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/Transforms/Utils/Local.h"
33 #include "llvm/Analysis/ValueTracking.h"
34 #include "llvm/IR/Attributes.h"
35 #include "llvm/IR/CallSite.h"
36 #include "llvm/IR/CallingConv.h"
37 #include "llvm/IR/DataLayout.h"
38 #include "llvm/IR/InlineAsm.h"
39 #include "llvm/IR/IntrinsicInst.h"
40 #include "llvm/IR/Intrinsics.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 /// Set calling convention for CUDA/HIP kernel.
259 static void setCUDAKernelCallingConvention(CanQualType &FTy, CodeGenModule &CGM,
260 const FunctionDecl *FD) {
261 if (FD->hasAttr<CUDAGlobalAttr>()) {
262 const FunctionType *FT = FTy->getAs<FunctionType>();
263 CGM.getTargetCodeGenInfo().setCUDAKernelCallingConvention(FT);
264 FTy = FT->getCanonicalTypeUnqualified();
268 /// Arrange the argument and result information for a declaration or
269 /// definition of the given C++ non-static member function. The
270 /// member function must be an ordinary function, i.e. not a
271 /// constructor or destructor.
272 const CGFunctionInfo &
273 CodeGenTypes::arrangeCXXMethodDeclaration(const CXXMethodDecl *MD) {
274 assert(!isa<CXXConstructorDecl>(MD) && "wrong method for constructors!");
275 assert(!isa<CXXDestructorDecl>(MD) && "wrong method for destructors!");
277 CanQualType FT = GetFormalType(MD).getAs<Type>();
278 setCUDAKernelCallingConvention(FT, CGM, MD);
279 auto prototype = FT.getAs<FunctionProtoType>();
281 if (MD->isInstance()) {
282 // The abstract case is perfectly fine.
283 const CXXRecordDecl *ThisType = TheCXXABI.getThisArgumentTypeForMethod(MD);
284 return arrangeCXXMethodType(ThisType, prototype.getTypePtr(), MD);
287 return arrangeFreeFunctionType(prototype, MD);
290 bool CodeGenTypes::inheritingCtorHasParams(
291 const InheritedConstructor &Inherited, CXXCtorType Type) {
292 // Parameters are unnecessary if we're constructing a base class subobject
293 // and the inherited constructor lives in a virtual base.
294 return Type == Ctor_Complete ||
295 !Inherited.getShadowDecl()->constructsVirtualBase() ||
296 !Target.getCXXABI().hasConstructorVariants();
299 const CGFunctionInfo &
300 CodeGenTypes::arrangeCXXStructorDeclaration(const CXXMethodDecl *MD,
303 SmallVector<CanQualType, 16> argTypes;
304 SmallVector<FunctionProtoType::ExtParameterInfo, 16> paramInfos;
305 argTypes.push_back(GetThisType(Context, MD->getParent()));
307 bool PassParams = true;
310 if (auto *CD = dyn_cast<CXXConstructorDecl>(MD)) {
311 GD = GlobalDecl(CD, toCXXCtorType(Type));
313 // A base class inheriting constructor doesn't get forwarded arguments
314 // needed to construct a virtual base (or base class thereof).
315 if (auto Inherited = CD->getInheritedConstructor())
316 PassParams = inheritingCtorHasParams(Inherited, toCXXCtorType(Type));
318 auto *DD = dyn_cast<CXXDestructorDecl>(MD);
319 GD = GlobalDecl(DD, toCXXDtorType(Type));
322 CanQual<FunctionProtoType> FTP = GetFormalType(MD);
324 // Add the formal parameters.
326 appendParameterTypes(*this, argTypes, paramInfos, FTP);
328 CGCXXABI::AddedStructorArgs AddedArgs =
329 TheCXXABI.buildStructorSignature(MD, Type, argTypes);
330 if (!paramInfos.empty()) {
331 // Note: prefix implies after the first param.
332 if (AddedArgs.Prefix)
333 paramInfos.insert(paramInfos.begin() + 1, AddedArgs.Prefix,
334 FunctionProtoType::ExtParameterInfo{});
335 if (AddedArgs.Suffix)
336 paramInfos.append(AddedArgs.Suffix,
337 FunctionProtoType::ExtParameterInfo{});
340 RequiredArgs required =
341 (PassParams && MD->isVariadic() ? RequiredArgs(argTypes.size())
342 : RequiredArgs::All);
344 FunctionType::ExtInfo extInfo = FTP->getExtInfo();
345 CanQualType resultType = TheCXXABI.HasThisReturn(GD)
347 : TheCXXABI.hasMostDerivedReturn(GD)
348 ? CGM.getContext().VoidPtrTy
350 return arrangeLLVMFunctionInfo(resultType, /*instanceMethod=*/true,
351 /*chainCall=*/false, argTypes, extInfo,
352 paramInfos, required);
355 static SmallVector<CanQualType, 16>
356 getArgTypesForCall(ASTContext &ctx, const CallArgList &args) {
357 SmallVector<CanQualType, 16> argTypes;
358 for (auto &arg : args)
359 argTypes.push_back(ctx.getCanonicalParamType(arg.Ty));
363 static SmallVector<CanQualType, 16>
364 getArgTypesForDeclaration(ASTContext &ctx, const FunctionArgList &args) {
365 SmallVector<CanQualType, 16> argTypes;
366 for (auto &arg : args)
367 argTypes.push_back(ctx.getCanonicalParamType(arg->getType()));
371 static llvm::SmallVector<FunctionProtoType::ExtParameterInfo, 16>
372 getExtParameterInfosForCall(const FunctionProtoType *proto,
373 unsigned prefixArgs, unsigned totalArgs) {
374 llvm::SmallVector<FunctionProtoType::ExtParameterInfo, 16> result;
375 if (proto->hasExtParameterInfos()) {
376 addExtParameterInfosForCall(result, proto, prefixArgs, totalArgs);
381 /// Arrange a call to a C++ method, passing the given arguments.
383 /// ExtraPrefixArgs is the number of ABI-specific args passed after the `this`
385 /// ExtraSuffixArgs is the number of ABI-specific args passed at the end of
387 /// PassProtoArgs indicates whether `args` has args for the parameters in the
388 /// given CXXConstructorDecl.
389 const CGFunctionInfo &
390 CodeGenTypes::arrangeCXXConstructorCall(const CallArgList &args,
391 const CXXConstructorDecl *D,
392 CXXCtorType CtorKind,
393 unsigned ExtraPrefixArgs,
394 unsigned ExtraSuffixArgs,
395 bool PassProtoArgs) {
397 SmallVector<CanQualType, 16> ArgTypes;
398 for (const auto &Arg : args)
399 ArgTypes.push_back(Context.getCanonicalParamType(Arg.Ty));
401 // +1 for implicit this, which should always be args[0].
402 unsigned TotalPrefixArgs = 1 + ExtraPrefixArgs;
404 CanQual<FunctionProtoType> FPT = GetFormalType(D);
405 RequiredArgs Required =
406 RequiredArgs::forPrototypePlus(FPT, TotalPrefixArgs + ExtraSuffixArgs, D);
407 GlobalDecl GD(D, CtorKind);
408 CanQualType ResultType = TheCXXABI.HasThisReturn(GD)
410 : TheCXXABI.hasMostDerivedReturn(GD)
411 ? CGM.getContext().VoidPtrTy
414 FunctionType::ExtInfo Info = FPT->getExtInfo();
415 llvm::SmallVector<FunctionProtoType::ExtParameterInfo, 16> ParamInfos;
416 // If the prototype args are elided, we should only have ABI-specific args,
417 // which never have param info.
418 if (PassProtoArgs && FPT->hasExtParameterInfos()) {
419 // ABI-specific suffix arguments are treated the same as variadic arguments.
420 addExtParameterInfosForCall(ParamInfos, FPT.getTypePtr(), TotalPrefixArgs,
423 return arrangeLLVMFunctionInfo(ResultType, /*instanceMethod=*/true,
424 /*chainCall=*/false, ArgTypes, Info,
425 ParamInfos, Required);
428 /// Arrange the argument and result information for the declaration or
429 /// definition of the given function.
430 const CGFunctionInfo &
431 CodeGenTypes::arrangeFunctionDeclaration(const FunctionDecl *FD) {
432 if (const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(FD))
433 if (MD->isInstance())
434 return arrangeCXXMethodDeclaration(MD);
436 CanQualType FTy = FD->getType()->getCanonicalTypeUnqualified();
438 assert(isa<FunctionType>(FTy));
439 setCUDAKernelCallingConvention(FTy, CGM, FD);
441 // When declaring a function without a prototype, always use a
442 // non-variadic type.
443 if (CanQual<FunctionNoProtoType> noProto = FTy.getAs<FunctionNoProtoType>()) {
444 return arrangeLLVMFunctionInfo(
445 noProto->getReturnType(), /*instanceMethod=*/false,
446 /*chainCall=*/false, None, noProto->getExtInfo(), {},RequiredArgs::All);
449 return arrangeFreeFunctionType(FTy.castAs<FunctionProtoType>(), FD);
452 /// Arrange the argument and result information for the declaration or
453 /// definition of an Objective-C method.
454 const CGFunctionInfo &
455 CodeGenTypes::arrangeObjCMethodDeclaration(const ObjCMethodDecl *MD) {
456 // It happens that this is the same as a call with no optional
457 // arguments, except also using the formal 'self' type.
458 return arrangeObjCMessageSendSignature(MD, MD->getSelfDecl()->getType());
461 /// Arrange the argument and result information for the function type
462 /// through which to perform a send to the given Objective-C method,
463 /// using the given receiver type. The receiver type is not always
464 /// the 'self' type of the method or even an Objective-C pointer type.
465 /// This is *not* the right method for actually performing such a
466 /// message send, due to the possibility of optional arguments.
467 const CGFunctionInfo &
468 CodeGenTypes::arrangeObjCMessageSendSignature(const ObjCMethodDecl *MD,
469 QualType receiverType) {
470 SmallVector<CanQualType, 16> argTys;
471 SmallVector<FunctionProtoType::ExtParameterInfo, 4> extParamInfos(2);
472 argTys.push_back(Context.getCanonicalParamType(receiverType));
473 argTys.push_back(Context.getCanonicalParamType(Context.getObjCSelType()));
475 for (const auto *I : MD->parameters()) {
476 argTys.push_back(Context.getCanonicalParamType(I->getType()));
477 auto extParamInfo = FunctionProtoType::ExtParameterInfo().withIsNoEscape(
478 I->hasAttr<NoEscapeAttr>());
479 extParamInfos.push_back(extParamInfo);
482 FunctionType::ExtInfo einfo;
483 bool IsWindows = getContext().getTargetInfo().getTriple().isOSWindows();
484 einfo = einfo.withCallingConv(getCallingConventionForDecl(MD, IsWindows));
486 if (getContext().getLangOpts().ObjCAutoRefCount &&
487 MD->hasAttr<NSReturnsRetainedAttr>())
488 einfo = einfo.withProducesResult(true);
490 RequiredArgs required =
491 (MD->isVariadic() ? RequiredArgs(argTys.size()) : RequiredArgs::All);
493 return arrangeLLVMFunctionInfo(
494 GetReturnType(MD->getReturnType()), /*instanceMethod=*/false,
495 /*chainCall=*/false, argTys, einfo, extParamInfos, required);
498 const CGFunctionInfo &
499 CodeGenTypes::arrangeUnprototypedObjCMessageSend(QualType returnType,
500 const CallArgList &args) {
501 auto argTypes = getArgTypesForCall(Context, args);
502 FunctionType::ExtInfo einfo;
504 return arrangeLLVMFunctionInfo(
505 GetReturnType(returnType), /*instanceMethod=*/false,
506 /*chainCall=*/false, argTypes, einfo, {}, RequiredArgs::All);
509 const CGFunctionInfo &
510 CodeGenTypes::arrangeGlobalDeclaration(GlobalDecl GD) {
511 // FIXME: Do we need to handle ObjCMethodDecl?
512 const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl());
514 if (const CXXConstructorDecl *CD = dyn_cast<CXXConstructorDecl>(FD))
515 return arrangeCXXStructorDeclaration(CD, getFromCtorType(GD.getCtorType()));
517 if (const CXXDestructorDecl *DD = dyn_cast<CXXDestructorDecl>(FD))
518 return arrangeCXXStructorDeclaration(DD, getFromDtorType(GD.getDtorType()));
520 return arrangeFunctionDeclaration(FD);
523 /// Arrange a thunk that takes 'this' as the first parameter followed by
524 /// varargs. Return a void pointer, regardless of the actual return type.
525 /// The body of the thunk will end in a musttail call to a function of the
526 /// correct type, and the caller will bitcast the function to the correct
528 const CGFunctionInfo &
529 CodeGenTypes::arrangeUnprototypedMustTailThunk(const CXXMethodDecl *MD) {
530 assert(MD->isVirtual() && "only methods have thunks");
531 CanQual<FunctionProtoType> FTP = GetFormalType(MD);
532 CanQualType ArgTys[] = { GetThisType(Context, MD->getParent()) };
533 return arrangeLLVMFunctionInfo(Context.VoidTy, /*instanceMethod=*/false,
534 /*chainCall=*/false, ArgTys,
535 FTP->getExtInfo(), {}, RequiredArgs(1));
538 const CGFunctionInfo &
539 CodeGenTypes::arrangeMSCtorClosure(const CXXConstructorDecl *CD,
541 assert(CT == Ctor_CopyingClosure || CT == Ctor_DefaultClosure);
543 CanQual<FunctionProtoType> FTP = GetFormalType(CD);
544 SmallVector<CanQualType, 2> ArgTys;
545 const CXXRecordDecl *RD = CD->getParent();
546 ArgTys.push_back(GetThisType(Context, RD));
547 if (CT == Ctor_CopyingClosure)
548 ArgTys.push_back(*FTP->param_type_begin());
549 if (RD->getNumVBases() > 0)
550 ArgTys.push_back(Context.IntTy);
551 CallingConv CC = Context.getDefaultCallingConvention(
552 /*IsVariadic=*/false, /*IsCXXMethod=*/true);
553 return arrangeLLVMFunctionInfo(Context.VoidTy, /*instanceMethod=*/true,
554 /*chainCall=*/false, ArgTys,
555 FunctionType::ExtInfo(CC), {},
559 /// Arrange a call as unto a free function, except possibly with an
560 /// additional number of formal parameters considered required.
561 static const CGFunctionInfo &
562 arrangeFreeFunctionLikeCall(CodeGenTypes &CGT,
564 const CallArgList &args,
565 const FunctionType *fnType,
566 unsigned numExtraRequiredArgs,
568 assert(args.size() >= numExtraRequiredArgs);
570 llvm::SmallVector<FunctionProtoType::ExtParameterInfo, 16> paramInfos;
572 // In most cases, there are no optional arguments.
573 RequiredArgs required = RequiredArgs::All;
575 // If we have a variadic prototype, the required arguments are the
576 // extra prefix plus the arguments in the prototype.
577 if (const FunctionProtoType *proto = dyn_cast<FunctionProtoType>(fnType)) {
578 if (proto->isVariadic())
579 required = RequiredArgs(proto->getNumParams() + numExtraRequiredArgs);
581 if (proto->hasExtParameterInfos())
582 addExtParameterInfosForCall(paramInfos, proto, numExtraRequiredArgs,
585 // If we don't have a prototype at all, but we're supposed to
586 // explicitly use the variadic convention for unprototyped calls,
587 // treat all of the arguments as required but preserve the nominal
588 // possibility of variadics.
589 } else if (CGM.getTargetCodeGenInfo()
590 .isNoProtoCallVariadic(args,
591 cast<FunctionNoProtoType>(fnType))) {
592 required = RequiredArgs(args.size());
596 SmallVector<CanQualType, 16> argTypes;
597 for (const auto &arg : args)
598 argTypes.push_back(CGT.getContext().getCanonicalParamType(arg.Ty));
599 return CGT.arrangeLLVMFunctionInfo(GetReturnType(fnType->getReturnType()),
600 /*instanceMethod=*/false, chainCall,
601 argTypes, fnType->getExtInfo(), paramInfos,
605 /// Figure out the rules for calling a function with the given formal
606 /// type using the given arguments. The arguments are necessary
607 /// because the function might be unprototyped, in which case it's
608 /// target-dependent in crazy ways.
609 const CGFunctionInfo &
610 CodeGenTypes::arrangeFreeFunctionCall(const CallArgList &args,
611 const FunctionType *fnType,
613 return arrangeFreeFunctionLikeCall(*this, CGM, args, fnType,
614 chainCall ? 1 : 0, chainCall);
617 /// A block function is essentially a free function with an
618 /// extra implicit argument.
619 const CGFunctionInfo &
620 CodeGenTypes::arrangeBlockFunctionCall(const CallArgList &args,
621 const FunctionType *fnType) {
622 return arrangeFreeFunctionLikeCall(*this, CGM, args, fnType, 1,
623 /*chainCall=*/false);
626 const CGFunctionInfo &
627 CodeGenTypes::arrangeBlockFunctionDeclaration(const FunctionProtoType *proto,
628 const FunctionArgList ¶ms) {
629 auto paramInfos = getExtParameterInfosForCall(proto, 1, params.size());
630 auto argTypes = getArgTypesForDeclaration(Context, params);
632 return arrangeLLVMFunctionInfo(
633 GetReturnType(proto->getReturnType()),
634 /*instanceMethod*/ false, /*chainCall*/ false, argTypes,
635 proto->getExtInfo(), paramInfos,
636 RequiredArgs::forPrototypePlus(proto, 1, nullptr));
639 const CGFunctionInfo &
640 CodeGenTypes::arrangeBuiltinFunctionCall(QualType resultType,
641 const CallArgList &args) {
643 SmallVector<CanQualType, 16> argTypes;
644 for (const auto &Arg : args)
645 argTypes.push_back(Context.getCanonicalParamType(Arg.Ty));
646 return arrangeLLVMFunctionInfo(
647 GetReturnType(resultType), /*instanceMethod=*/false,
648 /*chainCall=*/false, argTypes, FunctionType::ExtInfo(),
649 /*paramInfos=*/ {}, RequiredArgs::All);
652 const CGFunctionInfo &
653 CodeGenTypes::arrangeBuiltinFunctionDeclaration(QualType resultType,
654 const FunctionArgList &args) {
655 auto argTypes = getArgTypesForDeclaration(Context, args);
657 return arrangeLLVMFunctionInfo(
658 GetReturnType(resultType), /*instanceMethod=*/false, /*chainCall=*/false,
659 argTypes, FunctionType::ExtInfo(), {}, RequiredArgs::All);
662 const CGFunctionInfo &
663 CodeGenTypes::arrangeBuiltinFunctionDeclaration(CanQualType resultType,
664 ArrayRef<CanQualType> argTypes) {
665 return arrangeLLVMFunctionInfo(
666 resultType, /*instanceMethod=*/false, /*chainCall=*/false,
667 argTypes, FunctionType::ExtInfo(), {}, RequiredArgs::All);
670 /// Arrange a call to a C++ method, passing the given arguments.
672 /// numPrefixArgs is the number of ABI-specific prefix arguments we have. It
673 /// does not count `this`.
674 const CGFunctionInfo &
675 CodeGenTypes::arrangeCXXMethodCall(const CallArgList &args,
676 const FunctionProtoType *proto,
677 RequiredArgs required,
678 unsigned numPrefixArgs) {
679 assert(numPrefixArgs + 1 <= args.size() &&
680 "Emitting a call with less args than the required prefix?");
681 // Add one to account for `this`. It's a bit awkward here, but we don't count
682 // `this` in similar places elsewhere.
684 getExtParameterInfosForCall(proto, numPrefixArgs + 1, args.size());
687 auto argTypes = getArgTypesForCall(Context, args);
689 FunctionType::ExtInfo info = proto->getExtInfo();
690 return arrangeLLVMFunctionInfo(
691 GetReturnType(proto->getReturnType()), /*instanceMethod=*/true,
692 /*chainCall=*/false, argTypes, info, paramInfos, required);
695 const CGFunctionInfo &CodeGenTypes::arrangeNullaryFunction() {
696 return arrangeLLVMFunctionInfo(
697 getContext().VoidTy, /*instanceMethod=*/false, /*chainCall=*/false,
698 None, FunctionType::ExtInfo(), {}, RequiredArgs::All);
701 const CGFunctionInfo &
702 CodeGenTypes::arrangeCall(const CGFunctionInfo &signature,
703 const CallArgList &args) {
704 assert(signature.arg_size() <= args.size());
705 if (signature.arg_size() == args.size())
708 SmallVector<FunctionProtoType::ExtParameterInfo, 16> paramInfos;
709 auto sigParamInfos = signature.getExtParameterInfos();
710 if (!sigParamInfos.empty()) {
711 paramInfos.append(sigParamInfos.begin(), sigParamInfos.end());
712 paramInfos.resize(args.size());
715 auto argTypes = getArgTypesForCall(Context, args);
717 assert(signature.getRequiredArgs().allowsOptionalArgs());
718 return arrangeLLVMFunctionInfo(signature.getReturnType(),
719 signature.isInstanceMethod(),
720 signature.isChainCall(),
722 signature.getExtInfo(),
724 signature.getRequiredArgs());
729 void computeSPIRKernelABIInfo(CodeGenModule &CGM, CGFunctionInfo &FI);
733 /// Arrange the argument and result information for an abstract value
734 /// of a given function type. This is the method which all of the
735 /// above functions ultimately defer to.
736 const CGFunctionInfo &
737 CodeGenTypes::arrangeLLVMFunctionInfo(CanQualType resultType,
740 ArrayRef<CanQualType> argTypes,
741 FunctionType::ExtInfo info,
742 ArrayRef<FunctionProtoType::ExtParameterInfo> paramInfos,
743 RequiredArgs required) {
744 assert(std::all_of(argTypes.begin(), argTypes.end(),
745 [](CanQualType T) { return T.isCanonicalAsParam(); }));
747 // Lookup or create unique function info.
748 llvm::FoldingSetNodeID ID;
749 CGFunctionInfo::Profile(ID, instanceMethod, chainCall, info, paramInfos,
750 required, resultType, argTypes);
752 void *insertPos = nullptr;
753 CGFunctionInfo *FI = FunctionInfos.FindNodeOrInsertPos(ID, insertPos);
757 unsigned CC = ClangCallConvToLLVMCallConv(info.getCC());
759 // Construct the function info. We co-allocate the ArgInfos.
760 FI = CGFunctionInfo::create(CC, instanceMethod, chainCall, info,
761 paramInfos, resultType, argTypes, required);
762 FunctionInfos.InsertNode(FI, insertPos);
764 bool inserted = FunctionsBeingProcessed.insert(FI).second;
766 assert(inserted && "Recursively being processed?");
768 // Compute ABI information.
769 if (CC == llvm::CallingConv::SPIR_KERNEL) {
770 // Force target independent argument handling for the host visible
772 computeSPIRKernelABIInfo(CGM, *FI);
773 } else if (info.getCC() == CC_Swift) {
774 swiftcall::computeABIInfo(CGM, *FI);
776 getABIInfo().computeInfo(*FI);
779 // Loop over all of the computed argument and return value info. If any of
780 // them are direct or extend without a specified coerce type, specify the
782 ABIArgInfo &retInfo = FI->getReturnInfo();
783 if (retInfo.canHaveCoerceToType() && retInfo.getCoerceToType() == nullptr)
784 retInfo.setCoerceToType(ConvertType(FI->getReturnType()));
786 for (auto &I : FI->arguments())
787 if (I.info.canHaveCoerceToType() && I.info.getCoerceToType() == nullptr)
788 I.info.setCoerceToType(ConvertType(I.type));
790 bool erased = FunctionsBeingProcessed.erase(FI); (void)erased;
791 assert(erased && "Not in set?");
796 CGFunctionInfo *CGFunctionInfo::create(unsigned llvmCC,
799 const FunctionType::ExtInfo &info,
800 ArrayRef<ExtParameterInfo> paramInfos,
801 CanQualType resultType,
802 ArrayRef<CanQualType> argTypes,
803 RequiredArgs required) {
804 assert(paramInfos.empty() || paramInfos.size() == argTypes.size());
807 operator new(totalSizeToAlloc<ArgInfo, ExtParameterInfo>(
808 argTypes.size() + 1, paramInfos.size()));
810 CGFunctionInfo *FI = new(buffer) CGFunctionInfo();
811 FI->CallingConvention = llvmCC;
812 FI->EffectiveCallingConvention = llvmCC;
813 FI->ASTCallingConvention = info.getCC();
814 FI->InstanceMethod = instanceMethod;
815 FI->ChainCall = chainCall;
816 FI->NoReturn = info.getNoReturn();
817 FI->ReturnsRetained = info.getProducesResult();
818 FI->NoCallerSavedRegs = info.getNoCallerSavedRegs();
819 FI->NoCfCheck = info.getNoCfCheck();
820 FI->Required = required;
821 FI->HasRegParm = info.getHasRegParm();
822 FI->RegParm = info.getRegParm();
823 FI->ArgStruct = nullptr;
824 FI->ArgStructAlign = 0;
825 FI->NumArgs = argTypes.size();
826 FI->HasExtParameterInfos = !paramInfos.empty();
827 FI->getArgsBuffer()[0].type = resultType;
828 for (unsigned i = 0, e = argTypes.size(); i != e; ++i)
829 FI->getArgsBuffer()[i + 1].type = argTypes[i];
830 for (unsigned i = 0, e = paramInfos.size(); i != e; ++i)
831 FI->getExtParameterInfosBuffer()[i] = paramInfos[i];
838 // ABIArgInfo::Expand implementation.
840 // Specifies the way QualType passed as ABIArgInfo::Expand is expanded.
841 struct TypeExpansion {
842 enum TypeExpansionKind {
843 // Elements of constant arrays are expanded recursively.
845 // Record fields are expanded recursively (but if record is a union, only
846 // the field with the largest size is expanded).
848 // For complex types, real and imaginary parts are expanded recursively.
850 // All other types are not expandable.
854 const TypeExpansionKind Kind;
856 TypeExpansion(TypeExpansionKind K) : Kind(K) {}
857 virtual ~TypeExpansion() {}
860 struct ConstantArrayExpansion : TypeExpansion {
864 ConstantArrayExpansion(QualType EltTy, uint64_t NumElts)
865 : TypeExpansion(TEK_ConstantArray), EltTy(EltTy), NumElts(NumElts) {}
866 static bool classof(const TypeExpansion *TE) {
867 return TE->Kind == TEK_ConstantArray;
871 struct RecordExpansion : TypeExpansion {
872 SmallVector<const CXXBaseSpecifier *, 1> Bases;
874 SmallVector<const FieldDecl *, 1> Fields;
876 RecordExpansion(SmallVector<const CXXBaseSpecifier *, 1> &&Bases,
877 SmallVector<const FieldDecl *, 1> &&Fields)
878 : TypeExpansion(TEK_Record), Bases(std::move(Bases)),
879 Fields(std::move(Fields)) {}
880 static bool classof(const TypeExpansion *TE) {
881 return TE->Kind == TEK_Record;
885 struct ComplexExpansion : TypeExpansion {
888 ComplexExpansion(QualType EltTy) : TypeExpansion(TEK_Complex), EltTy(EltTy) {}
889 static bool classof(const TypeExpansion *TE) {
890 return TE->Kind == TEK_Complex;
894 struct NoExpansion : TypeExpansion {
895 NoExpansion() : TypeExpansion(TEK_None) {}
896 static bool classof(const TypeExpansion *TE) {
897 return TE->Kind == TEK_None;
902 static std::unique_ptr<TypeExpansion>
903 getTypeExpansion(QualType Ty, const ASTContext &Context) {
904 if (const ConstantArrayType *AT = Context.getAsConstantArrayType(Ty)) {
905 return llvm::make_unique<ConstantArrayExpansion>(
906 AT->getElementType(), AT->getSize().getZExtValue());
908 if (const RecordType *RT = Ty->getAs<RecordType>()) {
909 SmallVector<const CXXBaseSpecifier *, 1> Bases;
910 SmallVector<const FieldDecl *, 1> Fields;
911 const RecordDecl *RD = RT->getDecl();
912 assert(!RD->hasFlexibleArrayMember() &&
913 "Cannot expand structure with flexible array.");
915 // Unions can be here only in degenerative cases - all the fields are same
916 // after flattening. Thus we have to use the "largest" field.
917 const FieldDecl *LargestFD = nullptr;
918 CharUnits UnionSize = CharUnits::Zero();
920 for (const auto *FD : RD->fields()) {
921 if (FD->isZeroLengthBitField(Context))
923 assert(!FD->isBitField() &&
924 "Cannot expand structure with bit-field members.");
925 CharUnits FieldSize = Context.getTypeSizeInChars(FD->getType());
926 if (UnionSize < FieldSize) {
927 UnionSize = FieldSize;
932 Fields.push_back(LargestFD);
934 if (const auto *CXXRD = dyn_cast<CXXRecordDecl>(RD)) {
935 assert(!CXXRD->isDynamicClass() &&
936 "cannot expand vtable pointers in dynamic classes");
937 for (const CXXBaseSpecifier &BS : CXXRD->bases())
938 Bases.push_back(&BS);
941 for (const auto *FD : RD->fields()) {
942 if (FD->isZeroLengthBitField(Context))
944 assert(!FD->isBitField() &&
945 "Cannot expand structure with bit-field members.");
946 Fields.push_back(FD);
949 return llvm::make_unique<RecordExpansion>(std::move(Bases),
952 if (const ComplexType *CT = Ty->getAs<ComplexType>()) {
953 return llvm::make_unique<ComplexExpansion>(CT->getElementType());
955 return llvm::make_unique<NoExpansion>();
958 static int getExpansionSize(QualType Ty, const ASTContext &Context) {
959 auto Exp = getTypeExpansion(Ty, Context);
960 if (auto CAExp = dyn_cast<ConstantArrayExpansion>(Exp.get())) {
961 return CAExp->NumElts * getExpansionSize(CAExp->EltTy, Context);
963 if (auto RExp = dyn_cast<RecordExpansion>(Exp.get())) {
965 for (auto BS : RExp->Bases)
966 Res += getExpansionSize(BS->getType(), Context);
967 for (auto FD : RExp->Fields)
968 Res += getExpansionSize(FD->getType(), Context);
971 if (isa<ComplexExpansion>(Exp.get()))
973 assert(isa<NoExpansion>(Exp.get()));
978 CodeGenTypes::getExpandedTypes(QualType Ty,
979 SmallVectorImpl<llvm::Type *>::iterator &TI) {
980 auto Exp = getTypeExpansion(Ty, Context);
981 if (auto CAExp = dyn_cast<ConstantArrayExpansion>(Exp.get())) {
982 for (int i = 0, n = CAExp->NumElts; i < n; i++) {
983 getExpandedTypes(CAExp->EltTy, TI);
985 } else if (auto RExp = dyn_cast<RecordExpansion>(Exp.get())) {
986 for (auto BS : RExp->Bases)
987 getExpandedTypes(BS->getType(), TI);
988 for (auto FD : RExp->Fields)
989 getExpandedTypes(FD->getType(), TI);
990 } else if (auto CExp = dyn_cast<ComplexExpansion>(Exp.get())) {
991 llvm::Type *EltTy = ConvertType(CExp->EltTy);
995 assert(isa<NoExpansion>(Exp.get()));
996 *TI++ = ConvertType(Ty);
1000 static void forConstantArrayExpansion(CodeGenFunction &CGF,
1001 ConstantArrayExpansion *CAE,
1003 llvm::function_ref<void(Address)> Fn) {
1004 CharUnits EltSize = CGF.getContext().getTypeSizeInChars(CAE->EltTy);
1005 CharUnits EltAlign =
1006 BaseAddr.getAlignment().alignmentOfArrayElement(EltSize);
1008 for (int i = 0, n = CAE->NumElts; i < n; i++) {
1009 llvm::Value *EltAddr =
1010 CGF.Builder.CreateConstGEP2_32(nullptr, BaseAddr.getPointer(), 0, i);
1011 Fn(Address(EltAddr, EltAlign));
1015 void CodeGenFunction::ExpandTypeFromArgs(
1016 QualType Ty, LValue LV, SmallVectorImpl<llvm::Value *>::iterator &AI) {
1017 assert(LV.isSimple() &&
1018 "Unexpected non-simple lvalue during struct expansion.");
1020 auto Exp = getTypeExpansion(Ty, getContext());
1021 if (auto CAExp = dyn_cast<ConstantArrayExpansion>(Exp.get())) {
1022 forConstantArrayExpansion(*this, CAExp, LV.getAddress(),
1023 [&](Address EltAddr) {
1024 LValue LV = MakeAddrLValue(EltAddr, CAExp->EltTy);
1025 ExpandTypeFromArgs(CAExp->EltTy, LV, AI);
1027 } else if (auto RExp = dyn_cast<RecordExpansion>(Exp.get())) {
1028 Address This = LV.getAddress();
1029 for (const CXXBaseSpecifier *BS : RExp->Bases) {
1030 // Perform a single step derived-to-base conversion.
1032 GetAddressOfBaseClass(This, Ty->getAsCXXRecordDecl(), &BS, &BS + 1,
1033 /*NullCheckValue=*/false, SourceLocation());
1034 LValue SubLV = MakeAddrLValue(Base, BS->getType());
1036 // Recurse onto bases.
1037 ExpandTypeFromArgs(BS->getType(), SubLV, AI);
1039 for (auto FD : RExp->Fields) {
1040 // FIXME: What are the right qualifiers here?
1041 LValue SubLV = EmitLValueForFieldInitialization(LV, FD);
1042 ExpandTypeFromArgs(FD->getType(), SubLV, AI);
1044 } else if (isa<ComplexExpansion>(Exp.get())) {
1045 auto realValue = *AI++;
1046 auto imagValue = *AI++;
1047 EmitStoreOfComplex(ComplexPairTy(realValue, imagValue), LV, /*init*/ true);
1049 assert(isa<NoExpansion>(Exp.get()));
1050 EmitStoreThroughLValue(RValue::get(*AI++), LV);
1054 void CodeGenFunction::ExpandTypeToArgs(
1055 QualType Ty, CallArg Arg, llvm::FunctionType *IRFuncTy,
1056 SmallVectorImpl<llvm::Value *> &IRCallArgs, unsigned &IRCallArgPos) {
1057 auto Exp = getTypeExpansion(Ty, getContext());
1058 if (auto CAExp = dyn_cast<ConstantArrayExpansion>(Exp.get())) {
1059 Address Addr = Arg.hasLValue() ? Arg.getKnownLValue().getAddress()
1060 : Arg.getKnownRValue().getAggregateAddress();
1061 forConstantArrayExpansion(
1062 *this, CAExp, Addr, [&](Address EltAddr) {
1063 CallArg EltArg = CallArg(
1064 convertTempToRValue(EltAddr, CAExp->EltTy, SourceLocation()),
1066 ExpandTypeToArgs(CAExp->EltTy, EltArg, IRFuncTy, IRCallArgs,
1069 } else if (auto RExp = dyn_cast<RecordExpansion>(Exp.get())) {
1070 Address This = Arg.hasLValue() ? Arg.getKnownLValue().getAddress()
1071 : Arg.getKnownRValue().getAggregateAddress();
1072 for (const CXXBaseSpecifier *BS : RExp->Bases) {
1073 // Perform a single step derived-to-base conversion.
1075 GetAddressOfBaseClass(This, Ty->getAsCXXRecordDecl(), &BS, &BS + 1,
1076 /*NullCheckValue=*/false, SourceLocation());
1077 CallArg BaseArg = CallArg(RValue::getAggregate(Base), BS->getType());
1079 // Recurse onto bases.
1080 ExpandTypeToArgs(BS->getType(), BaseArg, IRFuncTy, IRCallArgs,
1084 LValue LV = MakeAddrLValue(This, Ty);
1085 for (auto FD : RExp->Fields) {
1087 CallArg(EmitRValueForField(LV, FD, SourceLocation()), FD->getType());
1088 ExpandTypeToArgs(FD->getType(), FldArg, IRFuncTy, IRCallArgs,
1091 } else if (isa<ComplexExpansion>(Exp.get())) {
1092 ComplexPairTy CV = Arg.getKnownRValue().getComplexVal();
1093 IRCallArgs[IRCallArgPos++] = CV.first;
1094 IRCallArgs[IRCallArgPos++] = CV.second;
1096 assert(isa<NoExpansion>(Exp.get()));
1097 auto RV = Arg.getKnownRValue();
1098 assert(RV.isScalar() &&
1099 "Unexpected non-scalar rvalue during struct expansion.");
1101 // Insert a bitcast as needed.
1102 llvm::Value *V = RV.getScalarVal();
1103 if (IRCallArgPos < IRFuncTy->getNumParams() &&
1104 V->getType() != IRFuncTy->getParamType(IRCallArgPos))
1105 V = Builder.CreateBitCast(V, IRFuncTy->getParamType(IRCallArgPos));
1107 IRCallArgs[IRCallArgPos++] = V;
1111 /// Create a temporary allocation for the purposes of coercion.
1112 static Address CreateTempAllocaForCoercion(CodeGenFunction &CGF, llvm::Type *Ty,
1113 CharUnits MinAlign) {
1114 // Don't use an alignment that's worse than what LLVM would prefer.
1115 auto PrefAlign = CGF.CGM.getDataLayout().getPrefTypeAlignment(Ty);
1116 CharUnits Align = std::max(MinAlign, CharUnits::fromQuantity(PrefAlign));
1118 return CGF.CreateTempAlloca(Ty, Align);
1121 /// EnterStructPointerForCoercedAccess - Given a struct pointer that we are
1122 /// accessing some number of bytes out of it, try to gep into the struct to get
1123 /// at its inner goodness. Dive as deep as possible without entering an element
1124 /// with an in-memory size smaller than DstSize.
1126 EnterStructPointerForCoercedAccess(Address SrcPtr,
1127 llvm::StructType *SrcSTy,
1128 uint64_t DstSize, CodeGenFunction &CGF) {
1129 // We can't dive into a zero-element struct.
1130 if (SrcSTy->getNumElements() == 0) return SrcPtr;
1132 llvm::Type *FirstElt = SrcSTy->getElementType(0);
1134 // If the first elt is at least as large as what we're looking for, or if the
1135 // first element is the same size as the whole struct, we can enter it. The
1136 // comparison must be made on the store size and not the alloca size. Using
1137 // the alloca size may overstate the size of the load.
1138 uint64_t FirstEltSize =
1139 CGF.CGM.getDataLayout().getTypeStoreSize(FirstElt);
1140 if (FirstEltSize < DstSize &&
1141 FirstEltSize < CGF.CGM.getDataLayout().getTypeStoreSize(SrcSTy))
1144 // GEP into the first element.
1145 SrcPtr = CGF.Builder.CreateStructGEP(SrcPtr, 0, CharUnits(), "coerce.dive");
1147 // If the first element is a struct, recurse.
1148 llvm::Type *SrcTy = SrcPtr.getElementType();
1149 if (llvm::StructType *SrcSTy = dyn_cast<llvm::StructType>(SrcTy))
1150 return EnterStructPointerForCoercedAccess(SrcPtr, SrcSTy, DstSize, CGF);
1155 /// CoerceIntOrPtrToIntOrPtr - Convert a value Val to the specific Ty where both
1156 /// are either integers or pointers. This does a truncation of the value if it
1157 /// is too large or a zero extension if it is too small.
1159 /// This behaves as if the value were coerced through memory, so on big-endian
1160 /// targets the high bits are preserved in a truncation, while little-endian
1161 /// targets preserve the low bits.
1162 static llvm::Value *CoerceIntOrPtrToIntOrPtr(llvm::Value *Val,
1164 CodeGenFunction &CGF) {
1165 if (Val->getType() == Ty)
1168 if (isa<llvm::PointerType>(Val->getType())) {
1169 // If this is Pointer->Pointer avoid conversion to and from int.
1170 if (isa<llvm::PointerType>(Ty))
1171 return CGF.Builder.CreateBitCast(Val, Ty, "coerce.val");
1173 // Convert the pointer to an integer so we can play with its width.
1174 Val = CGF.Builder.CreatePtrToInt(Val, CGF.IntPtrTy, "coerce.val.pi");
1177 llvm::Type *DestIntTy = Ty;
1178 if (isa<llvm::PointerType>(DestIntTy))
1179 DestIntTy = CGF.IntPtrTy;
1181 if (Val->getType() != DestIntTy) {
1182 const llvm::DataLayout &DL = CGF.CGM.getDataLayout();
1183 if (DL.isBigEndian()) {
1184 // Preserve the high bits on big-endian targets.
1185 // That is what memory coercion does.
1186 uint64_t SrcSize = DL.getTypeSizeInBits(Val->getType());
1187 uint64_t DstSize = DL.getTypeSizeInBits(DestIntTy);
1189 if (SrcSize > DstSize) {
1190 Val = CGF.Builder.CreateLShr(Val, SrcSize - DstSize, "coerce.highbits");
1191 Val = CGF.Builder.CreateTrunc(Val, DestIntTy, "coerce.val.ii");
1193 Val = CGF.Builder.CreateZExt(Val, DestIntTy, "coerce.val.ii");
1194 Val = CGF.Builder.CreateShl(Val, DstSize - SrcSize, "coerce.highbits");
1197 // Little-endian targets preserve the low bits. No shifts required.
1198 Val = CGF.Builder.CreateIntCast(Val, DestIntTy, false, "coerce.val.ii");
1202 if (isa<llvm::PointerType>(Ty))
1203 Val = CGF.Builder.CreateIntToPtr(Val, Ty, "coerce.val.ip");
1209 /// CreateCoercedLoad - Create a load from \arg SrcPtr interpreted as
1210 /// a pointer to an object of type \arg Ty, known to be aligned to
1211 /// \arg SrcAlign bytes.
1213 /// This safely handles the case when the src type is smaller than the
1214 /// destination type; in this situation the values of bits which not
1215 /// present in the src are undefined.
1216 static llvm::Value *CreateCoercedLoad(Address Src, llvm::Type *Ty,
1217 CodeGenFunction &CGF) {
1218 llvm::Type *SrcTy = Src.getElementType();
1220 // If SrcTy and Ty are the same, just do a load.
1222 return CGF.Builder.CreateLoad(Src);
1224 uint64_t DstSize = CGF.CGM.getDataLayout().getTypeAllocSize(Ty);
1226 if (llvm::StructType *SrcSTy = dyn_cast<llvm::StructType>(SrcTy)) {
1227 Src = EnterStructPointerForCoercedAccess(Src, SrcSTy, DstSize, CGF);
1228 SrcTy = Src.getType()->getElementType();
1231 uint64_t SrcSize = CGF.CGM.getDataLayout().getTypeAllocSize(SrcTy);
1233 // If the source and destination are integer or pointer types, just do an
1234 // extension or truncation to the desired type.
1235 if ((isa<llvm::IntegerType>(Ty) || isa<llvm::PointerType>(Ty)) &&
1236 (isa<llvm::IntegerType>(SrcTy) || isa<llvm::PointerType>(SrcTy))) {
1237 llvm::Value *Load = CGF.Builder.CreateLoad(Src);
1238 return CoerceIntOrPtrToIntOrPtr(Load, Ty, CGF);
1241 // If load is legal, just bitcast the src pointer.
1242 if (SrcSize >= DstSize) {
1243 // Generally SrcSize is never greater than DstSize, since this means we are
1244 // losing bits. However, this can happen in cases where the structure has
1245 // additional padding, for example due to a user specified alignment.
1247 // FIXME: Assert that we aren't truncating non-padding bits when have access
1248 // to that information.
1249 Src = CGF.Builder.CreateBitCast(Src,
1250 Ty->getPointerTo(Src.getAddressSpace()));
1251 return CGF.Builder.CreateLoad(Src);
1254 // Otherwise do coercion through memory. This is stupid, but simple.
1255 Address Tmp = CreateTempAllocaForCoercion(CGF, Ty, Src.getAlignment());
1256 Address Casted = CGF.Builder.CreateBitCast(Tmp, CGF.AllocaInt8PtrTy);
1257 Address SrcCasted = CGF.Builder.CreateBitCast(Src, CGF.AllocaInt8PtrTy);
1258 CGF.Builder.CreateMemCpy(Casted, SrcCasted,
1259 llvm::ConstantInt::get(CGF.IntPtrTy, SrcSize),
1261 return CGF.Builder.CreateLoad(Tmp);
1264 // Function to store a first-class aggregate into memory. We prefer to
1265 // store the elements rather than the aggregate to be more friendly to
1267 // FIXME: Do we need to recurse here?
1268 static void BuildAggStore(CodeGenFunction &CGF, llvm::Value *Val,
1269 Address Dest, bool DestIsVolatile) {
1270 // Prefer scalar stores to first-class aggregate stores.
1271 if (llvm::StructType *STy =
1272 dyn_cast<llvm::StructType>(Val->getType())) {
1273 const llvm::StructLayout *Layout =
1274 CGF.CGM.getDataLayout().getStructLayout(STy);
1276 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1277 auto EltOffset = CharUnits::fromQuantity(Layout->getElementOffset(i));
1278 Address EltPtr = CGF.Builder.CreateStructGEP(Dest, i, EltOffset);
1279 llvm::Value *Elt = CGF.Builder.CreateExtractValue(Val, i);
1280 CGF.Builder.CreateStore(Elt, EltPtr, DestIsVolatile);
1283 CGF.Builder.CreateStore(Val, Dest, DestIsVolatile);
1287 /// CreateCoercedStore - Create a store to \arg DstPtr from \arg Src,
1288 /// where the source and destination may have different types. The
1289 /// destination is known to be aligned to \arg DstAlign bytes.
1291 /// This safely handles the case when the src type is larger than the
1292 /// destination type; the upper bits of the src will be lost.
1293 static void CreateCoercedStore(llvm::Value *Src,
1296 CodeGenFunction &CGF) {
1297 llvm::Type *SrcTy = Src->getType();
1298 llvm::Type *DstTy = Dst.getType()->getElementType();
1299 if (SrcTy == DstTy) {
1300 CGF.Builder.CreateStore(Src, Dst, DstIsVolatile);
1304 uint64_t SrcSize = CGF.CGM.getDataLayout().getTypeAllocSize(SrcTy);
1306 if (llvm::StructType *DstSTy = dyn_cast<llvm::StructType>(DstTy)) {
1307 Dst = EnterStructPointerForCoercedAccess(Dst, DstSTy, SrcSize, CGF);
1308 DstTy = Dst.getType()->getElementType();
1311 // If the source and destination are integer or pointer types, just do an
1312 // extension or truncation to the desired type.
1313 if ((isa<llvm::IntegerType>(SrcTy) || isa<llvm::PointerType>(SrcTy)) &&
1314 (isa<llvm::IntegerType>(DstTy) || isa<llvm::PointerType>(DstTy))) {
1315 Src = CoerceIntOrPtrToIntOrPtr(Src, DstTy, CGF);
1316 CGF.Builder.CreateStore(Src, Dst, DstIsVolatile);
1320 uint64_t DstSize = CGF.CGM.getDataLayout().getTypeAllocSize(DstTy);
1322 // If store is legal, just bitcast the src pointer.
1323 if (SrcSize <= DstSize) {
1324 Dst = CGF.Builder.CreateElementBitCast(Dst, SrcTy);
1325 BuildAggStore(CGF, Src, Dst, DstIsVolatile);
1327 // Otherwise do coercion through memory. This is stupid, but
1330 // Generally SrcSize is never greater than DstSize, since this means we are
1331 // losing bits. However, this can happen in cases where the structure has
1332 // additional padding, for example due to a user specified alignment.
1334 // FIXME: Assert that we aren't truncating non-padding bits when have access
1335 // to that information.
1336 Address Tmp = CreateTempAllocaForCoercion(CGF, SrcTy, Dst.getAlignment());
1337 CGF.Builder.CreateStore(Src, Tmp);
1338 Address Casted = CGF.Builder.CreateBitCast(Tmp, CGF.AllocaInt8PtrTy);
1339 Address DstCasted = CGF.Builder.CreateBitCast(Dst, CGF.AllocaInt8PtrTy);
1340 CGF.Builder.CreateMemCpy(DstCasted, Casted,
1341 llvm::ConstantInt::get(CGF.IntPtrTy, DstSize),
1346 static Address emitAddressAtOffset(CodeGenFunction &CGF, Address addr,
1347 const ABIArgInfo &info) {
1348 if (unsigned offset = info.getDirectOffset()) {
1349 addr = CGF.Builder.CreateElementBitCast(addr, CGF.Int8Ty);
1350 addr = CGF.Builder.CreateConstInBoundsByteGEP(addr,
1351 CharUnits::fromQuantity(offset));
1352 addr = CGF.Builder.CreateElementBitCast(addr, info.getCoerceToType());
1359 /// Encapsulates information about the way function arguments from
1360 /// CGFunctionInfo should be passed to actual LLVM IR function.
1361 class ClangToLLVMArgMapping {
1362 static const unsigned InvalidIndex = ~0U;
1363 unsigned InallocaArgNo;
1365 unsigned TotalIRArgs;
1367 /// Arguments of LLVM IR function corresponding to single Clang argument.
1369 unsigned PaddingArgIndex;
1370 // Argument is expanded to IR arguments at positions
1371 // [FirstArgIndex, FirstArgIndex + NumberOfArgs).
1372 unsigned FirstArgIndex;
1373 unsigned NumberOfArgs;
1376 : PaddingArgIndex(InvalidIndex), FirstArgIndex(InvalidIndex),
1380 SmallVector<IRArgs, 8> ArgInfo;
1383 ClangToLLVMArgMapping(const ASTContext &Context, const CGFunctionInfo &FI,
1384 bool OnlyRequiredArgs = false)
1385 : InallocaArgNo(InvalidIndex), SRetArgNo(InvalidIndex), TotalIRArgs(0),
1386 ArgInfo(OnlyRequiredArgs ? FI.getNumRequiredArgs() : FI.arg_size()) {
1387 construct(Context, FI, OnlyRequiredArgs);
1390 bool hasInallocaArg() const { return InallocaArgNo != InvalidIndex; }
1391 unsigned getInallocaArgNo() const {
1392 assert(hasInallocaArg());
1393 return InallocaArgNo;
1396 bool hasSRetArg() const { return SRetArgNo != InvalidIndex; }
1397 unsigned getSRetArgNo() const {
1398 assert(hasSRetArg());
1402 unsigned totalIRArgs() const { return TotalIRArgs; }
1404 bool hasPaddingArg(unsigned ArgNo) const {
1405 assert(ArgNo < ArgInfo.size());
1406 return ArgInfo[ArgNo].PaddingArgIndex != InvalidIndex;
1408 unsigned getPaddingArgNo(unsigned ArgNo) const {
1409 assert(hasPaddingArg(ArgNo));
1410 return ArgInfo[ArgNo].PaddingArgIndex;
1413 /// Returns index of first IR argument corresponding to ArgNo, and their
1415 std::pair<unsigned, unsigned> getIRArgs(unsigned ArgNo) const {
1416 assert(ArgNo < ArgInfo.size());
1417 return std::make_pair(ArgInfo[ArgNo].FirstArgIndex,
1418 ArgInfo[ArgNo].NumberOfArgs);
1422 void construct(const ASTContext &Context, const CGFunctionInfo &FI,
1423 bool OnlyRequiredArgs);
1426 void ClangToLLVMArgMapping::construct(const ASTContext &Context,
1427 const CGFunctionInfo &FI,
1428 bool OnlyRequiredArgs) {
1429 unsigned IRArgNo = 0;
1430 bool SwapThisWithSRet = false;
1431 const ABIArgInfo &RetAI = FI.getReturnInfo();
1433 if (RetAI.getKind() == ABIArgInfo::Indirect) {
1434 SwapThisWithSRet = RetAI.isSRetAfterThis();
1435 SRetArgNo = SwapThisWithSRet ? 1 : IRArgNo++;
1439 unsigned NumArgs = OnlyRequiredArgs ? FI.getNumRequiredArgs() : FI.arg_size();
1440 for (CGFunctionInfo::const_arg_iterator I = FI.arg_begin(); ArgNo < NumArgs;
1442 assert(I != FI.arg_end());
1443 QualType ArgType = I->type;
1444 const ABIArgInfo &AI = I->info;
1445 // Collect data about IR arguments corresponding to Clang argument ArgNo.
1446 auto &IRArgs = ArgInfo[ArgNo];
1448 if (AI.getPaddingType())
1449 IRArgs.PaddingArgIndex = IRArgNo++;
1451 switch (AI.getKind()) {
1452 case ABIArgInfo::Extend:
1453 case ABIArgInfo::Direct: {
1454 // FIXME: handle sseregparm someday...
1455 llvm::StructType *STy = dyn_cast<llvm::StructType>(AI.getCoerceToType());
1456 if (AI.isDirect() && AI.getCanBeFlattened() && STy) {
1457 IRArgs.NumberOfArgs = STy->getNumElements();
1459 IRArgs.NumberOfArgs = 1;
1463 case ABIArgInfo::Indirect:
1464 IRArgs.NumberOfArgs = 1;
1466 case ABIArgInfo::Ignore:
1467 case ABIArgInfo::InAlloca:
1468 // ignore and inalloca doesn't have matching LLVM parameters.
1469 IRArgs.NumberOfArgs = 0;
1471 case ABIArgInfo::CoerceAndExpand:
1472 IRArgs.NumberOfArgs = AI.getCoerceAndExpandTypeSequence().size();
1474 case ABIArgInfo::Expand:
1475 IRArgs.NumberOfArgs = getExpansionSize(ArgType, Context);
1479 if (IRArgs.NumberOfArgs > 0) {
1480 IRArgs.FirstArgIndex = IRArgNo;
1481 IRArgNo += IRArgs.NumberOfArgs;
1484 // Skip over the sret parameter when it comes second. We already handled it
1486 if (IRArgNo == 1 && SwapThisWithSRet)
1489 assert(ArgNo == ArgInfo.size());
1491 if (FI.usesInAlloca())
1492 InallocaArgNo = IRArgNo++;
1494 TotalIRArgs = IRArgNo;
1500 bool CodeGenModule::ReturnTypeUsesSRet(const CGFunctionInfo &FI) {
1501 const auto &RI = FI.getReturnInfo();
1502 return RI.isIndirect() || (RI.isInAlloca() && RI.getInAllocaSRet());
1505 bool CodeGenModule::ReturnSlotInterferesWithArgs(const CGFunctionInfo &FI) {
1506 return ReturnTypeUsesSRet(FI) &&
1507 getTargetCodeGenInfo().doesReturnSlotInterfereWithArgs();
1510 bool CodeGenModule::ReturnTypeUsesFPRet(QualType ResultType) {
1511 if (const BuiltinType *BT = ResultType->getAs<BuiltinType>()) {
1512 switch (BT->getKind()) {
1515 case BuiltinType::Float:
1516 return getTarget().useObjCFPRetForRealType(TargetInfo::Float);
1517 case BuiltinType::Double:
1518 return getTarget().useObjCFPRetForRealType(TargetInfo::Double);
1519 case BuiltinType::LongDouble:
1520 return getTarget().useObjCFPRetForRealType(TargetInfo::LongDouble);
1527 bool CodeGenModule::ReturnTypeUsesFP2Ret(QualType ResultType) {
1528 if (const ComplexType *CT = ResultType->getAs<ComplexType>()) {
1529 if (const BuiltinType *BT = CT->getElementType()->getAs<BuiltinType>()) {
1530 if (BT->getKind() == BuiltinType::LongDouble)
1531 return getTarget().useObjCFP2RetForComplexLongDouble();
1538 llvm::FunctionType *CodeGenTypes::GetFunctionType(GlobalDecl GD) {
1539 const CGFunctionInfo &FI = arrangeGlobalDeclaration(GD);
1540 return GetFunctionType(FI);
1543 llvm::FunctionType *
1544 CodeGenTypes::GetFunctionType(const CGFunctionInfo &FI) {
1546 bool Inserted = FunctionsBeingProcessed.insert(&FI).second;
1548 assert(Inserted && "Recursively being processed?");
1550 llvm::Type *resultType = nullptr;
1551 const ABIArgInfo &retAI = FI.getReturnInfo();
1552 switch (retAI.getKind()) {
1553 case ABIArgInfo::Expand:
1554 llvm_unreachable("Invalid ABI kind for return argument");
1556 case ABIArgInfo::Extend:
1557 case ABIArgInfo::Direct:
1558 resultType = retAI.getCoerceToType();
1561 case ABIArgInfo::InAlloca:
1562 if (retAI.getInAllocaSRet()) {
1563 // sret things on win32 aren't void, they return the sret pointer.
1564 QualType ret = FI.getReturnType();
1565 llvm::Type *ty = ConvertType(ret);
1566 unsigned addressSpace = Context.getTargetAddressSpace(ret);
1567 resultType = llvm::PointerType::get(ty, addressSpace);
1569 resultType = llvm::Type::getVoidTy(getLLVMContext());
1573 case ABIArgInfo::Indirect:
1574 case ABIArgInfo::Ignore:
1575 resultType = llvm::Type::getVoidTy(getLLVMContext());
1578 case ABIArgInfo::CoerceAndExpand:
1579 resultType = retAI.getUnpaddedCoerceAndExpandType();
1583 ClangToLLVMArgMapping IRFunctionArgs(getContext(), FI, true);
1584 SmallVector<llvm::Type*, 8> ArgTypes(IRFunctionArgs.totalIRArgs());
1586 // Add type for sret argument.
1587 if (IRFunctionArgs.hasSRetArg()) {
1588 QualType Ret = FI.getReturnType();
1589 llvm::Type *Ty = ConvertType(Ret);
1590 unsigned AddressSpace = Context.getTargetAddressSpace(Ret);
1591 ArgTypes[IRFunctionArgs.getSRetArgNo()] =
1592 llvm::PointerType::get(Ty, AddressSpace);
1595 // Add type for inalloca argument.
1596 if (IRFunctionArgs.hasInallocaArg()) {
1597 auto ArgStruct = FI.getArgStruct();
1599 ArgTypes[IRFunctionArgs.getInallocaArgNo()] = ArgStruct->getPointerTo();
1602 // Add in all of the required arguments.
1604 CGFunctionInfo::const_arg_iterator it = FI.arg_begin(),
1605 ie = it + FI.getNumRequiredArgs();
1606 for (; it != ie; ++it, ++ArgNo) {
1607 const ABIArgInfo &ArgInfo = it->info;
1609 // Insert a padding type to ensure proper alignment.
1610 if (IRFunctionArgs.hasPaddingArg(ArgNo))
1611 ArgTypes[IRFunctionArgs.getPaddingArgNo(ArgNo)] =
1612 ArgInfo.getPaddingType();
1614 unsigned FirstIRArg, NumIRArgs;
1615 std::tie(FirstIRArg, NumIRArgs) = IRFunctionArgs.getIRArgs(ArgNo);
1617 switch (ArgInfo.getKind()) {
1618 case ABIArgInfo::Ignore:
1619 case ABIArgInfo::InAlloca:
1620 assert(NumIRArgs == 0);
1623 case ABIArgInfo::Indirect: {
1624 assert(NumIRArgs == 1);
1625 // indirect arguments are always on the stack, which is alloca addr space.
1626 llvm::Type *LTy = ConvertTypeForMem(it->type);
1627 ArgTypes[FirstIRArg] = LTy->getPointerTo(
1628 CGM.getDataLayout().getAllocaAddrSpace());
1632 case ABIArgInfo::Extend:
1633 case ABIArgInfo::Direct: {
1634 // Fast-isel and the optimizer generally like scalar values better than
1635 // FCAs, so we flatten them if this is safe to do for this argument.
1636 llvm::Type *argType = ArgInfo.getCoerceToType();
1637 llvm::StructType *st = dyn_cast<llvm::StructType>(argType);
1638 if (st && ArgInfo.isDirect() && ArgInfo.getCanBeFlattened()) {
1639 assert(NumIRArgs == st->getNumElements());
1640 for (unsigned i = 0, e = st->getNumElements(); i != e; ++i)
1641 ArgTypes[FirstIRArg + i] = st->getElementType(i);
1643 assert(NumIRArgs == 1);
1644 ArgTypes[FirstIRArg] = argType;
1649 case ABIArgInfo::CoerceAndExpand: {
1650 auto ArgTypesIter = ArgTypes.begin() + FirstIRArg;
1651 for (auto EltTy : ArgInfo.getCoerceAndExpandTypeSequence()) {
1652 *ArgTypesIter++ = EltTy;
1654 assert(ArgTypesIter == ArgTypes.begin() + FirstIRArg + NumIRArgs);
1658 case ABIArgInfo::Expand:
1659 auto ArgTypesIter = ArgTypes.begin() + FirstIRArg;
1660 getExpandedTypes(it->type, ArgTypesIter);
1661 assert(ArgTypesIter == ArgTypes.begin() + FirstIRArg + NumIRArgs);
1666 bool Erased = FunctionsBeingProcessed.erase(&FI); (void)Erased;
1667 assert(Erased && "Not in set?");
1669 return llvm::FunctionType::get(resultType, ArgTypes, FI.isVariadic());
1672 llvm::Type *CodeGenTypes::GetFunctionTypeForVTable(GlobalDecl GD) {
1673 const CXXMethodDecl *MD = cast<CXXMethodDecl>(GD.getDecl());
1674 const FunctionProtoType *FPT = MD->getType()->getAs<FunctionProtoType>();
1676 if (!isFuncTypeConvertible(FPT))
1677 return llvm::StructType::get(getLLVMContext());
1679 const CGFunctionInfo *Info;
1680 if (isa<CXXDestructorDecl>(MD))
1682 &arrangeCXXStructorDeclaration(MD, getFromDtorType(GD.getDtorType()));
1684 Info = &arrangeCXXMethodDeclaration(MD);
1685 return GetFunctionType(*Info);
1688 static void AddAttributesFromFunctionProtoType(ASTContext &Ctx,
1689 llvm::AttrBuilder &FuncAttrs,
1690 const FunctionProtoType *FPT) {
1694 if (!isUnresolvedExceptionSpec(FPT->getExceptionSpecType()) &&
1696 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1699 void CodeGenModule::ConstructDefaultFnAttrList(StringRef Name, bool HasOptnone,
1700 bool AttrOnCallSite,
1701 llvm::AttrBuilder &FuncAttrs) {
1702 // OptimizeNoneAttr takes precedence over -Os or -Oz. No warning needed.
1704 if (CodeGenOpts.OptimizeSize)
1705 FuncAttrs.addAttribute(llvm::Attribute::OptimizeForSize);
1706 if (CodeGenOpts.OptimizeSize == 2)
1707 FuncAttrs.addAttribute(llvm::Attribute::MinSize);
1710 if (CodeGenOpts.DisableRedZone)
1711 FuncAttrs.addAttribute(llvm::Attribute::NoRedZone);
1712 if (CodeGenOpts.NoImplicitFloat)
1713 FuncAttrs.addAttribute(llvm::Attribute::NoImplicitFloat);
1715 if (AttrOnCallSite) {
1716 // Attributes that should go on the call site only.
1717 if (!CodeGenOpts.SimplifyLibCalls ||
1718 CodeGenOpts.isNoBuiltinFunc(Name.data()))
1719 FuncAttrs.addAttribute(llvm::Attribute::NoBuiltin);
1720 if (!CodeGenOpts.TrapFuncName.empty())
1721 FuncAttrs.addAttribute("trap-func-name", CodeGenOpts.TrapFuncName);
1723 // Attributes that should go on the function, but not the call site.
1724 if (!CodeGenOpts.DisableFPElim) {
1725 FuncAttrs.addAttribute("no-frame-pointer-elim", "false");
1726 } else if (CodeGenOpts.OmitLeafFramePointer) {
1727 FuncAttrs.addAttribute("no-frame-pointer-elim", "false");
1728 FuncAttrs.addAttribute("no-frame-pointer-elim-non-leaf");
1730 FuncAttrs.addAttribute("no-frame-pointer-elim", "true");
1731 FuncAttrs.addAttribute("no-frame-pointer-elim-non-leaf");
1734 FuncAttrs.addAttribute("less-precise-fpmad",
1735 llvm::toStringRef(CodeGenOpts.LessPreciseFPMAD));
1737 if (CodeGenOpts.NullPointerIsValid)
1738 FuncAttrs.addAttribute("null-pointer-is-valid", "true");
1739 if (!CodeGenOpts.FPDenormalMode.empty())
1740 FuncAttrs.addAttribute("denormal-fp-math", CodeGenOpts.FPDenormalMode);
1742 FuncAttrs.addAttribute("no-trapping-math",
1743 llvm::toStringRef(CodeGenOpts.NoTrappingMath));
1745 // Strict (compliant) code is the default, so only add this attribute to
1746 // indicate that we are trying to workaround a problem case.
1747 if (!CodeGenOpts.StrictFloatCastOverflow)
1748 FuncAttrs.addAttribute("strict-float-cast-overflow", "false");
1750 // TODO: Are these all needed?
1751 // unsafe/inf/nan/nsz are handled by instruction-level FastMathFlags.
1752 FuncAttrs.addAttribute("no-infs-fp-math",
1753 llvm::toStringRef(CodeGenOpts.NoInfsFPMath));
1754 FuncAttrs.addAttribute("no-nans-fp-math",
1755 llvm::toStringRef(CodeGenOpts.NoNaNsFPMath));
1756 FuncAttrs.addAttribute("unsafe-fp-math",
1757 llvm::toStringRef(CodeGenOpts.UnsafeFPMath));
1758 FuncAttrs.addAttribute("use-soft-float",
1759 llvm::toStringRef(CodeGenOpts.SoftFloat));
1760 FuncAttrs.addAttribute("stack-protector-buffer-size",
1761 llvm::utostr(CodeGenOpts.SSPBufferSize));
1762 FuncAttrs.addAttribute("no-signed-zeros-fp-math",
1763 llvm::toStringRef(CodeGenOpts.NoSignedZeros));
1764 FuncAttrs.addAttribute(
1765 "correctly-rounded-divide-sqrt-fp-math",
1766 llvm::toStringRef(CodeGenOpts.CorrectlyRoundedDivSqrt));
1768 if (getLangOpts().OpenCL)
1769 FuncAttrs.addAttribute("denorms-are-zero",
1770 llvm::toStringRef(CodeGenOpts.FlushDenorm));
1772 // TODO: Reciprocal estimate codegen options should apply to instructions?
1773 const std::vector<std::string> &Recips = CodeGenOpts.Reciprocals;
1774 if (!Recips.empty())
1775 FuncAttrs.addAttribute("reciprocal-estimates",
1776 llvm::join(Recips, ","));
1778 if (!CodeGenOpts.PreferVectorWidth.empty() &&
1779 CodeGenOpts.PreferVectorWidth != "none")
1780 FuncAttrs.addAttribute("prefer-vector-width",
1781 CodeGenOpts.PreferVectorWidth);
1783 if (CodeGenOpts.StackRealignment)
1784 FuncAttrs.addAttribute("stackrealign");
1785 if (CodeGenOpts.Backchain)
1786 FuncAttrs.addAttribute("backchain");
1789 if (getLangOpts().assumeFunctionsAreConvergent()) {
1790 // Conservatively, mark all functions and calls in CUDA and OpenCL as
1791 // convergent (meaning, they may call an intrinsically convergent op, such
1792 // as __syncthreads() / barrier(), and so can't have certain optimizations
1793 // applied around them). LLVM will remove this attribute where it safely
1795 FuncAttrs.addAttribute(llvm::Attribute::Convergent);
1798 if (getLangOpts().CUDA && getLangOpts().CUDAIsDevice) {
1799 // Exceptions aren't supported in CUDA device code.
1800 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1802 // Respect -fcuda-flush-denormals-to-zero.
1803 if (CodeGenOpts.FlushDenorm)
1804 FuncAttrs.addAttribute("nvptx-f32ftz", "true");
1808 void CodeGenModule::AddDefaultFnAttrs(llvm::Function &F) {
1809 llvm::AttrBuilder FuncAttrs;
1810 ConstructDefaultFnAttrList(F.getName(),
1811 F.hasFnAttribute(llvm::Attribute::OptimizeNone),
1812 /* AttrOnCallsite = */ false, FuncAttrs);
1813 F.addAttributes(llvm::AttributeList::FunctionIndex, FuncAttrs);
1816 void CodeGenModule::ConstructAttributeList(
1817 StringRef Name, const CGFunctionInfo &FI, CGCalleeInfo CalleeInfo,
1818 llvm::AttributeList &AttrList, unsigned &CallingConv, bool AttrOnCallSite) {
1819 llvm::AttrBuilder FuncAttrs;
1820 llvm::AttrBuilder RetAttrs;
1822 CallingConv = FI.getEffectiveCallingConvention();
1823 if (FI.isNoReturn())
1824 FuncAttrs.addAttribute(llvm::Attribute::NoReturn);
1826 // If we have information about the function prototype, we can learn
1827 // attributes from there.
1828 AddAttributesFromFunctionProtoType(getContext(), FuncAttrs,
1829 CalleeInfo.getCalleeFunctionProtoType());
1831 const Decl *TargetDecl = CalleeInfo.getCalleeDecl();
1833 bool HasOptnone = false;
1834 // FIXME: handle sseregparm someday...
1836 if (TargetDecl->hasAttr<ReturnsTwiceAttr>())
1837 FuncAttrs.addAttribute(llvm::Attribute::ReturnsTwice);
1838 if (TargetDecl->hasAttr<NoThrowAttr>())
1839 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1840 if (TargetDecl->hasAttr<NoReturnAttr>())
1841 FuncAttrs.addAttribute(llvm::Attribute::NoReturn);
1842 if (TargetDecl->hasAttr<ColdAttr>())
1843 FuncAttrs.addAttribute(llvm::Attribute::Cold);
1844 if (TargetDecl->hasAttr<NoDuplicateAttr>())
1845 FuncAttrs.addAttribute(llvm::Attribute::NoDuplicate);
1846 if (TargetDecl->hasAttr<ConvergentAttr>())
1847 FuncAttrs.addAttribute(llvm::Attribute::Convergent);
1849 if (const FunctionDecl *Fn = dyn_cast<FunctionDecl>(TargetDecl)) {
1850 AddAttributesFromFunctionProtoType(
1851 getContext(), FuncAttrs, Fn->getType()->getAs<FunctionProtoType>());
1852 // Don't use [[noreturn]] or _Noreturn for a call to a virtual function.
1853 // These attributes are not inherited by overloads.
1854 const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(Fn);
1855 if (Fn->isNoReturn() && !(AttrOnCallSite && MD && MD->isVirtual()))
1856 FuncAttrs.addAttribute(llvm::Attribute::NoReturn);
1859 // 'const', 'pure' and 'noalias' attributed functions are also nounwind.
1860 if (TargetDecl->hasAttr<ConstAttr>()) {
1861 FuncAttrs.addAttribute(llvm::Attribute::ReadNone);
1862 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1863 } else if (TargetDecl->hasAttr<PureAttr>()) {
1864 FuncAttrs.addAttribute(llvm::Attribute::ReadOnly);
1865 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1866 } else if (TargetDecl->hasAttr<NoAliasAttr>()) {
1867 FuncAttrs.addAttribute(llvm::Attribute::ArgMemOnly);
1868 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1870 if (TargetDecl->hasAttr<RestrictAttr>())
1871 RetAttrs.addAttribute(llvm::Attribute::NoAlias);
1872 if (TargetDecl->hasAttr<ReturnsNonNullAttr>() &&
1873 !CodeGenOpts.NullPointerIsValid)
1874 RetAttrs.addAttribute(llvm::Attribute::NonNull);
1875 if (TargetDecl->hasAttr<AnyX86NoCallerSavedRegistersAttr>())
1876 FuncAttrs.addAttribute("no_caller_saved_registers");
1877 if (TargetDecl->hasAttr<AnyX86NoCfCheckAttr>())
1878 FuncAttrs.addAttribute(llvm::Attribute::NoCfCheck);
1880 HasOptnone = TargetDecl->hasAttr<OptimizeNoneAttr>();
1881 if (auto *AllocSize = TargetDecl->getAttr<AllocSizeAttr>()) {
1882 Optional<unsigned> NumElemsParam;
1883 if (AllocSize->getNumElemsParam().isValid())
1884 NumElemsParam = AllocSize->getNumElemsParam().getLLVMIndex();
1885 FuncAttrs.addAllocSizeAttr(AllocSize->getElemSizeParam().getLLVMIndex(),
1890 ConstructDefaultFnAttrList(Name, HasOptnone, AttrOnCallSite, FuncAttrs);
1892 if (CodeGenOpts.EnableSegmentedStacks &&
1893 !(TargetDecl && TargetDecl->hasAttr<NoSplitStackAttr>()))
1894 FuncAttrs.addAttribute("split-stack");
1896 // Add NonLazyBind attribute to function declarations when -fno-plt
1898 if (TargetDecl && CodeGenOpts.NoPLT) {
1899 if (auto *Fn = dyn_cast<FunctionDecl>(TargetDecl)) {
1900 if (!Fn->isDefined() && !AttrOnCallSite) {
1901 FuncAttrs.addAttribute(llvm::Attribute::NonLazyBind);
1906 if (TargetDecl && TargetDecl->hasAttr<OpenCLKernelAttr>()) {
1907 if (getLangOpts().OpenCLVersion <= 120) {
1908 // OpenCL v1.2 Work groups are always uniform
1909 FuncAttrs.addAttribute("uniform-work-group-size", "true");
1911 // OpenCL v2.0 Work groups may be whether uniform or not.
1912 // '-cl-uniform-work-group-size' compile option gets a hint
1913 // to the compiler that the global work-size be a multiple of
1914 // the work-group size specified to clEnqueueNDRangeKernel
1915 // (i.e. work groups are uniform).
1916 FuncAttrs.addAttribute("uniform-work-group-size",
1917 llvm::toStringRef(CodeGenOpts.UniformWGSize));
1921 if (!AttrOnCallSite) {
1922 bool DisableTailCalls = false;
1924 if (CodeGenOpts.DisableTailCalls)
1925 DisableTailCalls = true;
1926 else if (TargetDecl) {
1927 if (TargetDecl->hasAttr<DisableTailCallsAttr>() ||
1928 TargetDecl->hasAttr<AnyX86InterruptAttr>())
1929 DisableTailCalls = true;
1930 else if (CodeGenOpts.NoEscapingBlockTailCalls) {
1931 if (const auto *BD = dyn_cast<BlockDecl>(TargetDecl))
1932 if (!BD->doesNotEscape())
1933 DisableTailCalls = true;
1937 FuncAttrs.addAttribute("disable-tail-calls",
1938 llvm::toStringRef(DisableTailCalls));
1939 GetCPUAndFeaturesAttributes(TargetDecl, FuncAttrs);
1942 ClangToLLVMArgMapping IRFunctionArgs(getContext(), FI);
1944 QualType RetTy = FI.getReturnType();
1945 const ABIArgInfo &RetAI = FI.getReturnInfo();
1946 switch (RetAI.getKind()) {
1947 case ABIArgInfo::Extend:
1948 if (RetAI.isSignExt())
1949 RetAttrs.addAttribute(llvm::Attribute::SExt);
1951 RetAttrs.addAttribute(llvm::Attribute::ZExt);
1953 case ABIArgInfo::Direct:
1954 if (RetAI.getInReg())
1955 RetAttrs.addAttribute(llvm::Attribute::InReg);
1957 case ABIArgInfo::Ignore:
1960 case ABIArgInfo::InAlloca:
1961 case ABIArgInfo::Indirect: {
1962 // inalloca and sret disable readnone and readonly
1963 FuncAttrs.removeAttribute(llvm::Attribute::ReadOnly)
1964 .removeAttribute(llvm::Attribute::ReadNone);
1968 case ABIArgInfo::CoerceAndExpand:
1971 case ABIArgInfo::Expand:
1972 llvm_unreachable("Invalid ABI kind for return argument");
1975 if (const auto *RefTy = RetTy->getAs<ReferenceType>()) {
1976 QualType PTy = RefTy->getPointeeType();
1977 if (!PTy->isIncompleteType() && PTy->isConstantSizeType())
1978 RetAttrs.addDereferenceableAttr(getContext().getTypeSizeInChars(PTy)
1980 else if (getContext().getTargetAddressSpace(PTy) == 0 &&
1981 !CodeGenOpts.NullPointerIsValid)
1982 RetAttrs.addAttribute(llvm::Attribute::NonNull);
1985 bool hasUsedSRet = false;
1986 SmallVector<llvm::AttributeSet, 4> ArgAttrs(IRFunctionArgs.totalIRArgs());
1988 // Attach attributes to sret.
1989 if (IRFunctionArgs.hasSRetArg()) {
1990 llvm::AttrBuilder SRETAttrs;
1991 if (!RetAI.getSuppressSRet())
1992 SRETAttrs.addAttribute(llvm::Attribute::StructRet);
1994 if (RetAI.getInReg())
1995 SRETAttrs.addAttribute(llvm::Attribute::InReg);
1996 ArgAttrs[IRFunctionArgs.getSRetArgNo()] =
1997 llvm::AttributeSet::get(getLLVMContext(), SRETAttrs);
2000 // Attach attributes to inalloca argument.
2001 if (IRFunctionArgs.hasInallocaArg()) {
2002 llvm::AttrBuilder Attrs;
2003 Attrs.addAttribute(llvm::Attribute::InAlloca);
2004 ArgAttrs[IRFunctionArgs.getInallocaArgNo()] =
2005 llvm::AttributeSet::get(getLLVMContext(), Attrs);
2009 for (CGFunctionInfo::const_arg_iterator I = FI.arg_begin(),
2011 I != E; ++I, ++ArgNo) {
2012 QualType ParamType = I->type;
2013 const ABIArgInfo &AI = I->info;
2014 llvm::AttrBuilder Attrs;
2016 // Add attribute for padding argument, if necessary.
2017 if (IRFunctionArgs.hasPaddingArg(ArgNo)) {
2018 if (AI.getPaddingInReg()) {
2019 ArgAttrs[IRFunctionArgs.getPaddingArgNo(ArgNo)] =
2020 llvm::AttributeSet::get(
2022 llvm::AttrBuilder().addAttribute(llvm::Attribute::InReg));
2026 // 'restrict' -> 'noalias' is done in EmitFunctionProlog when we
2027 // have the corresponding parameter variable. It doesn't make
2028 // sense to do it here because parameters are so messed up.
2029 switch (AI.getKind()) {
2030 case ABIArgInfo::Extend:
2032 Attrs.addAttribute(llvm::Attribute::SExt);
2034 Attrs.addAttribute(llvm::Attribute::ZExt);
2036 case ABIArgInfo::Direct:
2037 if (ArgNo == 0 && FI.isChainCall())
2038 Attrs.addAttribute(llvm::Attribute::Nest);
2039 else if (AI.getInReg())
2040 Attrs.addAttribute(llvm::Attribute::InReg);
2043 case ABIArgInfo::Indirect: {
2045 Attrs.addAttribute(llvm::Attribute::InReg);
2047 if (AI.getIndirectByVal())
2048 Attrs.addAttribute(llvm::Attribute::ByVal);
2050 CharUnits Align = AI.getIndirectAlign();
2052 // In a byval argument, it is important that the required
2053 // alignment of the type is honored, as LLVM might be creating a
2054 // *new* stack object, and needs to know what alignment to give
2055 // it. (Sometimes it can deduce a sensible alignment on its own,
2056 // but not if clang decides it must emit a packed struct, or the
2057 // user specifies increased alignment requirements.)
2059 // This is different from indirect *not* byval, where the object
2060 // exists already, and the align attribute is purely
2062 assert(!Align.isZero());
2064 // For now, only add this when we have a byval argument.
2065 // TODO: be less lazy about updating test cases.
2066 if (AI.getIndirectByVal())
2067 Attrs.addAlignmentAttr(Align.getQuantity());
2069 // byval disables readnone and readonly.
2070 FuncAttrs.removeAttribute(llvm::Attribute::ReadOnly)
2071 .removeAttribute(llvm::Attribute::ReadNone);
2074 case ABIArgInfo::Ignore:
2075 case ABIArgInfo::Expand:
2076 case ABIArgInfo::CoerceAndExpand:
2079 case ABIArgInfo::InAlloca:
2080 // inalloca disables readnone and readonly.
2081 FuncAttrs.removeAttribute(llvm::Attribute::ReadOnly)
2082 .removeAttribute(llvm::Attribute::ReadNone);
2086 if (const auto *RefTy = ParamType->getAs<ReferenceType>()) {
2087 QualType PTy = RefTy->getPointeeType();
2088 if (!PTy->isIncompleteType() && PTy->isConstantSizeType())
2089 Attrs.addDereferenceableAttr(getContext().getTypeSizeInChars(PTy)
2091 else if (getContext().getTargetAddressSpace(PTy) == 0 &&
2092 !CodeGenOpts.NullPointerIsValid)
2093 Attrs.addAttribute(llvm::Attribute::NonNull);
2096 switch (FI.getExtParameterInfo(ArgNo).getABI()) {
2097 case ParameterABI::Ordinary:
2100 case ParameterABI::SwiftIndirectResult: {
2101 // Add 'sret' if we haven't already used it for something, but
2102 // only if the result is void.
2103 if (!hasUsedSRet && RetTy->isVoidType()) {
2104 Attrs.addAttribute(llvm::Attribute::StructRet);
2108 // Add 'noalias' in either case.
2109 Attrs.addAttribute(llvm::Attribute::NoAlias);
2111 // Add 'dereferenceable' and 'alignment'.
2112 auto PTy = ParamType->getPointeeType();
2113 if (!PTy->isIncompleteType() && PTy->isConstantSizeType()) {
2114 auto info = getContext().getTypeInfoInChars(PTy);
2115 Attrs.addDereferenceableAttr(info.first.getQuantity());
2116 Attrs.addAttribute(llvm::Attribute::getWithAlignment(getLLVMContext(),
2117 info.second.getQuantity()));
2122 case ParameterABI::SwiftErrorResult:
2123 Attrs.addAttribute(llvm::Attribute::SwiftError);
2126 case ParameterABI::SwiftContext:
2127 Attrs.addAttribute(llvm::Attribute::SwiftSelf);
2131 if (FI.getExtParameterInfo(ArgNo).isNoEscape())
2132 Attrs.addAttribute(llvm::Attribute::NoCapture);
2134 if (Attrs.hasAttributes()) {
2135 unsigned FirstIRArg, NumIRArgs;
2136 std::tie(FirstIRArg, NumIRArgs) = IRFunctionArgs.getIRArgs(ArgNo);
2137 for (unsigned i = 0; i < NumIRArgs; i++)
2138 ArgAttrs[FirstIRArg + i] =
2139 llvm::AttributeSet::get(getLLVMContext(), Attrs);
2142 assert(ArgNo == FI.arg_size());
2144 AttrList = llvm::AttributeList::get(
2145 getLLVMContext(), llvm::AttributeSet::get(getLLVMContext(), FuncAttrs),
2146 llvm::AttributeSet::get(getLLVMContext(), RetAttrs), ArgAttrs);
2149 /// An argument came in as a promoted argument; demote it back to its
2151 static llvm::Value *emitArgumentDemotion(CodeGenFunction &CGF,
2153 llvm::Value *value) {
2154 llvm::Type *varType = CGF.ConvertType(var->getType());
2156 // This can happen with promotions that actually don't change the
2157 // underlying type, like the enum promotions.
2158 if (value->getType() == varType) return value;
2160 assert((varType->isIntegerTy() || varType->isFloatingPointTy())
2161 && "unexpected promotion type");
2163 if (isa<llvm::IntegerType>(varType))
2164 return CGF.Builder.CreateTrunc(value, varType, "arg.unpromote");
2166 return CGF.Builder.CreateFPCast(value, varType, "arg.unpromote");
2169 /// Returns the attribute (either parameter attribute, or function
2170 /// attribute), which declares argument ArgNo to be non-null.
2171 static const NonNullAttr *getNonNullAttr(const Decl *FD, const ParmVarDecl *PVD,
2172 QualType ArgType, unsigned ArgNo) {
2173 // FIXME: __attribute__((nonnull)) can also be applied to:
2174 // - references to pointers, where the pointee is known to be
2175 // nonnull (apparently a Clang extension)
2176 // - transparent unions containing pointers
2177 // In the former case, LLVM IR cannot represent the constraint. In
2178 // the latter case, we have no guarantee that the transparent union
2179 // is in fact passed as a pointer.
2180 if (!ArgType->isAnyPointerType() && !ArgType->isBlockPointerType())
2182 // First, check attribute on parameter itself.
2184 if (auto ParmNNAttr = PVD->getAttr<NonNullAttr>())
2187 // Check function attributes.
2190 for (const auto *NNAttr : FD->specific_attrs<NonNullAttr>()) {
2191 if (NNAttr->isNonNull(ArgNo))
2198 struct CopyBackSwiftError final : EHScopeStack::Cleanup {
2201 CopyBackSwiftError(Address temp, Address arg) : Temp(temp), Arg(arg) {}
2202 void Emit(CodeGenFunction &CGF, Flags flags) override {
2203 llvm::Value *errorValue = CGF.Builder.CreateLoad(Temp);
2204 CGF.Builder.CreateStore(errorValue, Arg);
2209 void CodeGenFunction::EmitFunctionProlog(const CGFunctionInfo &FI,
2211 const FunctionArgList &Args) {
2212 if (CurCodeDecl && CurCodeDecl->hasAttr<NakedAttr>())
2213 // Naked functions don't have prologues.
2216 // If this is an implicit-return-zero function, go ahead and
2217 // initialize the return value. TODO: it might be nice to have
2218 // a more general mechanism for this that didn't require synthesized
2219 // return statements.
2220 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(CurCodeDecl)) {
2221 if (FD->hasImplicitReturnZero()) {
2222 QualType RetTy = FD->getReturnType().getUnqualifiedType();
2223 llvm::Type* LLVMTy = CGM.getTypes().ConvertType(RetTy);
2224 llvm::Constant* Zero = llvm::Constant::getNullValue(LLVMTy);
2225 Builder.CreateStore(Zero, ReturnValue);
2229 // FIXME: We no longer need the types from FunctionArgList; lift up and
2232 ClangToLLVMArgMapping IRFunctionArgs(CGM.getContext(), FI);
2233 // Flattened function arguments.
2234 SmallVector<llvm::Value *, 16> FnArgs;
2235 FnArgs.reserve(IRFunctionArgs.totalIRArgs());
2236 for (auto &Arg : Fn->args()) {
2237 FnArgs.push_back(&Arg);
2239 assert(FnArgs.size() == IRFunctionArgs.totalIRArgs());
2241 // If we're using inalloca, all the memory arguments are GEPs off of the last
2242 // parameter, which is a pointer to the complete memory area.
2243 Address ArgStruct = Address::invalid();
2244 const llvm::StructLayout *ArgStructLayout = nullptr;
2245 if (IRFunctionArgs.hasInallocaArg()) {
2246 ArgStructLayout = CGM.getDataLayout().getStructLayout(FI.getArgStruct());
2247 ArgStruct = Address(FnArgs[IRFunctionArgs.getInallocaArgNo()],
2248 FI.getArgStructAlignment());
2250 assert(ArgStruct.getType() == FI.getArgStruct()->getPointerTo());
2253 // Name the struct return parameter.
2254 if (IRFunctionArgs.hasSRetArg()) {
2255 auto AI = cast<llvm::Argument>(FnArgs[IRFunctionArgs.getSRetArgNo()]);
2256 AI->setName("agg.result");
2257 AI->addAttr(llvm::Attribute::NoAlias);
2260 // Track if we received the parameter as a pointer (indirect, byval, or
2261 // inalloca). If already have a pointer, EmitParmDecl doesn't need to copy it
2262 // into a local alloca for us.
2263 SmallVector<ParamValue, 16> ArgVals;
2264 ArgVals.reserve(Args.size());
2266 // Create a pointer value for every parameter declaration. This usually
2267 // entails copying one or more LLVM IR arguments into an alloca. Don't push
2268 // any cleanups or do anything that might unwind. We do that separately, so
2269 // we can push the cleanups in the correct order for the ABI.
2270 assert(FI.arg_size() == Args.size() &&
2271 "Mismatch between function signature & arguments.");
2273 CGFunctionInfo::const_arg_iterator info_it = FI.arg_begin();
2274 for (FunctionArgList::const_iterator i = Args.begin(), e = Args.end();
2275 i != e; ++i, ++info_it, ++ArgNo) {
2276 const VarDecl *Arg = *i;
2277 const ABIArgInfo &ArgI = info_it->info;
2280 isa<ParmVarDecl>(Arg) && cast<ParmVarDecl>(Arg)->isKNRPromoted();
2281 // We are converting from ABIArgInfo type to VarDecl type directly, unless
2282 // the parameter is promoted. In this case we convert to
2283 // CGFunctionInfo::ArgInfo type with subsequent argument demotion.
2284 QualType Ty = isPromoted ? info_it->type : Arg->getType();
2285 assert(hasScalarEvaluationKind(Ty) ==
2286 hasScalarEvaluationKind(Arg->getType()));
2288 unsigned FirstIRArg, NumIRArgs;
2289 std::tie(FirstIRArg, NumIRArgs) = IRFunctionArgs.getIRArgs(ArgNo);
2291 switch (ArgI.getKind()) {
2292 case ABIArgInfo::InAlloca: {
2293 assert(NumIRArgs == 0);
2294 auto FieldIndex = ArgI.getInAllocaFieldIndex();
2295 CharUnits FieldOffset =
2296 CharUnits::fromQuantity(ArgStructLayout->getElementOffset(FieldIndex));
2297 Address V = Builder.CreateStructGEP(ArgStruct, FieldIndex, FieldOffset,
2299 ArgVals.push_back(ParamValue::forIndirect(V));
2303 case ABIArgInfo::Indirect: {
2304 assert(NumIRArgs == 1);
2305 Address ParamAddr = Address(FnArgs[FirstIRArg], ArgI.getIndirectAlign());
2307 if (!hasScalarEvaluationKind(Ty)) {
2308 // Aggregates and complex variables are accessed by reference. All we
2309 // need to do is realign the value, if requested.
2310 Address V = ParamAddr;
2311 if (ArgI.getIndirectRealign()) {
2312 Address AlignedTemp = CreateMemTemp(Ty, "coerce");
2314 // Copy from the incoming argument pointer to the temporary with the
2315 // appropriate alignment.
2317 // FIXME: We should have a common utility for generating an aggregate
2319 CharUnits Size = getContext().getTypeSizeInChars(Ty);
2320 auto SizeVal = llvm::ConstantInt::get(IntPtrTy, Size.getQuantity());
2321 Address Dst = Builder.CreateBitCast(AlignedTemp, Int8PtrTy);
2322 Address Src = Builder.CreateBitCast(ParamAddr, Int8PtrTy);
2323 Builder.CreateMemCpy(Dst, Src, SizeVal, false);
2326 ArgVals.push_back(ParamValue::forIndirect(V));
2328 // Load scalar value from indirect argument.
2330 EmitLoadOfScalar(ParamAddr, false, Ty, Arg->getLocStart());
2333 V = emitArgumentDemotion(*this, Arg, V);
2334 ArgVals.push_back(ParamValue::forDirect(V));
2339 case ABIArgInfo::Extend:
2340 case ABIArgInfo::Direct: {
2342 // If we have the trivial case, handle it with no muss and fuss.
2343 if (!isa<llvm::StructType>(ArgI.getCoerceToType()) &&
2344 ArgI.getCoerceToType() == ConvertType(Ty) &&
2345 ArgI.getDirectOffset() == 0) {
2346 assert(NumIRArgs == 1);
2347 llvm::Value *V = FnArgs[FirstIRArg];
2348 auto AI = cast<llvm::Argument>(V);
2350 if (const ParmVarDecl *PVD = dyn_cast<ParmVarDecl>(Arg)) {
2351 if (getNonNullAttr(CurCodeDecl, PVD, PVD->getType(),
2352 PVD->getFunctionScopeIndex()) &&
2353 !CGM.getCodeGenOpts().NullPointerIsValid)
2354 AI->addAttr(llvm::Attribute::NonNull);
2356 QualType OTy = PVD->getOriginalType();
2357 if (const auto *ArrTy =
2358 getContext().getAsConstantArrayType(OTy)) {
2359 // A C99 array parameter declaration with the static keyword also
2360 // indicates dereferenceability, and if the size is constant we can
2361 // use the dereferenceable attribute (which requires the size in
2363 if (ArrTy->getSizeModifier() == ArrayType::Static) {
2364 QualType ETy = ArrTy->getElementType();
2365 uint64_t ArrSize = ArrTy->getSize().getZExtValue();
2366 if (!ETy->isIncompleteType() && ETy->isConstantSizeType() &&
2368 llvm::AttrBuilder Attrs;
2369 Attrs.addDereferenceableAttr(
2370 getContext().getTypeSizeInChars(ETy).getQuantity()*ArrSize);
2371 AI->addAttrs(Attrs);
2372 } else if (getContext().getTargetAddressSpace(ETy) == 0 &&
2373 !CGM.getCodeGenOpts().NullPointerIsValid) {
2374 AI->addAttr(llvm::Attribute::NonNull);
2377 } else if (const auto *ArrTy =
2378 getContext().getAsVariableArrayType(OTy)) {
2379 // For C99 VLAs with the static keyword, we don't know the size so
2380 // we can't use the dereferenceable attribute, but in addrspace(0)
2381 // we know that it must be nonnull.
2382 if (ArrTy->getSizeModifier() == VariableArrayType::Static &&
2383 !getContext().getTargetAddressSpace(ArrTy->getElementType()) &&
2384 !CGM.getCodeGenOpts().NullPointerIsValid)
2385 AI->addAttr(llvm::Attribute::NonNull);
2388 const auto *AVAttr = PVD->getAttr<AlignValueAttr>();
2390 if (const auto *TOTy = dyn_cast<TypedefType>(OTy))
2391 AVAttr = TOTy->getDecl()->getAttr<AlignValueAttr>();
2393 llvm::Value *AlignmentValue =
2394 EmitScalarExpr(AVAttr->getAlignment());
2395 llvm::ConstantInt *AlignmentCI =
2396 cast<llvm::ConstantInt>(AlignmentValue);
2397 unsigned Alignment = std::min((unsigned)AlignmentCI->getZExtValue(),
2398 +llvm::Value::MaximumAlignment);
2399 AI->addAttrs(llvm::AttrBuilder().addAlignmentAttr(Alignment));
2403 if (Arg->getType().isRestrictQualified())
2404 AI->addAttr(llvm::Attribute::NoAlias);
2406 // LLVM expects swifterror parameters to be used in very restricted
2407 // ways. Copy the value into a less-restricted temporary.
2408 if (FI.getExtParameterInfo(ArgNo).getABI()
2409 == ParameterABI::SwiftErrorResult) {
2410 QualType pointeeTy = Ty->getPointeeType();
2411 assert(pointeeTy->isPointerType());
2413 CreateMemTemp(pointeeTy, getPointerAlign(), "swifterror.temp");
2414 Address arg = Address(V, getContext().getTypeAlignInChars(pointeeTy));
2415 llvm::Value *incomingErrorValue = Builder.CreateLoad(arg);
2416 Builder.CreateStore(incomingErrorValue, temp);
2417 V = temp.getPointer();
2419 // Push a cleanup to copy the value back at the end of the function.
2420 // The convention does not guarantee that the value will be written
2421 // back if the function exits with an unwind exception.
2422 EHStack.pushCleanup<CopyBackSwiftError>(NormalCleanup, temp, arg);
2425 // Ensure the argument is the correct type.
2426 if (V->getType() != ArgI.getCoerceToType())
2427 V = Builder.CreateBitCast(V, ArgI.getCoerceToType());
2430 V = emitArgumentDemotion(*this, Arg, V);
2432 // Because of merging of function types from multiple decls it is
2433 // possible for the type of an argument to not match the corresponding
2434 // type in the function type. Since we are codegening the callee
2435 // in here, add a cast to the argument type.
2436 llvm::Type *LTy = ConvertType(Arg->getType());
2437 if (V->getType() != LTy)
2438 V = Builder.CreateBitCast(V, LTy);
2440 ArgVals.push_back(ParamValue::forDirect(V));
2444 Address Alloca = CreateMemTemp(Ty, getContext().getDeclAlign(Arg),
2447 // Pointer to store into.
2448 Address Ptr = emitAddressAtOffset(*this, Alloca, ArgI);
2450 // Fast-isel and the optimizer generally like scalar values better than
2451 // FCAs, so we flatten them if this is safe to do for this argument.
2452 llvm::StructType *STy = dyn_cast<llvm::StructType>(ArgI.getCoerceToType());
2453 if (ArgI.isDirect() && ArgI.getCanBeFlattened() && STy &&
2454 STy->getNumElements() > 1) {
2455 auto SrcLayout = CGM.getDataLayout().getStructLayout(STy);
2456 uint64_t SrcSize = CGM.getDataLayout().getTypeAllocSize(STy);
2457 llvm::Type *DstTy = Ptr.getElementType();
2458 uint64_t DstSize = CGM.getDataLayout().getTypeAllocSize(DstTy);
2460 Address AddrToStoreInto = Address::invalid();
2461 if (SrcSize <= DstSize) {
2462 AddrToStoreInto = Builder.CreateElementBitCast(Ptr, STy);
2465 CreateTempAlloca(STy, Alloca.getAlignment(), "coerce");
2468 assert(STy->getNumElements() == NumIRArgs);
2469 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
2470 auto AI = FnArgs[FirstIRArg + i];
2471 AI->setName(Arg->getName() + ".coerce" + Twine(i));
2472 auto Offset = CharUnits::fromQuantity(SrcLayout->getElementOffset(i));
2474 Builder.CreateStructGEP(AddrToStoreInto, i, Offset);
2475 Builder.CreateStore(AI, EltPtr);
2478 if (SrcSize > DstSize) {
2479 Builder.CreateMemCpy(Ptr, AddrToStoreInto, DstSize);
2483 // Simple case, just do a coerced store of the argument into the alloca.
2484 assert(NumIRArgs == 1);
2485 auto AI = FnArgs[FirstIRArg];
2486 AI->setName(Arg->getName() + ".coerce");
2487 CreateCoercedStore(AI, Ptr, /*DestIsVolatile=*/false, *this);
2490 // Match to what EmitParmDecl is expecting for this type.
2491 if (CodeGenFunction::hasScalarEvaluationKind(Ty)) {
2493 EmitLoadOfScalar(Alloca, false, Ty, Arg->getLocStart());
2495 V = emitArgumentDemotion(*this, Arg, V);
2496 ArgVals.push_back(ParamValue::forDirect(V));
2498 ArgVals.push_back(ParamValue::forIndirect(Alloca));
2503 case ABIArgInfo::CoerceAndExpand: {
2504 // Reconstruct into a temporary.
2505 Address alloca = CreateMemTemp(Ty, getContext().getDeclAlign(Arg));
2506 ArgVals.push_back(ParamValue::forIndirect(alloca));
2508 auto coercionType = ArgI.getCoerceAndExpandType();
2509 alloca = Builder.CreateElementBitCast(alloca, coercionType);
2510 auto layout = CGM.getDataLayout().getStructLayout(coercionType);
2512 unsigned argIndex = FirstIRArg;
2513 for (unsigned i = 0, e = coercionType->getNumElements(); i != e; ++i) {
2514 llvm::Type *eltType = coercionType->getElementType(i);
2515 if (ABIArgInfo::isPaddingForCoerceAndExpand(eltType))
2518 auto eltAddr = Builder.CreateStructGEP(alloca, i, layout);
2519 auto elt = FnArgs[argIndex++];
2520 Builder.CreateStore(elt, eltAddr);
2522 assert(argIndex == FirstIRArg + NumIRArgs);
2526 case ABIArgInfo::Expand: {
2527 // If this structure was expanded into multiple arguments then
2528 // we need to create a temporary and reconstruct it from the
2530 Address Alloca = CreateMemTemp(Ty, getContext().getDeclAlign(Arg));
2531 LValue LV = MakeAddrLValue(Alloca, Ty);
2532 ArgVals.push_back(ParamValue::forIndirect(Alloca));
2534 auto FnArgIter = FnArgs.begin() + FirstIRArg;
2535 ExpandTypeFromArgs(Ty, LV, FnArgIter);
2536 assert(FnArgIter == FnArgs.begin() + FirstIRArg + NumIRArgs);
2537 for (unsigned i = 0, e = NumIRArgs; i != e; ++i) {
2538 auto AI = FnArgs[FirstIRArg + i];
2539 AI->setName(Arg->getName() + "." + Twine(i));
2544 case ABIArgInfo::Ignore:
2545 assert(NumIRArgs == 0);
2546 // Initialize the local variable appropriately.
2547 if (!hasScalarEvaluationKind(Ty)) {
2548 ArgVals.push_back(ParamValue::forIndirect(CreateMemTemp(Ty)));
2550 llvm::Value *U = llvm::UndefValue::get(ConvertType(Arg->getType()));
2551 ArgVals.push_back(ParamValue::forDirect(U));
2557 if (getTarget().getCXXABI().areArgsDestroyedLeftToRightInCallee()) {
2558 for (int I = Args.size() - 1; I >= 0; --I)
2559 EmitParmDecl(*Args[I], ArgVals[I], I + 1);
2561 for (unsigned I = 0, E = Args.size(); I != E; ++I)
2562 EmitParmDecl(*Args[I], ArgVals[I], I + 1);
2566 static void eraseUnusedBitCasts(llvm::Instruction *insn) {
2567 while (insn->use_empty()) {
2568 llvm::BitCastInst *bitcast = dyn_cast<llvm::BitCastInst>(insn);
2569 if (!bitcast) return;
2571 // This is "safe" because we would have used a ConstantExpr otherwise.
2572 insn = cast<llvm::Instruction>(bitcast->getOperand(0));
2573 bitcast->eraseFromParent();
2577 /// Try to emit a fused autorelease of a return result.
2578 static llvm::Value *tryEmitFusedAutoreleaseOfResult(CodeGenFunction &CGF,
2579 llvm::Value *result) {
2580 // We must be immediately followed the cast.
2581 llvm::BasicBlock *BB = CGF.Builder.GetInsertBlock();
2582 if (BB->empty()) return nullptr;
2583 if (&BB->back() != result) return nullptr;
2585 llvm::Type *resultType = result->getType();
2587 // result is in a BasicBlock and is therefore an Instruction.
2588 llvm::Instruction *generator = cast<llvm::Instruction>(result);
2590 SmallVector<llvm::Instruction *, 4> InstsToKill;
2593 // %generator = bitcast %type1* %generator2 to %type2*
2594 while (llvm::BitCastInst *bitcast = dyn_cast<llvm::BitCastInst>(generator)) {
2595 // We would have emitted this as a constant if the operand weren't
2597 generator = cast<llvm::Instruction>(bitcast->getOperand(0));
2599 // Require the generator to be immediately followed by the cast.
2600 if (generator->getNextNode() != bitcast)
2603 InstsToKill.push_back(bitcast);
2607 // %generator = call i8* @objc_retain(i8* %originalResult)
2609 // %generator = call i8* @objc_retainAutoreleasedReturnValue(i8* %originalResult)
2610 llvm::CallInst *call = dyn_cast<llvm::CallInst>(generator);
2611 if (!call) return nullptr;
2613 bool doRetainAutorelease;
2615 if (call->getCalledValue() == CGF.CGM.getObjCEntrypoints().objc_retain) {
2616 doRetainAutorelease = true;
2617 } else if (call->getCalledValue() == CGF.CGM.getObjCEntrypoints()
2618 .objc_retainAutoreleasedReturnValue) {
2619 doRetainAutorelease = false;
2621 // If we emitted an assembly marker for this call (and the
2622 // ARCEntrypoints field should have been set if so), go looking
2623 // for that call. If we can't find it, we can't do this
2624 // optimization. But it should always be the immediately previous
2625 // instruction, unless we needed bitcasts around the call.
2626 if (CGF.CGM.getObjCEntrypoints().retainAutoreleasedReturnValueMarker) {
2627 llvm::Instruction *prev = call->getPrevNode();
2629 if (isa<llvm::BitCastInst>(prev)) {
2630 prev = prev->getPrevNode();
2633 assert(isa<llvm::CallInst>(prev));
2634 assert(cast<llvm::CallInst>(prev)->getCalledValue() ==
2635 CGF.CGM.getObjCEntrypoints().retainAutoreleasedReturnValueMarker);
2636 InstsToKill.push_back(prev);
2642 result = call->getArgOperand(0);
2643 InstsToKill.push_back(call);
2645 // Keep killing bitcasts, for sanity. Note that we no longer care
2646 // about precise ordering as long as there's exactly one use.
2647 while (llvm::BitCastInst *bitcast = dyn_cast<llvm::BitCastInst>(result)) {
2648 if (!bitcast->hasOneUse()) break;
2649 InstsToKill.push_back(bitcast);
2650 result = bitcast->getOperand(0);
2653 // Delete all the unnecessary instructions, from latest to earliest.
2654 for (auto *I : InstsToKill)
2655 I->eraseFromParent();
2657 // Do the fused retain/autorelease if we were asked to.
2658 if (doRetainAutorelease)
2659 result = CGF.EmitARCRetainAutoreleaseReturnValue(result);
2661 // Cast back to the result type.
2662 return CGF.Builder.CreateBitCast(result, resultType);
2665 /// If this is a +1 of the value of an immutable 'self', remove it.
2666 static llvm::Value *tryRemoveRetainOfSelf(CodeGenFunction &CGF,
2667 llvm::Value *result) {
2668 // This is only applicable to a method with an immutable 'self'.
2669 const ObjCMethodDecl *method =
2670 dyn_cast_or_null<ObjCMethodDecl>(CGF.CurCodeDecl);
2671 if (!method) return nullptr;
2672 const VarDecl *self = method->getSelfDecl();
2673 if (!self->getType().isConstQualified()) return nullptr;
2675 // Look for a retain call.
2676 llvm::CallInst *retainCall =
2677 dyn_cast<llvm::CallInst>(result->stripPointerCasts());
2679 retainCall->getCalledValue() != CGF.CGM.getObjCEntrypoints().objc_retain)
2682 // Look for an ordinary load of 'self'.
2683 llvm::Value *retainedValue = retainCall->getArgOperand(0);
2684 llvm::LoadInst *load =
2685 dyn_cast<llvm::LoadInst>(retainedValue->stripPointerCasts());
2686 if (!load || load->isAtomic() || load->isVolatile() ||
2687 load->getPointerOperand() != CGF.GetAddrOfLocalVar(self).getPointer())
2690 // Okay! Burn it all down. This relies for correctness on the
2691 // assumption that the retain is emitted as part of the return and
2692 // that thereafter everything is used "linearly".
2693 llvm::Type *resultType = result->getType();
2694 eraseUnusedBitCasts(cast<llvm::Instruction>(result));
2695 assert(retainCall->use_empty());
2696 retainCall->eraseFromParent();
2697 eraseUnusedBitCasts(cast<llvm::Instruction>(retainedValue));
2699 return CGF.Builder.CreateBitCast(load, resultType);
2702 /// Emit an ARC autorelease of the result of a function.
2704 /// \return the value to actually return from the function
2705 static llvm::Value *emitAutoreleaseOfResult(CodeGenFunction &CGF,
2706 llvm::Value *result) {
2707 // If we're returning 'self', kill the initial retain. This is a
2708 // heuristic attempt to "encourage correctness" in the really unfortunate
2709 // case where we have a return of self during a dealloc and we desperately
2710 // need to avoid the possible autorelease.
2711 if (llvm::Value *self = tryRemoveRetainOfSelf(CGF, result))
2714 // At -O0, try to emit a fused retain/autorelease.
2715 if (CGF.shouldUseFusedARCCalls())
2716 if (llvm::Value *fused = tryEmitFusedAutoreleaseOfResult(CGF, result))
2719 return CGF.EmitARCAutoreleaseReturnValue(result);
2722 /// Heuristically search for a dominating store to the return-value slot.
2723 static llvm::StoreInst *findDominatingStoreToReturnValue(CodeGenFunction &CGF) {
2724 // Check if a User is a store which pointerOperand is the ReturnValue.
2725 // We are looking for stores to the ReturnValue, not for stores of the
2726 // ReturnValue to some other location.
2727 auto GetStoreIfValid = [&CGF](llvm::User *U) -> llvm::StoreInst * {
2728 auto *SI = dyn_cast<llvm::StoreInst>(U);
2729 if (!SI || SI->getPointerOperand() != CGF.ReturnValue.getPointer())
2731 // These aren't actually possible for non-coerced returns, and we
2732 // only care about non-coerced returns on this code path.
2733 assert(!SI->isAtomic() && !SI->isVolatile());
2736 // If there are multiple uses of the return-value slot, just check
2737 // for something immediately preceding the IP. Sometimes this can
2738 // happen with how we generate implicit-returns; it can also happen
2739 // with noreturn cleanups.
2740 if (!CGF.ReturnValue.getPointer()->hasOneUse()) {
2741 llvm::BasicBlock *IP = CGF.Builder.GetInsertBlock();
2742 if (IP->empty()) return nullptr;
2743 llvm::Instruction *I = &IP->back();
2745 // Skip lifetime markers
2746 for (llvm::BasicBlock::reverse_iterator II = IP->rbegin(),
2749 if (llvm::IntrinsicInst *Intrinsic =
2750 dyn_cast<llvm::IntrinsicInst>(&*II)) {
2751 if (Intrinsic->getIntrinsicID() == llvm::Intrinsic::lifetime_end) {
2752 const llvm::Value *CastAddr = Intrinsic->getArgOperand(1);
2756 if (isa<llvm::BitCastInst>(&*II) && (CastAddr == &*II))
2764 return GetStoreIfValid(I);
2767 llvm::StoreInst *store =
2768 GetStoreIfValid(CGF.ReturnValue.getPointer()->user_back());
2769 if (!store) return nullptr;
2771 // Now do a first-and-dirty dominance check: just walk up the
2772 // single-predecessors chain from the current insertion point.
2773 llvm::BasicBlock *StoreBB = store->getParent();
2774 llvm::BasicBlock *IP = CGF.Builder.GetInsertBlock();
2775 while (IP != StoreBB) {
2776 if (!(IP = IP->getSinglePredecessor()))
2780 // Okay, the store's basic block dominates the insertion point; we
2781 // can do our thing.
2785 void CodeGenFunction::EmitFunctionEpilog(const CGFunctionInfo &FI,
2787 SourceLocation EndLoc) {
2788 if (FI.isNoReturn()) {
2789 // Noreturn functions don't return.
2790 EmitUnreachable(EndLoc);
2794 if (CurCodeDecl && CurCodeDecl->hasAttr<NakedAttr>()) {
2795 // Naked functions don't have epilogues.
2796 Builder.CreateUnreachable();
2800 // Functions with no result always return void.
2801 if (!ReturnValue.isValid()) {
2802 Builder.CreateRetVoid();
2806 llvm::DebugLoc RetDbgLoc;
2807 llvm::Value *RV = nullptr;
2808 QualType RetTy = FI.getReturnType();
2809 const ABIArgInfo &RetAI = FI.getReturnInfo();
2811 switch (RetAI.getKind()) {
2812 case ABIArgInfo::InAlloca:
2813 // Aggregrates get evaluated directly into the destination. Sometimes we
2814 // need to return the sret value in a register, though.
2815 assert(hasAggregateEvaluationKind(RetTy));
2816 if (RetAI.getInAllocaSRet()) {
2817 llvm::Function::arg_iterator EI = CurFn->arg_end();
2819 llvm::Value *ArgStruct = &*EI;
2820 llvm::Value *SRet = Builder.CreateStructGEP(
2821 nullptr, ArgStruct, RetAI.getInAllocaFieldIndex());
2822 RV = Builder.CreateAlignedLoad(SRet, getPointerAlign(), "sret");
2826 case ABIArgInfo::Indirect: {
2827 auto AI = CurFn->arg_begin();
2828 if (RetAI.isSRetAfterThis())
2830 switch (getEvaluationKind(RetTy)) {
2833 EmitLoadOfComplex(MakeAddrLValue(ReturnValue, RetTy), EndLoc);
2834 EmitStoreOfComplex(RT, MakeNaturalAlignAddrLValue(&*AI, RetTy),
2839 // Do nothing; aggregrates get evaluated directly into the destination.
2842 EmitStoreOfScalar(Builder.CreateLoad(ReturnValue),
2843 MakeNaturalAlignAddrLValue(&*AI, RetTy),
2850 case ABIArgInfo::Extend:
2851 case ABIArgInfo::Direct:
2852 if (RetAI.getCoerceToType() == ConvertType(RetTy) &&
2853 RetAI.getDirectOffset() == 0) {
2854 // The internal return value temp always will have pointer-to-return-type
2855 // type, just do a load.
2857 // If there is a dominating store to ReturnValue, we can elide
2858 // the load, zap the store, and usually zap the alloca.
2859 if (llvm::StoreInst *SI =
2860 findDominatingStoreToReturnValue(*this)) {
2861 // Reuse the debug location from the store unless there is
2862 // cleanup code to be emitted between the store and return
2864 if (EmitRetDbgLoc && !AutoreleaseResult)
2865 RetDbgLoc = SI->getDebugLoc();
2866 // Get the stored value and nuke the now-dead store.
2867 RV = SI->getValueOperand();
2868 SI->eraseFromParent();
2870 // If that was the only use of the return value, nuke it as well now.
2871 auto returnValueInst = ReturnValue.getPointer();
2872 if (returnValueInst->use_empty()) {
2873 if (auto alloca = dyn_cast<llvm::AllocaInst>(returnValueInst)) {
2874 alloca->eraseFromParent();
2875 ReturnValue = Address::invalid();
2879 // Otherwise, we have to do a simple load.
2881 RV = Builder.CreateLoad(ReturnValue);
2884 // If the value is offset in memory, apply the offset now.
2885 Address V = emitAddressAtOffset(*this, ReturnValue, RetAI);
2887 RV = CreateCoercedLoad(V, RetAI.getCoerceToType(), *this);
2890 // In ARC, end functions that return a retainable type with a call
2891 // to objc_autoreleaseReturnValue.
2892 if (AutoreleaseResult) {
2894 // Type::isObjCRetainabletype has to be called on a QualType that hasn't
2895 // been stripped of the typedefs, so we cannot use RetTy here. Get the
2896 // original return type of FunctionDecl, CurCodeDecl, and BlockDecl from
2897 // CurCodeDecl or BlockInfo.
2900 if (auto *FD = dyn_cast<FunctionDecl>(CurCodeDecl))
2901 RT = FD->getReturnType();
2902 else if (auto *MD = dyn_cast<ObjCMethodDecl>(CurCodeDecl))
2903 RT = MD->getReturnType();
2904 else if (isa<BlockDecl>(CurCodeDecl))
2905 RT = BlockInfo->BlockExpression->getFunctionType()->getReturnType();
2907 llvm_unreachable("Unexpected function/method type");
2909 assert(getLangOpts().ObjCAutoRefCount &&
2910 !FI.isReturnsRetained() &&
2911 RT->isObjCRetainableType());
2913 RV = emitAutoreleaseOfResult(*this, RV);
2918 case ABIArgInfo::Ignore:
2921 case ABIArgInfo::CoerceAndExpand: {
2922 auto coercionType = RetAI.getCoerceAndExpandType();
2923 auto layout = CGM.getDataLayout().getStructLayout(coercionType);
2925 // Load all of the coerced elements out into results.
2926 llvm::SmallVector<llvm::Value*, 4> results;
2927 Address addr = Builder.CreateElementBitCast(ReturnValue, coercionType);
2928 for (unsigned i = 0, e = coercionType->getNumElements(); i != e; ++i) {
2929 auto coercedEltType = coercionType->getElementType(i);
2930 if (ABIArgInfo::isPaddingForCoerceAndExpand(coercedEltType))
2933 auto eltAddr = Builder.CreateStructGEP(addr, i, layout);
2934 auto elt = Builder.CreateLoad(eltAddr);
2935 results.push_back(elt);
2938 // If we have one result, it's the single direct result type.
2939 if (results.size() == 1) {
2942 // Otherwise, we need to make a first-class aggregate.
2944 // Construct a return type that lacks padding elements.
2945 llvm::Type *returnType = RetAI.getUnpaddedCoerceAndExpandType();
2947 RV = llvm::UndefValue::get(returnType);
2948 for (unsigned i = 0, e = results.size(); i != e; ++i) {
2949 RV = Builder.CreateInsertValue(RV, results[i], i);
2955 case ABIArgInfo::Expand:
2956 llvm_unreachable("Invalid ABI kind for return argument");
2959 llvm::Instruction *Ret;
2961 EmitReturnValueCheck(RV);
2962 Ret = Builder.CreateRet(RV);
2964 Ret = Builder.CreateRetVoid();
2968 Ret->setDebugLoc(std::move(RetDbgLoc));
2971 void CodeGenFunction::EmitReturnValueCheck(llvm::Value *RV) {
2972 // A current decl may not be available when emitting vtable thunks.
2976 ReturnsNonNullAttr *RetNNAttr = nullptr;
2977 if (SanOpts.has(SanitizerKind::ReturnsNonnullAttribute))
2978 RetNNAttr = CurCodeDecl->getAttr<ReturnsNonNullAttr>();
2980 if (!RetNNAttr && !requiresReturnValueNullabilityCheck())
2983 // Prefer the returns_nonnull attribute if it's present.
2984 SourceLocation AttrLoc;
2985 SanitizerMask CheckKind;
2986 SanitizerHandler Handler;
2988 assert(!requiresReturnValueNullabilityCheck() &&
2989 "Cannot check nullability and the nonnull attribute");
2990 AttrLoc = RetNNAttr->getLocation();
2991 CheckKind = SanitizerKind::ReturnsNonnullAttribute;
2992 Handler = SanitizerHandler::NonnullReturn;
2994 if (auto *DD = dyn_cast<DeclaratorDecl>(CurCodeDecl))
2995 if (auto *TSI = DD->getTypeSourceInfo())
2996 if (auto FTL = TSI->getTypeLoc().castAs<FunctionTypeLoc>())
2997 AttrLoc = FTL.getReturnLoc().findNullabilityLoc();
2998 CheckKind = SanitizerKind::NullabilityReturn;
2999 Handler = SanitizerHandler::NullabilityReturn;
3002 SanitizerScope SanScope(this);
3004 // Make sure the "return" source location is valid. If we're checking a
3005 // nullability annotation, make sure the preconditions for the check are met.
3006 llvm::BasicBlock *Check = createBasicBlock("nullcheck");
3007 llvm::BasicBlock *NoCheck = createBasicBlock("no.nullcheck");
3008 llvm::Value *SLocPtr = Builder.CreateLoad(ReturnLocation, "return.sloc.load");
3009 llvm::Value *CanNullCheck = Builder.CreateIsNotNull(SLocPtr);
3010 if (requiresReturnValueNullabilityCheck())
3012 Builder.CreateAnd(CanNullCheck, RetValNullabilityPrecondition);
3013 Builder.CreateCondBr(CanNullCheck, Check, NoCheck);
3016 // Now do the null check.
3017 llvm::Value *Cond = Builder.CreateIsNotNull(RV);
3018 llvm::Constant *StaticData[] = {EmitCheckSourceLocation(AttrLoc)};
3019 llvm::Value *DynamicData[] = {SLocPtr};
3020 EmitCheck(std::make_pair(Cond, CheckKind), Handler, StaticData, DynamicData);
3025 // The return location should not be used after the check has been emitted.
3026 ReturnLocation = Address::invalid();
3030 static bool isInAllocaArgument(CGCXXABI &ABI, QualType type) {
3031 const CXXRecordDecl *RD = type->getAsCXXRecordDecl();
3032 return RD && ABI.getRecordArgABI(RD) == CGCXXABI::RAA_DirectInMemory;
3035 static AggValueSlot createPlaceholderSlot(CodeGenFunction &CGF,
3037 // FIXME: Generate IR in one pass, rather than going back and fixing up these
3039 llvm::Type *IRTy = CGF.ConvertTypeForMem(Ty);
3040 llvm::Type *IRPtrTy = IRTy->getPointerTo();
3041 llvm::Value *Placeholder = llvm::UndefValue::get(IRPtrTy->getPointerTo());
3043 // FIXME: When we generate this IR in one pass, we shouldn't need
3044 // this win32-specific alignment hack.
3045 CharUnits Align = CharUnits::fromQuantity(4);
3046 Placeholder = CGF.Builder.CreateAlignedLoad(IRPtrTy, Placeholder, Align);
3048 return AggValueSlot::forAddr(Address(Placeholder, Align),
3050 AggValueSlot::IsNotDestructed,
3051 AggValueSlot::DoesNotNeedGCBarriers,
3052 AggValueSlot::IsNotAliased,
3053 AggValueSlot::DoesNotOverlap);
3056 void CodeGenFunction::EmitDelegateCallArg(CallArgList &args,
3057 const VarDecl *param,
3058 SourceLocation loc) {
3059 // StartFunction converted the ABI-lowered parameter(s) into a
3060 // local alloca. We need to turn that into an r-value suitable
3062 Address local = GetAddrOfLocalVar(param);
3064 QualType type = param->getType();
3066 assert(!isInAllocaArgument(CGM.getCXXABI(), type) &&
3067 "cannot emit delegate call arguments for inalloca arguments!");
3069 // GetAddrOfLocalVar returns a pointer-to-pointer for references,
3070 // but the argument needs to be the original pointer.
3071 if (type->isReferenceType()) {
3072 args.add(RValue::get(Builder.CreateLoad(local)), type);
3074 // In ARC, move out of consumed arguments so that the release cleanup
3075 // entered by StartFunction doesn't cause an over-release. This isn't
3076 // optimal -O0 code generation, but it should get cleaned up when
3077 // optimization is enabled. This also assumes that delegate calls are
3078 // performed exactly once for a set of arguments, but that should be safe.
3079 } else if (getLangOpts().ObjCAutoRefCount &&
3080 param->hasAttr<NSConsumedAttr>() &&
3081 type->isObjCRetainableType()) {
3082 llvm::Value *ptr = Builder.CreateLoad(local);
3084 llvm::ConstantPointerNull::get(cast<llvm::PointerType>(ptr->getType()));
3085 Builder.CreateStore(null, local);
3086 args.add(RValue::get(ptr), type);
3088 // For the most part, we just need to load the alloca, except that
3089 // aggregate r-values are actually pointers to temporaries.
3091 args.add(convertTempToRValue(local, type, loc), type);
3094 // Deactivate the cleanup for the callee-destructed param that was pushed.
3095 if (hasAggregateEvaluationKind(type) && !CurFuncIsThunk &&
3096 type->getAs<RecordType>()->getDecl()->isParamDestroyedInCallee() &&
3097 type.isDestructedType()) {
3098 EHScopeStack::stable_iterator cleanup =
3099 CalleeDestructedParamCleanups.lookup(cast<ParmVarDecl>(param));
3100 assert(cleanup.isValid() &&
3101 "cleanup for callee-destructed param not recorded");
3102 // This unreachable is a temporary marker which will be removed later.
3103 llvm::Instruction *isActive = Builder.CreateUnreachable();
3104 args.addArgCleanupDeactivation(cleanup, isActive);
3108 static bool isProvablyNull(llvm::Value *addr) {
3109 return isa<llvm::ConstantPointerNull>(addr);
3112 /// Emit the actual writing-back of a writeback.
3113 static void emitWriteback(CodeGenFunction &CGF,
3114 const CallArgList::Writeback &writeback) {
3115 const LValue &srcLV = writeback.Source;
3116 Address srcAddr = srcLV.getAddress();
3117 assert(!isProvablyNull(srcAddr.getPointer()) &&
3118 "shouldn't have writeback for provably null argument");
3120 llvm::BasicBlock *contBB = nullptr;
3122 // If the argument wasn't provably non-null, we need to null check
3123 // before doing the store.
3124 bool provablyNonNull = llvm::isKnownNonZero(srcAddr.getPointer(),
3125 CGF.CGM.getDataLayout());
3126 if (!provablyNonNull) {
3127 llvm::BasicBlock *writebackBB = CGF.createBasicBlock("icr.writeback");
3128 contBB = CGF.createBasicBlock("icr.done");
3130 llvm::Value *isNull =
3131 CGF.Builder.CreateIsNull(srcAddr.getPointer(), "icr.isnull");
3132 CGF.Builder.CreateCondBr(isNull, contBB, writebackBB);
3133 CGF.EmitBlock(writebackBB);
3136 // Load the value to writeback.
3137 llvm::Value *value = CGF.Builder.CreateLoad(writeback.Temporary);
3139 // Cast it back, in case we're writing an id to a Foo* or something.
3140 value = CGF.Builder.CreateBitCast(value, srcAddr.getElementType(),
3141 "icr.writeback-cast");
3143 // Perform the writeback.
3145 // If we have a "to use" value, it's something we need to emit a use
3146 // of. This has to be carefully threaded in: if it's done after the
3147 // release it's potentially undefined behavior (and the optimizer
3148 // will ignore it), and if it happens before the retain then the
3149 // optimizer could move the release there.
3150 if (writeback.ToUse) {
3151 assert(srcLV.getObjCLifetime() == Qualifiers::OCL_Strong);
3153 // Retain the new value. No need to block-copy here: the block's
3154 // being passed up the stack.
3155 value = CGF.EmitARCRetainNonBlock(value);
3157 // Emit the intrinsic use here.
3158 CGF.EmitARCIntrinsicUse(writeback.ToUse);
3160 // Load the old value (primitively).
3161 llvm::Value *oldValue = CGF.EmitLoadOfScalar(srcLV, SourceLocation());
3163 // Put the new value in place (primitively).
3164 CGF.EmitStoreOfScalar(value, srcLV, /*init*/ false);
3166 // Release the old value.
3167 CGF.EmitARCRelease(oldValue, srcLV.isARCPreciseLifetime());
3169 // Otherwise, we can just do a normal lvalue store.
3171 CGF.EmitStoreThroughLValue(RValue::get(value), srcLV);
3174 // Jump to the continuation block.
3175 if (!provablyNonNull)
3176 CGF.EmitBlock(contBB);
3179 static void emitWritebacks(CodeGenFunction &CGF,
3180 const CallArgList &args) {
3181 for (const auto &I : args.writebacks())
3182 emitWriteback(CGF, I);
3185 static void deactivateArgCleanupsBeforeCall(CodeGenFunction &CGF,
3186 const CallArgList &CallArgs) {
3187 ArrayRef<CallArgList::CallArgCleanup> Cleanups =
3188 CallArgs.getCleanupsToDeactivate();
3189 // Iterate in reverse to increase the likelihood of popping the cleanup.
3190 for (const auto &I : llvm::reverse(Cleanups)) {
3191 CGF.DeactivateCleanupBlock(I.Cleanup, I.IsActiveIP);
3192 I.IsActiveIP->eraseFromParent();
3196 static const Expr *maybeGetUnaryAddrOfOperand(const Expr *E) {
3197 if (const UnaryOperator *uop = dyn_cast<UnaryOperator>(E->IgnoreParens()))
3198 if (uop->getOpcode() == UO_AddrOf)
3199 return uop->getSubExpr();
3203 /// Emit an argument that's being passed call-by-writeback. That is,
3204 /// we are passing the address of an __autoreleased temporary; it
3205 /// might be copy-initialized with the current value of the given
3206 /// address, but it will definitely be copied out of after the call.
3207 static void emitWritebackArg(CodeGenFunction &CGF, CallArgList &args,
3208 const ObjCIndirectCopyRestoreExpr *CRE) {
3211 // Make an optimistic effort to emit the address as an l-value.
3212 // This can fail if the argument expression is more complicated.
3213 if (const Expr *lvExpr = maybeGetUnaryAddrOfOperand(CRE->getSubExpr())) {
3214 srcLV = CGF.EmitLValue(lvExpr);
3216 // Otherwise, just emit it as a scalar.
3218 Address srcAddr = CGF.EmitPointerWithAlignment(CRE->getSubExpr());
3220 QualType srcAddrType =
3221 CRE->getSubExpr()->getType()->castAs<PointerType>()->getPointeeType();
3222 srcLV = CGF.MakeAddrLValue(srcAddr, srcAddrType);
3224 Address srcAddr = srcLV.getAddress();
3226 // The dest and src types don't necessarily match in LLVM terms
3227 // because of the crazy ObjC compatibility rules.
3229 llvm::PointerType *destType =
3230 cast<llvm::PointerType>(CGF.ConvertType(CRE->getType()));
3232 // If the address is a constant null, just pass the appropriate null.
3233 if (isProvablyNull(srcAddr.getPointer())) {
3234 args.add(RValue::get(llvm::ConstantPointerNull::get(destType)),
3239 // Create the temporary.
3240 Address temp = CGF.CreateTempAlloca(destType->getElementType(),
3241 CGF.getPointerAlign(),
3243 // Loading an l-value can introduce a cleanup if the l-value is __weak,
3244 // and that cleanup will be conditional if we can't prove that the l-value
3245 // isn't null, so we need to register a dominating point so that the cleanups
3246 // system will make valid IR.
3247 CodeGenFunction::ConditionalEvaluation condEval(CGF);
3249 // Zero-initialize it if we're not doing a copy-initialization.
3250 bool shouldCopy = CRE->shouldCopy();
3253 llvm::ConstantPointerNull::get(
3254 cast<llvm::PointerType>(destType->getElementType()));
3255 CGF.Builder.CreateStore(null, temp);
3258 llvm::BasicBlock *contBB = nullptr;
3259 llvm::BasicBlock *originBB = nullptr;
3261 // If the address is *not* known to be non-null, we need to switch.
3262 llvm::Value *finalArgument;
3264 bool provablyNonNull = llvm::isKnownNonZero(srcAddr.getPointer(),
3265 CGF.CGM.getDataLayout());
3266 if (provablyNonNull) {
3267 finalArgument = temp.getPointer();
3269 llvm::Value *isNull =
3270 CGF.Builder.CreateIsNull(srcAddr.getPointer(), "icr.isnull");
3272 finalArgument = CGF.Builder.CreateSelect(isNull,
3273 llvm::ConstantPointerNull::get(destType),
3274 temp.getPointer(), "icr.argument");
3276 // If we need to copy, then the load has to be conditional, which
3277 // means we need control flow.
3279 originBB = CGF.Builder.GetInsertBlock();
3280 contBB = CGF.createBasicBlock("icr.cont");
3281 llvm::BasicBlock *copyBB = CGF.createBasicBlock("icr.copy");
3282 CGF.Builder.CreateCondBr(isNull, contBB, copyBB);
3283 CGF.EmitBlock(copyBB);
3284 condEval.begin(CGF);
3288 llvm::Value *valueToUse = nullptr;
3290 // Perform a copy if necessary.
3292 RValue srcRV = CGF.EmitLoadOfLValue(srcLV, SourceLocation());
3293 assert(srcRV.isScalar());
3295 llvm::Value *src = srcRV.getScalarVal();
3296 src = CGF.Builder.CreateBitCast(src, destType->getElementType(),
3299 // Use an ordinary store, not a store-to-lvalue.
3300 CGF.Builder.CreateStore(src, temp);
3302 // If optimization is enabled, and the value was held in a
3303 // __strong variable, we need to tell the optimizer that this
3304 // value has to stay alive until we're doing the store back.
3305 // This is because the temporary is effectively unretained,
3306 // and so otherwise we can violate the high-level semantics.
3307 if (CGF.CGM.getCodeGenOpts().OptimizationLevel != 0 &&
3308 srcLV.getObjCLifetime() == Qualifiers::OCL_Strong) {
3313 // Finish the control flow if we needed it.
3314 if (shouldCopy && !provablyNonNull) {
3315 llvm::BasicBlock *copyBB = CGF.Builder.GetInsertBlock();
3316 CGF.EmitBlock(contBB);
3318 // Make a phi for the value to intrinsically use.
3320 llvm::PHINode *phiToUse = CGF.Builder.CreatePHI(valueToUse->getType(), 2,
3322 phiToUse->addIncoming(valueToUse, copyBB);
3323 phiToUse->addIncoming(llvm::UndefValue::get(valueToUse->getType()),
3325 valueToUse = phiToUse;
3331 args.addWriteback(srcLV, temp, valueToUse);
3332 args.add(RValue::get(finalArgument), CRE->getType());
3335 void CallArgList::allocateArgumentMemory(CodeGenFunction &CGF) {
3339 llvm::Function *F = CGF.CGM.getIntrinsic(llvm::Intrinsic::stacksave);
3340 StackBase = CGF.Builder.CreateCall(F, {}, "inalloca.save");
3343 void CallArgList::freeArgumentMemory(CodeGenFunction &CGF) const {
3345 // Restore the stack after the call.
3346 llvm::Value *F = CGF.CGM.getIntrinsic(llvm::Intrinsic::stackrestore);
3347 CGF.Builder.CreateCall(F, StackBase);
3351 void CodeGenFunction::EmitNonNullArgCheck(RValue RV, QualType ArgType,
3352 SourceLocation ArgLoc,
3355 if (!AC.getDecl() || !(SanOpts.has(SanitizerKind::NonnullAttribute) ||
3356 SanOpts.has(SanitizerKind::NullabilityArg)))
3359 // The param decl may be missing in a variadic function.
3360 auto PVD = ParmNum < AC.getNumParams() ? AC.getParamDecl(ParmNum) : nullptr;
3361 unsigned ArgNo = PVD ? PVD->getFunctionScopeIndex() : ParmNum;
3363 // Prefer the nonnull attribute if it's present.
3364 const NonNullAttr *NNAttr = nullptr;
3365 if (SanOpts.has(SanitizerKind::NonnullAttribute))
3366 NNAttr = getNonNullAttr(AC.getDecl(), PVD, ArgType, ArgNo);
3368 bool CanCheckNullability = false;
3369 if (SanOpts.has(SanitizerKind::NullabilityArg) && !NNAttr && PVD) {
3370 auto Nullability = PVD->getType()->getNullability(getContext());
3371 CanCheckNullability = Nullability &&
3372 *Nullability == NullabilityKind::NonNull &&
3373 PVD->getTypeSourceInfo();
3376 if (!NNAttr && !CanCheckNullability)
3379 SourceLocation AttrLoc;
3380 SanitizerMask CheckKind;
3381 SanitizerHandler Handler;
3383 AttrLoc = NNAttr->getLocation();
3384 CheckKind = SanitizerKind::NonnullAttribute;
3385 Handler = SanitizerHandler::NonnullArg;
3387 AttrLoc = PVD->getTypeSourceInfo()->getTypeLoc().findNullabilityLoc();
3388 CheckKind = SanitizerKind::NullabilityArg;
3389 Handler = SanitizerHandler::NullabilityArg;
3392 SanitizerScope SanScope(this);
3393 assert(RV.isScalar());
3394 llvm::Value *V = RV.getScalarVal();
3396 Builder.CreateICmpNE(V, llvm::Constant::getNullValue(V->getType()));
3397 llvm::Constant *StaticData[] = {
3398 EmitCheckSourceLocation(ArgLoc), EmitCheckSourceLocation(AttrLoc),
3399 llvm::ConstantInt::get(Int32Ty, ArgNo + 1),
3401 EmitCheck(std::make_pair(Cond, CheckKind), Handler, StaticData, None);
3404 void CodeGenFunction::EmitCallArgs(
3405 CallArgList &Args, ArrayRef<QualType> ArgTypes,
3406 llvm::iterator_range<CallExpr::const_arg_iterator> ArgRange,
3407 AbstractCallee AC, unsigned ParamsToSkip, EvaluationOrder Order) {
3408 assert((int)ArgTypes.size() == (ArgRange.end() - ArgRange.begin()));
3410 // We *have* to evaluate arguments from right to left in the MS C++ ABI,
3411 // because arguments are destroyed left to right in the callee. As a special
3412 // case, there are certain language constructs that require left-to-right
3413 // evaluation, and in those cases we consider the evaluation order requirement
3414 // to trump the "destruction order is reverse construction order" guarantee.
3416 CGM.getTarget().getCXXABI().areArgsDestroyedLeftToRightInCallee()
3417 ? Order == EvaluationOrder::ForceLeftToRight
3418 : Order != EvaluationOrder::ForceRightToLeft;
3420 auto MaybeEmitImplicitObjectSize = [&](unsigned I, const Expr *Arg,
3421 RValue EmittedArg) {
3422 if (!AC.hasFunctionDecl() || I >= AC.getNumParams())
3424 auto *PS = AC.getParamDecl(I)->getAttr<PassObjectSizeAttr>();
3428 const auto &Context = getContext();
3429 auto SizeTy = Context.getSizeType();
3430 auto T = Builder.getIntNTy(Context.getTypeSize(SizeTy));
3431 assert(EmittedArg.getScalarVal() && "We emitted nothing for the arg?");
3432 llvm::Value *V = evaluateOrEmitBuiltinObjectSize(Arg, PS->getType(), T,
3433 EmittedArg.getScalarVal());
3434 Args.add(RValue::get(V), SizeTy);
3435 // If we're emitting args in reverse, be sure to do so with
3436 // pass_object_size, as well.
3438 std::swap(Args.back(), *(&Args.back() - 1));
3441 // Insert a stack save if we're going to need any inalloca args.
3442 bool HasInAllocaArgs = false;
3443 if (CGM.getTarget().getCXXABI().isMicrosoft()) {
3444 for (ArrayRef<QualType>::iterator I = ArgTypes.begin(), E = ArgTypes.end();
3445 I != E && !HasInAllocaArgs; ++I)
3446 HasInAllocaArgs = isInAllocaArgument(CGM.getCXXABI(), *I);
3447 if (HasInAllocaArgs) {
3448 assert(getTarget().getTriple().getArch() == llvm::Triple::x86);
3449 Args.allocateArgumentMemory(*this);
3453 // Evaluate each argument in the appropriate order.
3454 size_t CallArgsStart = Args.size();
3455 for (unsigned I = 0, E = ArgTypes.size(); I != E; ++I) {
3456 unsigned Idx = LeftToRight ? I : E - I - 1;
3457 CallExpr::const_arg_iterator Arg = ArgRange.begin() + Idx;
3458 unsigned InitialArgSize = Args.size();
3459 // If *Arg is an ObjCIndirectCopyRestoreExpr, check that either the types of
3460 // the argument and parameter match or the objc method is parameterized.
3461 assert((!isa<ObjCIndirectCopyRestoreExpr>(*Arg) ||
3462 getContext().hasSameUnqualifiedType((*Arg)->getType(),
3464 (isa<ObjCMethodDecl>(AC.getDecl()) &&
3465 isObjCMethodWithTypeParams(cast<ObjCMethodDecl>(AC.getDecl())))) &&
3466 "Argument and parameter types don't match");
3467 EmitCallArg(Args, *Arg, ArgTypes[Idx]);
3468 // In particular, we depend on it being the last arg in Args, and the
3469 // objectsize bits depend on there only being one arg if !LeftToRight.
3470 assert(InitialArgSize + 1 == Args.size() &&
3471 "The code below depends on only adding one arg per EmitCallArg");
3472 (void)InitialArgSize;
3473 // Since pointer argument are never emitted as LValue, it is safe to emit
3474 // non-null argument check for r-value only.
3475 if (!Args.back().hasLValue()) {
3476 RValue RVArg = Args.back().getKnownRValue();
3477 EmitNonNullArgCheck(RVArg, ArgTypes[Idx], (*Arg)->getExprLoc(), AC,
3478 ParamsToSkip + Idx);
3479 // @llvm.objectsize should never have side-effects and shouldn't need
3480 // destruction/cleanups, so we can safely "emit" it after its arg,
3481 // regardless of right-to-leftness
3482 MaybeEmitImplicitObjectSize(Idx, *Arg, RVArg);
3487 // Un-reverse the arguments we just evaluated so they match up with the LLVM
3489 std::reverse(Args.begin() + CallArgsStart, Args.end());
3495 struct DestroyUnpassedArg final : EHScopeStack::Cleanup {
3496 DestroyUnpassedArg(Address Addr, QualType Ty)
3497 : Addr(Addr), Ty(Ty) {}
3502 void Emit(CodeGenFunction &CGF, Flags flags) override {
3503 QualType::DestructionKind DtorKind = Ty.isDestructedType();
3504 if (DtorKind == QualType::DK_cxx_destructor) {
3505 const CXXDestructorDecl *Dtor = Ty->getAsCXXRecordDecl()->getDestructor();
3506 assert(!Dtor->isTrivial());
3507 CGF.EmitCXXDestructorCall(Dtor, Dtor_Complete, /*for vbase*/ false,
3508 /*Delegating=*/false, Addr);
3510 CGF.callCStructDestructor(CGF.MakeAddrLValue(Addr, Ty));
3515 struct DisableDebugLocationUpdates {
3516 CodeGenFunction &CGF;
3517 bool disabledDebugInfo;
3518 DisableDebugLocationUpdates(CodeGenFunction &CGF, const Expr *E) : CGF(CGF) {
3519 if ((disabledDebugInfo = isa<CXXDefaultArgExpr>(E) && CGF.getDebugInfo()))
3520 CGF.disableDebugInfo();
3522 ~DisableDebugLocationUpdates() {
3523 if (disabledDebugInfo)
3524 CGF.enableDebugInfo();
3528 } // end anonymous namespace
3530 RValue CallArg::getRValue(CodeGenFunction &CGF) const {
3533 LValue Copy = CGF.MakeAddrLValue(CGF.CreateMemTemp(Ty), Ty);
3534 CGF.EmitAggregateCopy(Copy, LV, Ty, AggValueSlot::DoesNotOverlap,
3537 return RValue::getAggregate(Copy.getAddress());
3540 void CallArg::copyInto(CodeGenFunction &CGF, Address Addr) const {
3541 LValue Dst = CGF.MakeAddrLValue(Addr, Ty);
3542 if (!HasLV && RV.isScalar())
3543 CGF.EmitStoreOfScalar(RV.getScalarVal(), Dst, /*init=*/true);
3544 else if (!HasLV && RV.isComplex())
3545 CGF.EmitStoreOfComplex(RV.getComplexVal(), Dst, /*init=*/true);
3547 auto Addr = HasLV ? LV.getAddress() : RV.getAggregateAddress();
3548 LValue SrcLV = CGF.MakeAddrLValue(Addr, Ty);
3549 // We assume that call args are never copied into subobjects.
3550 CGF.EmitAggregateCopy(Dst, SrcLV, Ty, AggValueSlot::DoesNotOverlap,
3551 HasLV ? LV.isVolatileQualified()
3552 : RV.isVolatileQualified());
3557 void CodeGenFunction::EmitCallArg(CallArgList &args, const Expr *E,
3559 DisableDebugLocationUpdates Dis(*this, E);
3560 if (const ObjCIndirectCopyRestoreExpr *CRE
3561 = dyn_cast<ObjCIndirectCopyRestoreExpr>(E)) {
3562 assert(getLangOpts().ObjCAutoRefCount);
3563 return emitWritebackArg(*this, args, CRE);
3566 assert(type->isReferenceType() == E->isGLValue() &&
3567 "reference binding to unmaterialized r-value!");
3569 if (E->isGLValue()) {
3570 assert(E->getObjectKind() == OK_Ordinary);
3571 return args.add(EmitReferenceBindingToExpr(E), type);
3574 bool HasAggregateEvalKind = hasAggregateEvaluationKind(type);
3576 // In the Microsoft C++ ABI, aggregate arguments are destructed by the callee.
3577 // However, we still have to push an EH-only cleanup in case we unwind before
3578 // we make it to the call.
3579 if (HasAggregateEvalKind &&
3580 type->getAs<RecordType>()->getDecl()->isParamDestroyedInCallee()) {
3581 // If we're using inalloca, use the argument memory. Otherwise, use a
3584 if (args.isUsingInAlloca())
3585 Slot = createPlaceholderSlot(*this, type);
3587 Slot = CreateAggTemp(type, "agg.tmp");
3589 bool DestroyedInCallee = true, NeedsEHCleanup = true;
3590 if (const auto *RD = type->getAsCXXRecordDecl())
3591 DestroyedInCallee = RD->hasNonTrivialDestructor();
3593 NeedsEHCleanup = needsEHCleanup(type.isDestructedType());
3595 if (DestroyedInCallee)
3596 Slot.setExternallyDestructed();
3598 EmitAggExpr(E, Slot);
3599 RValue RV = Slot.asRValue();
3602 if (DestroyedInCallee && NeedsEHCleanup) {
3603 // Create a no-op GEP between the placeholder and the cleanup so we can
3604 // RAUW it successfully. It also serves as a marker of the first
3605 // instruction where the cleanup is active.
3606 pushFullExprCleanup<DestroyUnpassedArg>(EHCleanup, Slot.getAddress(),
3608 // This unreachable is a temporary marker which will be removed later.
3609 llvm::Instruction *IsActive = Builder.CreateUnreachable();
3610 args.addArgCleanupDeactivation(EHStack.getInnermostEHScope(), IsActive);
3615 if (HasAggregateEvalKind && isa<ImplicitCastExpr>(E) &&
3616 cast<CastExpr>(E)->getCastKind() == CK_LValueToRValue) {
3617 LValue L = EmitLValue(cast<CastExpr>(E)->getSubExpr());
3618 assert(L.isSimple());
3619 args.addUncopiedAggregate(L, type);
3623 args.add(EmitAnyExprToTemp(E), type);
3626 QualType CodeGenFunction::getVarArgType(const Expr *Arg) {
3627 // System headers on Windows define NULL to 0 instead of 0LL on Win64. MSVC
3628 // implicitly widens null pointer constants that are arguments to varargs
3629 // functions to pointer-sized ints.
3630 if (!getTarget().getTriple().isOSWindows())
3631 return Arg->getType();
3633 if (Arg->getType()->isIntegerType() &&
3634 getContext().getTypeSize(Arg->getType()) <
3635 getContext().getTargetInfo().getPointerWidth(0) &&
3636 Arg->isNullPointerConstant(getContext(),
3637 Expr::NPC_ValueDependentIsNotNull)) {
3638 return getContext().getIntPtrType();
3641 return Arg->getType();
3644 // In ObjC ARC mode with no ObjC ARC exception safety, tell the ARC
3645 // optimizer it can aggressively ignore unwind edges.
3647 CodeGenFunction::AddObjCARCExceptionMetadata(llvm::Instruction *Inst) {
3648 if (CGM.getCodeGenOpts().OptimizationLevel != 0 &&
3649 !CGM.getCodeGenOpts().ObjCAutoRefCountExceptions)
3650 Inst->setMetadata("clang.arc.no_objc_arc_exceptions",
3651 CGM.getNoObjCARCExceptionsMetadata());
3654 /// Emits a call to the given no-arguments nounwind runtime function.
3656 CodeGenFunction::EmitNounwindRuntimeCall(llvm::Value *callee,
3657 const llvm::Twine &name) {
3658 return EmitNounwindRuntimeCall(callee, None, name);
3661 /// Emits a call to the given nounwind runtime function.
3663 CodeGenFunction::EmitNounwindRuntimeCall(llvm::Value *callee,
3664 ArrayRef<llvm::Value*> args,
3665 const llvm::Twine &name) {
3666 llvm::CallInst *call = EmitRuntimeCall(callee, args, name);
3667 call->setDoesNotThrow();
3671 /// Emits a simple call (never an invoke) to the given no-arguments
3672 /// runtime function.
3674 CodeGenFunction::EmitRuntimeCall(llvm::Value *callee,
3675 const llvm::Twine &name) {
3676 return EmitRuntimeCall(callee, None, name);
3679 // Calls which may throw must have operand bundles indicating which funclet
3680 // they are nested within.
3681 SmallVector<llvm::OperandBundleDef, 1>
3682 CodeGenFunction::getBundlesForFunclet(llvm::Value *Callee) {
3683 SmallVector<llvm::OperandBundleDef, 1> BundleList;
3684 // There is no need for a funclet operand bundle if we aren't inside a
3686 if (!CurrentFuncletPad)
3689 // Skip intrinsics which cannot throw.
3690 auto *CalleeFn = dyn_cast<llvm::Function>(Callee->stripPointerCasts());
3691 if (CalleeFn && CalleeFn->isIntrinsic() && CalleeFn->doesNotThrow())
3694 BundleList.emplace_back("funclet", CurrentFuncletPad);
3698 /// Emits a simple call (never an invoke) to the given runtime function.
3700 CodeGenFunction::EmitRuntimeCall(llvm::Value *callee,
3701 ArrayRef<llvm::Value*> args,
3702 const llvm::Twine &name) {
3703 llvm::CallInst *call =
3704 Builder.CreateCall(callee, args, getBundlesForFunclet(callee), name);
3705 call->setCallingConv(getRuntimeCC());
3709 /// Emits a call or invoke to the given noreturn runtime function.
3710 void CodeGenFunction::EmitNoreturnRuntimeCallOrInvoke(llvm::Value *callee,
3711 ArrayRef<llvm::Value*> args) {
3712 SmallVector<llvm::OperandBundleDef, 1> BundleList =
3713 getBundlesForFunclet(callee);
3715 if (getInvokeDest()) {
3716 llvm::InvokeInst *invoke =
3717 Builder.CreateInvoke(callee,
3718 getUnreachableBlock(),
3722 invoke->setDoesNotReturn();
3723 invoke->setCallingConv(getRuntimeCC());
3725 llvm::CallInst *call = Builder.CreateCall(callee, args, BundleList);
3726 call->setDoesNotReturn();
3727 call->setCallingConv(getRuntimeCC());
3728 Builder.CreateUnreachable();
3732 /// Emits a call or invoke instruction to the given nullary runtime function.
3734 CodeGenFunction::EmitRuntimeCallOrInvoke(llvm::Value *callee,
3735 const Twine &name) {
3736 return EmitRuntimeCallOrInvoke(callee, None, name);
3739 /// Emits a call or invoke instruction to the given runtime function.
3741 CodeGenFunction::EmitRuntimeCallOrInvoke(llvm::Value *callee,
3742 ArrayRef<llvm::Value*> args,
3743 const Twine &name) {
3744 llvm::CallSite callSite = EmitCallOrInvoke(callee, args, name);
3745 callSite.setCallingConv(getRuntimeCC());
3749 /// Emits a call or invoke instruction to the given function, depending
3750 /// on the current state of the EH stack.
3752 CodeGenFunction::EmitCallOrInvoke(llvm::Value *Callee,
3753 ArrayRef<llvm::Value *> Args,
3754 const Twine &Name) {
3755 llvm::BasicBlock *InvokeDest = getInvokeDest();
3756 SmallVector<llvm::OperandBundleDef, 1> BundleList =
3757 getBundlesForFunclet(Callee);
3759 llvm::Instruction *Inst;
3761 Inst = Builder.CreateCall(Callee, Args, BundleList, Name);
3763 llvm::BasicBlock *ContBB = createBasicBlock("invoke.cont");
3764 Inst = Builder.CreateInvoke(Callee, ContBB, InvokeDest, Args, BundleList,
3769 // In ObjC ARC mode with no ObjC ARC exception safety, tell the ARC
3770 // optimizer it can aggressively ignore unwind edges.
3771 if (CGM.getLangOpts().ObjCAutoRefCount)
3772 AddObjCARCExceptionMetadata(Inst);
3774 return llvm::CallSite(Inst);
3777 void CodeGenFunction::deferPlaceholderReplacement(llvm::Instruction *Old,
3779 DeferredReplacements.push_back(std::make_pair(Old, New));
3782 RValue CodeGenFunction::EmitCall(const CGFunctionInfo &CallInfo,
3783 const CGCallee &Callee,
3784 ReturnValueSlot ReturnValue,
3785 const CallArgList &CallArgs,
3786 llvm::Instruction **callOrInvoke,
3787 SourceLocation Loc) {
3788 // FIXME: We no longer need the types from CallArgs; lift up and simplify.
3790 assert(Callee.isOrdinary() || Callee.isVirtual());
3792 // Handle struct-return functions by passing a pointer to the
3793 // location that we would like to return into.
3794 QualType RetTy = CallInfo.getReturnType();
3795 const ABIArgInfo &RetAI = CallInfo.getReturnInfo();
3797 llvm::FunctionType *IRFuncTy = Callee.getFunctionType();
3799 // 1. Set up the arguments.
3801 // If we're using inalloca, insert the allocation after the stack save.
3802 // FIXME: Do this earlier rather than hacking it in here!
3803 Address ArgMemory = Address::invalid();
3804 const llvm::StructLayout *ArgMemoryLayout = nullptr;
3805 if (llvm::StructType *ArgStruct = CallInfo.getArgStruct()) {
3806 const llvm::DataLayout &DL = CGM.getDataLayout();
3807 ArgMemoryLayout = DL.getStructLayout(ArgStruct);
3808 llvm::Instruction *IP = CallArgs.getStackBase();
3809 llvm::AllocaInst *AI;
3811 IP = IP->getNextNode();
3812 AI = new llvm::AllocaInst(ArgStruct, DL.getAllocaAddrSpace(),
3815 AI = CreateTempAlloca(ArgStruct, "argmem");
3817 auto Align = CallInfo.getArgStructAlignment();
3818 AI->setAlignment(Align.getQuantity());
3819 AI->setUsedWithInAlloca(true);
3820 assert(AI->isUsedWithInAlloca() && !AI->isStaticAlloca());
3821 ArgMemory = Address(AI, Align);
3824 // Helper function to drill into the inalloca allocation.
3825 auto createInAllocaStructGEP = [&](unsigned FieldIndex) -> Address {
3827 CharUnits::fromQuantity(ArgMemoryLayout->getElementOffset(FieldIndex));
3828 return Builder.CreateStructGEP(ArgMemory, FieldIndex, FieldOffset);
3831 ClangToLLVMArgMapping IRFunctionArgs(CGM.getContext(), CallInfo);
3832 SmallVector<llvm::Value *, 16> IRCallArgs(IRFunctionArgs.totalIRArgs());
3834 // If the call returns a temporary with struct return, create a temporary
3835 // alloca to hold the result, unless one is given to us.
3836 Address SRetPtr = Address::invalid();
3837 Address SRetAlloca = Address::invalid();
3838 llvm::Value *UnusedReturnSizePtr = nullptr;
3839 if (RetAI.isIndirect() || RetAI.isInAlloca() || RetAI.isCoerceAndExpand()) {
3840 if (!ReturnValue.isNull()) {
3841 SRetPtr = ReturnValue.getValue();
3843 SRetPtr = CreateMemTemp(RetTy, "tmp", &SRetAlloca);
3844 if (HaveInsertPoint() && ReturnValue.isUnused()) {
3846 CGM.getDataLayout().getTypeAllocSize(ConvertTypeForMem(RetTy));
3847 UnusedReturnSizePtr = EmitLifetimeStart(size, SRetAlloca.getPointer());
3850 if (IRFunctionArgs.hasSRetArg()) {
3851 IRCallArgs[IRFunctionArgs.getSRetArgNo()] = SRetPtr.getPointer();
3852 } else if (RetAI.isInAlloca()) {
3853 Address Addr = createInAllocaStructGEP(RetAI.getInAllocaFieldIndex());
3854 Builder.CreateStore(SRetPtr.getPointer(), Addr);
3858 Address swiftErrorTemp = Address::invalid();
3859 Address swiftErrorArg = Address::invalid();
3861 // Translate all of the arguments as necessary to match the IR lowering.
3862 assert(CallInfo.arg_size() == CallArgs.size() &&
3863 "Mismatch between function signature & arguments.");
3865 CGFunctionInfo::const_arg_iterator info_it = CallInfo.arg_begin();
3866 for (CallArgList::const_iterator I = CallArgs.begin(), E = CallArgs.end();
3867 I != E; ++I, ++info_it, ++ArgNo) {
3868 const ABIArgInfo &ArgInfo = info_it->info;
3870 // Insert a padding argument to ensure proper alignment.
3871 if (IRFunctionArgs.hasPaddingArg(ArgNo))
3872 IRCallArgs[IRFunctionArgs.getPaddingArgNo(ArgNo)] =
3873 llvm::UndefValue::get(ArgInfo.getPaddingType());
3875 unsigned FirstIRArg, NumIRArgs;
3876 std::tie(FirstIRArg, NumIRArgs) = IRFunctionArgs.getIRArgs(ArgNo);
3878 switch (ArgInfo.getKind()) {
3879 case ABIArgInfo::InAlloca: {
3880 assert(NumIRArgs == 0);
3881 assert(getTarget().getTriple().getArch() == llvm::Triple::x86);
3882 if (I->isAggregate()) {
3883 // Replace the placeholder with the appropriate argument slot GEP.
3884 Address Addr = I->hasLValue()
3885 ? I->getKnownLValue().getAddress()
3886 : I->getKnownRValue().getAggregateAddress();
3887 llvm::Instruction *Placeholder =
3888 cast<llvm::Instruction>(Addr.getPointer());
3889 CGBuilderTy::InsertPoint IP = Builder.saveIP();
3890 Builder.SetInsertPoint(Placeholder);
3891 Addr = createInAllocaStructGEP(ArgInfo.getInAllocaFieldIndex());
3892 Builder.restoreIP(IP);
3893 deferPlaceholderReplacement(Placeholder, Addr.getPointer());
3895 // Store the RValue into the argument struct.
3896 Address Addr = createInAllocaStructGEP(ArgInfo.getInAllocaFieldIndex());
3897 unsigned AS = Addr.getType()->getPointerAddressSpace();
3898 llvm::Type *MemType = ConvertTypeForMem(I->Ty)->getPointerTo(AS);
3899 // There are some cases where a trivial bitcast is not avoidable. The
3900 // definition of a type later in a translation unit may change it's type
3901 // from {}* to (%struct.foo*)*.
3902 if (Addr.getType() != MemType)
3903 Addr = Builder.CreateBitCast(Addr, MemType);
3904 I->copyInto(*this, Addr);
3909 case ABIArgInfo::Indirect: {
3910 assert(NumIRArgs == 1);
3911 if (!I->isAggregate()) {
3912 // Make a temporary alloca to pass the argument.
3913 Address Addr = CreateMemTempWithoutCast(
3914 I->Ty, ArgInfo.getIndirectAlign(), "indirect-arg-temp");
3915 IRCallArgs[FirstIRArg] = Addr.getPointer();
3917 I->copyInto(*this, Addr);
3919 // We want to avoid creating an unnecessary temporary+copy here;
3920 // however, we need one in three cases:
3921 // 1. If the argument is not byval, and we are required to copy the
3922 // source. (This case doesn't occur on any common architecture.)
3923 // 2. If the argument is byval, RV is not sufficiently aligned, and
3924 // we cannot force it to be sufficiently aligned.
3925 // 3. If the argument is byval, but RV is not located in default
3926 // or alloca address space.
3927 Address Addr = I->hasLValue()
3928 ? I->getKnownLValue().getAddress()
3929 : I->getKnownRValue().getAggregateAddress();
3930 llvm::Value *V = Addr.getPointer();
3931 CharUnits Align = ArgInfo.getIndirectAlign();
3932 const llvm::DataLayout *TD = &CGM.getDataLayout();
3934 assert((FirstIRArg >= IRFuncTy->getNumParams() ||
3935 IRFuncTy->getParamType(FirstIRArg)->getPointerAddressSpace() ==
3936 TD->getAllocaAddrSpace()) &&
3937 "indirect argument must be in alloca address space");
3939 bool NeedCopy = false;
3941 if (Addr.getAlignment() < Align &&
3942 llvm::getOrEnforceKnownAlignment(V, Align.getQuantity(), *TD) <
3943 Align.getQuantity()) {
3945 } else if (I->hasLValue()) {
3946 auto LV = I->getKnownLValue();
3947 auto AS = LV.getAddressSpace();
3948 if ((!ArgInfo.getIndirectByVal() &&
3949 (LV.getAlignment() >=
3950 getContext().getTypeAlignInChars(I->Ty))) ||
3951 (ArgInfo.getIndirectByVal() &&
3952 ((AS != LangAS::Default && AS != LangAS::opencl_private &&
3953 AS != CGM.getASTAllocaAddressSpace())))) {
3958 // Create an aligned temporary, and copy to it.
3959 Address AI = CreateMemTempWithoutCast(
3960 I->Ty, ArgInfo.getIndirectAlign(), "byval-temp");
3961 IRCallArgs[FirstIRArg] = AI.getPointer();
3962 I->copyInto(*this, AI);
3964 // Skip the extra memcpy call.
3965 auto *T = V->getType()->getPointerElementType()->getPointerTo(
3966 CGM.getDataLayout().getAllocaAddrSpace());
3967 IRCallArgs[FirstIRArg] = getTargetHooks().performAddrSpaceCast(
3968 *this, V, LangAS::Default, CGM.getASTAllocaAddressSpace(), T,
3975 case ABIArgInfo::Ignore:
3976 assert(NumIRArgs == 0);
3979 case ABIArgInfo::Extend:
3980 case ABIArgInfo::Direct: {
3981 if (!isa<llvm::StructType>(ArgInfo.getCoerceToType()) &&
3982 ArgInfo.getCoerceToType() == ConvertType(info_it->type) &&
3983 ArgInfo.getDirectOffset() == 0) {
3984 assert(NumIRArgs == 1);
3986 if (!I->isAggregate())
3987 V = I->getKnownRValue().getScalarVal();
3989 V = Builder.CreateLoad(
3990 I->hasLValue() ? I->getKnownLValue().getAddress()
3991 : I->getKnownRValue().getAggregateAddress());
3993 // Implement swifterror by copying into a new swifterror argument.
3994 // We'll write back in the normal path out of the call.
3995 if (CallInfo.getExtParameterInfo(ArgNo).getABI()
3996 == ParameterABI::SwiftErrorResult) {
3997 assert(!swiftErrorTemp.isValid() && "multiple swifterror args");
3999 QualType pointeeTy = I->Ty->getPointeeType();
4001 Address(V, getContext().getTypeAlignInChars(pointeeTy));
4004 CreateMemTemp(pointeeTy, getPointerAlign(), "swifterror.temp");
4005 V = swiftErrorTemp.getPointer();
4006 cast<llvm::AllocaInst>(V)->setSwiftError(true);
4008 llvm::Value *errorValue = Builder.CreateLoad(swiftErrorArg);
4009 Builder.CreateStore(errorValue, swiftErrorTemp);
4012 // We might have to widen integers, but we should never truncate.
4013 if (ArgInfo.getCoerceToType() != V->getType() &&
4014 V->getType()->isIntegerTy())
4015 V = Builder.CreateZExt(V, ArgInfo.getCoerceToType());
4017 // If the argument doesn't match, perform a bitcast to coerce it. This
4018 // can happen due to trivial type mismatches.
4019 if (FirstIRArg < IRFuncTy->getNumParams() &&
4020 V->getType() != IRFuncTy->getParamType(FirstIRArg))
4021 V = Builder.CreateBitCast(V, IRFuncTy->getParamType(FirstIRArg));
4023 IRCallArgs[FirstIRArg] = V;
4027 // FIXME: Avoid the conversion through memory if possible.
4028 Address Src = Address::invalid();
4029 if (!I->isAggregate()) {
4030 Src = CreateMemTemp(I->Ty, "coerce");
4031 I->copyInto(*this, Src);
4033 Src = I->hasLValue() ? I->getKnownLValue().getAddress()
4034 : I->getKnownRValue().getAggregateAddress();
4037 // If the value is offset in memory, apply the offset now.
4038 Src = emitAddressAtOffset(*this, Src, ArgInfo);
4040 // Fast-isel and the optimizer generally like scalar values better than
4041 // FCAs, so we flatten them if this is safe to do for this argument.
4042 llvm::StructType *STy =
4043 dyn_cast<llvm::StructType>(ArgInfo.getCoerceToType());
4044 if (STy && ArgInfo.isDirect() && ArgInfo.getCanBeFlattened()) {
4045 llvm::Type *SrcTy = Src.getType()->getElementType();
4046 uint64_t SrcSize = CGM.getDataLayout().getTypeAllocSize(SrcTy);
4047 uint64_t DstSize = CGM.getDataLayout().getTypeAllocSize(STy);
4049 // If the source type is smaller than the destination type of the
4050 // coerce-to logic, copy the source value into a temp alloca the size
4051 // of the destination type to allow loading all of it. The bits past
4052 // the source value are left undef.
4053 if (SrcSize < DstSize) {
4055 = CreateTempAlloca(STy, Src.getAlignment(),
4056 Src.getName() + ".coerce");
4057 Builder.CreateMemCpy(TempAlloca, Src, SrcSize);
4060 Src = Builder.CreateBitCast(Src,
4061 STy->getPointerTo(Src.getAddressSpace()));
4064 auto SrcLayout = CGM.getDataLayout().getStructLayout(STy);
4065 assert(NumIRArgs == STy->getNumElements());
4066 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
4067 auto Offset = CharUnits::fromQuantity(SrcLayout->getElementOffset(i));
4068 Address EltPtr = Builder.CreateStructGEP(Src, i, Offset);
4069 llvm::Value *LI = Builder.CreateLoad(EltPtr);
4070 IRCallArgs[FirstIRArg + i] = LI;
4073 // In the simple case, just pass the coerced loaded value.
4074 assert(NumIRArgs == 1);
4075 IRCallArgs[FirstIRArg] =
4076 CreateCoercedLoad(Src, ArgInfo.getCoerceToType(), *this);
4082 case ABIArgInfo::CoerceAndExpand: {
4083 auto coercionType = ArgInfo.getCoerceAndExpandType();
4084 auto layout = CGM.getDataLayout().getStructLayout(coercionType);
4086 llvm::Value *tempSize = nullptr;
4087 Address addr = Address::invalid();
4088 Address AllocaAddr = Address::invalid();
4089 if (I->isAggregate()) {
4090 addr = I->hasLValue() ? I->getKnownLValue().getAddress()
4091 : I->getKnownRValue().getAggregateAddress();
4094 RValue RV = I->getKnownRValue();
4095 assert(RV.isScalar()); // complex should always just be direct
4097 llvm::Type *scalarType = RV.getScalarVal()->getType();
4098 auto scalarSize = CGM.getDataLayout().getTypeAllocSize(scalarType);
4099 auto scalarAlign = CGM.getDataLayout().getPrefTypeAlignment(scalarType);
4101 // Materialize to a temporary.
4102 addr = CreateTempAlloca(RV.getScalarVal()->getType(),
4103 CharUnits::fromQuantity(std::max(
4104 layout->getAlignment(), scalarAlign)),
4106 /*ArraySize=*/nullptr, &AllocaAddr);
4107 tempSize = EmitLifetimeStart(scalarSize, AllocaAddr.getPointer());
4109 Builder.CreateStore(RV.getScalarVal(), addr);
4112 addr = Builder.CreateElementBitCast(addr, coercionType);
4114 unsigned IRArgPos = FirstIRArg;
4115 for (unsigned i = 0, e = coercionType->getNumElements(); i != e; ++i) {
4116 llvm::Type *eltType = coercionType->getElementType(i);
4117 if (ABIArgInfo::isPaddingForCoerceAndExpand(eltType)) continue;
4118 Address eltAddr = Builder.CreateStructGEP(addr, i, layout);
4119 llvm::Value *elt = Builder.CreateLoad(eltAddr);
4120 IRCallArgs[IRArgPos++] = elt;
4122 assert(IRArgPos == FirstIRArg + NumIRArgs);
4125 EmitLifetimeEnd(tempSize, AllocaAddr.getPointer());
4131 case ABIArgInfo::Expand:
4132 unsigned IRArgPos = FirstIRArg;
4133 ExpandTypeToArgs(I->Ty, *I, IRFuncTy, IRCallArgs, IRArgPos);
4134 assert(IRArgPos == FirstIRArg + NumIRArgs);
4139 const CGCallee &ConcreteCallee = Callee.prepareConcreteCallee(*this);
4140 llvm::Value *CalleePtr = ConcreteCallee.getFunctionPointer();
4142 // If we're using inalloca, set up that argument.
4143 if (ArgMemory.isValid()) {
4144 llvm::Value *Arg = ArgMemory.getPointer();
4145 if (CallInfo.isVariadic()) {
4146 // When passing non-POD arguments by value to variadic functions, we will
4147 // end up with a variadic prototype and an inalloca call site. In such
4148 // cases, we can't do any parameter mismatch checks. Give up and bitcast
4150 unsigned CalleeAS = CalleePtr->getType()->getPointerAddressSpace();
4151 auto FnTy = getTypes().GetFunctionType(CallInfo)->getPointerTo(CalleeAS);
4152 CalleePtr = Builder.CreateBitCast(CalleePtr, FnTy);
4154 llvm::Type *LastParamTy =
4155 IRFuncTy->getParamType(IRFuncTy->getNumParams() - 1);
4156 if (Arg->getType() != LastParamTy) {
4158 // Assert that these structs have equivalent element types.
4159 llvm::StructType *FullTy = CallInfo.getArgStruct();
4160 llvm::StructType *DeclaredTy = cast<llvm::StructType>(
4161 cast<llvm::PointerType>(LastParamTy)->getElementType());
4162 assert(DeclaredTy->getNumElements() == FullTy->getNumElements());
4163 for (llvm::StructType::element_iterator DI = DeclaredTy->element_begin(),
4164 DE = DeclaredTy->element_end(),
4165 FI = FullTy->element_begin();
4166 DI != DE; ++DI, ++FI)
4169 Arg = Builder.CreateBitCast(Arg, LastParamTy);
4172 assert(IRFunctionArgs.hasInallocaArg());
4173 IRCallArgs[IRFunctionArgs.getInallocaArgNo()] = Arg;
4176 // 2. Prepare the function pointer.
4178 // If the callee is a bitcast of a non-variadic function to have a
4179 // variadic function pointer type, check to see if we can remove the
4180 // bitcast. This comes up with unprototyped functions.
4182 // This makes the IR nicer, but more importantly it ensures that we
4183 // can inline the function at -O0 if it is marked always_inline.
4184 auto simplifyVariadicCallee = [](llvm::Value *Ptr) -> llvm::Value* {
4185 llvm::FunctionType *CalleeFT =
4186 cast<llvm::FunctionType>(Ptr->getType()->getPointerElementType());
4187 if (!CalleeFT->isVarArg())
4190 llvm::ConstantExpr *CE = dyn_cast<llvm::ConstantExpr>(Ptr);
4191 if (!CE || CE->getOpcode() != llvm::Instruction::BitCast)
4194 llvm::Function *OrigFn = dyn_cast<llvm::Function>(CE->getOperand(0));
4198 llvm::FunctionType *OrigFT = OrigFn->getFunctionType();
4200 // If the original type is variadic, or if any of the component types
4201 // disagree, we cannot remove the cast.
4202 if (OrigFT->isVarArg() ||
4203 OrigFT->getNumParams() != CalleeFT->getNumParams() ||
4204 OrigFT->getReturnType() != CalleeFT->getReturnType())
4207 for (unsigned i = 0, e = OrigFT->getNumParams(); i != e; ++i)
4208 if (OrigFT->getParamType(i) != CalleeFT->getParamType(i))
4213 CalleePtr = simplifyVariadicCallee(CalleePtr);
4215 // 3. Perform the actual call.
4217 // Deactivate any cleanups that we're supposed to do immediately before
4219 if (!CallArgs.getCleanupsToDeactivate().empty())
4220 deactivateArgCleanupsBeforeCall(*this, CallArgs);
4222 // Assert that the arguments we computed match up. The IR verifier
4223 // will catch this, but this is a common enough source of problems
4224 // during IRGen changes that it's way better for debugging to catch
4225 // it ourselves here.
4227 assert(IRCallArgs.size() == IRFuncTy->getNumParams() || IRFuncTy->isVarArg());
4228 for (unsigned i = 0; i < IRCallArgs.size(); ++i) {
4229 // Inalloca argument can have different type.
4230 if (IRFunctionArgs.hasInallocaArg() &&
4231 i == IRFunctionArgs.getInallocaArgNo())
4233 if (i < IRFuncTy->getNumParams())
4234 assert(IRCallArgs[i]->getType() == IRFuncTy->getParamType(i));
4238 // Compute the calling convention and attributes.
4239 unsigned CallingConv;
4240 llvm::AttributeList Attrs;
4241 CGM.ConstructAttributeList(CalleePtr->getName(), CallInfo,
4242 Callee.getAbstractInfo(), Attrs, CallingConv,
4243 /*AttrOnCallSite=*/true);
4245 // Apply some call-site-specific attributes.
4246 // TODO: work this into building the attribute set.
4248 // Apply always_inline to all calls within flatten functions.
4249 // FIXME: should this really take priority over __try, below?
4250 if (CurCodeDecl && CurCodeDecl->hasAttr<FlattenAttr>() &&
4251 !(Callee.getAbstractInfo().getCalleeDecl() &&
4252 Callee.getAbstractInfo().getCalleeDecl()->hasAttr<NoInlineAttr>())) {
4254 Attrs.addAttribute(getLLVMContext(), llvm::AttributeList::FunctionIndex,
4255 llvm::Attribute::AlwaysInline);
4258 // Disable inlining inside SEH __try blocks.
4259 if (isSEHTryScope()) {
4261 Attrs.addAttribute(getLLVMContext(), llvm::AttributeList::FunctionIndex,
4262 llvm::Attribute::NoInline);
4265 // Decide whether to use a call or an invoke.
4267 if (currentFunctionUsesSEHTry()) {
4268 // SEH cares about asynchronous exceptions, so everything can "throw."
4269 CannotThrow = false;
4270 } else if (isCleanupPadScope() &&
4271 EHPersonality::get(*this).isMSVCXXPersonality()) {
4272 // The MSVC++ personality will implicitly terminate the program if an
4273 // exception is thrown during a cleanup outside of a try/catch.
4274 // We don't need to model anything in IR to get this behavior.
4277 // Otherwise, nounwind call sites will never throw.
4278 CannotThrow = Attrs.hasAttribute(llvm::AttributeList::FunctionIndex,
4279 llvm::Attribute::NoUnwind);
4282 // If we made a temporary, be sure to clean up after ourselves. Note that we
4283 // can't depend on being inside of an ExprWithCleanups, so we need to manually
4284 // pop this cleanup later on. Being eager about this is OK, since this
4285 // temporary is 'invisible' outside of the callee.
4286 if (UnusedReturnSizePtr)
4287 pushFullExprCleanup<CallLifetimeEnd>(NormalEHLifetimeMarker, SRetAlloca,
4288 UnusedReturnSizePtr);
4290 llvm::BasicBlock *InvokeDest = CannotThrow ? nullptr : getInvokeDest();
4292 SmallVector<llvm::OperandBundleDef, 1> BundleList =
4293 getBundlesForFunclet(CalleePtr);
4295 // Emit the actual call/invoke instruction.
4298 CS = Builder.CreateCall(CalleePtr, IRCallArgs, BundleList);
4300 llvm::BasicBlock *Cont = createBasicBlock("invoke.cont");
4301 CS = Builder.CreateInvoke(CalleePtr, Cont, InvokeDest, IRCallArgs,
4305 llvm::Instruction *CI = CS.getInstruction();
4309 // Apply the attributes and calling convention.
4310 CS.setAttributes(Attrs);
4311 CS.setCallingConv(static_cast<llvm::CallingConv::ID>(CallingConv));
4313 // Apply various metadata.
4315 if (!CI->getType()->isVoidTy())
4316 CI->setName("call");
4318 // Insert instrumentation or attach profile metadata at indirect call sites.
4319 // For more details, see the comment before the definition of
4320 // IPVK_IndirectCallTarget in InstrProfData.inc.
4321 if (!CS.getCalledFunction())
4322 PGO.valueProfile(Builder, llvm::IPVK_IndirectCallTarget,
4325 // In ObjC ARC mode with no ObjC ARC exception safety, tell the ARC
4326 // optimizer it can aggressively ignore unwind edges.
4327 if (CGM.getLangOpts().ObjCAutoRefCount)
4328 AddObjCARCExceptionMetadata(CI);
4330 // Suppress tail calls if requested.
4331 if (llvm::CallInst *Call = dyn_cast<llvm::CallInst>(CI)) {
4332 const Decl *TargetDecl = Callee.getAbstractInfo().getCalleeDecl();
4333 if (TargetDecl && TargetDecl->hasAttr<NotTailCalledAttr>())
4334 Call->setTailCallKind(llvm::CallInst::TCK_NoTail);
4337 // 4. Finish the call.
4339 // If the call doesn't return, finish the basic block and clear the
4340 // insertion point; this allows the rest of IRGen to discard
4341 // unreachable code.
4342 if (CS.doesNotReturn()) {
4343 if (UnusedReturnSizePtr)
4346 // Strip away the noreturn attribute to better diagnose unreachable UB.
4347 if (SanOpts.has(SanitizerKind::Unreachable)) {
4348 if (auto *F = CS.getCalledFunction())
4349 F->removeFnAttr(llvm::Attribute::NoReturn);
4350 CS.removeAttribute(llvm::AttributeList::FunctionIndex,
4351 llvm::Attribute::NoReturn);
4354 EmitUnreachable(Loc);
4355 Builder.ClearInsertionPoint();
4357 // FIXME: For now, emit a dummy basic block because expr emitters in
4358 // generally are not ready to handle emitting expressions at unreachable
4360 EnsureInsertPoint();
4362 // Return a reasonable RValue.
4363 return GetUndefRValue(RetTy);
4366 // Perform the swifterror writeback.
4367 if (swiftErrorTemp.isValid()) {
4368 llvm::Value *errorResult = Builder.CreateLoad(swiftErrorTemp);
4369 Builder.CreateStore(errorResult, swiftErrorArg);
4372 // Emit any call-associated writebacks immediately. Arguably this
4373 // should happen after any return-value munging.
4374 if (CallArgs.hasWritebacks())
4375 emitWritebacks(*this, CallArgs);
4377 // The stack cleanup for inalloca arguments has to run out of the normal
4378 // lexical order, so deactivate it and run it manually here.
4379 CallArgs.freeArgumentMemory(*this);
4381 // Extract the return value.
4383 switch (RetAI.getKind()) {
4384 case ABIArgInfo::CoerceAndExpand: {
4385 auto coercionType = RetAI.getCoerceAndExpandType();
4386 auto layout = CGM.getDataLayout().getStructLayout(coercionType);
4388 Address addr = SRetPtr;
4389 addr = Builder.CreateElementBitCast(addr, coercionType);
4391 assert(CI->getType() == RetAI.getUnpaddedCoerceAndExpandType());
4392 bool requiresExtract = isa<llvm::StructType>(CI->getType());
4394 unsigned unpaddedIndex = 0;
4395 for (unsigned i = 0, e = coercionType->getNumElements(); i != e; ++i) {
4396 llvm::Type *eltType = coercionType->getElementType(i);
4397 if (ABIArgInfo::isPaddingForCoerceAndExpand(eltType)) continue;
4398 Address eltAddr = Builder.CreateStructGEP(addr, i, layout);
4399 llvm::Value *elt = CI;
4400 if (requiresExtract)
4401 elt = Builder.CreateExtractValue(elt, unpaddedIndex++);
4403 assert(unpaddedIndex == 0);
4404 Builder.CreateStore(elt, eltAddr);
4410 case ABIArgInfo::InAlloca:
4411 case ABIArgInfo::Indirect: {
4412 RValue ret = convertTempToRValue(SRetPtr, RetTy, SourceLocation());
4413 if (UnusedReturnSizePtr)
4418 case ABIArgInfo::Ignore:
4419 // If we are ignoring an argument that had a result, make sure to
4420 // construct the appropriate return value for our caller.
4421 return GetUndefRValue(RetTy);
4423 case ABIArgInfo::Extend:
4424 case ABIArgInfo::Direct: {
4425 llvm::Type *RetIRTy = ConvertType(RetTy);
4426 if (RetAI.getCoerceToType() == RetIRTy && RetAI.getDirectOffset() == 0) {
4427 switch (getEvaluationKind(RetTy)) {
4429 llvm::Value *Real = Builder.CreateExtractValue(CI, 0);
4430 llvm::Value *Imag = Builder.CreateExtractValue(CI, 1);
4431 return RValue::getComplex(std::make_pair(Real, Imag));
4433 case TEK_Aggregate: {
4434 Address DestPtr = ReturnValue.getValue();
4435 bool DestIsVolatile = ReturnValue.isVolatile();
4437 if (!DestPtr.isValid()) {
4438 DestPtr = CreateMemTemp(RetTy, "agg.tmp");
4439 DestIsVolatile = false;
4441 BuildAggStore(*this, CI, DestPtr, DestIsVolatile);
4442 return RValue::getAggregate(DestPtr);
4445 // If the argument doesn't match, perform a bitcast to coerce it. This
4446 // can happen due to trivial type mismatches.
4447 llvm::Value *V = CI;
4448 if (V->getType() != RetIRTy)
4449 V = Builder.CreateBitCast(V, RetIRTy);
4450 return RValue::get(V);
4453 llvm_unreachable("bad evaluation kind");
4456 Address DestPtr = ReturnValue.getValue();
4457 bool DestIsVolatile = ReturnValue.isVolatile();
4459 if (!DestPtr.isValid()) {
4460 DestPtr = CreateMemTemp(RetTy, "coerce");
4461 DestIsVolatile = false;
4464 // If the value is offset in memory, apply the offset now.
4465 Address StorePtr = emitAddressAtOffset(*this, DestPtr, RetAI);
4466 CreateCoercedStore(CI, StorePtr, DestIsVolatile, *this);
4468 return convertTempToRValue(DestPtr, RetTy, SourceLocation());
4471 case ABIArgInfo::Expand:
4472 llvm_unreachable("Invalid ABI kind for return argument");
4475 llvm_unreachable("Unhandled ABIArgInfo::Kind");
4478 // Emit the assume_aligned check on the return value.
4479 const Decl *TargetDecl = Callee.getAbstractInfo().getCalleeDecl();
4480 if (Ret.isScalar() && TargetDecl) {
4481 if (const auto *AA = TargetDecl->getAttr<AssumeAlignedAttr>()) {
4482 llvm::Value *OffsetValue = nullptr;
4483 if (const auto *Offset = AA->getOffset())
4484 OffsetValue = EmitScalarExpr(Offset);
4486 llvm::Value *Alignment = EmitScalarExpr(AA->getAlignment());
4487 llvm::ConstantInt *AlignmentCI = cast<llvm::ConstantInt>(Alignment);
4488 EmitAlignmentAssumption(Ret.getScalarVal(), AlignmentCI->getZExtValue(),
4490 } else if (const auto *AA = TargetDecl->getAttr<AllocAlignAttr>()) {
4491 llvm::Value *ParamVal =
4492 CallArgs[AA->getParamIndex().getLLVMIndex()].getRValue(
4493 *this).getScalarVal();
4494 EmitAlignmentAssumption(Ret.getScalarVal(), ParamVal);
4501 CGCallee CGCallee::prepareConcreteCallee(CodeGenFunction &CGF) const {
4503 const CallExpr *CE = getVirtualCallExpr();
4504 return CGF.CGM.getCXXABI().getVirtualFunctionPointer(
4505 CGF, getVirtualMethodDecl(), getThisAddress(),
4506 getFunctionType(), CE ? CE->getLocStart() : SourceLocation());
4512 /* VarArg handling */
4514 Address CodeGenFunction::EmitVAArg(VAArgExpr *VE, Address &VAListAddr) {
4515 VAListAddr = VE->isMicrosoftABI()
4516 ? EmitMSVAListRef(VE->getSubExpr())
4517 : EmitVAListRef(VE->getSubExpr());
4518 QualType Ty = VE->getType();
4519 if (VE->isMicrosoftABI())
4520 return CGM.getTypes().getABIInfo().EmitMSVAArg(*this, VAListAddr, Ty);
4521 return CGM.getTypes().getABIInfo().EmitVAArg(*this, VAListAddr, Ty);