1 //===--- CGCall.cpp - Encapsulate calling convention details --------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // These classes wrap the information about a call or function
11 // definition used to handle ABI compliancy.
13 //===----------------------------------------------------------------------===//
19 #include "CGCleanup.h"
20 #include "CodeGenFunction.h"
21 #include "CodeGenModule.h"
22 #include "TargetInfo.h"
23 #include "clang/AST/Decl.h"
24 #include "clang/AST/DeclCXX.h"
25 #include "clang/AST/DeclObjC.h"
26 #include "clang/Basic/TargetBuiltins.h"
27 #include "clang/Basic/TargetInfo.h"
28 #include "clang/CodeGen/CGFunctionInfo.h"
29 #include "clang/CodeGen/SwiftCallingConv.h"
30 #include "clang/Frontend/CodeGenOptions.h"
31 #include "llvm/ADT/StringExtras.h"
32 #include "llvm/Analysis/ValueTracking.h"
33 #include "llvm/IR/Attributes.h"
34 #include "llvm/IR/CallingConv.h"
35 #include "llvm/IR/CallSite.h"
36 #include "llvm/IR/DataLayout.h"
37 #include "llvm/IR/InlineAsm.h"
38 #include "llvm/IR/Intrinsics.h"
39 #include "llvm/IR/IntrinsicInst.h"
40 #include "llvm/Transforms/Utils/Local.h"
41 using namespace clang;
42 using namespace CodeGen;
46 unsigned CodeGenTypes::ClangCallConvToLLVMCallConv(CallingConv CC) {
48 default: return llvm::CallingConv::C;
49 case CC_X86StdCall: return llvm::CallingConv::X86_StdCall;
50 case CC_X86FastCall: return llvm::CallingConv::X86_FastCall;
51 case CC_X86RegCall: return llvm::CallingConv::X86_RegCall;
52 case CC_X86ThisCall: return llvm::CallingConv::X86_ThisCall;
53 case CC_X86_64Win64: return llvm::CallingConv::X86_64_Win64;
54 case CC_X86_64SysV: return llvm::CallingConv::X86_64_SysV;
55 case CC_AAPCS: return llvm::CallingConv::ARM_AAPCS;
56 case CC_AAPCS_VFP: return llvm::CallingConv::ARM_AAPCS_VFP;
57 case CC_IntelOclBicc: return llvm::CallingConv::Intel_OCL_BI;
58 // TODO: Add support for __pascal to LLVM.
59 case CC_X86Pascal: return llvm::CallingConv::C;
60 // TODO: Add support for __vectorcall to LLVM.
61 case CC_X86VectorCall: return llvm::CallingConv::X86_VectorCall;
62 case CC_SpirFunction: return llvm::CallingConv::SPIR_FUNC;
63 case CC_OpenCLKernel: return CGM.getTargetCodeGenInfo().getOpenCLKernelCallingConv();
64 case CC_PreserveMost: return llvm::CallingConv::PreserveMost;
65 case CC_PreserveAll: return llvm::CallingConv::PreserveAll;
66 case CC_Swift: return llvm::CallingConv::Swift;
70 /// Derives the 'this' type for codegen purposes, i.e. ignoring method
72 /// FIXME: address space qualification?
73 static CanQualType GetThisType(ASTContext &Context, const CXXRecordDecl *RD) {
74 QualType RecTy = Context.getTagDeclType(RD)->getCanonicalTypeInternal();
75 return Context.getPointerType(CanQualType::CreateUnsafe(RecTy));
78 /// Returns the canonical formal type of the given C++ method.
79 static CanQual<FunctionProtoType> GetFormalType(const CXXMethodDecl *MD) {
80 return MD->getType()->getCanonicalTypeUnqualified()
81 .getAs<FunctionProtoType>();
84 /// Returns the "extra-canonicalized" return type, which discards
85 /// qualifiers on the return type. Codegen doesn't care about them,
86 /// and it makes ABI code a little easier to be able to assume that
87 /// all parameter and return types are top-level unqualified.
88 static CanQualType GetReturnType(QualType RetTy) {
89 return RetTy->getCanonicalTypeUnqualified().getUnqualifiedType();
92 /// Arrange the argument and result information for a value of the given
93 /// unprototyped freestanding function type.
94 const CGFunctionInfo &
95 CodeGenTypes::arrangeFreeFunctionType(CanQual<FunctionNoProtoType> FTNP) {
96 // When translating an unprototyped function type, always use a
98 return arrangeLLVMFunctionInfo(FTNP->getReturnType().getUnqualifiedType(),
99 /*instanceMethod=*/false,
100 /*chainCall=*/false, None,
101 FTNP->getExtInfo(), {}, RequiredArgs(0));
104 static void addExtParameterInfosForCall(
105 llvm::SmallVectorImpl<FunctionProtoType::ExtParameterInfo> ¶mInfos,
106 const FunctionProtoType *proto,
108 unsigned totalArgs) {
109 assert(proto->hasExtParameterInfos());
110 assert(paramInfos.size() <= prefixArgs);
111 assert(proto->getNumParams() + prefixArgs <= totalArgs);
113 paramInfos.reserve(totalArgs);
115 // Add default infos for any prefix args that don't already have infos.
116 paramInfos.resize(prefixArgs);
118 // Add infos for the prototype.
119 for (const auto &ParamInfo : proto->getExtParameterInfos()) {
120 paramInfos.push_back(ParamInfo);
121 // pass_object_size params have no parameter info.
122 if (ParamInfo.hasPassObjectSize())
123 paramInfos.emplace_back();
126 assert(paramInfos.size() <= totalArgs &&
127 "Did we forget to insert pass_object_size args?");
128 // Add default infos for the variadic and/or suffix arguments.
129 paramInfos.resize(totalArgs);
132 /// Adds the formal paramaters in FPT to the given prefix. If any parameter in
133 /// FPT has pass_object_size attrs, then we'll add parameters for those, too.
134 static void appendParameterTypes(const CodeGenTypes &CGT,
135 SmallVectorImpl<CanQualType> &prefix,
136 SmallVectorImpl<FunctionProtoType::ExtParameterInfo> ¶mInfos,
137 CanQual<FunctionProtoType> FPT) {
138 // Fast path: don't touch param info if we don't need to.
139 if (!FPT->hasExtParameterInfos()) {
140 assert(paramInfos.empty() &&
141 "We have paramInfos, but the prototype doesn't?");
142 prefix.append(FPT->param_type_begin(), FPT->param_type_end());
146 unsigned PrefixSize = prefix.size();
147 // In the vast majority of cases, we'll have precisely FPT->getNumParams()
148 // parameters; the only thing that can change this is the presence of
149 // pass_object_size. So, we preallocate for the common case.
150 prefix.reserve(prefix.size() + FPT->getNumParams());
152 auto ExtInfos = FPT->getExtParameterInfos();
153 assert(ExtInfos.size() == FPT->getNumParams());
154 for (unsigned I = 0, E = FPT->getNumParams(); I != E; ++I) {
155 prefix.push_back(FPT->getParamType(I));
156 if (ExtInfos[I].hasPassObjectSize())
157 prefix.push_back(CGT.getContext().getSizeType());
160 addExtParameterInfosForCall(paramInfos, FPT.getTypePtr(), PrefixSize,
164 /// Arrange the LLVM function layout for a value of the given function
165 /// type, on top of any implicit parameters already stored.
166 static const CGFunctionInfo &
167 arrangeLLVMFunctionInfo(CodeGenTypes &CGT, bool instanceMethod,
168 SmallVectorImpl<CanQualType> &prefix,
169 CanQual<FunctionProtoType> FTP,
170 const FunctionDecl *FD) {
171 SmallVector<FunctionProtoType::ExtParameterInfo, 16> paramInfos;
172 RequiredArgs Required =
173 RequiredArgs::forPrototypePlus(FTP, prefix.size(), FD);
175 appendParameterTypes(CGT, prefix, paramInfos, FTP);
176 CanQualType resultType = FTP->getReturnType().getUnqualifiedType();
178 return CGT.arrangeLLVMFunctionInfo(resultType, instanceMethod,
179 /*chainCall=*/false, prefix,
180 FTP->getExtInfo(), paramInfos,
184 /// Arrange the argument and result information for a value of the
185 /// given freestanding function type.
186 const CGFunctionInfo &
187 CodeGenTypes::arrangeFreeFunctionType(CanQual<FunctionProtoType> FTP,
188 const FunctionDecl *FD) {
189 SmallVector<CanQualType, 16> argTypes;
190 return ::arrangeLLVMFunctionInfo(*this, /*instanceMethod=*/false, argTypes,
194 static CallingConv getCallingConventionForDecl(const Decl *D, bool IsWindows) {
195 // Set the appropriate calling convention for the Function.
196 if (D->hasAttr<StdCallAttr>())
197 return CC_X86StdCall;
199 if (D->hasAttr<FastCallAttr>())
200 return CC_X86FastCall;
202 if (D->hasAttr<RegCallAttr>())
203 return CC_X86RegCall;
205 if (D->hasAttr<ThisCallAttr>())
206 return CC_X86ThisCall;
208 if (D->hasAttr<VectorCallAttr>())
209 return CC_X86VectorCall;
211 if (D->hasAttr<PascalAttr>())
214 if (PcsAttr *PCS = D->getAttr<PcsAttr>())
215 return (PCS->getPCS() == PcsAttr::AAPCS ? CC_AAPCS : CC_AAPCS_VFP);
217 if (D->hasAttr<IntelOclBiccAttr>())
218 return CC_IntelOclBicc;
220 if (D->hasAttr<MSABIAttr>())
221 return IsWindows ? CC_C : CC_X86_64Win64;
223 if (D->hasAttr<SysVABIAttr>())
224 return IsWindows ? CC_X86_64SysV : CC_C;
226 if (D->hasAttr<PreserveMostAttr>())
227 return CC_PreserveMost;
229 if (D->hasAttr<PreserveAllAttr>())
230 return CC_PreserveAll;
235 /// Arrange the argument and result information for a call to an
236 /// unknown C++ non-static member function of the given abstract type.
237 /// (Zero value of RD means we don't have any meaningful "this" argument type,
238 /// so fall back to a generic pointer type).
239 /// The member function must be an ordinary function, i.e. not a
240 /// constructor or destructor.
241 const CGFunctionInfo &
242 CodeGenTypes::arrangeCXXMethodType(const CXXRecordDecl *RD,
243 const FunctionProtoType *FTP,
244 const CXXMethodDecl *MD) {
245 SmallVector<CanQualType, 16> argTypes;
247 // Add the 'this' pointer.
249 argTypes.push_back(GetThisType(Context, RD));
251 argTypes.push_back(Context.VoidPtrTy);
253 return ::arrangeLLVMFunctionInfo(
254 *this, true, argTypes,
255 FTP->getCanonicalTypeUnqualified().getAs<FunctionProtoType>(), MD);
258 /// Arrange the argument and result information for a declaration or
259 /// definition of the given C++ non-static member function. The
260 /// member function must be an ordinary function, i.e. not a
261 /// constructor or destructor.
262 const CGFunctionInfo &
263 CodeGenTypes::arrangeCXXMethodDeclaration(const CXXMethodDecl *MD) {
264 assert(!isa<CXXConstructorDecl>(MD) && "wrong method for constructors!");
265 assert(!isa<CXXDestructorDecl>(MD) && "wrong method for destructors!");
267 CanQual<FunctionProtoType> prototype = GetFormalType(MD);
269 if (MD->isInstance()) {
270 // The abstract case is perfectly fine.
271 const CXXRecordDecl *ThisType = TheCXXABI.getThisArgumentTypeForMethod(MD);
272 return arrangeCXXMethodType(ThisType, prototype.getTypePtr(), MD);
275 return arrangeFreeFunctionType(prototype, MD);
278 bool CodeGenTypes::inheritingCtorHasParams(
279 const InheritedConstructor &Inherited, CXXCtorType Type) {
280 // Parameters are unnecessary if we're constructing a base class subobject
281 // and the inherited constructor lives in a virtual base.
282 return Type == Ctor_Complete ||
283 !Inherited.getShadowDecl()->constructsVirtualBase() ||
284 !Target.getCXXABI().hasConstructorVariants();
287 const CGFunctionInfo &
288 CodeGenTypes::arrangeCXXStructorDeclaration(const CXXMethodDecl *MD,
291 SmallVector<CanQualType, 16> argTypes;
292 SmallVector<FunctionProtoType::ExtParameterInfo, 16> paramInfos;
293 argTypes.push_back(GetThisType(Context, MD->getParent()));
295 bool PassParams = true;
298 if (auto *CD = dyn_cast<CXXConstructorDecl>(MD)) {
299 GD = GlobalDecl(CD, toCXXCtorType(Type));
301 // A base class inheriting constructor doesn't get forwarded arguments
302 // needed to construct a virtual base (or base class thereof).
303 if (auto Inherited = CD->getInheritedConstructor())
304 PassParams = inheritingCtorHasParams(Inherited, toCXXCtorType(Type));
306 auto *DD = dyn_cast<CXXDestructorDecl>(MD);
307 GD = GlobalDecl(DD, toCXXDtorType(Type));
310 CanQual<FunctionProtoType> FTP = GetFormalType(MD);
312 // Add the formal parameters.
314 appendParameterTypes(*this, argTypes, paramInfos, FTP);
316 CGCXXABI::AddedStructorArgs AddedArgs =
317 TheCXXABI.buildStructorSignature(MD, Type, argTypes);
318 if (!paramInfos.empty()) {
319 // Note: prefix implies after the first param.
320 if (AddedArgs.Prefix)
321 paramInfos.insert(paramInfos.begin() + 1, AddedArgs.Prefix,
322 FunctionProtoType::ExtParameterInfo{});
323 if (AddedArgs.Suffix)
324 paramInfos.append(AddedArgs.Suffix,
325 FunctionProtoType::ExtParameterInfo{});
328 RequiredArgs required =
329 (PassParams && MD->isVariadic() ? RequiredArgs(argTypes.size())
330 : RequiredArgs::All);
332 FunctionType::ExtInfo extInfo = FTP->getExtInfo();
333 CanQualType resultType = TheCXXABI.HasThisReturn(GD)
335 : TheCXXABI.hasMostDerivedReturn(GD)
336 ? CGM.getContext().VoidPtrTy
338 return arrangeLLVMFunctionInfo(resultType, /*instanceMethod=*/true,
339 /*chainCall=*/false, argTypes, extInfo,
340 paramInfos, required);
343 static SmallVector<CanQualType, 16>
344 getArgTypesForCall(ASTContext &ctx, const CallArgList &args) {
345 SmallVector<CanQualType, 16> argTypes;
346 for (auto &arg : args)
347 argTypes.push_back(ctx.getCanonicalParamType(arg.Ty));
351 static SmallVector<CanQualType, 16>
352 getArgTypesForDeclaration(ASTContext &ctx, const FunctionArgList &args) {
353 SmallVector<CanQualType, 16> argTypes;
354 for (auto &arg : args)
355 argTypes.push_back(ctx.getCanonicalParamType(arg->getType()));
359 static llvm::SmallVector<FunctionProtoType::ExtParameterInfo, 16>
360 getExtParameterInfosForCall(const FunctionProtoType *proto,
361 unsigned prefixArgs, unsigned totalArgs) {
362 llvm::SmallVector<FunctionProtoType::ExtParameterInfo, 16> result;
363 if (proto->hasExtParameterInfos()) {
364 addExtParameterInfosForCall(result, proto, prefixArgs, totalArgs);
369 /// Arrange a call to a C++ method, passing the given arguments.
371 /// ExtraPrefixArgs is the number of ABI-specific args passed after the `this`
373 /// ExtraSuffixArgs is the number of ABI-specific args passed at the end of
375 /// PassProtoArgs indicates whether `args` has args for the parameters in the
376 /// given CXXConstructorDecl.
377 const CGFunctionInfo &
378 CodeGenTypes::arrangeCXXConstructorCall(const CallArgList &args,
379 const CXXConstructorDecl *D,
380 CXXCtorType CtorKind,
381 unsigned ExtraPrefixArgs,
382 unsigned ExtraSuffixArgs,
383 bool PassProtoArgs) {
385 SmallVector<CanQualType, 16> ArgTypes;
386 for (const auto &Arg : args)
387 ArgTypes.push_back(Context.getCanonicalParamType(Arg.Ty));
389 // +1 for implicit this, which should always be args[0].
390 unsigned TotalPrefixArgs = 1 + ExtraPrefixArgs;
392 CanQual<FunctionProtoType> FPT = GetFormalType(D);
393 RequiredArgs Required =
394 RequiredArgs::forPrototypePlus(FPT, TotalPrefixArgs + ExtraSuffixArgs, D);
395 GlobalDecl GD(D, CtorKind);
396 CanQualType ResultType = TheCXXABI.HasThisReturn(GD)
398 : TheCXXABI.hasMostDerivedReturn(GD)
399 ? CGM.getContext().VoidPtrTy
402 FunctionType::ExtInfo Info = FPT->getExtInfo();
403 llvm::SmallVector<FunctionProtoType::ExtParameterInfo, 16> ParamInfos;
404 // If the prototype args are elided, we should only have ABI-specific args,
405 // which never have param info.
406 if (PassProtoArgs && FPT->hasExtParameterInfos()) {
407 // ABI-specific suffix arguments are treated the same as variadic arguments.
408 addExtParameterInfosForCall(ParamInfos, FPT.getTypePtr(), TotalPrefixArgs,
411 return arrangeLLVMFunctionInfo(ResultType, /*instanceMethod=*/true,
412 /*chainCall=*/false, ArgTypes, Info,
413 ParamInfos, Required);
416 /// Arrange the argument and result information for the declaration or
417 /// definition of the given function.
418 const CGFunctionInfo &
419 CodeGenTypes::arrangeFunctionDeclaration(const FunctionDecl *FD) {
420 if (const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(FD))
421 if (MD->isInstance())
422 return arrangeCXXMethodDeclaration(MD);
424 CanQualType FTy = FD->getType()->getCanonicalTypeUnqualified();
426 assert(isa<FunctionType>(FTy));
428 // When declaring a function without a prototype, always use a
429 // non-variadic type.
430 if (CanQual<FunctionNoProtoType> noProto = FTy.getAs<FunctionNoProtoType>()) {
431 return arrangeLLVMFunctionInfo(
432 noProto->getReturnType(), /*instanceMethod=*/false,
433 /*chainCall=*/false, None, noProto->getExtInfo(), {},RequiredArgs::All);
436 return arrangeFreeFunctionType(FTy.castAs<FunctionProtoType>(), FD);
439 /// Arrange the argument and result information for the declaration or
440 /// definition of an Objective-C method.
441 const CGFunctionInfo &
442 CodeGenTypes::arrangeObjCMethodDeclaration(const ObjCMethodDecl *MD) {
443 // It happens that this is the same as a call with no optional
444 // arguments, except also using the formal 'self' type.
445 return arrangeObjCMessageSendSignature(MD, MD->getSelfDecl()->getType());
448 /// Arrange the argument and result information for the function type
449 /// through which to perform a send to the given Objective-C method,
450 /// using the given receiver type. The receiver type is not always
451 /// the 'self' type of the method or even an Objective-C pointer type.
452 /// This is *not* the right method for actually performing such a
453 /// message send, due to the possibility of optional arguments.
454 const CGFunctionInfo &
455 CodeGenTypes::arrangeObjCMessageSendSignature(const ObjCMethodDecl *MD,
456 QualType receiverType) {
457 SmallVector<CanQualType, 16> argTys;
458 argTys.push_back(Context.getCanonicalParamType(receiverType));
459 argTys.push_back(Context.getCanonicalParamType(Context.getObjCSelType()));
461 for (const auto *I : MD->parameters()) {
462 argTys.push_back(Context.getCanonicalParamType(I->getType()));
465 FunctionType::ExtInfo einfo;
466 bool IsWindows = getContext().getTargetInfo().getTriple().isOSWindows();
467 einfo = einfo.withCallingConv(getCallingConventionForDecl(MD, IsWindows));
469 if (getContext().getLangOpts().ObjCAutoRefCount &&
470 MD->hasAttr<NSReturnsRetainedAttr>())
471 einfo = einfo.withProducesResult(true);
473 RequiredArgs required =
474 (MD->isVariadic() ? RequiredArgs(argTys.size()) : RequiredArgs::All);
476 return arrangeLLVMFunctionInfo(
477 GetReturnType(MD->getReturnType()), /*instanceMethod=*/false,
478 /*chainCall=*/false, argTys, einfo, {}, required);
481 const CGFunctionInfo &
482 CodeGenTypes::arrangeUnprototypedObjCMessageSend(QualType returnType,
483 const CallArgList &args) {
484 auto argTypes = getArgTypesForCall(Context, args);
485 FunctionType::ExtInfo einfo;
487 return arrangeLLVMFunctionInfo(
488 GetReturnType(returnType), /*instanceMethod=*/false,
489 /*chainCall=*/false, argTypes, einfo, {}, RequiredArgs::All);
492 const CGFunctionInfo &
493 CodeGenTypes::arrangeGlobalDeclaration(GlobalDecl GD) {
494 // FIXME: Do we need to handle ObjCMethodDecl?
495 const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl());
497 if (const CXXConstructorDecl *CD = dyn_cast<CXXConstructorDecl>(FD))
498 return arrangeCXXStructorDeclaration(CD, getFromCtorType(GD.getCtorType()));
500 if (const CXXDestructorDecl *DD = dyn_cast<CXXDestructorDecl>(FD))
501 return arrangeCXXStructorDeclaration(DD, getFromDtorType(GD.getDtorType()));
503 return arrangeFunctionDeclaration(FD);
506 /// Arrange a thunk that takes 'this' as the first parameter followed by
507 /// varargs. Return a void pointer, regardless of the actual return type.
508 /// The body of the thunk will end in a musttail call to a function of the
509 /// correct type, and the caller will bitcast the function to the correct
511 const CGFunctionInfo &
512 CodeGenTypes::arrangeMSMemberPointerThunk(const CXXMethodDecl *MD) {
513 assert(MD->isVirtual() && "only virtual memptrs have thunks");
514 CanQual<FunctionProtoType> FTP = GetFormalType(MD);
515 CanQualType ArgTys[] = { GetThisType(Context, MD->getParent()) };
516 return arrangeLLVMFunctionInfo(Context.VoidTy, /*instanceMethod=*/false,
517 /*chainCall=*/false, ArgTys,
518 FTP->getExtInfo(), {}, RequiredArgs(1));
521 const CGFunctionInfo &
522 CodeGenTypes::arrangeMSCtorClosure(const CXXConstructorDecl *CD,
524 assert(CT == Ctor_CopyingClosure || CT == Ctor_DefaultClosure);
526 CanQual<FunctionProtoType> FTP = GetFormalType(CD);
527 SmallVector<CanQualType, 2> ArgTys;
528 const CXXRecordDecl *RD = CD->getParent();
529 ArgTys.push_back(GetThisType(Context, RD));
530 if (CT == Ctor_CopyingClosure)
531 ArgTys.push_back(*FTP->param_type_begin());
532 if (RD->getNumVBases() > 0)
533 ArgTys.push_back(Context.IntTy);
534 CallingConv CC = Context.getDefaultCallingConvention(
535 /*IsVariadic=*/false, /*IsCXXMethod=*/true);
536 return arrangeLLVMFunctionInfo(Context.VoidTy, /*instanceMethod=*/true,
537 /*chainCall=*/false, ArgTys,
538 FunctionType::ExtInfo(CC), {},
542 /// Arrange a call as unto a free function, except possibly with an
543 /// additional number of formal parameters considered required.
544 static const CGFunctionInfo &
545 arrangeFreeFunctionLikeCall(CodeGenTypes &CGT,
547 const CallArgList &args,
548 const FunctionType *fnType,
549 unsigned numExtraRequiredArgs,
551 assert(args.size() >= numExtraRequiredArgs);
553 llvm::SmallVector<FunctionProtoType::ExtParameterInfo, 16> paramInfos;
555 // In most cases, there are no optional arguments.
556 RequiredArgs required = RequiredArgs::All;
558 // If we have a variadic prototype, the required arguments are the
559 // extra prefix plus the arguments in the prototype.
560 if (const FunctionProtoType *proto = dyn_cast<FunctionProtoType>(fnType)) {
561 if (proto->isVariadic())
562 required = RequiredArgs(proto->getNumParams() + numExtraRequiredArgs);
564 if (proto->hasExtParameterInfos())
565 addExtParameterInfosForCall(paramInfos, proto, numExtraRequiredArgs,
568 // If we don't have a prototype at all, but we're supposed to
569 // explicitly use the variadic convention for unprototyped calls,
570 // treat all of the arguments as required but preserve the nominal
571 // possibility of variadics.
572 } else if (CGM.getTargetCodeGenInfo()
573 .isNoProtoCallVariadic(args,
574 cast<FunctionNoProtoType>(fnType))) {
575 required = RequiredArgs(args.size());
579 SmallVector<CanQualType, 16> argTypes;
580 for (const auto &arg : args)
581 argTypes.push_back(CGT.getContext().getCanonicalParamType(arg.Ty));
582 return CGT.arrangeLLVMFunctionInfo(GetReturnType(fnType->getReturnType()),
583 /*instanceMethod=*/false, chainCall,
584 argTypes, fnType->getExtInfo(), paramInfos,
588 /// Figure out the rules for calling a function with the given formal
589 /// type using the given arguments. The arguments are necessary
590 /// because the function might be unprototyped, in which case it's
591 /// target-dependent in crazy ways.
592 const CGFunctionInfo &
593 CodeGenTypes::arrangeFreeFunctionCall(const CallArgList &args,
594 const FunctionType *fnType,
596 return arrangeFreeFunctionLikeCall(*this, CGM, args, fnType,
597 chainCall ? 1 : 0, chainCall);
600 /// A block function is essentially a free function with an
601 /// extra implicit argument.
602 const CGFunctionInfo &
603 CodeGenTypes::arrangeBlockFunctionCall(const CallArgList &args,
604 const FunctionType *fnType) {
605 return arrangeFreeFunctionLikeCall(*this, CGM, args, fnType, 1,
606 /*chainCall=*/false);
609 const CGFunctionInfo &
610 CodeGenTypes::arrangeBlockFunctionDeclaration(const FunctionProtoType *proto,
611 const FunctionArgList ¶ms) {
612 auto paramInfos = getExtParameterInfosForCall(proto, 1, params.size());
613 auto argTypes = getArgTypesForDeclaration(Context, params);
615 return arrangeLLVMFunctionInfo(
616 GetReturnType(proto->getReturnType()),
617 /*instanceMethod*/ false, /*chainCall*/ false, argTypes,
618 proto->getExtInfo(), paramInfos,
619 RequiredArgs::forPrototypePlus(proto, 1, nullptr));
622 const CGFunctionInfo &
623 CodeGenTypes::arrangeBuiltinFunctionCall(QualType resultType,
624 const CallArgList &args) {
626 SmallVector<CanQualType, 16> argTypes;
627 for (const auto &Arg : args)
628 argTypes.push_back(Context.getCanonicalParamType(Arg.Ty));
629 return arrangeLLVMFunctionInfo(
630 GetReturnType(resultType), /*instanceMethod=*/false,
631 /*chainCall=*/false, argTypes, FunctionType::ExtInfo(),
632 /*paramInfos=*/ {}, RequiredArgs::All);
635 const CGFunctionInfo &
636 CodeGenTypes::arrangeBuiltinFunctionDeclaration(QualType resultType,
637 const FunctionArgList &args) {
638 auto argTypes = getArgTypesForDeclaration(Context, args);
640 return arrangeLLVMFunctionInfo(
641 GetReturnType(resultType), /*instanceMethod=*/false, /*chainCall=*/false,
642 argTypes, FunctionType::ExtInfo(), {}, RequiredArgs::All);
645 const CGFunctionInfo &
646 CodeGenTypes::arrangeBuiltinFunctionDeclaration(CanQualType resultType,
647 ArrayRef<CanQualType> argTypes) {
648 return arrangeLLVMFunctionInfo(
649 resultType, /*instanceMethod=*/false, /*chainCall=*/false,
650 argTypes, FunctionType::ExtInfo(), {}, RequiredArgs::All);
653 /// Arrange a call to a C++ method, passing the given arguments.
655 /// numPrefixArgs is the number of ABI-specific prefix arguments we have. It
656 /// does not count `this`.
657 const CGFunctionInfo &
658 CodeGenTypes::arrangeCXXMethodCall(const CallArgList &args,
659 const FunctionProtoType *proto,
660 RequiredArgs required,
661 unsigned numPrefixArgs) {
662 assert(numPrefixArgs + 1 <= args.size() &&
663 "Emitting a call with less args than the required prefix?");
664 // Add one to account for `this`. It's a bit awkward here, but we don't count
665 // `this` in similar places elsewhere.
667 getExtParameterInfosForCall(proto, numPrefixArgs + 1, args.size());
670 auto argTypes = getArgTypesForCall(Context, args);
672 FunctionType::ExtInfo info = proto->getExtInfo();
673 return arrangeLLVMFunctionInfo(
674 GetReturnType(proto->getReturnType()), /*instanceMethod=*/true,
675 /*chainCall=*/false, argTypes, info, paramInfos, required);
678 const CGFunctionInfo &CodeGenTypes::arrangeNullaryFunction() {
679 return arrangeLLVMFunctionInfo(
680 getContext().VoidTy, /*instanceMethod=*/false, /*chainCall=*/false,
681 None, FunctionType::ExtInfo(), {}, RequiredArgs::All);
684 const CGFunctionInfo &
685 CodeGenTypes::arrangeCall(const CGFunctionInfo &signature,
686 const CallArgList &args) {
687 assert(signature.arg_size() <= args.size());
688 if (signature.arg_size() == args.size())
691 SmallVector<FunctionProtoType::ExtParameterInfo, 16> paramInfos;
692 auto sigParamInfos = signature.getExtParameterInfos();
693 if (!sigParamInfos.empty()) {
694 paramInfos.append(sigParamInfos.begin(), sigParamInfos.end());
695 paramInfos.resize(args.size());
698 auto argTypes = getArgTypesForCall(Context, args);
700 assert(signature.getRequiredArgs().allowsOptionalArgs());
701 return arrangeLLVMFunctionInfo(signature.getReturnType(),
702 signature.isInstanceMethod(),
703 signature.isChainCall(),
705 signature.getExtInfo(),
707 signature.getRequiredArgs());
710 /// Arrange the argument and result information for an abstract value
711 /// of a given function type. This is the method which all of the
712 /// above functions ultimately defer to.
713 const CGFunctionInfo &
714 CodeGenTypes::arrangeLLVMFunctionInfo(CanQualType resultType,
717 ArrayRef<CanQualType> argTypes,
718 FunctionType::ExtInfo info,
719 ArrayRef<FunctionProtoType::ExtParameterInfo> paramInfos,
720 RequiredArgs required) {
721 assert(std::all_of(argTypes.begin(), argTypes.end(),
722 [](CanQualType T) { return T.isCanonicalAsParam(); }));
724 // Lookup or create unique function info.
725 llvm::FoldingSetNodeID ID;
726 CGFunctionInfo::Profile(ID, instanceMethod, chainCall, info, paramInfos,
727 required, resultType, argTypes);
729 void *insertPos = nullptr;
730 CGFunctionInfo *FI = FunctionInfos.FindNodeOrInsertPos(ID, insertPos);
734 unsigned CC = ClangCallConvToLLVMCallConv(info.getCC());
736 // Construct the function info. We co-allocate the ArgInfos.
737 FI = CGFunctionInfo::create(CC, instanceMethod, chainCall, info,
738 paramInfos, resultType, argTypes, required);
739 FunctionInfos.InsertNode(FI, insertPos);
741 bool inserted = FunctionsBeingProcessed.insert(FI).second;
743 assert(inserted && "Recursively being processed?");
745 // Compute ABI information.
746 if (info.getCC() != CC_Swift) {
747 getABIInfo().computeInfo(*FI);
749 swiftcall::computeABIInfo(CGM, *FI);
752 // Loop over all of the computed argument and return value info. If any of
753 // them are direct or extend without a specified coerce type, specify the
755 ABIArgInfo &retInfo = FI->getReturnInfo();
756 if (retInfo.canHaveCoerceToType() && retInfo.getCoerceToType() == nullptr)
757 retInfo.setCoerceToType(ConvertType(FI->getReturnType()));
759 for (auto &I : FI->arguments())
760 if (I.info.canHaveCoerceToType() && I.info.getCoerceToType() == nullptr)
761 I.info.setCoerceToType(ConvertType(I.type));
763 bool erased = FunctionsBeingProcessed.erase(FI); (void)erased;
764 assert(erased && "Not in set?");
769 CGFunctionInfo *CGFunctionInfo::create(unsigned llvmCC,
772 const FunctionType::ExtInfo &info,
773 ArrayRef<ExtParameterInfo> paramInfos,
774 CanQualType resultType,
775 ArrayRef<CanQualType> argTypes,
776 RequiredArgs required) {
777 assert(paramInfos.empty() || paramInfos.size() == argTypes.size());
780 operator new(totalSizeToAlloc<ArgInfo, ExtParameterInfo>(
781 argTypes.size() + 1, paramInfos.size()));
783 CGFunctionInfo *FI = new(buffer) CGFunctionInfo();
784 FI->CallingConvention = llvmCC;
785 FI->EffectiveCallingConvention = llvmCC;
786 FI->ASTCallingConvention = info.getCC();
787 FI->InstanceMethod = instanceMethod;
788 FI->ChainCall = chainCall;
789 FI->NoReturn = info.getNoReturn();
790 FI->ReturnsRetained = info.getProducesResult();
791 FI->NoCallerSavedRegs = info.getNoCallerSavedRegs();
792 FI->Required = required;
793 FI->HasRegParm = info.getHasRegParm();
794 FI->RegParm = info.getRegParm();
795 FI->ArgStruct = nullptr;
796 FI->ArgStructAlign = 0;
797 FI->NumArgs = argTypes.size();
798 FI->HasExtParameterInfos = !paramInfos.empty();
799 FI->getArgsBuffer()[0].type = resultType;
800 for (unsigned i = 0, e = argTypes.size(); i != e; ++i)
801 FI->getArgsBuffer()[i + 1].type = argTypes[i];
802 for (unsigned i = 0, e = paramInfos.size(); i != e; ++i)
803 FI->getExtParameterInfosBuffer()[i] = paramInfos[i];
810 // ABIArgInfo::Expand implementation.
812 // Specifies the way QualType passed as ABIArgInfo::Expand is expanded.
813 struct TypeExpansion {
814 enum TypeExpansionKind {
815 // Elements of constant arrays are expanded recursively.
817 // Record fields are expanded recursively (but if record is a union, only
818 // the field with the largest size is expanded).
820 // For complex types, real and imaginary parts are expanded recursively.
822 // All other types are not expandable.
826 const TypeExpansionKind Kind;
828 TypeExpansion(TypeExpansionKind K) : Kind(K) {}
829 virtual ~TypeExpansion() {}
832 struct ConstantArrayExpansion : TypeExpansion {
836 ConstantArrayExpansion(QualType EltTy, uint64_t NumElts)
837 : TypeExpansion(TEK_ConstantArray), EltTy(EltTy), NumElts(NumElts) {}
838 static bool classof(const TypeExpansion *TE) {
839 return TE->Kind == TEK_ConstantArray;
843 struct RecordExpansion : TypeExpansion {
844 SmallVector<const CXXBaseSpecifier *, 1> Bases;
846 SmallVector<const FieldDecl *, 1> Fields;
848 RecordExpansion(SmallVector<const CXXBaseSpecifier *, 1> &&Bases,
849 SmallVector<const FieldDecl *, 1> &&Fields)
850 : TypeExpansion(TEK_Record), Bases(std::move(Bases)),
851 Fields(std::move(Fields)) {}
852 static bool classof(const TypeExpansion *TE) {
853 return TE->Kind == TEK_Record;
857 struct ComplexExpansion : TypeExpansion {
860 ComplexExpansion(QualType EltTy) : TypeExpansion(TEK_Complex), EltTy(EltTy) {}
861 static bool classof(const TypeExpansion *TE) {
862 return TE->Kind == TEK_Complex;
866 struct NoExpansion : TypeExpansion {
867 NoExpansion() : TypeExpansion(TEK_None) {}
868 static bool classof(const TypeExpansion *TE) {
869 return TE->Kind == TEK_None;
874 static std::unique_ptr<TypeExpansion>
875 getTypeExpansion(QualType Ty, const ASTContext &Context) {
876 if (const ConstantArrayType *AT = Context.getAsConstantArrayType(Ty)) {
877 return llvm::make_unique<ConstantArrayExpansion>(
878 AT->getElementType(), AT->getSize().getZExtValue());
880 if (const RecordType *RT = Ty->getAs<RecordType>()) {
881 SmallVector<const CXXBaseSpecifier *, 1> Bases;
882 SmallVector<const FieldDecl *, 1> Fields;
883 const RecordDecl *RD = RT->getDecl();
884 assert(!RD->hasFlexibleArrayMember() &&
885 "Cannot expand structure with flexible array.");
887 // Unions can be here only in degenerative cases - all the fields are same
888 // after flattening. Thus we have to use the "largest" field.
889 const FieldDecl *LargestFD = nullptr;
890 CharUnits UnionSize = CharUnits::Zero();
892 for (const auto *FD : RD->fields()) {
893 // Skip zero length bitfields.
894 if (FD->isBitField() && FD->getBitWidthValue(Context) == 0)
896 assert(!FD->isBitField() &&
897 "Cannot expand structure with bit-field members.");
898 CharUnits FieldSize = Context.getTypeSizeInChars(FD->getType());
899 if (UnionSize < FieldSize) {
900 UnionSize = FieldSize;
905 Fields.push_back(LargestFD);
907 if (const auto *CXXRD = dyn_cast<CXXRecordDecl>(RD)) {
908 assert(!CXXRD->isDynamicClass() &&
909 "cannot expand vtable pointers in dynamic classes");
910 for (const CXXBaseSpecifier &BS : CXXRD->bases())
911 Bases.push_back(&BS);
914 for (const auto *FD : RD->fields()) {
915 // Skip zero length bitfields.
916 if (FD->isBitField() && FD->getBitWidthValue(Context) == 0)
918 assert(!FD->isBitField() &&
919 "Cannot expand structure with bit-field members.");
920 Fields.push_back(FD);
923 return llvm::make_unique<RecordExpansion>(std::move(Bases),
926 if (const ComplexType *CT = Ty->getAs<ComplexType>()) {
927 return llvm::make_unique<ComplexExpansion>(CT->getElementType());
929 return llvm::make_unique<NoExpansion>();
932 static int getExpansionSize(QualType Ty, const ASTContext &Context) {
933 auto Exp = getTypeExpansion(Ty, Context);
934 if (auto CAExp = dyn_cast<ConstantArrayExpansion>(Exp.get())) {
935 return CAExp->NumElts * getExpansionSize(CAExp->EltTy, Context);
937 if (auto RExp = dyn_cast<RecordExpansion>(Exp.get())) {
939 for (auto BS : RExp->Bases)
940 Res += getExpansionSize(BS->getType(), Context);
941 for (auto FD : RExp->Fields)
942 Res += getExpansionSize(FD->getType(), Context);
945 if (isa<ComplexExpansion>(Exp.get()))
947 assert(isa<NoExpansion>(Exp.get()));
952 CodeGenTypes::getExpandedTypes(QualType Ty,
953 SmallVectorImpl<llvm::Type *>::iterator &TI) {
954 auto Exp = getTypeExpansion(Ty, Context);
955 if (auto CAExp = dyn_cast<ConstantArrayExpansion>(Exp.get())) {
956 for (int i = 0, n = CAExp->NumElts; i < n; i++) {
957 getExpandedTypes(CAExp->EltTy, TI);
959 } else if (auto RExp = dyn_cast<RecordExpansion>(Exp.get())) {
960 for (auto BS : RExp->Bases)
961 getExpandedTypes(BS->getType(), TI);
962 for (auto FD : RExp->Fields)
963 getExpandedTypes(FD->getType(), TI);
964 } else if (auto CExp = dyn_cast<ComplexExpansion>(Exp.get())) {
965 llvm::Type *EltTy = ConvertType(CExp->EltTy);
969 assert(isa<NoExpansion>(Exp.get()));
970 *TI++ = ConvertType(Ty);
974 static void forConstantArrayExpansion(CodeGenFunction &CGF,
975 ConstantArrayExpansion *CAE,
977 llvm::function_ref<void(Address)> Fn) {
978 CharUnits EltSize = CGF.getContext().getTypeSizeInChars(CAE->EltTy);
980 BaseAddr.getAlignment().alignmentOfArrayElement(EltSize);
982 for (int i = 0, n = CAE->NumElts; i < n; i++) {
983 llvm::Value *EltAddr =
984 CGF.Builder.CreateConstGEP2_32(nullptr, BaseAddr.getPointer(), 0, i);
985 Fn(Address(EltAddr, EltAlign));
989 void CodeGenFunction::ExpandTypeFromArgs(
990 QualType Ty, LValue LV, SmallVectorImpl<llvm::Value *>::iterator &AI) {
991 assert(LV.isSimple() &&
992 "Unexpected non-simple lvalue during struct expansion.");
994 auto Exp = getTypeExpansion(Ty, getContext());
995 if (auto CAExp = dyn_cast<ConstantArrayExpansion>(Exp.get())) {
996 forConstantArrayExpansion(*this, CAExp, LV.getAddress(),
997 [&](Address EltAddr) {
998 LValue LV = MakeAddrLValue(EltAddr, CAExp->EltTy);
999 ExpandTypeFromArgs(CAExp->EltTy, LV, AI);
1001 } else if (auto RExp = dyn_cast<RecordExpansion>(Exp.get())) {
1002 Address This = LV.getAddress();
1003 for (const CXXBaseSpecifier *BS : RExp->Bases) {
1004 // Perform a single step derived-to-base conversion.
1006 GetAddressOfBaseClass(This, Ty->getAsCXXRecordDecl(), &BS, &BS + 1,
1007 /*NullCheckValue=*/false, SourceLocation());
1008 LValue SubLV = MakeAddrLValue(Base, BS->getType());
1010 // Recurse onto bases.
1011 ExpandTypeFromArgs(BS->getType(), SubLV, AI);
1013 for (auto FD : RExp->Fields) {
1014 // FIXME: What are the right qualifiers here?
1015 LValue SubLV = EmitLValueForFieldInitialization(LV, FD);
1016 ExpandTypeFromArgs(FD->getType(), SubLV, AI);
1018 } else if (isa<ComplexExpansion>(Exp.get())) {
1019 auto realValue = *AI++;
1020 auto imagValue = *AI++;
1021 EmitStoreOfComplex(ComplexPairTy(realValue, imagValue), LV, /*init*/ true);
1023 assert(isa<NoExpansion>(Exp.get()));
1024 EmitStoreThroughLValue(RValue::get(*AI++), LV);
1028 void CodeGenFunction::ExpandTypeToArgs(
1029 QualType Ty, RValue RV, llvm::FunctionType *IRFuncTy,
1030 SmallVectorImpl<llvm::Value *> &IRCallArgs, unsigned &IRCallArgPos) {
1031 auto Exp = getTypeExpansion(Ty, getContext());
1032 if (auto CAExp = dyn_cast<ConstantArrayExpansion>(Exp.get())) {
1033 forConstantArrayExpansion(*this, CAExp, RV.getAggregateAddress(),
1034 [&](Address EltAddr) {
1036 convertTempToRValue(EltAddr, CAExp->EltTy, SourceLocation());
1037 ExpandTypeToArgs(CAExp->EltTy, EltRV, IRFuncTy, IRCallArgs, IRCallArgPos);
1039 } else if (auto RExp = dyn_cast<RecordExpansion>(Exp.get())) {
1040 Address This = RV.getAggregateAddress();
1041 for (const CXXBaseSpecifier *BS : RExp->Bases) {
1042 // Perform a single step derived-to-base conversion.
1044 GetAddressOfBaseClass(This, Ty->getAsCXXRecordDecl(), &BS, &BS + 1,
1045 /*NullCheckValue=*/false, SourceLocation());
1046 RValue BaseRV = RValue::getAggregate(Base);
1048 // Recurse onto bases.
1049 ExpandTypeToArgs(BS->getType(), BaseRV, IRFuncTy, IRCallArgs,
1053 LValue LV = MakeAddrLValue(This, Ty);
1054 for (auto FD : RExp->Fields) {
1055 RValue FldRV = EmitRValueForField(LV, FD, SourceLocation());
1056 ExpandTypeToArgs(FD->getType(), FldRV, IRFuncTy, IRCallArgs,
1059 } else if (isa<ComplexExpansion>(Exp.get())) {
1060 ComplexPairTy CV = RV.getComplexVal();
1061 IRCallArgs[IRCallArgPos++] = CV.first;
1062 IRCallArgs[IRCallArgPos++] = CV.second;
1064 assert(isa<NoExpansion>(Exp.get()));
1065 assert(RV.isScalar() &&
1066 "Unexpected non-scalar rvalue during struct expansion.");
1068 // Insert a bitcast as needed.
1069 llvm::Value *V = RV.getScalarVal();
1070 if (IRCallArgPos < IRFuncTy->getNumParams() &&
1071 V->getType() != IRFuncTy->getParamType(IRCallArgPos))
1072 V = Builder.CreateBitCast(V, IRFuncTy->getParamType(IRCallArgPos));
1074 IRCallArgs[IRCallArgPos++] = V;
1078 /// Create a temporary allocation for the purposes of coercion.
1079 static Address CreateTempAllocaForCoercion(CodeGenFunction &CGF, llvm::Type *Ty,
1080 CharUnits MinAlign) {
1081 // Don't use an alignment that's worse than what LLVM would prefer.
1082 auto PrefAlign = CGF.CGM.getDataLayout().getPrefTypeAlignment(Ty);
1083 CharUnits Align = std::max(MinAlign, CharUnits::fromQuantity(PrefAlign));
1085 return CGF.CreateTempAlloca(Ty, Align);
1088 /// EnterStructPointerForCoercedAccess - Given a struct pointer that we are
1089 /// accessing some number of bytes out of it, try to gep into the struct to get
1090 /// at its inner goodness. Dive as deep as possible without entering an element
1091 /// with an in-memory size smaller than DstSize.
1093 EnterStructPointerForCoercedAccess(Address SrcPtr,
1094 llvm::StructType *SrcSTy,
1095 uint64_t DstSize, CodeGenFunction &CGF) {
1096 // We can't dive into a zero-element struct.
1097 if (SrcSTy->getNumElements() == 0) return SrcPtr;
1099 llvm::Type *FirstElt = SrcSTy->getElementType(0);
1101 // If the first elt is at least as large as what we're looking for, or if the
1102 // first element is the same size as the whole struct, we can enter it. The
1103 // comparison must be made on the store size and not the alloca size. Using
1104 // the alloca size may overstate the size of the load.
1105 uint64_t FirstEltSize =
1106 CGF.CGM.getDataLayout().getTypeStoreSize(FirstElt);
1107 if (FirstEltSize < DstSize &&
1108 FirstEltSize < CGF.CGM.getDataLayout().getTypeStoreSize(SrcSTy))
1111 // GEP into the first element.
1112 SrcPtr = CGF.Builder.CreateStructGEP(SrcPtr, 0, CharUnits(), "coerce.dive");
1114 // If the first element is a struct, recurse.
1115 llvm::Type *SrcTy = SrcPtr.getElementType();
1116 if (llvm::StructType *SrcSTy = dyn_cast<llvm::StructType>(SrcTy))
1117 return EnterStructPointerForCoercedAccess(SrcPtr, SrcSTy, DstSize, CGF);
1122 /// CoerceIntOrPtrToIntOrPtr - Convert a value Val to the specific Ty where both
1123 /// are either integers or pointers. This does a truncation of the value if it
1124 /// is too large or a zero extension if it is too small.
1126 /// This behaves as if the value were coerced through memory, so on big-endian
1127 /// targets the high bits are preserved in a truncation, while little-endian
1128 /// targets preserve the low bits.
1129 static llvm::Value *CoerceIntOrPtrToIntOrPtr(llvm::Value *Val,
1131 CodeGenFunction &CGF) {
1132 if (Val->getType() == Ty)
1135 if (isa<llvm::PointerType>(Val->getType())) {
1136 // If this is Pointer->Pointer avoid conversion to and from int.
1137 if (isa<llvm::PointerType>(Ty))
1138 return CGF.Builder.CreateBitCast(Val, Ty, "coerce.val");
1140 // Convert the pointer to an integer so we can play with its width.
1141 Val = CGF.Builder.CreatePtrToInt(Val, CGF.IntPtrTy, "coerce.val.pi");
1144 llvm::Type *DestIntTy = Ty;
1145 if (isa<llvm::PointerType>(DestIntTy))
1146 DestIntTy = CGF.IntPtrTy;
1148 if (Val->getType() != DestIntTy) {
1149 const llvm::DataLayout &DL = CGF.CGM.getDataLayout();
1150 if (DL.isBigEndian()) {
1151 // Preserve the high bits on big-endian targets.
1152 // That is what memory coercion does.
1153 uint64_t SrcSize = DL.getTypeSizeInBits(Val->getType());
1154 uint64_t DstSize = DL.getTypeSizeInBits(DestIntTy);
1156 if (SrcSize > DstSize) {
1157 Val = CGF.Builder.CreateLShr(Val, SrcSize - DstSize, "coerce.highbits");
1158 Val = CGF.Builder.CreateTrunc(Val, DestIntTy, "coerce.val.ii");
1160 Val = CGF.Builder.CreateZExt(Val, DestIntTy, "coerce.val.ii");
1161 Val = CGF.Builder.CreateShl(Val, DstSize - SrcSize, "coerce.highbits");
1164 // Little-endian targets preserve the low bits. No shifts required.
1165 Val = CGF.Builder.CreateIntCast(Val, DestIntTy, false, "coerce.val.ii");
1169 if (isa<llvm::PointerType>(Ty))
1170 Val = CGF.Builder.CreateIntToPtr(Val, Ty, "coerce.val.ip");
1176 /// CreateCoercedLoad - Create a load from \arg SrcPtr interpreted as
1177 /// a pointer to an object of type \arg Ty, known to be aligned to
1178 /// \arg SrcAlign bytes.
1180 /// This safely handles the case when the src type is smaller than the
1181 /// destination type; in this situation the values of bits which not
1182 /// present in the src are undefined.
1183 static llvm::Value *CreateCoercedLoad(Address Src, llvm::Type *Ty,
1184 CodeGenFunction &CGF) {
1185 llvm::Type *SrcTy = Src.getElementType();
1187 // If SrcTy and Ty are the same, just do a load.
1189 return CGF.Builder.CreateLoad(Src);
1191 uint64_t DstSize = CGF.CGM.getDataLayout().getTypeAllocSize(Ty);
1193 if (llvm::StructType *SrcSTy = dyn_cast<llvm::StructType>(SrcTy)) {
1194 Src = EnterStructPointerForCoercedAccess(Src, SrcSTy, DstSize, CGF);
1195 SrcTy = Src.getType()->getElementType();
1198 uint64_t SrcSize = CGF.CGM.getDataLayout().getTypeAllocSize(SrcTy);
1200 // If the source and destination are integer or pointer types, just do an
1201 // extension or truncation to the desired type.
1202 if ((isa<llvm::IntegerType>(Ty) || isa<llvm::PointerType>(Ty)) &&
1203 (isa<llvm::IntegerType>(SrcTy) || isa<llvm::PointerType>(SrcTy))) {
1204 llvm::Value *Load = CGF.Builder.CreateLoad(Src);
1205 return CoerceIntOrPtrToIntOrPtr(Load, Ty, CGF);
1208 // If load is legal, just bitcast the src pointer.
1209 if (SrcSize >= DstSize) {
1210 // Generally SrcSize is never greater than DstSize, since this means we are
1211 // losing bits. However, this can happen in cases where the structure has
1212 // additional padding, for example due to a user specified alignment.
1214 // FIXME: Assert that we aren't truncating non-padding bits when have access
1215 // to that information.
1216 Src = CGF.Builder.CreateBitCast(Src, llvm::PointerType::getUnqual(Ty));
1217 return CGF.Builder.CreateLoad(Src);
1220 // Otherwise do coercion through memory. This is stupid, but simple.
1221 Address Tmp = CreateTempAllocaForCoercion(CGF, Ty, Src.getAlignment());
1222 Address Casted = CGF.Builder.CreateBitCast(Tmp, CGF.Int8PtrTy);
1223 Address SrcCasted = CGF.Builder.CreateBitCast(Src, CGF.Int8PtrTy);
1224 CGF.Builder.CreateMemCpy(Casted, SrcCasted,
1225 llvm::ConstantInt::get(CGF.IntPtrTy, SrcSize),
1227 return CGF.Builder.CreateLoad(Tmp);
1230 // Function to store a first-class aggregate into memory. We prefer to
1231 // store the elements rather than the aggregate to be more friendly to
1233 // FIXME: Do we need to recurse here?
1234 static void BuildAggStore(CodeGenFunction &CGF, llvm::Value *Val,
1235 Address Dest, bool DestIsVolatile) {
1236 // Prefer scalar stores to first-class aggregate stores.
1237 if (llvm::StructType *STy =
1238 dyn_cast<llvm::StructType>(Val->getType())) {
1239 const llvm::StructLayout *Layout =
1240 CGF.CGM.getDataLayout().getStructLayout(STy);
1242 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1243 auto EltOffset = CharUnits::fromQuantity(Layout->getElementOffset(i));
1244 Address EltPtr = CGF.Builder.CreateStructGEP(Dest, i, EltOffset);
1245 llvm::Value *Elt = CGF.Builder.CreateExtractValue(Val, i);
1246 CGF.Builder.CreateStore(Elt, EltPtr, DestIsVolatile);
1249 CGF.Builder.CreateStore(Val, Dest, DestIsVolatile);
1253 /// CreateCoercedStore - Create a store to \arg DstPtr from \arg Src,
1254 /// where the source and destination may have different types. The
1255 /// destination is known to be aligned to \arg DstAlign bytes.
1257 /// This safely handles the case when the src type is larger than the
1258 /// destination type; the upper bits of the src will be lost.
1259 static void CreateCoercedStore(llvm::Value *Src,
1262 CodeGenFunction &CGF) {
1263 llvm::Type *SrcTy = Src->getType();
1264 llvm::Type *DstTy = Dst.getType()->getElementType();
1265 if (SrcTy == DstTy) {
1266 CGF.Builder.CreateStore(Src, Dst, DstIsVolatile);
1270 uint64_t SrcSize = CGF.CGM.getDataLayout().getTypeAllocSize(SrcTy);
1272 if (llvm::StructType *DstSTy = dyn_cast<llvm::StructType>(DstTy)) {
1273 Dst = EnterStructPointerForCoercedAccess(Dst, DstSTy, SrcSize, CGF);
1274 DstTy = Dst.getType()->getElementType();
1277 // If the source and destination are integer or pointer types, just do an
1278 // extension or truncation to the desired type.
1279 if ((isa<llvm::IntegerType>(SrcTy) || isa<llvm::PointerType>(SrcTy)) &&
1280 (isa<llvm::IntegerType>(DstTy) || isa<llvm::PointerType>(DstTy))) {
1281 Src = CoerceIntOrPtrToIntOrPtr(Src, DstTy, CGF);
1282 CGF.Builder.CreateStore(Src, Dst, DstIsVolatile);
1286 uint64_t DstSize = CGF.CGM.getDataLayout().getTypeAllocSize(DstTy);
1288 // If store is legal, just bitcast the src pointer.
1289 if (SrcSize <= DstSize) {
1290 Dst = CGF.Builder.CreateBitCast(Dst, llvm::PointerType::getUnqual(SrcTy));
1291 BuildAggStore(CGF, Src, Dst, DstIsVolatile);
1293 // Otherwise do coercion through memory. This is stupid, but
1296 // Generally SrcSize is never greater than DstSize, since this means we are
1297 // losing bits. However, this can happen in cases where the structure has
1298 // additional padding, for example due to a user specified alignment.
1300 // FIXME: Assert that we aren't truncating non-padding bits when have access
1301 // to that information.
1302 Address Tmp = CreateTempAllocaForCoercion(CGF, SrcTy, Dst.getAlignment());
1303 CGF.Builder.CreateStore(Src, Tmp);
1304 Address Casted = CGF.Builder.CreateBitCast(Tmp, CGF.Int8PtrTy);
1305 Address DstCasted = CGF.Builder.CreateBitCast(Dst, CGF.Int8PtrTy);
1306 CGF.Builder.CreateMemCpy(DstCasted, Casted,
1307 llvm::ConstantInt::get(CGF.IntPtrTy, DstSize),
1312 static Address emitAddressAtOffset(CodeGenFunction &CGF, Address addr,
1313 const ABIArgInfo &info) {
1314 if (unsigned offset = info.getDirectOffset()) {
1315 addr = CGF.Builder.CreateElementBitCast(addr, CGF.Int8Ty);
1316 addr = CGF.Builder.CreateConstInBoundsByteGEP(addr,
1317 CharUnits::fromQuantity(offset));
1318 addr = CGF.Builder.CreateElementBitCast(addr, info.getCoerceToType());
1325 /// Encapsulates information about the way function arguments from
1326 /// CGFunctionInfo should be passed to actual LLVM IR function.
1327 class ClangToLLVMArgMapping {
1328 static const unsigned InvalidIndex = ~0U;
1329 unsigned InallocaArgNo;
1331 unsigned TotalIRArgs;
1333 /// Arguments of LLVM IR function corresponding to single Clang argument.
1335 unsigned PaddingArgIndex;
1336 // Argument is expanded to IR arguments at positions
1337 // [FirstArgIndex, FirstArgIndex + NumberOfArgs).
1338 unsigned FirstArgIndex;
1339 unsigned NumberOfArgs;
1342 : PaddingArgIndex(InvalidIndex), FirstArgIndex(InvalidIndex),
1346 SmallVector<IRArgs, 8> ArgInfo;
1349 ClangToLLVMArgMapping(const ASTContext &Context, const CGFunctionInfo &FI,
1350 bool OnlyRequiredArgs = false)
1351 : InallocaArgNo(InvalidIndex), SRetArgNo(InvalidIndex), TotalIRArgs(0),
1352 ArgInfo(OnlyRequiredArgs ? FI.getNumRequiredArgs() : FI.arg_size()) {
1353 construct(Context, FI, OnlyRequiredArgs);
1356 bool hasInallocaArg() const { return InallocaArgNo != InvalidIndex; }
1357 unsigned getInallocaArgNo() const {
1358 assert(hasInallocaArg());
1359 return InallocaArgNo;
1362 bool hasSRetArg() const { return SRetArgNo != InvalidIndex; }
1363 unsigned getSRetArgNo() const {
1364 assert(hasSRetArg());
1368 unsigned totalIRArgs() const { return TotalIRArgs; }
1370 bool hasPaddingArg(unsigned ArgNo) const {
1371 assert(ArgNo < ArgInfo.size());
1372 return ArgInfo[ArgNo].PaddingArgIndex != InvalidIndex;
1374 unsigned getPaddingArgNo(unsigned ArgNo) const {
1375 assert(hasPaddingArg(ArgNo));
1376 return ArgInfo[ArgNo].PaddingArgIndex;
1379 /// Returns index of first IR argument corresponding to ArgNo, and their
1381 std::pair<unsigned, unsigned> getIRArgs(unsigned ArgNo) const {
1382 assert(ArgNo < ArgInfo.size());
1383 return std::make_pair(ArgInfo[ArgNo].FirstArgIndex,
1384 ArgInfo[ArgNo].NumberOfArgs);
1388 void construct(const ASTContext &Context, const CGFunctionInfo &FI,
1389 bool OnlyRequiredArgs);
1392 void ClangToLLVMArgMapping::construct(const ASTContext &Context,
1393 const CGFunctionInfo &FI,
1394 bool OnlyRequiredArgs) {
1395 unsigned IRArgNo = 0;
1396 bool SwapThisWithSRet = false;
1397 const ABIArgInfo &RetAI = FI.getReturnInfo();
1399 if (RetAI.getKind() == ABIArgInfo::Indirect) {
1400 SwapThisWithSRet = RetAI.isSRetAfterThis();
1401 SRetArgNo = SwapThisWithSRet ? 1 : IRArgNo++;
1405 unsigned NumArgs = OnlyRequiredArgs ? FI.getNumRequiredArgs() : FI.arg_size();
1406 for (CGFunctionInfo::const_arg_iterator I = FI.arg_begin(); ArgNo < NumArgs;
1408 assert(I != FI.arg_end());
1409 QualType ArgType = I->type;
1410 const ABIArgInfo &AI = I->info;
1411 // Collect data about IR arguments corresponding to Clang argument ArgNo.
1412 auto &IRArgs = ArgInfo[ArgNo];
1414 if (AI.getPaddingType())
1415 IRArgs.PaddingArgIndex = IRArgNo++;
1417 switch (AI.getKind()) {
1418 case ABIArgInfo::Extend:
1419 case ABIArgInfo::Direct: {
1420 // FIXME: handle sseregparm someday...
1421 llvm::StructType *STy = dyn_cast<llvm::StructType>(AI.getCoerceToType());
1422 if (AI.isDirect() && AI.getCanBeFlattened() && STy) {
1423 IRArgs.NumberOfArgs = STy->getNumElements();
1425 IRArgs.NumberOfArgs = 1;
1429 case ABIArgInfo::Indirect:
1430 IRArgs.NumberOfArgs = 1;
1432 case ABIArgInfo::Ignore:
1433 case ABIArgInfo::InAlloca:
1434 // ignore and inalloca doesn't have matching LLVM parameters.
1435 IRArgs.NumberOfArgs = 0;
1437 case ABIArgInfo::CoerceAndExpand:
1438 IRArgs.NumberOfArgs = AI.getCoerceAndExpandTypeSequence().size();
1440 case ABIArgInfo::Expand:
1441 IRArgs.NumberOfArgs = getExpansionSize(ArgType, Context);
1445 if (IRArgs.NumberOfArgs > 0) {
1446 IRArgs.FirstArgIndex = IRArgNo;
1447 IRArgNo += IRArgs.NumberOfArgs;
1450 // Skip over the sret parameter when it comes second. We already handled it
1452 if (IRArgNo == 1 && SwapThisWithSRet)
1455 assert(ArgNo == ArgInfo.size());
1457 if (FI.usesInAlloca())
1458 InallocaArgNo = IRArgNo++;
1460 TotalIRArgs = IRArgNo;
1466 bool CodeGenModule::ReturnTypeUsesSRet(const CGFunctionInfo &FI) {
1467 return FI.getReturnInfo().isIndirect();
1470 bool CodeGenModule::ReturnSlotInterferesWithArgs(const CGFunctionInfo &FI) {
1471 return ReturnTypeUsesSRet(FI) &&
1472 getTargetCodeGenInfo().doesReturnSlotInterfereWithArgs();
1475 bool CodeGenModule::ReturnTypeUsesFPRet(QualType ResultType) {
1476 if (const BuiltinType *BT = ResultType->getAs<BuiltinType>()) {
1477 switch (BT->getKind()) {
1480 case BuiltinType::Float:
1481 return getTarget().useObjCFPRetForRealType(TargetInfo::Float);
1482 case BuiltinType::Double:
1483 return getTarget().useObjCFPRetForRealType(TargetInfo::Double);
1484 case BuiltinType::LongDouble:
1485 return getTarget().useObjCFPRetForRealType(TargetInfo::LongDouble);
1492 bool CodeGenModule::ReturnTypeUsesFP2Ret(QualType ResultType) {
1493 if (const ComplexType *CT = ResultType->getAs<ComplexType>()) {
1494 if (const BuiltinType *BT = CT->getElementType()->getAs<BuiltinType>()) {
1495 if (BT->getKind() == BuiltinType::LongDouble)
1496 return getTarget().useObjCFP2RetForComplexLongDouble();
1503 llvm::FunctionType *CodeGenTypes::GetFunctionType(GlobalDecl GD) {
1504 const CGFunctionInfo &FI = arrangeGlobalDeclaration(GD);
1505 return GetFunctionType(FI);
1508 llvm::FunctionType *
1509 CodeGenTypes::GetFunctionType(const CGFunctionInfo &FI) {
1511 bool Inserted = FunctionsBeingProcessed.insert(&FI).second;
1513 assert(Inserted && "Recursively being processed?");
1515 llvm::Type *resultType = nullptr;
1516 const ABIArgInfo &retAI = FI.getReturnInfo();
1517 switch (retAI.getKind()) {
1518 case ABIArgInfo::Expand:
1519 llvm_unreachable("Invalid ABI kind for return argument");
1521 case ABIArgInfo::Extend:
1522 case ABIArgInfo::Direct:
1523 resultType = retAI.getCoerceToType();
1526 case ABIArgInfo::InAlloca:
1527 if (retAI.getInAllocaSRet()) {
1528 // sret things on win32 aren't void, they return the sret pointer.
1529 QualType ret = FI.getReturnType();
1530 llvm::Type *ty = ConvertType(ret);
1531 unsigned addressSpace = Context.getTargetAddressSpace(ret);
1532 resultType = llvm::PointerType::get(ty, addressSpace);
1534 resultType = llvm::Type::getVoidTy(getLLVMContext());
1538 case ABIArgInfo::Indirect:
1539 case ABIArgInfo::Ignore:
1540 resultType = llvm::Type::getVoidTy(getLLVMContext());
1543 case ABIArgInfo::CoerceAndExpand:
1544 resultType = retAI.getUnpaddedCoerceAndExpandType();
1548 ClangToLLVMArgMapping IRFunctionArgs(getContext(), FI, true);
1549 SmallVector<llvm::Type*, 8> ArgTypes(IRFunctionArgs.totalIRArgs());
1551 // Add type for sret argument.
1552 if (IRFunctionArgs.hasSRetArg()) {
1553 QualType Ret = FI.getReturnType();
1554 llvm::Type *Ty = ConvertType(Ret);
1555 unsigned AddressSpace = Context.getTargetAddressSpace(Ret);
1556 ArgTypes[IRFunctionArgs.getSRetArgNo()] =
1557 llvm::PointerType::get(Ty, AddressSpace);
1560 // Add type for inalloca argument.
1561 if (IRFunctionArgs.hasInallocaArg()) {
1562 auto ArgStruct = FI.getArgStruct();
1564 ArgTypes[IRFunctionArgs.getInallocaArgNo()] = ArgStruct->getPointerTo();
1567 // Add in all of the required arguments.
1569 CGFunctionInfo::const_arg_iterator it = FI.arg_begin(),
1570 ie = it + FI.getNumRequiredArgs();
1571 for (; it != ie; ++it, ++ArgNo) {
1572 const ABIArgInfo &ArgInfo = it->info;
1574 // Insert a padding type to ensure proper alignment.
1575 if (IRFunctionArgs.hasPaddingArg(ArgNo))
1576 ArgTypes[IRFunctionArgs.getPaddingArgNo(ArgNo)] =
1577 ArgInfo.getPaddingType();
1579 unsigned FirstIRArg, NumIRArgs;
1580 std::tie(FirstIRArg, NumIRArgs) = IRFunctionArgs.getIRArgs(ArgNo);
1582 switch (ArgInfo.getKind()) {
1583 case ABIArgInfo::Ignore:
1584 case ABIArgInfo::InAlloca:
1585 assert(NumIRArgs == 0);
1588 case ABIArgInfo::Indirect: {
1589 assert(NumIRArgs == 1);
1590 // indirect arguments are always on the stack, which is alloca addr space.
1591 llvm::Type *LTy = ConvertTypeForMem(it->type);
1592 ArgTypes[FirstIRArg] = LTy->getPointerTo(
1593 CGM.getDataLayout().getAllocaAddrSpace());
1597 case ABIArgInfo::Extend:
1598 case ABIArgInfo::Direct: {
1599 // Fast-isel and the optimizer generally like scalar values better than
1600 // FCAs, so we flatten them if this is safe to do for this argument.
1601 llvm::Type *argType = ArgInfo.getCoerceToType();
1602 llvm::StructType *st = dyn_cast<llvm::StructType>(argType);
1603 if (st && ArgInfo.isDirect() && ArgInfo.getCanBeFlattened()) {
1604 assert(NumIRArgs == st->getNumElements());
1605 for (unsigned i = 0, e = st->getNumElements(); i != e; ++i)
1606 ArgTypes[FirstIRArg + i] = st->getElementType(i);
1608 assert(NumIRArgs == 1);
1609 ArgTypes[FirstIRArg] = argType;
1614 case ABIArgInfo::CoerceAndExpand: {
1615 auto ArgTypesIter = ArgTypes.begin() + FirstIRArg;
1616 for (auto EltTy : ArgInfo.getCoerceAndExpandTypeSequence()) {
1617 *ArgTypesIter++ = EltTy;
1619 assert(ArgTypesIter == ArgTypes.begin() + FirstIRArg + NumIRArgs);
1623 case ABIArgInfo::Expand:
1624 auto ArgTypesIter = ArgTypes.begin() + FirstIRArg;
1625 getExpandedTypes(it->type, ArgTypesIter);
1626 assert(ArgTypesIter == ArgTypes.begin() + FirstIRArg + NumIRArgs);
1631 bool Erased = FunctionsBeingProcessed.erase(&FI); (void)Erased;
1632 assert(Erased && "Not in set?");
1634 return llvm::FunctionType::get(resultType, ArgTypes, FI.isVariadic());
1637 llvm::Type *CodeGenTypes::GetFunctionTypeForVTable(GlobalDecl GD) {
1638 const CXXMethodDecl *MD = cast<CXXMethodDecl>(GD.getDecl());
1639 const FunctionProtoType *FPT = MD->getType()->getAs<FunctionProtoType>();
1641 if (!isFuncTypeConvertible(FPT))
1642 return llvm::StructType::get(getLLVMContext());
1644 const CGFunctionInfo *Info;
1645 if (isa<CXXDestructorDecl>(MD))
1647 &arrangeCXXStructorDeclaration(MD, getFromDtorType(GD.getDtorType()));
1649 Info = &arrangeCXXMethodDeclaration(MD);
1650 return GetFunctionType(*Info);
1653 static void AddAttributesFromFunctionProtoType(ASTContext &Ctx,
1654 llvm::AttrBuilder &FuncAttrs,
1655 const FunctionProtoType *FPT) {
1659 if (!isUnresolvedExceptionSpec(FPT->getExceptionSpecType()) &&
1660 FPT->isNothrow(Ctx))
1661 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1664 void CodeGenModule::ConstructDefaultFnAttrList(StringRef Name, bool HasOptnone,
1665 bool AttrOnCallSite,
1666 llvm::AttrBuilder &FuncAttrs) {
1667 // OptimizeNoneAttr takes precedence over -Os or -Oz. No warning needed.
1669 if (CodeGenOpts.OptimizeSize)
1670 FuncAttrs.addAttribute(llvm::Attribute::OptimizeForSize);
1671 if (CodeGenOpts.OptimizeSize == 2)
1672 FuncAttrs.addAttribute(llvm::Attribute::MinSize);
1675 if (CodeGenOpts.DisableRedZone)
1676 FuncAttrs.addAttribute(llvm::Attribute::NoRedZone);
1677 if (CodeGenOpts.NoImplicitFloat)
1678 FuncAttrs.addAttribute(llvm::Attribute::NoImplicitFloat);
1680 if (AttrOnCallSite) {
1681 // Attributes that should go on the call site only.
1682 if (!CodeGenOpts.SimplifyLibCalls ||
1683 CodeGenOpts.isNoBuiltinFunc(Name.data()))
1684 FuncAttrs.addAttribute(llvm::Attribute::NoBuiltin);
1685 if (!CodeGenOpts.TrapFuncName.empty())
1686 FuncAttrs.addAttribute("trap-func-name", CodeGenOpts.TrapFuncName);
1688 // Attributes that should go on the function, but not the call site.
1689 if (!CodeGenOpts.DisableFPElim) {
1690 FuncAttrs.addAttribute("no-frame-pointer-elim", "false");
1691 } else if (CodeGenOpts.OmitLeafFramePointer) {
1692 FuncAttrs.addAttribute("no-frame-pointer-elim", "false");
1693 FuncAttrs.addAttribute("no-frame-pointer-elim-non-leaf");
1695 FuncAttrs.addAttribute("no-frame-pointer-elim", "true");
1696 FuncAttrs.addAttribute("no-frame-pointer-elim-non-leaf");
1699 FuncAttrs.addAttribute("less-precise-fpmad",
1700 llvm::toStringRef(CodeGenOpts.LessPreciseFPMAD));
1702 if (!CodeGenOpts.FPDenormalMode.empty())
1703 FuncAttrs.addAttribute("denormal-fp-math", CodeGenOpts.FPDenormalMode);
1705 FuncAttrs.addAttribute("no-trapping-math",
1706 llvm::toStringRef(CodeGenOpts.NoTrappingMath));
1708 // TODO: Are these all needed?
1709 // unsafe/inf/nan/nsz are handled by instruction-level FastMathFlags.
1710 FuncAttrs.addAttribute("no-infs-fp-math",
1711 llvm::toStringRef(CodeGenOpts.NoInfsFPMath));
1712 FuncAttrs.addAttribute("no-nans-fp-math",
1713 llvm::toStringRef(CodeGenOpts.NoNaNsFPMath));
1714 FuncAttrs.addAttribute("unsafe-fp-math",
1715 llvm::toStringRef(CodeGenOpts.UnsafeFPMath));
1716 FuncAttrs.addAttribute("use-soft-float",
1717 llvm::toStringRef(CodeGenOpts.SoftFloat));
1718 FuncAttrs.addAttribute("stack-protector-buffer-size",
1719 llvm::utostr(CodeGenOpts.SSPBufferSize));
1720 FuncAttrs.addAttribute("no-signed-zeros-fp-math",
1721 llvm::toStringRef(CodeGenOpts.NoSignedZeros));
1722 FuncAttrs.addAttribute(
1723 "correctly-rounded-divide-sqrt-fp-math",
1724 llvm::toStringRef(CodeGenOpts.CorrectlyRoundedDivSqrt));
1726 // TODO: Reciprocal estimate codegen options should apply to instructions?
1727 std::vector<std::string> &Recips = getTarget().getTargetOpts().Reciprocals;
1728 if (!Recips.empty())
1729 FuncAttrs.addAttribute("reciprocal-estimates",
1730 llvm::join(Recips.begin(), Recips.end(), ","));
1732 if (CodeGenOpts.StackRealignment)
1733 FuncAttrs.addAttribute("stackrealign");
1734 if (CodeGenOpts.Backchain)
1735 FuncAttrs.addAttribute("backchain");
1738 if (getLangOpts().CUDA && getLangOpts().CUDAIsDevice) {
1739 // Conservatively, mark all functions and calls in CUDA as convergent
1740 // (meaning, they may call an intrinsically convergent op, such as
1741 // __syncthreads(), and so can't have certain optimizations applied around
1742 // them). LLVM will remove this attribute where it safely can.
1743 FuncAttrs.addAttribute(llvm::Attribute::Convergent);
1745 // Exceptions aren't supported in CUDA device code.
1746 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1748 // Respect -fcuda-flush-denormals-to-zero.
1749 if (getLangOpts().CUDADeviceFlushDenormalsToZero)
1750 FuncAttrs.addAttribute("nvptx-f32ftz", "true");
1754 void CodeGenModule::AddDefaultFnAttrs(llvm::Function &F) {
1755 llvm::AttrBuilder FuncAttrs;
1756 ConstructDefaultFnAttrList(F.getName(),
1757 F.hasFnAttribute(llvm::Attribute::OptimizeNone),
1758 /* AttrOnCallsite = */ false, FuncAttrs);
1759 F.addAttributes(llvm::AttributeList::FunctionIndex, FuncAttrs);
1762 void CodeGenModule::ConstructAttributeList(
1763 StringRef Name, const CGFunctionInfo &FI, CGCalleeInfo CalleeInfo,
1764 llvm::AttributeList &AttrList, unsigned &CallingConv, bool AttrOnCallSite) {
1765 llvm::AttrBuilder FuncAttrs;
1766 llvm::AttrBuilder RetAttrs;
1768 CallingConv = FI.getEffectiveCallingConvention();
1769 if (FI.isNoReturn())
1770 FuncAttrs.addAttribute(llvm::Attribute::NoReturn);
1772 // If we have information about the function prototype, we can learn
1773 // attributes form there.
1774 AddAttributesFromFunctionProtoType(getContext(), FuncAttrs,
1775 CalleeInfo.getCalleeFunctionProtoType());
1777 const Decl *TargetDecl = CalleeInfo.getCalleeDecl();
1779 bool HasOptnone = false;
1780 // FIXME: handle sseregparm someday...
1782 if (TargetDecl->hasAttr<ReturnsTwiceAttr>())
1783 FuncAttrs.addAttribute(llvm::Attribute::ReturnsTwice);
1784 if (TargetDecl->hasAttr<NoThrowAttr>())
1785 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1786 if (TargetDecl->hasAttr<NoReturnAttr>())
1787 FuncAttrs.addAttribute(llvm::Attribute::NoReturn);
1788 if (TargetDecl->hasAttr<NoDuplicateAttr>())
1789 FuncAttrs.addAttribute(llvm::Attribute::NoDuplicate);
1790 if (TargetDecl->hasAttr<ConvergentAttr>())
1791 FuncAttrs.addAttribute(llvm::Attribute::Convergent);
1793 if (const FunctionDecl *Fn = dyn_cast<FunctionDecl>(TargetDecl)) {
1794 AddAttributesFromFunctionProtoType(
1795 getContext(), FuncAttrs, Fn->getType()->getAs<FunctionProtoType>());
1796 // Don't use [[noreturn]] or _Noreturn for a call to a virtual function.
1797 // These attributes are not inherited by overloads.
1798 const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(Fn);
1799 if (Fn->isNoReturn() && !(AttrOnCallSite && MD && MD->isVirtual()))
1800 FuncAttrs.addAttribute(llvm::Attribute::NoReturn);
1803 // 'const', 'pure' and 'noalias' attributed functions are also nounwind.
1804 if (TargetDecl->hasAttr<ConstAttr>()) {
1805 FuncAttrs.addAttribute(llvm::Attribute::ReadNone);
1806 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1807 } else if (TargetDecl->hasAttr<PureAttr>()) {
1808 FuncAttrs.addAttribute(llvm::Attribute::ReadOnly);
1809 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1810 } else if (TargetDecl->hasAttr<NoAliasAttr>()) {
1811 FuncAttrs.addAttribute(llvm::Attribute::ArgMemOnly);
1812 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1814 if (TargetDecl->hasAttr<RestrictAttr>())
1815 RetAttrs.addAttribute(llvm::Attribute::NoAlias);
1816 if (TargetDecl->hasAttr<ReturnsNonNullAttr>())
1817 RetAttrs.addAttribute(llvm::Attribute::NonNull);
1818 if (TargetDecl->hasAttr<AnyX86NoCallerSavedRegistersAttr>())
1819 FuncAttrs.addAttribute("no_caller_saved_registers");
1821 HasOptnone = TargetDecl->hasAttr<OptimizeNoneAttr>();
1822 if (auto *AllocSize = TargetDecl->getAttr<AllocSizeAttr>()) {
1823 Optional<unsigned> NumElemsParam;
1824 // alloc_size args are base-1, 0 means not present.
1825 if (unsigned N = AllocSize->getNumElemsParam())
1826 NumElemsParam = N - 1;
1827 FuncAttrs.addAllocSizeAttr(AllocSize->getElemSizeParam() - 1,
1832 ConstructDefaultFnAttrList(Name, HasOptnone, AttrOnCallSite, FuncAttrs);
1834 if (CodeGenOpts.EnableSegmentedStacks &&
1835 !(TargetDecl && TargetDecl->hasAttr<NoSplitStackAttr>()))
1836 FuncAttrs.addAttribute("split-stack");
1838 if (!AttrOnCallSite) {
1839 bool DisableTailCalls =
1840 CodeGenOpts.DisableTailCalls ||
1841 (TargetDecl && (TargetDecl->hasAttr<DisableTailCallsAttr>() ||
1842 TargetDecl->hasAttr<AnyX86InterruptAttr>()));
1843 FuncAttrs.addAttribute("disable-tail-calls",
1844 llvm::toStringRef(DisableTailCalls));
1846 // Add target-cpu and target-features attributes to functions. If
1847 // we have a decl for the function and it has a target attribute then
1848 // parse that and add it to the feature set.
1849 StringRef TargetCPU = getTarget().getTargetOpts().CPU;
1850 const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(TargetDecl);
1851 if (FD && FD->hasAttr<TargetAttr>()) {
1852 llvm::StringMap<bool> FeatureMap;
1853 getFunctionFeatureMap(FeatureMap, FD);
1855 // Produce the canonical string for this set of features.
1856 std::vector<std::string> Features;
1857 for (llvm::StringMap<bool>::const_iterator it = FeatureMap.begin(),
1858 ie = FeatureMap.end();
1860 Features.push_back((it->second ? "+" : "-") + it->first().str());
1862 // Now add the target-cpu and target-features to the function.
1863 // While we populated the feature map above, we still need to
1864 // get and parse the target attribute so we can get the cpu for
1866 const auto *TD = FD->getAttr<TargetAttr>();
1867 TargetAttr::ParsedTargetAttr ParsedAttr = TD->parse();
1868 if (ParsedAttr.second != "")
1869 TargetCPU = ParsedAttr.second;
1870 if (TargetCPU != "")
1871 FuncAttrs.addAttribute("target-cpu", TargetCPU);
1872 if (!Features.empty()) {
1873 std::sort(Features.begin(), Features.end());
1874 FuncAttrs.addAttribute(
1876 llvm::join(Features.begin(), Features.end(), ","));
1879 // Otherwise just add the existing target cpu and target features to the
1881 std::vector<std::string> &Features = getTarget().getTargetOpts().Features;
1882 if (TargetCPU != "")
1883 FuncAttrs.addAttribute("target-cpu", TargetCPU);
1884 if (!Features.empty()) {
1885 std::sort(Features.begin(), Features.end());
1886 FuncAttrs.addAttribute(
1888 llvm::join(Features.begin(), Features.end(), ","));
1893 ClangToLLVMArgMapping IRFunctionArgs(getContext(), FI);
1895 QualType RetTy = FI.getReturnType();
1896 const ABIArgInfo &RetAI = FI.getReturnInfo();
1897 switch (RetAI.getKind()) {
1898 case ABIArgInfo::Extend:
1899 if (RetTy->hasSignedIntegerRepresentation())
1900 RetAttrs.addAttribute(llvm::Attribute::SExt);
1901 else if (RetTy->hasUnsignedIntegerRepresentation())
1902 RetAttrs.addAttribute(llvm::Attribute::ZExt);
1904 case ABIArgInfo::Direct:
1905 if (RetAI.getInReg())
1906 RetAttrs.addAttribute(llvm::Attribute::InReg);
1908 case ABIArgInfo::Ignore:
1911 case ABIArgInfo::InAlloca:
1912 case ABIArgInfo::Indirect: {
1913 // inalloca and sret disable readnone and readonly
1914 FuncAttrs.removeAttribute(llvm::Attribute::ReadOnly)
1915 .removeAttribute(llvm::Attribute::ReadNone);
1919 case ABIArgInfo::CoerceAndExpand:
1922 case ABIArgInfo::Expand:
1923 llvm_unreachable("Invalid ABI kind for return argument");
1926 if (const auto *RefTy = RetTy->getAs<ReferenceType>()) {
1927 QualType PTy = RefTy->getPointeeType();
1928 if (!PTy->isIncompleteType() && PTy->isConstantSizeType())
1929 RetAttrs.addDereferenceableAttr(getContext().getTypeSizeInChars(PTy)
1931 else if (getContext().getTargetAddressSpace(PTy) == 0)
1932 RetAttrs.addAttribute(llvm::Attribute::NonNull);
1935 bool hasUsedSRet = false;
1936 SmallVector<llvm::AttributeSet, 4> ArgAttrs(IRFunctionArgs.totalIRArgs());
1938 // Attach attributes to sret.
1939 if (IRFunctionArgs.hasSRetArg()) {
1940 llvm::AttrBuilder SRETAttrs;
1941 SRETAttrs.addAttribute(llvm::Attribute::StructRet);
1943 if (RetAI.getInReg())
1944 SRETAttrs.addAttribute(llvm::Attribute::InReg);
1945 ArgAttrs[IRFunctionArgs.getSRetArgNo()] =
1946 llvm::AttributeSet::get(getLLVMContext(), SRETAttrs);
1949 // Attach attributes to inalloca argument.
1950 if (IRFunctionArgs.hasInallocaArg()) {
1951 llvm::AttrBuilder Attrs;
1952 Attrs.addAttribute(llvm::Attribute::InAlloca);
1953 ArgAttrs[IRFunctionArgs.getInallocaArgNo()] =
1954 llvm::AttributeSet::get(getLLVMContext(), Attrs);
1958 for (CGFunctionInfo::const_arg_iterator I = FI.arg_begin(),
1960 I != E; ++I, ++ArgNo) {
1961 QualType ParamType = I->type;
1962 const ABIArgInfo &AI = I->info;
1963 llvm::AttrBuilder Attrs;
1965 // Add attribute for padding argument, if necessary.
1966 if (IRFunctionArgs.hasPaddingArg(ArgNo)) {
1967 if (AI.getPaddingInReg()) {
1968 ArgAttrs[IRFunctionArgs.getPaddingArgNo(ArgNo)] =
1969 llvm::AttributeSet::get(
1971 llvm::AttrBuilder().addAttribute(llvm::Attribute::InReg));
1975 // 'restrict' -> 'noalias' is done in EmitFunctionProlog when we
1976 // have the corresponding parameter variable. It doesn't make
1977 // sense to do it here because parameters are so messed up.
1978 switch (AI.getKind()) {
1979 case ABIArgInfo::Extend:
1980 if (ParamType->isSignedIntegerOrEnumerationType())
1981 Attrs.addAttribute(llvm::Attribute::SExt);
1982 else if (ParamType->isUnsignedIntegerOrEnumerationType()) {
1983 if (getTypes().getABIInfo().shouldSignExtUnsignedType(ParamType))
1984 Attrs.addAttribute(llvm::Attribute::SExt);
1986 Attrs.addAttribute(llvm::Attribute::ZExt);
1989 case ABIArgInfo::Direct:
1990 if (ArgNo == 0 && FI.isChainCall())
1991 Attrs.addAttribute(llvm::Attribute::Nest);
1992 else if (AI.getInReg())
1993 Attrs.addAttribute(llvm::Attribute::InReg);
1996 case ABIArgInfo::Indirect: {
1998 Attrs.addAttribute(llvm::Attribute::InReg);
2000 if (AI.getIndirectByVal())
2001 Attrs.addAttribute(llvm::Attribute::ByVal);
2003 CharUnits Align = AI.getIndirectAlign();
2005 // In a byval argument, it is important that the required
2006 // alignment of the type is honored, as LLVM might be creating a
2007 // *new* stack object, and needs to know what alignment to give
2008 // it. (Sometimes it can deduce a sensible alignment on its own,
2009 // but not if clang decides it must emit a packed struct, or the
2010 // user specifies increased alignment requirements.)
2012 // This is different from indirect *not* byval, where the object
2013 // exists already, and the align attribute is purely
2015 assert(!Align.isZero());
2017 // For now, only add this when we have a byval argument.
2018 // TODO: be less lazy about updating test cases.
2019 if (AI.getIndirectByVal())
2020 Attrs.addAlignmentAttr(Align.getQuantity());
2022 // byval disables readnone and readonly.
2023 FuncAttrs.removeAttribute(llvm::Attribute::ReadOnly)
2024 .removeAttribute(llvm::Attribute::ReadNone);
2027 case ABIArgInfo::Ignore:
2028 case ABIArgInfo::Expand:
2029 case ABIArgInfo::CoerceAndExpand:
2032 case ABIArgInfo::InAlloca:
2033 // inalloca disables readnone and readonly.
2034 FuncAttrs.removeAttribute(llvm::Attribute::ReadOnly)
2035 .removeAttribute(llvm::Attribute::ReadNone);
2039 if (const auto *RefTy = ParamType->getAs<ReferenceType>()) {
2040 QualType PTy = RefTy->getPointeeType();
2041 if (!PTy->isIncompleteType() && PTy->isConstantSizeType())
2042 Attrs.addDereferenceableAttr(getContext().getTypeSizeInChars(PTy)
2044 else if (getContext().getTargetAddressSpace(PTy) == 0)
2045 Attrs.addAttribute(llvm::Attribute::NonNull);
2048 switch (FI.getExtParameterInfo(ArgNo).getABI()) {
2049 case ParameterABI::Ordinary:
2052 case ParameterABI::SwiftIndirectResult: {
2053 // Add 'sret' if we haven't already used it for something, but
2054 // only if the result is void.
2055 if (!hasUsedSRet && RetTy->isVoidType()) {
2056 Attrs.addAttribute(llvm::Attribute::StructRet);
2060 // Add 'noalias' in either case.
2061 Attrs.addAttribute(llvm::Attribute::NoAlias);
2063 // Add 'dereferenceable' and 'alignment'.
2064 auto PTy = ParamType->getPointeeType();
2065 if (!PTy->isIncompleteType() && PTy->isConstantSizeType()) {
2066 auto info = getContext().getTypeInfoInChars(PTy);
2067 Attrs.addDereferenceableAttr(info.first.getQuantity());
2068 Attrs.addAttribute(llvm::Attribute::getWithAlignment(getLLVMContext(),
2069 info.second.getQuantity()));
2074 case ParameterABI::SwiftErrorResult:
2075 Attrs.addAttribute(llvm::Attribute::SwiftError);
2078 case ParameterABI::SwiftContext:
2079 Attrs.addAttribute(llvm::Attribute::SwiftSelf);
2083 if (Attrs.hasAttributes()) {
2084 unsigned FirstIRArg, NumIRArgs;
2085 std::tie(FirstIRArg, NumIRArgs) = IRFunctionArgs.getIRArgs(ArgNo);
2086 for (unsigned i = 0; i < NumIRArgs; i++)
2087 ArgAttrs[FirstIRArg + i] =
2088 llvm::AttributeSet::get(getLLVMContext(), Attrs);
2091 assert(ArgNo == FI.arg_size());
2093 AttrList = llvm::AttributeList::get(
2094 getLLVMContext(), llvm::AttributeSet::get(getLLVMContext(), FuncAttrs),
2095 llvm::AttributeSet::get(getLLVMContext(), RetAttrs), ArgAttrs);
2098 /// An argument came in as a promoted argument; demote it back to its
2100 static llvm::Value *emitArgumentDemotion(CodeGenFunction &CGF,
2102 llvm::Value *value) {
2103 llvm::Type *varType = CGF.ConvertType(var->getType());
2105 // This can happen with promotions that actually don't change the
2106 // underlying type, like the enum promotions.
2107 if (value->getType() == varType) return value;
2109 assert((varType->isIntegerTy() || varType->isFloatingPointTy())
2110 && "unexpected promotion type");
2112 if (isa<llvm::IntegerType>(varType))
2113 return CGF.Builder.CreateTrunc(value, varType, "arg.unpromote");
2115 return CGF.Builder.CreateFPCast(value, varType, "arg.unpromote");
2118 /// Returns the attribute (either parameter attribute, or function
2119 /// attribute), which declares argument ArgNo to be non-null.
2120 static const NonNullAttr *getNonNullAttr(const Decl *FD, const ParmVarDecl *PVD,
2121 QualType ArgType, unsigned ArgNo) {
2122 // FIXME: __attribute__((nonnull)) can also be applied to:
2123 // - references to pointers, where the pointee is known to be
2124 // nonnull (apparently a Clang extension)
2125 // - transparent unions containing pointers
2126 // In the former case, LLVM IR cannot represent the constraint. In
2127 // the latter case, we have no guarantee that the transparent union
2128 // is in fact passed as a pointer.
2129 if (!ArgType->isAnyPointerType() && !ArgType->isBlockPointerType())
2131 // First, check attribute on parameter itself.
2133 if (auto ParmNNAttr = PVD->getAttr<NonNullAttr>())
2136 // Check function attributes.
2139 for (const auto *NNAttr : FD->specific_attrs<NonNullAttr>()) {
2140 if (NNAttr->isNonNull(ArgNo))
2147 struct CopyBackSwiftError final : EHScopeStack::Cleanup {
2150 CopyBackSwiftError(Address temp, Address arg) : Temp(temp), Arg(arg) {}
2151 void Emit(CodeGenFunction &CGF, Flags flags) override {
2152 llvm::Value *errorValue = CGF.Builder.CreateLoad(Temp);
2153 CGF.Builder.CreateStore(errorValue, Arg);
2158 void CodeGenFunction::EmitFunctionProlog(const CGFunctionInfo &FI,
2160 const FunctionArgList &Args) {
2161 if (CurCodeDecl && CurCodeDecl->hasAttr<NakedAttr>())
2162 // Naked functions don't have prologues.
2165 // If this is an implicit-return-zero function, go ahead and
2166 // initialize the return value. TODO: it might be nice to have
2167 // a more general mechanism for this that didn't require synthesized
2168 // return statements.
2169 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(CurCodeDecl)) {
2170 if (FD->hasImplicitReturnZero()) {
2171 QualType RetTy = FD->getReturnType().getUnqualifiedType();
2172 llvm::Type* LLVMTy = CGM.getTypes().ConvertType(RetTy);
2173 llvm::Constant* Zero = llvm::Constant::getNullValue(LLVMTy);
2174 Builder.CreateStore(Zero, ReturnValue);
2178 // FIXME: We no longer need the types from FunctionArgList; lift up and
2181 ClangToLLVMArgMapping IRFunctionArgs(CGM.getContext(), FI);
2182 // Flattened function arguments.
2183 SmallVector<llvm::Value *, 16> FnArgs;
2184 FnArgs.reserve(IRFunctionArgs.totalIRArgs());
2185 for (auto &Arg : Fn->args()) {
2186 FnArgs.push_back(&Arg);
2188 assert(FnArgs.size() == IRFunctionArgs.totalIRArgs());
2190 // If we're using inalloca, all the memory arguments are GEPs off of the last
2191 // parameter, which is a pointer to the complete memory area.
2192 Address ArgStruct = Address::invalid();
2193 const llvm::StructLayout *ArgStructLayout = nullptr;
2194 if (IRFunctionArgs.hasInallocaArg()) {
2195 ArgStructLayout = CGM.getDataLayout().getStructLayout(FI.getArgStruct());
2196 ArgStruct = Address(FnArgs[IRFunctionArgs.getInallocaArgNo()],
2197 FI.getArgStructAlignment());
2199 assert(ArgStruct.getType() == FI.getArgStruct()->getPointerTo());
2202 // Name the struct return parameter.
2203 if (IRFunctionArgs.hasSRetArg()) {
2204 auto AI = cast<llvm::Argument>(FnArgs[IRFunctionArgs.getSRetArgNo()]);
2205 AI->setName("agg.result");
2206 AI->addAttr(llvm::Attribute::NoAlias);
2209 // Track if we received the parameter as a pointer (indirect, byval, or
2210 // inalloca). If already have a pointer, EmitParmDecl doesn't need to copy it
2211 // into a local alloca for us.
2212 SmallVector<ParamValue, 16> ArgVals;
2213 ArgVals.reserve(Args.size());
2215 // Create a pointer value for every parameter declaration. This usually
2216 // entails copying one or more LLVM IR arguments into an alloca. Don't push
2217 // any cleanups or do anything that might unwind. We do that separately, so
2218 // we can push the cleanups in the correct order for the ABI.
2219 assert(FI.arg_size() == Args.size() &&
2220 "Mismatch between function signature & arguments.");
2222 CGFunctionInfo::const_arg_iterator info_it = FI.arg_begin();
2223 for (FunctionArgList::const_iterator i = Args.begin(), e = Args.end();
2224 i != e; ++i, ++info_it, ++ArgNo) {
2225 const VarDecl *Arg = *i;
2226 QualType Ty = info_it->type;
2227 const ABIArgInfo &ArgI = info_it->info;
2230 isa<ParmVarDecl>(Arg) && cast<ParmVarDecl>(Arg)->isKNRPromoted();
2232 unsigned FirstIRArg, NumIRArgs;
2233 std::tie(FirstIRArg, NumIRArgs) = IRFunctionArgs.getIRArgs(ArgNo);
2235 switch (ArgI.getKind()) {
2236 case ABIArgInfo::InAlloca: {
2237 assert(NumIRArgs == 0);
2238 auto FieldIndex = ArgI.getInAllocaFieldIndex();
2239 CharUnits FieldOffset =
2240 CharUnits::fromQuantity(ArgStructLayout->getElementOffset(FieldIndex));
2241 Address V = Builder.CreateStructGEP(ArgStruct, FieldIndex, FieldOffset,
2243 ArgVals.push_back(ParamValue::forIndirect(V));
2247 case ABIArgInfo::Indirect: {
2248 assert(NumIRArgs == 1);
2249 Address ParamAddr = Address(FnArgs[FirstIRArg], ArgI.getIndirectAlign());
2251 if (!hasScalarEvaluationKind(Ty)) {
2252 // Aggregates and complex variables are accessed by reference. All we
2253 // need to do is realign the value, if requested.
2254 Address V = ParamAddr;
2255 if (ArgI.getIndirectRealign()) {
2256 Address AlignedTemp = CreateMemTemp(Ty, "coerce");
2258 // Copy from the incoming argument pointer to the temporary with the
2259 // appropriate alignment.
2261 // FIXME: We should have a common utility for generating an aggregate
2263 CharUnits Size = getContext().getTypeSizeInChars(Ty);
2264 auto SizeVal = llvm::ConstantInt::get(IntPtrTy, Size.getQuantity());
2265 Address Dst = Builder.CreateBitCast(AlignedTemp, Int8PtrTy);
2266 Address Src = Builder.CreateBitCast(ParamAddr, Int8PtrTy);
2267 Builder.CreateMemCpy(Dst, Src, SizeVal, false);
2270 ArgVals.push_back(ParamValue::forIndirect(V));
2272 // Load scalar value from indirect argument.
2274 EmitLoadOfScalar(ParamAddr, false, Ty, Arg->getLocStart());
2277 V = emitArgumentDemotion(*this, Arg, V);
2278 ArgVals.push_back(ParamValue::forDirect(V));
2283 case ABIArgInfo::Extend:
2284 case ABIArgInfo::Direct: {
2286 // If we have the trivial case, handle it with no muss and fuss.
2287 if (!isa<llvm::StructType>(ArgI.getCoerceToType()) &&
2288 ArgI.getCoerceToType() == ConvertType(Ty) &&
2289 ArgI.getDirectOffset() == 0) {
2290 assert(NumIRArgs == 1);
2291 llvm::Value *V = FnArgs[FirstIRArg];
2292 auto AI = cast<llvm::Argument>(V);
2294 if (const ParmVarDecl *PVD = dyn_cast<ParmVarDecl>(Arg)) {
2295 if (getNonNullAttr(CurCodeDecl, PVD, PVD->getType(),
2296 PVD->getFunctionScopeIndex()))
2297 AI->addAttr(llvm::Attribute::NonNull);
2299 QualType OTy = PVD->getOriginalType();
2300 if (const auto *ArrTy =
2301 getContext().getAsConstantArrayType(OTy)) {
2302 // A C99 array parameter declaration with the static keyword also
2303 // indicates dereferenceability, and if the size is constant we can
2304 // use the dereferenceable attribute (which requires the size in
2306 if (ArrTy->getSizeModifier() == ArrayType::Static) {
2307 QualType ETy = ArrTy->getElementType();
2308 uint64_t ArrSize = ArrTy->getSize().getZExtValue();
2309 if (!ETy->isIncompleteType() && ETy->isConstantSizeType() &&
2311 llvm::AttrBuilder Attrs;
2312 Attrs.addDereferenceableAttr(
2313 getContext().getTypeSizeInChars(ETy).getQuantity()*ArrSize);
2314 AI->addAttrs(Attrs);
2315 } else if (getContext().getTargetAddressSpace(ETy) == 0) {
2316 AI->addAttr(llvm::Attribute::NonNull);
2319 } else if (const auto *ArrTy =
2320 getContext().getAsVariableArrayType(OTy)) {
2321 // For C99 VLAs with the static keyword, we don't know the size so
2322 // we can't use the dereferenceable attribute, but in addrspace(0)
2323 // we know that it must be nonnull.
2324 if (ArrTy->getSizeModifier() == VariableArrayType::Static &&
2325 !getContext().getTargetAddressSpace(ArrTy->getElementType()))
2326 AI->addAttr(llvm::Attribute::NonNull);
2329 const auto *AVAttr = PVD->getAttr<AlignValueAttr>();
2331 if (const auto *TOTy = dyn_cast<TypedefType>(OTy))
2332 AVAttr = TOTy->getDecl()->getAttr<AlignValueAttr>();
2334 llvm::Value *AlignmentValue =
2335 EmitScalarExpr(AVAttr->getAlignment());
2336 llvm::ConstantInt *AlignmentCI =
2337 cast<llvm::ConstantInt>(AlignmentValue);
2338 unsigned Alignment = std::min((unsigned)AlignmentCI->getZExtValue(),
2339 +llvm::Value::MaximumAlignment);
2340 AI->addAttrs(llvm::AttrBuilder().addAlignmentAttr(Alignment));
2344 if (Arg->getType().isRestrictQualified())
2345 AI->addAttr(llvm::Attribute::NoAlias);
2347 // LLVM expects swifterror parameters to be used in very restricted
2348 // ways. Copy the value into a less-restricted temporary.
2349 if (FI.getExtParameterInfo(ArgNo).getABI()
2350 == ParameterABI::SwiftErrorResult) {
2351 QualType pointeeTy = Ty->getPointeeType();
2352 assert(pointeeTy->isPointerType());
2354 CreateMemTemp(pointeeTy, getPointerAlign(), "swifterror.temp");
2355 Address arg = Address(V, getContext().getTypeAlignInChars(pointeeTy));
2356 llvm::Value *incomingErrorValue = Builder.CreateLoad(arg);
2357 Builder.CreateStore(incomingErrorValue, temp);
2358 V = temp.getPointer();
2360 // Push a cleanup to copy the value back at the end of the function.
2361 // The convention does not guarantee that the value will be written
2362 // back if the function exits with an unwind exception.
2363 EHStack.pushCleanup<CopyBackSwiftError>(NormalCleanup, temp, arg);
2366 // Ensure the argument is the correct type.
2367 if (V->getType() != ArgI.getCoerceToType())
2368 V = Builder.CreateBitCast(V, ArgI.getCoerceToType());
2371 V = emitArgumentDemotion(*this, Arg, V);
2373 // Because of merging of function types from multiple decls it is
2374 // possible for the type of an argument to not match the corresponding
2375 // type in the function type. Since we are codegening the callee
2376 // in here, add a cast to the argument type.
2377 llvm::Type *LTy = ConvertType(Arg->getType());
2378 if (V->getType() != LTy)
2379 V = Builder.CreateBitCast(V, LTy);
2381 ArgVals.push_back(ParamValue::forDirect(V));
2385 Address Alloca = CreateMemTemp(Ty, getContext().getDeclAlign(Arg),
2388 // Pointer to store into.
2389 Address Ptr = emitAddressAtOffset(*this, Alloca, ArgI);
2391 // Fast-isel and the optimizer generally like scalar values better than
2392 // FCAs, so we flatten them if this is safe to do for this argument.
2393 llvm::StructType *STy = dyn_cast<llvm::StructType>(ArgI.getCoerceToType());
2394 if (ArgI.isDirect() && ArgI.getCanBeFlattened() && STy &&
2395 STy->getNumElements() > 1) {
2396 auto SrcLayout = CGM.getDataLayout().getStructLayout(STy);
2397 uint64_t SrcSize = CGM.getDataLayout().getTypeAllocSize(STy);
2398 llvm::Type *DstTy = Ptr.getElementType();
2399 uint64_t DstSize = CGM.getDataLayout().getTypeAllocSize(DstTy);
2401 Address AddrToStoreInto = Address::invalid();
2402 if (SrcSize <= DstSize) {
2404 Builder.CreateBitCast(Ptr, llvm::PointerType::getUnqual(STy));
2407 CreateTempAlloca(STy, Alloca.getAlignment(), "coerce");
2410 assert(STy->getNumElements() == NumIRArgs);
2411 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
2412 auto AI = FnArgs[FirstIRArg + i];
2413 AI->setName(Arg->getName() + ".coerce" + Twine(i));
2414 auto Offset = CharUnits::fromQuantity(SrcLayout->getElementOffset(i));
2416 Builder.CreateStructGEP(AddrToStoreInto, i, Offset);
2417 Builder.CreateStore(AI, EltPtr);
2420 if (SrcSize > DstSize) {
2421 Builder.CreateMemCpy(Ptr, AddrToStoreInto, DstSize);
2425 // Simple case, just do a coerced store of the argument into the alloca.
2426 assert(NumIRArgs == 1);
2427 auto AI = FnArgs[FirstIRArg];
2428 AI->setName(Arg->getName() + ".coerce");
2429 CreateCoercedStore(AI, Ptr, /*DestIsVolatile=*/false, *this);
2432 // Match to what EmitParmDecl is expecting for this type.
2433 if (CodeGenFunction::hasScalarEvaluationKind(Ty)) {
2435 EmitLoadOfScalar(Alloca, false, Ty, Arg->getLocStart());
2437 V = emitArgumentDemotion(*this, Arg, V);
2438 ArgVals.push_back(ParamValue::forDirect(V));
2440 ArgVals.push_back(ParamValue::forIndirect(Alloca));
2445 case ABIArgInfo::CoerceAndExpand: {
2446 // Reconstruct into a temporary.
2447 Address alloca = CreateMemTemp(Ty, getContext().getDeclAlign(Arg));
2448 ArgVals.push_back(ParamValue::forIndirect(alloca));
2450 auto coercionType = ArgI.getCoerceAndExpandType();
2451 alloca = Builder.CreateElementBitCast(alloca, coercionType);
2452 auto layout = CGM.getDataLayout().getStructLayout(coercionType);
2454 unsigned argIndex = FirstIRArg;
2455 for (unsigned i = 0, e = coercionType->getNumElements(); i != e; ++i) {
2456 llvm::Type *eltType = coercionType->getElementType(i);
2457 if (ABIArgInfo::isPaddingForCoerceAndExpand(eltType))
2460 auto eltAddr = Builder.CreateStructGEP(alloca, i, layout);
2461 auto elt = FnArgs[argIndex++];
2462 Builder.CreateStore(elt, eltAddr);
2464 assert(argIndex == FirstIRArg + NumIRArgs);
2468 case ABIArgInfo::Expand: {
2469 // If this structure was expanded into multiple arguments then
2470 // we need to create a temporary and reconstruct it from the
2472 Address Alloca = CreateMemTemp(Ty, getContext().getDeclAlign(Arg));
2473 LValue LV = MakeAddrLValue(Alloca, Ty);
2474 ArgVals.push_back(ParamValue::forIndirect(Alloca));
2476 auto FnArgIter = FnArgs.begin() + FirstIRArg;
2477 ExpandTypeFromArgs(Ty, LV, FnArgIter);
2478 assert(FnArgIter == FnArgs.begin() + FirstIRArg + NumIRArgs);
2479 for (unsigned i = 0, e = NumIRArgs; i != e; ++i) {
2480 auto AI = FnArgs[FirstIRArg + i];
2481 AI->setName(Arg->getName() + "." + Twine(i));
2486 case ABIArgInfo::Ignore:
2487 assert(NumIRArgs == 0);
2488 // Initialize the local variable appropriately.
2489 if (!hasScalarEvaluationKind(Ty)) {
2490 ArgVals.push_back(ParamValue::forIndirect(CreateMemTemp(Ty)));
2492 llvm::Value *U = llvm::UndefValue::get(ConvertType(Arg->getType()));
2493 ArgVals.push_back(ParamValue::forDirect(U));
2499 if (getTarget().getCXXABI().areArgsDestroyedLeftToRightInCallee()) {
2500 for (int I = Args.size() - 1; I >= 0; --I)
2501 EmitParmDecl(*Args[I], ArgVals[I], I + 1);
2503 for (unsigned I = 0, E = Args.size(); I != E; ++I)
2504 EmitParmDecl(*Args[I], ArgVals[I], I + 1);
2508 static void eraseUnusedBitCasts(llvm::Instruction *insn) {
2509 while (insn->use_empty()) {
2510 llvm::BitCastInst *bitcast = dyn_cast<llvm::BitCastInst>(insn);
2511 if (!bitcast) return;
2513 // This is "safe" because we would have used a ConstantExpr otherwise.
2514 insn = cast<llvm::Instruction>(bitcast->getOperand(0));
2515 bitcast->eraseFromParent();
2519 /// Try to emit a fused autorelease of a return result.
2520 static llvm::Value *tryEmitFusedAutoreleaseOfResult(CodeGenFunction &CGF,
2521 llvm::Value *result) {
2522 // We must be immediately followed the cast.
2523 llvm::BasicBlock *BB = CGF.Builder.GetInsertBlock();
2524 if (BB->empty()) return nullptr;
2525 if (&BB->back() != result) return nullptr;
2527 llvm::Type *resultType = result->getType();
2529 // result is in a BasicBlock and is therefore an Instruction.
2530 llvm::Instruction *generator = cast<llvm::Instruction>(result);
2532 SmallVector<llvm::Instruction *, 4> InstsToKill;
2535 // %generator = bitcast %type1* %generator2 to %type2*
2536 while (llvm::BitCastInst *bitcast = dyn_cast<llvm::BitCastInst>(generator)) {
2537 // We would have emitted this as a constant if the operand weren't
2539 generator = cast<llvm::Instruction>(bitcast->getOperand(0));
2541 // Require the generator to be immediately followed by the cast.
2542 if (generator->getNextNode() != bitcast)
2545 InstsToKill.push_back(bitcast);
2549 // %generator = call i8* @objc_retain(i8* %originalResult)
2551 // %generator = call i8* @objc_retainAutoreleasedReturnValue(i8* %originalResult)
2552 llvm::CallInst *call = dyn_cast<llvm::CallInst>(generator);
2553 if (!call) return nullptr;
2555 bool doRetainAutorelease;
2557 if (call->getCalledValue() == CGF.CGM.getObjCEntrypoints().objc_retain) {
2558 doRetainAutorelease = true;
2559 } else if (call->getCalledValue() == CGF.CGM.getObjCEntrypoints()
2560 .objc_retainAutoreleasedReturnValue) {
2561 doRetainAutorelease = false;
2563 // If we emitted an assembly marker for this call (and the
2564 // ARCEntrypoints field should have been set if so), go looking
2565 // for that call. If we can't find it, we can't do this
2566 // optimization. But it should always be the immediately previous
2567 // instruction, unless we needed bitcasts around the call.
2568 if (CGF.CGM.getObjCEntrypoints().retainAutoreleasedReturnValueMarker) {
2569 llvm::Instruction *prev = call->getPrevNode();
2571 if (isa<llvm::BitCastInst>(prev)) {
2572 prev = prev->getPrevNode();
2575 assert(isa<llvm::CallInst>(prev));
2576 assert(cast<llvm::CallInst>(prev)->getCalledValue() ==
2577 CGF.CGM.getObjCEntrypoints().retainAutoreleasedReturnValueMarker);
2578 InstsToKill.push_back(prev);
2584 result = call->getArgOperand(0);
2585 InstsToKill.push_back(call);
2587 // Keep killing bitcasts, for sanity. Note that we no longer care
2588 // about precise ordering as long as there's exactly one use.
2589 while (llvm::BitCastInst *bitcast = dyn_cast<llvm::BitCastInst>(result)) {
2590 if (!bitcast->hasOneUse()) break;
2591 InstsToKill.push_back(bitcast);
2592 result = bitcast->getOperand(0);
2595 // Delete all the unnecessary instructions, from latest to earliest.
2596 for (auto *I : InstsToKill)
2597 I->eraseFromParent();
2599 // Do the fused retain/autorelease if we were asked to.
2600 if (doRetainAutorelease)
2601 result = CGF.EmitARCRetainAutoreleaseReturnValue(result);
2603 // Cast back to the result type.
2604 return CGF.Builder.CreateBitCast(result, resultType);
2607 /// If this is a +1 of the value of an immutable 'self', remove it.
2608 static llvm::Value *tryRemoveRetainOfSelf(CodeGenFunction &CGF,
2609 llvm::Value *result) {
2610 // This is only applicable to a method with an immutable 'self'.
2611 const ObjCMethodDecl *method =
2612 dyn_cast_or_null<ObjCMethodDecl>(CGF.CurCodeDecl);
2613 if (!method) return nullptr;
2614 const VarDecl *self = method->getSelfDecl();
2615 if (!self->getType().isConstQualified()) return nullptr;
2617 // Look for a retain call.
2618 llvm::CallInst *retainCall =
2619 dyn_cast<llvm::CallInst>(result->stripPointerCasts());
2621 retainCall->getCalledValue() != CGF.CGM.getObjCEntrypoints().objc_retain)
2624 // Look for an ordinary load of 'self'.
2625 llvm::Value *retainedValue = retainCall->getArgOperand(0);
2626 llvm::LoadInst *load =
2627 dyn_cast<llvm::LoadInst>(retainedValue->stripPointerCasts());
2628 if (!load || load->isAtomic() || load->isVolatile() ||
2629 load->getPointerOperand() != CGF.GetAddrOfLocalVar(self).getPointer())
2632 // Okay! Burn it all down. This relies for correctness on the
2633 // assumption that the retain is emitted as part of the return and
2634 // that thereafter everything is used "linearly".
2635 llvm::Type *resultType = result->getType();
2636 eraseUnusedBitCasts(cast<llvm::Instruction>(result));
2637 assert(retainCall->use_empty());
2638 retainCall->eraseFromParent();
2639 eraseUnusedBitCasts(cast<llvm::Instruction>(retainedValue));
2641 return CGF.Builder.CreateBitCast(load, resultType);
2644 /// Emit an ARC autorelease of the result of a function.
2646 /// \return the value to actually return from the function
2647 static llvm::Value *emitAutoreleaseOfResult(CodeGenFunction &CGF,
2648 llvm::Value *result) {
2649 // If we're returning 'self', kill the initial retain. This is a
2650 // heuristic attempt to "encourage correctness" in the really unfortunate
2651 // case where we have a return of self during a dealloc and we desperately
2652 // need to avoid the possible autorelease.
2653 if (llvm::Value *self = tryRemoveRetainOfSelf(CGF, result))
2656 // At -O0, try to emit a fused retain/autorelease.
2657 if (CGF.shouldUseFusedARCCalls())
2658 if (llvm::Value *fused = tryEmitFusedAutoreleaseOfResult(CGF, result))
2661 return CGF.EmitARCAutoreleaseReturnValue(result);
2664 /// Heuristically search for a dominating store to the return-value slot.
2665 static llvm::StoreInst *findDominatingStoreToReturnValue(CodeGenFunction &CGF) {
2666 // Check if a User is a store which pointerOperand is the ReturnValue.
2667 // We are looking for stores to the ReturnValue, not for stores of the
2668 // ReturnValue to some other location.
2669 auto GetStoreIfValid = [&CGF](llvm::User *U) -> llvm::StoreInst * {
2670 auto *SI = dyn_cast<llvm::StoreInst>(U);
2671 if (!SI || SI->getPointerOperand() != CGF.ReturnValue.getPointer())
2673 // These aren't actually possible for non-coerced returns, and we
2674 // only care about non-coerced returns on this code path.
2675 assert(!SI->isAtomic() && !SI->isVolatile());
2678 // If there are multiple uses of the return-value slot, just check
2679 // for something immediately preceding the IP. Sometimes this can
2680 // happen with how we generate implicit-returns; it can also happen
2681 // with noreturn cleanups.
2682 if (!CGF.ReturnValue.getPointer()->hasOneUse()) {
2683 llvm::BasicBlock *IP = CGF.Builder.GetInsertBlock();
2684 if (IP->empty()) return nullptr;
2685 llvm::Instruction *I = &IP->back();
2687 // Skip lifetime markers
2688 for (llvm::BasicBlock::reverse_iterator II = IP->rbegin(),
2691 if (llvm::IntrinsicInst *Intrinsic =
2692 dyn_cast<llvm::IntrinsicInst>(&*II)) {
2693 if (Intrinsic->getIntrinsicID() == llvm::Intrinsic::lifetime_end) {
2694 const llvm::Value *CastAddr = Intrinsic->getArgOperand(1);
2698 if (isa<llvm::BitCastInst>(&*II) && (CastAddr == &*II))
2706 return GetStoreIfValid(I);
2709 llvm::StoreInst *store =
2710 GetStoreIfValid(CGF.ReturnValue.getPointer()->user_back());
2711 if (!store) return nullptr;
2713 // Now do a first-and-dirty dominance check: just walk up the
2714 // single-predecessors chain from the current insertion point.
2715 llvm::BasicBlock *StoreBB = store->getParent();
2716 llvm::BasicBlock *IP = CGF.Builder.GetInsertBlock();
2717 while (IP != StoreBB) {
2718 if (!(IP = IP->getSinglePredecessor()))
2722 // Okay, the store's basic block dominates the insertion point; we
2723 // can do our thing.
2727 void CodeGenFunction::EmitFunctionEpilog(const CGFunctionInfo &FI,
2729 SourceLocation EndLoc) {
2730 if (CurCodeDecl && CurCodeDecl->hasAttr<NakedAttr>()) {
2731 // Naked functions don't have epilogues.
2732 Builder.CreateUnreachable();
2736 // Functions with no result always return void.
2737 if (!ReturnValue.isValid()) {
2738 Builder.CreateRetVoid();
2742 llvm::DebugLoc RetDbgLoc;
2743 llvm::Value *RV = nullptr;
2744 QualType RetTy = FI.getReturnType();
2745 const ABIArgInfo &RetAI = FI.getReturnInfo();
2747 switch (RetAI.getKind()) {
2748 case ABIArgInfo::InAlloca:
2749 // Aggregrates get evaluated directly into the destination. Sometimes we
2750 // need to return the sret value in a register, though.
2751 assert(hasAggregateEvaluationKind(RetTy));
2752 if (RetAI.getInAllocaSRet()) {
2753 llvm::Function::arg_iterator EI = CurFn->arg_end();
2755 llvm::Value *ArgStruct = &*EI;
2756 llvm::Value *SRet = Builder.CreateStructGEP(
2757 nullptr, ArgStruct, RetAI.getInAllocaFieldIndex());
2758 RV = Builder.CreateAlignedLoad(SRet, getPointerAlign(), "sret");
2762 case ABIArgInfo::Indirect: {
2763 auto AI = CurFn->arg_begin();
2764 if (RetAI.isSRetAfterThis())
2766 switch (getEvaluationKind(RetTy)) {
2769 EmitLoadOfComplex(MakeAddrLValue(ReturnValue, RetTy), EndLoc);
2770 EmitStoreOfComplex(RT, MakeNaturalAlignAddrLValue(&*AI, RetTy),
2775 // Do nothing; aggregrates get evaluated directly into the destination.
2778 EmitStoreOfScalar(Builder.CreateLoad(ReturnValue),
2779 MakeNaturalAlignAddrLValue(&*AI, RetTy),
2786 case ABIArgInfo::Extend:
2787 case ABIArgInfo::Direct:
2788 if (RetAI.getCoerceToType() == ConvertType(RetTy) &&
2789 RetAI.getDirectOffset() == 0) {
2790 // The internal return value temp always will have pointer-to-return-type
2791 // type, just do a load.
2793 // If there is a dominating store to ReturnValue, we can elide
2794 // the load, zap the store, and usually zap the alloca.
2795 if (llvm::StoreInst *SI =
2796 findDominatingStoreToReturnValue(*this)) {
2797 // Reuse the debug location from the store unless there is
2798 // cleanup code to be emitted between the store and return
2800 if (EmitRetDbgLoc && !AutoreleaseResult)
2801 RetDbgLoc = SI->getDebugLoc();
2802 // Get the stored value and nuke the now-dead store.
2803 RV = SI->getValueOperand();
2804 SI->eraseFromParent();
2806 // If that was the only use of the return value, nuke it as well now.
2807 auto returnValueInst = ReturnValue.getPointer();
2808 if (returnValueInst->use_empty()) {
2809 if (auto alloca = dyn_cast<llvm::AllocaInst>(returnValueInst)) {
2810 alloca->eraseFromParent();
2811 ReturnValue = Address::invalid();
2815 // Otherwise, we have to do a simple load.
2817 RV = Builder.CreateLoad(ReturnValue);
2820 // If the value is offset in memory, apply the offset now.
2821 Address V = emitAddressAtOffset(*this, ReturnValue, RetAI);
2823 RV = CreateCoercedLoad(V, RetAI.getCoerceToType(), *this);
2826 // In ARC, end functions that return a retainable type with a call
2827 // to objc_autoreleaseReturnValue.
2828 if (AutoreleaseResult) {
2830 // Type::isObjCRetainabletype has to be called on a QualType that hasn't
2831 // been stripped of the typedefs, so we cannot use RetTy here. Get the
2832 // original return type of FunctionDecl, CurCodeDecl, and BlockDecl from
2833 // CurCodeDecl or BlockInfo.
2836 if (auto *FD = dyn_cast<FunctionDecl>(CurCodeDecl))
2837 RT = FD->getReturnType();
2838 else if (auto *MD = dyn_cast<ObjCMethodDecl>(CurCodeDecl))
2839 RT = MD->getReturnType();
2840 else if (isa<BlockDecl>(CurCodeDecl))
2841 RT = BlockInfo->BlockExpression->getFunctionType()->getReturnType();
2843 llvm_unreachable("Unexpected function/method type");
2845 assert(getLangOpts().ObjCAutoRefCount &&
2846 !FI.isReturnsRetained() &&
2847 RT->isObjCRetainableType());
2849 RV = emitAutoreleaseOfResult(*this, RV);
2854 case ABIArgInfo::Ignore:
2857 case ABIArgInfo::CoerceAndExpand: {
2858 auto coercionType = RetAI.getCoerceAndExpandType();
2859 auto layout = CGM.getDataLayout().getStructLayout(coercionType);
2861 // Load all of the coerced elements out into results.
2862 llvm::SmallVector<llvm::Value*, 4> results;
2863 Address addr = Builder.CreateElementBitCast(ReturnValue, coercionType);
2864 for (unsigned i = 0, e = coercionType->getNumElements(); i != e; ++i) {
2865 auto coercedEltType = coercionType->getElementType(i);
2866 if (ABIArgInfo::isPaddingForCoerceAndExpand(coercedEltType))
2869 auto eltAddr = Builder.CreateStructGEP(addr, i, layout);
2870 auto elt = Builder.CreateLoad(eltAddr);
2871 results.push_back(elt);
2874 // If we have one result, it's the single direct result type.
2875 if (results.size() == 1) {
2878 // Otherwise, we need to make a first-class aggregate.
2880 // Construct a return type that lacks padding elements.
2881 llvm::Type *returnType = RetAI.getUnpaddedCoerceAndExpandType();
2883 RV = llvm::UndefValue::get(returnType);
2884 for (unsigned i = 0, e = results.size(); i != e; ++i) {
2885 RV = Builder.CreateInsertValue(RV, results[i], i);
2891 case ABIArgInfo::Expand:
2892 llvm_unreachable("Invalid ABI kind for return argument");
2895 llvm::Instruction *Ret;
2897 EmitReturnValueCheck(RV, EndLoc);
2898 Ret = Builder.CreateRet(RV);
2900 Ret = Builder.CreateRetVoid();
2904 Ret->setDebugLoc(std::move(RetDbgLoc));
2907 void CodeGenFunction::EmitReturnValueCheck(llvm::Value *RV,
2908 SourceLocation EndLoc) {
2909 // A current decl may not be available when emitting vtable thunks.
2913 ReturnsNonNullAttr *RetNNAttr = nullptr;
2914 if (SanOpts.has(SanitizerKind::ReturnsNonnullAttribute))
2915 RetNNAttr = CurCodeDecl->getAttr<ReturnsNonNullAttr>();
2917 if (!RetNNAttr && !requiresReturnValueNullabilityCheck())
2920 // Prefer the returns_nonnull attribute if it's present.
2921 SourceLocation AttrLoc;
2922 SanitizerMask CheckKind;
2923 SanitizerHandler Handler;
2925 assert(!requiresReturnValueNullabilityCheck() &&
2926 "Cannot check nullability and the nonnull attribute");
2927 AttrLoc = RetNNAttr->getLocation();
2928 CheckKind = SanitizerKind::ReturnsNonnullAttribute;
2929 Handler = SanitizerHandler::NonnullReturn;
2931 if (auto *DD = dyn_cast<DeclaratorDecl>(CurCodeDecl))
2932 if (auto *TSI = DD->getTypeSourceInfo())
2933 if (auto FTL = TSI->getTypeLoc().castAs<FunctionTypeLoc>())
2934 AttrLoc = FTL.getReturnLoc().findNullabilityLoc();
2935 CheckKind = SanitizerKind::NullabilityReturn;
2936 Handler = SanitizerHandler::NullabilityReturn;
2939 SanitizerScope SanScope(this);
2941 llvm::BasicBlock *Check = nullptr;
2942 llvm::BasicBlock *NoCheck = nullptr;
2943 if (requiresReturnValueNullabilityCheck()) {
2944 // Before doing the nullability check, make sure that the preconditions for
2945 // the check are met.
2946 Check = createBasicBlock("nullcheck");
2947 NoCheck = createBasicBlock("no.nullcheck");
2948 Builder.CreateCondBr(RetValNullabilityPrecondition, Check, NoCheck);
2952 // Now do the null check. If the returns_nonnull attribute is present, this
2953 // is done unconditionally.
2954 llvm::Value *Cond = Builder.CreateIsNotNull(RV);
2955 llvm::Constant *StaticData[] = {
2956 EmitCheckSourceLocation(EndLoc), EmitCheckSourceLocation(AttrLoc),
2958 EmitCheck(std::make_pair(Cond, CheckKind), Handler, StaticData, None);
2960 if (requiresReturnValueNullabilityCheck())
2964 static bool isInAllocaArgument(CGCXXABI &ABI, QualType type) {
2965 const CXXRecordDecl *RD = type->getAsCXXRecordDecl();
2966 return RD && ABI.getRecordArgABI(RD) == CGCXXABI::RAA_DirectInMemory;
2969 static AggValueSlot createPlaceholderSlot(CodeGenFunction &CGF,
2971 // FIXME: Generate IR in one pass, rather than going back and fixing up these
2973 llvm::Type *IRTy = CGF.ConvertTypeForMem(Ty);
2974 llvm::Type *IRPtrTy = IRTy->getPointerTo();
2975 llvm::Value *Placeholder = llvm::UndefValue::get(IRPtrTy->getPointerTo());
2977 // FIXME: When we generate this IR in one pass, we shouldn't need
2978 // this win32-specific alignment hack.
2979 CharUnits Align = CharUnits::fromQuantity(4);
2980 Placeholder = CGF.Builder.CreateAlignedLoad(IRPtrTy, Placeholder, Align);
2982 return AggValueSlot::forAddr(Address(Placeholder, Align),
2984 AggValueSlot::IsNotDestructed,
2985 AggValueSlot::DoesNotNeedGCBarriers,
2986 AggValueSlot::IsNotAliased);
2989 void CodeGenFunction::EmitDelegateCallArg(CallArgList &args,
2990 const VarDecl *param,
2991 SourceLocation loc) {
2992 // StartFunction converted the ABI-lowered parameter(s) into a
2993 // local alloca. We need to turn that into an r-value suitable
2995 Address local = GetAddrOfLocalVar(param);
2997 QualType type = param->getType();
2999 assert(!isInAllocaArgument(CGM.getCXXABI(), type) &&
3000 "cannot emit delegate call arguments for inalloca arguments!");
3002 // GetAddrOfLocalVar returns a pointer-to-pointer for references,
3003 // but the argument needs to be the original pointer.
3004 if (type->isReferenceType()) {
3005 args.add(RValue::get(Builder.CreateLoad(local)), type);
3007 // In ARC, move out of consumed arguments so that the release cleanup
3008 // entered by StartFunction doesn't cause an over-release. This isn't
3009 // optimal -O0 code generation, but it should get cleaned up when
3010 // optimization is enabled. This also assumes that delegate calls are
3011 // performed exactly once for a set of arguments, but that should be safe.
3012 } else if (getLangOpts().ObjCAutoRefCount &&
3013 param->hasAttr<NSConsumedAttr>() &&
3014 type->isObjCRetainableType()) {
3015 llvm::Value *ptr = Builder.CreateLoad(local);
3017 llvm::ConstantPointerNull::get(cast<llvm::PointerType>(ptr->getType()));
3018 Builder.CreateStore(null, local);
3019 args.add(RValue::get(ptr), type);
3021 // For the most part, we just need to load the alloca, except that
3022 // aggregate r-values are actually pointers to temporaries.
3024 args.add(convertTempToRValue(local, type, loc), type);
3028 static bool isProvablyNull(llvm::Value *addr) {
3029 return isa<llvm::ConstantPointerNull>(addr);
3032 /// Emit the actual writing-back of a writeback.
3033 static void emitWriteback(CodeGenFunction &CGF,
3034 const CallArgList::Writeback &writeback) {
3035 const LValue &srcLV = writeback.Source;
3036 Address srcAddr = srcLV.getAddress();
3037 assert(!isProvablyNull(srcAddr.getPointer()) &&
3038 "shouldn't have writeback for provably null argument");
3040 llvm::BasicBlock *contBB = nullptr;
3042 // If the argument wasn't provably non-null, we need to null check
3043 // before doing the store.
3044 bool provablyNonNull = llvm::isKnownNonNull(srcAddr.getPointer());
3045 if (!provablyNonNull) {
3046 llvm::BasicBlock *writebackBB = CGF.createBasicBlock("icr.writeback");
3047 contBB = CGF.createBasicBlock("icr.done");
3049 llvm::Value *isNull =
3050 CGF.Builder.CreateIsNull(srcAddr.getPointer(), "icr.isnull");
3051 CGF.Builder.CreateCondBr(isNull, contBB, writebackBB);
3052 CGF.EmitBlock(writebackBB);
3055 // Load the value to writeback.
3056 llvm::Value *value = CGF.Builder.CreateLoad(writeback.Temporary);
3058 // Cast it back, in case we're writing an id to a Foo* or something.
3059 value = CGF.Builder.CreateBitCast(value, srcAddr.getElementType(),
3060 "icr.writeback-cast");
3062 // Perform the writeback.
3064 // If we have a "to use" value, it's something we need to emit a use
3065 // of. This has to be carefully threaded in: if it's done after the
3066 // release it's potentially undefined behavior (and the optimizer
3067 // will ignore it), and if it happens before the retain then the
3068 // optimizer could move the release there.
3069 if (writeback.ToUse) {
3070 assert(srcLV.getObjCLifetime() == Qualifiers::OCL_Strong);
3072 // Retain the new value. No need to block-copy here: the block's
3073 // being passed up the stack.
3074 value = CGF.EmitARCRetainNonBlock(value);
3076 // Emit the intrinsic use here.
3077 CGF.EmitARCIntrinsicUse(writeback.ToUse);
3079 // Load the old value (primitively).
3080 llvm::Value *oldValue = CGF.EmitLoadOfScalar(srcLV, SourceLocation());
3082 // Put the new value in place (primitively).
3083 CGF.EmitStoreOfScalar(value, srcLV, /*init*/ false);
3085 // Release the old value.
3086 CGF.EmitARCRelease(oldValue, srcLV.isARCPreciseLifetime());
3088 // Otherwise, we can just do a normal lvalue store.
3090 CGF.EmitStoreThroughLValue(RValue::get(value), srcLV);
3093 // Jump to the continuation block.
3094 if (!provablyNonNull)
3095 CGF.EmitBlock(contBB);
3098 static void emitWritebacks(CodeGenFunction &CGF,
3099 const CallArgList &args) {
3100 for (const auto &I : args.writebacks())
3101 emitWriteback(CGF, I);
3104 static void deactivateArgCleanupsBeforeCall(CodeGenFunction &CGF,
3105 const CallArgList &CallArgs) {
3106 assert(CGF.getTarget().getCXXABI().areArgsDestroyedLeftToRightInCallee());
3107 ArrayRef<CallArgList::CallArgCleanup> Cleanups =
3108 CallArgs.getCleanupsToDeactivate();
3109 // Iterate in reverse to increase the likelihood of popping the cleanup.
3110 for (const auto &I : llvm::reverse(Cleanups)) {
3111 CGF.DeactivateCleanupBlock(I.Cleanup, I.IsActiveIP);
3112 I.IsActiveIP->eraseFromParent();
3116 static const Expr *maybeGetUnaryAddrOfOperand(const Expr *E) {
3117 if (const UnaryOperator *uop = dyn_cast<UnaryOperator>(E->IgnoreParens()))
3118 if (uop->getOpcode() == UO_AddrOf)
3119 return uop->getSubExpr();
3123 /// Emit an argument that's being passed call-by-writeback. That is,
3124 /// we are passing the address of an __autoreleased temporary; it
3125 /// might be copy-initialized with the current value of the given
3126 /// address, but it will definitely be copied out of after the call.
3127 static void emitWritebackArg(CodeGenFunction &CGF, CallArgList &args,
3128 const ObjCIndirectCopyRestoreExpr *CRE) {
3131 // Make an optimistic effort to emit the address as an l-value.
3132 // This can fail if the argument expression is more complicated.
3133 if (const Expr *lvExpr = maybeGetUnaryAddrOfOperand(CRE->getSubExpr())) {
3134 srcLV = CGF.EmitLValue(lvExpr);
3136 // Otherwise, just emit it as a scalar.
3138 Address srcAddr = CGF.EmitPointerWithAlignment(CRE->getSubExpr());
3140 QualType srcAddrType =
3141 CRE->getSubExpr()->getType()->castAs<PointerType>()->getPointeeType();
3142 srcLV = CGF.MakeAddrLValue(srcAddr, srcAddrType);
3144 Address srcAddr = srcLV.getAddress();
3146 // The dest and src types don't necessarily match in LLVM terms
3147 // because of the crazy ObjC compatibility rules.
3149 llvm::PointerType *destType =
3150 cast<llvm::PointerType>(CGF.ConvertType(CRE->getType()));
3152 // If the address is a constant null, just pass the appropriate null.
3153 if (isProvablyNull(srcAddr.getPointer())) {
3154 args.add(RValue::get(llvm::ConstantPointerNull::get(destType)),
3159 // Create the temporary.
3160 Address temp = CGF.CreateTempAlloca(destType->getElementType(),
3161 CGF.getPointerAlign(),
3163 // Loading an l-value can introduce a cleanup if the l-value is __weak,
3164 // and that cleanup will be conditional if we can't prove that the l-value
3165 // isn't null, so we need to register a dominating point so that the cleanups
3166 // system will make valid IR.
3167 CodeGenFunction::ConditionalEvaluation condEval(CGF);
3169 // Zero-initialize it if we're not doing a copy-initialization.
3170 bool shouldCopy = CRE->shouldCopy();
3173 llvm::ConstantPointerNull::get(
3174 cast<llvm::PointerType>(destType->getElementType()));
3175 CGF.Builder.CreateStore(null, temp);
3178 llvm::BasicBlock *contBB = nullptr;
3179 llvm::BasicBlock *originBB = nullptr;
3181 // If the address is *not* known to be non-null, we need to switch.
3182 llvm::Value *finalArgument;
3184 bool provablyNonNull = llvm::isKnownNonNull(srcAddr.getPointer());
3185 if (provablyNonNull) {
3186 finalArgument = temp.getPointer();
3188 llvm::Value *isNull =
3189 CGF.Builder.CreateIsNull(srcAddr.getPointer(), "icr.isnull");
3191 finalArgument = CGF.Builder.CreateSelect(isNull,
3192 llvm::ConstantPointerNull::get(destType),
3193 temp.getPointer(), "icr.argument");
3195 // If we need to copy, then the load has to be conditional, which
3196 // means we need control flow.
3198 originBB = CGF.Builder.GetInsertBlock();
3199 contBB = CGF.createBasicBlock("icr.cont");
3200 llvm::BasicBlock *copyBB = CGF.createBasicBlock("icr.copy");
3201 CGF.Builder.CreateCondBr(isNull, contBB, copyBB);
3202 CGF.EmitBlock(copyBB);
3203 condEval.begin(CGF);
3207 llvm::Value *valueToUse = nullptr;
3209 // Perform a copy if necessary.
3211 RValue srcRV = CGF.EmitLoadOfLValue(srcLV, SourceLocation());
3212 assert(srcRV.isScalar());
3214 llvm::Value *src = srcRV.getScalarVal();
3215 src = CGF.Builder.CreateBitCast(src, destType->getElementType(),
3218 // Use an ordinary store, not a store-to-lvalue.
3219 CGF.Builder.CreateStore(src, temp);
3221 // If optimization is enabled, and the value was held in a
3222 // __strong variable, we need to tell the optimizer that this
3223 // value has to stay alive until we're doing the store back.
3224 // This is because the temporary is effectively unretained,
3225 // and so otherwise we can violate the high-level semantics.
3226 if (CGF.CGM.getCodeGenOpts().OptimizationLevel != 0 &&
3227 srcLV.getObjCLifetime() == Qualifiers::OCL_Strong) {
3232 // Finish the control flow if we needed it.
3233 if (shouldCopy && !provablyNonNull) {
3234 llvm::BasicBlock *copyBB = CGF.Builder.GetInsertBlock();
3235 CGF.EmitBlock(contBB);
3237 // Make a phi for the value to intrinsically use.
3239 llvm::PHINode *phiToUse = CGF.Builder.CreatePHI(valueToUse->getType(), 2,
3241 phiToUse->addIncoming(valueToUse, copyBB);
3242 phiToUse->addIncoming(llvm::UndefValue::get(valueToUse->getType()),
3244 valueToUse = phiToUse;
3250 args.addWriteback(srcLV, temp, valueToUse);
3251 args.add(RValue::get(finalArgument), CRE->getType());
3254 void CallArgList::allocateArgumentMemory(CodeGenFunction &CGF) {
3258 llvm::Function *F = CGF.CGM.getIntrinsic(llvm::Intrinsic::stacksave);
3259 StackBase = CGF.Builder.CreateCall(F, {}, "inalloca.save");
3262 void CallArgList::freeArgumentMemory(CodeGenFunction &CGF) const {
3264 // Restore the stack after the call.
3265 llvm::Value *F = CGF.CGM.getIntrinsic(llvm::Intrinsic::stackrestore);
3266 CGF.Builder.CreateCall(F, StackBase);
3270 void CodeGenFunction::EmitNonNullArgCheck(RValue RV, QualType ArgType,
3271 SourceLocation ArgLoc,
3274 if (!AC.getDecl() || !(SanOpts.has(SanitizerKind::NonnullAttribute) ||
3275 SanOpts.has(SanitizerKind::NullabilityArg)))
3278 // The param decl may be missing in a variadic function.
3279 auto PVD = ParmNum < AC.getNumParams() ? AC.getParamDecl(ParmNum) : nullptr;
3280 unsigned ArgNo = PVD ? PVD->getFunctionScopeIndex() : ParmNum;
3282 // Prefer the nonnull attribute if it's present.
3283 const NonNullAttr *NNAttr = nullptr;
3284 if (SanOpts.has(SanitizerKind::NonnullAttribute))
3285 NNAttr = getNonNullAttr(AC.getDecl(), PVD, ArgType, ArgNo);
3287 bool CanCheckNullability = false;
3288 if (SanOpts.has(SanitizerKind::NullabilityArg) && !NNAttr && PVD) {
3289 auto Nullability = PVD->getType()->getNullability(getContext());
3290 CanCheckNullability = Nullability &&
3291 *Nullability == NullabilityKind::NonNull &&
3292 PVD->getTypeSourceInfo();
3295 if (!NNAttr && !CanCheckNullability)
3298 SourceLocation AttrLoc;
3299 SanitizerMask CheckKind;
3300 SanitizerHandler Handler;
3302 AttrLoc = NNAttr->getLocation();
3303 CheckKind = SanitizerKind::NonnullAttribute;
3304 Handler = SanitizerHandler::NonnullArg;
3306 AttrLoc = PVD->getTypeSourceInfo()->getTypeLoc().findNullabilityLoc();
3307 CheckKind = SanitizerKind::NullabilityArg;
3308 Handler = SanitizerHandler::NullabilityArg;
3311 SanitizerScope SanScope(this);
3312 assert(RV.isScalar());
3313 llvm::Value *V = RV.getScalarVal();
3315 Builder.CreateICmpNE(V, llvm::Constant::getNullValue(V->getType()));
3316 llvm::Constant *StaticData[] = {
3317 EmitCheckSourceLocation(ArgLoc), EmitCheckSourceLocation(AttrLoc),
3318 llvm::ConstantInt::get(Int32Ty, ArgNo + 1),
3320 EmitCheck(std::make_pair(Cond, CheckKind), Handler, StaticData, None);
3323 void CodeGenFunction::EmitCallArgs(
3324 CallArgList &Args, ArrayRef<QualType> ArgTypes,
3325 llvm::iterator_range<CallExpr::const_arg_iterator> ArgRange,
3326 AbstractCallee AC, unsigned ParamsToSkip, EvaluationOrder Order) {
3327 assert((int)ArgTypes.size() == (ArgRange.end() - ArgRange.begin()));
3329 // We *have* to evaluate arguments from right to left in the MS C++ ABI,
3330 // because arguments are destroyed left to right in the callee. As a special
3331 // case, there are certain language constructs that require left-to-right
3332 // evaluation, and in those cases we consider the evaluation order requirement
3333 // to trump the "destruction order is reverse construction order" guarantee.
3335 CGM.getTarget().getCXXABI().areArgsDestroyedLeftToRightInCallee()
3336 ? Order == EvaluationOrder::ForceLeftToRight
3337 : Order != EvaluationOrder::ForceRightToLeft;
3339 auto MaybeEmitImplicitObjectSize = [&](unsigned I, const Expr *Arg,
3340 RValue EmittedArg) {
3341 if (!AC.hasFunctionDecl() || I >= AC.getNumParams())
3343 auto *PS = AC.getParamDecl(I)->getAttr<PassObjectSizeAttr>();
3347 const auto &Context = getContext();
3348 auto SizeTy = Context.getSizeType();
3349 auto T = Builder.getIntNTy(Context.getTypeSize(SizeTy));
3350 assert(EmittedArg.getScalarVal() && "We emitted nothing for the arg?");
3351 llvm::Value *V = evaluateOrEmitBuiltinObjectSize(Arg, PS->getType(), T,
3352 EmittedArg.getScalarVal());
3353 Args.add(RValue::get(V), SizeTy);
3354 // If we're emitting args in reverse, be sure to do so with
3355 // pass_object_size, as well.
3357 std::swap(Args.back(), *(&Args.back() - 1));
3360 // Insert a stack save if we're going to need any inalloca args.
3361 bool HasInAllocaArgs = false;
3362 if (CGM.getTarget().getCXXABI().isMicrosoft()) {
3363 for (ArrayRef<QualType>::iterator I = ArgTypes.begin(), E = ArgTypes.end();
3364 I != E && !HasInAllocaArgs; ++I)
3365 HasInAllocaArgs = isInAllocaArgument(CGM.getCXXABI(), *I);
3366 if (HasInAllocaArgs) {
3367 assert(getTarget().getTriple().getArch() == llvm::Triple::x86);
3368 Args.allocateArgumentMemory(*this);
3372 // Evaluate each argument in the appropriate order.
3373 size_t CallArgsStart = Args.size();
3374 for (unsigned I = 0, E = ArgTypes.size(); I != E; ++I) {
3375 unsigned Idx = LeftToRight ? I : E - I - 1;
3376 CallExpr::const_arg_iterator Arg = ArgRange.begin() + Idx;
3377 unsigned InitialArgSize = Args.size();
3378 EmitCallArg(Args, *Arg, ArgTypes[Idx]);
3379 // In particular, we depend on it being the last arg in Args, and the
3380 // objectsize bits depend on there only being one arg if !LeftToRight.
3381 assert(InitialArgSize + 1 == Args.size() &&
3382 "The code below depends on only adding one arg per EmitCallArg");
3383 (void)InitialArgSize;
3384 RValue RVArg = Args.back().RV;
3385 EmitNonNullArgCheck(RVArg, ArgTypes[Idx], (*Arg)->getExprLoc(), AC,
3386 ParamsToSkip + Idx);
3387 // @llvm.objectsize should never have side-effects and shouldn't need
3388 // destruction/cleanups, so we can safely "emit" it after its arg,
3389 // regardless of right-to-leftness
3390 MaybeEmitImplicitObjectSize(Idx, *Arg, RVArg);
3394 // Un-reverse the arguments we just evaluated so they match up with the LLVM
3396 std::reverse(Args.begin() + CallArgsStart, Args.end());
3402 struct DestroyUnpassedArg final : EHScopeStack::Cleanup {
3403 DestroyUnpassedArg(Address Addr, QualType Ty)
3404 : Addr(Addr), Ty(Ty) {}
3409 void Emit(CodeGenFunction &CGF, Flags flags) override {
3410 const CXXDestructorDecl *Dtor = Ty->getAsCXXRecordDecl()->getDestructor();
3411 assert(!Dtor->isTrivial());
3412 CGF.EmitCXXDestructorCall(Dtor, Dtor_Complete, /*for vbase*/ false,
3413 /*Delegating=*/false, Addr);
3417 struct DisableDebugLocationUpdates {
3418 CodeGenFunction &CGF;
3419 bool disabledDebugInfo;
3420 DisableDebugLocationUpdates(CodeGenFunction &CGF, const Expr *E) : CGF(CGF) {
3421 if ((disabledDebugInfo = isa<CXXDefaultArgExpr>(E) && CGF.getDebugInfo()))
3422 CGF.disableDebugInfo();
3424 ~DisableDebugLocationUpdates() {
3425 if (disabledDebugInfo)
3426 CGF.enableDebugInfo();
3430 } // end anonymous namespace
3432 void CodeGenFunction::EmitCallArg(CallArgList &args, const Expr *E,
3434 DisableDebugLocationUpdates Dis(*this, E);
3435 if (const ObjCIndirectCopyRestoreExpr *CRE
3436 = dyn_cast<ObjCIndirectCopyRestoreExpr>(E)) {
3437 assert(getLangOpts().ObjCAutoRefCount);
3438 assert(getContext().hasSameUnqualifiedType(E->getType(), type));
3439 return emitWritebackArg(*this, args, CRE);
3442 assert(type->isReferenceType() == E->isGLValue() &&
3443 "reference binding to unmaterialized r-value!");
3445 if (E->isGLValue()) {
3446 assert(E->getObjectKind() == OK_Ordinary);
3447 return args.add(EmitReferenceBindingToExpr(E), type);
3450 bool HasAggregateEvalKind = hasAggregateEvaluationKind(type);
3452 // In the Microsoft C++ ABI, aggregate arguments are destructed by the callee.
3453 // However, we still have to push an EH-only cleanup in case we unwind before
3454 // we make it to the call.
3455 if (HasAggregateEvalKind &&
3456 CGM.getTarget().getCXXABI().areArgsDestroyedLeftToRightInCallee()) {
3457 // If we're using inalloca, use the argument memory. Otherwise, use a
3460 if (args.isUsingInAlloca())
3461 Slot = createPlaceholderSlot(*this, type);
3463 Slot = CreateAggTemp(type, "agg.tmp");
3465 const CXXRecordDecl *RD = type->getAsCXXRecordDecl();
3466 bool DestroyedInCallee =
3467 RD && RD->hasNonTrivialDestructor() &&
3468 CGM.getCXXABI().getRecordArgABI(RD) != CGCXXABI::RAA_Default;
3469 if (DestroyedInCallee)
3470 Slot.setExternallyDestructed();
3472 EmitAggExpr(E, Slot);
3473 RValue RV = Slot.asRValue();
3476 if (DestroyedInCallee) {
3477 // Create a no-op GEP between the placeholder and the cleanup so we can
3478 // RAUW it successfully. It also serves as a marker of the first
3479 // instruction where the cleanup is active.
3480 pushFullExprCleanup<DestroyUnpassedArg>(EHCleanup, Slot.getAddress(),
3482 // This unreachable is a temporary marker which will be removed later.
3483 llvm::Instruction *IsActive = Builder.CreateUnreachable();
3484 args.addArgCleanupDeactivation(EHStack.getInnermostEHScope(), IsActive);
3489 if (HasAggregateEvalKind && isa<ImplicitCastExpr>(E) &&
3490 cast<CastExpr>(E)->getCastKind() == CK_LValueToRValue) {
3491 LValue L = EmitLValue(cast<CastExpr>(E)->getSubExpr());
3492 assert(L.isSimple());
3493 if (L.getAlignment() >= getContext().getTypeAlignInChars(type)) {
3494 args.add(L.asAggregateRValue(), type, /*NeedsCopy*/true);
3496 // We can't represent a misaligned lvalue in the CallArgList, so copy
3497 // to an aligned temporary now.
3498 Address tmp = CreateMemTemp(type);
3499 EmitAggregateCopy(tmp, L.getAddress(), type, L.isVolatile());
3500 args.add(RValue::getAggregate(tmp), type);
3505 args.add(EmitAnyExprToTemp(E), type);
3508 QualType CodeGenFunction::getVarArgType(const Expr *Arg) {
3509 // System headers on Windows define NULL to 0 instead of 0LL on Win64. MSVC
3510 // implicitly widens null pointer constants that are arguments to varargs
3511 // functions to pointer-sized ints.
3512 if (!getTarget().getTriple().isOSWindows())
3513 return Arg->getType();
3515 if (Arg->getType()->isIntegerType() &&
3516 getContext().getTypeSize(Arg->getType()) <
3517 getContext().getTargetInfo().getPointerWidth(0) &&
3518 Arg->isNullPointerConstant(getContext(),
3519 Expr::NPC_ValueDependentIsNotNull)) {
3520 return getContext().getIntPtrType();
3523 return Arg->getType();
3526 // In ObjC ARC mode with no ObjC ARC exception safety, tell the ARC
3527 // optimizer it can aggressively ignore unwind edges.
3529 CodeGenFunction::AddObjCARCExceptionMetadata(llvm::Instruction *Inst) {
3530 if (CGM.getCodeGenOpts().OptimizationLevel != 0 &&
3531 !CGM.getCodeGenOpts().ObjCAutoRefCountExceptions)
3532 Inst->setMetadata("clang.arc.no_objc_arc_exceptions",
3533 CGM.getNoObjCARCExceptionsMetadata());
3536 /// Emits a call to the given no-arguments nounwind runtime function.
3538 CodeGenFunction::EmitNounwindRuntimeCall(llvm::Value *callee,
3539 const llvm::Twine &name) {
3540 return EmitNounwindRuntimeCall(callee, None, name);
3543 /// Emits a call to the given nounwind runtime function.
3545 CodeGenFunction::EmitNounwindRuntimeCall(llvm::Value *callee,
3546 ArrayRef<llvm::Value*> args,
3547 const llvm::Twine &name) {
3548 llvm::CallInst *call = EmitRuntimeCall(callee, args, name);
3549 call->setDoesNotThrow();
3553 /// Emits a simple call (never an invoke) to the given no-arguments
3554 /// runtime function.
3556 CodeGenFunction::EmitRuntimeCall(llvm::Value *callee,
3557 const llvm::Twine &name) {
3558 return EmitRuntimeCall(callee, None, name);
3561 // Calls which may throw must have operand bundles indicating which funclet
3562 // they are nested within.
3564 getBundlesForFunclet(llvm::Value *Callee, llvm::Instruction *CurrentFuncletPad,
3565 SmallVectorImpl<llvm::OperandBundleDef> &BundleList) {
3566 // There is no need for a funclet operand bundle if we aren't inside a
3568 if (!CurrentFuncletPad)
3571 // Skip intrinsics which cannot throw.
3572 auto *CalleeFn = dyn_cast<llvm::Function>(Callee->stripPointerCasts());
3573 if (CalleeFn && CalleeFn->isIntrinsic() && CalleeFn->doesNotThrow())
3576 BundleList.emplace_back("funclet", CurrentFuncletPad);
3579 /// Emits a simple call (never an invoke) to the given runtime function.
3581 CodeGenFunction::EmitRuntimeCall(llvm::Value *callee,
3582 ArrayRef<llvm::Value*> args,
3583 const llvm::Twine &name) {
3584 SmallVector<llvm::OperandBundleDef, 1> BundleList;
3585 getBundlesForFunclet(callee, CurrentFuncletPad, BundleList);
3587 llvm::CallInst *call = Builder.CreateCall(callee, args, BundleList, name);
3588 call->setCallingConv(getRuntimeCC());
3592 /// Emits a call or invoke to the given noreturn runtime function.
3593 void CodeGenFunction::EmitNoreturnRuntimeCallOrInvoke(llvm::Value *callee,
3594 ArrayRef<llvm::Value*> args) {
3595 SmallVector<llvm::OperandBundleDef, 1> BundleList;
3596 getBundlesForFunclet(callee, CurrentFuncletPad, BundleList);
3598 if (getInvokeDest()) {
3599 llvm::InvokeInst *invoke =
3600 Builder.CreateInvoke(callee,
3601 getUnreachableBlock(),
3605 invoke->setDoesNotReturn();
3606 invoke->setCallingConv(getRuntimeCC());
3608 llvm::CallInst *call = Builder.CreateCall(callee, args, BundleList);
3609 call->setDoesNotReturn();
3610 call->setCallingConv(getRuntimeCC());
3611 Builder.CreateUnreachable();
3615 /// Emits a call or invoke instruction to the given nullary runtime function.
3617 CodeGenFunction::EmitRuntimeCallOrInvoke(llvm::Value *callee,
3618 const Twine &name) {
3619 return EmitRuntimeCallOrInvoke(callee, None, name);
3622 /// Emits a call or invoke instruction to the given runtime function.
3624 CodeGenFunction::EmitRuntimeCallOrInvoke(llvm::Value *callee,
3625 ArrayRef<llvm::Value*> args,
3626 const Twine &name) {
3627 llvm::CallSite callSite = EmitCallOrInvoke(callee, args, name);
3628 callSite.setCallingConv(getRuntimeCC());
3632 /// Emits a call or invoke instruction to the given function, depending
3633 /// on the current state of the EH stack.
3635 CodeGenFunction::EmitCallOrInvoke(llvm::Value *Callee,
3636 ArrayRef<llvm::Value *> Args,
3637 const Twine &Name) {
3638 llvm::BasicBlock *InvokeDest = getInvokeDest();
3639 SmallVector<llvm::OperandBundleDef, 1> BundleList;
3640 getBundlesForFunclet(Callee, CurrentFuncletPad, BundleList);
3642 llvm::Instruction *Inst;
3644 Inst = Builder.CreateCall(Callee, Args, BundleList, Name);
3646 llvm::BasicBlock *ContBB = createBasicBlock("invoke.cont");
3647 Inst = Builder.CreateInvoke(Callee, ContBB, InvokeDest, Args, BundleList,
3652 // In ObjC ARC mode with no ObjC ARC exception safety, tell the ARC
3653 // optimizer it can aggressively ignore unwind edges.
3654 if (CGM.getLangOpts().ObjCAutoRefCount)
3655 AddObjCARCExceptionMetadata(Inst);
3657 return llvm::CallSite(Inst);
3660 /// \brief Store a non-aggregate value to an address to initialize it. For
3661 /// initialization, a non-atomic store will be used.
3662 static void EmitInitStoreOfNonAggregate(CodeGenFunction &CGF, RValue Src,
3665 CGF.EmitStoreOfScalar(Src.getScalarVal(), Dst, /*init=*/true);
3667 CGF.EmitStoreOfComplex(Src.getComplexVal(), Dst, /*init=*/true);
3670 void CodeGenFunction::deferPlaceholderReplacement(llvm::Instruction *Old,
3672 DeferredReplacements.push_back(std::make_pair(Old, New));
3675 RValue CodeGenFunction::EmitCall(const CGFunctionInfo &CallInfo,
3676 const CGCallee &Callee,
3677 ReturnValueSlot ReturnValue,
3678 const CallArgList &CallArgs,
3679 llvm::Instruction **callOrInvoke) {
3680 // FIXME: We no longer need the types from CallArgs; lift up and simplify.
3682 assert(Callee.isOrdinary());
3684 // Handle struct-return functions by passing a pointer to the
3685 // location that we would like to return into.
3686 QualType RetTy = CallInfo.getReturnType();
3687 const ABIArgInfo &RetAI = CallInfo.getReturnInfo();
3689 llvm::FunctionType *IRFuncTy = Callee.getFunctionType();
3691 // 1. Set up the arguments.
3693 // If we're using inalloca, insert the allocation after the stack save.
3694 // FIXME: Do this earlier rather than hacking it in here!
3695 Address ArgMemory = Address::invalid();
3696 const llvm::StructLayout *ArgMemoryLayout = nullptr;
3697 if (llvm::StructType *ArgStruct = CallInfo.getArgStruct()) {
3698 const llvm::DataLayout &DL = CGM.getDataLayout();
3699 ArgMemoryLayout = DL.getStructLayout(ArgStruct);
3700 llvm::Instruction *IP = CallArgs.getStackBase();
3701 llvm::AllocaInst *AI;
3703 IP = IP->getNextNode();
3704 AI = new llvm::AllocaInst(ArgStruct, DL.getAllocaAddrSpace(),
3707 AI = CreateTempAlloca(ArgStruct, "argmem");
3709 auto Align = CallInfo.getArgStructAlignment();
3710 AI->setAlignment(Align.getQuantity());
3711 AI->setUsedWithInAlloca(true);
3712 assert(AI->isUsedWithInAlloca() && !AI->isStaticAlloca());
3713 ArgMemory = Address(AI, Align);
3716 // Helper function to drill into the inalloca allocation.
3717 auto createInAllocaStructGEP = [&](unsigned FieldIndex) -> Address {
3719 CharUnits::fromQuantity(ArgMemoryLayout->getElementOffset(FieldIndex));
3720 return Builder.CreateStructGEP(ArgMemory, FieldIndex, FieldOffset);
3723 ClangToLLVMArgMapping IRFunctionArgs(CGM.getContext(), CallInfo);
3724 SmallVector<llvm::Value *, 16> IRCallArgs(IRFunctionArgs.totalIRArgs());
3726 // If the call returns a temporary with struct return, create a temporary
3727 // alloca to hold the result, unless one is given to us.
3728 Address SRetPtr = Address::invalid();
3729 size_t UnusedReturnSize = 0;
3730 if (RetAI.isIndirect() || RetAI.isInAlloca() || RetAI.isCoerceAndExpand()) {
3731 if (!ReturnValue.isNull()) {
3732 SRetPtr = ReturnValue.getValue();
3734 SRetPtr = CreateMemTemp(RetTy);
3735 if (HaveInsertPoint() && ReturnValue.isUnused()) {
3737 CGM.getDataLayout().getTypeAllocSize(ConvertTypeForMem(RetTy));
3738 if (EmitLifetimeStart(size, SRetPtr.getPointer()))
3739 UnusedReturnSize = size;
3742 if (IRFunctionArgs.hasSRetArg()) {
3743 IRCallArgs[IRFunctionArgs.getSRetArgNo()] = SRetPtr.getPointer();
3744 } else if (RetAI.isInAlloca()) {
3745 Address Addr = createInAllocaStructGEP(RetAI.getInAllocaFieldIndex());
3746 Builder.CreateStore(SRetPtr.getPointer(), Addr);
3750 Address swiftErrorTemp = Address::invalid();
3751 Address swiftErrorArg = Address::invalid();
3753 // Translate all of the arguments as necessary to match the IR lowering.
3754 assert(CallInfo.arg_size() == CallArgs.size() &&
3755 "Mismatch between function signature & arguments.");
3757 CGFunctionInfo::const_arg_iterator info_it = CallInfo.arg_begin();
3758 for (CallArgList::const_iterator I = CallArgs.begin(), E = CallArgs.end();
3759 I != E; ++I, ++info_it, ++ArgNo) {
3760 const ABIArgInfo &ArgInfo = info_it->info;
3763 // Insert a padding argument to ensure proper alignment.
3764 if (IRFunctionArgs.hasPaddingArg(ArgNo))
3765 IRCallArgs[IRFunctionArgs.getPaddingArgNo(ArgNo)] =
3766 llvm::UndefValue::get(ArgInfo.getPaddingType());
3768 unsigned FirstIRArg, NumIRArgs;
3769 std::tie(FirstIRArg, NumIRArgs) = IRFunctionArgs.getIRArgs(ArgNo);
3771 switch (ArgInfo.getKind()) {
3772 case ABIArgInfo::InAlloca: {
3773 assert(NumIRArgs == 0);
3774 assert(getTarget().getTriple().getArch() == llvm::Triple::x86);
3775 if (RV.isAggregate()) {
3776 // Replace the placeholder with the appropriate argument slot GEP.
3777 llvm::Instruction *Placeholder =
3778 cast<llvm::Instruction>(RV.getAggregatePointer());
3779 CGBuilderTy::InsertPoint IP = Builder.saveIP();
3780 Builder.SetInsertPoint(Placeholder);
3781 Address Addr = createInAllocaStructGEP(ArgInfo.getInAllocaFieldIndex());
3782 Builder.restoreIP(IP);
3783 deferPlaceholderReplacement(Placeholder, Addr.getPointer());
3785 // Store the RValue into the argument struct.
3786 Address Addr = createInAllocaStructGEP(ArgInfo.getInAllocaFieldIndex());
3787 unsigned AS = Addr.getType()->getPointerAddressSpace();
3788 llvm::Type *MemType = ConvertTypeForMem(I->Ty)->getPointerTo(AS);
3789 // There are some cases where a trivial bitcast is not avoidable. The
3790 // definition of a type later in a translation unit may change it's type
3791 // from {}* to (%struct.foo*)*.
3792 if (Addr.getType() != MemType)
3793 Addr = Builder.CreateBitCast(Addr, MemType);
3794 LValue argLV = MakeAddrLValue(Addr, I->Ty);
3795 EmitInitStoreOfNonAggregate(*this, RV, argLV);
3800 case ABIArgInfo::Indirect: {
3801 assert(NumIRArgs == 1);
3802 if (RV.isScalar() || RV.isComplex()) {
3803 // Make a temporary alloca to pass the argument.
3804 Address Addr = CreateMemTemp(I->Ty, ArgInfo.getIndirectAlign());
3805 IRCallArgs[FirstIRArg] = Addr.getPointer();
3807 LValue argLV = MakeAddrLValue(Addr, I->Ty);
3808 EmitInitStoreOfNonAggregate(*this, RV, argLV);
3810 // We want to avoid creating an unnecessary temporary+copy here;
3811 // however, we need one in three cases:
3812 // 1. If the argument is not byval, and we are required to copy the
3813 // source. (This case doesn't occur on any common architecture.)
3814 // 2. If the argument is byval, RV is not sufficiently aligned, and
3815 // we cannot force it to be sufficiently aligned.
3816 // 3. If the argument is byval, but RV is located in an address space
3817 // different than that of the argument (0).
3818 Address Addr = RV.getAggregateAddress();
3819 CharUnits Align = ArgInfo.getIndirectAlign();
3820 const llvm::DataLayout *TD = &CGM.getDataLayout();
3821 const unsigned RVAddrSpace = Addr.getType()->getAddressSpace();
3822 const unsigned ArgAddrSpace =
3823 (FirstIRArg < IRFuncTy->getNumParams()
3824 ? IRFuncTy->getParamType(FirstIRArg)->getPointerAddressSpace()
3826 if ((!ArgInfo.getIndirectByVal() && I->NeedsCopy) ||
3827 (ArgInfo.getIndirectByVal() && Addr.getAlignment() < Align &&
3828 llvm::getOrEnforceKnownAlignment(Addr.getPointer(),
3829 Align.getQuantity(), *TD)
3830 < Align.getQuantity()) ||
3831 (ArgInfo.getIndirectByVal() && (RVAddrSpace != ArgAddrSpace))) {
3832 // Create an aligned temporary, and copy to it.
3833 Address AI = CreateMemTemp(I->Ty, ArgInfo.getIndirectAlign());
3834 IRCallArgs[FirstIRArg] = AI.getPointer();
3835 EmitAggregateCopy(AI, Addr, I->Ty, RV.isVolatileQualified());
3837 // Skip the extra memcpy call.
3838 IRCallArgs[FirstIRArg] = Addr.getPointer();
3844 case ABIArgInfo::Ignore:
3845 assert(NumIRArgs == 0);
3848 case ABIArgInfo::Extend:
3849 case ABIArgInfo::Direct: {
3850 if (!isa<llvm::StructType>(ArgInfo.getCoerceToType()) &&
3851 ArgInfo.getCoerceToType() == ConvertType(info_it->type) &&
3852 ArgInfo.getDirectOffset() == 0) {
3853 assert(NumIRArgs == 1);
3856 V = RV.getScalarVal();
3858 V = Builder.CreateLoad(RV.getAggregateAddress());
3860 // Implement swifterror by copying into a new swifterror argument.
3861 // We'll write back in the normal path out of the call.
3862 if (CallInfo.getExtParameterInfo(ArgNo).getABI()
3863 == ParameterABI::SwiftErrorResult) {
3864 assert(!swiftErrorTemp.isValid() && "multiple swifterror args");
3866 QualType pointeeTy = I->Ty->getPointeeType();
3868 Address(V, getContext().getTypeAlignInChars(pointeeTy));
3871 CreateMemTemp(pointeeTy, getPointerAlign(), "swifterror.temp");
3872 V = swiftErrorTemp.getPointer();
3873 cast<llvm::AllocaInst>(V)->setSwiftError(true);
3875 llvm::Value *errorValue = Builder.CreateLoad(swiftErrorArg);
3876 Builder.CreateStore(errorValue, swiftErrorTemp);
3879 // We might have to widen integers, but we should never truncate.
3880 if (ArgInfo.getCoerceToType() != V->getType() &&
3881 V->getType()->isIntegerTy())
3882 V = Builder.CreateZExt(V, ArgInfo.getCoerceToType());
3884 // If the argument doesn't match, perform a bitcast to coerce it. This
3885 // can happen due to trivial type mismatches.
3886 if (FirstIRArg < IRFuncTy->getNumParams() &&
3887 V->getType() != IRFuncTy->getParamType(FirstIRArg))
3888 V = Builder.CreateBitCast(V, IRFuncTy->getParamType(FirstIRArg));
3890 IRCallArgs[FirstIRArg] = V;
3894 // FIXME: Avoid the conversion through memory if possible.
3895 Address Src = Address::invalid();
3896 if (RV.isScalar() || RV.isComplex()) {
3897 Src = CreateMemTemp(I->Ty, "coerce");
3898 LValue SrcLV = MakeAddrLValue(Src, I->Ty);
3899 EmitInitStoreOfNonAggregate(*this, RV, SrcLV);
3901 Src = RV.getAggregateAddress();
3904 // If the value is offset in memory, apply the offset now.
3905 Src = emitAddressAtOffset(*this, Src, ArgInfo);
3907 // Fast-isel and the optimizer generally like scalar values better than
3908 // FCAs, so we flatten them if this is safe to do for this argument.
3909 llvm::StructType *STy =
3910 dyn_cast<llvm::StructType>(ArgInfo.getCoerceToType());
3911 if (STy && ArgInfo.isDirect() && ArgInfo.getCanBeFlattened()) {
3912 llvm::Type *SrcTy = Src.getType()->getElementType();
3913 uint64_t SrcSize = CGM.getDataLayout().getTypeAllocSize(SrcTy);
3914 uint64_t DstSize = CGM.getDataLayout().getTypeAllocSize(STy);
3916 // If the source type is smaller than the destination type of the
3917 // coerce-to logic, copy the source value into a temp alloca the size
3918 // of the destination type to allow loading all of it. The bits past
3919 // the source value are left undef.
3920 if (SrcSize < DstSize) {
3922 = CreateTempAlloca(STy, Src.getAlignment(),
3923 Src.getName() + ".coerce");
3924 Builder.CreateMemCpy(TempAlloca, Src, SrcSize);
3927 Src = Builder.CreateBitCast(Src, llvm::PointerType::getUnqual(STy));
3930 auto SrcLayout = CGM.getDataLayout().getStructLayout(STy);
3931 assert(NumIRArgs == STy->getNumElements());
3932 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
3933 auto Offset = CharUnits::fromQuantity(SrcLayout->getElementOffset(i));
3934 Address EltPtr = Builder.CreateStructGEP(Src, i, Offset);
3935 llvm::Value *LI = Builder.CreateLoad(EltPtr);
3936 IRCallArgs[FirstIRArg + i] = LI;
3939 // In the simple case, just pass the coerced loaded value.
3940 assert(NumIRArgs == 1);
3941 IRCallArgs[FirstIRArg] =
3942 CreateCoercedLoad(Src, ArgInfo.getCoerceToType(), *this);
3948 case ABIArgInfo::CoerceAndExpand: {
3949 auto coercionType = ArgInfo.getCoerceAndExpandType();
3950 auto layout = CGM.getDataLayout().getStructLayout(coercionType);
3952 llvm::Value *tempSize = nullptr;
3953 Address addr = Address::invalid();
3954 if (RV.isAggregate()) {
3955 addr = RV.getAggregateAddress();
3957 assert(RV.isScalar()); // complex should always just be direct
3959 llvm::Type *scalarType = RV.getScalarVal()->getType();
3960 auto scalarSize = CGM.getDataLayout().getTypeAllocSize(scalarType);
3961 auto scalarAlign = CGM.getDataLayout().getPrefTypeAlignment(scalarType);
3963 tempSize = llvm::ConstantInt::get(CGM.Int64Ty, scalarSize);
3965 // Materialize to a temporary.
3966 addr = CreateTempAlloca(RV.getScalarVal()->getType(),
3967 CharUnits::fromQuantity(std::max(layout->getAlignment(),
3969 EmitLifetimeStart(scalarSize, addr.getPointer());
3971 Builder.CreateStore(RV.getScalarVal(), addr);
3974 addr = Builder.CreateElementBitCast(addr, coercionType);
3976 unsigned IRArgPos = FirstIRArg;
3977 for (unsigned i = 0, e = coercionType->getNumElements(); i != e; ++i) {
3978 llvm::Type *eltType = coercionType->getElementType(i);
3979 if (ABIArgInfo::isPaddingForCoerceAndExpand(eltType)) continue;
3980 Address eltAddr = Builder.CreateStructGEP(addr, i, layout);
3981 llvm::Value *elt = Builder.CreateLoad(eltAddr);
3982 IRCallArgs[IRArgPos++] = elt;
3984 assert(IRArgPos == FirstIRArg + NumIRArgs);
3987 EmitLifetimeEnd(tempSize, addr.getPointer());
3993 case ABIArgInfo::Expand:
3994 unsigned IRArgPos = FirstIRArg;
3995 ExpandTypeToArgs(I->Ty, RV, IRFuncTy, IRCallArgs, IRArgPos);
3996 assert(IRArgPos == FirstIRArg + NumIRArgs);
4001 llvm::Value *CalleePtr = Callee.getFunctionPointer();
4003 // If we're using inalloca, set up that argument.
4004 if (ArgMemory.isValid()) {
4005 llvm::Value *Arg = ArgMemory.getPointer();
4006 if (CallInfo.isVariadic()) {
4007 // When passing non-POD arguments by value to variadic functions, we will
4008 // end up with a variadic prototype and an inalloca call site. In such
4009 // cases, we can't do any parameter mismatch checks. Give up and bitcast
4011 unsigned CalleeAS = CalleePtr->getType()->getPointerAddressSpace();
4012 auto FnTy = getTypes().GetFunctionType(CallInfo)->getPointerTo(CalleeAS);
4013 CalleePtr = Builder.CreateBitCast(CalleePtr, FnTy);
4015 llvm::Type *LastParamTy =
4016 IRFuncTy->getParamType(IRFuncTy->getNumParams() - 1);
4017 if (Arg->getType() != LastParamTy) {
4019 // Assert that these structs have equivalent element types.
4020 llvm::StructType *FullTy = CallInfo.getArgStruct();
4021 llvm::StructType *DeclaredTy = cast<llvm::StructType>(
4022 cast<llvm::PointerType>(LastParamTy)->getElementType());
4023 assert(DeclaredTy->getNumElements() == FullTy->getNumElements());
4024 for (llvm::StructType::element_iterator DI = DeclaredTy->element_begin(),
4025 DE = DeclaredTy->element_end(),
4026 FI = FullTy->element_begin();
4027 DI != DE; ++DI, ++FI)
4030 Arg = Builder.CreateBitCast(Arg, LastParamTy);
4033 assert(IRFunctionArgs.hasInallocaArg());
4034 IRCallArgs[IRFunctionArgs.getInallocaArgNo()] = Arg;
4037 // 2. Prepare the function pointer.
4039 // If the callee is a bitcast of a non-variadic function to have a
4040 // variadic function pointer type, check to see if we can remove the
4041 // bitcast. This comes up with unprototyped functions.
4043 // This makes the IR nicer, but more importantly it ensures that we
4044 // can inline the function at -O0 if it is marked always_inline.
4045 auto simplifyVariadicCallee = [](llvm::Value *Ptr) -> llvm::Value* {
4046 llvm::FunctionType *CalleeFT =
4047 cast<llvm::FunctionType>(Ptr->getType()->getPointerElementType());
4048 if (!CalleeFT->isVarArg())
4051 llvm::ConstantExpr *CE = dyn_cast<llvm::ConstantExpr>(Ptr);
4052 if (!CE || CE->getOpcode() != llvm::Instruction::BitCast)
4055 llvm::Function *OrigFn = dyn_cast<llvm::Function>(CE->getOperand(0));
4059 llvm::FunctionType *OrigFT = OrigFn->getFunctionType();
4061 // If the original type is variadic, or if any of the component types
4062 // disagree, we cannot remove the cast.
4063 if (OrigFT->isVarArg() ||
4064 OrigFT->getNumParams() != CalleeFT->getNumParams() ||
4065 OrigFT->getReturnType() != CalleeFT->getReturnType())
4068 for (unsigned i = 0, e = OrigFT->getNumParams(); i != e; ++i)
4069 if (OrigFT->getParamType(i) != CalleeFT->getParamType(i))
4074 CalleePtr = simplifyVariadicCallee(CalleePtr);
4076 // 3. Perform the actual call.
4078 // Deactivate any cleanups that we're supposed to do immediately before
4080 if (!CallArgs.getCleanupsToDeactivate().empty())
4081 deactivateArgCleanupsBeforeCall(*this, CallArgs);
4083 // Assert that the arguments we computed match up. The IR verifier
4084 // will catch this, but this is a common enough source of problems
4085 // during IRGen changes that it's way better for debugging to catch
4086 // it ourselves here.
4088 assert(IRCallArgs.size() == IRFuncTy->getNumParams() || IRFuncTy->isVarArg());
4089 for (unsigned i = 0; i < IRCallArgs.size(); ++i) {
4090 // Inalloca argument can have different type.
4091 if (IRFunctionArgs.hasInallocaArg() &&
4092 i == IRFunctionArgs.getInallocaArgNo())
4094 if (i < IRFuncTy->getNumParams())
4095 assert(IRCallArgs[i]->getType() == IRFuncTy->getParamType(i));
4099 // Compute the calling convention and attributes.
4100 unsigned CallingConv;
4101 llvm::AttributeList Attrs;
4102 CGM.ConstructAttributeList(CalleePtr->getName(), CallInfo,
4103 Callee.getAbstractInfo(), Attrs, CallingConv,
4104 /*AttrOnCallSite=*/true);
4106 // Apply some call-site-specific attributes.
4107 // TODO: work this into building the attribute set.
4109 // Apply always_inline to all calls within flatten functions.
4110 // FIXME: should this really take priority over __try, below?
4111 if (CurCodeDecl && CurCodeDecl->hasAttr<FlattenAttr>() &&
4112 !(Callee.getAbstractInfo().getCalleeDecl() &&
4113 Callee.getAbstractInfo().getCalleeDecl()->hasAttr<NoInlineAttr>())) {
4115 Attrs.addAttribute(getLLVMContext(), llvm::AttributeList::FunctionIndex,
4116 llvm::Attribute::AlwaysInline);
4119 // Disable inlining inside SEH __try blocks.
4120 if (isSEHTryScope()) {
4122 Attrs.addAttribute(getLLVMContext(), llvm::AttributeList::FunctionIndex,
4123 llvm::Attribute::NoInline);
4126 // Decide whether to use a call or an invoke.
4128 if (currentFunctionUsesSEHTry()) {
4129 // SEH cares about asynchronous exceptions, so everything can "throw."
4130 CannotThrow = false;
4131 } else if (isCleanupPadScope() &&
4132 EHPersonality::get(*this).isMSVCXXPersonality()) {
4133 // The MSVC++ personality will implicitly terminate the program if an
4134 // exception is thrown during a cleanup outside of a try/catch.
4135 // We don't need to model anything in IR to get this behavior.
4138 // Otherwise, nounwind call sites will never throw.
4139 CannotThrow = Attrs.hasAttribute(llvm::AttributeList::FunctionIndex,
4140 llvm::Attribute::NoUnwind);
4142 llvm::BasicBlock *InvokeDest = CannotThrow ? nullptr : getInvokeDest();
4144 SmallVector<llvm::OperandBundleDef, 1> BundleList;
4145 getBundlesForFunclet(CalleePtr, CurrentFuncletPad, BundleList);
4147 // Emit the actual call/invoke instruction.
4150 CS = Builder.CreateCall(CalleePtr, IRCallArgs, BundleList);
4152 llvm::BasicBlock *Cont = createBasicBlock("invoke.cont");
4153 CS = Builder.CreateInvoke(CalleePtr, Cont, InvokeDest, IRCallArgs,
4157 llvm::Instruction *CI = CS.getInstruction();
4161 // Apply the attributes and calling convention.
4162 CS.setAttributes(Attrs);
4163 CS.setCallingConv(static_cast<llvm::CallingConv::ID>(CallingConv));
4165 // Apply various metadata.
4167 if (!CI->getType()->isVoidTy())
4168 CI->setName("call");
4170 // Insert instrumentation or attach profile metadata at indirect call sites.
4171 // For more details, see the comment before the definition of
4172 // IPVK_IndirectCallTarget in InstrProfData.inc.
4173 if (!CS.getCalledFunction())
4174 PGO.valueProfile(Builder, llvm::IPVK_IndirectCallTarget,
4177 // In ObjC ARC mode with no ObjC ARC exception safety, tell the ARC
4178 // optimizer it can aggressively ignore unwind edges.
4179 if (CGM.getLangOpts().ObjCAutoRefCount)
4180 AddObjCARCExceptionMetadata(CI);
4182 // Suppress tail calls if requested.
4183 if (llvm::CallInst *Call = dyn_cast<llvm::CallInst>(CI)) {
4184 const Decl *TargetDecl = Callee.getAbstractInfo().getCalleeDecl();
4185 if (TargetDecl && TargetDecl->hasAttr<NotTailCalledAttr>())
4186 Call->setTailCallKind(llvm::CallInst::TCK_NoTail);
4189 // 4. Finish the call.
4191 // If the call doesn't return, finish the basic block and clear the
4192 // insertion point; this allows the rest of IRGen to discard
4193 // unreachable code.
4194 if (CS.doesNotReturn()) {
4195 if (UnusedReturnSize)
4196 EmitLifetimeEnd(llvm::ConstantInt::get(Int64Ty, UnusedReturnSize),
4197 SRetPtr.getPointer());
4199 Builder.CreateUnreachable();
4200 Builder.ClearInsertionPoint();
4202 // FIXME: For now, emit a dummy basic block because expr emitters in
4203 // generally are not ready to handle emitting expressions at unreachable
4205 EnsureInsertPoint();
4207 // Return a reasonable RValue.
4208 return GetUndefRValue(RetTy);
4211 // Perform the swifterror writeback.
4212 if (swiftErrorTemp.isValid()) {
4213 llvm::Value *errorResult = Builder.CreateLoad(swiftErrorTemp);
4214 Builder.CreateStore(errorResult, swiftErrorArg);
4217 // Emit any call-associated writebacks immediately. Arguably this
4218 // should happen after any return-value munging.
4219 if (CallArgs.hasWritebacks())
4220 emitWritebacks(*this, CallArgs);
4222 // The stack cleanup for inalloca arguments has to run out of the normal
4223 // lexical order, so deactivate it and run it manually here.
4224 CallArgs.freeArgumentMemory(*this);
4226 // Extract the return value.
4228 switch (RetAI.getKind()) {
4229 case ABIArgInfo::CoerceAndExpand: {
4230 auto coercionType = RetAI.getCoerceAndExpandType();
4231 auto layout = CGM.getDataLayout().getStructLayout(coercionType);
4233 Address addr = SRetPtr;
4234 addr = Builder.CreateElementBitCast(addr, coercionType);
4236 assert(CI->getType() == RetAI.getUnpaddedCoerceAndExpandType());
4237 bool requiresExtract = isa<llvm::StructType>(CI->getType());
4239 unsigned unpaddedIndex = 0;
4240 for (unsigned i = 0, e = coercionType->getNumElements(); i != e; ++i) {
4241 llvm::Type *eltType = coercionType->getElementType(i);
4242 if (ABIArgInfo::isPaddingForCoerceAndExpand(eltType)) continue;
4243 Address eltAddr = Builder.CreateStructGEP(addr, i, layout);
4244 llvm::Value *elt = CI;
4245 if (requiresExtract)
4246 elt = Builder.CreateExtractValue(elt, unpaddedIndex++);
4248 assert(unpaddedIndex == 0);
4249 Builder.CreateStore(elt, eltAddr);
4254 case ABIArgInfo::InAlloca:
4255 case ABIArgInfo::Indirect: {
4256 RValue ret = convertTempToRValue(SRetPtr, RetTy, SourceLocation());
4257 if (UnusedReturnSize)
4258 EmitLifetimeEnd(llvm::ConstantInt::get(Int64Ty, UnusedReturnSize),
4259 SRetPtr.getPointer());
4263 case ABIArgInfo::Ignore:
4264 // If we are ignoring an argument that had a result, make sure to
4265 // construct the appropriate return value for our caller.
4266 return GetUndefRValue(RetTy);
4268 case ABIArgInfo::Extend:
4269 case ABIArgInfo::Direct: {
4270 llvm::Type *RetIRTy = ConvertType(RetTy);
4271 if (RetAI.getCoerceToType() == RetIRTy && RetAI.getDirectOffset() == 0) {
4272 switch (getEvaluationKind(RetTy)) {
4274 llvm::Value *Real = Builder.CreateExtractValue(CI, 0);
4275 llvm::Value *Imag = Builder.CreateExtractValue(CI, 1);
4276 return RValue::getComplex(std::make_pair(Real, Imag));
4278 case TEK_Aggregate: {
4279 Address DestPtr = ReturnValue.getValue();
4280 bool DestIsVolatile = ReturnValue.isVolatile();
4282 if (!DestPtr.isValid()) {
4283 DestPtr = CreateMemTemp(RetTy, "agg.tmp");
4284 DestIsVolatile = false;
4286 BuildAggStore(*this, CI, DestPtr, DestIsVolatile);
4287 return RValue::getAggregate(DestPtr);
4290 // If the argument doesn't match, perform a bitcast to coerce it. This
4291 // can happen due to trivial type mismatches.
4292 llvm::Value *V = CI;
4293 if (V->getType() != RetIRTy)
4294 V = Builder.CreateBitCast(V, RetIRTy);
4295 return RValue::get(V);
4298 llvm_unreachable("bad evaluation kind");
4301 Address DestPtr = ReturnValue.getValue();
4302 bool DestIsVolatile = ReturnValue.isVolatile();
4304 if (!DestPtr.isValid()) {
4305 DestPtr = CreateMemTemp(RetTy, "coerce");
4306 DestIsVolatile = false;
4309 // If the value is offset in memory, apply the offset now.
4310 Address StorePtr = emitAddressAtOffset(*this, DestPtr, RetAI);
4311 CreateCoercedStore(CI, StorePtr, DestIsVolatile, *this);
4313 return convertTempToRValue(DestPtr, RetTy, SourceLocation());
4316 case ABIArgInfo::Expand:
4317 llvm_unreachable("Invalid ABI kind for return argument");
4320 llvm_unreachable("Unhandled ABIArgInfo::Kind");
4323 // Emit the assume_aligned check on the return value.
4324 const Decl *TargetDecl = Callee.getAbstractInfo().getCalleeDecl();
4325 if (Ret.isScalar() && TargetDecl) {
4326 if (const auto *AA = TargetDecl->getAttr<AssumeAlignedAttr>()) {
4327 llvm::Value *OffsetValue = nullptr;
4328 if (const auto *Offset = AA->getOffset())
4329 OffsetValue = EmitScalarExpr(Offset);
4331 llvm::Value *Alignment = EmitScalarExpr(AA->getAlignment());
4332 llvm::ConstantInt *AlignmentCI = cast<llvm::ConstantInt>(Alignment);
4333 EmitAlignmentAssumption(Ret.getScalarVal(), AlignmentCI->getZExtValue(),
4335 } else if (const auto *AA = TargetDecl->getAttr<AllocAlignAttr>()) {
4336 llvm::Value *ParamVal =
4337 CallArgs[AA->getParamIndex() - 1].RV.getScalarVal();
4338 EmitAlignmentAssumption(Ret.getScalarVal(), ParamVal);
4345 /* VarArg handling */
4347 Address CodeGenFunction::EmitVAArg(VAArgExpr *VE, Address &VAListAddr) {
4348 VAListAddr = VE->isMicrosoftABI()
4349 ? EmitMSVAListRef(VE->getSubExpr())
4350 : EmitVAListRef(VE->getSubExpr());
4351 QualType Ty = VE->getType();
4352 if (VE->isMicrosoftABI())
4353 return CGM.getTypes().getABIInfo().EmitMSVAArg(*this, VAListAddr, Ty);
4354 return CGM.getTypes().getABIInfo().EmitVAArg(*this, VAListAddr, Ty);