltAddr, EltAlign));
999 void CodeGenFunction::ExpandTypeFromArgs(
1000 QualType Ty, LValue LV, SmallVectorImpl<llvm::Value *>::iterator &AI) {
1001 assert(LV.isSimple() &&
1002 "Unexpected non-simple lvalue during struct expansion.");
1004 auto Exp = getTypeExpansion(Ty, getContext());
1005 if (auto CAExp = dyn_cast<ConstantArrayExpansion>(Exp.get())) {
1006 forConstantArrayExpansion(*this, CAExp, LV.getAddress(),
1007 [&](Address EltAddr) {
1008 LValue LV = MakeAddrLValue(EltAddr, CAExp->EltTy);
1009 ExpandTypeFromArgs(CAExp->EltTy, LV, AI);
1011 } else if (auto RExp = dyn_cast<RecordExpansion>(Exp.get())) {
1012 Address This = LV.getAddress();
1013 for (const CXXBaseSpecifier *BS : RExp->Bases) {
1014 // Perform a single step derived-to-base conversion.
1016 GetAddressOfBaseClass(This, Ty->getAsCXXRecordDecl(), &BS, &BS + 1,
1017 /*NullCheckValue=*/false, SourceLocation());
1018 LValue SubLV = MakeAddrLValue(Base, BS->getType());
1020 // Recurse onto bases.
1021 ExpandTypeFromArgs(BS->getType(), SubLV, AI);
1023 for (auto FD : RExp->Fields) {
1024 // FIXME: What are the right qualifiers here?
1025 LValue SubLV = EmitLValueForFieldInitialization(LV, FD);
1026 ExpandTypeFromArgs(FD->getType(), SubLV, AI);
1028 } else if (isa<ComplexExpansion>(Exp.get())) {
1029 auto realValue = *AI++;
1030 auto imagValue = *AI++;
1031 EmitStoreOfComplex(ComplexPairTy(realValue, imagValue), LV, /*init*/ true);
1033 assert(isa<NoExpansion>(Exp.get()));
1034 EmitStoreThroughLValue(RValue::get(*AI++), LV);
1038 void CodeGenFunction::ExpandTypeToArgs(
1039 QualType Ty, RValue RV, llvm::FunctionType *IRFuncTy,
1040 SmallVectorImpl<llvm::Value *> &IRCallArgs, unsigned &IRCallArgPos) {
1041 auto Exp = getTypeExpansion(Ty, getContext());
1042 if (auto CAExp = dyn_cast<ConstantArrayExpansion>(Exp.get())) {
1043 forConstantArrayExpansion(*this, CAExp, RV.getAggregateAddress(),
1044 [&](Address EltAddr) {
1046 convertTempToRValue(EltAddr, CAExp->EltTy, SourceLocation());
1047 ExpandTypeToArgs(CAExp->EltTy, EltRV, IRFuncTy, IRCallArgs, IRCallArgPos);
1049 } else if (auto RExp = dyn_cast<RecordExpansion>(Exp.get())) {
1050 Address This = RV.getAggregateAddress();
1051 for (const CXXBaseSpecifier *BS : RExp->Bases) {
1052 // Perform a single step derived-to-base conversion.
1054 GetAddressOfBaseClass(This, Ty->getAsCXXRecordDecl(), &BS, &BS + 1,
1055 /*NullCheckValue=*/false, SourceLocation());
1056 RValue BaseRV = RValue::getAggregate(Base);
1058 // Recurse onto bases.
1059 ExpandTypeToArgs(BS->getType(), BaseRV, IRFuncTy, IRCallArgs,
1063 LValue LV = MakeAddrLValue(This, Ty);
1064 for (auto FD : RExp->Fields) {
1065 RValue FldRV = EmitRValueForField(LV, FD, SourceLocation());
1066 ExpandTypeToArgs(FD->getType(), FldRV, IRFuncTy, IRCallArgs,
1069 } else if (isa<ComplexExpansion>(Exp.get())) {
1070 ComplexPairTy CV = RV.getComplexVal();
1071 IRCallArgs[IRCallArgPos++] = CV.first;
1072 IRCallArgs[IRCallArgPos++] = CV.second;
1074 assert(isa<NoExpansion>(Exp.get()));
1075 assert(RV.isScalar() &&
1076 "Unexpected non-scalar rvalue during struct expansion.");
1078 // Insert a bitcast as needed.
1079 llvm::Value *V = RV.getScalarVal();
1080 if (IRCallArgPos < IRFuncTy->getNumParams() &&
1081 V->getType() != IRFuncTy->getParamType(IRCallArgPos))
1082 V = Builder.CreateBitCast(V, IRFuncTy->getParamType(IRCallArgPos));
1084 IRCallArgs[IRCallArgPos++] = V;
1088 /// Create a temporary allocation for the purposes of coercion.
1089 static Address CreateTempAllocaForCoercion(CodeGenFunction &CGF, llvm::Type *Ty,
1090 CharUnits MinAlign) {
1091 // Don't use an alignment that's worse than what LLVM would prefer.
1092 auto PrefAlign = CGF.CGM.getDataLayout().getPrefTypeAlignment(Ty);
1093 CharUnits Align = std::max(MinAlign, CharUnits::fromQuantity(PrefAlign));
1095 return CGF.CreateTempAlloca(Ty, Align);
1098 /// EnterStructPointerForCoercedAccess - Given a struct pointer that we are
1099 /// accessing some number of bytes out of it, try to gep into the struct to get
1100 /// at its inner goodness. Dive as deep as possible without entering an element
1101 /// with an in-memory size smaller than DstSize.
1103 EnterStructPointerForCoercedAccess(Address SrcPtr,
1104 llvm::StructType *SrcSTy,
1105 uint64_t DstSize, CodeGenFunction &CGF) {
1106 // We can't dive into a zero-element struct.
1107 if (SrcSTy->getNumElements() == 0) return SrcPtr;
1109 llvm::Type *FirstElt = SrcSTy->getElementType(0);
1111 // If the first elt is at least as large as what we're looking for, or if the
1112 // first element is the same size as the whole struct, we can enter it. The
1113 // comparison must be made on the store size and not the alloca size. Using
1114 // the alloca size may overstate the size of the load.
1115 uint64_t FirstEltSize =
1116 CGF.CGM.getDataLayout().getTypeStoreSize(FirstElt);
1117 if (FirstEltSize < DstSize &&
1118 FirstEltSize < CGF.CGM.getDataLayout().getTypeStoreSize(SrcSTy))
1121 // GEP into the first element.
1122 SrcPtr = CGF.Builder.CreateStructGEP(SrcPtr, 0, CharUnits(), "coerce.dive");
1124 // If the first element is a struct, recurse.
1125 llvm::Type *SrcTy = SrcPtr.getElementType();
1126 if (llvm::StructType *SrcSTy = dyn_cast<llvm::StructType>(SrcTy))
1127 return EnterStructPointerForCoercedAccess(SrcPtr, SrcSTy, DstSize, CGF);
1132 /// CoerceIntOrPtrToIntOrPtr - Convert a value Val to the specific Ty where both
1133 /// are either integers or pointers. This does a truncation of the value if it
1134 /// is too large or a zero extension if it is too small.
1136 /// This behaves as if the value were coerced through memory, so on big-endian
1137 /// targets the high bits are preserved in a truncation, while little-endian
1138 /// targets preserve the low bits.
1139 static llvm::Value *CoerceIntOrPtrToIntOrPtr(llvm::Value *Val,
1141 CodeGenFunction &CGF) {
1142 if (Val->getType() == Ty)
1145 if (isa<llvm::PointerType>(Val->getType())) {
1146 // If this is Pointer->Pointer avoid conversion to and from int.
1147 if (isa<llvm::PointerType>(Ty))
1148 return CGF.Builder.CreateBitCast(Val, Ty, "coerce.val");
1150 // Convert the pointer to an integer so we can play with its width.
1151 Val = CGF.Builder.CreatePtrToInt(Val, CGF.IntPtrTy, "coerce.val.pi");
1154 llvm::Type *DestIntTy = Ty;
1155 if (isa<llvm::PointerType>(DestIntTy))
1156 DestIntTy = CGF.IntPtrTy;
1158 if (Val->getType() != DestIntTy) {
1159 const llvm::DataLayout &DL = CGF.CGM.getDataLayout();
1160 if (DL.isBigEndian()) {
1161 // Preserve the high bits on big-endian targets.
1162 // That is what memory coercion does.
1163 uint64_t SrcSize = DL.getTypeSizeInBits(Val->getType());
1164 uint64_t DstSize = DL.getTypeSizeInBits(DestIntTy);
1166 if (SrcSize > DstSize) {
1167 Val = CGF.Builder.CreateLShr(Val, SrcSize - DstSize, "coerce.highbits");
1168 Val = CGF.Builder.CreateTrunc(Val, DestIntTy, "coerce.val.ii");
1170 Val = CGF.Builder.CreateZExt(Val, DestIntTy, "coerce.val.ii");
1171 Val = CGF.Builder.CreateShl(Val, DstSize - SrcSize, "coerce.highbits");
1174 // Little-endian targets preserve the low bits. No shifts required.
1175 Val = CGF.Builder.CreateIntCast(Val, DestIntTy, false, "coerce.val.ii");
1179 if (isa<llvm::PointerType>(Ty))
1180 Val = CGF.Builder.CreateIntToPtr(Val, Ty, "coerce.val.ip");
1186 /// CreateCoercedLoad - Create a load from \arg SrcPtr interpreted as
1187 /// a pointer to an object of type \arg Ty, known to be aligned to
1188 /// \arg SrcAlign bytes.
1190 /// This safely handles the case when the src type is smaller than the
1191 /// destination type; in this situation the values of bits which not
1192 /// present in the src are undefined.
1193 static llvm::Value *CreateCoercedLoad(Address Src, llvm::Type *Ty,
1194 CodeGenFunction &CGF) {
1195 llvm::Type *SrcTy = Src.getElementType();
1197 // If SrcTy and Ty are the same, just do a load.
1199 return CGF.Builder.CreateLoad(Src);
1201 uint64_t DstSize = CGF.CGM.getDataLayout().getTypeAllocSize(Ty);
1203 if (llvm::StructType *SrcSTy = dyn_cast<llvm::StructType>(SrcTy)) {
1204 Src = EnterStructPointerForCoercedAccess(Src, SrcSTy, DstSize, CGF);
1205 SrcTy = Src.getType()->getElementType();
1208 uint64_t SrcSize = CGF.CGM.getDataLayout().getTypeAllocSize(SrcTy);
1210 // If the source and destination are integer or pointer types, just do an
1211 // extension or truncation to the desired type.
1212 if ((isa<llvm::IntegerType>(Ty) || isa<llvm::PointerType>(Ty)) &&
1213 (isa<llvm::IntegerType>(SrcTy) || isa<llvm::PointerType>(SrcTy))) {
1214 llvm::Value *Load = CGF.Builder.CreateLoad(Src);
1215 return CoerceIntOrPtrToIntOrPtr(Load, Ty, CGF);
1218 // If load is legal, just bitcast the src pointer.
1219 if (SrcSize >= DstSize) {
1220 // Generally SrcSize is never greater than DstSize, since this means we are
1221 // losing bits. However, this can happen in cases where the structure has
1222 // additional padding, for example due to a user specified alignment.
1224 // FIXME: Assert that we aren't truncating non-padding bits when have access
1225 // to that information.
1226 Src = CGF.Builder.CreateBitCast(Src, llvm::PointerType::getUnqual(Ty));
1227 return CGF.Builder.CreateLoad(Src);
1230 // Otherwise do coercion through memory. This is stupid, but simple.
1231 Address Tmp = CreateTempAllocaForCoercion(CGF, Ty, Src.getAlignment());
1232 Address Casted = CGF.Builder.CreateBitCast(Tmp, CGF.Int8PtrTy);
1233 Address SrcCasted = CGF.Builder.CreateBitCast(Src, CGF.Int8PtrTy);
1234 CGF.Builder.CreateMemCpy(Casted, SrcCasted,
1235 llvm::ConstantInt::get(CGF.IntPtrTy, SrcSize),
1237 return CGF.Builder.CreateLoad(Tmp);
1240 // Function to store a first-class aggregate into memory. We prefer to
1241 // store the elements rather than the aggregate to be more friendly to
1243 // FIXME: Do we need to recurse here?
1244 static void BuildAggStore(CodeGenFunction &CGF, llvm::Value *Val,
1245 Address Dest, bool DestIsVolatile) {
1246 // Prefer scalar stores to first-class aggregate stores.
1247 if (llvm::StructType *STy =
1248 dyn_cast<llvm::StructType>(Val->getType())) {
1249 const llvm::StructLayout *Layout =
1250 CGF.CGM.getDataLayout().getStructLayout(STy);
1252 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1253 auto EltOffset = CharUnits::fromQuantity(Layout->getElementOffset(i));
1254 Address EltPtr = CGF.Builder.CreateStructGEP(Dest, i, EltOffset);
1255 llvm::Value *Elt = CGF.Builder.CreateExtractValue(Val, i);
1256 CGF.Builder.CreateStore(Elt, EltPtr, DestIsVolatile);
1259 CGF.Builder.CreateStore(Val, Dest, DestIsVolatile);
1263 /// CreateCoercedStore - Create a store to \arg DstPtr from \arg Src,
1264 /// where the source and destination may have different types. The
1265 /// destination is known to be aligned to \arg DstAlign bytes.
1267 /// This safely handles the case when the src type is larger than the
1268 /// destination type; the upper bits of the src will be lost.
1269 static void CreateCoercedStore(llvm::Value *Src,
1272 CodeGenFunction &CGF) {
1273 llvm::Type *SrcTy = Src->getType();
1274 llvm::Type *DstTy = Dst.getType()->getElementType();
1275 if (SrcTy == DstTy) {
1276 CGF.Builder.CreateStore(Src, Dst, DstIsVolatile);
1280 uint64_t SrcSize = CGF.CGM.getDataLayout().getTypeAllocSize(SrcTy);
1282 if (llvm::StructType *DstSTy = dyn_cast<llvm::StructType>(DstTy)) {
1283 Dst = EnterStructPointerForCoercedAccess(Dst, DstSTy, SrcSize, CGF);
1284 DstTy = Dst.getType()->getElementType();
1287 // If the source and destination are integer or pointer types, just do an
1288 // extension or truncation to the desired type.
1289 if ((isa<llvm::IntegerType>(SrcTy) || isa<llvm::PointerType>(SrcTy)) &&
1290 (isa<llvm::IntegerType>(DstTy) || isa<llvm::PointerType>(DstTy))) {
1291 Src = CoerceIntOrPtrToIntOrPtr(Src, DstTy, CGF);
1292 CGF.Builder.CreateStore(Src, Dst, DstIsVolatile);
1296 uint64_t DstSize = CGF.CGM.getDataLayout().getTypeAllocSize(DstTy);
1298 // If store is legal, just bitcast the src pointer.
1299 if (SrcSize <= DstSize) {
1300 Dst = CGF.Builder.CreateBitCast(Dst, llvm::PointerType::getUnqual(SrcTy));
1301 BuildAggStore(CGF, Src, Dst, DstIsVolatile);
1303 // Otherwise do coercion through memory. This is stupid, but
1306 // Generally SrcSize is never greater than DstSize, since this means we are
1307 // losing bits. However, this can happen in cases where the structure has
1308 // additional padding, for example due to a user specified alignment.
1310 // FIXME: Assert that we aren't truncating non-padding bits when have access
1311 // to that information.
1312 Address Tmp = CreateTempAllocaForCoercion(CGF, SrcTy, Dst.getAlignment());
1313 CGF.Builder.CreateStore(Src, Tmp);
1314 Address Casted = CGF.Builder.CreateBitCast(Tmp, CGF.Int8PtrTy);
1315 Address DstCasted = CGF.Builder.CreateBitCast(Dst, CGF.Int8PtrTy);
1316 CGF.Builder.CreateMemCpy(DstCasted, Casted,
1317 llvm::ConstantInt::get(CGF.IntPtrTy, DstSize),
1322 static Address emitAddressAtOffset(CodeGenFunction &CGF, Address addr,
1323 const ABIArgInfo &info) {
1324 if (unsigned offset = info.getDirectOffset()) {
1325 addr = CGF.Builder.CreateElementBitCast(addr, CGF.Int8Ty);
1326 addr = CGF.Builder.CreateConstInBoundsByteGEP(addr,
1327 CharUnits::fromQuantity(offset));
1328 addr = CGF.Builder.CreateElementBitCast(addr, info.getCoerceToType());
1335 /// Encapsulates information about the way function arguments from
1336 /// CGFunctionInfo should be passed to actual LLVM IR function.
1337 class ClangToLLVMArgMapping {
1338 static const unsigned InvalidIndex = ~0U;
1339 unsigned InallocaArgNo;
1341 unsigned TotalIRArgs;
1343 /// Arguments of LLVM IR function corresponding to single Clang argument.
1345 unsigned PaddingArgIndex;
1346 // Argument is expanded to IR arguments at positions
1347 // [FirstArgIndex, FirstArgIndex + NumberOfArgs).
1348 unsigned FirstArgIndex;
1349 unsigned NumberOfArgs;
1352 : PaddingArgIndex(InvalidIndex), FirstArgIndex(InvalidIndex),
1356 SmallVector<IRArgs, 8> ArgInfo;
1359 ClangToLLVMArgMapping(const ASTContext &Context, const CGFunctionInfo &FI,
1360 bool OnlyRequiredArgs = false)
1361 : InallocaArgNo(InvalidIndex), SRetArgNo(InvalidIndex), TotalIRArgs(0),
1362 ArgInfo(OnlyRequiredArgs ? FI.getNumRequiredArgs() : FI.arg_size()) {
1363 construct(Context, FI, OnlyRequiredArgs);
1366 bool hasInallocaArg() const { return InallocaArgNo != InvalidIndex; }
1367 unsigned getInallocaArgNo() const {
1368 assert(hasInallocaArg());
1369 return InallocaArgNo;
1372 bool hasSRetArg() const { return SRetArgNo != InvalidIndex; }
1373 unsigned getSRetArgNo() const {
1374 assert(hasSRetArg());
1378 unsigned totalIRArgs() const { return TotalIRArgs; }
1380 bool hasPaddingArg(unsigned ArgNo) const {
1381 assert(ArgNo < ArgInfo.size());
1382 return ArgInfo[ArgNo].PaddingArgIndex != InvalidIndex;
1384 unsigned getPaddingArgNo(unsigned ArgNo) const {
1385 assert(hasPaddingArg(ArgNo));
1386 return ArgInfo[ArgNo].PaddingArgIndex;
1389 /// Returns index of first IR argument corresponding to ArgNo, and their
1391 std::pair<unsigned, unsigned> getIRArgs(unsigned ArgNo) const {
1392 assert(ArgNo < ArgInfo.size());
1393 return std::make_pair(ArgInfo[ArgNo].FirstArgIndex,
1394 ArgInfo[ArgNo].NumberOfArgs);
1398 void construct(const ASTContext &Context, const CGFunctionInfo &FI,
1399 bool OnlyRequiredArgs);
1402 void ClangToLLVMArgMapping::construct(const ASTContext &Context,
1403 const CGFunctionInfo &FI,
1404 bool OnlyRequiredArgs) {
1405 unsigned IRArgNo = 0;
1406 bool SwapThisWithSRet = false;
1407 const ABIArgInfo &RetAI = FI.getReturnInfo();
1409 if (RetAI.getKind() == ABIArgInfo::Indirect) {
1410 SwapThisWithSRet = RetAI.isSRetAfterThis();
1411 SRetArgNo = SwapThisWithSRet ? 1 : IRArgNo++;
1415 unsigned NumArgs = OnlyRequiredArgs ? FI.getNumRequiredArgs() : FI.arg_size();
1416 for (CGFunctionInfo::const_arg_iterator I = FI.arg_begin(); ArgNo < NumArgs;
1418 assert(I != FI.arg_end());
1419 QualType ArgType = I->type;
1420 const ABIArgInfo &AI = I->info;
1421 // Collect data about IR arguments corresponding to Clang argument ArgNo.
1422 auto &IRArgs = ArgInfo[ArgNo];
1424 if (AI.getPaddingType())
1425 IRArgs.PaddingArgIndex = IRArgNo++;
1427 switch (AI.getKind()) {
1428 case ABIArgInfo::Extend:
1429 case ABIArgInfo::Direct: {
1430 // FIXME: handle sseregparm someday...
1431 llvm::StructType *STy = dyn_cast<llvm::StructType>(AI.getCoerceToType());
1432 if (AI.isDirect() && AI.getCanBeFlattened() && STy) {
1433 IRArgs.NumberOfArgs = STy->getNumElements();
1435 IRArgs.NumberOfArgs = 1;
1439 case ABIArgInfo::Indirect:
1440 IRArgs.NumberOfArgs = 1;
1442 case ABIArgInfo::Ignore:
1443 case ABIArgInfo::InAlloca:
1444 // ignore and inalloca doesn't have matching LLVM parameters.
1445 IRArgs.NumberOfArgs = 0;
1447 case ABIArgInfo::CoerceAndExpand:
1448 IRArgs.NumberOfArgs = AI.getCoerceAndExpandTypeSequence().size();
1450 case ABIArgInfo::Expand:
1451 IRArgs.NumberOfArgs = getExpansionSize(ArgType, Context);
1455 if (IRArgs.NumberOfArgs > 0) {
1456 IRArgs.FirstArgIndex = IRArgNo;
1457 IRArgNo += IRArgs.NumberOfArgs;
1460 // Skip over the sret parameter when it comes second. We already handled it
1462 if (IRArgNo == 1 && SwapThisWithSRet)
1465 assert(ArgNo == ArgInfo.size());
1467 if (FI.usesInAlloca())
1468 InallocaArgNo = IRArgNo++;
1470 TotalIRArgs = IRArgNo;
1476 bool CodeGenModule::ReturnTypeUsesSRet(const CGFunctionInfo &FI) {
1477 return FI.getReturnInfo().isIndirect();
1480 bool CodeGenModule::ReturnSlotInterferesWithArgs(const CGFunctionInfo &FI) {
1481 return ReturnTypeUsesSRet(FI) &&
1482 getTargetCodeGenInfo().doesReturnSlotInterfereWithArgs();
1485 bool CodeGenModule::ReturnTypeUsesFPRet(QualType ResultType) {
1486 if (const BuiltinType *BT = ResultType->getAs<BuiltinType>()) {
1487 switch (BT->getKind()) {
1490 case BuiltinType::Float:
1491 return getTarget().useObjCFPRetForRealType(TargetInfo::Float);
1492 case BuiltinType::Double:
1493 return getTarget().useObjCFPRetForRealType(TargetInfo::Double);
1494 case BuiltinType::LongDouble:
1495 return getTarget().useObjCFPRetForRealType(TargetInfo::LongDouble);
1502 bool CodeGenModule::ReturnTypeUsesFP2Ret(QualType ResultType) {
1503 if (const ComplexType *CT = ResultType->getAs<ComplexType>()) {
1504 if (const BuiltinType *BT = CT->getElementType()->getAs<BuiltinType>()) {
1505 if (BT->getKind() == BuiltinType::LongDouble)
1506 return getTarget().useObjCFP2RetForComplexLongDouble();
1513 llvm::FunctionType *CodeGenTypes::GetFunctionType(GlobalDecl GD) {
1514 const CGFunctionInfo &FI = arrangeGlobalDeclaration(GD);
1515 return GetFunctionType(FI);
1518 llvm::FunctionType *
1519 CodeGenTypes::GetFunctionType(const CGFunctionInfo &FI) {
1521 bool Inserted = FunctionsBeingProcessed.insert(&FI).second;
1523 assert(Inserted && "Recursively being processed?");
1525 llvm::Type *resultType = nullptr;
1526 const ABIArgInfo &retAI = FI.getReturnInfo();
1527 switch (retAI.getKind()) {
1528 case ABIArgInfo::Expand:
1529 llvm_unreachable("Invalid ABI kind for return argument");
1531 case ABIArgInfo::Extend:
1532 case ABIArgInfo::Direct:
1533 resultType = retAI.getCoerceToType();
1536 case ABIArgInfo::InAlloca:
1537 if (retAI.getInAllocaSRet()) {
1538 // sret things on win32 aren't void, they return the sret pointer.
1539 QualType ret = FI.getReturnType();
1540 llvm::Type *ty = ConvertType(ret);
1541 unsigned addressSpace = Context.getTargetAddressSpace(ret);
1542 resultType = llvm::PointerType::get(ty, addressSpace);
1544 resultType = llvm::Type::getVoidTy(getLLVMContext());
1548 case ABIArgInfo::Indirect:
1549 case ABIArgInfo::Ignore:
1550 resultType = llvm::Type::getVoidTy(getLLVMContext());
1553 case ABIArgInfo::CoerceAndExpand:
1554 resultType = retAI.getUnpaddedCoerceAndExpandType();
1558 ClangToLLVMArgMapping IRFunctionArgs(getContext(), FI, true);
1559 SmallVector<llvm::Type*, 8> ArgTypes(IRFunctionArgs.totalIRArgs());
1561 // Add type for sret argument.
1562 if (IRFunctionArgs.hasSRetArg()) {
1563 QualType Ret = FI.getReturnType();
1564 llvm::Type *Ty = ConvertType(Ret);
1565 unsigned AddressSpace = Context.getTargetAddressSpace(Ret);
1566 ArgTypes[IRFunctionArgs.getSRetArgNo()] =
1567 llvm::PointerType::get(Ty, AddressSpace);
1570 // Add type for inalloca argument.
1571 if (IRFunctionArgs.hasInallocaArg()) {
1572 auto ArgStruct = FI.getArgStruct();
1574 ArgTypes[IRFunctionArgs.getInallocaArgNo()] = ArgStruct->getPointerTo();
1577 // Add in all of the required arguments.
1579 CGFunctionInfo::const_arg_iterator it = FI.arg_begin(),
1580 ie = it + FI.getNumRequiredArgs();
1581 for (; it != ie; ++it, ++ArgNo) {
1582 const ABIArgInfo &ArgInfo = it->info;
1584 // Insert a padding type to ensure proper alignment.
1585 if (IRFunctionArgs.hasPaddingArg(ArgNo))
1586 ArgTypes[IRFunctionArgs.getPaddingArgNo(ArgNo)] =
1587 ArgInfo.getPaddingType();
1589 unsigned FirstIRArg, NumIRArgs;
1590 std::tie(FirstIRArg, NumIRArgs) = IRFunctionArgs.getIRArgs(ArgNo);
1592 switch (ArgInfo.getKind()) {
1593 case ABIArgInfo::Ignore:
1594 case ABIArgInfo::InAlloca:
1595 assert(NumIRArgs == 0);
1598 case ABIArgInfo::Indirect: {
1599 assert(NumIRArgs == 1);
1600 // indirect arguments are always on the stack, which is alloca addr space.
1601 llvm::Type *LTy = ConvertTypeForMem(it->type);
1602 ArgTypes[FirstIRArg] = LTy->getPointerTo(
1603 CGM.getDataLayout().getAllocaAddrSpace());
1607 case ABIArgInfo::Extend:
1608 case ABIArgInfo::Direct: {
1609 // Fast-isel and the optimizer generally like scalar values better than
1610 // FCAs, so we flatten them if this is safe to do for this argument.
1611 llvm::Type *argType = ArgInfo.getCoerceToType();
1612 llvm::StructType *st = dyn_cast<llvm::StructType>(argType);
1613 if (st && ArgInfo.isDirect() && ArgInfo.getCanBeFlattened()) {
1614 assert(NumIRArgs == st->getNumElements());
1615 for (unsigned i = 0, e = st->getNumElements(); i != e; ++i)
1616 ArgTypes[FirstIRArg + i] = st->getElementType(i);
1618 assert(NumIRArgs == 1);
1619 ArgTypes[FirstIRArg] = argType;
1624 case ABIArgInfo::CoerceAndExpand: {
1625 auto ArgTypesIter = ArgTypes.begin() + FirstIRArg;
1626 for (auto EltTy : ArgInfo.getCoerceAndExpandTypeSequence()) {
1627 *ArgTypesIter++ = EltTy;
1629 assert(ArgTypesIter == ArgTypes.begin() + FirstIRArg + NumIRArgs);
1633 case ABIArgInfo::Expand:
1634 auto ArgTypesIter = ArgTypes.begin() + FirstIRArg;
1635 getExpandedTypes(it->type, ArgTypesIter);
1636 assert(ArgTypesIter == ArgTypes.begin() + FirstIRArg + NumIRArgs);
1641 bool Erased = FunctionsBeingProcessed.erase(&FI); (void)Erased;
1642 assert(Erased && "Not in set?");
1644 return llvm::FunctionType::get(resultType, ArgTypes, FI.isVariadic());
1647 llvm::Type *CodeGenTypes::GetFunctionTypeForVTable(GlobalDecl GD) {
1648 const CXXMethodDecl *MD = cast<CXXMethodDecl>(GD.getDecl());
1649 const FunctionProtoType *FPT = MD->getType()->getAs<FunctionProtoType>();
1651 if (!isFuncTypeConvertible(FPT))
1652 return llvm::StructType::get(getLLVMContext());
1654 const CGFunctionInfo *Info;
1655 if (isa<CXXDestructorDecl>(MD))
1657 &arrangeCXXStructorDeclaration(MD, getFromDtorType(GD.getDtorType()));
1659 Info = &arrangeCXXMethodDeclaration(MD);
1660 return GetFunctionType(*Info);
1663 static void AddAttributesFromFunctionProtoType(ASTContext &Ctx,
1664 llvm::AttrBuilder &FuncAttrs,
1665 const FunctionProtoType *FPT) {
1669 if (!isUnresolvedExceptionSpec(FPT->getExceptionSpecType()) &&
1670 FPT->isNothrow(Ctx))
1671 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1674 void CodeGenModule::ConstructDefaultFnAttrList(StringRef Name, bool HasOptnone,
1675 bool AttrOnCallSite,
1676 llvm::AttrBuilder &FuncAttrs) {
1677 // OptimizeNoneAttr takes precedence over -Os or -Oz. No warning needed.
1679 if (CodeGenOpts.OptimizeSize)
1680 FuncAttrs.addAttribute(llvm::Attribute::OptimizeForSize);
1681 if (CodeGenOpts.OptimizeSize == 2)
1682 FuncAttrs.addAttribute(llvm::Attribute::MinSize);
1685 if (CodeGenOpts.DisableRedZone)
1686 FuncAttrs.addAttribute(llvm::Attribute::NoRedZone);
1687 if (CodeGenOpts.NoImplicitFloat)
1688 FuncAttrs.addAttribute(llvm::Attribute::NoImplicitFloat);
1690 if (AttrOnCallSite) {
1691 // Attributes that should go on the call site only.
1692 if (!CodeGenOpts.SimplifyLibCalls ||
1693 CodeGenOpts.isNoBuiltinFunc(Name.data()))
1694 FuncAttrs.addAttribute(llvm::Attribute::NoBuiltin);
1695 if (!CodeGenOpts.TrapFuncName.empty())
1696 FuncAttrs.addAttribute("trap-func-name", CodeGenOpts.TrapFuncName);
1698 // Attributes that should go on the function, but not the call site.
1699 if (!CodeGenOpts.DisableFPElim) {
1700 FuncAttrs.addAttribute("no-frame-pointer-elim", "false");
1701 } else if (CodeGenOpts.OmitLeafFramePointer) {
1702 FuncAttrs.addAttribute("no-frame-pointer-elim", "false");
1703 FuncAttrs.addAttribute("no-frame-pointer-elim-non-leaf");
1705 FuncAttrs.addAttribute("no-frame-pointer-elim", "true");
1706 FuncAttrs.addAttribute("no-frame-pointer-elim-non-leaf");
1709 FuncAttrs.addAttribute("less-precise-fpmad",
1710 llvm::toStringRef(CodeGenOpts.LessPreciseFPMAD));
1712 if (!CodeGenOpts.FPDenormalMode.empty())
1713 FuncAttrs.addAttribute("denormal-fp-math", CodeGenOpts.FPDenormalMode);
1715 FuncAttrs.addAttribute("no-trapping-math",
1716 llvm::toStringRef(CodeGenOpts.NoTrappingMath));
1718 // TODO: Are these all needed?
1719 // unsafe/inf/nan/nsz are handled by instruction-level FastMathFlags.
1720 FuncAttrs.addAttribute("no-infs-fp-math",
1721 llvm::toStringRef(CodeGenOpts.NoInfsFPMath));
1722 FuncAttrs.addAttribute("no-nans-fp-math",
1723 llvm::toStringRef(CodeGenOpts.NoNaNsFPMath));
1724 FuncAttrs.addAttribute("unsafe-fp-math",
1725 llvm::toStringRef(CodeGenOpts.UnsafeFPMath));
1726 FuncAttrs.addAttribute("use-soft-float",
1727 llvm::toStringRef(CodeGenOpts.SoftFloat));
1728 FuncAttrs.addAttribute("stack-protector-buffer-size",
1729 llvm::utostr(CodeGenOpts.SSPBufferSize));
1730 FuncAttrs.addAttribute("no-signed-zeros-fp-math",
1731 llvm::toStringRef(CodeGenOpts.NoSignedZeros));
1732 FuncAttrs.addAttribute(
1733 "correctly-rounded-divide-sqrt-fp-math",
1734 llvm::toStringRef(CodeGenOpts.CorrectlyRoundedDivSqrt));
1736 // TODO: Reciprocal estimate codegen options should apply to instructions?
1737 std::vector<std::string> &Recips = getTarget().getTargetOpts().Reciprocals;
1738 if (!Recips.empty())
1739 FuncAttrs.addAttribute("reciprocal-estimates",
1740 llvm::join(Recips.begin(), Recips.end(), ","));
1742 if (CodeGenOpts.StackRealignment)
1743 FuncAttrs.addAttribute("stackrealign");
1744 if (CodeGenOpts.Backchain)
1745 FuncAttrs.addAttribute("backchain");
1748 if (getLangOpts().CUDA && getLangOpts().CUDAIsDevice) {
1749 // Conservatively, mark all functions and calls in CUDA as convergent
1750 // (meaning, they may call an intrinsically convergent op, such as
1751 // __syncthreads(), and so can't have certain optimizations applied around
1752 // them). LLVM will remove this attribute where it safely can.
1753 FuncAttrs.addAttribute(llvm::Attribute::Convergent);
1755 // Exceptions aren't supported in CUDA device code.
1756 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1758 // Respect -fcuda-flush-denormals-to-zero.
1759 if (getLangOpts().CUDADeviceFlushDenormalsToZero)
1760 FuncAttrs.addAttribute("nvptx-f32ftz", "true");
1764 void CodeGenModule::AddDefaultFnAttrs(llvm::Function &F) {
1765 llvm::AttrBuilder FuncAttrs;
1766 ConstructDefaultFnAttrList(F.getName(),
1767 F.hasFnAttribute(llvm::Attribute::OptimizeNone),
1768 /* AttrOnCallsite = */ false, FuncAttrs);
1769 F.addAttributes(llvm::AttributeList::FunctionIndex, FuncAttrs);
1772 void CodeGenModule::ConstructAttributeList(
1773 StringRef Name, const CGFunctionInfo &FI, CGCalleeInfo CalleeInfo,
1774 llvm::AttributeList &AttrList, unsigned &CallingConv, bool AttrOnCallSite) {
1775 llvm::AttrBuilder FuncAttrs;
1776 llvm::AttrBuilder RetAttrs;
1778 CallingConv = FI.getEffectiveCallingConvention();
1779 if (FI.isNoReturn())
1780 FuncAttrs.addAttribute(llvm::Attribute::NoReturn);
1782 // If we have information about the function prototype, we can learn
1783 // attributes form there.
1784 AddAttributesFromFunctionProtoType(getContext(), FuncAttrs,
1785 CalleeInfo.getCalleeFunctionProtoType());
1787 const Decl *TargetDecl = CalleeInfo.getCalleeDecl();
1789 bool HasOptnone = false;
1790 // FIXME: handle sseregparm someday...
1792 if (TargetDecl->hasAttr<ReturnsTwiceAttr>())
1793 FuncAttrs.addAttribute(llvm::Attribute::ReturnsTwice);
1794 if (TargetDecl->hasAttr<NoThrowAttr>())
1795 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1796 if (TargetDecl->hasAttr<NoReturnAttr>())
1797 FuncAttrs.addAttribute(llvm::Attribute::NoReturn);
1798 if (TargetDecl->hasAttr<ColdAttr>())
1799 FuncAttrs.addAttribute(llvm::Attribute::Cold);
1800 if (TargetDecl->hasAttr<NoDuplicateAttr>())
1801 FuncAttrs.addAttribute(llvm::Attribute::NoDuplicate);
1802 if (TargetDecl->hasAttr<ConvergentAttr>())
1803 FuncAttrs.addAttribute(llvm::Attribute::Convergent);
1805 if (const FunctionDecl *Fn = dyn_cast<FunctionDecl>(TargetDecl)) {
1806 AddAttributesFromFunctionProtoType(
1807 getContext(), FuncAttrs, Fn->getType()->getAs<FunctionProtoType>());
1808 // Don't use [[noreturn]] or _Noreturn for a call to a virtual function.
1809 // These attributes are not inherited by overloads.
1810 const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(Fn);
1811 if (Fn->isNoReturn() && !(AttrOnCallSite && MD && MD->isVirtual()))
1812 FuncAttrs.addAttribute(llvm::Attribute::NoReturn);
1815 // 'const', 'pure' and 'noalias' attributed functions are also nounwind.
1816 if (TargetDecl->hasAttr<ConstAttr>()) {
1817 FuncAttrs.addAttribute(llvm::Attribute::ReadNone);
1818 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1819 } else if (TargetDecl->hasAttr<PureAttr>()) {
1820 FuncAttrs.addAttribute(llvm::Attribute::ReadOnly);
1821 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1822 } else if (TargetDecl->hasAttr<NoAliasAttr>()) {
1823 FuncAttrs.addAttribute(llvm::Attribute::ArgMemOnly);
1824 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1826 if (TargetDecl->hasAttr<RestrictAttr>())
1827 RetAttrs.addAttribute(llvm::Attribute::NoAlias);
1828 if (TargetDecl->hasAttr<ReturnsNonNullAttr>())
1829 RetAttrs.addAttribute(llvm::Attribute::NonNull);
1830 if (TargetDecl->hasAttr<AnyX86NoCallerSavedRegistersAttr>())
1831 FuncAttrs.addAttribute("no_caller_saved_registers");
1833 HasOptnone = TargetDecl->hasAttr<OptimizeNoneAttr>();
1834 if (auto *AllocSize = TargetDecl->getAttr<AllocSizeAttr>()) {
1835 Optional<unsigned> NumElemsParam;
1836 // alloc_size args are base-1, 0 means not present.
1837 if (unsigned N = AllocSize->getNumElemsParam())
1838 NumElemsParam = N - 1;
1839 FuncAttrs.addAllocSizeAttr(AllocSize->getElemSizeParam() - 1,
1844 ConstructDefaultFnAttrList(Name, HasOptnone, AttrOnCallSite, FuncAttrs);
1846 if (CodeGenOpts.EnableSegmentedStacks &&
1847 !(TargetDecl && TargetDecl->hasAttr<NoSplitStackAttr>()))
1848 FuncAttrs.addAttribute("split-stack");
1850 if (!AttrOnCallSite) {
1851 bool DisableTailCalls =
1852 CodeGenOpts.DisableTailCalls ||
1853 (TargetDecl && (TargetDecl->hasAttr<DisableTailCallsAttr>() ||
1854 TargetDecl->hasAttr<AnyX86InterruptAttr>()));
1855 FuncAttrs.addAttribute("disable-tail-calls",
1856 llvm::toStringRef(DisableTailCalls));
1858 // Add target-cpu and target-features attributes to functions. If
1859 // we have a decl for the function and it has a target attribute then
1860 // parse that and add it to the feature set.
1861 StringRef TargetCPU = getTarget().getTargetOpts().CPU;
1862 const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(TargetDecl);
1863 if (FD && FD->hasAttr<TargetAttr>()) {
1864 llvm::StringMap<bool> FeatureMap;
1865 getFunctionFeatureMap(FeatureMap, FD);
1867 // Produce the canonical string for this set of features.
1868 std::vector<std::string> Features;
1869 for (llvm::StringMap<bool>::const_iterator it = FeatureMap.begin(),
1870 ie = FeatureMap.end();
1872 Features.push_back((it->second ? "+" : "-") + it->first().str());
1874 // Now add the target-cpu and target-features to the function.
1875 // While we populated the feature map above, we still need to
1876 // get and parse the target attribute so we can get the cpu for
1878 const auto *TD = FD->getAttr<TargetAttr>();
1879 TargetAttr::ParsedTargetAttr ParsedAttr = TD->parse();
1880 if (ParsedAttr.second != "")
1881 TargetCPU = ParsedAttr.second;
1882 if (TargetCPU != "")
1883 FuncAttrs.addAttribute("target-cpu", TargetCPU);
1884 if (!Features.empty()) {
1885 std::sort(Features.begin(), Features.end());
1886 FuncAttrs.addAttribute(
1888 llvm::join(Features.begin(), Features.end(), ","));
1891 // Otherwise just add the existing target cpu and target features to the
1893 std::vector<std::string> &Features = getTarget().getTargetOpts().Features;
1894 if (TargetCPU != "")
1895 FuncAttrs.addAttribute("target-cpu", TargetCPU);
1896 if (!Features.empty()) {
1897 std::sort(Features.begin(), Features.end());
1898 FuncAttrs.addAttribute(
1900 llvm::join(Features.begin(), Features.end(), ","));
1905 ClangToLLVMArgMapping IRFunctionArgs(getContext(), FI);
1907 QualType RetTy = FI.getReturnType();
1908 const ABIArgInfo &RetAI = FI.getReturnInfo();
1909 switch (RetAI.getKind()) {
1910 case ABIArgInfo::Extend:
1911 if (RetTy->hasSignedIntegerRepresentation())
1912 RetAttrs.addAttribute(llvm::Attribute::SExt);
1913 else if (RetTy->hasUnsignedIntegerRepresentation())
1914 RetAttrs.addAttribute(llvm::Attribute::ZExt);
1916 case ABIArgInfo::Direct:
1917 if (RetAI.getInReg())
1918 RetAttrs.addAttribute(llvm::Attribute::InReg);
1920 case ABIArgInfo::Ignore:
1923 case ABIArgInfo::InAlloca:
1924 case ABIArgInfo::Indirect: {
1925 // inalloca and sret disable readnone and readonly
1926 FuncAttrs.removeAttribute(llvm::Attribute::ReadOnly)
1927 .removeAttribute(llvm::Attribute::ReadNone);
1931 case ABIArgInfo::CoerceAndExpand:
1934 case ABIArgInfo::Expand:
1935 llvm_unreachable("Invalid ABI kind for return argument");
1938 if (const auto *RefTy = RetTy->getAs<ReferenceType>()) {
1939 QualType PTy = RefTy->getPointeeType();
1940 if (!PTy->isIncompleteType() && PTy->isConstantSizeType())
1941 RetAttrs.addDereferenceableAttr(getContext().getTypeSizeInChars(PTy)
1943 else if (getContext().getTargetAddressSpace(PTy) == 0)
1944 RetAttrs.addAttribute(llvm::Attribute::NonNull);
1947 bool hasUsedSRet = false;
1948 SmallVector<llvm::AttributeSet, 4> ArgAttrs(IRFunctionArgs.totalIRArgs());
1950 // Attach attributes to sret.
1951 if (IRFunctionArgs.hasSRetArg()) {
1952 llvm::AttrBuilder SRETAttrs;
1953 SRETAttrs.addAttribute(llvm::Attribute::StructRet);
1955 if (RetAI.getInReg())
1956 SRETAttrs.addAttribute(llvm::Attribute::InReg);
1957 ArgAttrs[IRFunctionArgs.getSRetArgNo()] =
1958 llvm::AttributeSet::get(getLLVMContext(), SRETAttrs);
1961 // Attach attributes to inalloca argument.
1962 if (IRFunctionArgs.hasInallocaArg()) {
1963 llvm::AttrBuilder Attrs;
1964 Attrs.addAttribute(llvm::Attribute::InAlloca);
1965 ArgAttrs[IRFunctionArgs.getInallocaArgNo()] =
1966 llvm::AttributeSet::get(getLLVMContext(), Attrs);
1970 for (CGFunctionInfo::const_arg_iterator I = FI.arg_begin(),
1972 I != E; ++I, ++ArgNo) {
1973 QualType ParamType = I->type;
1974 const ABIArgInfo &AI = I->info;
1975 llvm::AttrBuilder Attrs;
1977 // Add attribute for padding argument, if necessary.
1978 if (IRFunctionArgs.hasPaddingArg(ArgNo)) {
1979 if (AI.getPaddingInReg()) {
1980 ArgAttrs[IRFunctionArgs.getPaddingArgNo(ArgNo)] =
1981 llvm::AttributeSet::get(
1983 llvm::AttrBuilder().addAttribute(llvm::Attribute::InReg));
1987 // 'restrict' -> 'noalias' is done in EmitFunctionProlog when we
1988 // have the corresponding parameter variable. It doesn't make
1989 // sense to do it here because parameters are so messed up.
1990 switch (AI.getKind()) {
1991 case ABIArgInfo::Extend:
1992 if (ParamType->isSignedIntegerOrEnumerationType())
1993 Attrs.addAttribute(llvm::Attribute::SExt);
1994 else if (ParamType->isUnsignedIntegerOrEnumerationType()) {
1995 if (getTypes().getABIInfo().shouldSignExtUnsignedType(ParamType))
1996 Attrs.addAttribute(llvm::Attribute::SExt);
1998 Attrs.addAttribute(llvm::Attribute::ZExt);
2001 case ABIArgInfo::Direct:
2002 if (ArgNo == 0 && FI.isChainCall())
2003 Attrs.addAttribute(llvm::Attribute::Nest);
2004 else if (AI.getInReg())
2005 Attrs.addAttribute(llvm::Attribute::InReg);
2008 case ABIArgInfo::Indirect: {
2010 Attrs.addAttribute(llvm::Attribute::InReg);
2012 if (AI.getIndirectByVal())
2013 Attrs.addAttribute(llvm::Attribute::ByVal);
2015 CharUnits Align = AI.getIndirectAlign();
2017 // In a byval argument, it is important that the required
2018 // alignment of the type is honored, as LLVM might be creating a
2019 // *new* stack object, and needs to know what alignment to give
2020 // it. (Sometimes it can deduce a sensible alignment on its own,
2021 // but not if clang decides it must emit a packed struct, or the
2022 // user specifies increased alignment requirements.)
2024 // This is different from indirect *not* byval, where the object
2025 // exists already, and the align attribute is purely
2027 assert(!Align.isZero());
2029 // For now, only add this when we have a byval argument.
2030 // TODO: be less lazy about updating test cases.
2031 if (AI.getIndirectByVal())
2032 Attrs.addAlignmentAttr(Align.getQuantity());
2034 // byval disables readnone and readonly.
2035 FuncAttrs.removeAttribute(llvm::Attribute::ReadOnly)
2036 .removeAttribute(llvm::Attribute::ReadNone);
2039 case ABIArgInfo::Ignore:
2040 case ABIArgInfo::Expand:
2041 case ABIArgInfo::CoerceAndExpand:
2044 case ABIArgInfo::InAlloca:
2045 // inalloca disables readnone and readonly.
2046 FuncAttrs.removeAttribute(llvm::Attribute::ReadOnly)
2047 .removeAttribute(llvm::Attribute::ReadNone);
2051 if (const auto *RefTy = ParamType->getAs<ReferenceType>()) {
2052 QualType PTy = RefTy->getPointeeType();
2053 if (!PTy->isIncompleteType() && PTy->isConstantSizeType())
2054 Attrs.addDereferenceableAttr(getContext().getTypeSizeInChars(PTy)
2056 else if (getContext().getTargetAddressSpace(PTy) == 0)
2057 Attrs.addAttribute(llvm::Attribute::NonNull);
2060 switch (FI.getExtParameterInfo(ArgNo).getABI()) {
2061 case ParameterABI::Ordinary:
2064 case ParameterABI::SwiftIndirectResult: {
2065 // Add 'sret' if we haven't already used it for something, but
2066 // only if the result is void.
2067 if (!hasUsedSRet && RetTy->isVoidType()) {
2068 Attrs.addAttribute(llvm::Attribute::StructRet);
2072 // Add 'noalias' in either case.
2073 Attrs.addAttribute(llvm::Attribute::NoAlias);
2075 // Add 'dereferenceable' and 'alignment'.
2076 auto PTy = ParamType->getPointeeType();
2077 if (!PTy->isIncompleteType() && PTy->isConstantSizeType()) {
2078 auto info = getContext().getTypeInfoInChars(PTy);
2079 Attrs.addDereferenceableAttr(info.first.getQuantity());
2080 Attrs.addAttribute(llvm::Attribute::getWithAlignment(getLLVMContext(),
2081 info.second.getQuantity()));
2086 case ParameterABI::SwiftErrorResult:
2087 Attrs.addAttribute(llvm::Attribute::SwiftError);
2090 case ParameterABI::SwiftContext:
2091 Attrs.addAttribute(llvm::Attribute::SwiftSelf);
2095 if (Attrs.hasAttributes()) {
2096 unsigned FirstIRArg, NumIRArgs;
2097 std::tie(FirstIRArg, NumIRArgs) = IRFunctionArgs.getIRArgs(ArgNo);
2098 for (unsigned i = 0; i < NumIRArgs; i++)
2099 ArgAttrs[FirstIRArg + i] =
2100 llvm::AttributeSet::get(getLLVMContext(), Attrs);
2103 assert(ArgNo == FI.arg_size());
2105 AttrList = llvm::AttributeList::get(
2106 getLLVMContext(), llvm::AttributeSet::get(getLLVMContext(), FuncAttrs),
2107 llvm::AttributeSet::get(getLLVMContext(), RetAttrs), ArgAttrs);
2110 /// An argument came in as a promoted argument; demote it back to its
2112 static llvm::Value *emitArgumentDemotion(CodeGenFunction &CGF,
2114 llvm::Value *value) {
2115 llvm::Type *varType = CGF.ConvertType(var->getType());
2117 // This can happen with promotions that actually don't change the
2118 // underlying type, like the enum promotions.
2119 if (value->getType() == varType) return value;
2121 assert((varType->isIntegerTy() || varType->isFloatingPointTy())
2122 && "unexpected promotion type");
2124 if (isa<llvm::IntegerType>(varType))
2125 return CGF.Builder.CreateTrunc(value, varType, "arg.unpromote");
2127 return CGF.Builder.CreateFPCast(value, varType, "arg.unpromote");
2130 /// Returns the attribute (either parameter attribute, or function
2131 /// attribute), which declares argument ArgNo to be non-null.
2132 static const NonNullAttr *getNonNullAttr(const Decl *FD, const ParmVarDecl *PVD,
2133 QualType ArgType, unsigned ArgNo) {
2134 // FIXME: __attribute__((nonnull)) can also be applied to:
2135 // - references to pointers, where the pointee is known to be
2136 // nonnull (apparently a Clang extension)
2137 // - transparent unions containing pointers
2138 // In the former case, LLVM IR cannot represent the constraint. In
2139 // the latter case, we have no guarantee that the transparent union
2140 // is in fact passed as a pointer.
2141 if (!ArgType->isAnyPointerType() && !ArgType->isBlockPointerType())
2143 // First, check attribute on parameter itself.
2145 if (auto ParmNNAttr = PVD->getAttr<NonNullAttr>())
2148 // Check function attributes.
2151 for (const auto *NNAttr : FD->specific_attrs<NonNullAttr>()) {
2152 if (NNAttr->isNonNull(ArgNo))
2159 struct CopyBackSwiftError final : EHScopeStack::Cleanup {
2162 CopyBackSwiftError(Address temp, Address arg) : Temp(temp), Arg(arg) {}
2163 void Emit(CodeGenFunction &CGF, Flags flags) override {
2164 llvm::Value *errorValue = CGF.Builder.CreateLoad(Temp);
2165 CGF.Builder.CreateStore(errorValue, Arg);
2170 void CodeGenFunction::EmitFunctionProlog(const CGFunctionInfo &FI,
2172 const FunctionArgList &Args) {
2173 if (CurCodeDecl && CurCodeDecl->hasAttr<NakedAttr>())
2174 // Naked functions don't have prologues.
2177 // If this is an implicit-return-zero function, go ahead and
2178 // initialize the return value. TODO: it might be nice to have
2179 // a more general mechanism for this that didn't require synthesized
2180 // return statements.
2181 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(CurCodeDecl)) {
2182 if (FD->hasImplicitReturnZero()) {
2183 QualType RetTy = FD->getReturnType().getUnqualifiedType();
2184 llvm::Type* LLVMTy = CGM.getTypes().ConvertType(RetTy);
2185 llvm::Constant* Zero = llvm::Constant::getNullValue(LLVMTy);
2186 Builder.CreateStore(Zero, ReturnValue);
2190 // FIXME: We no longer need the types from FunctionArgList; lift up and
2193 ClangToLLVMArgMapping IRFunctionArgs(CGM.getContext(), FI);
2194 // Flattened function arguments.
2195 SmallVector<llvm::Value *, 16> FnArgs;
2196 FnArgs.reserve(IRFunctionArgs.totalIRArgs());
2197 for (auto &Arg : Fn->args()) {
2198 FnArgs.push_back(&Arg);
2200 assert(FnArgs.size() == IRFunctionArgs.totalIRArgs());
2202 // If we're using inalloca, all the memory arguments are GEPs off of the last
2203 // parameter, which is a pointer to the complete memory area.
2204 Address ArgStruct = Address::invalid();
2205 const llvm::StructLayout *ArgStructLayout = nullptr;
2206 if (IRFunctionArgs.hasInallocaArg()) {
2207 ArgStructLayout = CGM.getDataLayout().getStructLayout(FI.getArgStruct());
2208 ArgStruct = Address(FnArgs[IRFunctionArgs.getInallocaArgNo()],
2209 FI.getArgStructAlignment());
2211 assert(ArgStruct.getType() == FI.getArgStruct()->getPointerTo());
2214 // Name the struct return parameter.
2215 if (IRFunctionArgs.hasSRetArg()) {
2216 auto AI = cast<llvm::Argument>(FnArgs[IRFunctionArgs.getSRetArgNo()]);
2217 AI->setName("agg.result");
2218 AI->addAttr(llvm::Attribute::NoAlias);
2221 // Track if we received the parameter as a pointer (indirect, byval, or
2222 // inalloca). If already have a pointer, EmitParmDecl doesn't need to copy it
2223 // into a local alloca for us.
2224 SmallVector<ParamValue, 16> ArgVals;
2225 ArgVals.reserve(Args.size());
2227 // Create a pointer value for every parameter declaration. This usually
2228 // entails copying one or more LLVM IR arguments into an alloca. Don't push
2229 // any cleanups or do anything that might unwind. We do that separately, so
2230 // we can push the cleanups in the correct order for the ABI.
2231 assert(FI.arg_size() == Args.size() &&
2232 "Mismatch between function signature & arguments.");
2234 CGFunctionInfo::const_arg_iterator info_it = FI.arg_begin();
2235 for (FunctionArgList::const_iterator i = Args.begin(), e = Args.end();
2236 i != e; ++i, ++info_it, ++ArgNo) {
2237 const VarDecl *Arg = *i;
2238 QualType Ty = info_it->type;
2239 const ABIArgInfo &ArgI = info_it->info;
2242 isa<ParmVarDecl>(Arg) && cast<ParmVarDecl>(Arg)->isKNRPromoted();
2244 unsigned FirstIRArg, NumIRArgs;
2245 std::tie(FirstIRArg, NumIRArgs) = IRFunctionArgs.getIRArgs(ArgNo);
2247 switch (ArgI.getKind()) {
2248 case ABIArgInfo::InAlloca: {
2249 assert(NumIRArgs == 0);
2250 auto FieldIndex = ArgI.getInAllocaFieldIndex();
2251 CharUnits FieldOffset =
2252 CharUnits::fromQuantity(ArgStructLayout->getElementOffset(FieldIndex));
2253 Address V = Builder.CreateStructGEP(ArgStruct, FieldIndex, FieldOffset,
2255 ArgVals.push_back(ParamValue::forIndirect(V));
2259 case ABIArgInfo::Indirect: {
2260 assert(NumIRArgs == 1);
2261 Address ParamAddr = Address(FnArgs[FirstIRArg], ArgI.getIndirectAlign());
2263 if (!hasScalarEvaluationKind(Ty)) {
2264 // Aggregates and complex variables are accessed by reference. All we
2265 // need to do is realign the value, if requested.
2266 Address V = ParamAddr;
2267 if (ArgI.getIndirectRealign()) {
2268 Address AlignedTemp = CreateMemTemp(Ty, "coerce");
2270 // Copy from the incoming argument pointer to the temporary with the
2271 // appropriate alignment.
2273 // FIXME: We should have a common utility for generating an aggregate
2275 CharUnits Size = getContext().getTypeSizeInChars(Ty);
2276 auto SizeVal = llvm::ConstantInt::get(IntPtrTy, Size.getQuantity());
2277 Address Dst = Builder.CreateBitCast(AlignedTemp, Int8PtrTy);
2278 Address Src = Builder.CreateBitCast(ParamAddr, Int8PtrTy);
2279 Builder.CreateMemCpy(Dst, Src, SizeVal, false);
2282 ArgVals.push_back(ParamValue::forIndirect(V));
2284 // Load scalar value from indirect argument.
2286 EmitLoadOfScalar(ParamAddr, false, Ty, Arg->getLocStart());
2289 V = emitArgumentDemotion(*this, Arg, V);
2290 ArgVals.push_back(ParamValue::forDirect(V));
2295 case ABIArgInfo::Extend:
2296 case ABIArgInfo::Direct: {
2298 // If we have the trivial case, handle it with no muss and fuss.
2299 if (!isa<llvm::StructType>(ArgI.getCoerceToType()) &&
2300 ArgI.getCoerceToType() == ConvertType(Ty) &&
2301 ArgI.getDirectOffset() == 0) {
2302 assert(NumIRArgs == 1);
2303 llvm::Value *V = FnArgs[FirstIRArg];
2304 auto AI = cast<llvm::Argument>(V);
2306 if (const ParmVarDecl *PVD = dyn_cast<ParmVarDecl>(Arg)) {
2307 if (getNonNullAttr(CurCodeDecl, PVD, PVD->getType(),
2308 PVD->getFunctionScopeIndex()))
2309 AI->addAttr(llvm::Attribute::NonNull);
2311 QualType OTy = PVD->getOriginalType();
2312 if (const auto *ArrTy =
2313 getContext().getAsConstantArrayType(OTy)) {
2314 // A C99 array parameter declaration with the static keyword also
2315 // indicates dereferenceability, and if the size is constant we can
2316 // use the dereferenceable attribute (which requires the size in
2318 if (ArrTy->getSizeModifier() == ArrayType::Static) {
2319 QualType ETy = ArrTy->getElementType();
2320 uint64_t ArrSize = ArrTy->getSize().getZExtValue();
2321 if (!ETy->isIncompleteType() && ETy->isConstantSizeType() &&
2323 llvm::AttrBuilder Attrs;
2324 Attrs.addDereferenceableAttr(
2325 getContext().getTypeSizeInChars(ETy).getQuantity()*ArrSize);
2326 AI->addAttrs(Attrs);
2327 } else if (getContext().getTargetAddressSpace(ETy) == 0) {
2328 AI->addAttr(llvm::Attribute::NonNull);
2331 } else if (const auto *ArrTy =
2332 getContext().getAsVariableArrayType(OTy)) {
2333 // For C99 VLAs with the static keyword, we don't know the size so
2334 // we can't use the dereferenceable attribute, but in addrspace(0)
2335 // we know that it must be nonnull.
2336 if (ArrTy->getSizeModifier() == VariableArrayType::Static &&
2337 !getContext().getTargetAddressSpace(ArrTy->getElementType()))
2338 AI->addAttr(llvm::Attribute::NonNull);
2341 const auto *AVAttr = PVD->getAttr<AlignValueAttr>();
2343 if (const auto *TOTy = dyn_cast<TypedefType>(OTy))
2344 AVAttr = TOTy->getDecl()->getAttr<AlignValueAttr>();
2346 llvm::Value *AlignmentValue =
2347 EmitScalarExpr(AVAttr->getAlignment());
2348 llvm::ConstantInt *AlignmentCI =
2349 cast<llvm::ConstantInt>(AlignmentValue);
2350 unsigned Alignment = std::min((unsigned)AlignmentCI->getZExtValue(),
2351 +llvm::Value::MaximumAlignment);
2352 AI->addAttrs(llvm::AttrBuilder().addAlignmentAttr(Alignment));
2356 if (Arg->getType().isRestrictQualified())
2357 AI->addAttr(llvm::Attribute::NoAlias);
2359 // LLVM expects swifterror parameters to be used in very restricted
2360 // ways. Copy the value into a less-restricted temporary.
2361 if (FI.getExtParameterInfo(ArgNo).getABI()
2362 == ParameterABI::SwiftErrorResult) {
2363 QualType pointeeTy = Ty->getPointeeType();
2364 assert(pointeeTy->isPointerType());
2366 CreateMemTemp(pointeeTy, getPointerAlign(), "swifterror.temp");
2367 Address arg = Address(V, getContext().getTypeAlignInChars(pointeeTy));
2368 llvm::Value *incomingErrorValue = Builder.CreateLoad(arg);
2369 Builder.CreateStore(incomingErrorValue, temp);
2370 V = temp.getPointer();
2372 // Push a cleanup to copy the value back at the end of the function.
2373 // The convention does not guarantee that the value will be written
2374 // back if the function exits with an unwind exception.
2375 EHStack.pushCleanup<CopyBackSwiftError>(NormalCleanup, temp, arg);
2378 // Ensure the argument is the correct type.
2379 if (V->getType() != ArgI.getCoerceToType())
2380 V = Builder.CreateBitCast(V, ArgI.getCoerceToType());
2383 V = emitArgumentDemotion(*this, Arg, V);
2385 // Because of merging of function types from multiple decls it is
2386 // possible for the type of an argument to not match the corresponding
2387 // type in the function type. Since we are codegening the callee
2388 // in here, add a cast to the argument type.
2389 llvm::Type *LTy = ConvertType(Arg->getType());
2390 if (V->getType() != LTy)
2391 V = Builder.CreateBitCast(V, LTy);
2393 ArgVals.push_back(ParamValue::forDirect(V));
2397 Address Alloca = CreateMemTemp(Ty, getContext().getDeclAlign(Arg),
2400 // Pointer to store into.
2401 Address Ptr = emitAddressAtOffset(*this, Alloca, ArgI);
2403 // Fast-isel and the optimizer generally like scalar values better than
2404 // FCAs, so we flatten them if this is safe to do for this argument.
2405 llvm::StructType *STy = dyn_cast<llvm::StructType>(ArgI.getCoerceToType());
2406 if (ArgI.isDirect() && ArgI.getCanBeFlattened() && STy &&
2407 STy->getNumElements() > 1) {
2408 auto SrcLayout = CGM.getDataLayout().getStructLayout(STy);
2409 uint64_t SrcSize = CGM.getDataLayout().getTypeAllocSize(STy);
2410 llvm::Type *DstTy = Ptr.getElementType();
2411 uint64_t DstSize = CGM.getDataLayout().getTypeAllocSize(DstTy);
2413 Address AddrToStoreInto = Address::invalid();
2414 if (SrcSize <= DstSize) {
2416 Builder.CreateBitCast(Ptr, llvm::PointerType::getUnqual(STy));
2419 CreateTempAlloca(STy, Alloca.getAlignment(), "coerce");
2422 assert(STy->getNumElements() == NumIRArgs);
2423 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
2424 auto AI = FnArgs[FirstIRArg + i];
2425 AI->setName(Arg->getName() + ".coerce" + Twine(i));
2426 auto Offset = CharUnits::fromQuantity(SrcLayout->getElementOffset(i));
2428 Builder.CreateStructGEP(AddrToStoreInto, i, Offset);
2429 Builder.CreateStore(AI, EltPtr);
2432 if (SrcSize > DstSize) {
2433 Builder.CreateMemCpy(Ptr, AddrToStoreInto, DstSize);
2437 // Simple case, just do a coerced store of the argument into the alloca.
2438 assert(NumIRArgs == 1);
2439 auto AI = FnArgs[FirstIRArg];
2440 AI->setName(Arg->getName() + ".coerce");
2441 CreateCoercedStore(AI, Ptr, /*DestIsVolatile=*/false, *this);
2444 // Match to what EmitParmDecl is expecting for this type.
2445 if (CodeGenFunction::hasScalarEvaluationKind(Ty)) {
2447 EmitLoadOfScalar(Alloca, false, Ty, Arg->getLocStart());
2449 V = emitArgumentDemotion(*this, Arg, V);
2450 ArgVals.push_back(ParamValue::forDirect(V));
2452 ArgVals.push_back(ParamValue::forIndirect(Alloca));
2457 case ABIArgInfo::CoerceAndExpand: {
2458 // Reconstruct into a temporary.
2459 Address alloca = CreateMemTemp(Ty, getContext().getDeclAlign(Arg));
2460 ArgVals.push_back(ParamValue::forIndirect(alloca));
2462 auto coercionType = ArgI.getCoerceAndExpandType();
2463 alloca = Builder.CreateElementBitCast(alloca, coercionType);
2464 auto layout = CGM.getDataLayout().getStructLayout(coercionType);
2466 unsigned argIndex = FirstIRArg;
2467 for (unsigned i = 0, e = coercionType->getNumElements(); i != e; ++i) {
2468 llvm::Type *eltType = coercionType->getElementType(i);
2469 if (ABIArgInfo::isPaddingForCoerceAndExpand(eltType))
2472 auto eltAddr = Builder.CreateStructGEP(alloca, i, layout);
2473 auto elt = FnArgs[argIndex++];
2474 Builder.CreateStore(elt, eltAddr);
2476 assert(argIndex == FirstIRArg + NumIRArgs);
2480 case ABIArgInfo::Expand: {
2481 // If this structure was expanded into multiple arguments then
2482 // we need to create a temporary and reconstruct it from the
2484 Address Alloca = CreateMemTemp(Ty, getContext().getDeclAlign(Arg));
2485 LValue LV = MakeAddrLValue(Alloca, Ty);
2486 ArgVals.push_back(ParamValue::forIndirect(Alloca));
2488 auto FnArgIter = FnArgs.begin() + FirstIRArg;
2489 ExpandTypeFromArgs(Ty, LV, FnArgIter);
2490 assert(FnArgIter == FnArgs.begin() + FirstIRArg + NumIRArgs);
2491 for (unsigned i = 0, e = NumIRArgs; i != e; ++i) {
2492 auto AI = FnArgs[FirstIRArg + i];
2493 AI->setName(Arg->getName() + "." + Twine(i));
2498 case ABIArgInfo::Ignore:
2499 assert(NumIRArgs == 0);
2500 // Initialize the local variable appropriately.
2501 if (!hasScalarEvaluationKind(Ty)) {
2502 ArgVals.push_back(ParamValue::forIndirect(CreateMemTemp(Ty)));
2504 llvm::Value *U = llvm::UndefValue::get(ConvertType(Arg->getType()));
2505 ArgVals.push_back(ParamValue::forDirect(U));
2511 if (getTarget().getCXXABI().areArgsDestroyedLeftToRightInCallee()) {
2512 for (int I = Args.size() - 1; I >= 0; --I)
2513 EmitParmDecl(*Args[I], ArgVals[I], I + 1);
2515 for (unsigned I = 0, E = Args.size(); I != E; ++I)
2516 EmitParmDecl(*Args[I], ArgVals[I], I + 1);
2520 static void eraseUnusedBitCasts(llvm::Instruction *insn) {
2521 while (insn->use_empty()) {
2522 llvm::BitCastInst *bitcast = dyn_cast<llvm::BitCastInst>(insn);
2523 if (!bitcast) return;
2525 // This is "safe" because we would have used a ConstantExpr otherwise.
2526 insn = cast<llvm::Instruction>(bitcast->getOperand(0));
2527 bitcast->eraseFromParent();
2531 /// Try to emit a fused autorelease of a return result.
2532 static llvm::Value *tryEmitFusedAutoreleaseOfResult(CodeGenFunction &CGF,
2533 llvm::Value *result) {
2534 // We must be immediately followed the cast.
2535 llvm::BasicBlock *BB = CGF.Builder.GetInsertBlock();
2536 if (BB->empty()) return nullptr;
2537 if (&BB->back() != result) return nullptr;
2539 llvm::Type *resultType = result->getType();
2541 // result is in a BasicBlock and is therefore an Instruction.
2542 llvm::Instruction *generator = cast<llvm::Instruction>(result);
2544 SmallVector<llvm::Instruction *, 4> InstsToKill;
2547 // %generator = bitcast %type1* %generator2 to %type2*
2548 while (llvm::BitCastInst *bitcast = dyn_cast<llvm::BitCastInst>(generator)) {
2549 // We would have emitted this as a constant if the operand weren't
2551 generator = cast<llvm::Instruction>(bitcast->getOperand(0));
2553 // Require the generator to be immediately followed by the cast.
2554 if (generator->getNextNode() != bitcast)
2557 InstsToKill.push_back(bitcast);
2561 // %generator = call i8* @objc_retain(i8* %originalResult)
2563 // %generator = call i8* @objc_retainAutoreleasedReturnValue(i8* %originalResult)
2564 llvm::CallInst *call = dyn_cast<llvm::CallInst>(generator);
2565 if (!call) return nullptr;
2567 bool doRetainAutorelease;
2569 if (call->getCalledValue() == CGF.CGM.getObjCEntrypoints().objc_retain) {
2570 doRetainAutorelease = true;
2571 } else if (call->getCalledValue() == CGF.CGM.getObjCEntrypoints()
2572 .objc_retainAutoreleasedReturnValue) {
2573 doRetainAutorelease = false;
2575 // If we emitted an assembly marker for this call (and the
2576 // ARCEntrypoints field should have been set if so), go looking
2577 // for that call. If we can't find it, we can't do this
2578 // optimization. But it should always be the immediately previous
2579 // instruction, unless we needed bitcasts around the call.
2580 if (CGF.CGM.getObjCEntrypoints().retainAutoreleasedReturnValueMarker) {
2581 llvm::Instruction *prev = call->getPrevNode();
2583 if (isa<llvm::BitCastInst>(prev)) {
2584 prev = prev->getPrevNode();
2587 assert(isa<llvm::CallInst>(prev));
2588 assert(cast<llvm::CallInst>(prev)->getCalledValue() ==
2589 CGF.CGM.getObjCEntrypoints().retainAutoreleasedReturnValueMarker);
2590 InstsToKill.push_back(prev);
2596 result = call->getArgOperand(0);
2597 InstsToKill.push_back(call);
2599 // Keep killing bitcasts, for sanity. Note that we no longer care
2600 // about precise ordering as long as there's exactly one use.
2601 while (llvm::BitCastInst *bitcast = dyn_cast<llvm::BitCastInst>(result)) {
2602 if (!bitcast->hasOneUse()) break;
2603 InstsToKill.push_back(bitcast);
2604 result = bitcast->getOperand(0);
2607 // Delete all the unnecessary instructions, from latest to earliest.
2608 for (auto *I : InstsToKill)
2609 I->eraseFromParent();
2611 // Do the fused retain/autorelease if we were asked to.
2612 if (doRetainAutorelease)
2613 result = CGF.EmitARCRetainAutoreleaseReturnValue(result);
2615 // Cast back to the result type.
2616 return CGF.Builder.CreateBitCast(result, resultType);
2619 /// If this is a +1 of the value of an immutable 'self', remove it.
2620 static llvm::Value *tryRemoveRetainOfSelf(CodeGenFunction &CGF,
2621 llvm::Value *result) {
2622 // This is only applicable to a method with an immutable 'self'.
2623 const ObjCMethodDecl *method =
2624 dyn_cast_or_null<ObjCMethodDecl>(CGF.CurCodeDecl);
2625 if (!method) return nullptr;
2626 const VarDecl *self = method->getSelfDecl();
2627 if (!self->getType().isConstQualified()) return nullptr;
2629 // Look for a retain call.
2630 llvm::CallInst *retainCall =
2631 dyn_cast<llvm::CallInst>(result->stripPointerCasts());
2633 retainCall->getCalledValue() != CGF.CGM.getObjCEntrypoints().objc_retain)
2636 // Look for an ordinary load of 'self'.
2637 llvm::Value *retainedValue = retainCall->getArgOperand(0);
2638 llvm::LoadInst *load =
2639 dyn_cast<llvm::LoadInst>(retainedValue->stripPointerCasts());
2640 if (!load || load->isAtomic() || load->isVolatile() ||
2641 load->getPointerOperand() != CGF.GetAddrOfLocalVar(self).getPointer())
2644 // Okay! Burn it all down. This relies for correctness on the
2645 // assumption that the retain is emitted as part of the return and
2646 // that thereafter everything is used "linearly".
2647 llvm::Type *resultType = result->getType();
2648 eraseUnusedBitCasts(cast<llvm::Instruction>(result));
2649 assert(retainCall->use_empty());
2650 retainCall->eraseFromParent();
2651 eraseUnusedBitCasts(cast<llvm::Instruction>(retainedValue));
2653 return CGF.Builder.CreateBitCast(load, resultType);
2656 /// Emit an ARC autorelease of the result of a function.
2658 /// \return the value to actually return from the function
2659 static llvm::Value *emitAutoreleaseOfResult(CodeGenFunction &CGF,
2660 llvm::Value *result) {
2661 // If we're returning 'self', kill the initial retain. This is a
2662 // heuristic attempt to "encourage correctness" in the really unfortunate
2663 // case where we have a return of self during a dealloc and we desperately
2664 // need to avoid the possible autorelease.
2665 if (llvm::Value *self = tryRemoveRetainOfSelf(CGF, result))
2668 // At -O0, try to emit a fused retain/autorelease.
2669 if (CGF.shouldUseFusedARCCalls())
2670 if (llvm::Value *fused = tryEmitFusedAutoreleaseOfResult(CGF, result))
2673 return CGF.EmitARCAutoreleaseReturnValue(result);
2676 /// Heuristically search for a dominating store to the return-value slot.
2677 static llvm::StoreInst *findDominatingStoreToReturnValue(CodeGenFunction &CGF) {
2678 // Check if a User is a store which pointerOperand is the ReturnValue.
2679 // We are looking for stores to the ReturnValue, not for stores of the
2680 // ReturnValue to some other location.
2681 auto GetStoreIfValid = [&CGF](llvm::User *U) -> llvm::StoreInst * {
2682 auto *SI = dyn_cast<llvm::StoreInst>(U);
2683 if (!SI || SI->getPointerOperand() != CGF.ReturnValue.getPointer())
2685 // These aren't actually possible for non-coerced returns, and we
2686 // only care about non-coerced returns on this code path.
2687 assert(!SI->isAtomic() && !SI->isVolatile());
2690 // If there are multiple uses of the return-value slot, just check
2691 // for something immediately preceding the IP. Sometimes this can
2692 // happen with how we generate implicit-returns; it can also happen
2693 // with noreturn cleanups.
2694 if (!CGF.ReturnValue.getPointer()->hasOneUse()) {
2695 llvm::BasicBlock *IP = CGF.Builder.GetInsertBlock();
2696 if (IP->empty()) return nullptr;
2697 llvm::Instruction *I = &IP->back();
2699 // Skip lifetime markers
2700 for (llvm::BasicBlock::reverse_iterator II = IP->rbegin(),
2703 if (llvm::IntrinsicInst *Intrinsic =
2704 dyn_cast<llvm::IntrinsicInst>(&*II)) {
2705 if (Intrinsic->getIntrinsicID() == llvm::Intrinsic::lifetime_end) {
2706 const llvm::Value *CastAddr = Intrinsic->getArgOperand(1);
2710 if (isa<llvm::BitCastInst>(&*II) && (CastAddr == &*II))
2718 return GetStoreIfValid(I);
2721 llvm::StoreInst *store =
2722 GetStoreIfValid(CGF.ReturnValue.getPointer()->user_back());
2723 if (!store) return nullptr;
2725 // Now do a first-and-dirty dominance check: just walk up the
2726 // single-predecessors chain from the current insertion point.
2727 llvm::BasicBlock *StoreBB = store->getParent();
2728 llvm::BasicBlock *IP = CGF.Builder.GetInsertBlock();
2729 while (IP != StoreBB) {
2730 if (!(IP = IP->getSinglePredecessor()))
2734 // Okay, the store's basic block dominates the insertion point; we
2735 // can do our thing.
2739 void CodeGenFunction::EmitFunctionEpilog(const CGFunctionInfo &FI,
2741 SourceLocation EndLoc) {
2742 if (CurCodeDecl && CurCodeDecl->hasAttr<NakedAttr>()) {
2743 // Naked functions don't have epilogues.
2744 Builder.CreateUnreachable();
2748 // Functions with no result always return void.
2749 if (!ReturnValue.isValid()) {
2750 Builder.CreateRetVoid();
2754 llvm::DebugLoc RetDbgLoc;
2755 llvm::Value *RV = nullptr;
2756 QualType RetTy = FI.getReturnType();
2757 const ABIArgInfo &RetAI = FI.getReturnInfo();
2759 switch (RetAI.getKind()) {
2760 case ABIArgInfo::InAlloca:
2761 // Aggregrates get evaluated directly into the destination. Sometimes we
2762 // need to return the sret value in a register, though.
2763 assert(hasAggregateEvaluationKind(RetTy));
2764 if (RetAI.getInAllocaSRet()) {
2765 llvm::Function::arg_iterator EI = CurFn->arg_end();
2767 llvm::Value *ArgStruct = &*EI;
2768 llvm::Value *SRet = Builder.CreateStructGEP(
2769 nullptr, ArgStruct, RetAI.getInAllocaFieldIndex());
2770 RV = Builder.CreateAlignedLoad(SRet, getPointerAlign(), "sret");
2774 case ABIArgInfo::Indirect: {
2775 auto AI = CurFn->arg_begin();
2776 if (RetAI.isSRetAfterThis())
2778 switch (getEvaluationKind(RetTy)) {
2781 EmitLoadOfComplex(MakeAddrLValue(ReturnValue, RetTy), EndLoc);
2782 EmitStoreOfComplex(RT, MakeNaturalAlignAddrLValue(&*AI, RetTy),
2787 // Do nothing; aggregrates get evaluated directly into the destination.
2790 EmitStoreOfScalar(Builder.CreateLoad(ReturnValue),
2791 MakeNaturalAlignAddrLValue(&*AI, RetTy),
2798 case ABIArgInfo::Extend:
2799 case ABIArgInfo::Direct:
2800 if (RetAI.getCoerceToType() == ConvertType(RetTy) &&
2801 RetAI.getDirectOffset() == 0) {
2802 // The internal return value temp always will have pointer-to-return-type
2803 // type, just do a load.
2805 // If there is a dominating store to ReturnValue, we can elide
2806 // the load, zap the store, and usually zap the alloca.
2807 if (llvm::StoreInst *SI =
2808 findDominatingStoreToReturnValue(*this)) {
2809 // Reuse the debug location from the store unless there is
2810 // cleanup code to be emitted between the store and return
2812 if (EmitRetDbgLoc && !AutoreleaseResult)
2813 RetDbgLoc = SI->getDebugLoc();
2814 // Get the stored value and nuke the now-dead store.
2815 RV = SI->getValueOperand();
2816 SI->eraseFromParent();
2818 // If that was the only use of the return value, nuke it as well now.
2819 auto returnValueInst = ReturnValue.getPointer();
2820 if (returnValueInst->use_empty()) {
2821 if (auto alloca = dyn_cast<llvm::AllocaInst>(returnValueInst)) {
2822 alloca->eraseFromParent();
2823 ReturnValue = Address::invalid();
2827 // Otherwise, we have to do a simple load.
2829 RV = Builder.CreateLoad(ReturnValue);
2832 // If the value is offset in memory, apply the offset now.
2833 Address V = emitAddressAtOffset(*this, ReturnValue, RetAI);
2835 RV = CreateCoercedLoad(V, RetAI.getCoerceToType(), *this);
2838 // In ARC, end functions that return a retainable type with a call
2839 // to objc_autoreleaseReturnValue.
2840 if (AutoreleaseResult) {
2842 // Type::isObjCRetainabletype has to be called on a QualType that hasn't
2843 // been stripped of the typedefs, so we cannot use RetTy here. Get the
2844 // original return type of FunctionDecl, CurCodeDecl, and BlockDecl from
2845 // CurCodeDecl or BlockInfo.
2848 if (auto *FD = dyn_cast<FunctionDecl>(CurCodeDecl))
2849 RT = FD->getReturnType();
2850 else if (auto *MD = dyn_cast<ObjCMethodDecl>(CurCodeDecl))
2851 RT = MD->getReturnType();
2852 else if (isa<BlockDecl>(CurCodeDecl))
2853 RT = BlockInfo->BlockExpression->getFunctionType()->getReturnType();
2855 llvm_unreachable("Unexpected function/method type");
2857 assert(getLangOpts().ObjCAutoRefCount &&
2858 !FI.isReturnsRetained() &&
2859 RT->isObjCRetainableType());
2861 RV = emitAutoreleaseOfResult(*this, RV);
2866 case ABIArgInfo::Ignore:
2869 case ABIArgInfo::CoerceAndExpand: {
2870 auto coercionType = RetAI.getCoerceAndExpandType();
2871 auto layout = CGM.getDataLayout().getStructLayout(coercionType);
2873 // Load all of the coerced elements out into results.
2874 llvm::SmallVector<llvm::Value*, 4> results;
2875 Address addr = Builder.CreateElementBitCast(ReturnValue, coercionType);
2876 for (unsigned i = 0, e = coercionType->getNumElements(); i != e; ++i) {
2877 auto coercedEltType = coercionType->getElementType(i);
2878 if (ABIArgInfo::isPaddingForCoerceAndExpand(coercedEltType))
2881 auto eltAddr = Builder.CreateStructGEP(addr, i, layout);
2882 auto elt = Builder.CreateLoad(eltAddr);
2883 results.push_back(elt);
2886 // If we have one result, it's the single direct result type.
2887 if (results.size() == 1) {
2890 // Otherwise, we need to make a first-class aggregate.
2892 // Construct a return type that lacks padding elements.
2893 llvm::Type *returnType = RetAI.getUnpaddedCoerceAndExpandType();
2895 RV = llvm::UndefValue::get(returnType);
2896 for (unsigned i = 0, e = results.size(); i != e; ++i) {
2897 RV = Builder.CreateInsertValue(RV, results[i], i);
2903 case ABIArgInfo::Expand:
2904 llvm_unreachable("Invalid ABI kind for return argument");
2907 llvm::Instruction *Ret;
2909 EmitReturnValueCheck(RV);
2910 Ret = Builder.CreateRet(RV);
2912 Ret = Builder.CreateRetVoid();
2916 Ret->setDebugLoc(std::move(RetDbgLoc));
2919 void CodeGenFunction::EmitReturnValueCheck(llvm::Value *RV) {
2920 // A current decl may not be available when emitting vtable thunks.
2924 ReturnsNonNullAttr *RetNNAttr = nullptr;
2925 if (SanOpts.has(SanitizerKind::ReturnsNonnullAttribute))
2926 RetNNAttr = CurCodeDecl->getAttr<ReturnsNonNullAttr>();
2928 if (!RetNNAttr && !requiresReturnValueNullabilityCheck())
2931 // Prefer the returns_nonnull attribute if it's present.
2932 SourceLocation AttrLoc;
2933 SanitizerMask CheckKind;
2934 SanitizerHandler Handler;
2936 assert(!requiresReturnValueNullabilityCheck() &&
2937 "Cannot check nullability and the nonnull attribute");
2938 AttrLoc = RetNNAttr->getLocation();
2939 CheckKind = SanitizerKind::ReturnsNonnullAttribute;
2940 Handler = SanitizerHandler::NonnullReturn;
2942 if (auto *DD = dyn_cast<DeclaratorDecl>(CurCodeDecl))
2943 if (auto *TSI = DD->getTypeSourceInfo())
2944 if (auto FTL = TSI->getTypeLoc().castAs<FunctionTypeLoc>())
2945 AttrLoc = FTL.getReturnLoc().findNullabilityLoc();
2946 CheckKind = SanitizerKind::NullabilityReturn;
2947 Handler = SanitizerHandler::NullabilityReturn;
2950 SanitizerScope SanScope(this);
2952 // Make sure the "return" source location is valid. If we're checking a
2953 // nullability annotation, make sure the preconditions for the check are met.
2954 llvm::BasicBlock *Check = createBasicBlock("nullcheck");
2955 llvm::BasicBlock *NoCheck = createBasicBlock("no.nullcheck");
2956 llvm::Value *SLocPtr = Builder.CreateLoad(ReturnLocation, "return.sloc.load");
2957 llvm::Value *CanNullCheck = Builder.CreateIsNotNull(SLocPtr);
2958 if (requiresReturnValueNullabilityCheck())
2960 Builder.CreateAnd(CanNullCheck, RetValNullabilityPrecondition);
2961 Builder.CreateCondBr(CanNullCheck, Check, NoCheck);
2964 // Now do the null check.
2965 llvm::Value *Cond = Builder.CreateIsNotNull(RV);
2966 llvm::Constant *StaticData[] = {EmitCheckSourceLocation(AttrLoc)};
2967 llvm::Value *DynamicData[] = {SLocPtr};
2968 EmitCheck(std::make_pair(Cond, CheckKind), Handler, StaticData, DynamicData);
2973 // The return location should not be used after the check has been emitted.
2974 ReturnLocation = Address::invalid();
2978 static bool isInAllocaArgument(CGCXXABI &ABI, QualType type) {
2979 const CXXRecordDecl *RD = type->getAsCXXRecordDecl();
2980 return RD && ABI.getRecordArgABI(RD) == CGCXXABI::RAA_DirectInMemory;
2983 static AggValueSlot createPlaceholderSlot(CodeGenFunction &CGF,
2985 // FIXME: Generate IR in one pass, rather than going back and fixing up these
2987 llvm::Type *IRTy = CGF.ConvertTypeForMem(Ty);
2988 llvm::Type *IRPtrTy = IRTy->getPointerTo();
2989 llvm::Value *Placeholder = llvm::UndefValue::get(IRPtrTy->getPointerTo());
2991 // FIXME: When we generate this IR in one pass, we shouldn't need
2992 // this win32-specific alignment hack.
2993 CharUnits Align = CharUnits::fromQuantity(4);
2994 Placeholder = CGF.Builder.CreateAlignedLoad(IRPtrTy, Placeholder, Align);
2996 return AggValueSlot::forAddr(Address(Placeholder, Align),
2998 AggValueSlot::IsNotDestructed,
2999 AggValueSlot::DoesNotNeedGCBarriers,
3000 AggValueSlot::IsNotAliased);
3003 void CodeGenFunction::EmitDelegateCallArg(CallArgList &args,
3004 const VarDecl *param,
3005 SourceLocation loc) {
3006 // StartFunction converted the ABI-lowered parameter(s) into a
3007 // local alloca. We need to turn that into an r-value suitable
3009 Address local = GetAddrOfLocalVar(param);
3011 QualType type = param->getType();
3013 assert(!isInAllocaArgument(CGM.getCXXABI(), type) &&
3014 "cannot emit delegate call arguments for inalloca arguments!");
3016 // GetAddrOfLocalVar returns a pointer-to-pointer for references,
3017 // but the argument needs to be the original pointer.
3018 if (type->isReferenceType()) {
3019 args.add(RValue::get(Builder.CreateLoad(local)), type);
3021 // In ARC, move out of consumed arguments so that the release cleanup
3022 // entered by StartFunction doesn't cause an over-release. This isn't
3023 // optimal -O0 code generation, but it should get cleaned up when
3024 // optimization is enabled. This also assumes that delegate calls are
3025 // performed exactly once for a set of arguments, but that should be safe.
3026 } else if (getLangOpts().ObjCAutoRefCount &&
3027 param->hasAttr<NSConsumedAttr>() &&
3028 type->isObjCRetainableType()) {
3029 llvm::Value *ptr = Builder.CreateLoad(local);
3031 llvm::ConstantPointerNull::get(cast<llvm::PointerType>(ptr->getType()));
3032 Builder.CreateStore(null, local);
3033 args.add(RValue::get(ptr), type);
3035 // For the most part, we just need to load the alloca, except that
3036 // aggregate r-values are actually pointers to temporaries.
3038 args.add(convertTempToRValue(local, type, loc), type);
3042 static bool isProvablyNull(llvm::Value *addr) {
3043 return isa<llvm::ConstantPointerNull>(addr);
3046 /// Emit the actual writing-back of a writeback.
3047 static void emitWriteback(CodeGenFunction &CGF,
3048 const CallArgList::Writeback &writeback) {
3049 const LValue &srcLV = writeback.Source;
3050 Address srcAddr = srcLV.getAddress();
3051 assert(!isProvablyNull(srcAddr.getPointer()) &&
3052 "shouldn't have writeback for provably null argument");
3054 llvm::BasicBlock *contBB = nullptr;
3056 // If the argument wasn't provably non-null, we need to null check
3057 // before doing the store.
3058 bool provablyNonNull = llvm::isKnownNonNull(srcAddr.getPointer());
3059 if (!provablyNonNull) {
3060 llvm::BasicBlock *writebackBB = CGF.createBasicBlock("icr.writeback");
3061 contBB = CGF.createBasicBlock("icr.done");
3063 llvm::Value *isNull =
3064 CGF.Builder.CreateIsNull(srcAddr.getPointer(), "icr.isnull");
3065 CGF.Builder.CreateCondBr(isNull, contBB, writebackBB);
3066 CGF.EmitBlock(writebackBB);
3069 // Load the value to writeback.
3070 llvm::Value *value = CGF.Builder.CreateLoad(writeback.Temporary);
3072 // Cast it back, in case we're writing an id to a Foo* or something.
3073 value = CGF.Builder.CreateBitCast(value, srcAddr.getElementType(),
3074 "icr.writeback-cast");
3076 // Perform the writeback.
3078 // If we have a "to use" value, it's something we need to emit a use
3079 // of. This has to be carefully threaded in: if it's done after the
3080 // release it's potentially undefined behavior (and the optimizer
3081 // will ignore it), and if it happens before the retain then the
3082 // optimizer could move the release there.
3083 if (writeback.ToUse) {
3084 assert(srcLV.getObjCLifetime() == Qualifiers::OCL_Strong);
3086 // Retain the new value. No need to block-copy here: the block's
3087 // being passed up the stack.
3088 value = CGF.EmitARCRetainNonBlock(value);
3090 // Emit the intrinsic use here.
3091 CGF.EmitARCIntrinsicUse(writeback.ToUse);
3093 // Load the old value (primitively).
3094 llvm::Value *oldValue = CGF.EmitLoadOfScalar(srcLV, SourceLocation());
3096 // Put the new value in place (primitively).
3097 CGF.EmitStoreOfScalar(value, srcLV, /*init*/ false);
3099 // Release the old value.
3100 CGF.EmitARCRelease(oldValue, srcLV.isARCPreciseLifetime());
3102 // Otherwise, we can just do a normal lvalue store.
3104 CGF.EmitStoreThroughLValue(RValue::get(value), srcLV);
3107 // Jump to the continuation block.
3108 if (!provablyNonNull)
3109 CGF.EmitBlock(contBB);
3112 static void emitWritebacks(CodeGenFunction &CGF,
3113 const CallArgList &args) {
3114 for (const auto &I : args.writebacks())
3115 emitWriteback(CGF, I);
3118 static void deactivateArgCleanupsBeforeCall(CodeGenFunction &CGF,
3119 const CallArgList &CallArgs) {
3120 assert(CGF.getTarget().getCXXABI().areArgsDestroyedLeftToRightInCallee());
3121 ArrayRef<CallArgList::CallArgCleanup> Cleanups =
3122 CallArgs.getCleanupsToDeactivate();
3123 // Iterate in reverse to increase the likelihood of popping the cleanup.
3124 for (const auto &I : llvm::reverse(Cleanups)) {
3125 CGF.DeactivateCleanupBlock(I.Cleanup, I.IsActiveIP);
3126 I.IsActiveIP->eraseFromParent();
3130 static const Expr *maybeGetUnaryAddrOfOperand(const Expr *E) {
3131 if (const UnaryOperator *uop = dyn_cast<UnaryOperator>(E->IgnoreParens()))
3132 if (uop->getOpcode() == UO_AddrOf)
3133 return uop->getSubExpr();
3137 /// Emit an argument that's being passed call-by-writeback. That is,
3138 /// we are passing the address of an __autoreleased temporary; it
3139 /// might be copy-initialized with the current value of the given
3140 /// address, but it will definitely be copied out of after the call.
3141 static void emitWritebackArg(CodeGenFunction &CGF, CallArgList &args,
3142 const ObjCIndirectCopyRestoreExpr *CRE) {
3145 // Make an optimistic effort to emit the address as an l-value.
3146 // This can fail if the argument expression is more complicated.
3147 if (const Expr *lvExpr = maybeGetUnaryAddrOfOperand(CRE->getSubExpr())) {
3148 srcLV = CGF.EmitLValue(lvExpr);
3150 // Otherwise, just emit it as a scalar.
3152 Address srcAddr = CGF.EmitPointerWithAlignment(CRE->getSubExpr());
3154 QualType srcAddrType =
3155 CRE->getSubExpr()->getType()->castAs<PointerType>()->getPointeeType();
3156 srcLV = CGF.MakeAddrLValue(srcAddr, srcAddrType);
3158 Address srcAddr = srcLV.getAddress();
3160 // The dest and src types don't necessarily match in LLVM terms
3161 // because of the crazy ObjC compatibility rules.
3163 llvm::PointerType *destType =
3164 cast<llvm::PointerType>(CGF.ConvertType(CRE->getType()));
3166 // If the address is a constant null, just pass the appropriate null.
3167 if (isProvablyNull(srcAddr.getPointer())) {
3168 args.add(RValue::get(llvm::ConstantPointerNull::get(destType)),
3173 // Create the temporary.
3174 Address temp = CGF.CreateTempAlloca(destType->getElementType(),
3175 CGF.getPointerAlign(),
3177 // Loading an l-value can introduce a cleanup if the l-value is __weak,
3178 // and that cleanup will be conditional if we can't prove that the l-value
3179 // isn't null, so we need to register a dominating point so that the cleanups
3180 // system will make valid IR.
3181 CodeGenFunction::ConditionalEvaluation condEval(CGF);
3183 // Zero-initialize it if we're not doing a copy-initialization.
3184 bool shouldCopy = CRE->shouldCopy();
3187 llvm::ConstantPointerNull::get(
3188 cast<llvm::PointerType>(destType->getElementType()));
3189 CGF.Builder.CreateStore(null, temp);
3192 llvm::BasicBlock *contBB = nullptr;
3193 llvm::BasicBlock *originBB = nullptr;
3195 // If the address is *not* known to be non-null, we need to switch.
3196 llvm::Value *finalArgument;
3198 bool provablyNonNull = llvm::isKnownNonNull(srcAddr.getPointer());
3199 if (provablyNonNull) {
3200 finalArgument = temp.getPointer();
3202 llvm::Value *isNull =
3203 CGF.Builder.CreateIsNull(srcAddr.getPointer(), "icr.isnull");
3205 finalArgument = CGF.Builder.CreateSelect(isNull,
3206 llvm::ConstantPointerNull::get(destType),
3207 temp.getPointer(), "icr.argument");
3209 // If we need to copy, then the load has to be conditional, which
3210 // means we need control flow.
3212 originBB = CGF.Builder.GetInsertBlock();
3213 contBB = CGF.createBasicBlock("icr.cont");
3214 llvm::BasicBlock *copyBB = CGF.createBasicBlock("icr.copy");
3215 CGF.Builder.CreateCondBr(isNull, contBB, copyBB);
3216 CGF.EmitBlock(copyBB);
3217 condEval.begin(CGF);
3221 llvm::Value *valueToUse = nullptr;
3223 // Perform a copy if necessary.
3225 RValue srcRV = CGF.EmitLoadOfLValue(srcLV, SourceLocation());
3226 assert(srcRV.isScalar());
3228 llvm::Value *src = srcRV.getScalarVal();
3229 src = CGF.Builder.CreateBitCast(src, destType->getElementType(),
3232 // Use an ordinary store, not a store-to-lvalue.
3233 CGF.Builder.CreateStore(src, temp);
3235 // If optimization is enabled, and the value was held in a
3236 // __strong variable, we need to tell the optimizer that this
3237 // value has to stay alive until we're doing the store back.
3238 // This is because the temporary is effectively unretained,
3239 // and so otherwise we can violate the high-level semantics.
3240 if (CGF.CGM.getCodeGenOpts().OptimizationLevel != 0 &&
3241 srcLV.getObjCLifetime() == Qualifiers::OCL_Strong) {
3246 // Finish the control flow if we needed it.
3247 if (shouldCopy && !provablyNonNull) {
3248 llvm::BasicBlock *copyBB = CGF.Builder.GetInsertBlock();
3249 CGF.EmitBlock(contBB);
3251 // Make a phi for the value to intrinsically use.
3253 llvm::PHINode *phiToUse = CGF.Builder.CreatePHI(valueToUse->getType(), 2,
3255 phiToUse->addIncoming(valueToUse, copyBB);
3256 phiToUse->addIncoming(llvm::UndefValue::get(valueToUse->getType()),
3258 valueToUse = phiToUse;
3264 args.addWriteback(srcLV, temp, valueToUse);
3265 args.add(RValue::get(finalArgument), CRE->getType());
3268 void CallArgList::allocateArgumentMemory(CodeGenFunction &CGF) {
3272 llvm::Function *F = CGF.CGM.getIntrinsic(llvm::Intrinsic::stacksave);
3273 StackBase = CGF.Builder.CreateCall(F, {}, "inalloca.save");
3276 void CallArgList::freeArgumentMemory(CodeGenFunction &CGF) const {
3278 // Restore the stack after the call.
3279 llvm::Value *F = CGF.CGM.getIntrinsic(llvm::Intrinsic::stackrestore);
3280 CGF.Builder.CreateCall(F, StackBase);
3284 void CodeGenFunction::EmitNonNullArgCheck(RValue RV, QualType ArgType,
3285 SourceLocation ArgLoc,
3288 if (!AC.getDecl() || !(SanOpts.has(SanitizerKind::NonnullAttribute) ||
3289 SanOpts.has(SanitizerKind::NullabilityArg)))
3292 // The param decl may be missing in a variadic function.
3293 auto PVD = ParmNum < AC.getNumParams() ? AC.getParamDecl(ParmNum) : nullptr;
3294 unsigned ArgNo = PVD ? PVD->getFunctionScopeIndex() : ParmNum;
3296 // Prefer the nonnull attribute if it's present.
3297 const NonNullAttr *NNAttr = nullptr;
3298 if (SanOpts.has(SanitizerKind::NonnullAttribute))
3299 NNAttr = getNonNullAttr(AC.getDecl(), PVD, ArgType, ArgNo);
3301 bool CanCheckNullability = false;
3302 if (SanOpts.has(SanitizerKind::NullabilityArg) && !NNAttr && PVD) {
3303 auto Nullability = PVD->getType()->getNullability(getContext());
3304 CanCheckNullability = Nullability &&
3305 *Nullability == NullabilityKind::NonNull &&
3306 PVD->getTypeSourceInfo();
3309 if (!NNAttr && !CanCheckNullability)
3312 SourceLocation AttrLoc;
3313 SanitizerMask CheckKind;
3314 SanitizerHandler Handler;
3316 AttrLoc = NNAttr->getLocation();
3317 CheckKind = SanitizerKind::NonnullAttribute;
3318 Handler = SanitizerHandler::NonnullArg;
3320 AttrLoc = PVD->getTypeSourceInfo()->getTypeLoc().findNullabilityLoc();
3321 CheckKind = SanitizerKind::NullabilityArg;
3322 Handler = SanitizerHandler::NullabilityArg;
3325 SanitizerScope SanScope(this);
3326 assert(RV.isScalar());
3327 llvm::Value *V = RV.getScalarVal();
3329 Builder.CreateICmpNE(V, llvm::Constant::getNullValue(V->getType()));
3330 llvm::Constant *StaticData[] = {
3331 EmitCheckSourceLocation(ArgLoc), EmitCheckSourceLocation(AttrLoc),
3332 llvm::ConstantInt::get(Int32Ty, ArgNo + 1),
3334 EmitCheck(std::make_pair(Cond, CheckKind), Handler, StaticData, None);
3337 void CodeGenFunction::EmitCallArgs(
3338 CallArgList &Args, ArrayRef<QualType> ArgTypes,
3339 llvm::iterator_range<CallExpr::const_arg_iterator> ArgRange,
3340 AbstractCallee AC, unsigned ParamsToSkip, EvaluationOrder Order) {
3341 assert((int)ArgTypes.size() == (ArgRange.end() - ArgRange.begin()));
3343 // We *have* to evaluate arguments from right to left in the MS C++ ABI,
3344 // because arguments are destroyed left to right in the callee. As a special
3345 // case, there are certain language constructs that require left-to-right
3346 // evaluation, and in those cases we consider the evaluation order requirement
3347 // to trump the "destruction order is reverse construction order" guarantee.
3349 CGM.getTarget().getCXXABI().areArgsDestroyedLeftToRightInCallee()
3350 ? Order == EvaluationOrder::ForceLeftToRight
3351 : Order != EvaluationOrder::ForceRightToLeft;
3353 auto MaybeEmitImplicitObjectSize = [&](unsigned I, const Expr *Arg,
3354 RValue EmittedArg) {
3355 if (!AC.hasFunctionDecl() || I >= AC.getNumParams())
3357 auto *PS = AC.getParamDecl(I)->getAttr<PassObjectSizeAttr>();
3361 const auto &Context = getContext();
3362 auto SizeTy = Context.getSizeType();
3363 auto T = Builder.getIntNTy(Context.getTypeSize(SizeTy));
3364 assert(EmittedArg.getScalarVal() && "We emitted nothing for the arg?");
3365 llvm::Value *V = evaluateOrEmitBuiltinObjectSize(Arg, PS->getType(), T,
3366 EmittedArg.getScalarVal());
3367 Args.add(RValue::get(V), SizeTy);
3368 // If we're emitting args in reverse, be sure to do so with
3369 // pass_object_size, as well.
3371 std::swap(Args.back(), *(&Args.back() - 1));
3374 // Insert a stack save if we're going to need any inalloca args.
3375 bool HasInAllocaArgs = false;
3376 if (CGM.getTarget().getCXXABI().isMicrosoft()) {
3377 for (ArrayRef<QualType>::iterator I = ArgTypes.begin(), E = ArgTypes.end();
3378 I != E && !HasInAllocaArgs; ++I)
3379 HasInAllocaArgs = isInAllocaArgument(CGM.getCXXABI(), *I);
3380 if (HasInAllocaArgs) {
3381 assert(getTarget().getTriple().getArch() == llvm::Triple::x86);
3382 Args.allocateArgumentMemory(*this);
3386 // Evaluate each argument in the appropriate order.
3387 size_t CallArgsStart = Args.size();
3388 for (unsigned I = 0, E = ArgTypes.size(); I != E; ++I) {
3389 unsigned Idx = LeftToRight ? I : E - I - 1;
3390 CallExpr::const_arg_iterator Arg = ArgRange.begin() + Idx;
3391 unsigned InitialArgSize = Args.size();
3392 EmitCallArg(Args, *Arg, ArgTypes[Idx]);
3393 // In particular, we depend on it being the last arg in Args, and the
3394 // objectsize bits depend on there only being one arg if !LeftToRight.
3395 assert(InitialArgSize + 1 == Args.size() &&
3396 "The code below depends on only adding one arg per EmitCallArg");
3397 (void)InitialArgSize;
3398 RValue RVArg = Args.back().RV;
3399 EmitNonNullArgCheck(RVArg, ArgTypes[Idx], (*Arg)->getExprLoc(), AC,
3400 ParamsToSkip + Idx);
3401 // @llvm.objectsize should never have side-effects and shouldn't need
3402 // destruction/cleanups, so we can safely "emit" it after its arg,
3403 // regardless of right-to-leftness
3404 MaybeEmitImplicitObjectSize(Idx, *Arg, RVArg);
3408 // Un-reverse the arguments we just evaluated so they match up with the LLVM
3410 std::reverse(Args.begin() + CallArgsStart, Args.end());
3416 struct DestroyUnpassedArg final : EHScopeStack::Cleanup {
3417 DestroyUnpassedArg(Address Addr, QualType Ty)
3418 : Addr(Addr), Ty(Ty) {}
3423 void Emit(CodeGenFunction &CGF, Flags flags) override {
3424 const CXXDestructorDecl *Dtor = Ty->getAsCXXRecordDecl()->getDestructor();
3425 assert(!Dtor->isTrivial());
3426 CGF.EmitCXXDestructorCall(Dtor, Dtor_Complete, /*for vbase*/ false,
3427 /*Delegating=*/false, Addr);
3431 struct DisableDebugLocationUpdates {
3432 CodeGenFunction &CGF;
3433 bool disabledDebugInfo;
3434 DisableDebugLocationUpdates(CodeGenFunction &CGF, const Expr *E) : CGF(CGF) {
3435 if ((disabledDebugInfo = isa<CXXDefaultArgExpr>(E) && CGF.getDebugInfo()))
3436 CGF.disableDebugInfo();
3438 ~DisableDebugLocationUpdates() {
3439 if (disabledDebugInfo)
3440 CGF.enableDebugInfo();
3444 } // end anonymous namespace
3446 void CodeGenFunction::EmitCallArg(CallArgList &args, const Expr *E,
3448 DisableDebugLocationUpdates Dis(*this, E);
3449 if (const ObjCIndirectCopyRestoreExpr *CRE
3450 = dyn_cast<ObjCIndirectCopyRestoreExpr>(E)) {
3451 assert(getLangOpts().ObjCAutoRefCount);
3452 assert(getContext().hasSameUnqualifiedType(E->getType(), type));
3453 return emitWritebackArg(*this, args, CRE);
3456 assert(type->isReferenceType() == E->isGLValue() &&
3457 "reference binding to unmaterialized r-value!");
3459 if (E->isGLValue()) {
3460 assert(E->getObjectKind() == OK_Ordinary);
3461 return args.add(EmitReferenceBindingToExpr(E), type);
3464 bool HasAggregateEvalKind = hasAggregateEvaluationKind(type);
3466 // In the Microsoft C++ ABI, aggregate arguments are destructed by the callee.
3467 // However, we still have to push an EH-only cleanup in case we unwind before
3468 // we make it to the call.
3469 if (HasAggregateEvalKind &&
3470 CGM.getTarget().getCXXABI().areArgsDestroyedLeftToRightInCallee()) {
3471 // If we're using inalloca, use the argument memory. Otherwise, use a
3474 if (args.isUsingInAlloca())
3475 Slot = createPlaceholderSlot(*this, type);
3477 Slot = CreateAggTemp(type, "agg.tmp");
3479 const CXXRecordDecl *RD = type->getAsCXXRecordDecl();
3480 bool DestroyedInCallee =
3481 RD && RD->hasNonTrivialDestructor() &&
3482 CGM.getCXXABI().getRecordArgABI(RD) != CGCXXABI::RAA_Default;
3483 if (DestroyedInCallee)
3484 Slot.setExternallyDestructed();
3486 EmitAggExpr(E, Slot);
3487 RValue RV = Slot.asRValue();
3490 if (DestroyedInCallee) {
3491 // Create a no-op GEP between the placeholder and the cleanup so we can
3492 // RAUW it successfully. It also serves as a marker of the first
3493 // instruction where the cleanup is active.
3494 pushFullExprCleanup<DestroyUnpassedArg>(EHCleanup, Slot.getAddress(),
3496 // This unreachable is a temporary marker which will be removed later.
3497 llvm::Instruction *IsActive = Builder.CreateUnreachable();
3498 args.addArgCleanupDeactivation(EHStack.getInnermostEHScope(), IsActive);
3503 if (HasAggregateEvalKind && isa<ImplicitCastExpr>(E) &&
3504 cast<CastExpr>(E)->getCastKind() == CK_LValueToRValue) {
3505 LValue L = EmitLValue(cast<CastExpr>(E)->getSubExpr());
3506 assert(L.isSimple());
3507 if (L.getAlignment() >= getContext().getTypeAlignInChars(type)) {
3508 args.add(L.asAggregateRValue(), type, /*NeedsCopy*/true);
3510 // We can't represent a misaligned lvalue in the CallArgList, so copy
3511 // to an aligned temporary now.
3512 Address tmp = CreateMemTemp(type);
3513 EmitAggregateCopy(tmp, L.getAddress(), type, L.isVolatile());
3514 args.add(RValue::getAggregate(tmp), type);
3519 args.add(EmitAnyExprToTemp(E), type);
3522 QualType CodeGenFunction::getVarArgType(const Expr *Arg) {
3523 // System headers on Windows define NULL to 0 instead of 0LL on Win64. MSVC
3524 // implicitly widens null pointer constants that are arguments to varargs
3525 // functions to pointer-sized ints.
3526 if (!getTarget().getTriple().isOSWindows())
3527 return Arg->getType();
3529 if (Arg->getType()->isIntegerType() &&
3530 getContext().getTypeSize(Arg->getType()) <
3531 getContext().getTargetInfo().getPointerWidth(0) &&
3532 Arg->isNullPointerConstant(getContext(),
3533 Expr::NPC_ValueDependentIsNotNull)) {
3534 return getContext().getIntPtrType();
3537 return Arg->getType();
3540 // In ObjC ARC mode with no ObjC ARC exception safety, tell the ARC
3541 // optimizer it can aggressively ignore unwind edges.
3543 CodeGenFunction::AddObjCARCExceptionMetadata(llvm::Instruction *Inst) {
3544 if (CGM.getCodeGenOpts().OptimizationLevel != 0 &&
3545 !CGM.getCodeGenOpts().ObjCAutoRefCountExceptions)
3546 Inst->setMetadata("clang.arc.no_objc_arc_exceptions",
3547 CGM.getNoObjCARCExceptionsMetadata());
3550 /// Emits a call to the given no-arguments nounwind runtime function.
3552 CodeGenFunction::EmitNounwindRuntimeCall(llvm::Value *callee,
3553 const llvm::Twine &name) {
3554 return EmitNounwindRuntimeCall(callee, None, name);
3557 /// Emits a call to the given nounwind runtime function.
3559 CodeGenFunction::EmitNounwindRuntimeCall(llvm::Value *callee,
3560 ArrayRef<llvm::Value*> args,
3561 const llvm::Twine &name) {
3562 llvm::CallInst *call = EmitRuntimeCall(callee, args, name);
3563 call->setDoesNotThrow();
3567 /// Emits a simple call (never an invoke) to the given no-arguments
3568 /// runtime function.
3570 CodeGenFunction::EmitRuntimeCall(llvm::Value *callee,
3571 const llvm::Twine &name) {
3572 return EmitRuntimeCall(callee, None, name);
3575 // Calls which may throw must have operand bundles indicating which funclet
3576 // they are nested within.
3578 getBundlesForFunclet(llvm::Value *Callee, llvm::Instruction *CurrentFuncletPad,
3579 SmallVectorImpl<llvm::OperandBundleDef> &BundleList) {
3580 // There is no need for a funclet operand bundle if we aren't inside a
3582 if (!CurrentFuncletPad)
3585 // Skip intrinsics which cannot throw.
3586 auto *CalleeFn = dyn_cast<llvm::Function>(Callee->stripPointerCasts());
3587 if (CalleeFn && CalleeFn->isIntrinsic() && CalleeFn->doesNotThrow())
3590 BundleList.emplace_back("funclet", CurrentFuncletPad);
3593 /// Emits a simple call (never an invoke) to the given runtime function.
3595 CodeGenFunction::EmitRuntimeCall(llvm::Value *callee,
3596 ArrayRef<llvm::Value*> args,
3597 const llvm::Twine &name) {
3598 SmallVector<llvm::OperandBundleDef, 1> BundleList;
3599 getBundlesForFunclet(callee, CurrentFuncletPad, BundleList);
3601 llvm::CallInst *call = Builder.CreateCall(callee, args, BundleList, name);
3602 call->setCallingConv(getRuntimeCC());
3606 /// Emits a call or invoke to the given noreturn runtime function.
3607 void CodeGenFunction::EmitNoreturnRuntimeCallOrInvoke(llvm::Value *callee,
3608 ArrayRef<llvm::Value*> args) {
3609 SmallVector<llvm::OperandBundleDef, 1> BundleList;
3610 getBundlesForFunclet(callee, CurrentFuncletPad, BundleList);
3612 if (getInvokeDest()) {
3613 llvm::InvokeInst *invoke =
3614 Builder.CreateInvoke(callee,
3615 getUnreachableBlock(),
3619 invoke->setDoesNotReturn();
3620 invoke->setCallingConv(getRuntimeCC());
3622 llvm::CallInst *call = Builder.CreateCall(callee, args, BundleList);
3623 call->setDoesNotReturn();
3624 call->setCallingConv(getRuntimeCC());
3625 Builder.CreateUnreachable();
3629 /// Emits a call or invoke instruction to the given nullary runtime function.
3631 CodeGenFunction::EmitRuntimeCallOrInvoke(llvm::Value *callee,
3632 const Twine &name) {
3633 return EmitRuntimeCallOrInvoke(callee, None, name);
3636 /// Emits a call or invoke instruction to the given runtime function.
3638 CodeGenFunction::EmitRuntimeCallOrInvoke(llvm::Value *callee,
3639 ArrayRef<llvm::Value*> args,
3640 const Twine &name) {
3641 llvm::CallSite callSite = EmitCallOrInvoke(callee, args, name);
3642 callSite.setCallingConv(getRuntimeCC());
3646 /// Emits a call or invoke instruction to the given function, depending
3647 /// on the current state of the EH stack.
3649 CodeGenFunction::EmitCallOrInvoke(llvm::Value *Callee,
3650 ArrayRef<llvm::Value *> Args,
3651 const Twine &Name) {
3652 llvm::BasicBlock *InvokeDest = getInvokeDest();
3653 SmallVector<llvm::OperandBundleDef, 1> BundleList;
3654 getBundlesForFunclet(Callee, CurrentFuncletPad, BundleList);
3656 llvm::Instruction *Inst;
3658 Inst = Builder.CreateCall(Callee, Args, BundleList, Name);
3660 llvm::BasicBlock *ContBB = createBasicBlock("invoke.cont");
3661 Inst = Builder.CreateInvoke(Callee, ContBB, InvokeDest, Args, BundleList,
3666 // In ObjC ARC mode with no ObjC ARC exception safety, tell the ARC
3667 // optimizer it can aggressively ignore unwind edges.
3668 if (CGM.getLangOpts().ObjCAutoRefCount)
3669 AddObjCARCExceptionMetadata(Inst);
3671 return llvm::CallSite(Inst);
3674 /// \brief Store a non-aggregate value to an address to initialize it. For
3675 /// initialization, a non-atomic store will be used.
3676 static void EmitInitStoreOfNonAggregate(CodeGenFunction &CGF, RValue Src,
3679 CGF.EmitStoreOfScalar(Src.getScalarVal(), Dst, /*init=*/true);
3681 CGF.EmitStoreOfComplex(Src.getComplexVal(), Dst, /*init=*/true);
3684 void CodeGenFunction::deferPlaceholderReplacement(llvm::Instruction *Old,
3686 DeferredReplacements.push_back(std::make_pair(Old, New));
3689 RValue CodeGenFunction::EmitCall(const CGFunctionInfo &CallInfo,
3690 const CGCallee &Callee,
3691 ReturnValueSlot ReturnValue,
3692 const CallArgList &CallArgs,
3693 llvm::Instruction **callOrInvoke) {
3694 // FIXME: We no longer need the types from CallArgs; lift up and simplify.
3696 assert(Callee.isOrdinary());
3698 // Handle struct-return functions by passing a pointer to the
3699 // location that we would like to return into.
3700 QualType RetTy = CallInfo.getReturnType();
3701 const ABIArgInfo &RetAI = CallInfo.getReturnInfo();
3703 llvm::FunctionType *IRFuncTy = Callee.getFunctionType();
3705 // 1. Set up the arguments.
3707 // If we're using inalloca, insert the allocation after the stack save.
3708 // FIXME: Do this earlier rather than hacking it in here!
3709 Address ArgMemory = Address::invalid();
3710 const llvm::StructLayout *ArgMemoryLayout = nullptr;
3711 if (llvm::StructType *ArgStruct = CallInfo.getArgStruct()) {
3712 const llvm::DataLayout &DL = CGM.getDataLayout();
3713 ArgMemoryLayout = DL.getStructLayout(ArgStruct);
3714 llvm::Instruction *IP = CallArgs.getStackBase();
3715 llvm::AllocaInst *AI;
3717 IP = IP->getNextNode();
3718 AI = new llvm::AllocaInst(ArgStruct, DL.getAllocaAddrSpace(),
3721 AI = CreateTempAlloca(ArgStruct, "argmem");
3723 auto Align = CallInfo.getArgStructAlignment();
3724 AI->setAlignment(Align.getQuantity());
3725 AI->setUsedWithInAlloca(true);
3726 assert(AI->isUsedWithInAlloca() && !AI->isStaticAlloca());
3727 ArgMemory = Address(AI, Align);
3730 // Helper function to drill into the inalloca allocation.
3731 auto createInAllocaStructGEP = [&](unsigned FieldIndex) -> Address {
3733 CharUnits::fromQuantity(ArgMemoryLayout->getElementOffset(FieldIndex));
3734 return Builder.CreateStructGEP(ArgMemory, FieldIndex, FieldOffset);
3737 ClangToLLVMArgMapping IRFunctionArgs(CGM.getContext(), CallInfo);
3738 SmallVector<llvm::Value *, 16> IRCallArgs(IRFunctionArgs.totalIRArgs());
3740 // If the call returns a temporary with struct return, create a temporary
3741 // alloca to hold the result, unless one is given to us.
3742 Address SRetPtr = Address::invalid();
3743 size_t UnusedReturnSize = 0;
3744 if (RetAI.isIndirect() || RetAI.isInAlloca() || RetAI.isCoerceAndExpand()) {
3745 if (!ReturnValue.isNull()) {
3746 SRetPtr = ReturnValue.getValue();
3748 SRetPtr = CreateMemTemp(RetTy);
3749 if (HaveInsertPoint() && ReturnValue.isUnused()) {
3751 CGM.getDataLayout().getTypeAllocSize(ConvertTypeForMem(RetTy));
3752 if (EmitLifetimeStart(size, SRetPtr.getPointer()))
3753 UnusedReturnSize = size;
3756 if (IRFunctionArgs.hasSRetArg()) {
3757 IRCallArgs[IRFunctionArgs.getSRetArgNo()] = SRetPtr.getPointer();
3758 } else if (RetAI.isInAlloca()) {
3759 Address Addr = createInAllocaStructGEP(RetAI.getInAllocaFieldIndex());
3760 Builder.CreateStore(SRetPtr.getPointer(), Addr);
3764 Address swiftErrorTemp = Address::invalid();
3765 Address swiftErrorArg = Address::invalid();
3767 // Translate all of the arguments as necessary to match the IR lowering.
3768 assert(CallInfo.arg_size() == CallArgs.size() &&
3769 "Mismatch between function signature & arguments.");
3771 CGFunctionInfo::const_arg_iterator info_it = CallInfo.arg_begin();
3772 for (CallArgList::const_iterator I = CallArgs.begin(), E = CallArgs.end();
3773 I != E; ++I, ++info_it, ++ArgNo) {
3774 const ABIArgInfo &ArgInfo = info_it->info;
3777 // Insert a padding argument to ensure proper alignment.
3778 if (IRFunctionArgs.hasPaddingArg(ArgNo))
3779 IRCallArgs[IRFunctionArgs.getPaddingArgNo(ArgNo)] =
3780 llvm::UndefValue::get(ArgInfo.getPaddingType());
3782 unsigned FirstIRArg, NumIRArgs;
3783 std::tie(FirstIRArg, NumIRArgs) = IRFunctionArgs.getIRArgs(ArgNo);
3785 switch (ArgInfo.getKind()) {
3786 case ABIArgInfo::InAlloca: {
3787 assert(NumIRArgs == 0);
3788 assert(getTarget().getTriple().getArch() == llvm::Triple::x86);
3789 if (RV.isAggregate()) {
3790 // Replace the placeholder with the appropriate argument slot GEP.
3791 llvm::Instruction *Placeholder =
3792 cast<llvm::Instruction>(RV.getAggregatePointer());
3793 CGBuilderTy::InsertPoint IP = Builder.saveIP();
3794 Builder.SetInsertPoint(Placeholder);
3795 Address Addr = createInAllocaStructGEP(ArgInfo.getInAllocaFieldIndex());
3796 Builder.restoreIP(IP);
3797 deferPlaceholderReplacement(Placeholder, Addr.getPointer());
3799 // Store the RValue into the argument struct.
3800 Address Addr = createInAllocaStructGEP(ArgInfo.getInAllocaFieldIndex());
3801 unsigned AS = Addr.getType()->getPointerAddressSpace();
3802 llvm::Type *MemType = ConvertTypeForMem(I->Ty)->getPointerTo(AS);
3803 // There are some cases where a trivial bitcast is not avoidable. The
3804 // definition of a type later in a translation unit may change it's type
3805 // from {}* to (%struct.foo*)*.
3806 if (Addr.getType() != MemType)
3807 Addr = Builder.CreateBitCast(Addr, MemType);
3808 LValue argLV = MakeAddrLValue(Addr, I->Ty);
3809 EmitInitStoreOfNonAggregate(*this, RV, argLV);
3814 case ABIArgInfo::Indirect: {
3815 assert(NumIRArgs == 1);
3816 if (RV.isScalar() || RV.isComplex()) {
3817 // Make a temporary alloca to pass the argument.
3818 Address Addr = CreateMemTemp(I->Ty, ArgInfo.getIndirectAlign(),
3819 "indirect-arg-temp", false);
3820 IRCallArgs[FirstIRArg] = Addr.getPointer();
3822 LValue argLV = MakeAddrLValue(Addr, I->Ty);
3823 EmitInitStoreOfNonAggregate(*this, RV, argLV);
3825 // We want to avoid creating an unnecessary temporary+copy here;
3826 // however, we need one in three cases:
3827 // 1. If the argument is not byval, and we are required to copy the
3828 // source. (This case doesn't occur on any common architecture.)
3829 // 2. If the argument is byval, RV is not sufficiently aligned, and
3830 // we cannot force it to be sufficiently aligned.
3831 // 3. If the argument is byval, but RV is located in an address space
3832 // different than that of the argument (0).
3833 Address Addr = RV.getAggregateAddress();
3834 CharUnits Align = ArgInfo.getIndirectAlign();
3835 const llvm::DataLayout *TD = &CGM.getDataLayout();
3836 const unsigned RVAddrSpace = Addr.getType()->getAddressSpace();
3837 const unsigned ArgAddrSpace =
3838 (FirstIRArg < IRFuncTy->getNumParams()
3839 ? IRFuncTy->getParamType(FirstIRArg)->getPointerAddressSpace()
3841 if ((!ArgInfo.getIndirectByVal() && I->NeedsCopy) ||
3842 (ArgInfo.getIndirectByVal() && Addr.getAlignment() < Align &&
3843 llvm::getOrEnforceKnownAlignment(Addr.getPointer(),
3844 Align.getQuantity(), *TD)
3845 < Align.getQuantity()) ||
3846 (ArgInfo.getIndirectByVal() && (RVAddrSpace != ArgAddrSpace))) {
3847 // Create an aligned temporary, and copy to it.
3848 Address AI = CreateMemTemp(I->Ty, ArgInfo.getIndirectAlign(),
3849 "byval-temp", false);
3850 IRCallArgs[FirstIRArg] = AI.getPointer();
3851 EmitAggregateCopy(AI, Addr, I->Ty, RV.isVolatileQualified());
3853 // Skip the extra memcpy call.
3854 IRCallArgs[FirstIRArg] = Addr.getPointer();
3860 case ABIArgInfo::Ignore:
3861 assert(NumIRArgs == 0);
3864 case ABIArgInfo::Extend:
3865 case ABIArgInfo::Direct: {
3866 if (!isa<llvm::StructType>(ArgInfo.getCoerceToType()) &&
3867 ArgInfo.getCoerceToType() == ConvertType(info_it->type) &&
3868 ArgInfo.getDirectOffset() == 0) {
3869 assert(NumIRArgs == 1);
3872 V = RV.getScalarVal();
3874 V = Builder.CreateLoad(RV.getAggregateAddress());
3876 // Implement swifterror by copying into a new swifterror argument.
3877 // We'll write back in the normal path out of the call.
3878 if (CallInfo.getExtParameterInfo(ArgNo).getABI()
3879 == ParameterABI::SwiftErrorResult) {
3880 assert(!swiftErrorTemp.isValid() && "multiple swifterror args");
3882 QualType pointeeTy = I->Ty->getPointeeType();
3884 Address(V, getContext().getTypeAlignInChars(pointeeTy));
3887 CreateMemTemp(pointeeTy, getPointerAlign(), "swifterror.temp");
3888 V = swiftErrorTemp.getPointer();
3889 cast<llvm::AllocaInst>(V)->setSwiftError(true);
3891 llvm::Value *errorValue = Builder.CreateLoad(swiftErrorArg);
3892 Builder.CreateStore(errorValue, swiftErrorTemp);
3895 // We might have to widen integers, but we should never truncate.
3896 if (ArgInfo.getCoerceToType() != V->getType() &&
3897 V->getType()->isIntegerTy())
3898 V = Builder.CreateZExt(V, ArgInfo.getCoerceToType());
3900 // If the argument doesn't match, perform a bitcast to coerce it. This
3901 // can happen due to trivial type mismatches.
3902 if (FirstIRArg < IRFuncTy->getNumParams() &&
3903 V->getType() != IRFuncTy->getParamType(FirstIRArg))
3904 V = Builder.CreateBitCast(V, IRFuncTy->getParamType(FirstIRArg));
3906 IRCallArgs[FirstIRArg] = V;
3910 // FIXME: Avoid the conversion through memory if possible.
3911 Address Src = Address::invalid();
3912 if (RV.isScalar() || RV.isComplex()) {
3913 Src = CreateMemTemp(I->Ty, "coerce");
3914 LValue SrcLV = MakeAddrLValue(Src, I->Ty);
3915 EmitInitStoreOfNonAggregate(*this, RV, SrcLV);
3917 Src = RV.getAggregateAddress();
3920 // If the value is offset in memory, apply the offset now.
3921 Src = emitAddressAtOffset(*this, Src, ArgInfo);
3923 // Fast-isel and the optimizer generally like scalar values better than
3924 // FCAs, so we flatten them if this is safe to do for this argument.
3925 llvm::StructType *STy =
3926 dyn_cast<llvm::StructType>(ArgInfo.getCoerceToType());
3927 if (STy && ArgInfo.isDirect() && ArgInfo.getCanBeFlattened()) {
3928 llvm::Type *SrcTy = Src.getType()->getElementType();
3929 uint64_t SrcSize = CGM.getDataLayout().getTypeAllocSize(SrcTy);
3930 uint64_t DstSize = CGM.getDataLayout().getTypeAllocSize(STy);
3932 // If the source type is smaller than the destination type of the
3933 // coerce-to logic, copy the source value into a temp alloca the size
3934 // of the destination type to allow loading all of it. The bits past
3935 // the source value are left undef.
3936 if (SrcSize < DstSize) {
3938 = CreateTempAlloca(STy, Src.getAlignment(),
3939 Src.getName() + ".coerce");
3940 Builder.CreateMemCpy(TempAlloca, Src, SrcSize);
3943 Src = Builder.CreateBitCast(Src, llvm::PointerType::getUnqual(STy));
3946 auto SrcLayout = CGM.getDataLayout().getStructLayout(STy);
3947 assert(NumIRArgs == STy->getNumElements());
3948 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
3949 auto Offset = CharUnits::fromQuantity(SrcLayout->getElementOffset(i));
3950 Address EltPtr = Builder.CreateStructGEP(Src, i, Offset);
3951 llvm::Value *LI = Builder.CreateLoad(EltPtr);
3952 IRCallArgs[FirstIRArg + i] = LI;
3955 // In the simple case, just pass the coerced loaded value.
3956 assert(NumIRArgs == 1);
3957 IRCallArgs[FirstIRArg] =
3958 CreateCoercedLoad(Src, ArgInfo.getCoerceToType(), *this);
3964 case ABIArgInfo::CoerceAndExpand: {
3965 auto coercionType = ArgInfo.getCoerceAndExpandType();
3966 auto layout = CGM.getDataLayout().getStructLayout(coercionType);
3968 llvm::Value *tempSize = nullptr;
3969 Address addr = Address::invalid();
3970 if (RV.isAggregate()) {
3971 addr = RV.getAggregateAddress();
3973 assert(RV.isScalar()); // complex should always just be direct
3975 llvm::Type *scalarType = RV.getScalarVal()->getType();
3976 auto scalarSize = CGM.getDataLayout().getTypeAllocSize(scalarType);
3977 auto scalarAlign = CGM.getDataLayout().getPrefTypeAlignment(scalarType);
3979 tempSize = llvm::ConstantInt::get(CGM.Int64Ty, scalarSize);
3981 // Materialize to a temporary.
3982 addr = CreateTempAlloca(RV.getScalarVal()->getType(),
3983 CharUnits::fromQuantity(std::max(layout->getAlignment(),
3985 EmitLifetimeStart(scalarSize, addr.getPointer());
3987 Builder.CreateStore(RV.getScalarVal(), addr);
3990 addr = Builder.CreateElementBitCast(addr, coercionType);
3992 unsigned IRArgPos = FirstIRArg;
3993 for (unsigned i = 0, e = coercionType->getNumElements(); i != e; ++i) {
3994 llvm::Type *eltType = coercionType->getElementType(i);
3995 if (ABIArgInfo::isPaddingForCoerceAndExpand(eltType)) continue;
3996 Address eltAddr = Builder.CreateStructGEP(addr, i, layout);
3997 llvm::Value *elt = Builder.CreateLoad(eltAddr);
3998 IRCallArgs[IRArgPos++] = elt;
4000 assert(IRArgPos == FirstIRArg + NumIRArgs);
4003 EmitLifetimeEnd(tempSize, addr.getPointer());
4009 case ABIArgInfo::Expand:
4010 unsigned IRArgPos = FirstIRArg;
4011 ExpandTypeToArgs(I->Ty, RV, IRFuncTy, IRCallArgs, IRArgPos);
4012 assert(IRArgPos == FirstIRArg + NumIRArgs);
4017 llvm::Value *CalleePtr = Callee.getFunctionPointer();
4019 // If we're using inalloca, set up that argument.
4020 if (ArgMemory.isValid()) {
4021 llvm::Value *Arg = ArgMemory.getPointer();
4022 if (CallInfo.isVariadic()) {
4023 // When passing non-POD arguments by value to variadic functions, we will
4024 // end up with a variadic prototype and an inalloca call site. In such
4025 // cases, we can't do any parameter mismatch checks. Give up and bitcast
4027 unsigned CalleeAS = CalleePtr->getType()->getPointerAddressSpace();
4028 auto FnTy = getTypes().GetFunctionType(CallInfo)->getPointerTo(CalleeAS);
4029 CalleePtr = Builder.CreateBitCast(CalleePtr, FnTy);
4031 llvm::Type *LastParamTy =
4032 IRFuncTy->getParamType(IRFuncTy->getNumParams() - 1);
4033 if (Arg->getType() != LastParamTy) {
4035 // Assert that these structs have equivalent element types.
4036 llvm::StructType *FullTy = CallInfo.getArgStruct();
4037 llvm::StructType *DeclaredTy = cast<llvm::StructType>(
4038 cast<llvm::PointerType>(LastParamTy)->getElementType());
4039 assert(DeclaredTy->getNumElements() == FullTy->getNumElements());
4040 for (llvm::StructType::element_iterator DI = DeclaredTy->element_begin(),
4041 DE = DeclaredTy->element_end(),
4042 FI = FullTy->element_begin();
4043 DI != DE; ++DI, ++FI)
4046 Arg = Builder.CreateBitCast(Arg, LastParamTy);
4049 assert(IRFunctionArgs.hasInallocaArg());
4050 IRCallArgs[IRFunctionArgs.getInallocaArgNo()] = Arg;
4053 // 2. Prepare the function pointer.
4055 // If the callee is a bitcast of a non-variadic function to have a
4056 // variadic function pointer type, check to see if we can remove the
4057 // bitcast. This comes up with unprototyped functions.
4059 // This makes the IR nicer, but more importantly it ensures that we
4060 // can inline the function at -O0 if it is marked always_inline.
4061 auto simplifyVariadicCallee = [](llvm::Value *Ptr) -> llvm::Value* {
4062 llvm::FunctionType *CalleeFT =
4063 cast<llvm::FunctionType>(Ptr->getType()->getPointerElementType());
4064 if (!CalleeFT->isVarArg())
4067 llvm::ConstantExpr *CE = dyn_cast<llvm::ConstantExpr>(Ptr);
4068 if (!CE || CE->getOpcode() != llvm::Instruction::BitCast)
4071 llvm::Function *OrigFn = dyn_cast<llvm::Function>(CE->getOperand(0));
4075 llvm::FunctionType *OrigFT = OrigFn->getFunctionType();
4077 // If the original type is variadic, or if any of the component types
4078 // disagree, we cannot remove the cast.
4079 if (OrigFT->isVarArg() ||
4080 OrigFT->getNumParams() != CalleeFT->getNumParams() ||
4081 OrigFT->getReturnType() != CalleeFT->getReturnType())
4084 for (unsigned i = 0, e = OrigFT->getNumParams(); i != e; ++i)
4085 if (OrigFT->getParamType(i) != CalleeFT->getParamType(i))
4090 CalleePtr = simplifyVariadicCallee(CalleePtr);
4092 // 3. Perform the actual call.
4094 // Deactivate any cleanups that we're supposed to do immediately before
4096 if (!CallArgs.getCleanupsToDeactivate().empty())
4097 deactivateArgCleanupsBeforeCall(*this, CallArgs);
4099 // Assert that the arguments we computed match up. The IR verifier
4100 // will catch this, but this is a common enough source of problems
4101 // during IRGen changes that it's way better for debugging to catch
4102 // it ourselves here.
4104 assert(IRCallArgs.size() == IRFuncTy->getNumParams() || IRFuncTy->isVarArg());
4105 for (unsigned i = 0; i < IRCallArgs.size(); ++i) {
4106 // Inalloca argument can have different type.
4107 if (IRFunctionArgs.hasInallocaArg() &&
4108 i == IRFunctionArgs.getInallocaArgNo())
4110 if (i < IRFuncTy->getNumParams())
4111 assert(IRCallArgs[i]->getType() == IRFuncTy->getParamType(i));
4115 // Compute the calling convention and attributes.
4116 unsigned CallingConv;
4117 llvm::AttributeList Attrs;
4118 CGM.ConstructAttributeList(CalleePtr->getName(), CallInfo,
4119 Callee.getAbstractInfo(), Attrs, CallingConv,
4120 /*AttrOnCallSite=*/true);
4122 // Apply some call-site-specific attributes.
4123 // TODO: work this into building the attribute set.
4125 // Apply always_inline to all calls within flatten functions.
4126 // FIXME: should this really take priority over __try, below?
4127 if (CurCodeDecl && CurCodeDecl->hasAttr<FlattenAttr>() &&
4128 !(Callee.getAbstractInfo().getCalleeDecl() &&
4129 Callee.getAbstractInfo().getCalleeDecl()->hasAttr<NoInlineAttr>())) {
4131 Attrs.addAttribute(getLLVMContext(), llvm::AttributeList::FunctionIndex,
4132 llvm::Attribute::AlwaysInline);
4135 // Disable inlining inside SEH __try blocks.
4136 if (isSEHTryScope()) {
4138 Attrs.addAttribute(getLLVMContext(), llvm::AttributeList::FunctionIndex,
4139 llvm::Attribute::NoInline);
4142 // Decide whether to use a call or an invoke.
4144 if (currentFunctionUsesSEHTry()) {
4145 // SEH cares about asynchronous exceptions, so everything can "throw."
4146 CannotThrow = false;
4147 } else if (isCleanupPadScope() &&
4148 EHPersonality::get(*this).isMSVCXXPersonality()) {
4149 // The MSVC++ personality will implicitly terminate the program if an
4150 // exception is thrown during a cleanup outside of a try/catch.
4151 // We don't need to model anything in IR to get this behavior.
4154 // Otherwise, nounwind call sites will never throw.
4155 CannotThrow = Attrs.hasAttribute(llvm::AttributeList::FunctionIndex,
4156 llvm::Attribute::NoUnwind);
4158 llvm::BasicBlock *InvokeDest = CannotThrow ? nullptr : getInvokeDest();
4160 SmallVector<llvm::OperandBundleDef, 1> BundleList;
4161 getBundlesForFunclet(CalleePtr, CurrentFuncletPad, BundleList);
4163 // Emit the actual call/invoke instruction.
4166 CS = Builder.CreateCall(CalleePtr, IRCallArgs, BundleList);
4168 llvm::BasicBlock *Cont = createBasicBlock("invoke.cont");
4169 CS = Builder.CreateInvoke(CalleePtr, Cont, InvokeDest, IRCallArgs,
4173 llvm::Instruction *CI = CS.getInstruction();
4177 // Apply the attributes and calling convention.
4178 CS.setAttributes(Attrs);
4179 CS.setCallingConv(static_cast<llvm::CallingConv::ID>(CallingConv));
4181 // Apply various metadata.
4183 if (!CI->getType()->isVoidTy())
4184 CI->setName("call");
4186 // Insert instrumentation or attach profile metadata at indirect call sites.
4187 // For more details, see the comment before the definition of
4188 // IPVK_IndirectCallTarget in InstrProfData.inc.
4189 if (!CS.getCalledFunction())
4190 PGO.valueProfile(Builder, llvm::IPVK_IndirectCallTarget,
4193 // In ObjC ARC mode with no ObjC ARC exception safety, tell the ARC
4194 // optimizer it can aggressively ignore unwind edges.
4195 if (CGM.getLangOpts().ObjCAutoRefCount)
4196 AddObjCARCExceptionMetadata(CI);
4198 // Suppress tail calls if requested.
4199 if (llvm::CallInst *Call = dyn_cast<llvm::CallInst>(CI)) {
4200 const Decl *TargetDecl = Callee.getAbstractInfo().getCalleeDecl();
4201 if (TargetDecl && TargetDecl->hasAttr<NotTailCalledAttr>())
4202 Call->setTailCallKind(llvm::CallInst::TCK_NoTail);
4205 // 4. Finish the call.
4207 // If the call doesn't return, finish the basic block and clear the
4208 // insertion point; this allows the rest of IRGen to discard
4209 // unreachable code.
4210 if (CS.doesNotReturn()) {
4211 if (UnusedReturnSize)
4212 EmitLifetimeEnd(llvm::ConstantInt::get(Int64Ty, UnusedReturnSize),
4213 SRetPtr.getPointer());
4215 Builder.CreateUnreachable();
4216 Builder.ClearInsertionPoint();
4218 // FIXME: For now, emit a dummy basic block because expr emitters in
4219 // generally are not ready to handle emitting expressions at unreachable
4221 EnsureInsertPoint();
4223 // Return a reasonable RValue.
4224 return GetUndefRValue(RetTy);
4227 // Perform the swifterror writeback.
4228 if (swiftErrorTemp.isValid()) {
4229 llvm::Value *errorResult = Builder.CreateLoad(swiftErrorTemp);
4230 Builder.CreateStore(errorResult, swiftErrorArg);
4233 // Emit any call-associated writebacks immediately. Arguably this
4234 // should happen after any return-value munging.
4235 if (CallArgs.hasWritebacks())
4236 emitWritebacks(*this, CallArgs);
4238 // The stack cleanup for inalloca arguments has to run out of the normal
4239 // lexical order, so deactivate it and run it manually here.
4240 CallArgs.freeArgumentMemory(*this);
4242 // Extract the return value.
4244 switch (RetAI.getKind()) {
4245 case ABIArgInfo::CoerceAndExpand: {
4246 auto coercionType = RetAI.getCoerceAndExpandType();
4247 auto layout = CGM.getDataLayout().getStructLayout(coercionType);
4249 Address addr = SRetPtr;
4250 addr = Builder.CreateElementBitCast(addr, coercionType);
4252 assert(CI->getType() == RetAI.getUnpaddedCoerceAndExpandType());
4253 bool requiresExtract = isa<llvm::StructType>(CI->getType());
4255 unsigned unpaddedIndex = 0;
4256 for (unsigned i = 0, e = coercionType->getNumElements(); i != e; ++i) {
4257 llvm::Type *eltType = coercionType->getElementType(i);
4258 if (ABIArgInfo::isPaddingForCoerceAndExpand(eltType)) continue;
4259 Address eltAddr = Builder.CreateStructGEP(addr, i, layout);
4260 llvm::Value *elt = CI;
4261 if (requiresExtract)
4262 elt = Builder.CreateExtractValue(elt, unpaddedIndex++);
4264 assert(unpaddedIndex == 0);
4265 Builder.CreateStore(elt, eltAddr);
4271 case ABIArgInfo::InAlloca:
4272 case ABIArgInfo::Indirect: {
4273 RValue ret = convertTempToRValue(SRetPtr, RetTy, SourceLocation());
4274 if (UnusedReturnSize)
4275 EmitLifetimeEnd(llvm::ConstantInt::get(Int64Ty, UnusedReturnSize),
4276 SRetPtr.getPointer());
4280 case ABIArgInfo::Ignore:
4281 // If we are ignoring an argument that had a result, make sure to
4282 // construct the appropriate return value for our caller.
4283 return GetUndefRValue(RetTy);
4285 case ABIArgInfo::Extend:
4286 case ABIArgInfo::Direct: {
4287 llvm::Type *RetIRTy = ConvertType(RetTy);
4288 if (RetAI.getCoerceToType() == RetIRTy && RetAI.getDirectOffset() == 0) {
4289 switch (getEvaluationKind(RetTy)) {
4291 llvm::Value *Real = Builder.CreateExtractValue(CI, 0);
4292 llvm::Value *Imag = Builder.CreateExtractValue(CI, 1);
4293 return RValue::getComplex(std::make_pair(Real, Imag));
4295 case TEK_Aggregate: {
4296 Address DestPtr = ReturnValue.getValue();
4297 bool DestIsVolatile = ReturnValue.isVolatile();
4299 if (!DestPtr.isValid()) {
4300 DestPtr = CreateMemTemp(RetTy, "agg.tmp");
4301 DestIsVolatile = false;
4303 BuildAggStore(*this, CI, DestPtr, DestIsVolatile);
4304 return RValue::getAggregate(DestPtr);
4307 // If the argument doesn't match, perform a bitcast to coerce it. This
4308 // can happen due to trivial type mismatches.
4309 llvm::Value *V = CI;
4310 if (V->getType() != RetIRTy)
4311 V = Builder.CreateBitCast(V, RetIRTy);
4312 return RValue::get(V);
4315 llvm_unreachable("bad evaluation kind");
4318 Address DestPtr = ReturnValue.getValue();
4319 bool DestIsVolatile = ReturnValue.isVolatile();
4321 if (!DestPtr.isValid()) {
4322 DestPtr = CreateMemTemp(RetTy, "coerce");
4323 DestIsVolatile = false;
4326 // If the value is offset in memory, apply the offset now.
4327 Address StorePtr = emitAddressAtOffset(*this, DestPtr, RetAI);
4328 CreateCoercedStore(CI, StorePtr, DestIsVolatile, *this);
4330 return convertTempToRValue(DestPtr, RetTy, SourceLocation());
4333 case ABIArgInfo::Expand:
4334 llvm_unreachable("Invalid ABI kind for return argument");
4337 llvm_unreachable("Unhandled ABIArgInfo::Kind");
4340 // Emit the assume_aligned check on the return value.
4341 const Decl *TargetDecl = Callee.getAbstractInfo().getCalleeDecl();
4342 if (Ret.isScalar() && TargetDecl) {
4343 if (const auto *AA = TargetDecl->getAttr<AssumeAlignedAttr>()) {
4344 llvm::Value *OffsetValue = nullptr;
4345 if (const auto *Offset = AA->getOffset())
4346 OffsetValue = EmitScalarExpr(Offset);
4348 llvm::Value *Alignment = EmitScalarExpr(AA->getAlignment());
4349 llvm::ConstantInt *AlignmentCI = cast<llvm::ConstantInt>(Alignment);
4350 EmitAlignmentAssumption(Ret.getScalarVal(), AlignmentCI->getZExtValue(),
4352 } else if (const auto *AA = TargetDecl->getAttr<AllocAlignAttr>()) {
4353 llvm::Value *ParamVal =
4354 CallArgs[AA->getParamIndex() - 1].RV.getScalarVal();
4355 EmitAlignmentAssumption(Ret.getScalarVal(), ParamVal);
4362 /* VarArg handling */
4364 Address CodeGenFunction::EmitVAArg(VAArgExpr *VE, Address &VAListAddr) {
4365 VAListAddr = VE->isMicrosoftABI()
4366 ? EmitMSVAListRef(VE->getSubExpr())
4367 : EmitVAListRef(VE->getSubExpr());
4368 QualType Ty = VE->getType();
4369 if (VE->isMicrosoftABI())
4370 return CGM.getTypes().getABIInfo().EmitMSVAArg(*this, VAListAddr, Ty);
4371 return CGM.getTypes().getABIInfo().EmitVAArg(*this, VAListAddr, Ty);