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_Win64: return llvm::CallingConv::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 parameters 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_Win64;
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 SmallVector<FunctionProtoType::ExtParameterInfo, 4> extParamInfos(2);
459 argTys.push_back(Context.getCanonicalParamType(receiverType));
460 argTys.push_back(Context.getCanonicalParamType(Context.getObjCSelType()));
462 for (const auto *I : MD->parameters()) {
463 argTys.push_back(Context.getCanonicalParamType(I->getType()));
464 auto extParamInfo = FunctionProtoType::ExtParameterInfo().withIsNoEscape(
465 I->hasAttr<NoEscapeAttr>());
466 extParamInfos.push_back(extParamInfo);
469 FunctionType::ExtInfo einfo;
470 bool IsWindows = getContext().getTargetInfo().getTriple().isOSWindows();
471 einfo = einfo.withCallingConv(getCallingConventionForDecl(MD, IsWindows));
473 if (getContext().getLangOpts().ObjCAutoRefCount &&
474 MD->hasAttr<NSReturnsRetainedAttr>())
475 einfo = einfo.withProducesResult(true);
477 RequiredArgs required =
478 (MD->isVariadic() ? RequiredArgs(argTys.size()) : RequiredArgs::All);
480 return arrangeLLVMFunctionInfo(
481 GetReturnType(MD->getReturnType()), /*instanceMethod=*/false,
482 /*chainCall=*/false, argTys, einfo, extParamInfos, required);
485 const CGFunctionInfo &
486 CodeGenTypes::arrangeUnprototypedObjCMessageSend(QualType returnType,
487 const CallArgList &args) {
488 auto argTypes = getArgTypesForCall(Context, args);
489 FunctionType::ExtInfo einfo;
491 return arrangeLLVMFunctionInfo(
492 GetReturnType(returnType), /*instanceMethod=*/false,
493 /*chainCall=*/false, argTypes, einfo, {}, RequiredArgs::All);
496 const CGFunctionInfo &
497 CodeGenTypes::arrangeGlobalDeclaration(GlobalDecl GD) {
498 // FIXME: Do we need to handle ObjCMethodDecl?
499 const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl());
501 if (const CXXConstructorDecl *CD = dyn_cast<CXXConstructorDecl>(FD))
502 return arrangeCXXStructorDeclaration(CD, getFromCtorType(GD.getCtorType()));
504 if (const CXXDestructorDecl *DD = dyn_cast<CXXDestructorDecl>(FD))
505 return arrangeCXXStructorDeclaration(DD, getFromDtorType(GD.getDtorType()));
507 return arrangeFunctionDeclaration(FD);
510 /// Arrange a thunk that takes 'this' as the first parameter followed by
511 /// varargs. Return a void pointer, regardless of the actual return type.
512 /// The body of the thunk will end in a musttail call to a function of the
513 /// correct type, and the caller will bitcast the function to the correct
515 const CGFunctionInfo &
516 CodeGenTypes::arrangeMSMemberPointerThunk(const CXXMethodDecl *MD) {
517 assert(MD->isVirtual() && "only virtual memptrs have thunks");
518 CanQual<FunctionProtoType> FTP = GetFormalType(MD);
519 CanQualType ArgTys[] = { GetThisType(Context, MD->getParent()) };
520 return arrangeLLVMFunctionInfo(Context.VoidTy, /*instanceMethod=*/false,
521 /*chainCall=*/false, ArgTys,
522 FTP->getExtInfo(), {}, RequiredArgs(1));
525 const CGFunctionInfo &
526 CodeGenTypes::arrangeMSCtorClosure(const CXXConstructorDecl *CD,
528 assert(CT == Ctor_CopyingClosure || CT == Ctor_DefaultClosure);
530 CanQual<FunctionProtoType> FTP = GetFormalType(CD);
531 SmallVector<CanQualType, 2> ArgTys;
532 const CXXRecordDecl *RD = CD->getParent();
533 ArgTys.push_back(GetThisType(Context, RD));
534 if (CT == Ctor_CopyingClosure)
535 ArgTys.push_back(*FTP->param_type_begin());
536 if (RD->getNumVBases() > 0)
537 ArgTys.push_back(Context.IntTy);
538 CallingConv CC = Context.getDefaultCallingConvention(
539 /*IsVariadic=*/false, /*IsCXXMethod=*/true);
540 return arrangeLLVMFunctionInfo(Context.VoidTy, /*instanceMethod=*/true,
541 /*chainCall=*/false, ArgTys,
542 FunctionType::ExtInfo(CC), {},
546 /// Arrange a call as unto a free function, except possibly with an
547 /// additional number of formal parameters considered required.
548 static const CGFunctionInfo &
549 arrangeFreeFunctionLikeCall(CodeGenTypes &CGT,
551 const CallArgList &args,
552 const FunctionType *fnType,
553 unsigned numExtraRequiredArgs,
555 assert(args.size() >= numExtraRequiredArgs);
557 llvm::SmallVector<FunctionProtoType::ExtParameterInfo, 16> paramInfos;
559 // In most cases, there are no optional arguments.
560 RequiredArgs required = RequiredArgs::All;
562 // If we have a variadic prototype, the required arguments are the
563 // extra prefix plus the arguments in the prototype.
564 if (const FunctionProtoType *proto = dyn_cast<FunctionProtoType>(fnType)) {
565 if (proto->isVariadic())
566 required = RequiredArgs(proto->getNumParams() + numExtraRequiredArgs);
568 if (proto->hasExtParameterInfos())
569 addExtParameterInfosForCall(paramInfos, proto, numExtraRequiredArgs,
572 // If we don't have a prototype at all, but we're supposed to
573 // explicitly use the variadic convention for unprototyped calls,
574 // treat all of the arguments as required but preserve the nominal
575 // possibility of variadics.
576 } else if (CGM.getTargetCodeGenInfo()
577 .isNoProtoCallVariadic(args,
578 cast<FunctionNoProtoType>(fnType))) {
579 required = RequiredArgs(args.size());
583 SmallVector<CanQualType, 16> argTypes;
584 for (const auto &arg : args)
585 argTypes.push_back(CGT.getContext().getCanonicalParamType(arg.Ty));
586 return CGT.arrangeLLVMFunctionInfo(GetReturnType(fnType->getReturnType()),
587 /*instanceMethod=*/false, chainCall,
588 argTypes, fnType->getExtInfo(), paramInfos,
592 /// Figure out the rules for calling a function with the given formal
593 /// type using the given arguments. The arguments are necessary
594 /// because the function might be unprototyped, in which case it's
595 /// target-dependent in crazy ways.
596 const CGFunctionInfo &
597 CodeGenTypes::arrangeFreeFunctionCall(const CallArgList &args,
598 const FunctionType *fnType,
600 return arrangeFreeFunctionLikeCall(*this, CGM, args, fnType,
601 chainCall ? 1 : 0, chainCall);
604 /// A block function is essentially a free function with an
605 /// extra implicit argument.
606 const CGFunctionInfo &
607 CodeGenTypes::arrangeBlockFunctionCall(const CallArgList &args,
608 const FunctionType *fnType) {
609 return arrangeFreeFunctionLikeCall(*this, CGM, args, fnType, 1,
610 /*chainCall=*/false);
613 const CGFunctionInfo &
614 CodeGenTypes::arrangeBlockFunctionDeclaration(const FunctionProtoType *proto,
615 const FunctionArgList ¶ms) {
616 auto paramInfos = getExtParameterInfosForCall(proto, 1, params.size());
617 auto argTypes = getArgTypesForDeclaration(Context, params);
619 return arrangeLLVMFunctionInfo(
620 GetReturnType(proto->getReturnType()),
621 /*instanceMethod*/ false, /*chainCall*/ false, argTypes,
622 proto->getExtInfo(), paramInfos,
623 RequiredArgs::forPrototypePlus(proto, 1, nullptr));
626 const CGFunctionInfo &
627 CodeGenTypes::arrangeBuiltinFunctionCall(QualType resultType,
628 const CallArgList &args) {
630 SmallVector<CanQualType, 16> argTypes;
631 for (const auto &Arg : args)
632 argTypes.push_back(Context.getCanonicalParamType(Arg.Ty));
633 return arrangeLLVMFunctionInfo(
634 GetReturnType(resultType), /*instanceMethod=*/false,
635 /*chainCall=*/false, argTypes, FunctionType::ExtInfo(),
636 /*paramInfos=*/ {}, RequiredArgs::All);
639 const CGFunctionInfo &
640 CodeGenTypes::arrangeBuiltinFunctionDeclaration(QualType resultType,
641 const FunctionArgList &args) {
642 auto argTypes = getArgTypesForDeclaration(Context, args);
644 return arrangeLLVMFunctionInfo(
645 GetReturnType(resultType), /*instanceMethod=*/false, /*chainCall=*/false,
646 argTypes, FunctionType::ExtInfo(), {}, RequiredArgs::All);
649 const CGFunctionInfo &
650 CodeGenTypes::arrangeBuiltinFunctionDeclaration(CanQualType resultType,
651 ArrayRef<CanQualType> argTypes) {
652 return arrangeLLVMFunctionInfo(
653 resultType, /*instanceMethod=*/false, /*chainCall=*/false,
654 argTypes, FunctionType::ExtInfo(), {}, RequiredArgs::All);
657 /// Arrange a call to a C++ method, passing the given arguments.
659 /// numPrefixArgs is the number of ABI-specific prefix arguments we have. It
660 /// does not count `this`.
661 const CGFunctionInfo &
662 CodeGenTypes::arrangeCXXMethodCall(const CallArgList &args,
663 const FunctionProtoType *proto,
664 RequiredArgs required,
665 unsigned numPrefixArgs) {
666 assert(numPrefixArgs + 1 <= args.size() &&
667 "Emitting a call with less args than the required prefix?");
668 // Add one to account for `this`. It's a bit awkward here, but we don't count
669 // `this` in similar places elsewhere.
671 getExtParameterInfosForCall(proto, numPrefixArgs + 1, args.size());
674 auto argTypes = getArgTypesForCall(Context, args);
676 FunctionType::ExtInfo info = proto->getExtInfo();
677 return arrangeLLVMFunctionInfo(
678 GetReturnType(proto->getReturnType()), /*instanceMethod=*/true,
679 /*chainCall=*/false, argTypes, info, paramInfos, required);
682 const CGFunctionInfo &CodeGenTypes::arrangeNullaryFunction() {
683 return arrangeLLVMFunctionInfo(
684 getContext().VoidTy, /*instanceMethod=*/false, /*chainCall=*/false,
685 None, FunctionType::ExtInfo(), {}, RequiredArgs::All);
688 const CGFunctionInfo &
689 CodeGenTypes::arrangeCall(const CGFunctionInfo &signature,
690 const CallArgList &args) {
691 assert(signature.arg_size() <= args.size());
692 if (signature.arg_size() == args.size())
695 SmallVector<FunctionProtoType::ExtParameterInfo, 16> paramInfos;
696 auto sigParamInfos = signature.getExtParameterInfos();
697 if (!sigParamInfos.empty()) {
698 paramInfos.append(sigParamInfos.begin(), sigParamInfos.end());
699 paramInfos.resize(args.size());
702 auto argTypes = getArgTypesForCall(Context, args);
704 assert(signature.getRequiredArgs().allowsOptionalArgs());
705 return arrangeLLVMFunctionInfo(signature.getReturnType(),
706 signature.isInstanceMethod(),
707 signature.isChainCall(),
709 signature.getExtInfo(),
711 signature.getRequiredArgs());
716 void computeSPIRKernelABIInfo(CodeGenModule &CGM, CGFunctionInfo &FI);
720 /// Arrange the argument and result information for an abstract value
721 /// of a given function type. This is the method which all of the
722 /// above functions ultimately defer to.
723 const CGFunctionInfo &
724 CodeGenTypes::arrangeLLVMFunctionInfo(CanQualType resultType,
727 ArrayRef<CanQualType> argTypes,
728 FunctionType::ExtInfo info,
729 ArrayRef<FunctionProtoType::ExtParameterInfo> paramInfos,
730 RequiredArgs required) {
731 assert(std::all_of(argTypes.begin(), argTypes.end(),
732 [](CanQualType T) { return T.isCanonicalAsParam(); }));
734 // Lookup or create unique function info.
735 llvm::FoldingSetNodeID ID;
736 CGFunctionInfo::Profile(ID, instanceMethod, chainCall, info, paramInfos,
737 required, resultType, argTypes);
739 void *insertPos = nullptr;
740 CGFunctionInfo *FI = FunctionInfos.FindNodeOrInsertPos(ID, insertPos);
744 unsigned CC = ClangCallConvToLLVMCallConv(info.getCC());
746 // Construct the function info. We co-allocate the ArgInfos.
747 FI = CGFunctionInfo::create(CC, instanceMethod, chainCall, info,
748 paramInfos, resultType, argTypes, required);
749 FunctionInfos.InsertNode(FI, insertPos);
751 bool inserted = FunctionsBeingProcessed.insert(FI).second;
753 assert(inserted && "Recursively being processed?");
755 // Compute ABI information.
756 if (CC == llvm::CallingConv::SPIR_KERNEL) {
757 // Force target independent argument handling for the host visible
759 computeSPIRKernelABIInfo(CGM, *FI);
760 } else if (info.getCC() == CC_Swift) {
761 swiftcall::computeABIInfo(CGM, *FI);
763 getABIInfo().computeInfo(*FI);
766 // Loop over all of the computed argument and return value info. If any of
767 // them are direct or extend without a specified coerce type, specify the
769 ABIArgInfo &retInfo = FI->getReturnInfo();
770 if (retInfo.canHaveCoerceToType() && retInfo.getCoerceToType() == nullptr)
771 retInfo.setCoerceToType(ConvertType(FI->getReturnType()));
773 for (auto &I : FI->arguments())
774 if (I.info.canHaveCoerceToType() && I.info.getCoerceToType() == nullptr)
775 I.info.setCoerceToType(ConvertType(I.type));
777 bool erased = FunctionsBeingProcessed.erase(FI); (void)erased;
778 assert(erased && "Not in set?");
783 CGFunctionInfo *CGFunctionInfo::create(unsigned llvmCC,
786 const FunctionType::ExtInfo &info,
787 ArrayRef<ExtParameterInfo> paramInfos,
788 CanQualType resultType,
789 ArrayRef<CanQualType> argTypes,
790 RequiredArgs required) {
791 assert(paramInfos.empty() || paramInfos.size() == argTypes.size());
794 operator new(totalSizeToAlloc<ArgInfo, ExtParameterInfo>(
795 argTypes.size() + 1, paramInfos.size()));
797 CGFunctionInfo *FI = new(buffer) CGFunctionInfo();
798 FI->CallingConvention = llvmCC;
799 FI->EffectiveCallingConvention = llvmCC;
800 FI->ASTCallingConvention = info.getCC();
801 FI->InstanceMethod = instanceMethod;
802 FI->ChainCall = chainCall;
803 FI->NoReturn = info.getNoReturn();
804 FI->ReturnsRetained = info.getProducesResult();
805 FI->NoCallerSavedRegs = info.getNoCallerSavedRegs();
806 FI->Required = required;
807 FI->HasRegParm = info.getHasRegParm();
808 FI->RegParm = info.getRegParm();
809 FI->ArgStruct = nullptr;
810 FI->ArgStructAlign = 0;
811 FI->NumArgs = argTypes.size();
812 FI->HasExtParameterInfos = !paramInfos.empty();
813 FI->getArgsBuffer()[0].type = resultType;
814 for (unsigned i = 0, e = argTypes.size(); i != e; ++i)
815 FI->getArgsBuffer()[i + 1].type = argTypes[i];
816 for (unsigned i = 0, e = paramInfos.size(); i != e; ++i)
817 FI->getExtParameterInfosBuffer()[i] = paramInfos[i];
824 // ABIArgInfo::Expand implementation.
826 // Specifies the way QualType passed as ABIArgInfo::Expand is expanded.
827 struct TypeExpansion {
828 enum TypeExpansionKind {
829 // Elements of constant arrays are expanded recursively.
831 // Record fields are expanded recursively (but if record is a union, only
832 // the field with the largest size is expanded).
834 // For complex types, real and imaginary parts are expanded recursively.
836 // All other types are not expandable.
840 const TypeExpansionKind Kind;
842 TypeExpansion(TypeExpansionKind K) : Kind(K) {}
843 virtual ~TypeExpansion() {}
846 struct ConstantArrayExpansion : TypeExpansion {
850 ConstantArrayExpansion(QualType EltTy, uint64_t NumElts)
851 : TypeExpansion(TEK_ConstantArray), EltTy(EltTy), NumElts(NumElts) {}
852 static bool classof(const TypeExpansion *TE) {
853 return TE->Kind == TEK_ConstantArray;
857 struct RecordExpansion : TypeExpansion {
858 SmallVector<const CXXBaseSpecifier *, 1> Bases;
860 SmallVector<const FieldDecl *, 1> Fields;
862 RecordExpansion(SmallVector<const CXXBaseSpecifier *, 1> &&Bases,
863 SmallVector<const FieldDecl *, 1> &&Fields)
864 : TypeExpansion(TEK_Record), Bases(std::move(Bases)),
865 Fields(std::move(Fields)) {}
866 static bool classof(const TypeExpansion *TE) {
867 return TE->Kind == TEK_Record;
871 struct ComplexExpansion : TypeExpansion {
874 ComplexExpansion(QualType EltTy) : TypeExpansion(TEK_Complex), EltTy(EltTy) {}
875 static bool classof(const TypeExpansion *TE) {
876 return TE->Kind == TEK_Complex;
880 struct NoExpansion : TypeExpansion {
881 NoExpansion() : TypeExpansion(TEK_None) {}
882 static bool classof(const TypeExpansion *TE) {
883 return TE->Kind == TEK_None;
888 static std::unique_ptr<TypeExpansion>
889 getTypeExpansion(QualType Ty, const ASTContext &Context) {
890 if (const ConstantArrayType *AT = Context.getAsConstantArrayType(Ty)) {
891 return llvm::make_unique<ConstantArrayExpansion>(
892 AT->getElementType(), AT->getSize().getZExtValue());
894 if (const RecordType *RT = Ty->getAs<RecordType>()) {
895 SmallVector<const CXXBaseSpecifier *, 1> Bases;
896 SmallVector<const FieldDecl *, 1> Fields;
897 const RecordDecl *RD = RT->getDecl();
898 assert(!RD->hasFlexibleArrayMember() &&
899 "Cannot expand structure with flexible array.");
901 // Unions can be here only in degenerative cases - all the fields are same
902 // after flattening. Thus we have to use the "largest" field.
903 const FieldDecl *LargestFD = nullptr;
904 CharUnits UnionSize = CharUnits::Zero();
906 for (const auto *FD : RD->fields()) {
907 // Skip zero length bitfields.
908 if (FD->isBitField() && FD->getBitWidthValue(Context) == 0)
910 assert(!FD->isBitField() &&
911 "Cannot expand structure with bit-field members.");
912 CharUnits FieldSize = Context.getTypeSizeInChars(FD->getType());
913 if (UnionSize < FieldSize) {
914 UnionSize = FieldSize;
919 Fields.push_back(LargestFD);
921 if (const auto *CXXRD = dyn_cast<CXXRecordDecl>(RD)) {
922 assert(!CXXRD->isDynamicClass() &&
923 "cannot expand vtable pointers in dynamic classes");
924 for (const CXXBaseSpecifier &BS : CXXRD->bases())
925 Bases.push_back(&BS);
928 for (const auto *FD : RD->fields()) {
929 // Skip zero length bitfields.
930 if (FD->isBitField() && FD->getBitWidthValue(Context) == 0)
932 assert(!FD->isBitField() &&
933 "Cannot expand structure with bit-field members.");
934 Fields.push_back(FD);
937 return llvm::make_unique<RecordExpansion>(std::move(Bases),
940 if (const ComplexType *CT = Ty->getAs<ComplexType>()) {
941 return llvm::make_unique<ComplexExpansion>(CT->getElementType());
943 return llvm::make_unique<NoExpansion>();
946 static int getExpansionSize(QualType Ty, const ASTContext &Context) {
947 auto Exp = getTypeExpansion(Ty, Context);
948 if (auto CAExp = dyn_cast<ConstantArrayExpansion>(Exp.get())) {
949 return CAExp->NumElts * getExpansionSize(CAExp->EltTy, Context);
951 if (auto RExp = dyn_cast<RecordExpansion>(Exp.get())) {
953 for (auto BS : RExp->Bases)
954 Res += getExpansionSize(BS->getType(), Context);
955 for (auto FD : RExp->Fields)
956 Res += getExpansionSize(FD->getType(), Context);
959 if (isa<ComplexExpansion>(Exp.get()))
961 assert(isa<NoExpansion>(Exp.get()));
966 CodeGenTypes::getExpandedTypes(QualType Ty,
967 SmallVectorImpl<llvm::Type *>::iterator &TI) {
968 auto Exp = getTypeExpansion(Ty, Context);
969 if (auto CAExp = dyn_cast<ConstantArrayExpansion>(Exp.get())) {
970 for (int i = 0, n = CAExp->NumElts; i < n; i++) {
971 getExpandedTypes(CAExp->EltTy, TI);
973 } else if (auto RExp = dyn_cast<RecordExpansion>(Exp.get())) {
974 for (auto BS : RExp->Bases)
975 getExpandedTypes(BS->getType(), TI);
976 for (auto FD : RExp->Fields)
977 getExpandedTypes(FD->getType(), TI);
978 } else if (auto CExp = dyn_cast<ComplexExpansion>(Exp.get())) {
979 llvm::Type *EltTy = ConvertType(CExp->EltTy);
983 assert(isa<NoExpansion>(Exp.get()));
984 *TI++ = ConvertType(Ty);
988 static void forConstantArrayExpansion(CodeGenFunction &CGF,
989 ConstantArrayExpansion *CAE,
991 llvm::function_ref<void(Address)> Fn) {
992 CharUnits EltSize = CGF.getContext().getTypeSizeInChars(CAE->EltTy);
994 BaseAddr.getAlignment().alignmentOfArrayElement(EltSize);
996 for (int i = 0, n = CAE->NumElts; i < n; i++) {
997 llvm::Value *EltAddr =
998 CGF.Builder.CreateConstGEP2_32(nullptr, BaseAddr.getPointer(), 0, i);
999 Fn(Address(EltAddr, EltAlign));
1003 void CodeGenFunction::ExpandTypeFromArgs(
1004 QualType Ty, LValue LV, SmallVectorImpl<llvm::Value *>::iterator &AI) {
1005 assert(LV.isSimple() &&
1006 "Unexpected non-simple lvalue during struct expansion.");
1008 auto Exp = getTypeExpansion(Ty, getContext());
1009 if (auto CAExp = dyn_cast<ConstantArrayExpansion>(Exp.get())) {
1010 forConstantArrayExpansion(*this, CAExp, LV.getAddress(),
1011 [&](Address EltAddr) {
1012 LValue LV = MakeAddrLValue(EltAddr, CAExp->EltTy);
1013 ExpandTypeFromArgs(CAExp->EltTy, LV, AI);
1015 } else if (auto RExp = dyn_cast<RecordExpansion>(Exp.get())) {
1016 Address This = LV.getAddress();
1017 for (const CXXBaseSpecifier *BS : RExp->Bases) {
1018 // Perform a single step derived-to-base conversion.
1020 GetAddressOfBaseClass(This, Ty->getAsCXXRecordDecl(), &BS, &BS + 1,
1021 /*NullCheckValue=*/false, SourceLocation());
1022 LValue SubLV = MakeAddrLValue(Base, BS->getType());
1024 // Recurse onto bases.
1025 ExpandTypeFromArgs(BS->getType(), SubLV, AI);
1027 for (auto FD : RExp->Fields) {
1028 // FIXME: What are the right qualifiers here?
1029 LValue SubLV = EmitLValueForFieldInitialization(LV, FD);
1030 ExpandTypeFromArgs(FD->getType(), SubLV, AI);
1032 } else if (isa<ComplexExpansion>(Exp.get())) {
1033 auto realValue = *AI++;
1034 auto imagValue = *AI++;
1035 EmitStoreOfComplex(ComplexPairTy(realValue, imagValue), LV, /*init*/ true);
1037 assert(isa<NoExpansion>(Exp.get()));
1038 EmitStoreThroughLValue(RValue::get(*AI++), LV);
1042 void CodeGenFunction::ExpandTypeToArgs(
1043 QualType Ty, RValue RV, llvm::FunctionType *IRFuncTy,
1044 SmallVectorImpl<llvm::Value *> &IRCallArgs, unsigned &IRCallArgPos) {
1045 auto Exp = getTypeExpansion(Ty, getContext());
1046 if (auto CAExp = dyn_cast<ConstantArrayExpansion>(Exp.get())) {
1047 forConstantArrayExpansion(*this, CAExp, RV.getAggregateAddress(),
1048 [&](Address EltAddr) {
1050 convertTempToRValue(EltAddr, CAExp->EltTy, SourceLocation());
1051 ExpandTypeToArgs(CAExp->EltTy, EltRV, IRFuncTy, IRCallArgs, IRCallArgPos);
1053 } else if (auto RExp = dyn_cast<RecordExpansion>(Exp.get())) {
1054 Address This = RV.getAggregateAddress();
1055 for (const CXXBaseSpecifier *BS : RExp->Bases) {
1056 // Perform a single step derived-to-base conversion.
1058 GetAddressOfBaseClass(This, Ty->getAsCXXRecordDecl(), &BS, &BS + 1,
1059 /*NullCheckValue=*/false, SourceLocation());
1060 RValue BaseRV = RValue::getAggregate(Base);
1062 // Recurse onto bases.
1063 ExpandTypeToArgs(BS->getType(), BaseRV, IRFuncTy, IRCallArgs,
1067 LValue LV = MakeAddrLValue(This, Ty);
1068 for (auto FD : RExp->Fields) {
1069 RValue FldRV = EmitRValueForField(LV, FD, SourceLocation());
1070 ExpandTypeToArgs(FD->getType(), FldRV, IRFuncTy, IRCallArgs,
1073 } else if (isa<ComplexExpansion>(Exp.get())) {
1074 ComplexPairTy CV = RV.getComplexVal();
1075 IRCallArgs[IRCallArgPos++] = CV.first;
1076 IRCallArgs[IRCallArgPos++] = CV.second;
1078 assert(isa<NoExpansion>(Exp.get()));
1079 assert(RV.isScalar() &&
1080 "Unexpected non-scalar rvalue during struct expansion.");
1082 // Insert a bitcast as needed.
1083 llvm::Value *V = RV.getScalarVal();
1084 if (IRCallArgPos < IRFuncTy->getNumParams() &&
1085 V->getType() != IRFuncTy->getParamType(IRCallArgPos))
1086 V = Builder.CreateBitCast(V, IRFuncTy->getParamType(IRCallArgPos));
1088 IRCallArgs[IRCallArgPos++] = V;
1092 /// Create a temporary allocation for the purposes of coercion.
1093 static Address CreateTempAllocaForCoercion(CodeGenFunction &CGF, llvm::Type *Ty,
1094 CharUnits MinAlign) {
1095 // Don't use an alignment that's worse than what LLVM would prefer.
1096 auto PrefAlign = CGF.CGM.getDataLayout().getPrefTypeAlignment(Ty);
1097 CharUnits Align = std::max(MinAlign, CharUnits::fromQuantity(PrefAlign));
1099 return CGF.CreateTempAlloca(Ty, Align);
1102 /// EnterStructPointerForCoercedAccess - Given a struct pointer that we are
1103 /// accessing some number of bytes out of it, try to gep into the struct to get
1104 /// at its inner goodness. Dive as deep as possible without entering an element
1105 /// with an in-memory size smaller than DstSize.
1107 EnterStructPointerForCoercedAccess(Address SrcPtr,
1108 llvm::StructType *SrcSTy,
1109 uint64_t DstSize, CodeGenFunction &CGF) {
1110 // We can't dive into a zero-element struct.
1111 if (SrcSTy->getNumElements() == 0) return SrcPtr;
1113 llvm::Type *FirstElt = SrcSTy->getElementType(0);
1115 // If the first elt is at least as large as what we're looking for, or if the
1116 // first element is the same size as the whole struct, we can enter it. The
1117 // comparison must be made on the store size and not the alloca size. Using
1118 // the alloca size may overstate the size of the load.
1119 uint64_t FirstEltSize =
1120 CGF.CGM.getDataLayout().getTypeStoreSize(FirstElt);
1121 if (FirstEltSize < DstSize &&
1122 FirstEltSize < CGF.CGM.getDataLayout().getTypeStoreSize(SrcSTy))
1125 // GEP into the first element.
1126 SrcPtr = CGF.Builder.CreateStructGEP(SrcPtr, 0, CharUnits(), "coerce.dive");
1128 // If the first element is a struct, recurse.
1129 llvm::Type *SrcTy = SrcPtr.getElementType();
1130 if (llvm::StructType *SrcSTy = dyn_cast<llvm::StructType>(SrcTy))
1131 return EnterStructPointerForCoercedAccess(SrcPtr, SrcSTy, DstSize, CGF);
1136 /// CoerceIntOrPtrToIntOrPtr - Convert a value Val to the specific Ty where both
1137 /// are either integers or pointers. This does a truncation of the value if it
1138 /// is too large or a zero extension if it is too small.
1140 /// This behaves as if the value were coerced through memory, so on big-endian
1141 /// targets the high bits are preserved in a truncation, while little-endian
1142 /// targets preserve the low bits.
1143 static llvm::Value *CoerceIntOrPtrToIntOrPtr(llvm::Value *Val,
1145 CodeGenFunction &CGF) {
1146 if (Val->getType() == Ty)
1149 if (isa<llvm::PointerType>(Val->getType())) {
1150 // If this is Pointer->Pointer avoid conversion to and from int.
1151 if (isa<llvm::PointerType>(Ty))
1152 return CGF.Builder.CreateBitCast(Val, Ty, "coerce.val");
1154 // Convert the pointer to an integer so we can play with its width.
1155 Val = CGF.Builder.CreatePtrToInt(Val, CGF.IntPtrTy, "coerce.val.pi");
1158 llvm::Type *DestIntTy = Ty;
1159 if (isa<llvm::PointerType>(DestIntTy))
1160 DestIntTy = CGF.IntPtrTy;
1162 if (Val->getType() != DestIntTy) {
1163 const llvm::DataLayout &DL = CGF.CGM.getDataLayout();
1164 if (DL.isBigEndian()) {
1165 // Preserve the high bits on big-endian targets.
1166 // That is what memory coercion does.
1167 uint64_t SrcSize = DL.getTypeSizeInBits(Val->getType());
1168 uint64_t DstSize = DL.getTypeSizeInBits(DestIntTy);
1170 if (SrcSize > DstSize) {
1171 Val = CGF.Builder.CreateLShr(Val, SrcSize - DstSize, "coerce.highbits");
1172 Val = CGF.Builder.CreateTrunc(Val, DestIntTy, "coerce.val.ii");
1174 Val = CGF.Builder.CreateZExt(Val, DestIntTy, "coerce.val.ii");
1175 Val = CGF.Builder.CreateShl(Val, DstSize - SrcSize, "coerce.highbits");
1178 // Little-endian targets preserve the low bits. No shifts required.
1179 Val = CGF.Builder.CreateIntCast(Val, DestIntTy, false, "coerce.val.ii");
1183 if (isa<llvm::PointerType>(Ty))
1184 Val = CGF.Builder.CreateIntToPtr(Val, Ty, "coerce.val.ip");
1190 /// CreateCoercedLoad - Create a load from \arg SrcPtr interpreted as
1191 /// a pointer to an object of type \arg Ty, known to be aligned to
1192 /// \arg SrcAlign bytes.
1194 /// This safely handles the case when the src type is smaller than the
1195 /// destination type; in this situation the values of bits which not
1196 /// present in the src are undefined.
1197 static llvm::Value *CreateCoercedLoad(Address Src, llvm::Type *Ty,
1198 CodeGenFunction &CGF) {
1199 llvm::Type *SrcTy = Src.getElementType();
1201 // If SrcTy and Ty are the same, just do a load.
1203 return CGF.Builder.CreateLoad(Src);
1205 uint64_t DstSize = CGF.CGM.getDataLayout().getTypeAllocSize(Ty);
1207 if (llvm::StructType *SrcSTy = dyn_cast<llvm::StructType>(SrcTy)) {
1208 Src = EnterStructPointerForCoercedAccess(Src, SrcSTy, DstSize, CGF);
1209 SrcTy = Src.getType()->getElementType();
1212 uint64_t SrcSize = CGF.CGM.getDataLayout().getTypeAllocSize(SrcTy);
1214 // If the source and destination are integer or pointer types, just do an
1215 // extension or truncation to the desired type.
1216 if ((isa<llvm::IntegerType>(Ty) || isa<llvm::PointerType>(Ty)) &&
1217 (isa<llvm::IntegerType>(SrcTy) || isa<llvm::PointerType>(SrcTy))) {
1218 llvm::Value *Load = CGF.Builder.CreateLoad(Src);
1219 return CoerceIntOrPtrToIntOrPtr(Load, Ty, CGF);
1222 // If load is legal, just bitcast the src pointer.
1223 if (SrcSize >= DstSize) {
1224 // Generally SrcSize is never greater than DstSize, since this means we are
1225 // losing bits. However, this can happen in cases where the structure has
1226 // additional padding, for example due to a user specified alignment.
1228 // FIXME: Assert that we aren't truncating non-padding bits when have access
1229 // to that information.
1230 Src = CGF.Builder.CreateBitCast(Src,
1231 Ty->getPointerTo(Src.getAddressSpace()));
1232 return CGF.Builder.CreateLoad(Src);
1235 // Otherwise do coercion through memory. This is stupid, but simple.
1236 Address Tmp = CreateTempAllocaForCoercion(CGF, Ty, Src.getAlignment());
1237 Address Casted = CGF.Builder.CreateBitCast(Tmp, CGF.AllocaInt8PtrTy);
1238 Address SrcCasted = CGF.Builder.CreateBitCast(Src, CGF.AllocaInt8PtrTy);
1239 CGF.Builder.CreateMemCpy(Casted, SrcCasted,
1240 llvm::ConstantInt::get(CGF.IntPtrTy, SrcSize),
1242 return CGF.Builder.CreateLoad(Tmp);
1245 // Function to store a first-class aggregate into memory. We prefer to
1246 // store the elements rather than the aggregate to be more friendly to
1248 // FIXME: Do we need to recurse here?
1249 static void BuildAggStore(CodeGenFunction &CGF, llvm::Value *Val,
1250 Address Dest, bool DestIsVolatile) {
1251 // Prefer scalar stores to first-class aggregate stores.
1252 if (llvm::StructType *STy =
1253 dyn_cast<llvm::StructType>(Val->getType())) {
1254 const llvm::StructLayout *Layout =
1255 CGF.CGM.getDataLayout().getStructLayout(STy);
1257 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1258 auto EltOffset = CharUnits::fromQuantity(Layout->getElementOffset(i));
1259 Address EltPtr = CGF.Builder.CreateStructGEP(Dest, i, EltOffset);
1260 llvm::Value *Elt = CGF.Builder.CreateExtractValue(Val, i);
1261 CGF.Builder.CreateStore(Elt, EltPtr, DestIsVolatile);
1264 CGF.Builder.CreateStore(Val, Dest, DestIsVolatile);
1268 /// CreateCoercedStore - Create a store to \arg DstPtr from \arg Src,
1269 /// where the source and destination may have different types. The
1270 /// destination is known to be aligned to \arg DstAlign bytes.
1272 /// This safely handles the case when the src type is larger than the
1273 /// destination type; the upper bits of the src will be lost.
1274 static void CreateCoercedStore(llvm::Value *Src,
1277 CodeGenFunction &CGF) {
1278 llvm::Type *SrcTy = Src->getType();
1279 llvm::Type *DstTy = Dst.getType()->getElementType();
1280 if (SrcTy == DstTy) {
1281 CGF.Builder.CreateStore(Src, Dst, DstIsVolatile);
1285 uint64_t SrcSize = CGF.CGM.getDataLayout().getTypeAllocSize(SrcTy);
1287 if (llvm::StructType *DstSTy = dyn_cast<llvm::StructType>(DstTy)) {
1288 Dst = EnterStructPointerForCoercedAccess(Dst, DstSTy, SrcSize, CGF);
1289 DstTy = Dst.getType()->getElementType();
1292 // If the source and destination are integer or pointer types, just do an
1293 // extension or truncation to the desired type.
1294 if ((isa<llvm::IntegerType>(SrcTy) || isa<llvm::PointerType>(SrcTy)) &&
1295 (isa<llvm::IntegerType>(DstTy) || isa<llvm::PointerType>(DstTy))) {
1296 Src = CoerceIntOrPtrToIntOrPtr(Src, DstTy, CGF);
1297 CGF.Builder.CreateStore(Src, Dst, DstIsVolatile);
1301 uint64_t DstSize = CGF.CGM.getDataLayout().getTypeAllocSize(DstTy);
1303 // If store is legal, just bitcast the src pointer.
1304 if (SrcSize <= DstSize) {
1305 Dst = CGF.Builder.CreateElementBitCast(Dst, SrcTy);
1306 BuildAggStore(CGF, Src, Dst, DstIsVolatile);
1308 // Otherwise do coercion through memory. This is stupid, but
1311 // Generally SrcSize is never greater than DstSize, since this means we are
1312 // losing bits. However, this can happen in cases where the structure has
1313 // additional padding, for example due to a user specified alignment.
1315 // FIXME: Assert that we aren't truncating non-padding bits when have access
1316 // to that information.
1317 Address Tmp = CreateTempAllocaForCoercion(CGF, SrcTy, Dst.getAlignment());
1318 CGF.Builder.CreateStore(Src, Tmp);
1319 Address Casted = CGF.Builder.CreateBitCast(Tmp, CGF.AllocaInt8PtrTy);
1320 Address DstCasted = CGF.Builder.CreateBitCast(Dst, CGF.AllocaInt8PtrTy);
1321 CGF.Builder.CreateMemCpy(DstCasted, Casted,
1322 llvm::ConstantInt::get(CGF.IntPtrTy, DstSize),
1327 static Address emitAddressAtOffset(CodeGenFunction &CGF, Address addr,
1328 const ABIArgInfo &info) {
1329 if (unsigned offset = info.getDirectOffset()) {
1330 addr = CGF.Builder.CreateElementBitCast(addr, CGF.Int8Ty);
1331 addr = CGF.Builder.CreateConstInBoundsByteGEP(addr,
1332 CharUnits::fromQuantity(offset));
1333 addr = CGF.Builder.CreateElementBitCast(addr, info.getCoerceToType());
1340 /// Encapsulates information about the way function arguments from
1341 /// CGFunctionInfo should be passed to actual LLVM IR function.
1342 class ClangToLLVMArgMapping {
1343 static const unsigned InvalidIndex = ~0U;
1344 unsigned InallocaArgNo;
1346 unsigned TotalIRArgs;
1348 /// Arguments of LLVM IR function corresponding to single Clang argument.
1350 unsigned PaddingArgIndex;
1351 // Argument is expanded to IR arguments at positions
1352 // [FirstArgIndex, FirstArgIndex + NumberOfArgs).
1353 unsigned FirstArgIndex;
1354 unsigned NumberOfArgs;
1357 : PaddingArgIndex(InvalidIndex), FirstArgIndex(InvalidIndex),
1361 SmallVector<IRArgs, 8> ArgInfo;
1364 ClangToLLVMArgMapping(const ASTContext &Context, const CGFunctionInfo &FI,
1365 bool OnlyRequiredArgs = false)
1366 : InallocaArgNo(InvalidIndex), SRetArgNo(InvalidIndex), TotalIRArgs(0),
1367 ArgInfo(OnlyRequiredArgs ? FI.getNumRequiredArgs() : FI.arg_size()) {
1368 construct(Context, FI, OnlyRequiredArgs);
1371 bool hasInallocaArg() const { return InallocaArgNo != InvalidIndex; }
1372 unsigned getInallocaArgNo() const {
1373 assert(hasInallocaArg());
1374 return InallocaArgNo;
1377 bool hasSRetArg() const { return SRetArgNo != InvalidIndex; }
1378 unsigned getSRetArgNo() const {
1379 assert(hasSRetArg());
1383 unsigned totalIRArgs() const { return TotalIRArgs; }
1385 bool hasPaddingArg(unsigned ArgNo) const {
1386 assert(ArgNo < ArgInfo.size());
1387 return ArgInfo[ArgNo].PaddingArgIndex != InvalidIndex;
1389 unsigned getPaddingArgNo(unsigned ArgNo) const {
1390 assert(hasPaddingArg(ArgNo));
1391 return ArgInfo[ArgNo].PaddingArgIndex;
1394 /// Returns index of first IR argument corresponding to ArgNo, and their
1396 std::pair<unsigned, unsigned> getIRArgs(unsigned ArgNo) const {
1397 assert(ArgNo < ArgInfo.size());
1398 return std::make_pair(ArgInfo[ArgNo].FirstArgIndex,
1399 ArgInfo[ArgNo].NumberOfArgs);
1403 void construct(const ASTContext &Context, const CGFunctionInfo &FI,
1404 bool OnlyRequiredArgs);
1407 void ClangToLLVMArgMapping::construct(const ASTContext &Context,
1408 const CGFunctionInfo &FI,
1409 bool OnlyRequiredArgs) {
1410 unsigned IRArgNo = 0;
1411 bool SwapThisWithSRet = false;
1412 const ABIArgInfo &RetAI = FI.getReturnInfo();
1414 if (RetAI.getKind() == ABIArgInfo::Indirect) {
1415 SwapThisWithSRet = RetAI.isSRetAfterThis();
1416 SRetArgNo = SwapThisWithSRet ? 1 : IRArgNo++;
1420 unsigned NumArgs = OnlyRequiredArgs ? FI.getNumRequiredArgs() : FI.arg_size();
1421 for (CGFunctionInfo::const_arg_iterator I = FI.arg_begin(); ArgNo < NumArgs;
1423 assert(I != FI.arg_end());
1424 QualType ArgType = I->type;
1425 const ABIArgInfo &AI = I->info;
1426 // Collect data about IR arguments corresponding to Clang argument ArgNo.
1427 auto &IRArgs = ArgInfo[ArgNo];
1429 if (AI.getPaddingType())
1430 IRArgs.PaddingArgIndex = IRArgNo++;
1432 switch (AI.getKind()) {
1433 case ABIArgInfo::Extend:
1434 case ABIArgInfo::Direct: {
1435 // FIXME: handle sseregparm someday...
1436 llvm::StructType *STy = dyn_cast<llvm::StructType>(AI.getCoerceToType());
1437 if (AI.isDirect() && AI.getCanBeFlattened() && STy) {
1438 IRArgs.NumberOfArgs = STy->getNumElements();
1440 IRArgs.NumberOfArgs = 1;
1444 case ABIArgInfo::Indirect:
1445 IRArgs.NumberOfArgs = 1;
1447 case ABIArgInfo::Ignore:
1448 case ABIArgInfo::InAlloca:
1449 // ignore and inalloca doesn't have matching LLVM parameters.
1450 IRArgs.NumberOfArgs = 0;
1452 case ABIArgInfo::CoerceAndExpand:
1453 IRArgs.NumberOfArgs = AI.getCoerceAndExpandTypeSequence().size();
1455 case ABIArgInfo::Expand:
1456 IRArgs.NumberOfArgs = getExpansionSize(ArgType, Context);
1460 if (IRArgs.NumberOfArgs > 0) {
1461 IRArgs.FirstArgIndex = IRArgNo;
1462 IRArgNo += IRArgs.NumberOfArgs;
1465 // Skip over the sret parameter when it comes second. We already handled it
1467 if (IRArgNo == 1 && SwapThisWithSRet)
1470 assert(ArgNo == ArgInfo.size());
1472 if (FI.usesInAlloca())
1473 InallocaArgNo = IRArgNo++;
1475 TotalIRArgs = IRArgNo;
1481 bool CodeGenModule::ReturnTypeUsesSRet(const CGFunctionInfo &FI) {
1482 return FI.getReturnInfo().isIndirect();
1485 bool CodeGenModule::ReturnSlotInterferesWithArgs(const CGFunctionInfo &FI) {
1486 return ReturnTypeUsesSRet(FI) &&
1487 getTargetCodeGenInfo().doesReturnSlotInterfereWithArgs();
1490 bool CodeGenModule::ReturnTypeUsesFPRet(QualType ResultType) {
1491 if (const BuiltinType *BT = ResultType->getAs<BuiltinType>()) {
1492 switch (BT->getKind()) {
1495 case BuiltinType::Float:
1496 return getTarget().useObjCFPRetForRealType(TargetInfo::Float);
1497 case BuiltinType::Double:
1498 return getTarget().useObjCFPRetForRealType(TargetInfo::Double);
1499 case BuiltinType::LongDouble:
1500 return getTarget().useObjCFPRetForRealType(TargetInfo::LongDouble);
1507 bool CodeGenModule::ReturnTypeUsesFP2Ret(QualType ResultType) {
1508 if (const ComplexType *CT = ResultType->getAs<ComplexType>()) {
1509 if (const BuiltinType *BT = CT->getElementType()->getAs<BuiltinType>()) {
1510 if (BT->getKind() == BuiltinType::LongDouble)
1511 return getTarget().useObjCFP2RetForComplexLongDouble();
1518 llvm::FunctionType *CodeGenTypes::GetFunctionType(GlobalDecl GD) {
1519 const CGFunctionInfo &FI = arrangeGlobalDeclaration(GD);
1520 return GetFunctionType(FI);
1523 llvm::FunctionType *
1524 CodeGenTypes::GetFunctionType(const CGFunctionInfo &FI) {
1526 bool Inserted = FunctionsBeingProcessed.insert(&FI).second;
1528 assert(Inserted && "Recursively being processed?");
1530 llvm::Type *resultType = nullptr;
1531 const ABIArgInfo &retAI = FI.getReturnInfo();
1532 switch (retAI.getKind()) {
1533 case ABIArgInfo::Expand:
1534 llvm_unreachable("Invalid ABI kind for return argument");
1536 case ABIArgInfo::Extend:
1537 case ABIArgInfo::Direct:
1538 resultType = retAI.getCoerceToType();
1541 case ABIArgInfo::InAlloca:
1542 if (retAI.getInAllocaSRet()) {
1543 // sret things on win32 aren't void, they return the sret pointer.
1544 QualType ret = FI.getReturnType();
1545 llvm::Type *ty = ConvertType(ret);
1546 unsigned addressSpace = Context.getTargetAddressSpace(ret);
1547 resultType = llvm::PointerType::get(ty, addressSpace);
1549 resultType = llvm::Type::getVoidTy(getLLVMContext());
1553 case ABIArgInfo::Indirect:
1554 case ABIArgInfo::Ignore:
1555 resultType = llvm::Type::getVoidTy(getLLVMContext());
1558 case ABIArgInfo::CoerceAndExpand:
1559 resultType = retAI.getUnpaddedCoerceAndExpandType();
1563 ClangToLLVMArgMapping IRFunctionArgs(getContext(), FI, true);
1564 SmallVector<llvm::Type*, 8> ArgTypes(IRFunctionArgs.totalIRArgs());
1566 // Add type for sret argument.
1567 if (IRFunctionArgs.hasSRetArg()) {
1568 QualType Ret = FI.getReturnType();
1569 llvm::Type *Ty = ConvertType(Ret);
1570 unsigned AddressSpace = Context.getTargetAddressSpace(Ret);
1571 ArgTypes[IRFunctionArgs.getSRetArgNo()] =
1572 llvm::PointerType::get(Ty, AddressSpace);
1575 // Add type for inalloca argument.
1576 if (IRFunctionArgs.hasInallocaArg()) {
1577 auto ArgStruct = FI.getArgStruct();
1579 ArgTypes[IRFunctionArgs.getInallocaArgNo()] = ArgStruct->getPointerTo();
1582 // Add in all of the required arguments.
1584 CGFunctionInfo::const_arg_iterator it = FI.arg_begin(),
1585 ie = it + FI.getNumRequiredArgs();
1586 for (; it != ie; ++it, ++ArgNo) {
1587 const ABIArgInfo &ArgInfo = it->info;
1589 // Insert a padding type to ensure proper alignment.
1590 if (IRFunctionArgs.hasPaddingArg(ArgNo))
1591 ArgTypes[IRFunctionArgs.getPaddingArgNo(ArgNo)] =
1592 ArgInfo.getPaddingType();
1594 unsigned FirstIRArg, NumIRArgs;
1595 std::tie(FirstIRArg, NumIRArgs) = IRFunctionArgs.getIRArgs(ArgNo);
1597 switch (ArgInfo.getKind()) {
1598 case ABIArgInfo::Ignore:
1599 case ABIArgInfo::InAlloca:
1600 assert(NumIRArgs == 0);
1603 case ABIArgInfo::Indirect: {
1604 assert(NumIRArgs == 1);
1605 // indirect arguments are always on the stack, which is alloca addr space.
1606 llvm::Type *LTy = ConvertTypeForMem(it->type);
1607 ArgTypes[FirstIRArg] = LTy->getPointerTo(
1608 CGM.getDataLayout().getAllocaAddrSpace());
1612 case ABIArgInfo::Extend:
1613 case ABIArgInfo::Direct: {
1614 // Fast-isel and the optimizer generally like scalar values better than
1615 // FCAs, so we flatten them if this is safe to do for this argument.
1616 llvm::Type *argType = ArgInfo.getCoerceToType();
1617 llvm::StructType *st = dyn_cast<llvm::StructType>(argType);
1618 if (st && ArgInfo.isDirect() && ArgInfo.getCanBeFlattened()) {
1619 assert(NumIRArgs == st->getNumElements());
1620 for (unsigned i = 0, e = st->getNumElements(); i != e; ++i)
1621 ArgTypes[FirstIRArg + i] = st->getElementType(i);
1623 assert(NumIRArgs == 1);
1624 ArgTypes[FirstIRArg] = argType;
1629 case ABIArgInfo::CoerceAndExpand: {
1630 auto ArgTypesIter = ArgTypes.begin() + FirstIRArg;
1631 for (auto EltTy : ArgInfo.getCoerceAndExpandTypeSequence()) {
1632 *ArgTypesIter++ = EltTy;
1634 assert(ArgTypesIter == ArgTypes.begin() + FirstIRArg + NumIRArgs);
1638 case ABIArgInfo::Expand:
1639 auto ArgTypesIter = ArgTypes.begin() + FirstIRArg;
1640 getExpandedTypes(it->type, ArgTypesIter);
1641 assert(ArgTypesIter == ArgTypes.begin() + FirstIRArg + NumIRArgs);
1646 bool Erased = FunctionsBeingProcessed.erase(&FI); (void)Erased;
1647 assert(Erased && "Not in set?");
1649 return llvm::FunctionType::get(resultType, ArgTypes, FI.isVariadic());
1652 llvm::Type *CodeGenTypes::GetFunctionTypeForVTable(GlobalDecl GD) {
1653 const CXXMethodDecl *MD = cast<CXXMethodDecl>(GD.getDecl());
1654 const FunctionProtoType *FPT = MD->getType()->getAs<FunctionProtoType>();
1656 if (!isFuncTypeConvertible(FPT))
1657 return llvm::StructType::get(getLLVMContext());
1659 const CGFunctionInfo *Info;
1660 if (isa<CXXDestructorDecl>(MD))
1662 &arrangeCXXStructorDeclaration(MD, getFromDtorType(GD.getDtorType()));
1664 Info = &arrangeCXXMethodDeclaration(MD);
1665 return GetFunctionType(*Info);
1668 static void AddAttributesFromFunctionProtoType(ASTContext &Ctx,
1669 llvm::AttrBuilder &FuncAttrs,
1670 const FunctionProtoType *FPT) {
1674 if (!isUnresolvedExceptionSpec(FPT->getExceptionSpecType()) &&
1675 FPT->isNothrow(Ctx))
1676 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1679 void CodeGenModule::ConstructDefaultFnAttrList(StringRef Name, bool HasOptnone,
1680 bool AttrOnCallSite,
1681 llvm::AttrBuilder &FuncAttrs) {
1682 // OptimizeNoneAttr takes precedence over -Os or -Oz. No warning needed.
1684 if (CodeGenOpts.OptimizeSize)
1685 FuncAttrs.addAttribute(llvm::Attribute::OptimizeForSize);
1686 if (CodeGenOpts.OptimizeSize == 2)
1687 FuncAttrs.addAttribute(llvm::Attribute::MinSize);
1690 if (CodeGenOpts.DisableRedZone)
1691 FuncAttrs.addAttribute(llvm::Attribute::NoRedZone);
1692 if (CodeGenOpts.NoImplicitFloat)
1693 FuncAttrs.addAttribute(llvm::Attribute::NoImplicitFloat);
1695 if (AttrOnCallSite) {
1696 // Attributes that should go on the call site only.
1697 if (!CodeGenOpts.SimplifyLibCalls ||
1698 CodeGenOpts.isNoBuiltinFunc(Name.data()))
1699 FuncAttrs.addAttribute(llvm::Attribute::NoBuiltin);
1700 if (!CodeGenOpts.TrapFuncName.empty())
1701 FuncAttrs.addAttribute("trap-func-name", CodeGenOpts.TrapFuncName);
1703 // Attributes that should go on the function, but not the call site.
1704 if (!CodeGenOpts.DisableFPElim) {
1705 FuncAttrs.addAttribute("no-frame-pointer-elim", "false");
1706 } else if (CodeGenOpts.OmitLeafFramePointer) {
1707 FuncAttrs.addAttribute("no-frame-pointer-elim", "false");
1708 FuncAttrs.addAttribute("no-frame-pointer-elim-non-leaf");
1710 FuncAttrs.addAttribute("no-frame-pointer-elim", "true");
1711 FuncAttrs.addAttribute("no-frame-pointer-elim-non-leaf");
1714 FuncAttrs.addAttribute("less-precise-fpmad",
1715 llvm::toStringRef(CodeGenOpts.LessPreciseFPMAD));
1717 if (!CodeGenOpts.FPDenormalMode.empty())
1718 FuncAttrs.addAttribute("denormal-fp-math", CodeGenOpts.FPDenormalMode);
1720 FuncAttrs.addAttribute("no-trapping-math",
1721 llvm::toStringRef(CodeGenOpts.NoTrappingMath));
1723 // TODO: Are these all needed?
1724 // unsafe/inf/nan/nsz are handled by instruction-level FastMathFlags.
1725 FuncAttrs.addAttribute("no-infs-fp-math",
1726 llvm::toStringRef(CodeGenOpts.NoInfsFPMath));
1727 FuncAttrs.addAttribute("no-nans-fp-math",
1728 llvm::toStringRef(CodeGenOpts.NoNaNsFPMath));
1729 FuncAttrs.addAttribute("unsafe-fp-math",
1730 llvm::toStringRef(CodeGenOpts.UnsafeFPMath));
1731 FuncAttrs.addAttribute("use-soft-float",
1732 llvm::toStringRef(CodeGenOpts.SoftFloat));
1733 FuncAttrs.addAttribute("stack-protector-buffer-size",
1734 llvm::utostr(CodeGenOpts.SSPBufferSize));
1735 FuncAttrs.addAttribute("no-signed-zeros-fp-math",
1736 llvm::toStringRef(CodeGenOpts.NoSignedZeros));
1737 FuncAttrs.addAttribute(
1738 "correctly-rounded-divide-sqrt-fp-math",
1739 llvm::toStringRef(CodeGenOpts.CorrectlyRoundedDivSqrt));
1741 // TODO: Reciprocal estimate codegen options should apply to instructions?
1742 const std::vector<std::string> &Recips = CodeGenOpts.Reciprocals;
1743 if (!Recips.empty())
1744 FuncAttrs.addAttribute("reciprocal-estimates",
1745 llvm::join(Recips, ","));
1747 if (!CodeGenOpts.PreferVectorWidth.empty() &&
1748 CodeGenOpts.PreferVectorWidth != "none")
1749 FuncAttrs.addAttribute("prefer-vector-width",
1750 CodeGenOpts.PreferVectorWidth);
1752 if (CodeGenOpts.StackRealignment)
1753 FuncAttrs.addAttribute("stackrealign");
1754 if (CodeGenOpts.Backchain)
1755 FuncAttrs.addAttribute("backchain");
1758 if (getLangOpts().assumeFunctionsAreConvergent()) {
1759 // Conservatively, mark all functions and calls in CUDA and OpenCL as
1760 // convergent (meaning, they may call an intrinsically convergent op, such
1761 // as __syncthreads() / barrier(), and so can't have certain optimizations
1762 // applied around them). LLVM will remove this attribute where it safely
1764 FuncAttrs.addAttribute(llvm::Attribute::Convergent);
1767 if (getLangOpts().CUDA && getLangOpts().CUDAIsDevice) {
1768 // Exceptions aren't supported in CUDA device code.
1769 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1771 // Respect -fcuda-flush-denormals-to-zero.
1772 if (getLangOpts().CUDADeviceFlushDenormalsToZero)
1773 FuncAttrs.addAttribute("nvptx-f32ftz", "true");
1777 void CodeGenModule::AddDefaultFnAttrs(llvm::Function &F) {
1778 llvm::AttrBuilder FuncAttrs;
1779 ConstructDefaultFnAttrList(F.getName(),
1780 F.hasFnAttribute(llvm::Attribute::OptimizeNone),
1781 /* AttrOnCallsite = */ false, FuncAttrs);
1782 F.addAttributes(llvm::AttributeList::FunctionIndex, FuncAttrs);
1785 void CodeGenModule::ConstructAttributeList(
1786 StringRef Name, const CGFunctionInfo &FI, CGCalleeInfo CalleeInfo,
1787 llvm::AttributeList &AttrList, unsigned &CallingConv, bool AttrOnCallSite) {
1788 llvm::AttrBuilder FuncAttrs;
1789 llvm::AttrBuilder RetAttrs;
1791 CallingConv = FI.getEffectiveCallingConvention();
1792 if (FI.isNoReturn())
1793 FuncAttrs.addAttribute(llvm::Attribute::NoReturn);
1795 // If we have information about the function prototype, we can learn
1796 // attributes form there.
1797 AddAttributesFromFunctionProtoType(getContext(), FuncAttrs,
1798 CalleeInfo.getCalleeFunctionProtoType());
1800 const Decl *TargetDecl = CalleeInfo.getCalleeDecl();
1802 bool HasOptnone = false;
1803 // FIXME: handle sseregparm someday...
1805 if (TargetDecl->hasAttr<ReturnsTwiceAttr>())
1806 FuncAttrs.addAttribute(llvm::Attribute::ReturnsTwice);
1807 if (TargetDecl->hasAttr<NoThrowAttr>())
1808 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1809 if (TargetDecl->hasAttr<NoReturnAttr>())
1810 FuncAttrs.addAttribute(llvm::Attribute::NoReturn);
1811 if (TargetDecl->hasAttr<ColdAttr>())
1812 FuncAttrs.addAttribute(llvm::Attribute::Cold);
1813 if (TargetDecl->hasAttr<NoDuplicateAttr>())
1814 FuncAttrs.addAttribute(llvm::Attribute::NoDuplicate);
1815 if (TargetDecl->hasAttr<ConvergentAttr>())
1816 FuncAttrs.addAttribute(llvm::Attribute::Convergent);
1818 if (const FunctionDecl *Fn = dyn_cast<FunctionDecl>(TargetDecl)) {
1819 AddAttributesFromFunctionProtoType(
1820 getContext(), FuncAttrs, Fn->getType()->getAs<FunctionProtoType>());
1821 // Don't use [[noreturn]] or _Noreturn for a call to a virtual function.
1822 // These attributes are not inherited by overloads.
1823 const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(Fn);
1824 if (Fn->isNoReturn() && !(AttrOnCallSite && MD && MD->isVirtual()))
1825 FuncAttrs.addAttribute(llvm::Attribute::NoReturn);
1828 // 'const', 'pure' and 'noalias' attributed functions are also nounwind.
1829 if (TargetDecl->hasAttr<ConstAttr>()) {
1830 FuncAttrs.addAttribute(llvm::Attribute::ReadNone);
1831 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1832 } else if (TargetDecl->hasAttr<PureAttr>()) {
1833 FuncAttrs.addAttribute(llvm::Attribute::ReadOnly);
1834 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1835 } else if (TargetDecl->hasAttr<NoAliasAttr>()) {
1836 FuncAttrs.addAttribute(llvm::Attribute::ArgMemOnly);
1837 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1839 if (TargetDecl->hasAttr<RestrictAttr>())
1840 RetAttrs.addAttribute(llvm::Attribute::NoAlias);
1841 if (TargetDecl->hasAttr<ReturnsNonNullAttr>())
1842 RetAttrs.addAttribute(llvm::Attribute::NonNull);
1843 if (TargetDecl->hasAttr<AnyX86NoCallerSavedRegistersAttr>())
1844 FuncAttrs.addAttribute("no_caller_saved_registers");
1846 HasOptnone = TargetDecl->hasAttr<OptimizeNoneAttr>();
1847 if (auto *AllocSize = TargetDecl->getAttr<AllocSizeAttr>()) {
1848 Optional<unsigned> NumElemsParam;
1849 // alloc_size args are base-1, 0 means not present.
1850 if (unsigned N = AllocSize->getNumElemsParam())
1851 NumElemsParam = N - 1;
1852 FuncAttrs.addAllocSizeAttr(AllocSize->getElemSizeParam() - 1,
1857 ConstructDefaultFnAttrList(Name, HasOptnone, AttrOnCallSite, FuncAttrs);
1859 if (CodeGenOpts.EnableSegmentedStacks &&
1860 !(TargetDecl && TargetDecl->hasAttr<NoSplitStackAttr>()))
1861 FuncAttrs.addAttribute("split-stack");
1863 // Add NonLazyBind attribute to function declarations when -fno-plt
1865 if (TargetDecl && CodeGenOpts.NoPLT) {
1866 if (auto *Fn = dyn_cast<FunctionDecl>(TargetDecl)) {
1867 if (!Fn->isDefined() && !AttrOnCallSite) {
1868 FuncAttrs.addAttribute(llvm::Attribute::NonLazyBind);
1873 if (!AttrOnCallSite) {
1874 bool DisableTailCalls =
1875 CodeGenOpts.DisableTailCalls ||
1876 (TargetDecl && (TargetDecl->hasAttr<DisableTailCallsAttr>() ||
1877 TargetDecl->hasAttr<AnyX86InterruptAttr>()));
1878 FuncAttrs.addAttribute("disable-tail-calls",
1879 llvm::toStringRef(DisableTailCalls));
1881 // Add target-cpu and target-features attributes to functions. If
1882 // we have a decl for the function and it has a target attribute then
1883 // parse that and add it to the feature set.
1884 StringRef TargetCPU = getTarget().getTargetOpts().CPU;
1885 std::vector<std::string> Features;
1886 const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(TargetDecl);
1887 if (FD && FD->hasAttr<TargetAttr>()) {
1888 llvm::StringMap<bool> FeatureMap;
1889 getFunctionFeatureMap(FeatureMap, FD);
1891 // Produce the canonical string for this set of features.
1892 for (llvm::StringMap<bool>::const_iterator it = FeatureMap.begin(),
1893 ie = FeatureMap.end();
1895 Features.push_back((it->second ? "+" : "-") + it->first().str());
1897 // Now add the target-cpu and target-features to the function.
1898 // While we populated the feature map above, we still need to
1899 // get and parse the target attribute so we can get the cpu for
1901 const auto *TD = FD->getAttr<TargetAttr>();
1902 TargetAttr::ParsedTargetAttr ParsedAttr = TD->parse();
1903 if (ParsedAttr.Architecture != "" &&
1904 getTarget().isValidCPUName(ParsedAttr.Architecture))
1905 TargetCPU = ParsedAttr.Architecture;
1907 // Otherwise just add the existing target cpu and target features to the
1909 Features = getTarget().getTargetOpts().Features;
1912 if (TargetCPU != "")
1913 FuncAttrs.addAttribute("target-cpu", TargetCPU);
1914 if (!Features.empty()) {
1915 std::sort(Features.begin(), Features.end());
1916 FuncAttrs.addAttribute(
1918 llvm::join(Features, ","));
1922 ClangToLLVMArgMapping IRFunctionArgs(getContext(), FI);
1924 QualType RetTy = FI.getReturnType();
1925 const ABIArgInfo &RetAI = FI.getReturnInfo();
1926 switch (RetAI.getKind()) {
1927 case ABIArgInfo::Extend:
1928 if (RetTy->hasSignedIntegerRepresentation())
1929 RetAttrs.addAttribute(llvm::Attribute::SExt);
1930 else if (RetTy->hasUnsignedIntegerRepresentation())
1931 RetAttrs.addAttribute(llvm::Attribute::ZExt);
1933 case ABIArgInfo::Direct:
1934 if (RetAI.getInReg())
1935 RetAttrs.addAttribute(llvm::Attribute::InReg);
1937 case ABIArgInfo::Ignore:
1940 case ABIArgInfo::InAlloca:
1941 case ABIArgInfo::Indirect: {
1942 // inalloca and sret disable readnone and readonly
1943 FuncAttrs.removeAttribute(llvm::Attribute::ReadOnly)
1944 .removeAttribute(llvm::Attribute::ReadNone);
1948 case ABIArgInfo::CoerceAndExpand:
1951 case ABIArgInfo::Expand:
1952 llvm_unreachable("Invalid ABI kind for return argument");
1955 if (const auto *RefTy = RetTy->getAs<ReferenceType>()) {
1956 QualType PTy = RefTy->getPointeeType();
1957 if (!PTy->isIncompleteType() && PTy->isConstantSizeType())
1958 RetAttrs.addDereferenceableAttr(getContext().getTypeSizeInChars(PTy)
1960 else if (getContext().getTargetAddressSpace(PTy) == 0)
1961 RetAttrs.addAttribute(llvm::Attribute::NonNull);
1964 bool hasUsedSRet = false;
1965 SmallVector<llvm::AttributeSet, 4> ArgAttrs(IRFunctionArgs.totalIRArgs());
1967 // Attach attributes to sret.
1968 if (IRFunctionArgs.hasSRetArg()) {
1969 llvm::AttrBuilder SRETAttrs;
1970 SRETAttrs.addAttribute(llvm::Attribute::StructRet);
1972 if (RetAI.getInReg())
1973 SRETAttrs.addAttribute(llvm::Attribute::InReg);
1974 ArgAttrs[IRFunctionArgs.getSRetArgNo()] =
1975 llvm::AttributeSet::get(getLLVMContext(), SRETAttrs);
1978 // Attach attributes to inalloca argument.
1979 if (IRFunctionArgs.hasInallocaArg()) {
1980 llvm::AttrBuilder Attrs;
1981 Attrs.addAttribute(llvm::Attribute::InAlloca);
1982 ArgAttrs[IRFunctionArgs.getInallocaArgNo()] =
1983 llvm::AttributeSet::get(getLLVMContext(), Attrs);
1987 for (CGFunctionInfo::const_arg_iterator I = FI.arg_begin(),
1989 I != E; ++I, ++ArgNo) {
1990 QualType ParamType = I->type;
1991 const ABIArgInfo &AI = I->info;
1992 llvm::AttrBuilder Attrs;
1994 // Add attribute for padding argument, if necessary.
1995 if (IRFunctionArgs.hasPaddingArg(ArgNo)) {
1996 if (AI.getPaddingInReg()) {
1997 ArgAttrs[IRFunctionArgs.getPaddingArgNo(ArgNo)] =
1998 llvm::AttributeSet::get(
2000 llvm::AttrBuilder().addAttribute(llvm::Attribute::InReg));
2004 // 'restrict' -> 'noalias' is done in EmitFunctionProlog when we
2005 // have the corresponding parameter variable. It doesn't make
2006 // sense to do it here because parameters are so messed up.
2007 switch (AI.getKind()) {
2008 case ABIArgInfo::Extend:
2009 if (ParamType->isSignedIntegerOrEnumerationType())
2010 Attrs.addAttribute(llvm::Attribute::SExt);
2011 else if (ParamType->isUnsignedIntegerOrEnumerationType()) {
2012 if (getTypes().getABIInfo().shouldSignExtUnsignedType(ParamType))
2013 Attrs.addAttribute(llvm::Attribute::SExt);
2015 Attrs.addAttribute(llvm::Attribute::ZExt);
2018 case ABIArgInfo::Direct:
2019 if (ArgNo == 0 && FI.isChainCall())
2020 Attrs.addAttribute(llvm::Attribute::Nest);
2021 else if (AI.getInReg())
2022 Attrs.addAttribute(llvm::Attribute::InReg);
2025 case ABIArgInfo::Indirect: {
2027 Attrs.addAttribute(llvm::Attribute::InReg);
2029 if (AI.getIndirectByVal())
2030 Attrs.addAttribute(llvm::Attribute::ByVal);
2032 CharUnits Align = AI.getIndirectAlign();
2034 // In a byval argument, it is important that the required
2035 // alignment of the type is honored, as LLVM might be creating a
2036 // *new* stack object, and needs to know what alignment to give
2037 // it. (Sometimes it can deduce a sensible alignment on its own,
2038 // but not if clang decides it must emit a packed struct, or the
2039 // user specifies increased alignment requirements.)
2041 // This is different from indirect *not* byval, where the object
2042 // exists already, and the align attribute is purely
2044 assert(!Align.isZero());
2046 // For now, only add this when we have a byval argument.
2047 // TODO: be less lazy about updating test cases.
2048 if (AI.getIndirectByVal())
2049 Attrs.addAlignmentAttr(Align.getQuantity());
2051 // byval disables readnone and readonly.
2052 FuncAttrs.removeAttribute(llvm::Attribute::ReadOnly)
2053 .removeAttribute(llvm::Attribute::ReadNone);
2056 case ABIArgInfo::Ignore:
2057 case ABIArgInfo::Expand:
2058 case ABIArgInfo::CoerceAndExpand:
2061 case ABIArgInfo::InAlloca:
2062 // inalloca disables readnone and readonly.
2063 FuncAttrs.removeAttribute(llvm::Attribute::ReadOnly)
2064 .removeAttribute(llvm::Attribute::ReadNone);
2068 if (const auto *RefTy = ParamType->getAs<ReferenceType>()) {
2069 QualType PTy = RefTy->getPointeeType();
2070 if (!PTy->isIncompleteType() && PTy->isConstantSizeType())
2071 Attrs.addDereferenceableAttr(getContext().getTypeSizeInChars(PTy)
2073 else if (getContext().getTargetAddressSpace(PTy) == 0)
2074 Attrs.addAttribute(llvm::Attribute::NonNull);
2077 switch (FI.getExtParameterInfo(ArgNo).getABI()) {
2078 case ParameterABI::Ordinary:
2081 case ParameterABI::SwiftIndirectResult: {
2082 // Add 'sret' if we haven't already used it for something, but
2083 // only if the result is void.
2084 if (!hasUsedSRet && RetTy->isVoidType()) {
2085 Attrs.addAttribute(llvm::Attribute::StructRet);
2089 // Add 'noalias' in either case.
2090 Attrs.addAttribute(llvm::Attribute::NoAlias);
2092 // Add 'dereferenceable' and 'alignment'.
2093 auto PTy = ParamType->getPointeeType();
2094 if (!PTy->isIncompleteType() && PTy->isConstantSizeType()) {
2095 auto info = getContext().getTypeInfoInChars(PTy);
2096 Attrs.addDereferenceableAttr(info.first.getQuantity());
2097 Attrs.addAttribute(llvm::Attribute::getWithAlignment(getLLVMContext(),
2098 info.second.getQuantity()));
2103 case ParameterABI::SwiftErrorResult:
2104 Attrs.addAttribute(llvm::Attribute::SwiftError);
2107 case ParameterABI::SwiftContext:
2108 Attrs.addAttribute(llvm::Attribute::SwiftSelf);
2112 if (FI.getExtParameterInfo(ArgNo).isNoEscape())
2113 Attrs.addAttribute(llvm::Attribute::NoCapture);
2115 if (Attrs.hasAttributes()) {
2116 unsigned FirstIRArg, NumIRArgs;
2117 std::tie(FirstIRArg, NumIRArgs) = IRFunctionArgs.getIRArgs(ArgNo);
2118 for (unsigned i = 0; i < NumIRArgs; i++)
2119 ArgAttrs[FirstIRArg + i] =
2120 llvm::AttributeSet::get(getLLVMContext(), Attrs);
2123 assert(ArgNo == FI.arg_size());
2125 AttrList = llvm::AttributeList::get(
2126 getLLVMContext(), llvm::AttributeSet::get(getLLVMContext(), FuncAttrs),
2127 llvm::AttributeSet::get(getLLVMContext(), RetAttrs), ArgAttrs);
2130 /// An argument came in as a promoted argument; demote it back to its
2132 static llvm::Value *emitArgumentDemotion(CodeGenFunction &CGF,
2134 llvm::Value *value) {
2135 llvm::Type *varType = CGF.ConvertType(var->getType());
2137 // This can happen with promotions that actually don't change the
2138 // underlying type, like the enum promotions.
2139 if (value->getType() == varType) return value;
2141 assert((varType->isIntegerTy() || varType->isFloatingPointTy())
2142 && "unexpected promotion type");
2144 if (isa<llvm::IntegerType>(varType))
2145 return CGF.Builder.CreateTrunc(value, varType, "arg.unpromote");
2147 return CGF.Builder.CreateFPCast(value, varType, "arg.unpromote");
2150 /// Returns the attribute (either parameter attribute, or function
2151 /// attribute), which declares argument ArgNo to be non-null.
2152 static const NonNullAttr *getNonNullAttr(const Decl *FD, const ParmVarDecl *PVD,
2153 QualType ArgType, unsigned ArgNo) {
2154 // FIXME: __attribute__((nonnull)) can also be applied to:
2155 // - references to pointers, where the pointee is known to be
2156 // nonnull (apparently a Clang extension)
2157 // - transparent unions containing pointers
2158 // In the former case, LLVM IR cannot represent the constraint. In
2159 // the latter case, we have no guarantee that the transparent union
2160 // is in fact passed as a pointer.
2161 if (!ArgType->isAnyPointerType() && !ArgType->isBlockPointerType())
2163 // First, check attribute on parameter itself.
2165 if (auto ParmNNAttr = PVD->getAttr<NonNullAttr>())
2168 // Check function attributes.
2171 for (const auto *NNAttr : FD->specific_attrs<NonNullAttr>()) {
2172 if (NNAttr->isNonNull(ArgNo))
2179 struct CopyBackSwiftError final : EHScopeStack::Cleanup {
2182 CopyBackSwiftError(Address temp, Address arg) : Temp(temp), Arg(arg) {}
2183 void Emit(CodeGenFunction &CGF, Flags flags) override {
2184 llvm::Value *errorValue = CGF.Builder.CreateLoad(Temp);
2185 CGF.Builder.CreateStore(errorValue, Arg);
2190 void CodeGenFunction::EmitFunctionProlog(const CGFunctionInfo &FI,
2192 const FunctionArgList &Args) {
2193 if (CurCodeDecl && CurCodeDecl->hasAttr<NakedAttr>())
2194 // Naked functions don't have prologues.
2197 // If this is an implicit-return-zero function, go ahead and
2198 // initialize the return value. TODO: it might be nice to have
2199 // a more general mechanism for this that didn't require synthesized
2200 // return statements.
2201 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(CurCodeDecl)) {
2202 if (FD->hasImplicitReturnZero()) {
2203 QualType RetTy = FD->getReturnType().getUnqualifiedType();
2204 llvm::Type* LLVMTy = CGM.getTypes().ConvertType(RetTy);
2205 llvm::Constant* Zero = llvm::Constant::getNullValue(LLVMTy);
2206 Builder.CreateStore(Zero, ReturnValue);
2210 // FIXME: We no longer need the types from FunctionArgList; lift up and
2213 ClangToLLVMArgMapping IRFunctionArgs(CGM.getContext(), FI);
2214 // Flattened function arguments.
2215 SmallVector<llvm::Value *, 16> FnArgs;
2216 FnArgs.reserve(IRFunctionArgs.totalIRArgs());
2217 for (auto &Arg : Fn->args()) {
2218 FnArgs.push_back(&Arg);
2220 assert(FnArgs.size() == IRFunctionArgs.totalIRArgs());
2222 // If we're using inalloca, all the memory arguments are GEPs off of the last
2223 // parameter, which is a pointer to the complete memory area.
2224 Address ArgStruct = Address::invalid();
2225 const llvm::StructLayout *ArgStructLayout = nullptr;
2226 if (IRFunctionArgs.hasInallocaArg()) {
2227 ArgStructLayout = CGM.getDataLayout().getStructLayout(FI.getArgStruct());
2228 ArgStruct = Address(FnArgs[IRFunctionArgs.getInallocaArgNo()],
2229 FI.getArgStructAlignment());
2231 assert(ArgStruct.getType() == FI.getArgStruct()->getPointerTo());
2234 // Name the struct return parameter.
2235 if (IRFunctionArgs.hasSRetArg()) {
2236 auto AI = cast<llvm::Argument>(FnArgs[IRFunctionArgs.getSRetArgNo()]);
2237 AI->setName("agg.result");
2238 AI->addAttr(llvm::Attribute::NoAlias);
2241 // Track if we received the parameter as a pointer (indirect, byval, or
2242 // inalloca). If already have a pointer, EmitParmDecl doesn't need to copy it
2243 // into a local alloca for us.
2244 SmallVector<ParamValue, 16> ArgVals;
2245 ArgVals.reserve(Args.size());
2247 // Create a pointer value for every parameter declaration. This usually
2248 // entails copying one or more LLVM IR arguments into an alloca. Don't push
2249 // any cleanups or do anything that might unwind. We do that separately, so
2250 // we can push the cleanups in the correct order for the ABI.
2251 assert(FI.arg_size() == Args.size() &&
2252 "Mismatch between function signature & arguments.");
2254 CGFunctionInfo::const_arg_iterator info_it = FI.arg_begin();
2255 for (FunctionArgList::const_iterator i = Args.begin(), e = Args.end();
2256 i != e; ++i, ++info_it, ++ArgNo) {
2257 const VarDecl *Arg = *i;
2258 QualType Ty = info_it->type;
2259 const ABIArgInfo &ArgI = info_it->info;
2262 isa<ParmVarDecl>(Arg) && cast<ParmVarDecl>(Arg)->isKNRPromoted();
2264 unsigned FirstIRArg, NumIRArgs;
2265 std::tie(FirstIRArg, NumIRArgs) = IRFunctionArgs.getIRArgs(ArgNo);
2267 switch (ArgI.getKind()) {
2268 case ABIArgInfo::InAlloca: {
2269 assert(NumIRArgs == 0);
2270 auto FieldIndex = ArgI.getInAllocaFieldIndex();
2271 CharUnits FieldOffset =
2272 CharUnits::fromQuantity(ArgStructLayout->getElementOffset(FieldIndex));
2273 Address V = Builder.CreateStructGEP(ArgStruct, FieldIndex, FieldOffset,
2275 ArgVals.push_back(ParamValue::forIndirect(V));
2279 case ABIArgInfo::Indirect: {
2280 assert(NumIRArgs == 1);
2281 Address ParamAddr = Address(FnArgs[FirstIRArg], ArgI.getIndirectAlign());
2283 if (!hasScalarEvaluationKind(Ty)) {
2284 // Aggregates and complex variables are accessed by reference. All we
2285 // need to do is realign the value, if requested.
2286 Address V = ParamAddr;
2287 if (ArgI.getIndirectRealign()) {
2288 Address AlignedTemp = CreateMemTemp(Ty, "coerce");
2290 // Copy from the incoming argument pointer to the temporary with the
2291 // appropriate alignment.
2293 // FIXME: We should have a common utility for generating an aggregate
2295 CharUnits Size = getContext().getTypeSizeInChars(Ty);
2296 auto SizeVal = llvm::ConstantInt::get(IntPtrTy, Size.getQuantity());
2297 Address Dst = Builder.CreateBitCast(AlignedTemp, Int8PtrTy);
2298 Address Src = Builder.CreateBitCast(ParamAddr, Int8PtrTy);
2299 Builder.CreateMemCpy(Dst, Src, SizeVal, false);
2302 ArgVals.push_back(ParamValue::forIndirect(V));
2304 // Load scalar value from indirect argument.
2306 EmitLoadOfScalar(ParamAddr, false, Ty, Arg->getLocStart());
2309 V = emitArgumentDemotion(*this, Arg, V);
2310 ArgVals.push_back(ParamValue::forDirect(V));
2315 case ABIArgInfo::Extend:
2316 case ABIArgInfo::Direct: {
2318 // If we have the trivial case, handle it with no muss and fuss.
2319 if (!isa<llvm::StructType>(ArgI.getCoerceToType()) &&
2320 ArgI.getCoerceToType() == ConvertType(Ty) &&
2321 ArgI.getDirectOffset() == 0) {
2322 assert(NumIRArgs == 1);
2323 llvm::Value *V = FnArgs[FirstIRArg];
2324 auto AI = cast<llvm::Argument>(V);
2326 if (const ParmVarDecl *PVD = dyn_cast<ParmVarDecl>(Arg)) {
2327 if (getNonNullAttr(CurCodeDecl, PVD, PVD->getType(),
2328 PVD->getFunctionScopeIndex()))
2329 AI->addAttr(llvm::Attribute::NonNull);
2331 QualType OTy = PVD->getOriginalType();
2332 if (const auto *ArrTy =
2333 getContext().getAsConstantArrayType(OTy)) {
2334 // A C99 array parameter declaration with the static keyword also
2335 // indicates dereferenceability, and if the size is constant we can
2336 // use the dereferenceable attribute (which requires the size in
2338 if (ArrTy->getSizeModifier() == ArrayType::Static) {
2339 QualType ETy = ArrTy->getElementType();
2340 uint64_t ArrSize = ArrTy->getSize().getZExtValue();
2341 if (!ETy->isIncompleteType() && ETy->isConstantSizeType() &&
2343 llvm::AttrBuilder Attrs;
2344 Attrs.addDereferenceableAttr(
2345 getContext().getTypeSizeInChars(ETy).getQuantity()*ArrSize);
2346 AI->addAttrs(Attrs);
2347 } else if (getContext().getTargetAddressSpace(ETy) == 0) {
2348 AI->addAttr(llvm::Attribute::NonNull);
2351 } else if (const auto *ArrTy =
2352 getContext().getAsVariableArrayType(OTy)) {
2353 // For C99 VLAs with the static keyword, we don't know the size so
2354 // we can't use the dereferenceable attribute, but in addrspace(0)
2355 // we know that it must be nonnull.
2356 if (ArrTy->getSizeModifier() == VariableArrayType::Static &&
2357 !getContext().getTargetAddressSpace(ArrTy->getElementType()))
2358 AI->addAttr(llvm::Attribute::NonNull);
2361 const auto *AVAttr = PVD->getAttr<AlignValueAttr>();
2363 if (const auto *TOTy = dyn_cast<TypedefType>(OTy))
2364 AVAttr = TOTy->getDecl()->getAttr<AlignValueAttr>();
2366 llvm::Value *AlignmentValue =
2367 EmitScalarExpr(AVAttr->getAlignment());
2368 llvm::ConstantInt *AlignmentCI =
2369 cast<llvm::ConstantInt>(AlignmentValue);
2370 unsigned Alignment = std::min((unsigned)AlignmentCI->getZExtValue(),
2371 +llvm::Value::MaximumAlignment);
2372 AI->addAttrs(llvm::AttrBuilder().addAlignmentAttr(Alignment));
2376 if (Arg->getType().isRestrictQualified())
2377 AI->addAttr(llvm::Attribute::NoAlias);
2379 // LLVM expects swifterror parameters to be used in very restricted
2380 // ways. Copy the value into a less-restricted temporary.
2381 if (FI.getExtParameterInfo(ArgNo).getABI()
2382 == ParameterABI::SwiftErrorResult) {
2383 QualType pointeeTy = Ty->getPointeeType();
2384 assert(pointeeTy->isPointerType());
2386 CreateMemTemp(pointeeTy, getPointerAlign(), "swifterror.temp");
2387 Address arg = Address(V, getContext().getTypeAlignInChars(pointeeTy));
2388 llvm::Value *incomingErrorValue = Builder.CreateLoad(arg);
2389 Builder.CreateStore(incomingErrorValue, temp);
2390 V = temp.getPointer();
2392 // Push a cleanup to copy the value back at the end of the function.
2393 // The convention does not guarantee that the value will be written
2394 // back if the function exits with an unwind exception.
2395 EHStack.pushCleanup<CopyBackSwiftError>(NormalCleanup, temp, arg);
2398 // Ensure the argument is the correct type.
2399 if (V->getType() != ArgI.getCoerceToType())
2400 V = Builder.CreateBitCast(V, ArgI.getCoerceToType());
2403 V = emitArgumentDemotion(*this, Arg, V);
2405 // Because of merging of function types from multiple decls it is
2406 // possible for the type of an argument to not match the corresponding
2407 // type in the function type. Since we are codegening the callee
2408 // in here, add a cast to the argument type.
2409 llvm::Type *LTy = ConvertType(Arg->getType());
2410 if (V->getType() != LTy)
2411 V = Builder.CreateBitCast(V, LTy);
2413 ArgVals.push_back(ParamValue::forDirect(V));
2417 Address Alloca = CreateMemTemp(Ty, getContext().getDeclAlign(Arg),
2420 // Pointer to store into.
2421 Address Ptr = emitAddressAtOffset(*this, Alloca, ArgI);
2423 // Fast-isel and the optimizer generally like scalar values better than
2424 // FCAs, so we flatten them if this is safe to do for this argument.
2425 llvm::StructType *STy = dyn_cast<llvm::StructType>(ArgI.getCoerceToType());
2426 if (ArgI.isDirect() && ArgI.getCanBeFlattened() && STy &&
2427 STy->getNumElements() > 1) {
2428 auto SrcLayout = CGM.getDataLayout().getStructLayout(STy);
2429 uint64_t SrcSize = CGM.getDataLayout().getTypeAllocSize(STy);
2430 llvm::Type *DstTy = Ptr.getElementType();
2431 uint64_t DstSize = CGM.getDataLayout().getTypeAllocSize(DstTy);
2433 Address AddrToStoreInto = Address::invalid();
2434 if (SrcSize <= DstSize) {
2435 AddrToStoreInto = Builder.CreateElementBitCast(Ptr, STy);
2438 CreateTempAlloca(STy, Alloca.getAlignment(), "coerce");
2441 assert(STy->getNumElements() == NumIRArgs);
2442 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
2443 auto AI = FnArgs[FirstIRArg + i];
2444 AI->setName(Arg->getName() + ".coerce" + Twine(i));
2445 auto Offset = CharUnits::fromQuantity(SrcLayout->getElementOffset(i));
2447 Builder.CreateStructGEP(AddrToStoreInto, i, Offset);
2448 Builder.CreateStore(AI, EltPtr);
2451 if (SrcSize > DstSize) {
2452 Builder.CreateMemCpy(Ptr, AddrToStoreInto, DstSize);
2456 // Simple case, just do a coerced store of the argument into the alloca.
2457 assert(NumIRArgs == 1);
2458 auto AI = FnArgs[FirstIRArg];
2459 AI->setName(Arg->getName() + ".coerce");
2460 CreateCoercedStore(AI, Ptr, /*DestIsVolatile=*/false, *this);
2463 // Match to what EmitParmDecl is expecting for this type.
2464 if (CodeGenFunction::hasScalarEvaluationKind(Ty)) {
2466 EmitLoadOfScalar(Alloca, false, Ty, Arg->getLocStart());
2468 V = emitArgumentDemotion(*this, Arg, V);
2469 ArgVals.push_back(ParamValue::forDirect(V));
2471 ArgVals.push_back(ParamValue::forIndirect(Alloca));
2476 case ABIArgInfo::CoerceAndExpand: {
2477 // Reconstruct into a temporary.
2478 Address alloca = CreateMemTemp(Ty, getContext().getDeclAlign(Arg));
2479 ArgVals.push_back(ParamValue::forIndirect(alloca));
2481 auto coercionType = ArgI.getCoerceAndExpandType();
2482 alloca = Builder.CreateElementBitCast(alloca, coercionType);
2483 auto layout = CGM.getDataLayout().getStructLayout(coercionType);
2485 unsigned argIndex = FirstIRArg;
2486 for (unsigned i = 0, e = coercionType->getNumElements(); i != e; ++i) {
2487 llvm::Type *eltType = coercionType->getElementType(i);
2488 if (ABIArgInfo::isPaddingForCoerceAndExpand(eltType))
2491 auto eltAddr = Builder.CreateStructGEP(alloca, i, layout);
2492 auto elt = FnArgs[argIndex++];
2493 Builder.CreateStore(elt, eltAddr);
2495 assert(argIndex == FirstIRArg + NumIRArgs);
2499 case ABIArgInfo::Expand: {
2500 // If this structure was expanded into multiple arguments then
2501 // we need to create a temporary and reconstruct it from the
2503 Address Alloca = CreateMemTemp(Ty, getContext().getDeclAlign(Arg));
2504 LValue LV = MakeAddrLValue(Alloca, Ty);
2505 ArgVals.push_back(ParamValue::forIndirect(Alloca));
2507 auto FnArgIter = FnArgs.begin() + FirstIRArg;
2508 ExpandTypeFromArgs(Ty, LV, FnArgIter);
2509 assert(FnArgIter == FnArgs.begin() + FirstIRArg + NumIRArgs);
2510 for (unsigned i = 0, e = NumIRArgs; i != e; ++i) {
2511 auto AI = FnArgs[FirstIRArg + i];
2512 AI->setName(Arg->getName() + "." + Twine(i));
2517 case ABIArgInfo::Ignore:
2518 assert(NumIRArgs == 0);
2519 // Initialize the local variable appropriately.
2520 if (!hasScalarEvaluationKind(Ty)) {
2521 ArgVals.push_back(ParamValue::forIndirect(CreateMemTemp(Ty)));
2523 llvm::Value *U = llvm::UndefValue::get(ConvertType(Arg->getType()));
2524 ArgVals.push_back(ParamValue::forDirect(U));
2530 if (getTarget().getCXXABI().areArgsDestroyedLeftToRightInCallee()) {
2531 for (int I = Args.size() - 1; I >= 0; --I)
2532 EmitParmDecl(*Args[I], ArgVals[I], I + 1);
2534 for (unsigned I = 0, E = Args.size(); I != E; ++I)
2535 EmitParmDecl(*Args[I], ArgVals[I], I + 1);
2539 static void eraseUnusedBitCasts(llvm::Instruction *insn) {
2540 while (insn->use_empty()) {
2541 llvm::BitCastInst *bitcast = dyn_cast<llvm::BitCastInst>(insn);
2542 if (!bitcast) return;
2544 // This is "safe" because we would have used a ConstantExpr otherwise.
2545 insn = cast<llvm::Instruction>(bitcast->getOperand(0));
2546 bitcast->eraseFromParent();
2550 /// Try to emit a fused autorelease of a return result.
2551 static llvm::Value *tryEmitFusedAutoreleaseOfResult(CodeGenFunction &CGF,
2552 llvm::Value *result) {
2553 // We must be immediately followed the cast.
2554 llvm::BasicBlock *BB = CGF.Builder.GetInsertBlock();
2555 if (BB->empty()) return nullptr;
2556 if (&BB->back() != result) return nullptr;
2558 llvm::Type *resultType = result->getType();
2560 // result is in a BasicBlock and is therefore an Instruction.
2561 llvm::Instruction *generator = cast<llvm::Instruction>(result);
2563 SmallVector<llvm::Instruction *, 4> InstsToKill;
2566 // %generator = bitcast %type1* %generator2 to %type2*
2567 while (llvm::BitCastInst *bitcast = dyn_cast<llvm::BitCastInst>(generator)) {
2568 // We would have emitted this as a constant if the operand weren't
2570 generator = cast<llvm::Instruction>(bitcast->getOperand(0));
2572 // Require the generator to be immediately followed by the cast.
2573 if (generator->getNextNode() != bitcast)
2576 InstsToKill.push_back(bitcast);
2580 // %generator = call i8* @objc_retain(i8* %originalResult)
2582 // %generator = call i8* @objc_retainAutoreleasedReturnValue(i8* %originalResult)
2583 llvm::CallInst *call = dyn_cast<llvm::CallInst>(generator);
2584 if (!call) return nullptr;
2586 bool doRetainAutorelease;
2588 if (call->getCalledValue() == CGF.CGM.getObjCEntrypoints().objc_retain) {
2589 doRetainAutorelease = true;
2590 } else if (call->getCalledValue() == CGF.CGM.getObjCEntrypoints()
2591 .objc_retainAutoreleasedReturnValue) {
2592 doRetainAutorelease = false;
2594 // If we emitted an assembly marker for this call (and the
2595 // ARCEntrypoints field should have been set if so), go looking
2596 // for that call. If we can't find it, we can't do this
2597 // optimization. But it should always be the immediately previous
2598 // instruction, unless we needed bitcasts around the call.
2599 if (CGF.CGM.getObjCEntrypoints().retainAutoreleasedReturnValueMarker) {
2600 llvm::Instruction *prev = call->getPrevNode();
2602 if (isa<llvm::BitCastInst>(prev)) {
2603 prev = prev->getPrevNode();
2606 assert(isa<llvm::CallInst>(prev));
2607 assert(cast<llvm::CallInst>(prev)->getCalledValue() ==
2608 CGF.CGM.getObjCEntrypoints().retainAutoreleasedReturnValueMarker);
2609 InstsToKill.push_back(prev);
2615 result = call->getArgOperand(0);
2616 InstsToKill.push_back(call);
2618 // Keep killing bitcasts, for sanity. Note that we no longer care
2619 // about precise ordering as long as there's exactly one use.
2620 while (llvm::BitCastInst *bitcast = dyn_cast<llvm::BitCastInst>(result)) {
2621 if (!bitcast->hasOneUse()) break;
2622 InstsToKill.push_back(bitcast);
2623 result = bitcast->getOperand(0);
2626 // Delete all the unnecessary instructions, from latest to earliest.
2627 for (auto *I : InstsToKill)
2628 I->eraseFromParent();
2630 // Do the fused retain/autorelease if we were asked to.
2631 if (doRetainAutorelease)
2632 result = CGF.EmitARCRetainAutoreleaseReturnValue(result);
2634 // Cast back to the result type.
2635 return CGF.Builder.CreateBitCast(result, resultType);
2638 /// If this is a +1 of the value of an immutable 'self', remove it.
2639 static llvm::Value *tryRemoveRetainOfSelf(CodeGenFunction &CGF,
2640 llvm::Value *result) {
2641 // This is only applicable to a method with an immutable 'self'.
2642 const ObjCMethodDecl *method =
2643 dyn_cast_or_null<ObjCMethodDecl>(CGF.CurCodeDecl);
2644 if (!method) return nullptr;
2645 const VarDecl *self = method->getSelfDecl();
2646 if (!self->getType().isConstQualified()) return nullptr;
2648 // Look for a retain call.
2649 llvm::CallInst *retainCall =
2650 dyn_cast<llvm::CallInst>(result->stripPointerCasts());
2652 retainCall->getCalledValue() != CGF.CGM.getObjCEntrypoints().objc_retain)
2655 // Look for an ordinary load of 'self'.
2656 llvm::Value *retainedValue = retainCall->getArgOperand(0);
2657 llvm::LoadInst *load =
2658 dyn_cast<llvm::LoadInst>(retainedValue->stripPointerCasts());
2659 if (!load || load->isAtomic() || load->isVolatile() ||
2660 load->getPointerOperand() != CGF.GetAddrOfLocalVar(self).getPointer())
2663 // Okay! Burn it all down. This relies for correctness on the
2664 // assumption that the retain is emitted as part of the return and
2665 // that thereafter everything is used "linearly".
2666 llvm::Type *resultType = result->getType();
2667 eraseUnusedBitCasts(cast<llvm::Instruction>(result));
2668 assert(retainCall->use_empty());
2669 retainCall->eraseFromParent();
2670 eraseUnusedBitCasts(cast<llvm::Instruction>(retainedValue));
2672 return CGF.Builder.CreateBitCast(load, resultType);
2675 /// Emit an ARC autorelease of the result of a function.
2677 /// \return the value to actually return from the function
2678 static llvm::Value *emitAutoreleaseOfResult(CodeGenFunction &CGF,
2679 llvm::Value *result) {
2680 // If we're returning 'self', kill the initial retain. This is a
2681 // heuristic attempt to "encourage correctness" in the really unfortunate
2682 // case where we have a return of self during a dealloc and we desperately
2683 // need to avoid the possible autorelease.
2684 if (llvm::Value *self = tryRemoveRetainOfSelf(CGF, result))
2687 // At -O0, try to emit a fused retain/autorelease.
2688 if (CGF.shouldUseFusedARCCalls())
2689 if (llvm::Value *fused = tryEmitFusedAutoreleaseOfResult(CGF, result))
2692 return CGF.EmitARCAutoreleaseReturnValue(result);
2695 /// Heuristically search for a dominating store to the return-value slot.
2696 static llvm::StoreInst *findDominatingStoreToReturnValue(CodeGenFunction &CGF) {
2697 // Check if a User is a store which pointerOperand is the ReturnValue.
2698 // We are looking for stores to the ReturnValue, not for stores of the
2699 // ReturnValue to some other location.
2700 auto GetStoreIfValid = [&CGF](llvm::User *U) -> llvm::StoreInst * {
2701 auto *SI = dyn_cast<llvm::StoreInst>(U);
2702 if (!SI || SI->getPointerOperand() != CGF.ReturnValue.getPointer())
2704 // These aren't actually possible for non-coerced returns, and we
2705 // only care about non-coerced returns on this code path.
2706 assert(!SI->isAtomic() && !SI->isVolatile());
2709 // If there are multiple uses of the return-value slot, just check
2710 // for something immediately preceding the IP. Sometimes this can
2711 // happen with how we generate implicit-returns; it can also happen
2712 // with noreturn cleanups.
2713 if (!CGF.ReturnValue.getPointer()->hasOneUse()) {
2714 llvm::BasicBlock *IP = CGF.Builder.GetInsertBlock();
2715 if (IP->empty()) return nullptr;
2716 llvm::Instruction *I = &IP->back();
2718 // Skip lifetime markers
2719 for (llvm::BasicBlock::reverse_iterator II = IP->rbegin(),
2722 if (llvm::IntrinsicInst *Intrinsic =
2723 dyn_cast<llvm::IntrinsicInst>(&*II)) {
2724 if (Intrinsic->getIntrinsicID() == llvm::Intrinsic::lifetime_end) {
2725 const llvm::Value *CastAddr = Intrinsic->getArgOperand(1);
2729 if (isa<llvm::BitCastInst>(&*II) && (CastAddr == &*II))
2737 return GetStoreIfValid(I);
2740 llvm::StoreInst *store =
2741 GetStoreIfValid(CGF.ReturnValue.getPointer()->user_back());
2742 if (!store) return nullptr;
2744 // Now do a first-and-dirty dominance check: just walk up the
2745 // single-predecessors chain from the current insertion point.
2746 llvm::BasicBlock *StoreBB = store->getParent();
2747 llvm::BasicBlock *IP = CGF.Builder.GetInsertBlock();
2748 while (IP != StoreBB) {
2749 if (!(IP = IP->getSinglePredecessor()))
2753 // Okay, the store's basic block dominates the insertion point; we
2754 // can do our thing.
2758 void CodeGenFunction::EmitFunctionEpilog(const CGFunctionInfo &FI,
2760 SourceLocation EndLoc) {
2761 if (FI.isNoReturn()) {
2762 // Noreturn functions don't return.
2763 EmitUnreachable(EndLoc);
2767 if (CurCodeDecl && CurCodeDecl->hasAttr<NakedAttr>()) {
2768 // Naked functions don't have epilogues.
2769 Builder.CreateUnreachable();
2773 // Functions with no result always return void.
2774 if (!ReturnValue.isValid()) {
2775 Builder.CreateRetVoid();
2779 llvm::DebugLoc RetDbgLoc;
2780 llvm::Value *RV = nullptr;
2781 QualType RetTy = FI.getReturnType();
2782 const ABIArgInfo &RetAI = FI.getReturnInfo();
2784 switch (RetAI.getKind()) {
2785 case ABIArgInfo::InAlloca:
2786 // Aggregrates get evaluated directly into the destination. Sometimes we
2787 // need to return the sret value in a register, though.
2788 assert(hasAggregateEvaluationKind(RetTy));
2789 if (RetAI.getInAllocaSRet()) {
2790 llvm::Function::arg_iterator EI = CurFn->arg_end();
2792 llvm::Value *ArgStruct = &*EI;
2793 llvm::Value *SRet = Builder.CreateStructGEP(
2794 nullptr, ArgStruct, RetAI.getInAllocaFieldIndex());
2795 RV = Builder.CreateAlignedLoad(SRet, getPointerAlign(), "sret");
2799 case ABIArgInfo::Indirect: {
2800 auto AI = CurFn->arg_begin();
2801 if (RetAI.isSRetAfterThis())
2803 switch (getEvaluationKind(RetTy)) {
2806 EmitLoadOfComplex(MakeAddrLValue(ReturnValue, RetTy), EndLoc);
2807 EmitStoreOfComplex(RT, MakeNaturalAlignAddrLValue(&*AI, RetTy),
2812 // Do nothing; aggregrates get evaluated directly into the destination.
2815 EmitStoreOfScalar(Builder.CreateLoad(ReturnValue),
2816 MakeNaturalAlignAddrLValue(&*AI, RetTy),
2823 case ABIArgInfo::Extend:
2824 case ABIArgInfo::Direct:
2825 if (RetAI.getCoerceToType() == ConvertType(RetTy) &&
2826 RetAI.getDirectOffset() == 0) {
2827 // The internal return value temp always will have pointer-to-return-type
2828 // type, just do a load.
2830 // If there is a dominating store to ReturnValue, we can elide
2831 // the load, zap the store, and usually zap the alloca.
2832 if (llvm::StoreInst *SI =
2833 findDominatingStoreToReturnValue(*this)) {
2834 // Reuse the debug location from the store unless there is
2835 // cleanup code to be emitted between the store and return
2837 if (EmitRetDbgLoc && !AutoreleaseResult)
2838 RetDbgLoc = SI->getDebugLoc();
2839 // Get the stored value and nuke the now-dead store.
2840 RV = SI->getValueOperand();
2841 SI->eraseFromParent();
2843 // If that was the only use of the return value, nuke it as well now.
2844 auto returnValueInst = ReturnValue.getPointer();
2845 if (returnValueInst->use_empty()) {
2846 if (auto alloca = dyn_cast<llvm::AllocaInst>(returnValueInst)) {
2847 alloca->eraseFromParent();
2848 ReturnValue = Address::invalid();
2852 // Otherwise, we have to do a simple load.
2854 RV = Builder.CreateLoad(ReturnValue);
2857 // If the value is offset in memory, apply the offset now.
2858 Address V = emitAddressAtOffset(*this, ReturnValue, RetAI);
2860 RV = CreateCoercedLoad(V, RetAI.getCoerceToType(), *this);
2863 // In ARC, end functions that return a retainable type with a call
2864 // to objc_autoreleaseReturnValue.
2865 if (AutoreleaseResult) {
2867 // Type::isObjCRetainabletype has to be called on a QualType that hasn't
2868 // been stripped of the typedefs, so we cannot use RetTy here. Get the
2869 // original return type of FunctionDecl, CurCodeDecl, and BlockDecl from
2870 // CurCodeDecl or BlockInfo.
2873 if (auto *FD = dyn_cast<FunctionDecl>(CurCodeDecl))
2874 RT = FD->getReturnType();
2875 else if (auto *MD = dyn_cast<ObjCMethodDecl>(CurCodeDecl))
2876 RT = MD->getReturnType();
2877 else if (isa<BlockDecl>(CurCodeDecl))
2878 RT = BlockInfo->BlockExpression->getFunctionType()->getReturnType();
2880 llvm_unreachable("Unexpected function/method type");
2882 assert(getLangOpts().ObjCAutoRefCount &&
2883 !FI.isReturnsRetained() &&
2884 RT->isObjCRetainableType());
2886 RV = emitAutoreleaseOfResult(*this, RV);
2891 case ABIArgInfo::Ignore:
2894 case ABIArgInfo::CoerceAndExpand: {
2895 auto coercionType = RetAI.getCoerceAndExpandType();
2896 auto layout = CGM.getDataLayout().getStructLayout(coercionType);
2898 // Load all of the coerced elements out into results.
2899 llvm::SmallVector<llvm::Value*, 4> results;
2900 Address addr = Builder.CreateElementBitCast(ReturnValue, coercionType);
2901 for (unsigned i = 0, e = coercionType->getNumElements(); i != e; ++i) {
2902 auto coercedEltType = coercionType->getElementType(i);
2903 if (ABIArgInfo::isPaddingForCoerceAndExpand(coercedEltType))
2906 auto eltAddr = Builder.CreateStructGEP(addr, i, layout);
2907 auto elt = Builder.CreateLoad(eltAddr);
2908 results.push_back(elt);
2911 // If we have one result, it's the single direct result type.
2912 if (results.size() == 1) {
2915 // Otherwise, we need to make a first-class aggregate.
2917 // Construct a return type that lacks padding elements.
2918 llvm::Type *returnType = RetAI.getUnpaddedCoerceAndExpandType();
2920 RV = llvm::UndefValue::get(returnType);
2921 for (unsigned i = 0, e = results.size(); i != e; ++i) {
2922 RV = Builder.CreateInsertValue(RV, results[i], i);
2928 case ABIArgInfo::Expand:
2929 llvm_unreachable("Invalid ABI kind for return argument");
2932 llvm::Instruction *Ret;
2934 EmitReturnValueCheck(RV);
2935 Ret = Builder.CreateRet(RV);
2937 Ret = Builder.CreateRetVoid();
2941 Ret->setDebugLoc(std::move(RetDbgLoc));
2944 void CodeGenFunction::EmitReturnValueCheck(llvm::Value *RV) {
2945 // A current decl may not be available when emitting vtable thunks.
2949 ReturnsNonNullAttr *RetNNAttr = nullptr;
2950 if (SanOpts.has(SanitizerKind::ReturnsNonnullAttribute))
2951 RetNNAttr = CurCodeDecl->getAttr<ReturnsNonNullAttr>();
2953 if (!RetNNAttr && !requiresReturnValueNullabilityCheck())
2956 // Prefer the returns_nonnull attribute if it's present.
2957 SourceLocation AttrLoc;
2958 SanitizerMask CheckKind;
2959 SanitizerHandler Handler;
2961 assert(!requiresReturnValueNullabilityCheck() &&
2962 "Cannot check nullability and the nonnull attribute");
2963 AttrLoc = RetNNAttr->getLocation();
2964 CheckKind = SanitizerKind::ReturnsNonnullAttribute;
2965 Handler = SanitizerHandler::NonnullReturn;
2967 if (auto *DD = dyn_cast<DeclaratorDecl>(CurCodeDecl))
2968 if (auto *TSI = DD->getTypeSourceInfo())
2969 if (auto FTL = TSI->getTypeLoc().castAs<FunctionTypeLoc>())
2970 AttrLoc = FTL.getReturnLoc().findNullabilityLoc();
2971 CheckKind = SanitizerKind::NullabilityReturn;
2972 Handler = SanitizerHandler::NullabilityReturn;
2975 SanitizerScope SanScope(this);
2977 // Make sure the "return" source location is valid. If we're checking a
2978 // nullability annotation, make sure the preconditions for the check are met.
2979 llvm::BasicBlock *Check = createBasicBlock("nullcheck");
2980 llvm::BasicBlock *NoCheck = createBasicBlock("no.nullcheck");
2981 llvm::Value *SLocPtr = Builder.CreateLoad(ReturnLocation, "return.sloc.load");
2982 llvm::Value *CanNullCheck = Builder.CreateIsNotNull(SLocPtr);
2983 if (requiresReturnValueNullabilityCheck())
2985 Builder.CreateAnd(CanNullCheck, RetValNullabilityPrecondition);
2986 Builder.CreateCondBr(CanNullCheck, Check, NoCheck);
2989 // Now do the null check.
2990 llvm::Value *Cond = Builder.CreateIsNotNull(RV);
2991 llvm::Constant *StaticData[] = {EmitCheckSourceLocation(AttrLoc)};
2992 llvm::Value *DynamicData[] = {SLocPtr};
2993 EmitCheck(std::make_pair(Cond, CheckKind), Handler, StaticData, DynamicData);
2998 // The return location should not be used after the check has been emitted.
2999 ReturnLocation = Address::invalid();
3003 static bool isInAllocaArgument(CGCXXABI &ABI, QualType type) {
3004 const CXXRecordDecl *RD = type->getAsCXXRecordDecl();
3005 return RD && ABI.getRecordArgABI(RD) == CGCXXABI::RAA_DirectInMemory;
3008 static AggValueSlot createPlaceholderSlot(CodeGenFunction &CGF,
3010 // FIXME: Generate IR in one pass, rather than going back and fixing up these
3012 llvm::Type *IRTy = CGF.ConvertTypeForMem(Ty);
3013 llvm::Type *IRPtrTy = IRTy->getPointerTo();
3014 llvm::Value *Placeholder = llvm::UndefValue::get(IRPtrTy->getPointerTo());
3016 // FIXME: When we generate this IR in one pass, we shouldn't need
3017 // this win32-specific alignment hack.
3018 CharUnits Align = CharUnits::fromQuantity(4);
3019 Placeholder = CGF.Builder.CreateAlignedLoad(IRPtrTy, Placeholder, Align);
3021 return AggValueSlot::forAddr(Address(Placeholder, Align),
3023 AggValueSlot::IsNotDestructed,
3024 AggValueSlot::DoesNotNeedGCBarriers,
3025 AggValueSlot::IsNotAliased);
3028 void CodeGenFunction::EmitDelegateCallArg(CallArgList &args,
3029 const VarDecl *param,
3030 SourceLocation loc) {
3031 // StartFunction converted the ABI-lowered parameter(s) into a
3032 // local alloca. We need to turn that into an r-value suitable
3034 Address local = GetAddrOfLocalVar(param);
3036 QualType type = param->getType();
3038 assert(!isInAllocaArgument(CGM.getCXXABI(), type) &&
3039 "cannot emit delegate call arguments for inalloca arguments!");
3041 // GetAddrOfLocalVar returns a pointer-to-pointer for references,
3042 // but the argument needs to be the original pointer.
3043 if (type->isReferenceType()) {
3044 args.add(RValue::get(Builder.CreateLoad(local)), type);
3046 // In ARC, move out of consumed arguments so that the release cleanup
3047 // entered by StartFunction doesn't cause an over-release. This isn't
3048 // optimal -O0 code generation, but it should get cleaned up when
3049 // optimization is enabled. This also assumes that delegate calls are
3050 // performed exactly once for a set of arguments, but that should be safe.
3051 } else if (getLangOpts().ObjCAutoRefCount &&
3052 param->hasAttr<NSConsumedAttr>() &&
3053 type->isObjCRetainableType()) {
3054 llvm::Value *ptr = Builder.CreateLoad(local);
3056 llvm::ConstantPointerNull::get(cast<llvm::PointerType>(ptr->getType()));
3057 Builder.CreateStore(null, local);
3058 args.add(RValue::get(ptr), type);
3060 // For the most part, we just need to load the alloca, except that
3061 // aggregate r-values are actually pointers to temporaries.
3063 args.add(convertTempToRValue(local, type, loc), type);
3067 static bool isProvablyNull(llvm::Value *addr) {
3068 return isa<llvm::ConstantPointerNull>(addr);
3071 /// Emit the actual writing-back of a writeback.
3072 static void emitWriteback(CodeGenFunction &CGF,
3073 const CallArgList::Writeback &writeback) {
3074 const LValue &srcLV = writeback.Source;
3075 Address srcAddr = srcLV.getAddress();
3076 assert(!isProvablyNull(srcAddr.getPointer()) &&
3077 "shouldn't have writeback for provably null argument");
3079 llvm::BasicBlock *contBB = nullptr;
3081 // If the argument wasn't provably non-null, we need to null check
3082 // before doing the store.
3083 bool provablyNonNull = llvm::isKnownNonZero(srcAddr.getPointer(),
3084 CGF.CGM.getDataLayout());
3085 if (!provablyNonNull) {
3086 llvm::BasicBlock *writebackBB = CGF.createBasicBlock("icr.writeback");
3087 contBB = CGF.createBasicBlock("icr.done");
3089 llvm::Value *isNull =
3090 CGF.Builder.CreateIsNull(srcAddr.getPointer(), "icr.isnull");
3091 CGF.Builder.CreateCondBr(isNull, contBB, writebackBB);
3092 CGF.EmitBlock(writebackBB);
3095 // Load the value to writeback.
3096 llvm::Value *value = CGF.Builder.CreateLoad(writeback.Temporary);
3098 // Cast it back, in case we're writing an id to a Foo* or something.
3099 value = CGF.Builder.CreateBitCast(value, srcAddr.getElementType(),
3100 "icr.writeback-cast");
3102 // Perform the writeback.
3104 // If we have a "to use" value, it's something we need to emit a use
3105 // of. This has to be carefully threaded in: if it's done after the
3106 // release it's potentially undefined behavior (and the optimizer
3107 // will ignore it), and if it happens before the retain then the
3108 // optimizer could move the release there.
3109 if (writeback.ToUse) {
3110 assert(srcLV.getObjCLifetime() == Qualifiers::OCL_Strong);
3112 // Retain the new value. No need to block-copy here: the block's
3113 // being passed up the stack.
3114 value = CGF.EmitARCRetainNonBlock(value);
3116 // Emit the intrinsic use here.
3117 CGF.EmitARCIntrinsicUse(writeback.ToUse);
3119 // Load the old value (primitively).
3120 llvm::Value *oldValue = CGF.EmitLoadOfScalar(srcLV, SourceLocation());
3122 // Put the new value in place (primitively).
3123 CGF.EmitStoreOfScalar(value, srcLV, /*init*/ false);
3125 // Release the old value.
3126 CGF.EmitARCRelease(oldValue, srcLV.isARCPreciseLifetime());
3128 // Otherwise, we can just do a normal lvalue store.
3130 CGF.EmitStoreThroughLValue(RValue::get(value), srcLV);
3133 // Jump to the continuation block.
3134 if (!provablyNonNull)
3135 CGF.EmitBlock(contBB);
3138 static void emitWritebacks(CodeGenFunction &CGF,
3139 const CallArgList &args) {
3140 for (const auto &I : args.writebacks())
3141 emitWriteback(CGF, I);
3144 static void deactivateArgCleanupsBeforeCall(CodeGenFunction &CGF,
3145 const CallArgList &CallArgs) {
3146 assert(CGF.getTarget().getCXXABI().areArgsDestroyedLeftToRightInCallee());
3147 ArrayRef<CallArgList::CallArgCleanup> Cleanups =
3148 CallArgs.getCleanupsToDeactivate();
3149 // Iterate in reverse to increase the likelihood of popping the cleanup.
3150 for (const auto &I : llvm::reverse(Cleanups)) {
3151 CGF.DeactivateCleanupBlock(I.Cleanup, I.IsActiveIP);
3152 I.IsActiveIP->eraseFromParent();
3156 static const Expr *maybeGetUnaryAddrOfOperand(const Expr *E) {
3157 if (const UnaryOperator *uop = dyn_cast<UnaryOperator>(E->IgnoreParens()))
3158 if (uop->getOpcode() == UO_AddrOf)
3159 return uop->getSubExpr();
3163 /// Emit an argument that's being passed call-by-writeback. That is,
3164 /// we are passing the address of an __autoreleased temporary; it
3165 /// might be copy-initialized with the current value of the given
3166 /// address, but it will definitely be copied out of after the call.
3167 static void emitWritebackArg(CodeGenFunction &CGF, CallArgList &args,
3168 const ObjCIndirectCopyRestoreExpr *CRE) {
3171 // Make an optimistic effort to emit the address as an l-value.
3172 // This can fail if the argument expression is more complicated.
3173 if (const Expr *lvExpr = maybeGetUnaryAddrOfOperand(CRE->getSubExpr())) {
3174 srcLV = CGF.EmitLValue(lvExpr);
3176 // Otherwise, just emit it as a scalar.
3178 Address srcAddr = CGF.EmitPointerWithAlignment(CRE->getSubExpr());
3180 QualType srcAddrType =
3181 CRE->getSubExpr()->getType()->castAs<PointerType>()->getPointeeType();
3182 srcLV = CGF.MakeAddrLValue(srcAddr, srcAddrType);
3184 Address srcAddr = srcLV.getAddress();
3186 // The dest and src types don't necessarily match in LLVM terms
3187 // because of the crazy ObjC compatibility rules.
3189 llvm::PointerType *destType =
3190 cast<llvm::PointerType>(CGF.ConvertType(CRE->getType()));
3192 // If the address is a constant null, just pass the appropriate null.
3193 if (isProvablyNull(srcAddr.getPointer())) {
3194 args.add(RValue::get(llvm::ConstantPointerNull::get(destType)),
3199 // Create the temporary.
3200 Address temp = CGF.CreateTempAlloca(destType->getElementType(),
3201 CGF.getPointerAlign(),
3203 // Loading an l-value can introduce a cleanup if the l-value is __weak,
3204 // and that cleanup will be conditional if we can't prove that the l-value
3205 // isn't null, so we need to register a dominating point so that the cleanups
3206 // system will make valid IR.
3207 CodeGenFunction::ConditionalEvaluation condEval(CGF);
3209 // Zero-initialize it if we're not doing a copy-initialization.
3210 bool shouldCopy = CRE->shouldCopy();
3213 llvm::ConstantPointerNull::get(
3214 cast<llvm::PointerType>(destType->getElementType()));
3215 CGF.Builder.CreateStore(null, temp);
3218 llvm::BasicBlock *contBB = nullptr;
3219 llvm::BasicBlock *originBB = nullptr;
3221 // If the address is *not* known to be non-null, we need to switch.
3222 llvm::Value *finalArgument;
3224 bool provablyNonNull = llvm::isKnownNonZero(srcAddr.getPointer(),
3225 CGF.CGM.getDataLayout());
3226 if (provablyNonNull) {
3227 finalArgument = temp.getPointer();
3229 llvm::Value *isNull =
3230 CGF.Builder.CreateIsNull(srcAddr.getPointer(), "icr.isnull");
3232 finalArgument = CGF.Builder.CreateSelect(isNull,
3233 llvm::ConstantPointerNull::get(destType),
3234 temp.getPointer(), "icr.argument");
3236 // If we need to copy, then the load has to be conditional, which
3237 // means we need control flow.
3239 originBB = CGF.Builder.GetInsertBlock();
3240 contBB = CGF.createBasicBlock("icr.cont");
3241 llvm::BasicBlock *copyBB = CGF.createBasicBlock("icr.copy");
3242 CGF.Builder.CreateCondBr(isNull, contBB, copyBB);
3243 CGF.EmitBlock(copyBB);
3244 condEval.begin(CGF);
3248 llvm::Value *valueToUse = nullptr;
3250 // Perform a copy if necessary.
3252 RValue srcRV = CGF.EmitLoadOfLValue(srcLV, SourceLocation());
3253 assert(srcRV.isScalar());
3255 llvm::Value *src = srcRV.getScalarVal();
3256 src = CGF.Builder.CreateBitCast(src, destType->getElementType(),
3259 // Use an ordinary store, not a store-to-lvalue.
3260 CGF.Builder.CreateStore(src, temp);
3262 // If optimization is enabled, and the value was held in a
3263 // __strong variable, we need to tell the optimizer that this
3264 // value has to stay alive until we're doing the store back.
3265 // This is because the temporary is effectively unretained,
3266 // and so otherwise we can violate the high-level semantics.
3267 if (CGF.CGM.getCodeGenOpts().OptimizationLevel != 0 &&
3268 srcLV.getObjCLifetime() == Qualifiers::OCL_Strong) {
3273 // Finish the control flow if we needed it.
3274 if (shouldCopy && !provablyNonNull) {
3275 llvm::BasicBlock *copyBB = CGF.Builder.GetInsertBlock();
3276 CGF.EmitBlock(contBB);
3278 // Make a phi for the value to intrinsically use.
3280 llvm::PHINode *phiToUse = CGF.Builder.CreatePHI(valueToUse->getType(), 2,
3282 phiToUse->addIncoming(valueToUse, copyBB);
3283 phiToUse->addIncoming(llvm::UndefValue::get(valueToUse->getType()),
3285 valueToUse = phiToUse;
3291 args.addWriteback(srcLV, temp, valueToUse);
3292 args.add(RValue::get(finalArgument), CRE->getType());
3295 void CallArgList::allocateArgumentMemory(CodeGenFunction &CGF) {
3299 llvm::Function *F = CGF.CGM.getIntrinsic(llvm::Intrinsic::stacksave);
3300 StackBase = CGF.Builder.CreateCall(F, {}, "inalloca.save");
3303 void CallArgList::freeArgumentMemory(CodeGenFunction &CGF) const {
3305 // Restore the stack after the call.
3306 llvm::Value *F = CGF.CGM.getIntrinsic(llvm::Intrinsic::stackrestore);
3307 CGF.Builder.CreateCall(F, StackBase);
3311 void CodeGenFunction::EmitNonNullArgCheck(RValue RV, QualType ArgType,
3312 SourceLocation ArgLoc,
3315 if (!AC.getDecl() || !(SanOpts.has(SanitizerKind::NonnullAttribute) ||
3316 SanOpts.has(SanitizerKind::NullabilityArg)))
3319 // The param decl may be missing in a variadic function.
3320 auto PVD = ParmNum < AC.getNumParams() ? AC.getParamDecl(ParmNum) : nullptr;
3321 unsigned ArgNo = PVD ? PVD->getFunctionScopeIndex() : ParmNum;
3323 // Prefer the nonnull attribute if it's present.
3324 const NonNullAttr *NNAttr = nullptr;
3325 if (SanOpts.has(SanitizerKind::NonnullAttribute))
3326 NNAttr = getNonNullAttr(AC.getDecl(), PVD, ArgType, ArgNo);
3328 bool CanCheckNullability = false;
3329 if (SanOpts.has(SanitizerKind::NullabilityArg) && !NNAttr && PVD) {
3330 auto Nullability = PVD->getType()->getNullability(getContext());
3331 CanCheckNullability = Nullability &&
3332 *Nullability == NullabilityKind::NonNull &&
3333 PVD->getTypeSourceInfo();
3336 if (!NNAttr && !CanCheckNullability)
3339 SourceLocation AttrLoc;
3340 SanitizerMask CheckKind;
3341 SanitizerHandler Handler;
3343 AttrLoc = NNAttr->getLocation();
3344 CheckKind = SanitizerKind::NonnullAttribute;
3345 Handler = SanitizerHandler::NonnullArg;
3347 AttrLoc = PVD->getTypeSourceInfo()->getTypeLoc().findNullabilityLoc();
3348 CheckKind = SanitizerKind::NullabilityArg;
3349 Handler = SanitizerHandler::NullabilityArg;
3352 SanitizerScope SanScope(this);
3353 assert(RV.isScalar());
3354 llvm::Value *V = RV.getScalarVal();
3356 Builder.CreateICmpNE(V, llvm::Constant::getNullValue(V->getType()));
3357 llvm::Constant *StaticData[] = {
3358 EmitCheckSourceLocation(ArgLoc), EmitCheckSourceLocation(AttrLoc),
3359 llvm::ConstantInt::get(Int32Ty, ArgNo + 1),
3361 EmitCheck(std::make_pair(Cond, CheckKind), Handler, StaticData, None);
3364 void CodeGenFunction::EmitCallArgs(
3365 CallArgList &Args, ArrayRef<QualType> ArgTypes,
3366 llvm::iterator_range<CallExpr::const_arg_iterator> ArgRange,
3367 AbstractCallee AC, unsigned ParamsToSkip, EvaluationOrder Order) {
3368 assert((int)ArgTypes.size() == (ArgRange.end() - ArgRange.begin()));
3370 // We *have* to evaluate arguments from right to left in the MS C++ ABI,
3371 // because arguments are destroyed left to right in the callee. As a special
3372 // case, there are certain language constructs that require left-to-right
3373 // evaluation, and in those cases we consider the evaluation order requirement
3374 // to trump the "destruction order is reverse construction order" guarantee.
3376 CGM.getTarget().getCXXABI().areArgsDestroyedLeftToRightInCallee()
3377 ? Order == EvaluationOrder::ForceLeftToRight
3378 : Order != EvaluationOrder::ForceRightToLeft;
3380 auto MaybeEmitImplicitObjectSize = [&](unsigned I, const Expr *Arg,
3381 RValue EmittedArg) {
3382 if (!AC.hasFunctionDecl() || I >= AC.getNumParams())
3384 auto *PS = AC.getParamDecl(I)->getAttr<PassObjectSizeAttr>();
3388 const auto &Context = getContext();
3389 auto SizeTy = Context.getSizeType();
3390 auto T = Builder.getIntNTy(Context.getTypeSize(SizeTy));
3391 assert(EmittedArg.getScalarVal() && "We emitted nothing for the arg?");
3392 llvm::Value *V = evaluateOrEmitBuiltinObjectSize(Arg, PS->getType(), T,
3393 EmittedArg.getScalarVal());
3394 Args.add(RValue::get(V), SizeTy);
3395 // If we're emitting args in reverse, be sure to do so with
3396 // pass_object_size, as well.
3398 std::swap(Args.back(), *(&Args.back() - 1));
3401 // Insert a stack save if we're going to need any inalloca args.
3402 bool HasInAllocaArgs = false;
3403 if (CGM.getTarget().getCXXABI().isMicrosoft()) {
3404 for (ArrayRef<QualType>::iterator I = ArgTypes.begin(), E = ArgTypes.end();
3405 I != E && !HasInAllocaArgs; ++I)
3406 HasInAllocaArgs = isInAllocaArgument(CGM.getCXXABI(), *I);
3407 if (HasInAllocaArgs) {
3408 assert(getTarget().getTriple().getArch() == llvm::Triple::x86);
3409 Args.allocateArgumentMemory(*this);
3413 // Evaluate each argument in the appropriate order.
3414 size_t CallArgsStart = Args.size();
3415 for (unsigned I = 0, E = ArgTypes.size(); I != E; ++I) {
3416 unsigned Idx = LeftToRight ? I : E - I - 1;
3417 CallExpr::const_arg_iterator Arg = ArgRange.begin() + Idx;
3418 unsigned InitialArgSize = Args.size();
3419 // If *Arg is an ObjCIndirectCopyRestoreExpr, check that either the types of
3420 // the argument and parameter match or the objc method is parameterized.
3421 assert((!isa<ObjCIndirectCopyRestoreExpr>(*Arg) ||
3422 getContext().hasSameUnqualifiedType((*Arg)->getType(),
3424 (isa<ObjCMethodDecl>(AC.getDecl()) &&
3425 isObjCMethodWithTypeParams(cast<ObjCMethodDecl>(AC.getDecl())))) &&
3426 "Argument and parameter types don't match");
3427 EmitCallArg(Args, *Arg, ArgTypes[Idx]);
3428 // In particular, we depend on it being the last arg in Args, and the
3429 // objectsize bits depend on there only being one arg if !LeftToRight.
3430 assert(InitialArgSize + 1 == Args.size() &&
3431 "The code below depends on only adding one arg per EmitCallArg");
3432 (void)InitialArgSize;
3433 RValue RVArg = Args.back().RV;
3434 EmitNonNullArgCheck(RVArg, ArgTypes[Idx], (*Arg)->getExprLoc(), AC,
3435 ParamsToSkip + Idx);
3436 // @llvm.objectsize should never have side-effects and shouldn't need
3437 // destruction/cleanups, so we can safely "emit" it after its arg,
3438 // regardless of right-to-leftness
3439 MaybeEmitImplicitObjectSize(Idx, *Arg, RVArg);
3443 // Un-reverse the arguments we just evaluated so they match up with the LLVM
3445 std::reverse(Args.begin() + CallArgsStart, Args.end());
3451 struct DestroyUnpassedArg final : EHScopeStack::Cleanup {
3452 DestroyUnpassedArg(Address Addr, QualType Ty)
3453 : Addr(Addr), Ty(Ty) {}
3458 void Emit(CodeGenFunction &CGF, Flags flags) override {
3459 const CXXDestructorDecl *Dtor = Ty->getAsCXXRecordDecl()->getDestructor();
3460 assert(!Dtor->isTrivial());
3461 CGF.EmitCXXDestructorCall(Dtor, Dtor_Complete, /*for vbase*/ false,
3462 /*Delegating=*/false, Addr);
3466 struct DisableDebugLocationUpdates {
3467 CodeGenFunction &CGF;
3468 bool disabledDebugInfo;
3469 DisableDebugLocationUpdates(CodeGenFunction &CGF, const Expr *E) : CGF(CGF) {
3470 if ((disabledDebugInfo = isa<CXXDefaultArgExpr>(E) && CGF.getDebugInfo()))
3471 CGF.disableDebugInfo();
3473 ~DisableDebugLocationUpdates() {
3474 if (disabledDebugInfo)
3475 CGF.enableDebugInfo();
3479 } // end anonymous namespace
3481 void CodeGenFunction::EmitCallArg(CallArgList &args, const Expr *E,
3483 DisableDebugLocationUpdates Dis(*this, E);
3484 if (const ObjCIndirectCopyRestoreExpr *CRE
3485 = dyn_cast<ObjCIndirectCopyRestoreExpr>(E)) {
3486 assert(getLangOpts().ObjCAutoRefCount);
3487 return emitWritebackArg(*this, args, CRE);
3490 assert(type->isReferenceType() == E->isGLValue() &&
3491 "reference binding to unmaterialized r-value!");
3493 if (E->isGLValue()) {
3494 assert(E->getObjectKind() == OK_Ordinary);
3495 return args.add(EmitReferenceBindingToExpr(E), type);
3498 bool HasAggregateEvalKind = hasAggregateEvaluationKind(type);
3500 // In the Microsoft C++ ABI, aggregate arguments are destructed by the callee.
3501 // However, we still have to push an EH-only cleanup in case we unwind before
3502 // we make it to the call.
3503 if (HasAggregateEvalKind &&
3504 CGM.getTarget().getCXXABI().areArgsDestroyedLeftToRightInCallee()) {
3505 // If we're using inalloca, use the argument memory. Otherwise, use a
3508 if (args.isUsingInAlloca())
3509 Slot = createPlaceholderSlot(*this, type);
3511 Slot = CreateAggTemp(type, "agg.tmp");
3513 const CXXRecordDecl *RD = type->getAsCXXRecordDecl();
3514 bool DestroyedInCallee =
3515 RD && RD->hasNonTrivialDestructor() &&
3516 CGM.getCXXABI().getRecordArgABI(RD) != CGCXXABI::RAA_Default;
3517 if (DestroyedInCallee)
3518 Slot.setExternallyDestructed();
3520 EmitAggExpr(E, Slot);
3521 RValue RV = Slot.asRValue();
3524 if (DestroyedInCallee) {
3525 // Create a no-op GEP between the placeholder and the cleanup so we can
3526 // RAUW it successfully. It also serves as a marker of the first
3527 // instruction where the cleanup is active.
3528 pushFullExprCleanup<DestroyUnpassedArg>(EHCleanup, Slot.getAddress(),
3530 // This unreachable is a temporary marker which will be removed later.
3531 llvm::Instruction *IsActive = Builder.CreateUnreachable();
3532 args.addArgCleanupDeactivation(EHStack.getInnermostEHScope(), IsActive);
3537 if (HasAggregateEvalKind && isa<ImplicitCastExpr>(E) &&
3538 cast<CastExpr>(E)->getCastKind() == CK_LValueToRValue) {
3539 LValue L = EmitLValue(cast<CastExpr>(E)->getSubExpr());
3540 assert(L.isSimple());
3541 if (L.getAlignment() >= getContext().getTypeAlignInChars(type)) {
3542 args.add(L.asAggregateRValue(), type, /*NeedsCopy*/true);
3544 // We can't represent a misaligned lvalue in the CallArgList, so copy
3545 // to an aligned temporary now.
3546 Address tmp = CreateMemTemp(type);
3547 EmitAggregateCopy(tmp, L.getAddress(), type, L.isVolatile());
3548 args.add(RValue::getAggregate(tmp), type);
3553 args.add(EmitAnyExprToTemp(E), type);
3556 QualType CodeGenFunction::getVarArgType(const Expr *Arg) {
3557 // System headers on Windows define NULL to 0 instead of 0LL on Win64. MSVC
3558 // implicitly widens null pointer constants that are arguments to varargs
3559 // functions to pointer-sized ints.
3560 if (!getTarget().getTriple().isOSWindows())
3561 return Arg->getType();
3563 if (Arg->getType()->isIntegerType() &&
3564 getContext().getTypeSize(Arg->getType()) <
3565 getContext().getTargetInfo().getPointerWidth(0) &&
3566 Arg->isNullPointerConstant(getContext(),
3567 Expr::NPC_ValueDependentIsNotNull)) {
3568 return getContext().getIntPtrType();
3571 return Arg->getType();
3574 // In ObjC ARC mode with no ObjC ARC exception safety, tell the ARC
3575 // optimizer it can aggressively ignore unwind edges.
3577 CodeGenFunction::AddObjCARCExceptionMetadata(llvm::Instruction *Inst) {
3578 if (CGM.getCodeGenOpts().OptimizationLevel != 0 &&
3579 !CGM.getCodeGenOpts().ObjCAutoRefCountExceptions)
3580 Inst->setMetadata("clang.arc.no_objc_arc_exceptions",
3581 CGM.getNoObjCARCExceptionsMetadata());
3584 /// Emits a call to the given no-arguments nounwind runtime function.
3586 CodeGenFunction::EmitNounwindRuntimeCall(llvm::Value *callee,
3587 const llvm::Twine &name) {
3588 return EmitNounwindRuntimeCall(callee, None, name);
3591 /// Emits a call to the given nounwind runtime function.
3593 CodeGenFunction::EmitNounwindRuntimeCall(llvm::Value *callee,
3594 ArrayRef<llvm::Value*> args,
3595 const llvm::Twine &name) {
3596 llvm::CallInst *call = EmitRuntimeCall(callee, args, name);
3597 call->setDoesNotThrow();
3601 /// Emits a simple call (never an invoke) to the given no-arguments
3602 /// runtime function.
3604 CodeGenFunction::EmitRuntimeCall(llvm::Value *callee,
3605 const llvm::Twine &name) {
3606 return EmitRuntimeCall(callee, None, name);
3609 // Calls which may throw must have operand bundles indicating which funclet
3610 // they are nested within.
3612 getBundlesForFunclet(llvm::Value *Callee, llvm::Instruction *CurrentFuncletPad,
3613 SmallVectorImpl<llvm::OperandBundleDef> &BundleList) {
3614 // There is no need for a funclet operand bundle if we aren't inside a
3616 if (!CurrentFuncletPad)
3619 // Skip intrinsics which cannot throw.
3620 auto *CalleeFn = dyn_cast<llvm::Function>(Callee->stripPointerCasts());
3621 if (CalleeFn && CalleeFn->isIntrinsic() && CalleeFn->doesNotThrow())
3624 BundleList.emplace_back("funclet", CurrentFuncletPad);
3627 /// Emits a simple call (never an invoke) to the given runtime function.
3629 CodeGenFunction::EmitRuntimeCall(llvm::Value *callee,
3630 ArrayRef<llvm::Value*> args,
3631 const llvm::Twine &name) {
3632 SmallVector<llvm::OperandBundleDef, 1> BundleList;
3633 getBundlesForFunclet(callee, CurrentFuncletPad, BundleList);
3635 llvm::CallInst *call = Builder.CreateCall(callee, args, BundleList, name);
3636 call->setCallingConv(getRuntimeCC());
3640 /// Emits a call or invoke to the given noreturn runtime function.
3641 void CodeGenFunction::EmitNoreturnRuntimeCallOrInvoke(llvm::Value *callee,
3642 ArrayRef<llvm::Value*> args) {
3643 SmallVector<llvm::OperandBundleDef, 1> BundleList;
3644 getBundlesForFunclet(callee, CurrentFuncletPad, BundleList);
3646 if (getInvokeDest()) {
3647 llvm::InvokeInst *invoke =
3648 Builder.CreateInvoke(callee,
3649 getUnreachableBlock(),
3653 invoke->setDoesNotReturn();
3654 invoke->setCallingConv(getRuntimeCC());
3656 llvm::CallInst *call = Builder.CreateCall(callee, args, BundleList);
3657 call->setDoesNotReturn();
3658 call->setCallingConv(getRuntimeCC());
3659 Builder.CreateUnreachable();
3663 /// Emits a call or invoke instruction to the given nullary runtime function.
3665 CodeGenFunction::EmitRuntimeCallOrInvoke(llvm::Value *callee,
3666 const Twine &name) {
3667 return EmitRuntimeCallOrInvoke(callee, None, name);
3670 /// Emits a call or invoke instruction to the given runtime function.
3672 CodeGenFunction::EmitRuntimeCallOrInvoke(llvm::Value *callee,
3673 ArrayRef<llvm::Value*> args,
3674 const Twine &name) {
3675 llvm::CallSite callSite = EmitCallOrInvoke(callee, args, name);
3676 callSite.setCallingConv(getRuntimeCC());
3680 /// Emits a call or invoke instruction to the given function, depending
3681 /// on the current state of the EH stack.
3683 CodeGenFunction::EmitCallOrInvoke(llvm::Value *Callee,
3684 ArrayRef<llvm::Value *> Args,
3685 const Twine &Name) {
3686 llvm::BasicBlock *InvokeDest = getInvokeDest();
3687 SmallVector<llvm::OperandBundleDef, 1> BundleList;
3688 getBundlesForFunclet(Callee, CurrentFuncletPad, BundleList);
3690 llvm::Instruction *Inst;
3692 Inst = Builder.CreateCall(Callee, Args, BundleList, Name);
3694 llvm::BasicBlock *ContBB = createBasicBlock("invoke.cont");
3695 Inst = Builder.CreateInvoke(Callee, ContBB, InvokeDest, Args, BundleList,
3700 // In ObjC ARC mode with no ObjC ARC exception safety, tell the ARC
3701 // optimizer it can aggressively ignore unwind edges.
3702 if (CGM.getLangOpts().ObjCAutoRefCount)
3703 AddObjCARCExceptionMetadata(Inst);
3705 return llvm::CallSite(Inst);
3708 /// \brief Store a non-aggregate value to an address to initialize it. For
3709 /// initialization, a non-atomic store will be used.
3710 static void EmitInitStoreOfNonAggregate(CodeGenFunction &CGF, RValue Src,
3713 CGF.EmitStoreOfScalar(Src.getScalarVal(), Dst, /*init=*/true);
3715 CGF.EmitStoreOfComplex(Src.getComplexVal(), Dst, /*init=*/true);
3718 void CodeGenFunction::deferPlaceholderReplacement(llvm::Instruction *Old,
3720 DeferredReplacements.push_back(std::make_pair(Old, New));
3723 RValue CodeGenFunction::EmitCall(const CGFunctionInfo &CallInfo,
3724 const CGCallee &Callee,
3725 ReturnValueSlot ReturnValue,
3726 const CallArgList &CallArgs,
3727 llvm::Instruction **callOrInvoke,
3728 SourceLocation Loc) {
3729 // FIXME: We no longer need the types from CallArgs; lift up and simplify.
3731 assert(Callee.isOrdinary());
3733 // Handle struct-return functions by passing a pointer to the
3734 // location that we would like to return into.
3735 QualType RetTy = CallInfo.getReturnType();
3736 const ABIArgInfo &RetAI = CallInfo.getReturnInfo();
3738 llvm::FunctionType *IRFuncTy = Callee.getFunctionType();
3740 // 1. Set up the arguments.
3742 // If we're using inalloca, insert the allocation after the stack save.
3743 // FIXME: Do this earlier rather than hacking it in here!
3744 Address ArgMemory = Address::invalid();
3745 const llvm::StructLayout *ArgMemoryLayout = nullptr;
3746 if (llvm::StructType *ArgStruct = CallInfo.getArgStruct()) {
3747 const llvm::DataLayout &DL = CGM.getDataLayout();
3748 ArgMemoryLayout = DL.getStructLayout(ArgStruct);
3749 llvm::Instruction *IP = CallArgs.getStackBase();
3750 llvm::AllocaInst *AI;
3752 IP = IP->getNextNode();
3753 AI = new llvm::AllocaInst(ArgStruct, DL.getAllocaAddrSpace(),
3756 AI = CreateTempAlloca(ArgStruct, "argmem");
3758 auto Align = CallInfo.getArgStructAlignment();
3759 AI->setAlignment(Align.getQuantity());
3760 AI->setUsedWithInAlloca(true);
3761 assert(AI->isUsedWithInAlloca() && !AI->isStaticAlloca());
3762 ArgMemory = Address(AI, Align);
3765 // Helper function to drill into the inalloca allocation.
3766 auto createInAllocaStructGEP = [&](unsigned FieldIndex) -> Address {
3768 CharUnits::fromQuantity(ArgMemoryLayout->getElementOffset(FieldIndex));
3769 return Builder.CreateStructGEP(ArgMemory, FieldIndex, FieldOffset);
3772 ClangToLLVMArgMapping IRFunctionArgs(CGM.getContext(), CallInfo);
3773 SmallVector<llvm::Value *, 16> IRCallArgs(IRFunctionArgs.totalIRArgs());
3775 // If the call returns a temporary with struct return, create a temporary
3776 // alloca to hold the result, unless one is given to us.
3777 Address SRetPtr = Address::invalid();
3778 size_t UnusedReturnSize = 0;
3779 if (RetAI.isIndirect() || RetAI.isInAlloca() || RetAI.isCoerceAndExpand()) {
3780 if (!ReturnValue.isNull()) {
3781 SRetPtr = ReturnValue.getValue();
3783 SRetPtr = CreateMemTemp(RetTy);
3784 if (HaveInsertPoint() && ReturnValue.isUnused()) {
3786 CGM.getDataLayout().getTypeAllocSize(ConvertTypeForMem(RetTy));
3787 if (EmitLifetimeStart(size, SRetPtr.getPointer()))
3788 UnusedReturnSize = size;
3791 if (IRFunctionArgs.hasSRetArg()) {
3792 IRCallArgs[IRFunctionArgs.getSRetArgNo()] = SRetPtr.getPointer();
3793 } else if (RetAI.isInAlloca()) {
3794 Address Addr = createInAllocaStructGEP(RetAI.getInAllocaFieldIndex());
3795 Builder.CreateStore(SRetPtr.getPointer(), Addr);
3799 Address swiftErrorTemp = Address::invalid();
3800 Address swiftErrorArg = Address::invalid();
3802 // Translate all of the arguments as necessary to match the IR lowering.
3803 assert(CallInfo.arg_size() == CallArgs.size() &&
3804 "Mismatch between function signature & arguments.");
3806 CGFunctionInfo::const_arg_iterator info_it = CallInfo.arg_begin();
3807 for (CallArgList::const_iterator I = CallArgs.begin(), E = CallArgs.end();
3808 I != E; ++I, ++info_it, ++ArgNo) {
3809 const ABIArgInfo &ArgInfo = info_it->info;
3812 // Insert a padding argument to ensure proper alignment.
3813 if (IRFunctionArgs.hasPaddingArg(ArgNo))
3814 IRCallArgs[IRFunctionArgs.getPaddingArgNo(ArgNo)] =
3815 llvm::UndefValue::get(ArgInfo.getPaddingType());
3817 unsigned FirstIRArg, NumIRArgs;
3818 std::tie(FirstIRArg, NumIRArgs) = IRFunctionArgs.getIRArgs(ArgNo);
3820 switch (ArgInfo.getKind()) {
3821 case ABIArgInfo::InAlloca: {
3822 assert(NumIRArgs == 0);
3823 assert(getTarget().getTriple().getArch() == llvm::Triple::x86);
3824 if (RV.isAggregate()) {
3825 // Replace the placeholder with the appropriate argument slot GEP.
3826 llvm::Instruction *Placeholder =
3827 cast<llvm::Instruction>(RV.getAggregatePointer());
3828 CGBuilderTy::InsertPoint IP = Builder.saveIP();
3829 Builder.SetInsertPoint(Placeholder);
3830 Address Addr = createInAllocaStructGEP(ArgInfo.getInAllocaFieldIndex());
3831 Builder.restoreIP(IP);
3832 deferPlaceholderReplacement(Placeholder, Addr.getPointer());
3834 // Store the RValue into the argument struct.
3835 Address Addr = createInAllocaStructGEP(ArgInfo.getInAllocaFieldIndex());
3836 unsigned AS = Addr.getType()->getPointerAddressSpace();
3837 llvm::Type *MemType = ConvertTypeForMem(I->Ty)->getPointerTo(AS);
3838 // There are some cases where a trivial bitcast is not avoidable. The
3839 // definition of a type later in a translation unit may change it's type
3840 // from {}* to (%struct.foo*)*.
3841 if (Addr.getType() != MemType)
3842 Addr = Builder.CreateBitCast(Addr, MemType);
3843 LValue argLV = MakeAddrLValue(Addr, I->Ty);
3844 EmitInitStoreOfNonAggregate(*this, RV, argLV);
3849 case ABIArgInfo::Indirect: {
3850 assert(NumIRArgs == 1);
3851 if (RV.isScalar() || RV.isComplex()) {
3852 // Make a temporary alloca to pass the argument.
3853 Address Addr = CreateMemTemp(I->Ty, ArgInfo.getIndirectAlign(),
3854 "indirect-arg-temp", false);
3855 IRCallArgs[FirstIRArg] = Addr.getPointer();
3857 LValue argLV = MakeAddrLValue(Addr, I->Ty);
3858 EmitInitStoreOfNonAggregate(*this, RV, argLV);
3860 // We want to avoid creating an unnecessary temporary+copy here;
3861 // however, we need one in three cases:
3862 // 1. If the argument is not byval, and we are required to copy the
3863 // source. (This case doesn't occur on any common architecture.)
3864 // 2. If the argument is byval, RV is not sufficiently aligned, and
3865 // we cannot force it to be sufficiently aligned.
3866 // 3. If the argument is byval, but RV is located in an address space
3867 // different than that of the argument (0).
3868 Address Addr = RV.getAggregateAddress();
3869 CharUnits Align = ArgInfo.getIndirectAlign();
3870 const llvm::DataLayout *TD = &CGM.getDataLayout();
3871 const unsigned RVAddrSpace = Addr.getType()->getAddressSpace();
3872 const unsigned ArgAddrSpace =
3873 (FirstIRArg < IRFuncTy->getNumParams()
3874 ? IRFuncTy->getParamType(FirstIRArg)->getPointerAddressSpace()
3876 if ((!ArgInfo.getIndirectByVal() && I->NeedsCopy) ||
3877 (ArgInfo.getIndirectByVal() && Addr.getAlignment() < Align &&
3878 llvm::getOrEnforceKnownAlignment(Addr.getPointer(),
3879 Align.getQuantity(), *TD)
3880 < Align.getQuantity()) ||
3881 (ArgInfo.getIndirectByVal() && (RVAddrSpace != ArgAddrSpace))) {
3882 // Create an aligned temporary, and copy to it.
3883 Address AI = CreateMemTemp(I->Ty, ArgInfo.getIndirectAlign(),
3884 "byval-temp", false);
3885 IRCallArgs[FirstIRArg] = AI.getPointer();
3886 EmitAggregateCopy(AI, Addr, I->Ty, RV.isVolatileQualified());
3888 // Skip the extra memcpy call.
3889 IRCallArgs[FirstIRArg] = Addr.getPointer();
3895 case ABIArgInfo::Ignore:
3896 assert(NumIRArgs == 0);
3899 case ABIArgInfo::Extend:
3900 case ABIArgInfo::Direct: {
3901 if (!isa<llvm::StructType>(ArgInfo.getCoerceToType()) &&
3902 ArgInfo.getCoerceToType() == ConvertType(info_it->type) &&
3903 ArgInfo.getDirectOffset() == 0) {
3904 assert(NumIRArgs == 1);
3907 V = RV.getScalarVal();
3909 V = Builder.CreateLoad(RV.getAggregateAddress());
3911 // Implement swifterror by copying into a new swifterror argument.
3912 // We'll write back in the normal path out of the call.
3913 if (CallInfo.getExtParameterInfo(ArgNo).getABI()
3914 == ParameterABI::SwiftErrorResult) {
3915 assert(!swiftErrorTemp.isValid() && "multiple swifterror args");
3917 QualType pointeeTy = I->Ty->getPointeeType();
3919 Address(V, getContext().getTypeAlignInChars(pointeeTy));
3922 CreateMemTemp(pointeeTy, getPointerAlign(), "swifterror.temp");
3923 V = swiftErrorTemp.getPointer();
3924 cast<llvm::AllocaInst>(V)->setSwiftError(true);
3926 llvm::Value *errorValue = Builder.CreateLoad(swiftErrorArg);
3927 Builder.CreateStore(errorValue, swiftErrorTemp);
3930 // We might have to widen integers, but we should never truncate.
3931 if (ArgInfo.getCoerceToType() != V->getType() &&
3932 V->getType()->isIntegerTy())
3933 V = Builder.CreateZExt(V, ArgInfo.getCoerceToType());
3935 // If the argument doesn't match, perform a bitcast to coerce it. This
3936 // can happen due to trivial type mismatches.
3937 if (FirstIRArg < IRFuncTy->getNumParams() &&
3938 V->getType() != IRFuncTy->getParamType(FirstIRArg))
3939 V = Builder.CreateBitCast(V, IRFuncTy->getParamType(FirstIRArg));
3941 IRCallArgs[FirstIRArg] = V;
3945 // FIXME: Avoid the conversion through memory if possible.
3946 Address Src = Address::invalid();
3947 if (RV.isScalar() || RV.isComplex()) {
3948 Src = CreateMemTemp(I->Ty, "coerce");
3949 LValue SrcLV = MakeAddrLValue(Src, I->Ty);
3950 EmitInitStoreOfNonAggregate(*this, RV, SrcLV);
3952 Src = RV.getAggregateAddress();
3955 // If the value is offset in memory, apply the offset now.
3956 Src = emitAddressAtOffset(*this, Src, ArgInfo);
3958 // Fast-isel and the optimizer generally like scalar values better than
3959 // FCAs, so we flatten them if this is safe to do for this argument.
3960 llvm::StructType *STy =
3961 dyn_cast<llvm::StructType>(ArgInfo.getCoerceToType());
3962 if (STy && ArgInfo.isDirect() && ArgInfo.getCanBeFlattened()) {
3963 llvm::Type *SrcTy = Src.getType()->getElementType();
3964 uint64_t SrcSize = CGM.getDataLayout().getTypeAllocSize(SrcTy);
3965 uint64_t DstSize = CGM.getDataLayout().getTypeAllocSize(STy);
3967 // If the source type is smaller than the destination type of the
3968 // coerce-to logic, copy the source value into a temp alloca the size
3969 // of the destination type to allow loading all of it. The bits past
3970 // the source value are left undef.
3971 if (SrcSize < DstSize) {
3973 = CreateTempAlloca(STy, Src.getAlignment(),
3974 Src.getName() + ".coerce");
3975 Builder.CreateMemCpy(TempAlloca, Src, SrcSize);
3978 Src = Builder.CreateBitCast(Src,
3979 STy->getPointerTo(Src.getAddressSpace()));
3982 auto SrcLayout = CGM.getDataLayout().getStructLayout(STy);
3983 assert(NumIRArgs == STy->getNumElements());
3984 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
3985 auto Offset = CharUnits::fromQuantity(SrcLayout->getElementOffset(i));
3986 Address EltPtr = Builder.CreateStructGEP(Src, i, Offset);
3987 llvm::Value *LI = Builder.CreateLoad(EltPtr);
3988 IRCallArgs[FirstIRArg + i] = LI;
3991 // In the simple case, just pass the coerced loaded value.
3992 assert(NumIRArgs == 1);
3993 IRCallArgs[FirstIRArg] =
3994 CreateCoercedLoad(Src, ArgInfo.getCoerceToType(), *this);
4000 case ABIArgInfo::CoerceAndExpand: {
4001 auto coercionType = ArgInfo.getCoerceAndExpandType();
4002 auto layout = CGM.getDataLayout().getStructLayout(coercionType);
4004 llvm::Value *tempSize = nullptr;
4005 Address addr = Address::invalid();
4006 if (RV.isAggregate()) {
4007 addr = RV.getAggregateAddress();
4009 assert(RV.isScalar()); // complex should always just be direct
4011 llvm::Type *scalarType = RV.getScalarVal()->getType();
4012 auto scalarSize = CGM.getDataLayout().getTypeAllocSize(scalarType);
4013 auto scalarAlign = CGM.getDataLayout().getPrefTypeAlignment(scalarType);
4015 tempSize = llvm::ConstantInt::get(CGM.Int64Ty, scalarSize);
4017 // Materialize to a temporary.
4018 addr = CreateTempAlloca(RV.getScalarVal()->getType(),
4019 CharUnits::fromQuantity(std::max(layout->getAlignment(),
4021 EmitLifetimeStart(scalarSize, addr.getPointer());
4023 Builder.CreateStore(RV.getScalarVal(), addr);
4026 addr = Builder.CreateElementBitCast(addr, coercionType);
4028 unsigned IRArgPos = FirstIRArg;
4029 for (unsigned i = 0, e = coercionType->getNumElements(); i != e; ++i) {
4030 llvm::Type *eltType = coercionType->getElementType(i);
4031 if (ABIArgInfo::isPaddingForCoerceAndExpand(eltType)) continue;
4032 Address eltAddr = Builder.CreateStructGEP(addr, i, layout);
4033 llvm::Value *elt = Builder.CreateLoad(eltAddr);
4034 IRCallArgs[IRArgPos++] = elt;
4036 assert(IRArgPos == FirstIRArg + NumIRArgs);
4039 EmitLifetimeEnd(tempSize, addr.getPointer());
4045 case ABIArgInfo::Expand:
4046 unsigned IRArgPos = FirstIRArg;
4047 ExpandTypeToArgs(I->Ty, RV, IRFuncTy, IRCallArgs, IRArgPos);
4048 assert(IRArgPos == FirstIRArg + NumIRArgs);
4053 llvm::Value *CalleePtr = Callee.getFunctionPointer();
4055 // If we're using inalloca, set up that argument.
4056 if (ArgMemory.isValid()) {
4057 llvm::Value *Arg = ArgMemory.getPointer();
4058 if (CallInfo.isVariadic()) {
4059 // When passing non-POD arguments by value to variadic functions, we will
4060 // end up with a variadic prototype and an inalloca call site. In such
4061 // cases, we can't do any parameter mismatch checks. Give up and bitcast
4063 unsigned CalleeAS = CalleePtr->getType()->getPointerAddressSpace();
4064 auto FnTy = getTypes().GetFunctionType(CallInfo)->getPointerTo(CalleeAS);
4065 CalleePtr = Builder.CreateBitCast(CalleePtr, FnTy);
4067 llvm::Type *LastParamTy =
4068 IRFuncTy->getParamType(IRFuncTy->getNumParams() - 1);
4069 if (Arg->getType() != LastParamTy) {
4071 // Assert that these structs have equivalent element types.
4072 llvm::StructType *FullTy = CallInfo.getArgStruct();
4073 llvm::StructType *DeclaredTy = cast<llvm::StructType>(
4074 cast<llvm::PointerType>(LastParamTy)->getElementType());
4075 assert(DeclaredTy->getNumElements() == FullTy->getNumElements());
4076 for (llvm::StructType::element_iterator DI = DeclaredTy->element_begin(),
4077 DE = DeclaredTy->element_end(),
4078 FI = FullTy->element_begin();
4079 DI != DE; ++DI, ++FI)
4082 Arg = Builder.CreateBitCast(Arg, LastParamTy);
4085 assert(IRFunctionArgs.hasInallocaArg());
4086 IRCallArgs[IRFunctionArgs.getInallocaArgNo()] = Arg;
4089 // 2. Prepare the function pointer.
4091 // If the callee is a bitcast of a non-variadic function to have a
4092 // variadic function pointer type, check to see if we can remove the
4093 // bitcast. This comes up with unprototyped functions.
4095 // This makes the IR nicer, but more importantly it ensures that we
4096 // can inline the function at -O0 if it is marked always_inline.
4097 auto simplifyVariadicCallee = [](llvm::Value *Ptr) -> llvm::Value* {
4098 llvm::FunctionType *CalleeFT =
4099 cast<llvm::FunctionType>(Ptr->getType()->getPointerElementType());
4100 if (!CalleeFT->isVarArg())
4103 llvm::ConstantExpr *CE = dyn_cast<llvm::ConstantExpr>(Ptr);
4104 if (!CE || CE->getOpcode() != llvm::Instruction::BitCast)
4107 llvm::Function *OrigFn = dyn_cast<llvm::Function>(CE->getOperand(0));
4111 llvm::FunctionType *OrigFT = OrigFn->getFunctionType();
4113 // If the original type is variadic, or if any of the component types
4114 // disagree, we cannot remove the cast.
4115 if (OrigFT->isVarArg() ||
4116 OrigFT->getNumParams() != CalleeFT->getNumParams() ||
4117 OrigFT->getReturnType() != CalleeFT->getReturnType())
4120 for (unsigned i = 0, e = OrigFT->getNumParams(); i != e; ++i)
4121 if (OrigFT->getParamType(i) != CalleeFT->getParamType(i))
4126 CalleePtr = simplifyVariadicCallee(CalleePtr);
4128 // 3. Perform the actual call.
4130 // Deactivate any cleanups that we're supposed to do immediately before
4132 if (!CallArgs.getCleanupsToDeactivate().empty())
4133 deactivateArgCleanupsBeforeCall(*this, CallArgs);
4135 // Assert that the arguments we computed match up. The IR verifier
4136 // will catch this, but this is a common enough source of problems
4137 // during IRGen changes that it's way better for debugging to catch
4138 // it ourselves here.
4140 assert(IRCallArgs.size() == IRFuncTy->getNumParams() || IRFuncTy->isVarArg());
4141 for (unsigned i = 0; i < IRCallArgs.size(); ++i) {
4142 // Inalloca argument can have different type.
4143 if (IRFunctionArgs.hasInallocaArg() &&
4144 i == IRFunctionArgs.getInallocaArgNo())
4146 if (i < IRFuncTy->getNumParams())
4147 assert(IRCallArgs[i]->getType() == IRFuncTy->getParamType(i));
4151 // Compute the calling convention and attributes.
4152 unsigned CallingConv;
4153 llvm::AttributeList Attrs;
4154 CGM.ConstructAttributeList(CalleePtr->getName(), CallInfo,
4155 Callee.getAbstractInfo(), Attrs, CallingConv,
4156 /*AttrOnCallSite=*/true);
4158 // Apply some call-site-specific attributes.
4159 // TODO: work this into building the attribute set.
4161 // Apply always_inline to all calls within flatten functions.
4162 // FIXME: should this really take priority over __try, below?
4163 if (CurCodeDecl && CurCodeDecl->hasAttr<FlattenAttr>() &&
4164 !(Callee.getAbstractInfo().getCalleeDecl() &&
4165 Callee.getAbstractInfo().getCalleeDecl()->hasAttr<NoInlineAttr>())) {
4167 Attrs.addAttribute(getLLVMContext(), llvm::AttributeList::FunctionIndex,
4168 llvm::Attribute::AlwaysInline);
4171 // Disable inlining inside SEH __try blocks.
4172 if (isSEHTryScope()) {
4174 Attrs.addAttribute(getLLVMContext(), llvm::AttributeList::FunctionIndex,
4175 llvm::Attribute::NoInline);
4178 // Decide whether to use a call or an invoke.
4180 if (currentFunctionUsesSEHTry()) {
4181 // SEH cares about asynchronous exceptions, so everything can "throw."
4182 CannotThrow = false;
4183 } else if (isCleanupPadScope() &&
4184 EHPersonality::get(*this).isMSVCXXPersonality()) {
4185 // The MSVC++ personality will implicitly terminate the program if an
4186 // exception is thrown during a cleanup outside of a try/catch.
4187 // We don't need to model anything in IR to get this behavior.
4190 // Otherwise, nounwind call sites will never throw.
4191 CannotThrow = Attrs.hasAttribute(llvm::AttributeList::FunctionIndex,
4192 llvm::Attribute::NoUnwind);
4194 llvm::BasicBlock *InvokeDest = CannotThrow ? nullptr : getInvokeDest();
4196 SmallVector<llvm::OperandBundleDef, 1> BundleList;
4197 getBundlesForFunclet(CalleePtr, CurrentFuncletPad, BundleList);
4199 // Emit the actual call/invoke instruction.
4202 CS = Builder.CreateCall(CalleePtr, IRCallArgs, BundleList);
4204 llvm::BasicBlock *Cont = createBasicBlock("invoke.cont");
4205 CS = Builder.CreateInvoke(CalleePtr, Cont, InvokeDest, IRCallArgs,
4209 llvm::Instruction *CI = CS.getInstruction();
4213 // Apply the attributes and calling convention.
4214 CS.setAttributes(Attrs);
4215 CS.setCallingConv(static_cast<llvm::CallingConv::ID>(CallingConv));
4217 // Apply various metadata.
4219 if (!CI->getType()->isVoidTy())
4220 CI->setName("call");
4222 // Insert instrumentation or attach profile metadata at indirect call sites.
4223 // For more details, see the comment before the definition of
4224 // IPVK_IndirectCallTarget in InstrProfData.inc.
4225 if (!CS.getCalledFunction())
4226 PGO.valueProfile(Builder, llvm::IPVK_IndirectCallTarget,
4229 // In ObjC ARC mode with no ObjC ARC exception safety, tell the ARC
4230 // optimizer it can aggressively ignore unwind edges.
4231 if (CGM.getLangOpts().ObjCAutoRefCount)
4232 AddObjCARCExceptionMetadata(CI);
4234 // Suppress tail calls if requested.
4235 if (llvm::CallInst *Call = dyn_cast<llvm::CallInst>(CI)) {
4236 const Decl *TargetDecl = Callee.getAbstractInfo().getCalleeDecl();
4237 if (TargetDecl && TargetDecl->hasAttr<NotTailCalledAttr>())
4238 Call->setTailCallKind(llvm::CallInst::TCK_NoTail);
4241 // 4. Finish the call.
4243 // If the call doesn't return, finish the basic block and clear the
4244 // insertion point; this allows the rest of IRGen to discard
4245 // unreachable code.
4246 if (CS.doesNotReturn()) {
4247 if (UnusedReturnSize)
4248 EmitLifetimeEnd(llvm::ConstantInt::get(Int64Ty, UnusedReturnSize),
4249 SRetPtr.getPointer());
4251 // Strip away the noreturn attribute to better diagnose unreachable UB.
4252 if (SanOpts.has(SanitizerKind::Unreachable)) {
4253 if (auto *F = CS.getCalledFunction())
4254 F->removeFnAttr(llvm::Attribute::NoReturn);
4255 CS.removeAttribute(llvm::AttributeList::FunctionIndex,
4256 llvm::Attribute::NoReturn);
4259 EmitUnreachable(Loc);
4260 Builder.ClearInsertionPoint();
4262 // FIXME: For now, emit a dummy basic block because expr emitters in
4263 // generally are not ready to handle emitting expressions at unreachable
4265 EnsureInsertPoint();
4267 // Return a reasonable RValue.
4268 return GetUndefRValue(RetTy);
4271 // Perform the swifterror writeback.
4272 if (swiftErrorTemp.isValid()) {
4273 llvm::Value *errorResult = Builder.CreateLoad(swiftErrorTemp);
4274 Builder.CreateStore(errorResult, swiftErrorArg);
4277 // Emit any call-associated writebacks immediately. Arguably this
4278 // should happen after any return-value munging.
4279 if (CallArgs.hasWritebacks())
4280 emitWritebacks(*this, CallArgs);
4282 // The stack cleanup for inalloca arguments has to run out of the normal
4283 // lexical order, so deactivate it and run it manually here.
4284 CallArgs.freeArgumentMemory(*this);
4286 // Extract the return value.
4288 switch (RetAI.getKind()) {
4289 case ABIArgInfo::CoerceAndExpand: {
4290 auto coercionType = RetAI.getCoerceAndExpandType();
4291 auto layout = CGM.getDataLayout().getStructLayout(coercionType);
4293 Address addr = SRetPtr;
4294 addr = Builder.CreateElementBitCast(addr, coercionType);
4296 assert(CI->getType() == RetAI.getUnpaddedCoerceAndExpandType());
4297 bool requiresExtract = isa<llvm::StructType>(CI->getType());
4299 unsigned unpaddedIndex = 0;
4300 for (unsigned i = 0, e = coercionType->getNumElements(); i != e; ++i) {
4301 llvm::Type *eltType = coercionType->getElementType(i);
4302 if (ABIArgInfo::isPaddingForCoerceAndExpand(eltType)) continue;
4303 Address eltAddr = Builder.CreateStructGEP(addr, i, layout);
4304 llvm::Value *elt = CI;
4305 if (requiresExtract)
4306 elt = Builder.CreateExtractValue(elt, unpaddedIndex++);
4308 assert(unpaddedIndex == 0);
4309 Builder.CreateStore(elt, eltAddr);
4315 case ABIArgInfo::InAlloca:
4316 case ABIArgInfo::Indirect: {
4317 RValue ret = convertTempToRValue(SRetPtr, RetTy, SourceLocation());
4318 if (UnusedReturnSize)
4319 EmitLifetimeEnd(llvm::ConstantInt::get(Int64Ty, UnusedReturnSize),
4320 SRetPtr.getPointer());
4324 case ABIArgInfo::Ignore:
4325 // If we are ignoring an argument that had a result, make sure to
4326 // construct the appropriate return value for our caller.
4327 return GetUndefRValue(RetTy);
4329 case ABIArgInfo::Extend:
4330 case ABIArgInfo::Direct: {
4331 llvm::Type *RetIRTy = ConvertType(RetTy);
4332 if (RetAI.getCoerceToType() == RetIRTy && RetAI.getDirectOffset() == 0) {
4333 switch (getEvaluationKind(RetTy)) {
4335 llvm::Value *Real = Builder.CreateExtractValue(CI, 0);
4336 llvm::Value *Imag = Builder.CreateExtractValue(CI, 1);
4337 return RValue::getComplex(std::make_pair(Real, Imag));
4339 case TEK_Aggregate: {
4340 Address DestPtr = ReturnValue.getValue();
4341 bool DestIsVolatile = ReturnValue.isVolatile();
4343 if (!DestPtr.isValid()) {
4344 DestPtr = CreateMemTemp(RetTy, "agg.tmp");
4345 DestIsVolatile = false;
4347 BuildAggStore(*this, CI, DestPtr, DestIsVolatile);
4348 return RValue::getAggregate(DestPtr);
4351 // If the argument doesn't match, perform a bitcast to coerce it. This
4352 // can happen due to trivial type mismatches.
4353 llvm::Value *V = CI;
4354 if (V->getType() != RetIRTy)
4355 V = Builder.CreateBitCast(V, RetIRTy);
4356 return RValue::get(V);
4359 llvm_unreachable("bad evaluation kind");
4362 Address DestPtr = ReturnValue.getValue();
4363 bool DestIsVolatile = ReturnValue.isVolatile();
4365 if (!DestPtr.isValid()) {
4366 DestPtr = CreateMemTemp(RetTy, "coerce");
4367 DestIsVolatile = false;
4370 // If the value is offset in memory, apply the offset now.
4371 Address StorePtr = emitAddressAtOffset(*this, DestPtr, RetAI);
4372 CreateCoercedStore(CI, StorePtr, DestIsVolatile, *this);
4374 return convertTempToRValue(DestPtr, RetTy, SourceLocation());
4377 case ABIArgInfo::Expand:
4378 llvm_unreachable("Invalid ABI kind for return argument");
4381 llvm_unreachable("Unhandled ABIArgInfo::Kind");
4384 // Emit the assume_aligned check on the return value.
4385 const Decl *TargetDecl = Callee.getAbstractInfo().getCalleeDecl();
4386 if (Ret.isScalar() && TargetDecl) {
4387 if (const auto *AA = TargetDecl->getAttr<AssumeAlignedAttr>()) {
4388 llvm::Value *OffsetValue = nullptr;
4389 if (const auto *Offset = AA->getOffset())
4390 OffsetValue = EmitScalarExpr(Offset);
4392 llvm::Value *Alignment = EmitScalarExpr(AA->getAlignment());
4393 llvm::ConstantInt *AlignmentCI = cast<llvm::ConstantInt>(Alignment);
4394 EmitAlignmentAssumption(Ret.getScalarVal(), AlignmentCI->getZExtValue(),
4396 } else if (const auto *AA = TargetDecl->getAttr<AllocAlignAttr>()) {
4397 llvm::Value *ParamVal =
4398 CallArgs[AA->getParamIndex() - 1].RV.getScalarVal();
4399 EmitAlignmentAssumption(Ret.getScalarVal(), ParamVal);
4406 /* VarArg handling */
4408 Address CodeGenFunction::EmitVAArg(VAArgExpr *VE, Address &VAListAddr) {
4409 VAListAddr = VE->isMicrosoftABI()
4410 ? EmitMSVAListRef(VE->getSubExpr())
4411 : EmitVAListRef(VE->getSubExpr());
4412 QualType Ty = VE->getType();
4413 if (VE->isMicrosoftABI())
4414 return CGM.getTypes().getABIInfo().EmitMSVAArg(*this, VAListAddr, Ty);
4415 return CGM.getTypes().getABIInfo().EmitVAArg(*this, VAListAddr, Ty);