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 /// Adds the formal paramaters in FPT to the given prefix. If any parameter in
105 /// FPT has pass_object_size attrs, then we'll add parameters for those, too.
106 static void appendParameterTypes(const CodeGenTypes &CGT,
107 SmallVectorImpl<CanQualType> &prefix,
108 SmallVectorImpl<FunctionProtoType::ExtParameterInfo> ¶mInfos,
109 CanQual<FunctionProtoType> FPT,
110 const FunctionDecl *FD) {
111 // Fill out paramInfos.
112 if (FPT->hasExtParameterInfos() || !paramInfos.empty()) {
113 assert(paramInfos.size() <= prefix.size());
114 auto protoParamInfos = FPT->getExtParameterInfos();
115 paramInfos.reserve(prefix.size() + protoParamInfos.size());
116 paramInfos.resize(prefix.size());
117 paramInfos.append(protoParamInfos.begin(), protoParamInfos.end());
120 // Fast path: unknown target.
122 prefix.append(FPT->param_type_begin(), FPT->param_type_end());
126 // In the vast majority cases, we'll have precisely FPT->getNumParams()
127 // parameters; the only thing that can change this is the presence of
128 // pass_object_size. So, we preallocate for the common case.
129 prefix.reserve(prefix.size() + FPT->getNumParams());
131 assert(FD->getNumParams() == FPT->getNumParams());
132 for (unsigned I = 0, E = FPT->getNumParams(); I != E; ++I) {
133 prefix.push_back(FPT->getParamType(I));
134 if (FD->getParamDecl(I)->hasAttr<PassObjectSizeAttr>())
135 prefix.push_back(CGT.getContext().getSizeType());
139 /// Arrange the LLVM function layout for a value of the given function
140 /// type, on top of any implicit parameters already stored.
141 static const CGFunctionInfo &
142 arrangeLLVMFunctionInfo(CodeGenTypes &CGT, bool instanceMethod,
143 SmallVectorImpl<CanQualType> &prefix,
144 CanQual<FunctionProtoType> FTP,
145 const FunctionDecl *FD) {
146 SmallVector<FunctionProtoType::ExtParameterInfo, 16> paramInfos;
147 RequiredArgs Required =
148 RequiredArgs::forPrototypePlus(FTP, prefix.size(), FD);
150 appendParameterTypes(CGT, prefix, paramInfos, FTP, FD);
151 CanQualType resultType = FTP->getReturnType().getUnqualifiedType();
153 return CGT.arrangeLLVMFunctionInfo(resultType, instanceMethod,
154 /*chainCall=*/false, prefix,
155 FTP->getExtInfo(), paramInfos,
159 /// Arrange the argument and result information for a value of the
160 /// given freestanding function type.
161 const CGFunctionInfo &
162 CodeGenTypes::arrangeFreeFunctionType(CanQual<FunctionProtoType> FTP,
163 const FunctionDecl *FD) {
164 SmallVector<CanQualType, 16> argTypes;
165 return ::arrangeLLVMFunctionInfo(*this, /*instanceMethod=*/false, argTypes,
169 static CallingConv getCallingConventionForDecl(const Decl *D, bool IsWindows) {
170 // Set the appropriate calling convention for the Function.
171 if (D->hasAttr<StdCallAttr>())
172 return CC_X86StdCall;
174 if (D->hasAttr<FastCallAttr>())
175 return CC_X86FastCall;
177 if (D->hasAttr<RegCallAttr>())
178 return CC_X86RegCall;
180 if (D->hasAttr<ThisCallAttr>())
181 return CC_X86ThisCall;
183 if (D->hasAttr<VectorCallAttr>())
184 return CC_X86VectorCall;
186 if (D->hasAttr<PascalAttr>())
189 if (PcsAttr *PCS = D->getAttr<PcsAttr>())
190 return (PCS->getPCS() == PcsAttr::AAPCS ? CC_AAPCS : CC_AAPCS_VFP);
192 if (D->hasAttr<IntelOclBiccAttr>())
193 return CC_IntelOclBicc;
195 if (D->hasAttr<MSABIAttr>())
196 return IsWindows ? CC_C : CC_X86_64Win64;
198 if (D->hasAttr<SysVABIAttr>())
199 return IsWindows ? CC_X86_64SysV : CC_C;
201 if (D->hasAttr<PreserveMostAttr>())
202 return CC_PreserveMost;
204 if (D->hasAttr<PreserveAllAttr>())
205 return CC_PreserveAll;
210 /// Arrange the argument and result information for a call to an
211 /// unknown C++ non-static member function of the given abstract type.
212 /// (Zero value of RD means we don't have any meaningful "this" argument type,
213 /// so fall back to a generic pointer type).
214 /// The member function must be an ordinary function, i.e. not a
215 /// constructor or destructor.
216 const CGFunctionInfo &
217 CodeGenTypes::arrangeCXXMethodType(const CXXRecordDecl *RD,
218 const FunctionProtoType *FTP,
219 const CXXMethodDecl *MD) {
220 SmallVector<CanQualType, 16> argTypes;
222 // Add the 'this' pointer.
224 argTypes.push_back(GetThisType(Context, RD));
226 argTypes.push_back(Context.VoidPtrTy);
228 return ::arrangeLLVMFunctionInfo(
229 *this, true, argTypes,
230 FTP->getCanonicalTypeUnqualified().getAs<FunctionProtoType>(), MD);
233 /// Arrange the argument and result information for a declaration or
234 /// definition of the given C++ non-static member function. The
235 /// member function must be an ordinary function, i.e. not a
236 /// constructor or destructor.
237 const CGFunctionInfo &
238 CodeGenTypes::arrangeCXXMethodDeclaration(const CXXMethodDecl *MD) {
239 assert(!isa<CXXConstructorDecl>(MD) && "wrong method for constructors!");
240 assert(!isa<CXXDestructorDecl>(MD) && "wrong method for destructors!");
242 CanQual<FunctionProtoType> prototype = GetFormalType(MD);
244 if (MD->isInstance()) {
245 // The abstract case is perfectly fine.
246 const CXXRecordDecl *ThisType = TheCXXABI.getThisArgumentTypeForMethod(MD);
247 return arrangeCXXMethodType(ThisType, prototype.getTypePtr(), MD);
250 return arrangeFreeFunctionType(prototype, MD);
253 bool CodeGenTypes::inheritingCtorHasParams(
254 const InheritedConstructor &Inherited, CXXCtorType Type) {
255 // Parameters are unnecessary if we're constructing a base class subobject
256 // and the inherited constructor lives in a virtual base.
257 return Type == Ctor_Complete ||
258 !Inherited.getShadowDecl()->constructsVirtualBase() ||
259 !Target.getCXXABI().hasConstructorVariants();
262 const CGFunctionInfo &
263 CodeGenTypes::arrangeCXXStructorDeclaration(const CXXMethodDecl *MD,
266 SmallVector<CanQualType, 16> argTypes;
267 SmallVector<FunctionProtoType::ExtParameterInfo, 16> paramInfos;
268 argTypes.push_back(GetThisType(Context, MD->getParent()));
270 bool PassParams = true;
273 if (auto *CD = dyn_cast<CXXConstructorDecl>(MD)) {
274 GD = GlobalDecl(CD, toCXXCtorType(Type));
276 // A base class inheriting constructor doesn't get forwarded arguments
277 // needed to construct a virtual base (or base class thereof).
278 if (auto Inherited = CD->getInheritedConstructor())
279 PassParams = inheritingCtorHasParams(Inherited, toCXXCtorType(Type));
281 auto *DD = dyn_cast<CXXDestructorDecl>(MD);
282 GD = GlobalDecl(DD, toCXXDtorType(Type));
285 CanQual<FunctionProtoType> FTP = GetFormalType(MD);
287 // Add the formal parameters.
289 appendParameterTypes(*this, argTypes, paramInfos, FTP, MD);
291 TheCXXABI.buildStructorSignature(MD, Type, argTypes);
293 RequiredArgs required =
294 (PassParams && MD->isVariadic() ? RequiredArgs(argTypes.size())
295 : RequiredArgs::All);
297 FunctionType::ExtInfo extInfo = FTP->getExtInfo();
298 CanQualType resultType = TheCXXABI.HasThisReturn(GD)
300 : TheCXXABI.hasMostDerivedReturn(GD)
301 ? CGM.getContext().VoidPtrTy
303 return arrangeLLVMFunctionInfo(resultType, /*instanceMethod=*/true,
304 /*chainCall=*/false, argTypes, extInfo,
305 paramInfos, required);
308 static SmallVector<CanQualType, 16>
309 getArgTypesForCall(ASTContext &ctx, const CallArgList &args) {
310 SmallVector<CanQualType, 16> argTypes;
311 for (auto &arg : args)
312 argTypes.push_back(ctx.getCanonicalParamType(arg.Ty));
316 static SmallVector<CanQualType, 16>
317 getArgTypesForDeclaration(ASTContext &ctx, const FunctionArgList &args) {
318 SmallVector<CanQualType, 16> argTypes;
319 for (auto &arg : args)
320 argTypes.push_back(ctx.getCanonicalParamType(arg->getType()));
324 static void addExtParameterInfosForCall(
325 llvm::SmallVectorImpl<FunctionProtoType::ExtParameterInfo> ¶mInfos,
326 const FunctionProtoType *proto,
328 unsigned totalArgs) {
329 assert(proto->hasExtParameterInfos());
330 assert(paramInfos.size() <= prefixArgs);
331 assert(proto->getNumParams() + prefixArgs <= totalArgs);
333 // Add default infos for any prefix args that don't already have infos.
334 paramInfos.resize(prefixArgs);
336 // Add infos for the prototype.
337 auto protoInfos = proto->getExtParameterInfos();
338 paramInfos.append(protoInfos.begin(), protoInfos.end());
340 // Add default infos for the variadic arguments.
341 paramInfos.resize(totalArgs);
344 static llvm::SmallVector<FunctionProtoType::ExtParameterInfo, 16>
345 getExtParameterInfosForCall(const FunctionProtoType *proto,
346 unsigned prefixArgs, unsigned totalArgs) {
347 llvm::SmallVector<FunctionProtoType::ExtParameterInfo, 16> result;
348 if (proto->hasExtParameterInfos()) {
349 addExtParameterInfosForCall(result, proto, prefixArgs, totalArgs);
354 /// Arrange a call to a C++ method, passing the given arguments.
355 const CGFunctionInfo &
356 CodeGenTypes::arrangeCXXConstructorCall(const CallArgList &args,
357 const CXXConstructorDecl *D,
358 CXXCtorType CtorKind,
359 unsigned ExtraArgs) {
361 SmallVector<CanQualType, 16> ArgTypes;
362 for (const auto &Arg : args)
363 ArgTypes.push_back(Context.getCanonicalParamType(Arg.Ty));
365 CanQual<FunctionProtoType> FPT = GetFormalType(D);
366 RequiredArgs Required = RequiredArgs::forPrototypePlus(FPT, 1 + ExtraArgs, D);
367 GlobalDecl GD(D, CtorKind);
368 CanQualType ResultType = TheCXXABI.HasThisReturn(GD)
370 : TheCXXABI.hasMostDerivedReturn(GD)
371 ? CGM.getContext().VoidPtrTy
374 FunctionType::ExtInfo Info = FPT->getExtInfo();
375 auto ParamInfos = getExtParameterInfosForCall(FPT.getTypePtr(), 1 + ExtraArgs,
377 return arrangeLLVMFunctionInfo(ResultType, /*instanceMethod=*/true,
378 /*chainCall=*/false, ArgTypes, Info,
379 ParamInfos, Required);
382 /// Arrange the argument and result information for the declaration or
383 /// definition of the given function.
384 const CGFunctionInfo &
385 CodeGenTypes::arrangeFunctionDeclaration(const FunctionDecl *FD) {
386 if (const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(FD))
387 if (MD->isInstance())
388 return arrangeCXXMethodDeclaration(MD);
390 CanQualType FTy = FD->getType()->getCanonicalTypeUnqualified();
392 assert(isa<FunctionType>(FTy));
394 // When declaring a function without a prototype, always use a
395 // non-variadic type.
396 if (CanQual<FunctionNoProtoType> noProto = FTy.getAs<FunctionNoProtoType>()) {
397 return arrangeLLVMFunctionInfo(
398 noProto->getReturnType(), /*instanceMethod=*/false,
399 /*chainCall=*/false, None, noProto->getExtInfo(), {},RequiredArgs::All);
402 return arrangeFreeFunctionType(FTy.castAs<FunctionProtoType>(), FD);
405 /// Arrange the argument and result information for the declaration or
406 /// definition of an Objective-C method.
407 const CGFunctionInfo &
408 CodeGenTypes::arrangeObjCMethodDeclaration(const ObjCMethodDecl *MD) {
409 // It happens that this is the same as a call with no optional
410 // arguments, except also using the formal 'self' type.
411 return arrangeObjCMessageSendSignature(MD, MD->getSelfDecl()->getType());
414 /// Arrange the argument and result information for the function type
415 /// through which to perform a send to the given Objective-C method,
416 /// using the given receiver type. The receiver type is not always
417 /// the 'self' type of the method or even an Objective-C pointer type.
418 /// This is *not* the right method for actually performing such a
419 /// message send, due to the possibility of optional arguments.
420 const CGFunctionInfo &
421 CodeGenTypes::arrangeObjCMessageSendSignature(const ObjCMethodDecl *MD,
422 QualType receiverType) {
423 SmallVector<CanQualType, 16> argTys;
424 argTys.push_back(Context.getCanonicalParamType(receiverType));
425 argTys.push_back(Context.getCanonicalParamType(Context.getObjCSelType()));
427 for (const auto *I : MD->parameters()) {
428 argTys.push_back(Context.getCanonicalParamType(I->getType()));
431 FunctionType::ExtInfo einfo;
432 bool IsWindows = getContext().getTargetInfo().getTriple().isOSWindows();
433 einfo = einfo.withCallingConv(getCallingConventionForDecl(MD, IsWindows));
435 if (getContext().getLangOpts().ObjCAutoRefCount &&
436 MD->hasAttr<NSReturnsRetainedAttr>())
437 einfo = einfo.withProducesResult(true);
439 RequiredArgs required =
440 (MD->isVariadic() ? RequiredArgs(argTys.size()) : RequiredArgs::All);
442 return arrangeLLVMFunctionInfo(
443 GetReturnType(MD->getReturnType()), /*instanceMethod=*/false,
444 /*chainCall=*/false, argTys, einfo, {}, required);
447 const CGFunctionInfo &
448 CodeGenTypes::arrangeUnprototypedObjCMessageSend(QualType returnType,
449 const CallArgList &args) {
450 auto argTypes = getArgTypesForCall(Context, args);
451 FunctionType::ExtInfo einfo;
453 return arrangeLLVMFunctionInfo(
454 GetReturnType(returnType), /*instanceMethod=*/false,
455 /*chainCall=*/false, argTypes, einfo, {}, RequiredArgs::All);
458 const CGFunctionInfo &
459 CodeGenTypes::arrangeGlobalDeclaration(GlobalDecl GD) {
460 // FIXME: Do we need to handle ObjCMethodDecl?
461 const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl());
463 if (const CXXConstructorDecl *CD = dyn_cast<CXXConstructorDecl>(FD))
464 return arrangeCXXStructorDeclaration(CD, getFromCtorType(GD.getCtorType()));
466 if (const CXXDestructorDecl *DD = dyn_cast<CXXDestructorDecl>(FD))
467 return arrangeCXXStructorDeclaration(DD, getFromDtorType(GD.getDtorType()));
469 return arrangeFunctionDeclaration(FD);
472 /// Arrange a thunk that takes 'this' as the first parameter followed by
473 /// varargs. Return a void pointer, regardless of the actual return type.
474 /// The body of the thunk will end in a musttail call to a function of the
475 /// correct type, and the caller will bitcast the function to the correct
477 const CGFunctionInfo &
478 CodeGenTypes::arrangeMSMemberPointerThunk(const CXXMethodDecl *MD) {
479 assert(MD->isVirtual() && "only virtual memptrs have thunks");
480 CanQual<FunctionProtoType> FTP = GetFormalType(MD);
481 CanQualType ArgTys[] = { GetThisType(Context, MD->getParent()) };
482 return arrangeLLVMFunctionInfo(Context.VoidTy, /*instanceMethod=*/false,
483 /*chainCall=*/false, ArgTys,
484 FTP->getExtInfo(), {}, RequiredArgs(1));
487 const CGFunctionInfo &
488 CodeGenTypes::arrangeMSCtorClosure(const CXXConstructorDecl *CD,
490 assert(CT == Ctor_CopyingClosure || CT == Ctor_DefaultClosure);
492 CanQual<FunctionProtoType> FTP = GetFormalType(CD);
493 SmallVector<CanQualType, 2> ArgTys;
494 const CXXRecordDecl *RD = CD->getParent();
495 ArgTys.push_back(GetThisType(Context, RD));
496 if (CT == Ctor_CopyingClosure)
497 ArgTys.push_back(*FTP->param_type_begin());
498 if (RD->getNumVBases() > 0)
499 ArgTys.push_back(Context.IntTy);
500 CallingConv CC = Context.getDefaultCallingConvention(
501 /*IsVariadic=*/false, /*IsCXXMethod=*/true);
502 return arrangeLLVMFunctionInfo(Context.VoidTy, /*instanceMethod=*/true,
503 /*chainCall=*/false, ArgTys,
504 FunctionType::ExtInfo(CC), {},
508 /// Arrange a call as unto a free function, except possibly with an
509 /// additional number of formal parameters considered required.
510 static const CGFunctionInfo &
511 arrangeFreeFunctionLikeCall(CodeGenTypes &CGT,
513 const CallArgList &args,
514 const FunctionType *fnType,
515 unsigned numExtraRequiredArgs,
517 assert(args.size() >= numExtraRequiredArgs);
519 llvm::SmallVector<FunctionProtoType::ExtParameterInfo, 16> paramInfos;
521 // In most cases, there are no optional arguments.
522 RequiredArgs required = RequiredArgs::All;
524 // If we have a variadic prototype, the required arguments are the
525 // extra prefix plus the arguments in the prototype.
526 if (const FunctionProtoType *proto = dyn_cast<FunctionProtoType>(fnType)) {
527 if (proto->isVariadic())
528 required = RequiredArgs(proto->getNumParams() + numExtraRequiredArgs);
530 if (proto->hasExtParameterInfos())
531 addExtParameterInfosForCall(paramInfos, proto, numExtraRequiredArgs,
534 // If we don't have a prototype at all, but we're supposed to
535 // explicitly use the variadic convention for unprototyped calls,
536 // treat all of the arguments as required but preserve the nominal
537 // possibility of variadics.
538 } else if (CGM.getTargetCodeGenInfo()
539 .isNoProtoCallVariadic(args,
540 cast<FunctionNoProtoType>(fnType))) {
541 required = RequiredArgs(args.size());
545 SmallVector<CanQualType, 16> argTypes;
546 for (const auto &arg : args)
547 argTypes.push_back(CGT.getContext().getCanonicalParamType(arg.Ty));
548 return CGT.arrangeLLVMFunctionInfo(GetReturnType(fnType->getReturnType()),
549 /*instanceMethod=*/false, chainCall,
550 argTypes, fnType->getExtInfo(), paramInfos,
554 /// Figure out the rules for calling a function with the given formal
555 /// type using the given arguments. The arguments are necessary
556 /// because the function might be unprototyped, in which case it's
557 /// target-dependent in crazy ways.
558 const CGFunctionInfo &
559 CodeGenTypes::arrangeFreeFunctionCall(const CallArgList &args,
560 const FunctionType *fnType,
562 return arrangeFreeFunctionLikeCall(*this, CGM, args, fnType,
563 chainCall ? 1 : 0, chainCall);
566 /// A block function is essentially a free function with an
567 /// extra implicit argument.
568 const CGFunctionInfo &
569 CodeGenTypes::arrangeBlockFunctionCall(const CallArgList &args,
570 const FunctionType *fnType) {
571 return arrangeFreeFunctionLikeCall(*this, CGM, args, fnType, 1,
572 /*chainCall=*/false);
575 const CGFunctionInfo &
576 CodeGenTypes::arrangeBlockFunctionDeclaration(const FunctionProtoType *proto,
577 const FunctionArgList ¶ms) {
578 auto paramInfos = getExtParameterInfosForCall(proto, 1, params.size());
579 auto argTypes = getArgTypesForDeclaration(Context, params);
581 return arrangeLLVMFunctionInfo(
582 GetReturnType(proto->getReturnType()),
583 /*instanceMethod*/ false, /*chainCall*/ false, argTypes,
584 proto->getExtInfo(), paramInfos,
585 RequiredArgs::forPrototypePlus(proto, 1, nullptr));
588 const CGFunctionInfo &
589 CodeGenTypes::arrangeBuiltinFunctionCall(QualType resultType,
590 const CallArgList &args) {
592 SmallVector<CanQualType, 16> argTypes;
593 for (const auto &Arg : args)
594 argTypes.push_back(Context.getCanonicalParamType(Arg.Ty));
595 return arrangeLLVMFunctionInfo(
596 GetReturnType(resultType), /*instanceMethod=*/false,
597 /*chainCall=*/false, argTypes, FunctionType::ExtInfo(),
598 /*paramInfos=*/ {}, RequiredArgs::All);
601 const CGFunctionInfo &
602 CodeGenTypes::arrangeBuiltinFunctionDeclaration(QualType resultType,
603 const FunctionArgList &args) {
604 auto argTypes = getArgTypesForDeclaration(Context, args);
606 return arrangeLLVMFunctionInfo(
607 GetReturnType(resultType), /*instanceMethod=*/false, /*chainCall=*/false,
608 argTypes, FunctionType::ExtInfo(), {}, RequiredArgs::All);
611 const CGFunctionInfo &
612 CodeGenTypes::arrangeBuiltinFunctionDeclaration(CanQualType resultType,
613 ArrayRef<CanQualType> argTypes) {
614 return arrangeLLVMFunctionInfo(
615 resultType, /*instanceMethod=*/false, /*chainCall=*/false,
616 argTypes, FunctionType::ExtInfo(), {}, RequiredArgs::All);
619 /// Arrange a call to a C++ method, passing the given arguments.
620 const CGFunctionInfo &
621 CodeGenTypes::arrangeCXXMethodCall(const CallArgList &args,
622 const FunctionProtoType *proto,
623 RequiredArgs required) {
624 unsigned numRequiredArgs =
625 (proto->isVariadic() ? required.getNumRequiredArgs() : args.size());
626 unsigned numPrefixArgs = numRequiredArgs - proto->getNumParams();
628 getExtParameterInfosForCall(proto, numPrefixArgs, args.size());
631 auto argTypes = getArgTypesForCall(Context, args);
633 FunctionType::ExtInfo info = proto->getExtInfo();
634 return arrangeLLVMFunctionInfo(
635 GetReturnType(proto->getReturnType()), /*instanceMethod=*/true,
636 /*chainCall=*/false, argTypes, info, paramInfos, required);
639 const CGFunctionInfo &CodeGenTypes::arrangeNullaryFunction() {
640 return arrangeLLVMFunctionInfo(
641 getContext().VoidTy, /*instanceMethod=*/false, /*chainCall=*/false,
642 None, FunctionType::ExtInfo(), {}, RequiredArgs::All);
645 const CGFunctionInfo &
646 CodeGenTypes::arrangeCall(const CGFunctionInfo &signature,
647 const CallArgList &args) {
648 assert(signature.arg_size() <= args.size());
649 if (signature.arg_size() == args.size())
652 SmallVector<FunctionProtoType::ExtParameterInfo, 16> paramInfos;
653 auto sigParamInfos = signature.getExtParameterInfos();
654 if (!sigParamInfos.empty()) {
655 paramInfos.append(sigParamInfos.begin(), sigParamInfos.end());
656 paramInfos.resize(args.size());
659 auto argTypes = getArgTypesForCall(Context, args);
661 assert(signature.getRequiredArgs().allowsOptionalArgs());
662 return arrangeLLVMFunctionInfo(signature.getReturnType(),
663 signature.isInstanceMethod(),
664 signature.isChainCall(),
666 signature.getExtInfo(),
668 signature.getRequiredArgs());
671 /// Arrange the argument and result information for an abstract value
672 /// of a given function type. This is the method which all of the
673 /// above functions ultimately defer to.
674 const CGFunctionInfo &
675 CodeGenTypes::arrangeLLVMFunctionInfo(CanQualType resultType,
678 ArrayRef<CanQualType> argTypes,
679 FunctionType::ExtInfo info,
680 ArrayRef<FunctionProtoType::ExtParameterInfo> paramInfos,
681 RequiredArgs required) {
682 assert(std::all_of(argTypes.begin(), argTypes.end(),
683 std::mem_fun_ref(&CanQualType::isCanonicalAsParam)));
685 // Lookup or create unique function info.
686 llvm::FoldingSetNodeID ID;
687 CGFunctionInfo::Profile(ID, instanceMethod, chainCall, info, paramInfos,
688 required, resultType, argTypes);
690 void *insertPos = nullptr;
691 CGFunctionInfo *FI = FunctionInfos.FindNodeOrInsertPos(ID, insertPos);
695 unsigned CC = ClangCallConvToLLVMCallConv(info.getCC());
697 // Construct the function info. We co-allocate the ArgInfos.
698 FI = CGFunctionInfo::create(CC, instanceMethod, chainCall, info,
699 paramInfos, resultType, argTypes, required);
700 FunctionInfos.InsertNode(FI, insertPos);
702 bool inserted = FunctionsBeingProcessed.insert(FI).second;
704 assert(inserted && "Recursively being processed?");
706 // Compute ABI information.
707 if (info.getCC() != CC_Swift) {
708 getABIInfo().computeInfo(*FI);
710 swiftcall::computeABIInfo(CGM, *FI);
713 // Loop over all of the computed argument and return value info. If any of
714 // them are direct or extend without a specified coerce type, specify the
716 ABIArgInfo &retInfo = FI->getReturnInfo();
717 if (retInfo.canHaveCoerceToType() && retInfo.getCoerceToType() == nullptr)
718 retInfo.setCoerceToType(ConvertType(FI->getReturnType()));
720 for (auto &I : FI->arguments())
721 if (I.info.canHaveCoerceToType() && I.info.getCoerceToType() == nullptr)
722 I.info.setCoerceToType(ConvertType(I.type));
724 bool erased = FunctionsBeingProcessed.erase(FI); (void)erased;
725 assert(erased && "Not in set?");
730 CGFunctionInfo *CGFunctionInfo::create(unsigned llvmCC,
733 const FunctionType::ExtInfo &info,
734 ArrayRef<ExtParameterInfo> paramInfos,
735 CanQualType resultType,
736 ArrayRef<CanQualType> argTypes,
737 RequiredArgs required) {
738 assert(paramInfos.empty() || paramInfos.size() == argTypes.size());
741 operator new(totalSizeToAlloc<ArgInfo, ExtParameterInfo>(
742 argTypes.size() + 1, paramInfos.size()));
744 CGFunctionInfo *FI = new(buffer) CGFunctionInfo();
745 FI->CallingConvention = llvmCC;
746 FI->EffectiveCallingConvention = llvmCC;
747 FI->ASTCallingConvention = info.getCC();
748 FI->InstanceMethod = instanceMethod;
749 FI->ChainCall = chainCall;
750 FI->NoReturn = info.getNoReturn();
751 FI->ReturnsRetained = info.getProducesResult();
752 FI->Required = required;
753 FI->HasRegParm = info.getHasRegParm();
754 FI->RegParm = info.getRegParm();
755 FI->ArgStruct = nullptr;
756 FI->ArgStructAlign = 0;
757 FI->NumArgs = argTypes.size();
758 FI->HasExtParameterInfos = !paramInfos.empty();
759 FI->getArgsBuffer()[0].type = resultType;
760 for (unsigned i = 0, e = argTypes.size(); i != e; ++i)
761 FI->getArgsBuffer()[i + 1].type = argTypes[i];
762 for (unsigned i = 0, e = paramInfos.size(); i != e; ++i)
763 FI->getExtParameterInfosBuffer()[i] = paramInfos[i];
770 // ABIArgInfo::Expand implementation.
772 // Specifies the way QualType passed as ABIArgInfo::Expand is expanded.
773 struct TypeExpansion {
774 enum TypeExpansionKind {
775 // Elements of constant arrays are expanded recursively.
777 // Record fields are expanded recursively (but if record is a union, only
778 // the field with the largest size is expanded).
780 // For complex types, real and imaginary parts are expanded recursively.
782 // All other types are not expandable.
786 const TypeExpansionKind Kind;
788 TypeExpansion(TypeExpansionKind K) : Kind(K) {}
789 virtual ~TypeExpansion() {}
792 struct ConstantArrayExpansion : TypeExpansion {
796 ConstantArrayExpansion(QualType EltTy, uint64_t NumElts)
797 : TypeExpansion(TEK_ConstantArray), EltTy(EltTy), NumElts(NumElts) {}
798 static bool classof(const TypeExpansion *TE) {
799 return TE->Kind == TEK_ConstantArray;
803 struct RecordExpansion : TypeExpansion {
804 SmallVector<const CXXBaseSpecifier *, 1> Bases;
806 SmallVector<const FieldDecl *, 1> Fields;
808 RecordExpansion(SmallVector<const CXXBaseSpecifier *, 1> &&Bases,
809 SmallVector<const FieldDecl *, 1> &&Fields)
810 : TypeExpansion(TEK_Record), Bases(std::move(Bases)),
811 Fields(std::move(Fields)) {}
812 static bool classof(const TypeExpansion *TE) {
813 return TE->Kind == TEK_Record;
817 struct ComplexExpansion : TypeExpansion {
820 ComplexExpansion(QualType EltTy) : TypeExpansion(TEK_Complex), EltTy(EltTy) {}
821 static bool classof(const TypeExpansion *TE) {
822 return TE->Kind == TEK_Complex;
826 struct NoExpansion : TypeExpansion {
827 NoExpansion() : TypeExpansion(TEK_None) {}
828 static bool classof(const TypeExpansion *TE) {
829 return TE->Kind == TEK_None;
834 static std::unique_ptr<TypeExpansion>
835 getTypeExpansion(QualType Ty, const ASTContext &Context) {
836 if (const ConstantArrayType *AT = Context.getAsConstantArrayType(Ty)) {
837 return llvm::make_unique<ConstantArrayExpansion>(
838 AT->getElementType(), AT->getSize().getZExtValue());
840 if (const RecordType *RT = Ty->getAs<RecordType>()) {
841 SmallVector<const CXXBaseSpecifier *, 1> Bases;
842 SmallVector<const FieldDecl *, 1> Fields;
843 const RecordDecl *RD = RT->getDecl();
844 assert(!RD->hasFlexibleArrayMember() &&
845 "Cannot expand structure with flexible array.");
847 // Unions can be here only in degenerative cases - all the fields are same
848 // after flattening. Thus we have to use the "largest" field.
849 const FieldDecl *LargestFD = nullptr;
850 CharUnits UnionSize = CharUnits::Zero();
852 for (const auto *FD : RD->fields()) {
853 // Skip zero length bitfields.
854 if (FD->isBitField() && FD->getBitWidthValue(Context) == 0)
856 assert(!FD->isBitField() &&
857 "Cannot expand structure with bit-field members.");
858 CharUnits FieldSize = Context.getTypeSizeInChars(FD->getType());
859 if (UnionSize < FieldSize) {
860 UnionSize = FieldSize;
865 Fields.push_back(LargestFD);
867 if (const auto *CXXRD = dyn_cast<CXXRecordDecl>(RD)) {
868 assert(!CXXRD->isDynamicClass() &&
869 "cannot expand vtable pointers in dynamic classes");
870 for (const CXXBaseSpecifier &BS : CXXRD->bases())
871 Bases.push_back(&BS);
874 for (const auto *FD : RD->fields()) {
875 // Skip zero length bitfields.
876 if (FD->isBitField() && FD->getBitWidthValue(Context) == 0)
878 assert(!FD->isBitField() &&
879 "Cannot expand structure with bit-field members.");
880 Fields.push_back(FD);
883 return llvm::make_unique<RecordExpansion>(std::move(Bases),
886 if (const ComplexType *CT = Ty->getAs<ComplexType>()) {
887 return llvm::make_unique<ComplexExpansion>(CT->getElementType());
889 return llvm::make_unique<NoExpansion>();
892 static int getExpansionSize(QualType Ty, const ASTContext &Context) {
893 auto Exp = getTypeExpansion(Ty, Context);
894 if (auto CAExp = dyn_cast<ConstantArrayExpansion>(Exp.get())) {
895 return CAExp->NumElts * getExpansionSize(CAExp->EltTy, Context);
897 if (auto RExp = dyn_cast<RecordExpansion>(Exp.get())) {
899 for (auto BS : RExp->Bases)
900 Res += getExpansionSize(BS->getType(), Context);
901 for (auto FD : RExp->Fields)
902 Res += getExpansionSize(FD->getType(), Context);
905 if (isa<ComplexExpansion>(Exp.get()))
907 assert(isa<NoExpansion>(Exp.get()));
912 CodeGenTypes::getExpandedTypes(QualType Ty,
913 SmallVectorImpl<llvm::Type *>::iterator &TI) {
914 auto Exp = getTypeExpansion(Ty, Context);
915 if (auto CAExp = dyn_cast<ConstantArrayExpansion>(Exp.get())) {
916 for (int i = 0, n = CAExp->NumElts; i < n; i++) {
917 getExpandedTypes(CAExp->EltTy, TI);
919 } else if (auto RExp = dyn_cast<RecordExpansion>(Exp.get())) {
920 for (auto BS : RExp->Bases)
921 getExpandedTypes(BS->getType(), TI);
922 for (auto FD : RExp->Fields)
923 getExpandedTypes(FD->getType(), TI);
924 } else if (auto CExp = dyn_cast<ComplexExpansion>(Exp.get())) {
925 llvm::Type *EltTy = ConvertType(CExp->EltTy);
929 assert(isa<NoExpansion>(Exp.get()));
930 *TI++ = ConvertType(Ty);
934 static void forConstantArrayExpansion(CodeGenFunction &CGF,
935 ConstantArrayExpansion *CAE,
937 llvm::function_ref<void(Address)> Fn) {
938 CharUnits EltSize = CGF.getContext().getTypeSizeInChars(CAE->EltTy);
940 BaseAddr.getAlignment().alignmentOfArrayElement(EltSize);
942 for (int i = 0, n = CAE->NumElts; i < n; i++) {
943 llvm::Value *EltAddr =
944 CGF.Builder.CreateConstGEP2_32(nullptr, BaseAddr.getPointer(), 0, i);
945 Fn(Address(EltAddr, EltAlign));
949 void CodeGenFunction::ExpandTypeFromArgs(
950 QualType Ty, LValue LV, SmallVectorImpl<llvm::Value *>::iterator &AI) {
951 assert(LV.isSimple() &&
952 "Unexpected non-simple lvalue during struct expansion.");
954 auto Exp = getTypeExpansion(Ty, getContext());
955 if (auto CAExp = dyn_cast<ConstantArrayExpansion>(Exp.get())) {
956 forConstantArrayExpansion(*this, CAExp, LV.getAddress(),
957 [&](Address EltAddr) {
958 LValue LV = MakeAddrLValue(EltAddr, CAExp->EltTy);
959 ExpandTypeFromArgs(CAExp->EltTy, LV, AI);
961 } else if (auto RExp = dyn_cast<RecordExpansion>(Exp.get())) {
962 Address This = LV.getAddress();
963 for (const CXXBaseSpecifier *BS : RExp->Bases) {
964 // Perform a single step derived-to-base conversion.
966 GetAddressOfBaseClass(This, Ty->getAsCXXRecordDecl(), &BS, &BS + 1,
967 /*NullCheckValue=*/false, SourceLocation());
968 LValue SubLV = MakeAddrLValue(Base, BS->getType());
970 // Recurse onto bases.
971 ExpandTypeFromArgs(BS->getType(), SubLV, AI);
973 for (auto FD : RExp->Fields) {
974 // FIXME: What are the right qualifiers here?
975 LValue SubLV = EmitLValueForFieldInitialization(LV, FD);
976 ExpandTypeFromArgs(FD->getType(), SubLV, AI);
978 } else if (isa<ComplexExpansion>(Exp.get())) {
979 auto realValue = *AI++;
980 auto imagValue = *AI++;
981 EmitStoreOfComplex(ComplexPairTy(realValue, imagValue), LV, /*init*/ true);
983 assert(isa<NoExpansion>(Exp.get()));
984 EmitStoreThroughLValue(RValue::get(*AI++), LV);
988 void CodeGenFunction::ExpandTypeToArgs(
989 QualType Ty, RValue RV, llvm::FunctionType *IRFuncTy,
990 SmallVectorImpl<llvm::Value *> &IRCallArgs, unsigned &IRCallArgPos) {
991 auto Exp = getTypeExpansion(Ty, getContext());
992 if (auto CAExp = dyn_cast<ConstantArrayExpansion>(Exp.get())) {
993 forConstantArrayExpansion(*this, CAExp, RV.getAggregateAddress(),
994 [&](Address EltAddr) {
996 convertTempToRValue(EltAddr, CAExp->EltTy, SourceLocation());
997 ExpandTypeToArgs(CAExp->EltTy, EltRV, IRFuncTy, IRCallArgs, IRCallArgPos);
999 } else if (auto RExp = dyn_cast<RecordExpansion>(Exp.get())) {
1000 Address This = RV.getAggregateAddress();
1001 for (const CXXBaseSpecifier *BS : RExp->Bases) {
1002 // Perform a single step derived-to-base conversion.
1004 GetAddressOfBaseClass(This, Ty->getAsCXXRecordDecl(), &BS, &BS + 1,
1005 /*NullCheckValue=*/false, SourceLocation());
1006 RValue BaseRV = RValue::getAggregate(Base);
1008 // Recurse onto bases.
1009 ExpandTypeToArgs(BS->getType(), BaseRV, IRFuncTy, IRCallArgs,
1013 LValue LV = MakeAddrLValue(This, Ty);
1014 for (auto FD : RExp->Fields) {
1015 RValue FldRV = EmitRValueForField(LV, FD, SourceLocation());
1016 ExpandTypeToArgs(FD->getType(), FldRV, IRFuncTy, IRCallArgs,
1019 } else if (isa<ComplexExpansion>(Exp.get())) {
1020 ComplexPairTy CV = RV.getComplexVal();
1021 IRCallArgs[IRCallArgPos++] = CV.first;
1022 IRCallArgs[IRCallArgPos++] = CV.second;
1024 assert(isa<NoExpansion>(Exp.get()));
1025 assert(RV.isScalar() &&
1026 "Unexpected non-scalar rvalue during struct expansion.");
1028 // Insert a bitcast as needed.
1029 llvm::Value *V = RV.getScalarVal();
1030 if (IRCallArgPos < IRFuncTy->getNumParams() &&
1031 V->getType() != IRFuncTy->getParamType(IRCallArgPos))
1032 V = Builder.CreateBitCast(V, IRFuncTy->getParamType(IRCallArgPos));
1034 IRCallArgs[IRCallArgPos++] = V;
1038 /// Create a temporary allocation for the purposes of coercion.
1039 static Address CreateTempAllocaForCoercion(CodeGenFunction &CGF, llvm::Type *Ty,
1040 CharUnits MinAlign) {
1041 // Don't use an alignment that's worse than what LLVM would prefer.
1042 auto PrefAlign = CGF.CGM.getDataLayout().getPrefTypeAlignment(Ty);
1043 CharUnits Align = std::max(MinAlign, CharUnits::fromQuantity(PrefAlign));
1045 return CGF.CreateTempAlloca(Ty, Align);
1048 /// EnterStructPointerForCoercedAccess - Given a struct pointer that we are
1049 /// accessing some number of bytes out of it, try to gep into the struct to get
1050 /// at its inner goodness. Dive as deep as possible without entering an element
1051 /// with an in-memory size smaller than DstSize.
1053 EnterStructPointerForCoercedAccess(Address SrcPtr,
1054 llvm::StructType *SrcSTy,
1055 uint64_t DstSize, CodeGenFunction &CGF) {
1056 // We can't dive into a zero-element struct.
1057 if (SrcSTy->getNumElements() == 0) return SrcPtr;
1059 llvm::Type *FirstElt = SrcSTy->getElementType(0);
1061 // If the first elt is at least as large as what we're looking for, or if the
1062 // first element is the same size as the whole struct, we can enter it. The
1063 // comparison must be made on the store size and not the alloca size. Using
1064 // the alloca size may overstate the size of the load.
1065 uint64_t FirstEltSize =
1066 CGF.CGM.getDataLayout().getTypeStoreSize(FirstElt);
1067 if (FirstEltSize < DstSize &&
1068 FirstEltSize < CGF.CGM.getDataLayout().getTypeStoreSize(SrcSTy))
1071 // GEP into the first element.
1072 SrcPtr = CGF.Builder.CreateStructGEP(SrcPtr, 0, CharUnits(), "coerce.dive");
1074 // If the first element is a struct, recurse.
1075 llvm::Type *SrcTy = SrcPtr.getElementType();
1076 if (llvm::StructType *SrcSTy = dyn_cast<llvm::StructType>(SrcTy))
1077 return EnterStructPointerForCoercedAccess(SrcPtr, SrcSTy, DstSize, CGF);
1082 /// CoerceIntOrPtrToIntOrPtr - Convert a value Val to the specific Ty where both
1083 /// are either integers or pointers. This does a truncation of the value if it
1084 /// is too large or a zero extension if it is too small.
1086 /// This behaves as if the value were coerced through memory, so on big-endian
1087 /// targets the high bits are preserved in a truncation, while little-endian
1088 /// targets preserve the low bits.
1089 static llvm::Value *CoerceIntOrPtrToIntOrPtr(llvm::Value *Val,
1091 CodeGenFunction &CGF) {
1092 if (Val->getType() == Ty)
1095 if (isa<llvm::PointerType>(Val->getType())) {
1096 // If this is Pointer->Pointer avoid conversion to and from int.
1097 if (isa<llvm::PointerType>(Ty))
1098 return CGF.Builder.CreateBitCast(Val, Ty, "coerce.val");
1100 // Convert the pointer to an integer so we can play with its width.
1101 Val = CGF.Builder.CreatePtrToInt(Val, CGF.IntPtrTy, "coerce.val.pi");
1104 llvm::Type *DestIntTy = Ty;
1105 if (isa<llvm::PointerType>(DestIntTy))
1106 DestIntTy = CGF.IntPtrTy;
1108 if (Val->getType() != DestIntTy) {
1109 const llvm::DataLayout &DL = CGF.CGM.getDataLayout();
1110 if (DL.isBigEndian()) {
1111 // Preserve the high bits on big-endian targets.
1112 // That is what memory coercion does.
1113 uint64_t SrcSize = DL.getTypeSizeInBits(Val->getType());
1114 uint64_t DstSize = DL.getTypeSizeInBits(DestIntTy);
1116 if (SrcSize > DstSize) {
1117 Val = CGF.Builder.CreateLShr(Val, SrcSize - DstSize, "coerce.highbits");
1118 Val = CGF.Builder.CreateTrunc(Val, DestIntTy, "coerce.val.ii");
1120 Val = CGF.Builder.CreateZExt(Val, DestIntTy, "coerce.val.ii");
1121 Val = CGF.Builder.CreateShl(Val, DstSize - SrcSize, "coerce.highbits");
1124 // Little-endian targets preserve the low bits. No shifts required.
1125 Val = CGF.Builder.CreateIntCast(Val, DestIntTy, false, "coerce.val.ii");
1129 if (isa<llvm::PointerType>(Ty))
1130 Val = CGF.Builder.CreateIntToPtr(Val, Ty, "coerce.val.ip");
1136 /// CreateCoercedLoad - Create a load from \arg SrcPtr interpreted as
1137 /// a pointer to an object of type \arg Ty, known to be aligned to
1138 /// \arg SrcAlign bytes.
1140 /// This safely handles the case when the src type is smaller than the
1141 /// destination type; in this situation the values of bits which not
1142 /// present in the src are undefined.
1143 static llvm::Value *CreateCoercedLoad(Address Src, llvm::Type *Ty,
1144 CodeGenFunction &CGF) {
1145 llvm::Type *SrcTy = Src.getElementType();
1147 // If SrcTy and Ty are the same, just do a load.
1149 return CGF.Builder.CreateLoad(Src);
1151 uint64_t DstSize = CGF.CGM.getDataLayout().getTypeAllocSize(Ty);
1153 if (llvm::StructType *SrcSTy = dyn_cast<llvm::StructType>(SrcTy)) {
1154 Src = EnterStructPointerForCoercedAccess(Src, SrcSTy, DstSize, CGF);
1155 SrcTy = Src.getType()->getElementType();
1158 uint64_t SrcSize = CGF.CGM.getDataLayout().getTypeAllocSize(SrcTy);
1160 // If the source and destination are integer or pointer types, just do an
1161 // extension or truncation to the desired type.
1162 if ((isa<llvm::IntegerType>(Ty) || isa<llvm::PointerType>(Ty)) &&
1163 (isa<llvm::IntegerType>(SrcTy) || isa<llvm::PointerType>(SrcTy))) {
1164 llvm::Value *Load = CGF.Builder.CreateLoad(Src);
1165 return CoerceIntOrPtrToIntOrPtr(Load, Ty, CGF);
1168 // If load is legal, just bitcast the src pointer.
1169 if (SrcSize >= DstSize) {
1170 // Generally SrcSize is never greater than DstSize, since this means we are
1171 // losing bits. However, this can happen in cases where the structure has
1172 // additional padding, for example due to a user specified alignment.
1174 // FIXME: Assert that we aren't truncating non-padding bits when have access
1175 // to that information.
1176 Src = CGF.Builder.CreateBitCast(Src, llvm::PointerType::getUnqual(Ty));
1177 return CGF.Builder.CreateLoad(Src);
1180 // Otherwise do coercion through memory. This is stupid, but simple.
1181 Address Tmp = CreateTempAllocaForCoercion(CGF, Ty, Src.getAlignment());
1182 Address Casted = CGF.Builder.CreateBitCast(Tmp, CGF.Int8PtrTy);
1183 Address SrcCasted = CGF.Builder.CreateBitCast(Src, CGF.Int8PtrTy);
1184 CGF.Builder.CreateMemCpy(Casted, SrcCasted,
1185 llvm::ConstantInt::get(CGF.IntPtrTy, SrcSize),
1187 return CGF.Builder.CreateLoad(Tmp);
1190 // Function to store a first-class aggregate into memory. We prefer to
1191 // store the elements rather than the aggregate to be more friendly to
1193 // FIXME: Do we need to recurse here?
1194 static void BuildAggStore(CodeGenFunction &CGF, llvm::Value *Val,
1195 Address Dest, bool DestIsVolatile) {
1196 // Prefer scalar stores to first-class aggregate stores.
1197 if (llvm::StructType *STy =
1198 dyn_cast<llvm::StructType>(Val->getType())) {
1199 const llvm::StructLayout *Layout =
1200 CGF.CGM.getDataLayout().getStructLayout(STy);
1202 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1203 auto EltOffset = CharUnits::fromQuantity(Layout->getElementOffset(i));
1204 Address EltPtr = CGF.Builder.CreateStructGEP(Dest, i, EltOffset);
1205 llvm::Value *Elt = CGF.Builder.CreateExtractValue(Val, i);
1206 CGF.Builder.CreateStore(Elt, EltPtr, DestIsVolatile);
1209 CGF.Builder.CreateStore(Val, Dest, DestIsVolatile);
1213 /// CreateCoercedStore - Create a store to \arg DstPtr from \arg Src,
1214 /// where the source and destination may have different types. The
1215 /// destination is known to be aligned to \arg DstAlign bytes.
1217 /// This safely handles the case when the src type is larger than the
1218 /// destination type; the upper bits of the src will be lost.
1219 static void CreateCoercedStore(llvm::Value *Src,
1222 CodeGenFunction &CGF) {
1223 llvm::Type *SrcTy = Src->getType();
1224 llvm::Type *DstTy = Dst.getType()->getElementType();
1225 if (SrcTy == DstTy) {
1226 CGF.Builder.CreateStore(Src, Dst, DstIsVolatile);
1230 uint64_t SrcSize = CGF.CGM.getDataLayout().getTypeAllocSize(SrcTy);
1232 if (llvm::StructType *DstSTy = dyn_cast<llvm::StructType>(DstTy)) {
1233 Dst = EnterStructPointerForCoercedAccess(Dst, DstSTy, SrcSize, CGF);
1234 DstTy = Dst.getType()->getElementType();
1237 // If the source and destination are integer or pointer types, just do an
1238 // extension or truncation to the desired type.
1239 if ((isa<llvm::IntegerType>(SrcTy) || isa<llvm::PointerType>(SrcTy)) &&
1240 (isa<llvm::IntegerType>(DstTy) || isa<llvm::PointerType>(DstTy))) {
1241 Src = CoerceIntOrPtrToIntOrPtr(Src, DstTy, CGF);
1242 CGF.Builder.CreateStore(Src, Dst, DstIsVolatile);
1246 uint64_t DstSize = CGF.CGM.getDataLayout().getTypeAllocSize(DstTy);
1248 // If store is legal, just bitcast the src pointer.
1249 if (SrcSize <= DstSize) {
1250 Dst = CGF.Builder.CreateBitCast(Dst, llvm::PointerType::getUnqual(SrcTy));
1251 BuildAggStore(CGF, Src, Dst, DstIsVolatile);
1253 // Otherwise do coercion through memory. This is stupid, but
1256 // Generally SrcSize is never greater than DstSize, since this means we are
1257 // losing bits. However, this can happen in cases where the structure has
1258 // additional padding, for example due to a user specified alignment.
1260 // FIXME: Assert that we aren't truncating non-padding bits when have access
1261 // to that information.
1262 Address Tmp = CreateTempAllocaForCoercion(CGF, SrcTy, Dst.getAlignment());
1263 CGF.Builder.CreateStore(Src, Tmp);
1264 Address Casted = CGF.Builder.CreateBitCast(Tmp, CGF.Int8PtrTy);
1265 Address DstCasted = CGF.Builder.CreateBitCast(Dst, CGF.Int8PtrTy);
1266 CGF.Builder.CreateMemCpy(DstCasted, Casted,
1267 llvm::ConstantInt::get(CGF.IntPtrTy, DstSize),
1272 static Address emitAddressAtOffset(CodeGenFunction &CGF, Address addr,
1273 const ABIArgInfo &info) {
1274 if (unsigned offset = info.getDirectOffset()) {
1275 addr = CGF.Builder.CreateElementBitCast(addr, CGF.Int8Ty);
1276 addr = CGF.Builder.CreateConstInBoundsByteGEP(addr,
1277 CharUnits::fromQuantity(offset));
1278 addr = CGF.Builder.CreateElementBitCast(addr, info.getCoerceToType());
1285 /// Encapsulates information about the way function arguments from
1286 /// CGFunctionInfo should be passed to actual LLVM IR function.
1287 class ClangToLLVMArgMapping {
1288 static const unsigned InvalidIndex = ~0U;
1289 unsigned InallocaArgNo;
1291 unsigned TotalIRArgs;
1293 /// Arguments of LLVM IR function corresponding to single Clang argument.
1295 unsigned PaddingArgIndex;
1296 // Argument is expanded to IR arguments at positions
1297 // [FirstArgIndex, FirstArgIndex + NumberOfArgs).
1298 unsigned FirstArgIndex;
1299 unsigned NumberOfArgs;
1302 : PaddingArgIndex(InvalidIndex), FirstArgIndex(InvalidIndex),
1306 SmallVector<IRArgs, 8> ArgInfo;
1309 ClangToLLVMArgMapping(const ASTContext &Context, const CGFunctionInfo &FI,
1310 bool OnlyRequiredArgs = false)
1311 : InallocaArgNo(InvalidIndex), SRetArgNo(InvalidIndex), TotalIRArgs(0),
1312 ArgInfo(OnlyRequiredArgs ? FI.getNumRequiredArgs() : FI.arg_size()) {
1313 construct(Context, FI, OnlyRequiredArgs);
1316 bool hasInallocaArg() const { return InallocaArgNo != InvalidIndex; }
1317 unsigned getInallocaArgNo() const {
1318 assert(hasInallocaArg());
1319 return InallocaArgNo;
1322 bool hasSRetArg() const { return SRetArgNo != InvalidIndex; }
1323 unsigned getSRetArgNo() const {
1324 assert(hasSRetArg());
1328 unsigned totalIRArgs() const { return TotalIRArgs; }
1330 bool hasPaddingArg(unsigned ArgNo) const {
1331 assert(ArgNo < ArgInfo.size());
1332 return ArgInfo[ArgNo].PaddingArgIndex != InvalidIndex;
1334 unsigned getPaddingArgNo(unsigned ArgNo) const {
1335 assert(hasPaddingArg(ArgNo));
1336 return ArgInfo[ArgNo].PaddingArgIndex;
1339 /// Returns index of first IR argument corresponding to ArgNo, and their
1341 std::pair<unsigned, unsigned> getIRArgs(unsigned ArgNo) const {
1342 assert(ArgNo < ArgInfo.size());
1343 return std::make_pair(ArgInfo[ArgNo].FirstArgIndex,
1344 ArgInfo[ArgNo].NumberOfArgs);
1348 void construct(const ASTContext &Context, const CGFunctionInfo &FI,
1349 bool OnlyRequiredArgs);
1352 void ClangToLLVMArgMapping::construct(const ASTContext &Context,
1353 const CGFunctionInfo &FI,
1354 bool OnlyRequiredArgs) {
1355 unsigned IRArgNo = 0;
1356 bool SwapThisWithSRet = false;
1357 const ABIArgInfo &RetAI = FI.getReturnInfo();
1359 if (RetAI.getKind() == ABIArgInfo::Indirect) {
1360 SwapThisWithSRet = RetAI.isSRetAfterThis();
1361 SRetArgNo = SwapThisWithSRet ? 1 : IRArgNo++;
1365 unsigned NumArgs = OnlyRequiredArgs ? FI.getNumRequiredArgs() : FI.arg_size();
1366 for (CGFunctionInfo::const_arg_iterator I = FI.arg_begin(); ArgNo < NumArgs;
1368 assert(I != FI.arg_end());
1369 QualType ArgType = I->type;
1370 const ABIArgInfo &AI = I->info;
1371 // Collect data about IR arguments corresponding to Clang argument ArgNo.
1372 auto &IRArgs = ArgInfo[ArgNo];
1374 if (AI.getPaddingType())
1375 IRArgs.PaddingArgIndex = IRArgNo++;
1377 switch (AI.getKind()) {
1378 case ABIArgInfo::Extend:
1379 case ABIArgInfo::Direct: {
1380 // FIXME: handle sseregparm someday...
1381 llvm::StructType *STy = dyn_cast<llvm::StructType>(AI.getCoerceToType());
1382 if (AI.isDirect() && AI.getCanBeFlattened() && STy) {
1383 IRArgs.NumberOfArgs = STy->getNumElements();
1385 IRArgs.NumberOfArgs = 1;
1389 case ABIArgInfo::Indirect:
1390 IRArgs.NumberOfArgs = 1;
1392 case ABIArgInfo::Ignore:
1393 case ABIArgInfo::InAlloca:
1394 // ignore and inalloca doesn't have matching LLVM parameters.
1395 IRArgs.NumberOfArgs = 0;
1397 case ABIArgInfo::CoerceAndExpand:
1398 IRArgs.NumberOfArgs = AI.getCoerceAndExpandTypeSequence().size();
1400 case ABIArgInfo::Expand:
1401 IRArgs.NumberOfArgs = getExpansionSize(ArgType, Context);
1405 if (IRArgs.NumberOfArgs > 0) {
1406 IRArgs.FirstArgIndex = IRArgNo;
1407 IRArgNo += IRArgs.NumberOfArgs;
1410 // Skip over the sret parameter when it comes second. We already handled it
1412 if (IRArgNo == 1 && SwapThisWithSRet)
1415 assert(ArgNo == ArgInfo.size());
1417 if (FI.usesInAlloca())
1418 InallocaArgNo = IRArgNo++;
1420 TotalIRArgs = IRArgNo;
1426 bool CodeGenModule::ReturnTypeUsesSRet(const CGFunctionInfo &FI) {
1427 return FI.getReturnInfo().isIndirect();
1430 bool CodeGenModule::ReturnSlotInterferesWithArgs(const CGFunctionInfo &FI) {
1431 return ReturnTypeUsesSRet(FI) &&
1432 getTargetCodeGenInfo().doesReturnSlotInterfereWithArgs();
1435 bool CodeGenModule::ReturnTypeUsesFPRet(QualType ResultType) {
1436 if (const BuiltinType *BT = ResultType->getAs<BuiltinType>()) {
1437 switch (BT->getKind()) {
1440 case BuiltinType::Float:
1441 return getTarget().useObjCFPRetForRealType(TargetInfo::Float);
1442 case BuiltinType::Double:
1443 return getTarget().useObjCFPRetForRealType(TargetInfo::Double);
1444 case BuiltinType::LongDouble:
1445 return getTarget().useObjCFPRetForRealType(TargetInfo::LongDouble);
1452 bool CodeGenModule::ReturnTypeUsesFP2Ret(QualType ResultType) {
1453 if (const ComplexType *CT = ResultType->getAs<ComplexType>()) {
1454 if (const BuiltinType *BT = CT->getElementType()->getAs<BuiltinType>()) {
1455 if (BT->getKind() == BuiltinType::LongDouble)
1456 return getTarget().useObjCFP2RetForComplexLongDouble();
1463 llvm::FunctionType *CodeGenTypes::GetFunctionType(GlobalDecl GD) {
1464 const CGFunctionInfo &FI = arrangeGlobalDeclaration(GD);
1465 return GetFunctionType(FI);
1468 llvm::FunctionType *
1469 CodeGenTypes::GetFunctionType(const CGFunctionInfo &FI) {
1471 bool Inserted = FunctionsBeingProcessed.insert(&FI).second;
1473 assert(Inserted && "Recursively being processed?");
1475 llvm::Type *resultType = nullptr;
1476 const ABIArgInfo &retAI = FI.getReturnInfo();
1477 switch (retAI.getKind()) {
1478 case ABIArgInfo::Expand:
1479 llvm_unreachable("Invalid ABI kind for return argument");
1481 case ABIArgInfo::Extend:
1482 case ABIArgInfo::Direct:
1483 resultType = retAI.getCoerceToType();
1486 case ABIArgInfo::InAlloca:
1487 if (retAI.getInAllocaSRet()) {
1488 // sret things on win32 aren't void, they return the sret pointer.
1489 QualType ret = FI.getReturnType();
1490 llvm::Type *ty = ConvertType(ret);
1491 unsigned addressSpace = Context.getTargetAddressSpace(ret);
1492 resultType = llvm::PointerType::get(ty, addressSpace);
1494 resultType = llvm::Type::getVoidTy(getLLVMContext());
1498 case ABIArgInfo::Indirect:
1499 case ABIArgInfo::Ignore:
1500 resultType = llvm::Type::getVoidTy(getLLVMContext());
1503 case ABIArgInfo::CoerceAndExpand:
1504 resultType = retAI.getUnpaddedCoerceAndExpandType();
1508 ClangToLLVMArgMapping IRFunctionArgs(getContext(), FI, true);
1509 SmallVector<llvm::Type*, 8> ArgTypes(IRFunctionArgs.totalIRArgs());
1511 // Add type for sret argument.
1512 if (IRFunctionArgs.hasSRetArg()) {
1513 QualType Ret = FI.getReturnType();
1514 llvm::Type *Ty = ConvertType(Ret);
1515 unsigned AddressSpace = Context.getTargetAddressSpace(Ret);
1516 ArgTypes[IRFunctionArgs.getSRetArgNo()] =
1517 llvm::PointerType::get(Ty, AddressSpace);
1520 // Add type for inalloca argument.
1521 if (IRFunctionArgs.hasInallocaArg()) {
1522 auto ArgStruct = FI.getArgStruct();
1524 ArgTypes[IRFunctionArgs.getInallocaArgNo()] = ArgStruct->getPointerTo();
1527 // Add in all of the required arguments.
1529 CGFunctionInfo::const_arg_iterator it = FI.arg_begin(),
1530 ie = it + FI.getNumRequiredArgs();
1531 for (; it != ie; ++it, ++ArgNo) {
1532 const ABIArgInfo &ArgInfo = it->info;
1534 // Insert a padding type to ensure proper alignment.
1535 if (IRFunctionArgs.hasPaddingArg(ArgNo))
1536 ArgTypes[IRFunctionArgs.getPaddingArgNo(ArgNo)] =
1537 ArgInfo.getPaddingType();
1539 unsigned FirstIRArg, NumIRArgs;
1540 std::tie(FirstIRArg, NumIRArgs) = IRFunctionArgs.getIRArgs(ArgNo);
1542 switch (ArgInfo.getKind()) {
1543 case ABIArgInfo::Ignore:
1544 case ABIArgInfo::InAlloca:
1545 assert(NumIRArgs == 0);
1548 case ABIArgInfo::Indirect: {
1549 assert(NumIRArgs == 1);
1550 // indirect arguments are always on the stack, which is addr space #0.
1551 llvm::Type *LTy = ConvertTypeForMem(it->type);
1552 ArgTypes[FirstIRArg] = LTy->getPointerTo();
1556 case ABIArgInfo::Extend:
1557 case ABIArgInfo::Direct: {
1558 // Fast-isel and the optimizer generally like scalar values better than
1559 // FCAs, so we flatten them if this is safe to do for this argument.
1560 llvm::Type *argType = ArgInfo.getCoerceToType();
1561 llvm::StructType *st = dyn_cast<llvm::StructType>(argType);
1562 if (st && ArgInfo.isDirect() && ArgInfo.getCanBeFlattened()) {
1563 assert(NumIRArgs == st->getNumElements());
1564 for (unsigned i = 0, e = st->getNumElements(); i != e; ++i)
1565 ArgTypes[FirstIRArg + i] = st->getElementType(i);
1567 assert(NumIRArgs == 1);
1568 ArgTypes[FirstIRArg] = argType;
1573 case ABIArgInfo::CoerceAndExpand: {
1574 auto ArgTypesIter = ArgTypes.begin() + FirstIRArg;
1575 for (auto EltTy : ArgInfo.getCoerceAndExpandTypeSequence()) {
1576 *ArgTypesIter++ = EltTy;
1578 assert(ArgTypesIter == ArgTypes.begin() + FirstIRArg + NumIRArgs);
1582 case ABIArgInfo::Expand:
1583 auto ArgTypesIter = ArgTypes.begin() + FirstIRArg;
1584 getExpandedTypes(it->type, ArgTypesIter);
1585 assert(ArgTypesIter == ArgTypes.begin() + FirstIRArg + NumIRArgs);
1590 bool Erased = FunctionsBeingProcessed.erase(&FI); (void)Erased;
1591 assert(Erased && "Not in set?");
1593 return llvm::FunctionType::get(resultType, ArgTypes, FI.isVariadic());
1596 llvm::Type *CodeGenTypes::GetFunctionTypeForVTable(GlobalDecl GD) {
1597 const CXXMethodDecl *MD = cast<CXXMethodDecl>(GD.getDecl());
1598 const FunctionProtoType *FPT = MD->getType()->getAs<FunctionProtoType>();
1600 if (!isFuncTypeConvertible(FPT))
1601 return llvm::StructType::get(getLLVMContext());
1603 const CGFunctionInfo *Info;
1604 if (isa<CXXDestructorDecl>(MD))
1606 &arrangeCXXStructorDeclaration(MD, getFromDtorType(GD.getDtorType()));
1608 Info = &arrangeCXXMethodDeclaration(MD);
1609 return GetFunctionType(*Info);
1612 static void AddAttributesFromFunctionProtoType(ASTContext &Ctx,
1613 llvm::AttrBuilder &FuncAttrs,
1614 const FunctionProtoType *FPT) {
1618 if (!isUnresolvedExceptionSpec(FPT->getExceptionSpecType()) &&
1619 FPT->isNothrow(Ctx))
1620 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1623 void CodeGenModule::ConstructAttributeList(
1624 StringRef Name, const CGFunctionInfo &FI, CGCalleeInfo CalleeInfo,
1625 AttributeListType &PAL, unsigned &CallingConv, bool AttrOnCallSite) {
1626 llvm::AttrBuilder FuncAttrs;
1627 llvm::AttrBuilder RetAttrs;
1628 bool HasOptnone = false;
1630 CallingConv = FI.getEffectiveCallingConvention();
1632 if (FI.isNoReturn())
1633 FuncAttrs.addAttribute(llvm::Attribute::NoReturn);
1635 // If we have information about the function prototype, we can learn
1636 // attributes form there.
1637 AddAttributesFromFunctionProtoType(getContext(), FuncAttrs,
1638 CalleeInfo.getCalleeFunctionProtoType());
1640 const Decl *TargetDecl = CalleeInfo.getCalleeDecl();
1642 bool HasAnyX86InterruptAttr = false;
1643 // FIXME: handle sseregparm someday...
1645 if (TargetDecl->hasAttr<ReturnsTwiceAttr>())
1646 FuncAttrs.addAttribute(llvm::Attribute::ReturnsTwice);
1647 if (TargetDecl->hasAttr<NoThrowAttr>())
1648 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1649 if (TargetDecl->hasAttr<NoReturnAttr>())
1650 FuncAttrs.addAttribute(llvm::Attribute::NoReturn);
1651 if (TargetDecl->hasAttr<NoDuplicateAttr>())
1652 FuncAttrs.addAttribute(llvm::Attribute::NoDuplicate);
1653 if (TargetDecl->hasAttr<ConvergentAttr>())
1654 FuncAttrs.addAttribute(llvm::Attribute::Convergent);
1656 if (const FunctionDecl *Fn = dyn_cast<FunctionDecl>(TargetDecl)) {
1657 AddAttributesFromFunctionProtoType(
1658 getContext(), FuncAttrs, Fn->getType()->getAs<FunctionProtoType>());
1659 // Don't use [[noreturn]] or _Noreturn for a call to a virtual function.
1660 // These attributes are not inherited by overloads.
1661 const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(Fn);
1662 if (Fn->isNoReturn() && !(AttrOnCallSite && MD && MD->isVirtual()))
1663 FuncAttrs.addAttribute(llvm::Attribute::NoReturn);
1666 // 'const', 'pure' and 'noalias' attributed functions are also nounwind.
1667 if (TargetDecl->hasAttr<ConstAttr>()) {
1668 FuncAttrs.addAttribute(llvm::Attribute::ReadNone);
1669 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1670 } else if (TargetDecl->hasAttr<PureAttr>()) {
1671 FuncAttrs.addAttribute(llvm::Attribute::ReadOnly);
1672 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1673 } else if (TargetDecl->hasAttr<NoAliasAttr>()) {
1674 FuncAttrs.addAttribute(llvm::Attribute::ArgMemOnly);
1675 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1677 if (TargetDecl->hasAttr<RestrictAttr>())
1678 RetAttrs.addAttribute(llvm::Attribute::NoAlias);
1679 if (TargetDecl->hasAttr<ReturnsNonNullAttr>())
1680 RetAttrs.addAttribute(llvm::Attribute::NonNull);
1682 HasAnyX86InterruptAttr = TargetDecl->hasAttr<AnyX86InterruptAttr>();
1683 HasOptnone = TargetDecl->hasAttr<OptimizeNoneAttr>();
1684 if (auto *AllocSize = TargetDecl->getAttr<AllocSizeAttr>()) {
1685 Optional<unsigned> NumElemsParam;
1686 // alloc_size args are base-1, 0 means not present.
1687 if (unsigned N = AllocSize->getNumElemsParam())
1688 NumElemsParam = N - 1;
1689 FuncAttrs.addAllocSizeAttr(AllocSize->getElemSizeParam() - 1,
1694 // OptimizeNoneAttr takes precedence over -Os or -Oz. No warning needed.
1696 if (CodeGenOpts.OptimizeSize)
1697 FuncAttrs.addAttribute(llvm::Attribute::OptimizeForSize);
1698 if (CodeGenOpts.OptimizeSize == 2)
1699 FuncAttrs.addAttribute(llvm::Attribute::MinSize);
1702 if (CodeGenOpts.DisableRedZone)
1703 FuncAttrs.addAttribute(llvm::Attribute::NoRedZone);
1704 if (CodeGenOpts.NoImplicitFloat)
1705 FuncAttrs.addAttribute(llvm::Attribute::NoImplicitFloat);
1706 if (CodeGenOpts.EnableSegmentedStacks &&
1707 !(TargetDecl && TargetDecl->hasAttr<NoSplitStackAttr>()))
1708 FuncAttrs.addAttribute("split-stack");
1710 if (AttrOnCallSite) {
1711 // Attributes that should go on the call site only.
1712 if (!CodeGenOpts.SimplifyLibCalls ||
1713 CodeGenOpts.isNoBuiltinFunc(Name.data()))
1714 FuncAttrs.addAttribute(llvm::Attribute::NoBuiltin);
1715 if (!CodeGenOpts.TrapFuncName.empty())
1716 FuncAttrs.addAttribute("trap-func-name", CodeGenOpts.TrapFuncName);
1718 // Attributes that should go on the function, but not the call site.
1719 if (!CodeGenOpts.DisableFPElim) {
1720 FuncAttrs.addAttribute("no-frame-pointer-elim", "false");
1721 } else if (CodeGenOpts.OmitLeafFramePointer) {
1722 FuncAttrs.addAttribute("no-frame-pointer-elim", "false");
1723 FuncAttrs.addAttribute("no-frame-pointer-elim-non-leaf");
1725 FuncAttrs.addAttribute("no-frame-pointer-elim", "true");
1726 FuncAttrs.addAttribute("no-frame-pointer-elim-non-leaf");
1729 bool DisableTailCalls =
1730 CodeGenOpts.DisableTailCalls || HasAnyX86InterruptAttr ||
1731 (TargetDecl && TargetDecl->hasAttr<DisableTailCallsAttr>());
1732 FuncAttrs.addAttribute(
1733 "disable-tail-calls",
1734 llvm::toStringRef(DisableTailCalls));
1736 FuncAttrs.addAttribute("less-precise-fpmad",
1737 llvm::toStringRef(CodeGenOpts.LessPreciseFPMAD));
1739 if (!CodeGenOpts.FPDenormalMode.empty())
1740 FuncAttrs.addAttribute("denormal-fp-math",
1741 CodeGenOpts.FPDenormalMode);
1743 FuncAttrs.addAttribute("no-trapping-math",
1744 llvm::toStringRef(CodeGenOpts.NoTrappingMath));
1746 // TODO: Are these all needed?
1747 // unsafe/inf/nan/nsz are handled by instruction-level FastMathFlags.
1748 FuncAttrs.addAttribute("no-infs-fp-math",
1749 llvm::toStringRef(CodeGenOpts.NoInfsFPMath));
1750 FuncAttrs.addAttribute("no-nans-fp-math",
1751 llvm::toStringRef(CodeGenOpts.NoNaNsFPMath));
1752 FuncAttrs.addAttribute("unsafe-fp-math",
1753 llvm::toStringRef(CodeGenOpts.UnsafeFPMath));
1754 FuncAttrs.addAttribute("use-soft-float",
1755 llvm::toStringRef(CodeGenOpts.SoftFloat));
1756 FuncAttrs.addAttribute("stack-protector-buffer-size",
1757 llvm::utostr(CodeGenOpts.SSPBufferSize));
1758 FuncAttrs.addAttribute("no-signed-zeros-fp-math",
1759 llvm::toStringRef(CodeGenOpts.NoSignedZeros));
1760 FuncAttrs.addAttribute(
1761 "correctly-rounded-divide-sqrt-fp-math",
1762 llvm::toStringRef(CodeGenOpts.CorrectlyRoundedDivSqrt));
1764 // TODO: Reciprocal estimate codegen options should apply to instructions?
1765 std::vector<std::string> &Recips = getTarget().getTargetOpts().Reciprocals;
1766 if (!Recips.empty())
1767 FuncAttrs.addAttribute("reciprocal-estimates",
1768 llvm::join(Recips.begin(), Recips.end(), ","));
1770 if (CodeGenOpts.StackRealignment)
1771 FuncAttrs.addAttribute("stackrealign");
1772 if (CodeGenOpts.Backchain)
1773 FuncAttrs.addAttribute("backchain");
1775 // Add target-cpu and target-features attributes to functions. If
1776 // we have a decl for the function and it has a target attribute then
1777 // parse that and add it to the feature set.
1778 StringRef TargetCPU = getTarget().getTargetOpts().CPU;
1779 const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(TargetDecl);
1780 if (FD && FD->hasAttr<TargetAttr>()) {
1781 llvm::StringMap<bool> FeatureMap;
1782 getFunctionFeatureMap(FeatureMap, FD);
1784 // Produce the canonical string for this set of features.
1785 std::vector<std::string> Features;
1786 for (llvm::StringMap<bool>::const_iterator it = FeatureMap.begin(),
1787 ie = FeatureMap.end();
1789 Features.push_back((it->second ? "+" : "-") + it->first().str());
1791 // Now add the target-cpu and target-features to the function.
1792 // While we populated the feature map above, we still need to
1793 // get and parse the target attribute so we can get the cpu for
1795 const auto *TD = FD->getAttr<TargetAttr>();
1796 TargetAttr::ParsedTargetAttr ParsedAttr = TD->parse();
1797 if (ParsedAttr.second != "")
1798 TargetCPU = ParsedAttr.second;
1799 if (TargetCPU != "")
1800 FuncAttrs.addAttribute("target-cpu", TargetCPU);
1801 if (!Features.empty()) {
1802 std::sort(Features.begin(), Features.end());
1803 FuncAttrs.addAttribute(
1805 llvm::join(Features.begin(), Features.end(), ","));
1808 // Otherwise just add the existing target cpu and target features to the
1810 std::vector<std::string> &Features = getTarget().getTargetOpts().Features;
1811 if (TargetCPU != "")
1812 FuncAttrs.addAttribute("target-cpu", TargetCPU);
1813 if (!Features.empty()) {
1814 std::sort(Features.begin(), Features.end());
1815 FuncAttrs.addAttribute(
1817 llvm::join(Features.begin(), Features.end(), ","));
1822 if (getLangOpts().CUDA && getLangOpts().CUDAIsDevice) {
1823 // Conservatively, mark all functions and calls in CUDA as convergent
1824 // (meaning, they may call an intrinsically convergent op, such as
1825 // __syncthreads(), and so can't have certain optimizations applied around
1826 // them). LLVM will remove this attribute where it safely can.
1827 FuncAttrs.addAttribute(llvm::Attribute::Convergent);
1829 // Exceptions aren't supported in CUDA device code.
1830 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1832 // Respect -fcuda-flush-denormals-to-zero.
1833 if (getLangOpts().CUDADeviceFlushDenormalsToZero)
1834 FuncAttrs.addAttribute("nvptx-f32ftz", "true");
1837 ClangToLLVMArgMapping IRFunctionArgs(getContext(), FI);
1839 QualType RetTy = FI.getReturnType();
1840 const ABIArgInfo &RetAI = FI.getReturnInfo();
1841 switch (RetAI.getKind()) {
1842 case ABIArgInfo::Extend:
1843 if (RetTy->hasSignedIntegerRepresentation())
1844 RetAttrs.addAttribute(llvm::Attribute::SExt);
1845 else if (RetTy->hasUnsignedIntegerRepresentation())
1846 RetAttrs.addAttribute(llvm::Attribute::ZExt);
1848 case ABIArgInfo::Direct:
1849 if (RetAI.getInReg())
1850 RetAttrs.addAttribute(llvm::Attribute::InReg);
1852 case ABIArgInfo::Ignore:
1855 case ABIArgInfo::InAlloca:
1856 case ABIArgInfo::Indirect: {
1857 // inalloca and sret disable readnone and readonly
1858 FuncAttrs.removeAttribute(llvm::Attribute::ReadOnly)
1859 .removeAttribute(llvm::Attribute::ReadNone);
1863 case ABIArgInfo::CoerceAndExpand:
1866 case ABIArgInfo::Expand:
1867 llvm_unreachable("Invalid ABI kind for return argument");
1870 if (const auto *RefTy = RetTy->getAs<ReferenceType>()) {
1871 QualType PTy = RefTy->getPointeeType();
1872 if (!PTy->isIncompleteType() && PTy->isConstantSizeType())
1873 RetAttrs.addDereferenceableAttr(getContext().getTypeSizeInChars(PTy)
1875 else if (getContext().getTargetAddressSpace(PTy) == 0)
1876 RetAttrs.addAttribute(llvm::Attribute::NonNull);
1879 // Attach return attributes.
1880 if (RetAttrs.hasAttributes()) {
1881 PAL.push_back(llvm::AttributeSet::get(
1882 getLLVMContext(), llvm::AttributeSet::ReturnIndex, RetAttrs));
1885 bool hasUsedSRet = false;
1887 // Attach attributes to sret.
1888 if (IRFunctionArgs.hasSRetArg()) {
1889 llvm::AttrBuilder SRETAttrs;
1890 SRETAttrs.addAttribute(llvm::Attribute::StructRet);
1892 if (RetAI.getInReg())
1893 SRETAttrs.addAttribute(llvm::Attribute::InReg);
1894 PAL.push_back(llvm::AttributeSet::get(
1895 getLLVMContext(), IRFunctionArgs.getSRetArgNo() + 1, SRETAttrs));
1898 // Attach attributes to inalloca argument.
1899 if (IRFunctionArgs.hasInallocaArg()) {
1900 llvm::AttrBuilder Attrs;
1901 Attrs.addAttribute(llvm::Attribute::InAlloca);
1902 PAL.push_back(llvm::AttributeSet::get(
1903 getLLVMContext(), IRFunctionArgs.getInallocaArgNo() + 1, Attrs));
1907 for (CGFunctionInfo::const_arg_iterator I = FI.arg_begin(),
1909 I != E; ++I, ++ArgNo) {
1910 QualType ParamType = I->type;
1911 const ABIArgInfo &AI = I->info;
1912 llvm::AttrBuilder Attrs;
1914 // Add attribute for padding argument, if necessary.
1915 if (IRFunctionArgs.hasPaddingArg(ArgNo)) {
1916 if (AI.getPaddingInReg())
1917 PAL.push_back(llvm::AttributeSet::get(
1918 getLLVMContext(), IRFunctionArgs.getPaddingArgNo(ArgNo) + 1,
1919 llvm::Attribute::InReg));
1922 // 'restrict' -> 'noalias' is done in EmitFunctionProlog when we
1923 // have the corresponding parameter variable. It doesn't make
1924 // sense to do it here because parameters are so messed up.
1925 switch (AI.getKind()) {
1926 case ABIArgInfo::Extend:
1927 if (ParamType->isSignedIntegerOrEnumerationType())
1928 Attrs.addAttribute(llvm::Attribute::SExt);
1929 else if (ParamType->isUnsignedIntegerOrEnumerationType()) {
1930 if (getTypes().getABIInfo().shouldSignExtUnsignedType(ParamType))
1931 Attrs.addAttribute(llvm::Attribute::SExt);
1933 Attrs.addAttribute(llvm::Attribute::ZExt);
1936 case ABIArgInfo::Direct:
1937 if (ArgNo == 0 && FI.isChainCall())
1938 Attrs.addAttribute(llvm::Attribute::Nest);
1939 else if (AI.getInReg())
1940 Attrs.addAttribute(llvm::Attribute::InReg);
1943 case ABIArgInfo::Indirect: {
1945 Attrs.addAttribute(llvm::Attribute::InReg);
1947 if (AI.getIndirectByVal())
1948 Attrs.addAttribute(llvm::Attribute::ByVal);
1950 CharUnits Align = AI.getIndirectAlign();
1952 // In a byval argument, it is important that the required
1953 // alignment of the type is honored, as LLVM might be creating a
1954 // *new* stack object, and needs to know what alignment to give
1955 // it. (Sometimes it can deduce a sensible alignment on its own,
1956 // but not if clang decides it must emit a packed struct, or the
1957 // user specifies increased alignment requirements.)
1959 // This is different from indirect *not* byval, where the object
1960 // exists already, and the align attribute is purely
1962 assert(!Align.isZero());
1964 // For now, only add this when we have a byval argument.
1965 // TODO: be less lazy about updating test cases.
1966 if (AI.getIndirectByVal())
1967 Attrs.addAlignmentAttr(Align.getQuantity());
1969 // byval disables readnone and readonly.
1970 FuncAttrs.removeAttribute(llvm::Attribute::ReadOnly)
1971 .removeAttribute(llvm::Attribute::ReadNone);
1974 case ABIArgInfo::Ignore:
1975 case ABIArgInfo::Expand:
1976 case ABIArgInfo::CoerceAndExpand:
1979 case ABIArgInfo::InAlloca:
1980 // inalloca disables readnone and readonly.
1981 FuncAttrs.removeAttribute(llvm::Attribute::ReadOnly)
1982 .removeAttribute(llvm::Attribute::ReadNone);
1986 if (const auto *RefTy = ParamType->getAs<ReferenceType>()) {
1987 QualType PTy = RefTy->getPointeeType();
1988 if (!PTy->isIncompleteType() && PTy->isConstantSizeType())
1989 Attrs.addDereferenceableAttr(getContext().getTypeSizeInChars(PTy)
1991 else if (getContext().getTargetAddressSpace(PTy) == 0)
1992 Attrs.addAttribute(llvm::Attribute::NonNull);
1995 switch (FI.getExtParameterInfo(ArgNo).getABI()) {
1996 case ParameterABI::Ordinary:
1999 case ParameterABI::SwiftIndirectResult: {
2000 // Add 'sret' if we haven't already used it for something, but
2001 // only if the result is void.
2002 if (!hasUsedSRet && RetTy->isVoidType()) {
2003 Attrs.addAttribute(llvm::Attribute::StructRet);
2007 // Add 'noalias' in either case.
2008 Attrs.addAttribute(llvm::Attribute::NoAlias);
2010 // Add 'dereferenceable' and 'alignment'.
2011 auto PTy = ParamType->getPointeeType();
2012 if (!PTy->isIncompleteType() && PTy->isConstantSizeType()) {
2013 auto info = getContext().getTypeInfoInChars(PTy);
2014 Attrs.addDereferenceableAttr(info.first.getQuantity());
2015 Attrs.addAttribute(llvm::Attribute::getWithAlignment(getLLVMContext(),
2016 info.second.getQuantity()));
2021 case ParameterABI::SwiftErrorResult:
2022 Attrs.addAttribute(llvm::Attribute::SwiftError);
2025 case ParameterABI::SwiftContext:
2026 Attrs.addAttribute(llvm::Attribute::SwiftSelf);
2030 if (Attrs.hasAttributes()) {
2031 unsigned FirstIRArg, NumIRArgs;
2032 std::tie(FirstIRArg, NumIRArgs) = IRFunctionArgs.getIRArgs(ArgNo);
2033 for (unsigned i = 0; i < NumIRArgs; i++)
2034 PAL.push_back(llvm::AttributeSet::get(getLLVMContext(),
2035 FirstIRArg + i + 1, Attrs));
2038 assert(ArgNo == FI.arg_size());
2040 if (FuncAttrs.hasAttributes())
2041 PAL.push_back(llvm::
2042 AttributeSet::get(getLLVMContext(),
2043 llvm::AttributeSet::FunctionIndex,
2047 /// An argument came in as a promoted argument; demote it back to its
2049 static llvm::Value *emitArgumentDemotion(CodeGenFunction &CGF,
2051 llvm::Value *value) {
2052 llvm::Type *varType = CGF.ConvertType(var->getType());
2054 // This can happen with promotions that actually don't change the
2055 // underlying type, like the enum promotions.
2056 if (value->getType() == varType) return value;
2058 assert((varType->isIntegerTy() || varType->isFloatingPointTy())
2059 && "unexpected promotion type");
2061 if (isa<llvm::IntegerType>(varType))
2062 return CGF.Builder.CreateTrunc(value, varType, "arg.unpromote");
2064 return CGF.Builder.CreateFPCast(value, varType, "arg.unpromote");
2067 /// Returns the attribute (either parameter attribute, or function
2068 /// attribute), which declares argument ArgNo to be non-null.
2069 static const NonNullAttr *getNonNullAttr(const Decl *FD, const ParmVarDecl *PVD,
2070 QualType ArgType, unsigned ArgNo) {
2071 // FIXME: __attribute__((nonnull)) can also be applied to:
2072 // - references to pointers, where the pointee is known to be
2073 // nonnull (apparently a Clang extension)
2074 // - transparent unions containing pointers
2075 // In the former case, LLVM IR cannot represent the constraint. In
2076 // the latter case, we have no guarantee that the transparent union
2077 // is in fact passed as a pointer.
2078 if (!ArgType->isAnyPointerType() && !ArgType->isBlockPointerType())
2080 // First, check attribute on parameter itself.
2082 if (auto ParmNNAttr = PVD->getAttr<NonNullAttr>())
2085 // Check function attributes.
2088 for (const auto *NNAttr : FD->specific_attrs<NonNullAttr>()) {
2089 if (NNAttr->isNonNull(ArgNo))
2096 struct CopyBackSwiftError final : EHScopeStack::Cleanup {
2099 CopyBackSwiftError(Address temp, Address arg) : Temp(temp), Arg(arg) {}
2100 void Emit(CodeGenFunction &CGF, Flags flags) override {
2101 llvm::Value *errorValue = CGF.Builder.CreateLoad(Temp);
2102 CGF.Builder.CreateStore(errorValue, Arg);
2107 void CodeGenFunction::EmitFunctionProlog(const CGFunctionInfo &FI,
2109 const FunctionArgList &Args) {
2110 if (CurCodeDecl && CurCodeDecl->hasAttr<NakedAttr>())
2111 // Naked functions don't have prologues.
2114 // If this is an implicit-return-zero function, go ahead and
2115 // initialize the return value. TODO: it might be nice to have
2116 // a more general mechanism for this that didn't require synthesized
2117 // return statements.
2118 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(CurCodeDecl)) {
2119 if (FD->hasImplicitReturnZero()) {
2120 QualType RetTy = FD->getReturnType().getUnqualifiedType();
2121 llvm::Type* LLVMTy = CGM.getTypes().ConvertType(RetTy);
2122 llvm::Constant* Zero = llvm::Constant::getNullValue(LLVMTy);
2123 Builder.CreateStore(Zero, ReturnValue);
2127 // FIXME: We no longer need the types from FunctionArgList; lift up and
2130 ClangToLLVMArgMapping IRFunctionArgs(CGM.getContext(), FI);
2131 // Flattened function arguments.
2132 SmallVector<llvm::Value *, 16> FnArgs;
2133 FnArgs.reserve(IRFunctionArgs.totalIRArgs());
2134 for (auto &Arg : Fn->args()) {
2135 FnArgs.push_back(&Arg);
2137 assert(FnArgs.size() == IRFunctionArgs.totalIRArgs());
2139 // If we're using inalloca, all the memory arguments are GEPs off of the last
2140 // parameter, which is a pointer to the complete memory area.
2141 Address ArgStruct = Address::invalid();
2142 const llvm::StructLayout *ArgStructLayout = nullptr;
2143 if (IRFunctionArgs.hasInallocaArg()) {
2144 ArgStructLayout = CGM.getDataLayout().getStructLayout(FI.getArgStruct());
2145 ArgStruct = Address(FnArgs[IRFunctionArgs.getInallocaArgNo()],
2146 FI.getArgStructAlignment());
2148 assert(ArgStruct.getType() == FI.getArgStruct()->getPointerTo());
2151 // Name the struct return parameter.
2152 if (IRFunctionArgs.hasSRetArg()) {
2153 auto AI = cast<llvm::Argument>(FnArgs[IRFunctionArgs.getSRetArgNo()]);
2154 AI->setName("agg.result");
2155 AI->addAttr(llvm::AttributeSet::get(getLLVMContext(), AI->getArgNo() + 1,
2156 llvm::Attribute::NoAlias));
2159 // Track if we received the parameter as a pointer (indirect, byval, or
2160 // inalloca). If already have a pointer, EmitParmDecl doesn't need to copy it
2161 // into a local alloca for us.
2162 SmallVector<ParamValue, 16> ArgVals;
2163 ArgVals.reserve(Args.size());
2165 // Create a pointer value for every parameter declaration. This usually
2166 // entails copying one or more LLVM IR arguments into an alloca. Don't push
2167 // any cleanups or do anything that might unwind. We do that separately, so
2168 // we can push the cleanups in the correct order for the ABI.
2169 assert(FI.arg_size() == Args.size() &&
2170 "Mismatch between function signature & arguments.");
2172 CGFunctionInfo::const_arg_iterator info_it = FI.arg_begin();
2173 for (FunctionArgList::const_iterator i = Args.begin(), e = Args.end();
2174 i != e; ++i, ++info_it, ++ArgNo) {
2175 const VarDecl *Arg = *i;
2176 QualType Ty = info_it->type;
2177 const ABIArgInfo &ArgI = info_it->info;
2180 isa<ParmVarDecl>(Arg) && cast<ParmVarDecl>(Arg)->isKNRPromoted();
2182 unsigned FirstIRArg, NumIRArgs;
2183 std::tie(FirstIRArg, NumIRArgs) = IRFunctionArgs.getIRArgs(ArgNo);
2185 switch (ArgI.getKind()) {
2186 case ABIArgInfo::InAlloca: {
2187 assert(NumIRArgs == 0);
2188 auto FieldIndex = ArgI.getInAllocaFieldIndex();
2189 CharUnits FieldOffset =
2190 CharUnits::fromQuantity(ArgStructLayout->getElementOffset(FieldIndex));
2191 Address V = Builder.CreateStructGEP(ArgStruct, FieldIndex, FieldOffset,
2193 ArgVals.push_back(ParamValue::forIndirect(V));
2197 case ABIArgInfo::Indirect: {
2198 assert(NumIRArgs == 1);
2199 Address ParamAddr = Address(FnArgs[FirstIRArg], ArgI.getIndirectAlign());
2201 if (!hasScalarEvaluationKind(Ty)) {
2202 // Aggregates and complex variables are accessed by reference. All we
2203 // need to do is realign the value, if requested.
2204 Address V = ParamAddr;
2205 if (ArgI.getIndirectRealign()) {
2206 Address AlignedTemp = CreateMemTemp(Ty, "coerce");
2208 // Copy from the incoming argument pointer to the temporary with the
2209 // appropriate alignment.
2211 // FIXME: We should have a common utility for generating an aggregate
2213 CharUnits Size = getContext().getTypeSizeInChars(Ty);
2214 auto SizeVal = llvm::ConstantInt::get(IntPtrTy, Size.getQuantity());
2215 Address Dst = Builder.CreateBitCast(AlignedTemp, Int8PtrTy);
2216 Address Src = Builder.CreateBitCast(ParamAddr, Int8PtrTy);
2217 Builder.CreateMemCpy(Dst, Src, SizeVal, false);
2220 ArgVals.push_back(ParamValue::forIndirect(V));
2222 // Load scalar value from indirect argument.
2224 EmitLoadOfScalar(ParamAddr, false, Ty, Arg->getLocStart());
2227 V = emitArgumentDemotion(*this, Arg, V);
2228 ArgVals.push_back(ParamValue::forDirect(V));
2233 case ABIArgInfo::Extend:
2234 case ABIArgInfo::Direct: {
2236 // If we have the trivial case, handle it with no muss and fuss.
2237 if (!isa<llvm::StructType>(ArgI.getCoerceToType()) &&
2238 ArgI.getCoerceToType() == ConvertType(Ty) &&
2239 ArgI.getDirectOffset() == 0) {
2240 assert(NumIRArgs == 1);
2241 llvm::Value *V = FnArgs[FirstIRArg];
2242 auto AI = cast<llvm::Argument>(V);
2244 if (const ParmVarDecl *PVD = dyn_cast<ParmVarDecl>(Arg)) {
2245 if (getNonNullAttr(CurCodeDecl, PVD, PVD->getType(),
2246 PVD->getFunctionScopeIndex()))
2247 AI->addAttr(llvm::AttributeSet::get(getLLVMContext(),
2249 llvm::Attribute::NonNull));
2251 QualType OTy = PVD->getOriginalType();
2252 if (const auto *ArrTy =
2253 getContext().getAsConstantArrayType(OTy)) {
2254 // A C99 array parameter declaration with the static keyword also
2255 // indicates dereferenceability, and if the size is constant we can
2256 // use the dereferenceable attribute (which requires the size in
2258 if (ArrTy->getSizeModifier() == ArrayType::Static) {
2259 QualType ETy = ArrTy->getElementType();
2260 uint64_t ArrSize = ArrTy->getSize().getZExtValue();
2261 if (!ETy->isIncompleteType() && ETy->isConstantSizeType() &&
2263 llvm::AttrBuilder Attrs;
2264 Attrs.addDereferenceableAttr(
2265 getContext().getTypeSizeInChars(ETy).getQuantity()*ArrSize);
2266 AI->addAttr(llvm::AttributeSet::get(getLLVMContext(),
2267 AI->getArgNo() + 1, Attrs));
2268 } else if (getContext().getTargetAddressSpace(ETy) == 0) {
2269 AI->addAttr(llvm::AttributeSet::get(getLLVMContext(),
2271 llvm::Attribute::NonNull));
2274 } else if (const auto *ArrTy =
2275 getContext().getAsVariableArrayType(OTy)) {
2276 // For C99 VLAs with the static keyword, we don't know the size so
2277 // we can't use the dereferenceable attribute, but in addrspace(0)
2278 // we know that it must be nonnull.
2279 if (ArrTy->getSizeModifier() == VariableArrayType::Static &&
2280 !getContext().getTargetAddressSpace(ArrTy->getElementType()))
2281 AI->addAttr(llvm::AttributeSet::get(getLLVMContext(),
2283 llvm::Attribute::NonNull));
2286 const auto *AVAttr = PVD->getAttr<AlignValueAttr>();
2288 if (const auto *TOTy = dyn_cast<TypedefType>(OTy))
2289 AVAttr = TOTy->getDecl()->getAttr<AlignValueAttr>();
2291 llvm::Value *AlignmentValue =
2292 EmitScalarExpr(AVAttr->getAlignment());
2293 llvm::ConstantInt *AlignmentCI =
2294 cast<llvm::ConstantInt>(AlignmentValue);
2295 unsigned Alignment =
2296 std::min((unsigned) AlignmentCI->getZExtValue(),
2297 +llvm::Value::MaximumAlignment);
2299 llvm::AttrBuilder Attrs;
2300 Attrs.addAlignmentAttr(Alignment);
2301 AI->addAttr(llvm::AttributeSet::get(getLLVMContext(),
2302 AI->getArgNo() + 1, Attrs));
2306 if (Arg->getType().isRestrictQualified())
2307 AI->addAttr(llvm::AttributeSet::get(getLLVMContext(),
2309 llvm::Attribute::NoAlias));
2311 // LLVM expects swifterror parameters to be used in very restricted
2312 // ways. Copy the value into a less-restricted temporary.
2313 if (FI.getExtParameterInfo(ArgNo).getABI()
2314 == ParameterABI::SwiftErrorResult) {
2315 QualType pointeeTy = Ty->getPointeeType();
2316 assert(pointeeTy->isPointerType());
2318 CreateMemTemp(pointeeTy, getPointerAlign(), "swifterror.temp");
2319 Address arg = Address(V, getContext().getTypeAlignInChars(pointeeTy));
2320 llvm::Value *incomingErrorValue = Builder.CreateLoad(arg);
2321 Builder.CreateStore(incomingErrorValue, temp);
2322 V = temp.getPointer();
2324 // Push a cleanup to copy the value back at the end of the function.
2325 // The convention does not guarantee that the value will be written
2326 // back if the function exits with an unwind exception.
2327 EHStack.pushCleanup<CopyBackSwiftError>(NormalCleanup, temp, arg);
2330 // Ensure the argument is the correct type.
2331 if (V->getType() != ArgI.getCoerceToType())
2332 V = Builder.CreateBitCast(V, ArgI.getCoerceToType());
2335 V = emitArgumentDemotion(*this, Arg, V);
2337 // Because of merging of function types from multiple decls it is
2338 // possible for the type of an argument to not match the corresponding
2339 // type in the function type. Since we are codegening the callee
2340 // in here, add a cast to the argument type.
2341 llvm::Type *LTy = ConvertType(Arg->getType());
2342 if (V->getType() != LTy)
2343 V = Builder.CreateBitCast(V, LTy);
2345 ArgVals.push_back(ParamValue::forDirect(V));
2349 Address Alloca = CreateMemTemp(Ty, getContext().getDeclAlign(Arg),
2352 // Pointer to store into.
2353 Address Ptr = emitAddressAtOffset(*this, Alloca, ArgI);
2355 // Fast-isel and the optimizer generally like scalar values better than
2356 // FCAs, so we flatten them if this is safe to do for this argument.
2357 llvm::StructType *STy = dyn_cast<llvm::StructType>(ArgI.getCoerceToType());
2358 if (ArgI.isDirect() && ArgI.getCanBeFlattened() && STy &&
2359 STy->getNumElements() > 1) {
2360 auto SrcLayout = CGM.getDataLayout().getStructLayout(STy);
2361 uint64_t SrcSize = CGM.getDataLayout().getTypeAllocSize(STy);
2362 llvm::Type *DstTy = Ptr.getElementType();
2363 uint64_t DstSize = CGM.getDataLayout().getTypeAllocSize(DstTy);
2365 Address AddrToStoreInto = Address::invalid();
2366 if (SrcSize <= DstSize) {
2368 Builder.CreateBitCast(Ptr, llvm::PointerType::getUnqual(STy));
2371 CreateTempAlloca(STy, Alloca.getAlignment(), "coerce");
2374 assert(STy->getNumElements() == NumIRArgs);
2375 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
2376 auto AI = FnArgs[FirstIRArg + i];
2377 AI->setName(Arg->getName() + ".coerce" + Twine(i));
2378 auto Offset = CharUnits::fromQuantity(SrcLayout->getElementOffset(i));
2380 Builder.CreateStructGEP(AddrToStoreInto, i, Offset);
2381 Builder.CreateStore(AI, EltPtr);
2384 if (SrcSize > DstSize) {
2385 Builder.CreateMemCpy(Ptr, AddrToStoreInto, DstSize);
2389 // Simple case, just do a coerced store of the argument into the alloca.
2390 assert(NumIRArgs == 1);
2391 auto AI = FnArgs[FirstIRArg];
2392 AI->setName(Arg->getName() + ".coerce");
2393 CreateCoercedStore(AI, Ptr, /*DestIsVolatile=*/false, *this);
2396 // Match to what EmitParmDecl is expecting for this type.
2397 if (CodeGenFunction::hasScalarEvaluationKind(Ty)) {
2399 EmitLoadOfScalar(Alloca, false, Ty, Arg->getLocStart());
2401 V = emitArgumentDemotion(*this, Arg, V);
2402 ArgVals.push_back(ParamValue::forDirect(V));
2404 ArgVals.push_back(ParamValue::forIndirect(Alloca));
2409 case ABIArgInfo::CoerceAndExpand: {
2410 // Reconstruct into a temporary.
2411 Address alloca = CreateMemTemp(Ty, getContext().getDeclAlign(Arg));
2412 ArgVals.push_back(ParamValue::forIndirect(alloca));
2414 auto coercionType = ArgI.getCoerceAndExpandType();
2415 alloca = Builder.CreateElementBitCast(alloca, coercionType);
2416 auto layout = CGM.getDataLayout().getStructLayout(coercionType);
2418 unsigned argIndex = FirstIRArg;
2419 for (unsigned i = 0, e = coercionType->getNumElements(); i != e; ++i) {
2420 llvm::Type *eltType = coercionType->getElementType(i);
2421 if (ABIArgInfo::isPaddingForCoerceAndExpand(eltType))
2424 auto eltAddr = Builder.CreateStructGEP(alloca, i, layout);
2425 auto elt = FnArgs[argIndex++];
2426 Builder.CreateStore(elt, eltAddr);
2428 assert(argIndex == FirstIRArg + NumIRArgs);
2432 case ABIArgInfo::Expand: {
2433 // If this structure was expanded into multiple arguments then
2434 // we need to create a temporary and reconstruct it from the
2436 Address Alloca = CreateMemTemp(Ty, getContext().getDeclAlign(Arg));
2437 LValue LV = MakeAddrLValue(Alloca, Ty);
2438 ArgVals.push_back(ParamValue::forIndirect(Alloca));
2440 auto FnArgIter = FnArgs.begin() + FirstIRArg;
2441 ExpandTypeFromArgs(Ty, LV, FnArgIter);
2442 assert(FnArgIter == FnArgs.begin() + FirstIRArg + NumIRArgs);
2443 for (unsigned i = 0, e = NumIRArgs; i != e; ++i) {
2444 auto AI = FnArgs[FirstIRArg + i];
2445 AI->setName(Arg->getName() + "." + Twine(i));
2450 case ABIArgInfo::Ignore:
2451 assert(NumIRArgs == 0);
2452 // Initialize the local variable appropriately.
2453 if (!hasScalarEvaluationKind(Ty)) {
2454 ArgVals.push_back(ParamValue::forIndirect(CreateMemTemp(Ty)));
2456 llvm::Value *U = llvm::UndefValue::get(ConvertType(Arg->getType()));
2457 ArgVals.push_back(ParamValue::forDirect(U));
2463 if (getTarget().getCXXABI().areArgsDestroyedLeftToRightInCallee()) {
2464 for (int I = Args.size() - 1; I >= 0; --I)
2465 EmitParmDecl(*Args[I], ArgVals[I], I + 1);
2467 for (unsigned I = 0, E = Args.size(); I != E; ++I)
2468 EmitParmDecl(*Args[I], ArgVals[I], I + 1);
2472 static void eraseUnusedBitCasts(llvm::Instruction *insn) {
2473 while (insn->use_empty()) {
2474 llvm::BitCastInst *bitcast = dyn_cast<llvm::BitCastInst>(insn);
2475 if (!bitcast) return;
2477 // This is "safe" because we would have used a ConstantExpr otherwise.
2478 insn = cast<llvm::Instruction>(bitcast->getOperand(0));
2479 bitcast->eraseFromParent();
2483 /// Try to emit a fused autorelease of a return result.
2484 static llvm::Value *tryEmitFusedAutoreleaseOfResult(CodeGenFunction &CGF,
2485 llvm::Value *result) {
2486 // We must be immediately followed the cast.
2487 llvm::BasicBlock *BB = CGF.Builder.GetInsertBlock();
2488 if (BB->empty()) return nullptr;
2489 if (&BB->back() != result) return nullptr;
2491 llvm::Type *resultType = result->getType();
2493 // result is in a BasicBlock and is therefore an Instruction.
2494 llvm::Instruction *generator = cast<llvm::Instruction>(result);
2496 SmallVector<llvm::Instruction *, 4> InstsToKill;
2499 // %generator = bitcast %type1* %generator2 to %type2*
2500 while (llvm::BitCastInst *bitcast = dyn_cast<llvm::BitCastInst>(generator)) {
2501 // We would have emitted this as a constant if the operand weren't
2503 generator = cast<llvm::Instruction>(bitcast->getOperand(0));
2505 // Require the generator to be immediately followed by the cast.
2506 if (generator->getNextNode() != bitcast)
2509 InstsToKill.push_back(bitcast);
2513 // %generator = call i8* @objc_retain(i8* %originalResult)
2515 // %generator = call i8* @objc_retainAutoreleasedReturnValue(i8* %originalResult)
2516 llvm::CallInst *call = dyn_cast<llvm::CallInst>(generator);
2517 if (!call) return nullptr;
2519 bool doRetainAutorelease;
2521 if (call->getCalledValue() == CGF.CGM.getObjCEntrypoints().objc_retain) {
2522 doRetainAutorelease = true;
2523 } else if (call->getCalledValue() == CGF.CGM.getObjCEntrypoints()
2524 .objc_retainAutoreleasedReturnValue) {
2525 doRetainAutorelease = false;
2527 // If we emitted an assembly marker for this call (and the
2528 // ARCEntrypoints field should have been set if so), go looking
2529 // for that call. If we can't find it, we can't do this
2530 // optimization. But it should always be the immediately previous
2531 // instruction, unless we needed bitcasts around the call.
2532 if (CGF.CGM.getObjCEntrypoints().retainAutoreleasedReturnValueMarker) {
2533 llvm::Instruction *prev = call->getPrevNode();
2535 if (isa<llvm::BitCastInst>(prev)) {
2536 prev = prev->getPrevNode();
2539 assert(isa<llvm::CallInst>(prev));
2540 assert(cast<llvm::CallInst>(prev)->getCalledValue() ==
2541 CGF.CGM.getObjCEntrypoints().retainAutoreleasedReturnValueMarker);
2542 InstsToKill.push_back(prev);
2548 result = call->getArgOperand(0);
2549 InstsToKill.push_back(call);
2551 // Keep killing bitcasts, for sanity. Note that we no longer care
2552 // about precise ordering as long as there's exactly one use.
2553 while (llvm::BitCastInst *bitcast = dyn_cast<llvm::BitCastInst>(result)) {
2554 if (!bitcast->hasOneUse()) break;
2555 InstsToKill.push_back(bitcast);
2556 result = bitcast->getOperand(0);
2559 // Delete all the unnecessary instructions, from latest to earliest.
2560 for (auto *I : InstsToKill)
2561 I->eraseFromParent();
2563 // Do the fused retain/autorelease if we were asked to.
2564 if (doRetainAutorelease)
2565 result = CGF.EmitARCRetainAutoreleaseReturnValue(result);
2567 // Cast back to the result type.
2568 return CGF.Builder.CreateBitCast(result, resultType);
2571 /// If this is a +1 of the value of an immutable 'self', remove it.
2572 static llvm::Value *tryRemoveRetainOfSelf(CodeGenFunction &CGF,
2573 llvm::Value *result) {
2574 // This is only applicable to a method with an immutable 'self'.
2575 const ObjCMethodDecl *method =
2576 dyn_cast_or_null<ObjCMethodDecl>(CGF.CurCodeDecl);
2577 if (!method) return nullptr;
2578 const VarDecl *self = method->getSelfDecl();
2579 if (!self->getType().isConstQualified()) return nullptr;
2581 // Look for a retain call.
2582 llvm::CallInst *retainCall =
2583 dyn_cast<llvm::CallInst>(result->stripPointerCasts());
2585 retainCall->getCalledValue() != CGF.CGM.getObjCEntrypoints().objc_retain)
2588 // Look for an ordinary load of 'self'.
2589 llvm::Value *retainedValue = retainCall->getArgOperand(0);
2590 llvm::LoadInst *load =
2591 dyn_cast<llvm::LoadInst>(retainedValue->stripPointerCasts());
2592 if (!load || load->isAtomic() || load->isVolatile() ||
2593 load->getPointerOperand() != CGF.GetAddrOfLocalVar(self).getPointer())
2596 // Okay! Burn it all down. This relies for correctness on the
2597 // assumption that the retain is emitted as part of the return and
2598 // that thereafter everything is used "linearly".
2599 llvm::Type *resultType = result->getType();
2600 eraseUnusedBitCasts(cast<llvm::Instruction>(result));
2601 assert(retainCall->use_empty());
2602 retainCall->eraseFromParent();
2603 eraseUnusedBitCasts(cast<llvm::Instruction>(retainedValue));
2605 return CGF.Builder.CreateBitCast(load, resultType);
2608 /// Emit an ARC autorelease of the result of a function.
2610 /// \return the value to actually return from the function
2611 static llvm::Value *emitAutoreleaseOfResult(CodeGenFunction &CGF,
2612 llvm::Value *result) {
2613 // If we're returning 'self', kill the initial retain. This is a
2614 // heuristic attempt to "encourage correctness" in the really unfortunate
2615 // case where we have a return of self during a dealloc and we desperately
2616 // need to avoid the possible autorelease.
2617 if (llvm::Value *self = tryRemoveRetainOfSelf(CGF, result))
2620 // At -O0, try to emit a fused retain/autorelease.
2621 if (CGF.shouldUseFusedARCCalls())
2622 if (llvm::Value *fused = tryEmitFusedAutoreleaseOfResult(CGF, result))
2625 return CGF.EmitARCAutoreleaseReturnValue(result);
2628 /// Heuristically search for a dominating store to the return-value slot.
2629 static llvm::StoreInst *findDominatingStoreToReturnValue(CodeGenFunction &CGF) {
2630 // Check if a User is a store which pointerOperand is the ReturnValue.
2631 // We are looking for stores to the ReturnValue, not for stores of the
2632 // ReturnValue to some other location.
2633 auto GetStoreIfValid = [&CGF](llvm::User *U) -> llvm::StoreInst * {
2634 auto *SI = dyn_cast<llvm::StoreInst>(U);
2635 if (!SI || SI->getPointerOperand() != CGF.ReturnValue.getPointer())
2637 // These aren't actually possible for non-coerced returns, and we
2638 // only care about non-coerced returns on this code path.
2639 assert(!SI->isAtomic() && !SI->isVolatile());
2642 // If there are multiple uses of the return-value slot, just check
2643 // for something immediately preceding the IP. Sometimes this can
2644 // happen with how we generate implicit-returns; it can also happen
2645 // with noreturn cleanups.
2646 if (!CGF.ReturnValue.getPointer()->hasOneUse()) {
2647 llvm::BasicBlock *IP = CGF.Builder.GetInsertBlock();
2648 if (IP->empty()) return nullptr;
2649 llvm::Instruction *I = &IP->back();
2651 // Skip lifetime markers
2652 for (llvm::BasicBlock::reverse_iterator II = IP->rbegin(),
2655 if (llvm::IntrinsicInst *Intrinsic =
2656 dyn_cast<llvm::IntrinsicInst>(&*II)) {
2657 if (Intrinsic->getIntrinsicID() == llvm::Intrinsic::lifetime_end) {
2658 const llvm::Value *CastAddr = Intrinsic->getArgOperand(1);
2662 if (isa<llvm::BitCastInst>(&*II) && (CastAddr == &*II))
2670 return GetStoreIfValid(I);
2673 llvm::StoreInst *store =
2674 GetStoreIfValid(CGF.ReturnValue.getPointer()->user_back());
2675 if (!store) return nullptr;
2677 // Now do a first-and-dirty dominance check: just walk up the
2678 // single-predecessors chain from the current insertion point.
2679 llvm::BasicBlock *StoreBB = store->getParent();
2680 llvm::BasicBlock *IP = CGF.Builder.GetInsertBlock();
2681 while (IP != StoreBB) {
2682 if (!(IP = IP->getSinglePredecessor()))
2686 // Okay, the store's basic block dominates the insertion point; we
2687 // can do our thing.
2691 void CodeGenFunction::EmitFunctionEpilog(const CGFunctionInfo &FI,
2693 SourceLocation EndLoc) {
2694 if (CurCodeDecl && CurCodeDecl->hasAttr<NakedAttr>()) {
2695 // Naked functions don't have epilogues.
2696 Builder.CreateUnreachable();
2700 // Functions with no result always return void.
2701 if (!ReturnValue.isValid()) {
2702 Builder.CreateRetVoid();
2706 llvm::DebugLoc RetDbgLoc;
2707 llvm::Value *RV = nullptr;
2708 QualType RetTy = FI.getReturnType();
2709 const ABIArgInfo &RetAI = FI.getReturnInfo();
2711 switch (RetAI.getKind()) {
2712 case ABIArgInfo::InAlloca:
2713 // Aggregrates get evaluated directly into the destination. Sometimes we
2714 // need to return the sret value in a register, though.
2715 assert(hasAggregateEvaluationKind(RetTy));
2716 if (RetAI.getInAllocaSRet()) {
2717 llvm::Function::arg_iterator EI = CurFn->arg_end();
2719 llvm::Value *ArgStruct = &*EI;
2720 llvm::Value *SRet = Builder.CreateStructGEP(
2721 nullptr, ArgStruct, RetAI.getInAllocaFieldIndex());
2722 RV = Builder.CreateAlignedLoad(SRet, getPointerAlign(), "sret");
2726 case ABIArgInfo::Indirect: {
2727 auto AI = CurFn->arg_begin();
2728 if (RetAI.isSRetAfterThis())
2730 switch (getEvaluationKind(RetTy)) {
2733 EmitLoadOfComplex(MakeAddrLValue(ReturnValue, RetTy), EndLoc);
2734 EmitStoreOfComplex(RT, MakeNaturalAlignAddrLValue(&*AI, RetTy),
2739 // Do nothing; aggregrates get evaluated directly into the destination.
2742 EmitStoreOfScalar(Builder.CreateLoad(ReturnValue),
2743 MakeNaturalAlignAddrLValue(&*AI, RetTy),
2750 case ABIArgInfo::Extend:
2751 case ABIArgInfo::Direct:
2752 if (RetAI.getCoerceToType() == ConvertType(RetTy) &&
2753 RetAI.getDirectOffset() == 0) {
2754 // The internal return value temp always will have pointer-to-return-type
2755 // type, just do a load.
2757 // If there is a dominating store to ReturnValue, we can elide
2758 // the load, zap the store, and usually zap the alloca.
2759 if (llvm::StoreInst *SI =
2760 findDominatingStoreToReturnValue(*this)) {
2761 // Reuse the debug location from the store unless there is
2762 // cleanup code to be emitted between the store and return
2764 if (EmitRetDbgLoc && !AutoreleaseResult)
2765 RetDbgLoc = SI->getDebugLoc();
2766 // Get the stored value and nuke the now-dead store.
2767 RV = SI->getValueOperand();
2768 SI->eraseFromParent();
2770 // If that was the only use of the return value, nuke it as well now.
2771 auto returnValueInst = ReturnValue.getPointer();
2772 if (returnValueInst->use_empty()) {
2773 if (auto alloca = dyn_cast<llvm::AllocaInst>(returnValueInst)) {
2774 alloca->eraseFromParent();
2775 ReturnValue = Address::invalid();
2779 // Otherwise, we have to do a simple load.
2781 RV = Builder.CreateLoad(ReturnValue);
2784 // If the value is offset in memory, apply the offset now.
2785 Address V = emitAddressAtOffset(*this, ReturnValue, RetAI);
2787 RV = CreateCoercedLoad(V, RetAI.getCoerceToType(), *this);
2790 // In ARC, end functions that return a retainable type with a call
2791 // to objc_autoreleaseReturnValue.
2792 if (AutoreleaseResult) {
2794 // Type::isObjCRetainabletype has to be called on a QualType that hasn't
2795 // been stripped of the typedefs, so we cannot use RetTy here. Get the
2796 // original return type of FunctionDecl, CurCodeDecl, and BlockDecl from
2797 // CurCodeDecl or BlockInfo.
2800 if (auto *FD = dyn_cast<FunctionDecl>(CurCodeDecl))
2801 RT = FD->getReturnType();
2802 else if (auto *MD = dyn_cast<ObjCMethodDecl>(CurCodeDecl))
2803 RT = MD->getReturnType();
2804 else if (isa<BlockDecl>(CurCodeDecl))
2805 RT = BlockInfo->BlockExpression->getFunctionType()->getReturnType();
2807 llvm_unreachable("Unexpected function/method type");
2809 assert(getLangOpts().ObjCAutoRefCount &&
2810 !FI.isReturnsRetained() &&
2811 RT->isObjCRetainableType());
2813 RV = emitAutoreleaseOfResult(*this, RV);
2818 case ABIArgInfo::Ignore:
2821 case ABIArgInfo::CoerceAndExpand: {
2822 auto coercionType = RetAI.getCoerceAndExpandType();
2823 auto layout = CGM.getDataLayout().getStructLayout(coercionType);
2825 // Load all of the coerced elements out into results.
2826 llvm::SmallVector<llvm::Value*, 4> results;
2827 Address addr = Builder.CreateElementBitCast(ReturnValue, coercionType);
2828 for (unsigned i = 0, e = coercionType->getNumElements(); i != e; ++i) {
2829 auto coercedEltType = coercionType->getElementType(i);
2830 if (ABIArgInfo::isPaddingForCoerceAndExpand(coercedEltType))
2833 auto eltAddr = Builder.CreateStructGEP(addr, i, layout);
2834 auto elt = Builder.CreateLoad(eltAddr);
2835 results.push_back(elt);
2838 // If we have one result, it's the single direct result type.
2839 if (results.size() == 1) {
2842 // Otherwise, we need to make a first-class aggregate.
2844 // Construct a return type that lacks padding elements.
2845 llvm::Type *returnType = RetAI.getUnpaddedCoerceAndExpandType();
2847 RV = llvm::UndefValue::get(returnType);
2848 for (unsigned i = 0, e = results.size(); i != e; ++i) {
2849 RV = Builder.CreateInsertValue(RV, results[i], i);
2855 case ABIArgInfo::Expand:
2856 llvm_unreachable("Invalid ABI kind for return argument");
2859 llvm::Instruction *Ret;
2861 if (CurCodeDecl && SanOpts.has(SanitizerKind::ReturnsNonnullAttribute)) {
2862 if (auto RetNNAttr = CurCodeDecl->getAttr<ReturnsNonNullAttr>()) {
2863 SanitizerScope SanScope(this);
2864 llvm::Value *Cond = Builder.CreateICmpNE(
2865 RV, llvm::Constant::getNullValue(RV->getType()));
2866 llvm::Constant *StaticData[] = {
2867 EmitCheckSourceLocation(EndLoc),
2868 EmitCheckSourceLocation(RetNNAttr->getLocation()),
2870 EmitCheck(std::make_pair(Cond, SanitizerKind::ReturnsNonnullAttribute),
2871 SanitizerHandler::NonnullReturn, StaticData, None);
2874 Ret = Builder.CreateRet(RV);
2876 Ret = Builder.CreateRetVoid();
2880 Ret->setDebugLoc(std::move(RetDbgLoc));
2883 static bool isInAllocaArgument(CGCXXABI &ABI, QualType type) {
2884 const CXXRecordDecl *RD = type->getAsCXXRecordDecl();
2885 return RD && ABI.getRecordArgABI(RD) == CGCXXABI::RAA_DirectInMemory;
2888 static AggValueSlot createPlaceholderSlot(CodeGenFunction &CGF,
2890 // FIXME: Generate IR in one pass, rather than going back and fixing up these
2892 llvm::Type *IRTy = CGF.ConvertTypeForMem(Ty);
2893 llvm::Type *IRPtrTy = IRTy->getPointerTo();
2894 llvm::Value *Placeholder = llvm::UndefValue::get(IRPtrTy->getPointerTo());
2896 // FIXME: When we generate this IR in one pass, we shouldn't need
2897 // this win32-specific alignment hack.
2898 CharUnits Align = CharUnits::fromQuantity(4);
2899 Placeholder = CGF.Builder.CreateAlignedLoad(IRPtrTy, Placeholder, Align);
2901 return AggValueSlot::forAddr(Address(Placeholder, Align),
2903 AggValueSlot::IsNotDestructed,
2904 AggValueSlot::DoesNotNeedGCBarriers,
2905 AggValueSlot::IsNotAliased);
2908 void CodeGenFunction::EmitDelegateCallArg(CallArgList &args,
2909 const VarDecl *param,
2910 SourceLocation loc) {
2911 // StartFunction converted the ABI-lowered parameter(s) into a
2912 // local alloca. We need to turn that into an r-value suitable
2914 Address local = GetAddrOfLocalVar(param);
2916 QualType type = param->getType();
2918 assert(!isInAllocaArgument(CGM.getCXXABI(), type) &&
2919 "cannot emit delegate call arguments for inalloca arguments!");
2921 // GetAddrOfLocalVar returns a pointer-to-pointer for references,
2922 // but the argument needs to be the original pointer.
2923 if (type->isReferenceType()) {
2924 args.add(RValue::get(Builder.CreateLoad(local)), type);
2926 // In ARC, move out of consumed arguments so that the release cleanup
2927 // entered by StartFunction doesn't cause an over-release. This isn't
2928 // optimal -O0 code generation, but it should get cleaned up when
2929 // optimization is enabled. This also assumes that delegate calls are
2930 // performed exactly once for a set of arguments, but that should be safe.
2931 } else if (getLangOpts().ObjCAutoRefCount &&
2932 param->hasAttr<NSConsumedAttr>() &&
2933 type->isObjCRetainableType()) {
2934 llvm::Value *ptr = Builder.CreateLoad(local);
2936 llvm::ConstantPointerNull::get(cast<llvm::PointerType>(ptr->getType()));
2937 Builder.CreateStore(null, local);
2938 args.add(RValue::get(ptr), type);
2940 // For the most part, we just need to load the alloca, except that
2941 // aggregate r-values are actually pointers to temporaries.
2943 args.add(convertTempToRValue(local, type, loc), type);
2947 static bool isProvablyNull(llvm::Value *addr) {
2948 return isa<llvm::ConstantPointerNull>(addr);
2951 /// Emit the actual writing-back of a writeback.
2952 static void emitWriteback(CodeGenFunction &CGF,
2953 const CallArgList::Writeback &writeback) {
2954 const LValue &srcLV = writeback.Source;
2955 Address srcAddr = srcLV.getAddress();
2956 assert(!isProvablyNull(srcAddr.getPointer()) &&
2957 "shouldn't have writeback for provably null argument");
2959 llvm::BasicBlock *contBB = nullptr;
2961 // If the argument wasn't provably non-null, we need to null check
2962 // before doing the store.
2963 bool provablyNonNull = llvm::isKnownNonNull(srcAddr.getPointer());
2964 if (!provablyNonNull) {
2965 llvm::BasicBlock *writebackBB = CGF.createBasicBlock("icr.writeback");
2966 contBB = CGF.createBasicBlock("icr.done");
2968 llvm::Value *isNull =
2969 CGF.Builder.CreateIsNull(srcAddr.getPointer(), "icr.isnull");
2970 CGF.Builder.CreateCondBr(isNull, contBB, writebackBB);
2971 CGF.EmitBlock(writebackBB);
2974 // Load the value to writeback.
2975 llvm::Value *value = CGF.Builder.CreateLoad(writeback.Temporary);
2977 // Cast it back, in case we're writing an id to a Foo* or something.
2978 value = CGF.Builder.CreateBitCast(value, srcAddr.getElementType(),
2979 "icr.writeback-cast");
2981 // Perform the writeback.
2983 // If we have a "to use" value, it's something we need to emit a use
2984 // of. This has to be carefully threaded in: if it's done after the
2985 // release it's potentially undefined behavior (and the optimizer
2986 // will ignore it), and if it happens before the retain then the
2987 // optimizer could move the release there.
2988 if (writeback.ToUse) {
2989 assert(srcLV.getObjCLifetime() == Qualifiers::OCL_Strong);
2991 // Retain the new value. No need to block-copy here: the block's
2992 // being passed up the stack.
2993 value = CGF.EmitARCRetainNonBlock(value);
2995 // Emit the intrinsic use here.
2996 CGF.EmitARCIntrinsicUse(writeback.ToUse);
2998 // Load the old value (primitively).
2999 llvm::Value *oldValue = CGF.EmitLoadOfScalar(srcLV, SourceLocation());
3001 // Put the new value in place (primitively).
3002 CGF.EmitStoreOfScalar(value, srcLV, /*init*/ false);
3004 // Release the old value.
3005 CGF.EmitARCRelease(oldValue, srcLV.isARCPreciseLifetime());
3007 // Otherwise, we can just do a normal lvalue store.
3009 CGF.EmitStoreThroughLValue(RValue::get(value), srcLV);
3012 // Jump to the continuation block.
3013 if (!provablyNonNull)
3014 CGF.EmitBlock(contBB);
3017 static void emitWritebacks(CodeGenFunction &CGF,
3018 const CallArgList &args) {
3019 for (const auto &I : args.writebacks())
3020 emitWriteback(CGF, I);
3023 static void deactivateArgCleanupsBeforeCall(CodeGenFunction &CGF,
3024 const CallArgList &CallArgs) {
3025 assert(CGF.getTarget().getCXXABI().areArgsDestroyedLeftToRightInCallee());
3026 ArrayRef<CallArgList::CallArgCleanup> Cleanups =
3027 CallArgs.getCleanupsToDeactivate();
3028 // Iterate in reverse to increase the likelihood of popping the cleanup.
3029 for (const auto &I : llvm::reverse(Cleanups)) {
3030 CGF.DeactivateCleanupBlock(I.Cleanup, I.IsActiveIP);
3031 I.IsActiveIP->eraseFromParent();
3035 static const Expr *maybeGetUnaryAddrOfOperand(const Expr *E) {
3036 if (const UnaryOperator *uop = dyn_cast<UnaryOperator>(E->IgnoreParens()))
3037 if (uop->getOpcode() == UO_AddrOf)
3038 return uop->getSubExpr();
3042 /// Emit an argument that's being passed call-by-writeback. That is,
3043 /// we are passing the address of an __autoreleased temporary; it
3044 /// might be copy-initialized with the current value of the given
3045 /// address, but it will definitely be copied out of after the call.
3046 static void emitWritebackArg(CodeGenFunction &CGF, CallArgList &args,
3047 const ObjCIndirectCopyRestoreExpr *CRE) {
3050 // Make an optimistic effort to emit the address as an l-value.
3051 // This can fail if the argument expression is more complicated.
3052 if (const Expr *lvExpr = maybeGetUnaryAddrOfOperand(CRE->getSubExpr())) {
3053 srcLV = CGF.EmitLValue(lvExpr);
3055 // Otherwise, just emit it as a scalar.
3057 Address srcAddr = CGF.EmitPointerWithAlignment(CRE->getSubExpr());
3059 QualType srcAddrType =
3060 CRE->getSubExpr()->getType()->castAs<PointerType>()->getPointeeType();
3061 srcLV = CGF.MakeAddrLValue(srcAddr, srcAddrType);
3063 Address srcAddr = srcLV.getAddress();
3065 // The dest and src types don't necessarily match in LLVM terms
3066 // because of the crazy ObjC compatibility rules.
3068 llvm::PointerType *destType =
3069 cast<llvm::PointerType>(CGF.ConvertType(CRE->getType()));
3071 // If the address is a constant null, just pass the appropriate null.
3072 if (isProvablyNull(srcAddr.getPointer())) {
3073 args.add(RValue::get(llvm::ConstantPointerNull::get(destType)),
3078 // Create the temporary.
3079 Address temp = CGF.CreateTempAlloca(destType->getElementType(),
3080 CGF.getPointerAlign(),
3082 // Loading an l-value can introduce a cleanup if the l-value is __weak,
3083 // and that cleanup will be conditional if we can't prove that the l-value
3084 // isn't null, so we need to register a dominating point so that the cleanups
3085 // system will make valid IR.
3086 CodeGenFunction::ConditionalEvaluation condEval(CGF);
3088 // Zero-initialize it if we're not doing a copy-initialization.
3089 bool shouldCopy = CRE->shouldCopy();
3092 llvm::ConstantPointerNull::get(
3093 cast<llvm::PointerType>(destType->getElementType()));
3094 CGF.Builder.CreateStore(null, temp);
3097 llvm::BasicBlock *contBB = nullptr;
3098 llvm::BasicBlock *originBB = nullptr;
3100 // If the address is *not* known to be non-null, we need to switch.
3101 llvm::Value *finalArgument;
3103 bool provablyNonNull = llvm::isKnownNonNull(srcAddr.getPointer());
3104 if (provablyNonNull) {
3105 finalArgument = temp.getPointer();
3107 llvm::Value *isNull =
3108 CGF.Builder.CreateIsNull(srcAddr.getPointer(), "icr.isnull");
3110 finalArgument = CGF.Builder.CreateSelect(isNull,
3111 llvm::ConstantPointerNull::get(destType),
3112 temp.getPointer(), "icr.argument");
3114 // If we need to copy, then the load has to be conditional, which
3115 // means we need control flow.
3117 originBB = CGF.Builder.GetInsertBlock();
3118 contBB = CGF.createBasicBlock("icr.cont");
3119 llvm::BasicBlock *copyBB = CGF.createBasicBlock("icr.copy");
3120 CGF.Builder.CreateCondBr(isNull, contBB, copyBB);
3121 CGF.EmitBlock(copyBB);
3122 condEval.begin(CGF);
3126 llvm::Value *valueToUse = nullptr;
3128 // Perform a copy if necessary.
3130 RValue srcRV = CGF.EmitLoadOfLValue(srcLV, SourceLocation());
3131 assert(srcRV.isScalar());
3133 llvm::Value *src = srcRV.getScalarVal();
3134 src = CGF.Builder.CreateBitCast(src, destType->getElementType(),
3137 // Use an ordinary store, not a store-to-lvalue.
3138 CGF.Builder.CreateStore(src, temp);
3140 // If optimization is enabled, and the value was held in a
3141 // __strong variable, we need to tell the optimizer that this
3142 // value has to stay alive until we're doing the store back.
3143 // This is because the temporary is effectively unretained,
3144 // and so otherwise we can violate the high-level semantics.
3145 if (CGF.CGM.getCodeGenOpts().OptimizationLevel != 0 &&
3146 srcLV.getObjCLifetime() == Qualifiers::OCL_Strong) {
3151 // Finish the control flow if we needed it.
3152 if (shouldCopy && !provablyNonNull) {
3153 llvm::BasicBlock *copyBB = CGF.Builder.GetInsertBlock();
3154 CGF.EmitBlock(contBB);
3156 // Make a phi for the value to intrinsically use.
3158 llvm::PHINode *phiToUse = CGF.Builder.CreatePHI(valueToUse->getType(), 2,
3160 phiToUse->addIncoming(valueToUse, copyBB);
3161 phiToUse->addIncoming(llvm::UndefValue::get(valueToUse->getType()),
3163 valueToUse = phiToUse;
3169 args.addWriteback(srcLV, temp, valueToUse);
3170 args.add(RValue::get(finalArgument), CRE->getType());
3173 void CallArgList::allocateArgumentMemory(CodeGenFunction &CGF) {
3177 llvm::Function *F = CGF.CGM.getIntrinsic(llvm::Intrinsic::stacksave);
3178 StackBase = CGF.Builder.CreateCall(F, {}, "inalloca.save");
3181 void CallArgList::freeArgumentMemory(CodeGenFunction &CGF) const {
3183 // Restore the stack after the call.
3184 llvm::Value *F = CGF.CGM.getIntrinsic(llvm::Intrinsic::stackrestore);
3185 CGF.Builder.CreateCall(F, StackBase);
3189 void CodeGenFunction::EmitNonNullArgCheck(RValue RV, QualType ArgType,
3190 SourceLocation ArgLoc,
3191 const FunctionDecl *FD,
3193 if (!SanOpts.has(SanitizerKind::NonnullAttribute) || !FD)
3195 auto PVD = ParmNum < FD->getNumParams() ? FD->getParamDecl(ParmNum) : nullptr;
3196 unsigned ArgNo = PVD ? PVD->getFunctionScopeIndex() : ParmNum;
3197 auto NNAttr = getNonNullAttr(FD, PVD, ArgType, ArgNo);
3200 SanitizerScope SanScope(this);
3201 assert(RV.isScalar());
3202 llvm::Value *V = RV.getScalarVal();
3204 Builder.CreateICmpNE(V, llvm::Constant::getNullValue(V->getType()));
3205 llvm::Constant *StaticData[] = {
3206 EmitCheckSourceLocation(ArgLoc),
3207 EmitCheckSourceLocation(NNAttr->getLocation()),
3208 llvm::ConstantInt::get(Int32Ty, ArgNo + 1),
3210 EmitCheck(std::make_pair(Cond, SanitizerKind::NonnullAttribute),
3211 SanitizerHandler::NonnullArg, StaticData, None);
3214 void CodeGenFunction::EmitCallArgs(
3215 CallArgList &Args, ArrayRef<QualType> ArgTypes,
3216 llvm::iterator_range<CallExpr::const_arg_iterator> ArgRange,
3217 const FunctionDecl *CalleeDecl, unsigned ParamsToSkip,
3218 EvaluationOrder Order) {
3219 assert((int)ArgTypes.size() == (ArgRange.end() - ArgRange.begin()));
3221 auto MaybeEmitImplicitObjectSize = [&](unsigned I, const Expr *Arg) {
3222 if (CalleeDecl == nullptr || I >= CalleeDecl->getNumParams())
3224 auto *PS = CalleeDecl->getParamDecl(I)->getAttr<PassObjectSizeAttr>();
3228 const auto &Context = getContext();
3229 auto SizeTy = Context.getSizeType();
3230 auto T = Builder.getIntNTy(Context.getTypeSize(SizeTy));
3231 llvm::Value *V = evaluateOrEmitBuiltinObjectSize(Arg, PS->getType(), T);
3232 Args.add(RValue::get(V), SizeTy);
3235 // We *have* to evaluate arguments from right to left in the MS C++ ABI,
3236 // because arguments are destroyed left to right in the callee. As a special
3237 // case, there are certain language constructs that require left-to-right
3238 // evaluation, and in those cases we consider the evaluation order requirement
3239 // to trump the "destruction order is reverse construction order" guarantee.
3241 CGM.getTarget().getCXXABI().areArgsDestroyedLeftToRightInCallee()
3242 ? Order == EvaluationOrder::ForceLeftToRight
3243 : Order != EvaluationOrder::ForceRightToLeft;
3245 // Insert a stack save if we're going to need any inalloca args.
3246 bool HasInAllocaArgs = false;
3247 if (CGM.getTarget().getCXXABI().isMicrosoft()) {
3248 for (ArrayRef<QualType>::iterator I = ArgTypes.begin(), E = ArgTypes.end();
3249 I != E && !HasInAllocaArgs; ++I)
3250 HasInAllocaArgs = isInAllocaArgument(CGM.getCXXABI(), *I);
3251 if (HasInAllocaArgs) {
3252 assert(getTarget().getTriple().getArch() == llvm::Triple::x86);
3253 Args.allocateArgumentMemory(*this);
3257 // Evaluate each argument in the appropriate order.
3258 size_t CallArgsStart = Args.size();
3259 for (unsigned I = 0, E = ArgTypes.size(); I != E; ++I) {
3260 unsigned Idx = LeftToRight ? I : E - I - 1;
3261 CallExpr::const_arg_iterator Arg = ArgRange.begin() + Idx;
3262 if (!LeftToRight) MaybeEmitImplicitObjectSize(Idx, *Arg);
3263 EmitCallArg(Args, *Arg, ArgTypes[Idx]);
3264 EmitNonNullArgCheck(Args.back().RV, ArgTypes[Idx], (*Arg)->getExprLoc(),
3265 CalleeDecl, ParamsToSkip + Idx);
3266 if (LeftToRight) MaybeEmitImplicitObjectSize(Idx, *Arg);
3270 // Un-reverse the arguments we just evaluated so they match up with the LLVM
3272 std::reverse(Args.begin() + CallArgsStart, Args.end());
3278 struct DestroyUnpassedArg final : EHScopeStack::Cleanup {
3279 DestroyUnpassedArg(Address Addr, QualType Ty)
3280 : Addr(Addr), Ty(Ty) {}
3285 void Emit(CodeGenFunction &CGF, Flags flags) override {
3286 const CXXDestructorDecl *Dtor = Ty->getAsCXXRecordDecl()->getDestructor();
3287 assert(!Dtor->isTrivial());
3288 CGF.EmitCXXDestructorCall(Dtor, Dtor_Complete, /*for vbase*/ false,
3289 /*Delegating=*/false, Addr);
3293 struct DisableDebugLocationUpdates {
3294 CodeGenFunction &CGF;
3295 bool disabledDebugInfo;
3296 DisableDebugLocationUpdates(CodeGenFunction &CGF, const Expr *E) : CGF(CGF) {
3297 if ((disabledDebugInfo = isa<CXXDefaultArgExpr>(E) && CGF.getDebugInfo()))
3298 CGF.disableDebugInfo();
3300 ~DisableDebugLocationUpdates() {
3301 if (disabledDebugInfo)
3302 CGF.enableDebugInfo();
3306 } // end anonymous namespace
3308 void CodeGenFunction::EmitCallArg(CallArgList &args, const Expr *E,
3310 DisableDebugLocationUpdates Dis(*this, E);
3311 if (const ObjCIndirectCopyRestoreExpr *CRE
3312 = dyn_cast<ObjCIndirectCopyRestoreExpr>(E)) {
3313 assert(getLangOpts().ObjCAutoRefCount);
3314 assert(getContext().hasSameUnqualifiedType(E->getType(), type));
3315 return emitWritebackArg(*this, args, CRE);
3318 assert(type->isReferenceType() == E->isGLValue() &&
3319 "reference binding to unmaterialized r-value!");
3321 if (E->isGLValue()) {
3322 assert(E->getObjectKind() == OK_Ordinary);
3323 return args.add(EmitReferenceBindingToExpr(E), type);
3326 bool HasAggregateEvalKind = hasAggregateEvaluationKind(type);
3328 // In the Microsoft C++ ABI, aggregate arguments are destructed by the callee.
3329 // However, we still have to push an EH-only cleanup in case we unwind before
3330 // we make it to the call.
3331 if (HasAggregateEvalKind &&
3332 CGM.getTarget().getCXXABI().areArgsDestroyedLeftToRightInCallee()) {
3333 // If we're using inalloca, use the argument memory. Otherwise, use a
3336 if (args.isUsingInAlloca())
3337 Slot = createPlaceholderSlot(*this, type);
3339 Slot = CreateAggTemp(type, "agg.tmp");
3341 const CXXRecordDecl *RD = type->getAsCXXRecordDecl();
3342 bool DestroyedInCallee =
3343 RD && RD->hasNonTrivialDestructor() &&
3344 CGM.getCXXABI().getRecordArgABI(RD) != CGCXXABI::RAA_Default;
3345 if (DestroyedInCallee)
3346 Slot.setExternallyDestructed();
3348 EmitAggExpr(E, Slot);
3349 RValue RV = Slot.asRValue();
3352 if (DestroyedInCallee) {
3353 // Create a no-op GEP between the placeholder and the cleanup so we can
3354 // RAUW it successfully. It also serves as a marker of the first
3355 // instruction where the cleanup is active.
3356 pushFullExprCleanup<DestroyUnpassedArg>(EHCleanup, Slot.getAddress(),
3358 // This unreachable is a temporary marker which will be removed later.
3359 llvm::Instruction *IsActive = Builder.CreateUnreachable();
3360 args.addArgCleanupDeactivation(EHStack.getInnermostEHScope(), IsActive);
3365 if (HasAggregateEvalKind && isa<ImplicitCastExpr>(E) &&
3366 cast<CastExpr>(E)->getCastKind() == CK_LValueToRValue) {
3367 LValue L = EmitLValue(cast<CastExpr>(E)->getSubExpr());
3368 assert(L.isSimple());
3369 if (L.getAlignment() >= getContext().getTypeAlignInChars(type)) {
3370 args.add(L.asAggregateRValue(), type, /*NeedsCopy*/true);
3372 // We can't represent a misaligned lvalue in the CallArgList, so copy
3373 // to an aligned temporary now.
3374 Address tmp = CreateMemTemp(type);
3375 EmitAggregateCopy(tmp, L.getAddress(), type, L.isVolatile());
3376 args.add(RValue::getAggregate(tmp), type);
3381 args.add(EmitAnyExprToTemp(E), type);
3384 QualType CodeGenFunction::getVarArgType(const Expr *Arg) {
3385 // System headers on Windows define NULL to 0 instead of 0LL on Win64. MSVC
3386 // implicitly widens null pointer constants that are arguments to varargs
3387 // functions to pointer-sized ints.
3388 if (!getTarget().getTriple().isOSWindows())
3389 return Arg->getType();
3391 if (Arg->getType()->isIntegerType() &&
3392 getContext().getTypeSize(Arg->getType()) <
3393 getContext().getTargetInfo().getPointerWidth(0) &&
3394 Arg->isNullPointerConstant(getContext(),
3395 Expr::NPC_ValueDependentIsNotNull)) {
3396 return getContext().getIntPtrType();
3399 return Arg->getType();
3402 // In ObjC ARC mode with no ObjC ARC exception safety, tell the ARC
3403 // optimizer it can aggressively ignore unwind edges.
3405 CodeGenFunction::AddObjCARCExceptionMetadata(llvm::Instruction *Inst) {
3406 if (CGM.getCodeGenOpts().OptimizationLevel != 0 &&
3407 !CGM.getCodeGenOpts().ObjCAutoRefCountExceptions)
3408 Inst->setMetadata("clang.arc.no_objc_arc_exceptions",
3409 CGM.getNoObjCARCExceptionsMetadata());
3412 /// Emits a call to the given no-arguments nounwind runtime function.
3414 CodeGenFunction::EmitNounwindRuntimeCall(llvm::Value *callee,
3415 const llvm::Twine &name) {
3416 return EmitNounwindRuntimeCall(callee, None, name);
3419 /// Emits a call to the given nounwind runtime function.
3421 CodeGenFunction::EmitNounwindRuntimeCall(llvm::Value *callee,
3422 ArrayRef<llvm::Value*> args,
3423 const llvm::Twine &name) {
3424 llvm::CallInst *call = EmitRuntimeCall(callee, args, name);
3425 call->setDoesNotThrow();
3429 /// Emits a simple call (never an invoke) to the given no-arguments
3430 /// runtime function.
3432 CodeGenFunction::EmitRuntimeCall(llvm::Value *callee,
3433 const llvm::Twine &name) {
3434 return EmitRuntimeCall(callee, None, name);
3437 // Calls which may throw must have operand bundles indicating which funclet
3438 // they are nested within.
3440 getBundlesForFunclet(llvm::Value *Callee, llvm::Instruction *CurrentFuncletPad,
3441 SmallVectorImpl<llvm::OperandBundleDef> &BundleList) {
3442 // There is no need for a funclet operand bundle if we aren't inside a
3444 if (!CurrentFuncletPad)
3447 // Skip intrinsics which cannot throw.
3448 auto *CalleeFn = dyn_cast<llvm::Function>(Callee->stripPointerCasts());
3449 if (CalleeFn && CalleeFn->isIntrinsic() && CalleeFn->doesNotThrow())
3452 BundleList.emplace_back("funclet", CurrentFuncletPad);
3455 /// Emits a simple call (never an invoke) to the given runtime function.
3457 CodeGenFunction::EmitRuntimeCall(llvm::Value *callee,
3458 ArrayRef<llvm::Value*> args,
3459 const llvm::Twine &name) {
3460 SmallVector<llvm::OperandBundleDef, 1> BundleList;
3461 getBundlesForFunclet(callee, CurrentFuncletPad, BundleList);
3463 llvm::CallInst *call = Builder.CreateCall(callee, args, BundleList, name);
3464 call->setCallingConv(getRuntimeCC());
3468 /// Emits a call or invoke to the given noreturn runtime function.
3469 void CodeGenFunction::EmitNoreturnRuntimeCallOrInvoke(llvm::Value *callee,
3470 ArrayRef<llvm::Value*> args) {
3471 SmallVector<llvm::OperandBundleDef, 1> BundleList;
3472 getBundlesForFunclet(callee, CurrentFuncletPad, BundleList);
3474 if (getInvokeDest()) {
3475 llvm::InvokeInst *invoke =
3476 Builder.CreateInvoke(callee,
3477 getUnreachableBlock(),
3481 invoke->setDoesNotReturn();
3482 invoke->setCallingConv(getRuntimeCC());
3484 llvm::CallInst *call = Builder.CreateCall(callee, args, BundleList);
3485 call->setDoesNotReturn();
3486 call->setCallingConv(getRuntimeCC());
3487 Builder.CreateUnreachable();
3491 /// Emits a call or invoke instruction to the given nullary runtime function.
3493 CodeGenFunction::EmitRuntimeCallOrInvoke(llvm::Value *callee,
3494 const Twine &name) {
3495 return EmitRuntimeCallOrInvoke(callee, None, name);
3498 /// Emits a call or invoke instruction to the given runtime function.
3500 CodeGenFunction::EmitRuntimeCallOrInvoke(llvm::Value *callee,
3501 ArrayRef<llvm::Value*> args,
3502 const Twine &name) {
3503 llvm::CallSite callSite = EmitCallOrInvoke(callee, args, name);
3504 callSite.setCallingConv(getRuntimeCC());
3508 /// Emits a call or invoke instruction to the given function, depending
3509 /// on the current state of the EH stack.
3511 CodeGenFunction::EmitCallOrInvoke(llvm::Value *Callee,
3512 ArrayRef<llvm::Value *> Args,
3513 const Twine &Name) {
3514 llvm::BasicBlock *InvokeDest = getInvokeDest();
3515 SmallVector<llvm::OperandBundleDef, 1> BundleList;
3516 getBundlesForFunclet(Callee, CurrentFuncletPad, BundleList);
3518 llvm::Instruction *Inst;
3520 Inst = Builder.CreateCall(Callee, Args, BundleList, Name);
3522 llvm::BasicBlock *ContBB = createBasicBlock("invoke.cont");
3523 Inst = Builder.CreateInvoke(Callee, ContBB, InvokeDest, Args, BundleList,
3528 // In ObjC ARC mode with no ObjC ARC exception safety, tell the ARC
3529 // optimizer it can aggressively ignore unwind edges.
3530 if (CGM.getLangOpts().ObjCAutoRefCount)
3531 AddObjCARCExceptionMetadata(Inst);
3533 return llvm::CallSite(Inst);
3536 /// \brief Store a non-aggregate value to an address to initialize it. For
3537 /// initialization, a non-atomic store will be used.
3538 static void EmitInitStoreOfNonAggregate(CodeGenFunction &CGF, RValue Src,
3541 CGF.EmitStoreOfScalar(Src.getScalarVal(), Dst, /*init=*/true);
3543 CGF.EmitStoreOfComplex(Src.getComplexVal(), Dst, /*init=*/true);
3546 void CodeGenFunction::deferPlaceholderReplacement(llvm::Instruction *Old,
3548 DeferredReplacements.push_back(std::make_pair(Old, New));
3551 RValue CodeGenFunction::EmitCall(const CGFunctionInfo &CallInfo,
3552 const CGCallee &Callee,
3553 ReturnValueSlot ReturnValue,
3554 const CallArgList &CallArgs,
3555 llvm::Instruction **callOrInvoke) {
3556 // FIXME: We no longer need the types from CallArgs; lift up and simplify.
3558 assert(Callee.isOrdinary());
3560 // Handle struct-return functions by passing a pointer to the
3561 // location that we would like to return into.
3562 QualType RetTy = CallInfo.getReturnType();
3563 const ABIArgInfo &RetAI = CallInfo.getReturnInfo();
3565 llvm::FunctionType *IRFuncTy = Callee.getFunctionType();
3567 // 1. Set up the arguments.
3569 // If we're using inalloca, insert the allocation after the stack save.
3570 // FIXME: Do this earlier rather than hacking it in here!
3571 Address ArgMemory = Address::invalid();
3572 const llvm::StructLayout *ArgMemoryLayout = nullptr;
3573 if (llvm::StructType *ArgStruct = CallInfo.getArgStruct()) {
3574 ArgMemoryLayout = CGM.getDataLayout().getStructLayout(ArgStruct);
3575 llvm::Instruction *IP = CallArgs.getStackBase();
3576 llvm::AllocaInst *AI;
3578 IP = IP->getNextNode();
3579 AI = new llvm::AllocaInst(ArgStruct, "argmem", IP);
3581 AI = CreateTempAlloca(ArgStruct, "argmem");
3583 auto Align = CallInfo.getArgStructAlignment();
3584 AI->setAlignment(Align.getQuantity());
3585 AI->setUsedWithInAlloca(true);
3586 assert(AI->isUsedWithInAlloca() && !AI->isStaticAlloca());
3587 ArgMemory = Address(AI, Align);
3590 // Helper function to drill into the inalloca allocation.
3591 auto createInAllocaStructGEP = [&](unsigned FieldIndex) -> Address {
3593 CharUnits::fromQuantity(ArgMemoryLayout->getElementOffset(FieldIndex));
3594 return Builder.CreateStructGEP(ArgMemory, FieldIndex, FieldOffset);
3597 ClangToLLVMArgMapping IRFunctionArgs(CGM.getContext(), CallInfo);
3598 SmallVector<llvm::Value *, 16> IRCallArgs(IRFunctionArgs.totalIRArgs());
3600 // If the call returns a temporary with struct return, create a temporary
3601 // alloca to hold the result, unless one is given to us.
3602 Address SRetPtr = Address::invalid();
3603 size_t UnusedReturnSize = 0;
3604 if (RetAI.isIndirect() || RetAI.isInAlloca() || RetAI.isCoerceAndExpand()) {
3605 if (!ReturnValue.isNull()) {
3606 SRetPtr = ReturnValue.getValue();
3608 SRetPtr = CreateMemTemp(RetTy);
3609 if (HaveInsertPoint() && ReturnValue.isUnused()) {
3611 CGM.getDataLayout().getTypeAllocSize(ConvertTypeForMem(RetTy));
3612 if (EmitLifetimeStart(size, SRetPtr.getPointer()))
3613 UnusedReturnSize = size;
3616 if (IRFunctionArgs.hasSRetArg()) {
3617 IRCallArgs[IRFunctionArgs.getSRetArgNo()] = SRetPtr.getPointer();
3618 } else if (RetAI.isInAlloca()) {
3619 Address Addr = createInAllocaStructGEP(RetAI.getInAllocaFieldIndex());
3620 Builder.CreateStore(SRetPtr.getPointer(), Addr);
3624 Address swiftErrorTemp = Address::invalid();
3625 Address swiftErrorArg = Address::invalid();
3627 // Translate all of the arguments as necessary to match the IR lowering.
3628 assert(CallInfo.arg_size() == CallArgs.size() &&
3629 "Mismatch between function signature & arguments.");
3631 CGFunctionInfo::const_arg_iterator info_it = CallInfo.arg_begin();
3632 for (CallArgList::const_iterator I = CallArgs.begin(), E = CallArgs.end();
3633 I != E; ++I, ++info_it, ++ArgNo) {
3634 const ABIArgInfo &ArgInfo = info_it->info;
3637 // Insert a padding argument to ensure proper alignment.
3638 if (IRFunctionArgs.hasPaddingArg(ArgNo))
3639 IRCallArgs[IRFunctionArgs.getPaddingArgNo(ArgNo)] =
3640 llvm::UndefValue::get(ArgInfo.getPaddingType());
3642 unsigned FirstIRArg, NumIRArgs;
3643 std::tie(FirstIRArg, NumIRArgs) = IRFunctionArgs.getIRArgs(ArgNo);
3645 switch (ArgInfo.getKind()) {
3646 case ABIArgInfo::InAlloca: {
3647 assert(NumIRArgs == 0);
3648 assert(getTarget().getTriple().getArch() == llvm::Triple::x86);
3649 if (RV.isAggregate()) {
3650 // Replace the placeholder with the appropriate argument slot GEP.
3651 llvm::Instruction *Placeholder =
3652 cast<llvm::Instruction>(RV.getAggregatePointer());
3653 CGBuilderTy::InsertPoint IP = Builder.saveIP();
3654 Builder.SetInsertPoint(Placeholder);
3655 Address Addr = createInAllocaStructGEP(ArgInfo.getInAllocaFieldIndex());
3656 Builder.restoreIP(IP);
3657 deferPlaceholderReplacement(Placeholder, Addr.getPointer());
3659 // Store the RValue into the argument struct.
3660 Address Addr = createInAllocaStructGEP(ArgInfo.getInAllocaFieldIndex());
3661 unsigned AS = Addr.getType()->getPointerAddressSpace();
3662 llvm::Type *MemType = ConvertTypeForMem(I->Ty)->getPointerTo(AS);
3663 // There are some cases where a trivial bitcast is not avoidable. The
3664 // definition of a type later in a translation unit may change it's type
3665 // from {}* to (%struct.foo*)*.
3666 if (Addr.getType() != MemType)
3667 Addr = Builder.CreateBitCast(Addr, MemType);
3668 LValue argLV = MakeAddrLValue(Addr, I->Ty);
3669 EmitInitStoreOfNonAggregate(*this, RV, argLV);
3674 case ABIArgInfo::Indirect: {
3675 assert(NumIRArgs == 1);
3676 if (RV.isScalar() || RV.isComplex()) {
3677 // Make a temporary alloca to pass the argument.
3678 Address Addr = CreateMemTemp(I->Ty, ArgInfo.getIndirectAlign());
3679 IRCallArgs[FirstIRArg] = Addr.getPointer();
3681 LValue argLV = MakeAddrLValue(Addr, I->Ty);
3682 EmitInitStoreOfNonAggregate(*this, RV, argLV);
3684 // We want to avoid creating an unnecessary temporary+copy here;
3685 // however, we need one in three cases:
3686 // 1. If the argument is not byval, and we are required to copy the
3687 // source. (This case doesn't occur on any common architecture.)
3688 // 2. If the argument is byval, RV is not sufficiently aligned, and
3689 // we cannot force it to be sufficiently aligned.
3690 // 3. If the argument is byval, but RV is located in an address space
3691 // different than that of the argument (0).
3692 Address Addr = RV.getAggregateAddress();
3693 CharUnits Align = ArgInfo.getIndirectAlign();
3694 const llvm::DataLayout *TD = &CGM.getDataLayout();
3695 const unsigned RVAddrSpace = Addr.getType()->getAddressSpace();
3696 const unsigned ArgAddrSpace =
3697 (FirstIRArg < IRFuncTy->getNumParams()
3698 ? IRFuncTy->getParamType(FirstIRArg)->getPointerAddressSpace()
3700 if ((!ArgInfo.getIndirectByVal() && I->NeedsCopy) ||
3701 (ArgInfo.getIndirectByVal() && Addr.getAlignment() < Align &&
3702 llvm::getOrEnforceKnownAlignment(Addr.getPointer(),
3703 Align.getQuantity(), *TD)
3704 < Align.getQuantity()) ||
3705 (ArgInfo.getIndirectByVal() && (RVAddrSpace != ArgAddrSpace))) {
3706 // Create an aligned temporary, and copy to it.
3707 Address AI = CreateMemTemp(I->Ty, ArgInfo.getIndirectAlign());
3708 IRCallArgs[FirstIRArg] = AI.getPointer();
3709 EmitAggregateCopy(AI, Addr, I->Ty, RV.isVolatileQualified());
3711 // Skip the extra memcpy call.
3712 IRCallArgs[FirstIRArg] = Addr.getPointer();
3718 case ABIArgInfo::Ignore:
3719 assert(NumIRArgs == 0);
3722 case ABIArgInfo::Extend:
3723 case ABIArgInfo::Direct: {
3724 if (!isa<llvm::StructType>(ArgInfo.getCoerceToType()) &&
3725 ArgInfo.getCoerceToType() == ConvertType(info_it->type) &&
3726 ArgInfo.getDirectOffset() == 0) {
3727 assert(NumIRArgs == 1);
3730 V = RV.getScalarVal();
3732 V = Builder.CreateLoad(RV.getAggregateAddress());
3734 // Implement swifterror by copying into a new swifterror argument.
3735 // We'll write back in the normal path out of the call.
3736 if (CallInfo.getExtParameterInfo(ArgNo).getABI()
3737 == ParameterABI::SwiftErrorResult) {
3738 assert(!swiftErrorTemp.isValid() && "multiple swifterror args");
3740 QualType pointeeTy = I->Ty->getPointeeType();
3742 Address(V, getContext().getTypeAlignInChars(pointeeTy));
3745 CreateMemTemp(pointeeTy, getPointerAlign(), "swifterror.temp");
3746 V = swiftErrorTemp.getPointer();
3747 cast<llvm::AllocaInst>(V)->setSwiftError(true);
3749 llvm::Value *errorValue = Builder.CreateLoad(swiftErrorArg);
3750 Builder.CreateStore(errorValue, swiftErrorTemp);
3753 // We might have to widen integers, but we should never truncate.
3754 if (ArgInfo.getCoerceToType() != V->getType() &&
3755 V->getType()->isIntegerTy())
3756 V = Builder.CreateZExt(V, ArgInfo.getCoerceToType());
3758 // If the argument doesn't match, perform a bitcast to coerce it. This
3759 // can happen due to trivial type mismatches.
3760 if (FirstIRArg < IRFuncTy->getNumParams() &&
3761 V->getType() != IRFuncTy->getParamType(FirstIRArg))
3762 V = Builder.CreateBitCast(V, IRFuncTy->getParamType(FirstIRArg));
3764 IRCallArgs[FirstIRArg] = V;
3768 // FIXME: Avoid the conversion through memory if possible.
3769 Address Src = Address::invalid();
3770 if (RV.isScalar() || RV.isComplex()) {
3771 Src = CreateMemTemp(I->Ty, "coerce");
3772 LValue SrcLV = MakeAddrLValue(Src, I->Ty);
3773 EmitInitStoreOfNonAggregate(*this, RV, SrcLV);
3775 Src = RV.getAggregateAddress();
3778 // If the value is offset in memory, apply the offset now.
3779 Src = emitAddressAtOffset(*this, Src, ArgInfo);
3781 // Fast-isel and the optimizer generally like scalar values better than
3782 // FCAs, so we flatten them if this is safe to do for this argument.
3783 llvm::StructType *STy =
3784 dyn_cast<llvm::StructType>(ArgInfo.getCoerceToType());
3785 if (STy && ArgInfo.isDirect() && ArgInfo.getCanBeFlattened()) {
3786 llvm::Type *SrcTy = Src.getType()->getElementType();
3787 uint64_t SrcSize = CGM.getDataLayout().getTypeAllocSize(SrcTy);
3788 uint64_t DstSize = CGM.getDataLayout().getTypeAllocSize(STy);
3790 // If the source type is smaller than the destination type of the
3791 // coerce-to logic, copy the source value into a temp alloca the size
3792 // of the destination type to allow loading all of it. The bits past
3793 // the source value are left undef.
3794 if (SrcSize < DstSize) {
3796 = CreateTempAlloca(STy, Src.getAlignment(),
3797 Src.getName() + ".coerce");
3798 Builder.CreateMemCpy(TempAlloca, Src, SrcSize);
3801 Src = Builder.CreateBitCast(Src, llvm::PointerType::getUnqual(STy));
3804 auto SrcLayout = CGM.getDataLayout().getStructLayout(STy);
3805 assert(NumIRArgs == STy->getNumElements());
3806 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
3807 auto Offset = CharUnits::fromQuantity(SrcLayout->getElementOffset(i));
3808 Address EltPtr = Builder.CreateStructGEP(Src, i, Offset);
3809 llvm::Value *LI = Builder.CreateLoad(EltPtr);
3810 IRCallArgs[FirstIRArg + i] = LI;
3813 // In the simple case, just pass the coerced loaded value.
3814 assert(NumIRArgs == 1);
3815 IRCallArgs[FirstIRArg] =
3816 CreateCoercedLoad(Src, ArgInfo.getCoerceToType(), *this);
3822 case ABIArgInfo::CoerceAndExpand: {
3823 auto coercionType = ArgInfo.getCoerceAndExpandType();
3824 auto layout = CGM.getDataLayout().getStructLayout(coercionType);
3826 llvm::Value *tempSize = nullptr;
3827 Address addr = Address::invalid();
3828 if (RV.isAggregate()) {
3829 addr = RV.getAggregateAddress();
3831 assert(RV.isScalar()); // complex should always just be direct
3833 llvm::Type *scalarType = RV.getScalarVal()->getType();
3834 auto scalarSize = CGM.getDataLayout().getTypeAllocSize(scalarType);
3835 auto scalarAlign = CGM.getDataLayout().getPrefTypeAlignment(scalarType);
3837 tempSize = llvm::ConstantInt::get(CGM.Int64Ty, scalarSize);
3839 // Materialize to a temporary.
3840 addr = CreateTempAlloca(RV.getScalarVal()->getType(),
3841 CharUnits::fromQuantity(std::max(layout->getAlignment(),
3843 EmitLifetimeStart(scalarSize, addr.getPointer());
3845 Builder.CreateStore(RV.getScalarVal(), addr);
3848 addr = Builder.CreateElementBitCast(addr, coercionType);
3850 unsigned IRArgPos = FirstIRArg;
3851 for (unsigned i = 0, e = coercionType->getNumElements(); i != e; ++i) {
3852 llvm::Type *eltType = coercionType->getElementType(i);
3853 if (ABIArgInfo::isPaddingForCoerceAndExpand(eltType)) continue;
3854 Address eltAddr = Builder.CreateStructGEP(addr, i, layout);
3855 llvm::Value *elt = Builder.CreateLoad(eltAddr);
3856 IRCallArgs[IRArgPos++] = elt;
3858 assert(IRArgPos == FirstIRArg + NumIRArgs);
3861 EmitLifetimeEnd(tempSize, addr.getPointer());
3867 case ABIArgInfo::Expand:
3868 unsigned IRArgPos = FirstIRArg;
3869 ExpandTypeToArgs(I->Ty, RV, IRFuncTy, IRCallArgs, IRArgPos);
3870 assert(IRArgPos == FirstIRArg + NumIRArgs);
3875 llvm::Value *CalleePtr = Callee.getFunctionPointer();
3877 // If we're using inalloca, set up that argument.
3878 if (ArgMemory.isValid()) {
3879 llvm::Value *Arg = ArgMemory.getPointer();
3880 if (CallInfo.isVariadic()) {
3881 // When passing non-POD arguments by value to variadic functions, we will
3882 // end up with a variadic prototype and an inalloca call site. In such
3883 // cases, we can't do any parameter mismatch checks. Give up and bitcast
3885 unsigned CalleeAS = CalleePtr->getType()->getPointerAddressSpace();
3886 auto FnTy = getTypes().GetFunctionType(CallInfo)->getPointerTo(CalleeAS);
3887 CalleePtr = Builder.CreateBitCast(CalleePtr, FnTy);
3889 llvm::Type *LastParamTy =
3890 IRFuncTy->getParamType(IRFuncTy->getNumParams() - 1);
3891 if (Arg->getType() != LastParamTy) {
3893 // Assert that these structs have equivalent element types.
3894 llvm::StructType *FullTy = CallInfo.getArgStruct();
3895 llvm::StructType *DeclaredTy = cast<llvm::StructType>(
3896 cast<llvm::PointerType>(LastParamTy)->getElementType());
3897 assert(DeclaredTy->getNumElements() == FullTy->getNumElements());
3898 for (llvm::StructType::element_iterator DI = DeclaredTy->element_begin(),
3899 DE = DeclaredTy->element_end(),
3900 FI = FullTy->element_begin();
3901 DI != DE; ++DI, ++FI)
3904 Arg = Builder.CreateBitCast(Arg, LastParamTy);
3907 assert(IRFunctionArgs.hasInallocaArg());
3908 IRCallArgs[IRFunctionArgs.getInallocaArgNo()] = Arg;
3911 // 2. Prepare the function pointer.
3913 // If the callee is a bitcast of a non-variadic function to have a
3914 // variadic function pointer type, check to see if we can remove the
3915 // bitcast. This comes up with unprototyped functions.
3917 // This makes the IR nicer, but more importantly it ensures that we
3918 // can inline the function at -O0 if it is marked always_inline.
3919 auto simplifyVariadicCallee = [](llvm::Value *Ptr) -> llvm::Value* {
3920 llvm::FunctionType *CalleeFT =
3921 cast<llvm::FunctionType>(Ptr->getType()->getPointerElementType());
3922 if (!CalleeFT->isVarArg())
3925 llvm::ConstantExpr *CE = dyn_cast<llvm::ConstantExpr>(Ptr);
3926 if (!CE || CE->getOpcode() != llvm::Instruction::BitCast)
3929 llvm::Function *OrigFn = dyn_cast<llvm::Function>(CE->getOperand(0));
3933 llvm::FunctionType *OrigFT = OrigFn->getFunctionType();
3935 // If the original type is variadic, or if any of the component types
3936 // disagree, we cannot remove the cast.
3937 if (OrigFT->isVarArg() ||
3938 OrigFT->getNumParams() != CalleeFT->getNumParams() ||
3939 OrigFT->getReturnType() != CalleeFT->getReturnType())
3942 for (unsigned i = 0, e = OrigFT->getNumParams(); i != e; ++i)
3943 if (OrigFT->getParamType(i) != CalleeFT->getParamType(i))
3948 CalleePtr = simplifyVariadicCallee(CalleePtr);
3950 // 3. Perform the actual call.
3952 // Deactivate any cleanups that we're supposed to do immediately before
3954 if (!CallArgs.getCleanupsToDeactivate().empty())
3955 deactivateArgCleanupsBeforeCall(*this, CallArgs);
3957 // Assert that the arguments we computed match up. The IR verifier
3958 // will catch this, but this is a common enough source of problems
3959 // during IRGen changes that it's way better for debugging to catch
3960 // it ourselves here.
3962 assert(IRCallArgs.size() == IRFuncTy->getNumParams() || IRFuncTy->isVarArg());
3963 for (unsigned i = 0; i < IRCallArgs.size(); ++i) {
3964 // Inalloca argument can have different type.
3965 if (IRFunctionArgs.hasInallocaArg() &&
3966 i == IRFunctionArgs.getInallocaArgNo())
3968 if (i < IRFuncTy->getNumParams())
3969 assert(IRCallArgs[i]->getType() == IRFuncTy->getParamType(i));
3973 // Compute the calling convention and attributes.
3974 unsigned CallingConv;
3975 CodeGen::AttributeListType AttributeList;
3976 CGM.ConstructAttributeList(CalleePtr->getName(), CallInfo,
3977 Callee.getAbstractInfo(),
3978 AttributeList, CallingConv,
3979 /*AttrOnCallSite=*/true);
3980 llvm::AttributeSet Attrs = llvm::AttributeSet::get(getLLVMContext(),
3983 // Apply some call-site-specific attributes.
3984 // TODO: work this into building the attribute set.
3986 // Apply always_inline to all calls within flatten functions.
3987 // FIXME: should this really take priority over __try, below?
3988 if (CurCodeDecl && CurCodeDecl->hasAttr<FlattenAttr>() &&
3989 !(Callee.getAbstractInfo().getCalleeDecl() &&
3990 Callee.getAbstractInfo().getCalleeDecl()->hasAttr<NoInlineAttr>())) {
3992 Attrs.addAttribute(getLLVMContext(),
3993 llvm::AttributeSet::FunctionIndex,
3994 llvm::Attribute::AlwaysInline);
3997 // Disable inlining inside SEH __try blocks.
3998 if (isSEHTryScope()) {
4000 Attrs.addAttribute(getLLVMContext(), llvm::AttributeSet::FunctionIndex,
4001 llvm::Attribute::NoInline);
4004 // Decide whether to use a call or an invoke.
4006 if (currentFunctionUsesSEHTry()) {
4007 // SEH cares about asynchronous exceptions, so everything can "throw."
4008 CannotThrow = false;
4009 } else if (isCleanupPadScope() &&
4010 EHPersonality::get(*this).isMSVCXXPersonality()) {
4011 // The MSVC++ personality will implicitly terminate the program if an
4012 // exception is thrown during a cleanup outside of a try/catch.
4013 // We don't need to model anything in IR to get this behavior.
4016 // Otherwise, nounwind call sites will never throw.
4017 CannotThrow = Attrs.hasAttribute(llvm::AttributeSet::FunctionIndex,
4018 llvm::Attribute::NoUnwind);
4020 llvm::BasicBlock *InvokeDest = CannotThrow ? nullptr : getInvokeDest();
4022 SmallVector<llvm::OperandBundleDef, 1> BundleList;
4023 getBundlesForFunclet(CalleePtr, CurrentFuncletPad, BundleList);
4025 // Emit the actual call/invoke instruction.
4028 CS = Builder.CreateCall(CalleePtr, IRCallArgs, BundleList);
4030 llvm::BasicBlock *Cont = createBasicBlock("invoke.cont");
4031 CS = Builder.CreateInvoke(CalleePtr, Cont, InvokeDest, IRCallArgs,
4035 llvm::Instruction *CI = CS.getInstruction();
4039 // Apply the attributes and calling convention.
4040 CS.setAttributes(Attrs);
4041 CS.setCallingConv(static_cast<llvm::CallingConv::ID>(CallingConv));
4043 // Apply various metadata.
4045 if (!CI->getType()->isVoidTy())
4046 CI->setName("call");
4048 // Insert instrumentation or attach profile metadata at indirect call sites.
4049 // For more details, see the comment before the definition of
4050 // IPVK_IndirectCallTarget in InstrProfData.inc.
4051 if (!CS.getCalledFunction())
4052 PGO.valueProfile(Builder, llvm::IPVK_IndirectCallTarget,
4055 // In ObjC ARC mode with no ObjC ARC exception safety, tell the ARC
4056 // optimizer it can aggressively ignore unwind edges.
4057 if (CGM.getLangOpts().ObjCAutoRefCount)
4058 AddObjCARCExceptionMetadata(CI);
4060 // Suppress tail calls if requested.
4061 if (llvm::CallInst *Call = dyn_cast<llvm::CallInst>(CI)) {
4062 const Decl *TargetDecl = Callee.getAbstractInfo().getCalleeDecl();
4063 if (TargetDecl && TargetDecl->hasAttr<NotTailCalledAttr>())
4064 Call->setTailCallKind(llvm::CallInst::TCK_NoTail);
4067 // 4. Finish the call.
4069 // If the call doesn't return, finish the basic block and clear the
4070 // insertion point; this allows the rest of IRGen to discard
4071 // unreachable code.
4072 if (CS.doesNotReturn()) {
4073 if (UnusedReturnSize)
4074 EmitLifetimeEnd(llvm::ConstantInt::get(Int64Ty, UnusedReturnSize),
4075 SRetPtr.getPointer());
4077 Builder.CreateUnreachable();
4078 Builder.ClearInsertionPoint();
4080 // FIXME: For now, emit a dummy basic block because expr emitters in
4081 // generally are not ready to handle emitting expressions at unreachable
4083 EnsureInsertPoint();
4085 // Return a reasonable RValue.
4086 return GetUndefRValue(RetTy);
4089 // Perform the swifterror writeback.
4090 if (swiftErrorTemp.isValid()) {
4091 llvm::Value *errorResult = Builder.CreateLoad(swiftErrorTemp);
4092 Builder.CreateStore(errorResult, swiftErrorArg);
4095 // Emit any call-associated writebacks immediately. Arguably this
4096 // should happen after any return-value munging.
4097 if (CallArgs.hasWritebacks())
4098 emitWritebacks(*this, CallArgs);
4100 // The stack cleanup for inalloca arguments has to run out of the normal
4101 // lexical order, so deactivate it and run it manually here.
4102 CallArgs.freeArgumentMemory(*this);
4104 // Extract the return value.
4106 switch (RetAI.getKind()) {
4107 case ABIArgInfo::CoerceAndExpand: {
4108 auto coercionType = RetAI.getCoerceAndExpandType();
4109 auto layout = CGM.getDataLayout().getStructLayout(coercionType);
4111 Address addr = SRetPtr;
4112 addr = Builder.CreateElementBitCast(addr, coercionType);
4114 assert(CI->getType() == RetAI.getUnpaddedCoerceAndExpandType());
4115 bool requiresExtract = isa<llvm::StructType>(CI->getType());
4117 unsigned unpaddedIndex = 0;
4118 for (unsigned i = 0, e = coercionType->getNumElements(); i != e; ++i) {
4119 llvm::Type *eltType = coercionType->getElementType(i);
4120 if (ABIArgInfo::isPaddingForCoerceAndExpand(eltType)) continue;
4121 Address eltAddr = Builder.CreateStructGEP(addr, i, layout);
4122 llvm::Value *elt = CI;
4123 if (requiresExtract)
4124 elt = Builder.CreateExtractValue(elt, unpaddedIndex++);
4126 assert(unpaddedIndex == 0);
4127 Builder.CreateStore(elt, eltAddr);
4132 case ABIArgInfo::InAlloca:
4133 case ABIArgInfo::Indirect: {
4134 RValue ret = convertTempToRValue(SRetPtr, RetTy, SourceLocation());
4135 if (UnusedReturnSize)
4136 EmitLifetimeEnd(llvm::ConstantInt::get(Int64Ty, UnusedReturnSize),
4137 SRetPtr.getPointer());
4141 case ABIArgInfo::Ignore:
4142 // If we are ignoring an argument that had a result, make sure to
4143 // construct the appropriate return value for our caller.
4144 return GetUndefRValue(RetTy);
4146 case ABIArgInfo::Extend:
4147 case ABIArgInfo::Direct: {
4148 llvm::Type *RetIRTy = ConvertType(RetTy);
4149 if (RetAI.getCoerceToType() == RetIRTy && RetAI.getDirectOffset() == 0) {
4150 switch (getEvaluationKind(RetTy)) {
4152 llvm::Value *Real = Builder.CreateExtractValue(CI, 0);
4153 llvm::Value *Imag = Builder.CreateExtractValue(CI, 1);
4154 return RValue::getComplex(std::make_pair(Real, Imag));
4156 case TEK_Aggregate: {
4157 Address DestPtr = ReturnValue.getValue();
4158 bool DestIsVolatile = ReturnValue.isVolatile();
4160 if (!DestPtr.isValid()) {
4161 DestPtr = CreateMemTemp(RetTy, "agg.tmp");
4162 DestIsVolatile = false;
4164 BuildAggStore(*this, CI, DestPtr, DestIsVolatile);
4165 return RValue::getAggregate(DestPtr);
4168 // If the argument doesn't match, perform a bitcast to coerce it. This
4169 // can happen due to trivial type mismatches.
4170 llvm::Value *V = CI;
4171 if (V->getType() != RetIRTy)
4172 V = Builder.CreateBitCast(V, RetIRTy);
4173 return RValue::get(V);
4176 llvm_unreachable("bad evaluation kind");
4179 Address DestPtr = ReturnValue.getValue();
4180 bool DestIsVolatile = ReturnValue.isVolatile();
4182 if (!DestPtr.isValid()) {
4183 DestPtr = CreateMemTemp(RetTy, "coerce");
4184 DestIsVolatile = false;
4187 // If the value is offset in memory, apply the offset now.
4188 Address StorePtr = emitAddressAtOffset(*this, DestPtr, RetAI);
4189 CreateCoercedStore(CI, StorePtr, DestIsVolatile, *this);
4191 return convertTempToRValue(DestPtr, RetTy, SourceLocation());
4194 case ABIArgInfo::Expand:
4195 llvm_unreachable("Invalid ABI kind for return argument");
4198 llvm_unreachable("Unhandled ABIArgInfo::Kind");
4201 // Emit the assume_aligned check on the return value.
4202 const Decl *TargetDecl = Callee.getAbstractInfo().getCalleeDecl();
4203 if (Ret.isScalar() && TargetDecl) {
4204 if (const auto *AA = TargetDecl->getAttr<AssumeAlignedAttr>()) {
4205 llvm::Value *OffsetValue = nullptr;
4206 if (const auto *Offset = AA->getOffset())
4207 OffsetValue = EmitScalarExpr(Offset);
4209 llvm::Value *Alignment = EmitScalarExpr(AA->getAlignment());
4210 llvm::ConstantInt *AlignmentCI = cast<llvm::ConstantInt>(Alignment);
4211 EmitAlignmentAssumption(Ret.getScalarVal(), AlignmentCI->getZExtValue(),
4219 /* VarArg handling */
4221 Address CodeGenFunction::EmitVAArg(VAArgExpr *VE, Address &VAListAddr) {
4222 VAListAddr = VE->isMicrosoftABI()
4223 ? EmitMSVAListRef(VE->getSubExpr())
4224 : EmitVAListRef(VE->getSubExpr());
4225 QualType Ty = VE->getType();
4226 if (VE->isMicrosoftABI())
4227 return CGM.getTypes().getABIInfo().EmitMSVAArg(*this, VAListAddr, Ty);
4228 return CGM.getTypes().getABIInfo().EmitVAArg(*this, VAListAddr, Ty);