Debian Codegraph @

llvm-project/clang/lib/CodeGen/CGCall.cpp

raw
1 //===--- CGCall.cpp - Encapsulate calling convention details --------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // These classes wrap the information about a call or function
10 // definition used to handle ABI compliancy.
11 //
12 //===----------------------------------------------------------------------===//
13
14 #include "CGCall.h"
15 #include "ABIInfo.h"
16 #include "CGBlocks.h"
17 #include "CGCXXABI.h"
18 #include "CGCleanup.h"
19 #include "CGRecordLayout.h"
20 #include "CodeGenFunction.h"
21 #include "CodeGenModule.h"
22 #include "TargetInfo.h"
23 #include "clang/AST/Attr.h"
24 #include "clang/AST/Decl.h"
25 #include "clang/AST/DeclCXX.h"
26 #include "clang/AST/DeclObjC.h"
27 #include "clang/Basic/CodeGenOptions.h"
28 #include "clang/Basic/TargetBuiltins.h"
29 #include "clang/Basic/TargetInfo.h"
30 #include "clang/CodeGen/CGFunctionInfo.h"
31 #include "clang/CodeGen/SwiftCallingConv.h"
32 #include "llvm/ADT/StringExtras.h"
33 #include "llvm/Analysis/ValueTracking.h"
34 #include "llvm/IR/Assumptions.h"
35 #include "llvm/IR/Attributes.h"
36 #include "llvm/IR/CallingConv.h"
37 #include "llvm/IR/DataLayout.h"
38 #include "llvm/IR/InlineAsm.h"
39 #include "llvm/IR/IntrinsicInst.h"
40 #include "llvm/IR/Intrinsics.h"
41 #include "llvm/IR/Type.h"
42 #include "llvm/Transforms/Utils/Local.h"
43 #include <optional>
44 using namespace clang;
45 using namespace CodeGen;
46
47 /***/
48
49 unsigned CodeGenTypes::ClangCallConvToLLVMCallConv(CallingConv CC) {
50 switch (CC) {
51 default: return llvm::CallingConv::C;
52 case CC_X86StdCall: return llvm::CallingConv::X86_StdCall;
53 case CC_X86FastCall: return llvm::CallingConv::X86_FastCall;
54 case CC_X86RegCall: return llvm::CallingConv::X86_RegCall;
55 case CC_X86ThisCall: return llvm::CallingConv::X86_ThisCall;
56 case CC_Win64: return llvm::CallingConv::Win64;
57 case CC_X86_64SysV: return llvm::CallingConv::X86_64_SysV;
58 case CC_AAPCS: return llvm::CallingConv::ARM_AAPCS;
59 case CC_AAPCS_VFP: return llvm::CallingConv::ARM_AAPCS_VFP;
60 case CC_IntelOclBicc: return llvm::CallingConv::Intel_OCL_BI;
61 // TODO: Add support for __pascal to LLVM.
62 case CC_X86Pascal: return llvm::CallingConv::C;
63 // TODO: Add support for __vectorcall to LLVM.
64 case CC_X86VectorCall: return llvm::CallingConv::X86_VectorCall;
65 case CC_AArch64VectorCall: return llvm::CallingConv::AArch64_VectorCall;
66 case CC_AArch64SVEPCS: return llvm::CallingConv::AArch64_SVE_VectorCall;
67 case CC_AMDGPUKernelCall: return llvm::CallingConv::AMDGPU_KERNEL;
68 case CC_SpirFunction: return llvm::CallingConv::SPIR_FUNC;
69 case CC_OpenCLKernel: return CGM.getTargetCodeGenInfo().getOpenCLKernelCallingConv();
70 case CC_PreserveMost: return llvm::CallingConv::PreserveMost;
71 case CC_PreserveAll: return llvm::CallingConv::PreserveAll;
72 case CC_Swift: return llvm::CallingConv::Swift;
73 case CC_SwiftAsync: return llvm::CallingConv::SwiftTail;
74 }
75 }
76
77 /// Derives the 'this' type for codegen purposes, i.e. ignoring method CVR
78 /// qualification. Either or both of RD and MD may be null. A null RD indicates
79 /// that there is no meaningful 'this' type, and a null MD can occur when
80 /// calling a method pointer.
81 CanQualType CodeGenTypes::DeriveThisType(const CXXRecordDecl *RD,
82 const CXXMethodDecl *MD) {
83 QualType RecTy;
84 if (RD)
85 RecTy = Context.getTagDeclType(RD)->getCanonicalTypeInternal();
86 else
87 RecTy = Context.VoidTy;
88
89 if (MD)
90 RecTy = Context.getAddrSpaceQualType(RecTy, MD->getMethodQualifiers().getAddressSpace());
91 return Context.getPointerType(CanQualType::CreateUnsafe(RecTy));
92 }
93
94 /// Returns the canonical formal type of the given C++ method.
95 static CanQual<FunctionProtoType> GetFormalType(const CXXMethodDecl *MD) {
96 return MD->getType()->getCanonicalTypeUnqualified()
97 .getAs<FunctionProtoType>();
98 }
99
100 /// Returns the "extra-canonicalized" return type, which discards
101 /// qualifiers on the return type. Codegen doesn't care about them,
102 /// and it makes ABI code a little easier to be able to assume that
103 /// all parameter and return types are top-level unqualified.
104 static CanQualType GetReturnType(QualType RetTy) {
105 return RetTy->getCanonicalTypeUnqualified().getUnqualifiedType();
106 }
107
108 /// Arrange the argument and result information for a value of the given
109 /// unprototyped freestanding function type.
110 const CGFunctionInfo &
111 CodeGenTypes::arrangeFreeFunctionType(CanQual<FunctionNoProtoType> FTNP) {
112 // When translating an unprototyped function type, always use a
113 // variadic type.
114 return arrangeLLVMFunctionInfo(FTNP->getReturnType().getUnqualifiedType(),
115 /*instanceMethod=*/false,
116 /*chainCall=*/false, std::nullopt,
117 FTNP->getExtInfo(), {}, RequiredArgs(0));
118 }
119
120 static void addExtParameterInfosForCall(
121 llvm::SmallVectorImpl<FunctionProtoType::ExtParameterInfo> &paramInfos,
122 const FunctionProtoType *proto,
123 unsigned prefixArgs,
124 unsigned totalArgs) {
125 assert(proto->hasExtParameterInfos());
126 assert(paramInfos.size() <= prefixArgs);
127 assert(proto->getNumParams() + prefixArgs <= totalArgs);
128
129 paramInfos.reserve(totalArgs);
130
131 // Add default infos for any prefix args that don't already have infos.
132 paramInfos.resize(prefixArgs);
133
134 // Add infos for the prototype.
135 for (const auto &ParamInfo : proto->getExtParameterInfos()) {
136 paramInfos.push_back(ParamInfo);
137 // pass_object_size params have no parameter info.
138 if (ParamInfo.hasPassObjectSize())
139 paramInfos.emplace_back();
140 }
141
142 assert(paramInfos.size() <= totalArgs &&
143 "Did we forget to insert pass_object_size args?");
144 // Add default infos for the variadic and/or suffix arguments.
145 paramInfos.resize(totalArgs);
146 }
147
148 /// Adds the formal parameters in FPT to the given prefix. If any parameter in
149 /// FPT has pass_object_size attrs, then we'll add parameters for those, too.
150 static void appendParameterTypes(const CodeGenTypes &CGT,
151 SmallVectorImpl<CanQualType> &prefix,
152 SmallVectorImpl<FunctionProtoType::ExtParameterInfo> &paramInfos,
153 CanQual<FunctionProtoType> FPT) {
154 // Fast path: don't touch param info if we don't need to.
155 if (!FPT->hasExtParameterInfos()) {
156 assert(paramInfos.empty() &&
157 "We have paramInfos, but the prototype doesn't?");
158 prefix.append(FPT->param_type_begin(), FPT->param_type_end());
159 return;
160 }
161
162 unsigned PrefixSize = prefix.size();
163 // In the vast majority of cases, we'll have precisely FPT->getNumParams()
164 // parameters; the only thing that can change this is the presence of
165 // pass_object_size. So, we preallocate for the common case.
166 prefix.reserve(prefix.size() + FPT->getNumParams());
167
168 auto ExtInfos = FPT->getExtParameterInfos();
169 assert(ExtInfos.size() == FPT->getNumParams());
170 for (unsigned I = 0, E = FPT->getNumParams(); I != E; ++I) {
171 prefix.push_back(FPT->getParamType(I));
172 if (ExtInfos[I].hasPassObjectSize())
173 prefix.push_back(CGT.getContext().getSizeType());
174 }
175
176 addExtParameterInfosForCall(paramInfos, FPT.getTypePtr(), PrefixSize,
177 prefix.size());
178 }
179
180 /// Arrange the LLVM function layout for a value of the given function
181 /// type, on top of any implicit parameters already stored.
182 static const CGFunctionInfo &
183 arrangeLLVMFunctionInfo(CodeGenTypes &CGT, bool instanceMethod,
184 SmallVectorImpl<CanQualType> &prefix,
185 CanQual<FunctionProtoType> FTP) {
186 SmallVector<FunctionProtoType::ExtParameterInfo, 16> paramInfos;
187 RequiredArgs Required = RequiredArgs::forPrototypePlus(FTP, prefix.size());
188 // FIXME: Kill copy.
189 appendParameterTypes(CGT, prefix, paramInfos, FTP);
190 CanQualType resultType = FTP->getReturnType().getUnqualifiedType();
191
192 return CGT.arrangeLLVMFunctionInfo(resultType, instanceMethod,
193 /*chainCall=*/false, prefix,
194 FTP->getExtInfo(), paramInfos,
195 Required);
196 }
197
198 /// Arrange the argument and result information for a value of the
199 /// given freestanding function type.
200 const CGFunctionInfo &
201 CodeGenTypes::arrangeFreeFunctionType(CanQual<FunctionProtoType> FTP) {
202 SmallVector<CanQualType, 16> argTypes;
203 return ::arrangeLLVMFunctionInfo(*this, /*instanceMethod=*/false, argTypes,
204 FTP);
205 }
206
207 static CallingConv getCallingConventionForDecl(const ObjCMethodDecl *D,
208 bool IsWindows) {
209 // Set the appropriate calling convention for the Function.
210 if (D->hasAttr<StdCallAttr>())
211 return CC_X86StdCall;
212
213 if (D->hasAttr<FastCallAttr>())
214 return CC_X86FastCall;
215
216 if (D->hasAttr<RegCallAttr>())
217 return CC_X86RegCall;
218
219 if (D->hasAttr<ThisCallAttr>())
220 return CC_X86ThisCall;
221
222 if (D->hasAttr<VectorCallAttr>())
223 return CC_X86VectorCall;
224
225 if (D->hasAttr<PascalAttr>())
226 return CC_X86Pascal;
227
228 if (PcsAttr *PCS = D->getAttr<PcsAttr>())
229 return (PCS->getPCS() == PcsAttr::AAPCS ? CC_AAPCS : CC_AAPCS_VFP);
230
231 if (D->hasAttr<AArch64VectorPcsAttr>())
232 return CC_AArch64VectorCall;
233
234 if (D->hasAttr<AArch64SVEPcsAttr>())
235 return CC_AArch64SVEPCS;
236
237 if (D->hasAttr<AMDGPUKernelCallAttr>())
238 return CC_AMDGPUKernelCall;
239
240 if (D->hasAttr<IntelOclBiccAttr>())
241 return CC_IntelOclBicc;
242
243 if (D->hasAttr<MSABIAttr>())
244 return IsWindows ? CC_C : CC_Win64;
245
246 if (D->hasAttr<SysVABIAttr>())
247 return IsWindows ? CC_X86_64SysV : CC_C;
248
249 if (D->hasAttr<PreserveMostAttr>())
250 return CC_PreserveMost;
251
252 if (D->hasAttr<PreserveAllAttr>())
253 return CC_PreserveAll;
254
255 return CC_C;
256 }
257
258 /// Arrange the argument and result information for a call to an
259 /// unknown C++ non-static member function of the given abstract type.
260 /// (A null RD means we don't have any meaningful "this" argument type,
261 /// so fall back to a generic pointer type).
262 /// The member function must be an ordinary function, i.e. not a
263 /// constructor or destructor.
264 const CGFunctionInfo &
265 CodeGenTypes::arrangeCXXMethodType(const CXXRecordDecl *RD,
266 const FunctionProtoType *FTP,
267 const CXXMethodDecl *MD) {
268 SmallVector<CanQualType, 16> argTypes;
269
270 // Add the 'this' pointer.
271 argTypes.push_back(DeriveThisType(RD, MD));
272
273 return ::arrangeLLVMFunctionInfo(
274 *this, true, argTypes,
275 FTP->getCanonicalTypeUnqualified().getAs<FunctionProtoType>());
276 }
277
278 /// Set calling convention for CUDA/HIP kernel.
279 static void setCUDAKernelCallingConvention(CanQualType &FTy, CodeGenModule &CGM,
280 const FunctionDecl *FD) {
281 if (FD->hasAttr<CUDAGlobalAttr>()) {
282 const FunctionType *FT = FTy->getAs<FunctionType>();
283 CGM.getTargetCodeGenInfo().setCUDAKernelCallingConvention(FT);
284 FTy = FT->getCanonicalTypeUnqualified();
285 }
286 }
287
288 /// Arrange the argument and result information for a declaration or
289 /// definition of the given C++ non-static member function. The
290 /// member function must be an ordinary function, i.e. not a
291 /// constructor or destructor.
292 const CGFunctionInfo &
293 CodeGenTypes::arrangeCXXMethodDeclaration(const CXXMethodDecl *MD) {
294 assert(!isa<CXXConstructorDecl>(MD) && "wrong method for constructors!");
295 assert(!isa<CXXDestructorDecl>(MD) && "wrong method for destructors!");
296
297 CanQualType FT = GetFormalType(MD).getAs<Type>();
298 setCUDAKernelCallingConvention(FT, CGM, MD);
299 auto prototype = FT.getAs<FunctionProtoType>();
300
301 if (MD->isInstance()) {
302 // The abstract case is perfectly fine.
303 const CXXRecordDecl *ThisType = TheCXXABI.getThisArgumentTypeForMethod(MD);
304 return arrangeCXXMethodType(ThisType, prototype.getTypePtr(), MD);
305 }
306
307 return arrangeFreeFunctionType(prototype);
308 }
309
310 bool CodeGenTypes::inheritingCtorHasParams(
311 const InheritedConstructor &Inherited, CXXCtorType Type) {
312 // Parameters are unnecessary if we're constructing a base class subobject
313 // and the inherited constructor lives in a virtual base.
314 return Type == Ctor_Complete ||
315 !Inherited.getShadowDecl()->constructsVirtualBase() ||
316 !Target.getCXXABI().hasConstructorVariants();
317 }
318
319 const CGFunctionInfo &
320 CodeGenTypes::arrangeCXXStructorDeclaration(GlobalDecl GD) {
321 auto *MD = cast<CXXMethodDecl>(GD.getDecl());
322
323 SmallVector<CanQualType, 16> argTypes;
324 SmallVector<FunctionProtoType::ExtParameterInfo, 16> paramInfos;
325
326 const CXXRecordDecl *ThisType = TheCXXABI.getThisArgumentTypeForMethod(GD);
327 argTypes.push_back(DeriveThisType(ThisType, MD));
328
329 bool PassParams = true;
330
331 if (auto *CD = dyn_cast<CXXConstructorDecl>(MD)) {
332 // A base class inheriting constructor doesn't get forwarded arguments
333 // needed to construct a virtual base (or base class thereof).
334 if (auto Inherited = CD->getInheritedConstructor())
335 PassParams = inheritingCtorHasParams(Inherited, GD.getCtorType());
336 }
337
338 CanQual<FunctionProtoType> FTP = GetFormalType(MD);
339
340 // Add the formal parameters.
341 if (PassParams)
342 appendParameterTypes(*this, argTypes, paramInfos, FTP);
343
344 CGCXXABI::AddedStructorArgCounts AddedArgs =
345 TheCXXABI.buildStructorSignature(GD, argTypes);
346 if (!paramInfos.empty()) {
347 // Note: prefix implies after the first param.
348 if (AddedArgs.Prefix)
349 paramInfos.insert(paramInfos.begin() + 1, AddedArgs.Prefix,
350 FunctionProtoType::ExtParameterInfo{});
351 if (AddedArgs.Suffix)
352 paramInfos.append(AddedArgs.Suffix,
353 FunctionProtoType::ExtParameterInfo{});
354 }
355
356 RequiredArgs required =
357 (PassParams && MD->isVariadic() ? RequiredArgs(argTypes.size())
358 : RequiredArgs::All);
359
360 FunctionType::ExtInfo extInfo = FTP->getExtInfo();
361 CanQualType resultType = TheCXXABI.HasThisReturn(GD)
362 ? argTypes.front()
363 : TheCXXABI.hasMostDerivedReturn(GD)
364 ? CGM.getContext().VoidPtrTy
365 : Context.VoidTy;
366 return arrangeLLVMFunctionInfo(resultType, /*instanceMethod=*/true,
367 /*chainCall=*/false, argTypes, extInfo,
368 paramInfos, required);
369 }
370
371 static SmallVector<CanQualType, 16>
372 getArgTypesForCall(ASTContext &ctx, const CallArgList &args) {
373 SmallVector<CanQualType, 16> argTypes;
374 for (auto &arg : args)
375 argTypes.push_back(ctx.getCanonicalParamType(arg.Ty));
376 return argTypes;
377 }
378
379 static SmallVector<CanQualType, 16>
380 getArgTypesForDeclaration(ASTContext &ctx, const FunctionArgList &args) {
381 SmallVector<CanQualType, 16> argTypes;
382 for (auto &arg : args)
383 argTypes.push_back(ctx.getCanonicalParamType(arg->getType()));
384 return argTypes;
385 }
386
387 static llvm::SmallVector<FunctionProtoType::ExtParameterInfo, 16>
388 getExtParameterInfosForCall(const FunctionProtoType *proto,
389 unsigned prefixArgs, unsigned totalArgs) {
390 llvm::SmallVector<FunctionProtoType::ExtParameterInfo, 16> result;
391 if (proto->hasExtParameterInfos()) {
392 addExtParameterInfosForCall(result, proto, prefixArgs, totalArgs);
393 }
394 return result;
395 }
396
397 /// Arrange a call to a C++ method, passing the given arguments.
398 ///
399 /// ExtraPrefixArgs is the number of ABI-specific args passed after the `this`
400 /// parameter.
401 /// ExtraSuffixArgs is the number of ABI-specific args passed at the end of
402 /// args.
403 /// PassProtoArgs indicates whether `args` has args for the parameters in the
404 /// given CXXConstructorDecl.
405 const CGFunctionInfo &
406 CodeGenTypes::arrangeCXXConstructorCall(const CallArgList &args,
407 const CXXConstructorDecl *D,
408 CXXCtorType CtorKind,
409 unsigned ExtraPrefixArgs,
410 unsigned ExtraSuffixArgs,
411 bool PassProtoArgs) {
412 // FIXME: Kill copy.
413 SmallVector<CanQualType, 16> ArgTypes;
414 for (const auto &Arg : args)
415 ArgTypes.push_back(Context.getCanonicalParamType(Arg.Ty));
416
417 // +1 for implicit this, which should always be args[0].
418 unsigned TotalPrefixArgs = 1 + ExtraPrefixArgs;
419
420 CanQual<FunctionProtoType> FPT = GetFormalType(D);
421 RequiredArgs Required = PassProtoArgs
422 ? RequiredArgs::forPrototypePlus(
423 FPT, TotalPrefixArgs + ExtraSuffixArgs)
424 : RequiredArgs::All;
425
426 GlobalDecl GD(D, CtorKind);
427 CanQualType ResultType = TheCXXABI.HasThisReturn(GD)
428 ? ArgTypes.front()
429 : TheCXXABI.hasMostDerivedReturn(GD)
430 ? CGM.getContext().VoidPtrTy
431 : Context.VoidTy;
432
433 FunctionType::ExtInfo Info = FPT->getExtInfo();
434 llvm::SmallVector<FunctionProtoType::ExtParameterInfo, 16> ParamInfos;
435 // If the prototype args are elided, we should only have ABI-specific args,
436 // which never have param info.
437 if (PassProtoArgs && FPT->hasExtParameterInfos()) {
438 // ABI-specific suffix arguments are treated the same as variadic arguments.
439 addExtParameterInfosForCall(ParamInfos, FPT.getTypePtr(), TotalPrefixArgs,
440 ArgTypes.size());
441 }
442 return arrangeLLVMFunctionInfo(ResultType, /*instanceMethod=*/true,
443 /*chainCall=*/false, ArgTypes, Info,
444 ParamInfos, Required);
445 }
446
447 /// Arrange the argument and result information for the declaration or
448 /// definition of the given function.
449 const CGFunctionInfo &
450 CodeGenTypes::arrangeFunctionDeclaration(const FunctionDecl *FD) {
451 if (const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(FD))
452 if (MD->isInstance())
453 return arrangeCXXMethodDeclaration(MD);
454
455 CanQualType FTy = FD->getType()->getCanonicalTypeUnqualified();
456
457 assert(isa<FunctionType>(FTy));
458 setCUDAKernelCallingConvention(FTy, CGM, FD);
459
460 // When declaring a function without a prototype, always use a
461 // non-variadic type.
462 if (CanQual<FunctionNoProtoType> noProto = FTy.getAs<FunctionNoProtoType>()) {
463 return arrangeLLVMFunctionInfo(
464 noProto->getReturnType(), /*instanceMethod=*/false,
465 /*chainCall=*/false, std::nullopt, noProto->getExtInfo(), {},
466 RequiredArgs::All);
467 }
468
469 return arrangeFreeFunctionType(FTy.castAs<FunctionProtoType>());
470 }
471
472 /// Arrange the argument and result information for the declaration or
473 /// definition of an Objective-C method.
474 const CGFunctionInfo &
475 CodeGenTypes::arrangeObjCMethodDeclaration(const ObjCMethodDecl *MD) {
476 // It happens that this is the same as a call with no optional
477 // arguments, except also using the formal 'self' type.
478 return arrangeObjCMessageSendSignature(MD, MD->getSelfDecl()->getType());
479 }
480
481 /// Arrange the argument and result information for the function type
482 /// through which to perform a send to the given Objective-C method,
483 /// using the given receiver type. The receiver type is not always
484 /// the 'self' type of the method or even an Objective-C pointer type.
485 /// This is *not* the right method for actually performing such a
486 /// message send, due to the possibility of optional arguments.
487 const CGFunctionInfo &
488 CodeGenTypes::arrangeObjCMessageSendSignature(const ObjCMethodDecl *MD,
489 QualType receiverType) {
490 SmallVector<CanQualType, 16> argTys;
491 SmallVector<FunctionProtoType::ExtParameterInfo, 4> extParamInfos(
492 MD->isDirectMethod() ? 1 : 2);
493 argTys.push_back(Context.getCanonicalParamType(receiverType));
494 if (!MD->isDirectMethod())
495 argTys.push_back(Context.getCanonicalParamType(Context.getObjCSelType()));
496 // FIXME: Kill copy?
497 for (const auto *I : MD->parameters()) {
498 argTys.push_back(Context.getCanonicalParamType(I->getType()));
499 auto extParamInfo = FunctionProtoType::ExtParameterInfo().withIsNoEscape(
500 I->hasAttr<NoEscapeAttr>());
501 extParamInfos.push_back(extParamInfo);
502 }
503
504 FunctionType::ExtInfo einfo;
505 bool IsWindows = getContext().getTargetInfo().getTriple().isOSWindows();
506 einfo = einfo.withCallingConv(getCallingConventionForDecl(MD, IsWindows));
507
508 if (getContext().getLangOpts().ObjCAutoRefCount &&
509 MD->hasAttr<NSReturnsRetainedAttr>())
510 einfo = einfo.withProducesResult(true);
511
512 RequiredArgs required =
513 (MD->isVariadic() ? RequiredArgs(argTys.size()) : RequiredArgs::All);
514
515 return arrangeLLVMFunctionInfo(
516 GetReturnType(MD->getReturnType()), /*instanceMethod=*/false,
517 /*chainCall=*/false, argTys, einfo, extParamInfos, required);
518 }
519
520 const CGFunctionInfo &
521 CodeGenTypes::arrangeUnprototypedObjCMessageSend(QualType returnType,
522 const CallArgList &args) {
523 auto argTypes = getArgTypesForCall(Context, args);
524 FunctionType::ExtInfo einfo;
525
526 return arrangeLLVMFunctionInfo(
527 GetReturnType(returnType), /*instanceMethod=*/false,
528 /*chainCall=*/false, argTypes, einfo, {}, RequiredArgs::All);
529 }
530
531 const CGFunctionInfo &
532 CodeGenTypes::arrangeGlobalDeclaration(GlobalDecl GD) {
533 // FIXME: Do we need to handle ObjCMethodDecl?
534 const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl());
535
536 if (isa<CXXConstructorDecl>(GD.getDecl()) ||
537 isa<CXXDestructorDecl>(GD.getDecl()))
538 return arrangeCXXStructorDeclaration(GD);
539
540 return arrangeFunctionDeclaration(FD);
541 }
542
543 /// Arrange a thunk that takes 'this' as the first parameter followed by
544 /// varargs. Return a void pointer, regardless of the actual return type.
545 /// The body of the thunk will end in a musttail call to a function of the
546 /// correct type, and the caller will bitcast the function to the correct
547 /// prototype.
548 const CGFunctionInfo &
549 CodeGenTypes::arrangeUnprototypedMustTailThunk(const CXXMethodDecl *MD) {
550 assert(MD->isVirtual() && "only methods have thunks");
551 CanQual<FunctionProtoType> FTP = GetFormalType(MD);
552 CanQualType ArgTys[] = {DeriveThisType(MD->getParent(), MD)};
553 return arrangeLLVMFunctionInfo(Context.VoidTy, /*instanceMethod=*/false,
554 /*chainCall=*/false, ArgTys,
555 FTP->getExtInfo(), {}, RequiredArgs(1));
556 }
557
558 const CGFunctionInfo &
559 CodeGenTypes::arrangeMSCtorClosure(const CXXConstructorDecl *CD,
560 CXXCtorType CT) {
561 assert(CT == Ctor_CopyingClosure || CT == Ctor_DefaultClosure);
562
563 CanQual<FunctionProtoType> FTP = GetFormalType(CD);
564 SmallVector<CanQualType, 2> ArgTys;
565 const CXXRecordDecl *RD = CD->getParent();
566 ArgTys.push_back(DeriveThisType(RD, CD));
567 if (CT == Ctor_CopyingClosure)
568 ArgTys.push_back(*FTP->param_type_begin());
569 if (RD->getNumVBases() > 0)
570 ArgTys.push_back(Context.IntTy);
571 CallingConv CC = Context.getDefaultCallingConvention(
572 /*IsVariadic=*/false, /*IsCXXMethod=*/true);
573 return arrangeLLVMFunctionInfo(Context.VoidTy, /*instanceMethod=*/true,
574 /*chainCall=*/false, ArgTys,
575 FunctionType::ExtInfo(CC), {},
576 RequiredArgs::All);
577 }
578
579 /// Arrange a call as unto a free function, except possibly with an
580 /// additional number of formal parameters considered required.
581 static const CGFunctionInfo &
582 arrangeFreeFunctionLikeCall(CodeGenTypes &CGT,
583 CodeGenModule &CGM,
584 const CallArgList &args,
585 const FunctionType *fnType,
586 unsigned numExtraRequiredArgs,
587 bool chainCall) {
588 assert(args.size() >= numExtraRequiredArgs);
589
590 llvm::SmallVector<FunctionProtoType::ExtParameterInfo, 16> paramInfos;
591
592 // In most cases, there are no optional arguments.
593 RequiredArgs required = RequiredArgs::All;
594
595 // If we have a variadic prototype, the required arguments are the
596 // extra prefix plus the arguments in the prototype.
597 if (const FunctionProtoType *proto = dyn_cast<FunctionProtoType>(fnType)) {
598 if (proto->isVariadic())
599 required = RequiredArgs::forPrototypePlus(proto, numExtraRequiredArgs);
600
601 if (proto->hasExtParameterInfos())
602 addExtParameterInfosForCall(paramInfos, proto, numExtraRequiredArgs,
603 args.size());
604
605 // If we don't have a prototype at all, but we're supposed to
606 // explicitly use the variadic convention for unprototyped calls,
607 // treat all of the arguments as required but preserve the nominal
608 // possibility of variadics.
609 } else if (CGM.getTargetCodeGenInfo()
610 .isNoProtoCallVariadic(args,
611 cast<FunctionNoProtoType>(fnType))) {
612 required = RequiredArgs(args.size());
613 }
614
615 // FIXME: Kill copy.
616 SmallVector<CanQualType, 16> argTypes;
617 for (const auto &arg : args)
618 argTypes.push_back(CGT.getContext().getCanonicalParamType(arg.Ty));
619 return CGT.arrangeLLVMFunctionInfo(GetReturnType(fnType->getReturnType()),
620 /*instanceMethod=*/false, chainCall,
621 argTypes, fnType->getExtInfo(), paramInfos,
622 required);
623 }
624
625 /// Figure out the rules for calling a function with the given formal
626 /// type using the given arguments. The arguments are necessary
627 /// because the function might be unprototyped, in which case it's
628 /// target-dependent in crazy ways.
629 const CGFunctionInfo &
630 CodeGenTypes::arrangeFreeFunctionCall(const CallArgList &args,
631 const FunctionType *fnType,
632 bool chainCall) {
633 return arrangeFreeFunctionLikeCall(*this, CGM, args, fnType,
634 chainCall ? 1 : 0, chainCall);
635 }
636
637 /// A block function is essentially a free function with an
638 /// extra implicit argument.
639 const CGFunctionInfo &
640 CodeGenTypes::arrangeBlockFunctionCall(const CallArgList &args,
641 const FunctionType *fnType) {
642 return arrangeFreeFunctionLikeCall(*this, CGM, args, fnType, 1,
643 /*chainCall=*/false);
644 }
645
646 const CGFunctionInfo &
647 CodeGenTypes::arrangeBlockFunctionDeclaration(const FunctionProtoType *proto,
648 const FunctionArgList &params) {
649 auto paramInfos = getExtParameterInfosForCall(proto, 1, params.size());
650 auto argTypes = getArgTypesForDeclaration(Context, params);
651
652 return arrangeLLVMFunctionInfo(GetReturnType(proto->getReturnType()),
653 /*instanceMethod*/ false, /*chainCall*/ false,
654 argTypes, proto->getExtInfo(), paramInfos,
655 RequiredArgs::forPrototypePlus(proto, 1));
656 }
657
658 const CGFunctionInfo &
659 CodeGenTypes::arrangeBuiltinFunctionCall(QualType resultType,
660 const CallArgList &args) {
661 // FIXME: Kill copy.
662 SmallVector<CanQualType, 16> argTypes;
663 for (const auto &Arg : args)
664 argTypes.push_back(Context.getCanonicalParamType(Arg.Ty));
665 return arrangeLLVMFunctionInfo(
666 GetReturnType(resultType), /*instanceMethod=*/false,
667 /*chainCall=*/false, argTypes, FunctionType::ExtInfo(),
668 /*paramInfos=*/ {}, RequiredArgs::All);
669 }
670
671 const CGFunctionInfo &
672 CodeGenTypes::arrangeBuiltinFunctionDeclaration(QualType resultType,
673 const FunctionArgList &args) {
674 auto argTypes = getArgTypesForDeclaration(Context, args);
675
676 return arrangeLLVMFunctionInfo(
677 GetReturnType(resultType), /*instanceMethod=*/false, /*chainCall=*/false,
678 argTypes, FunctionType::ExtInfo(), {}, RequiredArgs::All);
679 }
680
681 const CGFunctionInfo &
682 CodeGenTypes::arrangeBuiltinFunctionDeclaration(CanQualType resultType,
683 ArrayRef<CanQualType> argTypes) {
684 return arrangeLLVMFunctionInfo(
685 resultType, /*instanceMethod=*/false, /*chainCall=*/false,
686 argTypes, FunctionType::ExtInfo(), {}, RequiredArgs::All);
687 }
688
689 /// Arrange a call to a C++ method, passing the given arguments.
690 ///
691 /// numPrefixArgs is the number of ABI-specific prefix arguments we have. It
692 /// does not count `this`.
693 const CGFunctionInfo &
694 CodeGenTypes::arrangeCXXMethodCall(const CallArgList &args,
695 const FunctionProtoType *proto,
696 RequiredArgs required,
697 unsigned numPrefixArgs) {
698 assert(numPrefixArgs + 1 <= args.size() &&
699 "Emitting a call with less args than the required prefix?");
700 // Add one to account for `this`. It's a bit awkward here, but we don't count
701 // `this` in similar places elsewhere.
702 auto paramInfos =
703 getExtParameterInfosForCall(proto, numPrefixArgs + 1, args.size());
704
705 // FIXME: Kill copy.
706 auto argTypes = getArgTypesForCall(Context, args);
707
708 FunctionType::ExtInfo info = proto->getExtInfo();
709 return arrangeLLVMFunctionInfo(
710 GetReturnType(proto->getReturnType()), /*instanceMethod=*/true,
711 /*chainCall=*/false, argTypes, info, paramInfos, required);
712 }
713
714 const CGFunctionInfo &CodeGenTypes::arrangeNullaryFunction() {
715 return arrangeLLVMFunctionInfo(
716 getContext().VoidTy, /*instanceMethod=*/false, /*chainCall=*/false,
717 std::nullopt, FunctionType::ExtInfo(), {}, RequiredArgs::All);
718 }
719
720 const CGFunctionInfo &
721 CodeGenTypes::arrangeCall(const CGFunctionInfo &signature,
722 const CallArgList &args) {
723 assert(signature.arg_size() <= args.size());
724 if (signature.arg_size() == args.size())
725 return signature;
726
727 SmallVector<FunctionProtoType::ExtParameterInfo, 16> paramInfos;
728 auto sigParamInfos = signature.getExtParameterInfos();
729 if (!sigParamInfos.empty()) {
730 paramInfos.append(sigParamInfos.begin(), sigParamInfos.end());
731 paramInfos.resize(args.size());
732 }
733
734 auto argTypes = getArgTypesForCall(Context, args);
735
736 assert(signature.getRequiredArgs().allowsOptionalArgs());
737 return arrangeLLVMFunctionInfo(signature.getReturnType(),
738 signature.isInstanceMethod(),
739 signature.isChainCall(),
740 argTypes,
741 signature.getExtInfo(),
742 paramInfos,
743 signature.getRequiredArgs());
744 }
745
746 namespace clang {
747 namespace CodeGen {
748 void computeSPIRKernelABIInfo(CodeGenModule &CGM, CGFunctionInfo &FI);
749 }
750 }
751
752 /// Arrange the argument and result information for an abstract value
753 /// of a given function type. This is the method which all of the
754 /// above functions ultimately defer to.
755 const CGFunctionInfo &
756 CodeGenTypes::arrangeLLVMFunctionInfo(CanQualType resultType,
757 bool instanceMethod,
758 bool chainCall,
759 ArrayRef<CanQualType> argTypes,
760 FunctionType::ExtInfo info,
761 ArrayRef<FunctionProtoType::ExtParameterInfo> paramInfos,
762 RequiredArgs required) {
763 assert(llvm::all_of(argTypes,
764 [](CanQualType T) { return T.isCanonicalAsParam(); }));
765
766 // Lookup or create unique function info.
767 llvm::FoldingSetNodeID ID;
768 CGFunctionInfo::Profile(ID, instanceMethod, chainCall, info, paramInfos,
769 required, resultType, argTypes);
770
771 void *insertPos = nullptr;
772 CGFunctionInfo *FI = FunctionInfos.FindNodeOrInsertPos(ID, insertPos);
773 if (FI)
774 return *FI;
775
776 unsigned CC = ClangCallConvToLLVMCallConv(info.getCC());
777
778 // Construct the function info. We co-allocate the ArgInfos.
779 FI = CGFunctionInfo::create(CC, instanceMethod, chainCall, info,
780 paramInfos, resultType, argTypes, required);
781 FunctionInfos.InsertNode(FI, insertPos);
782
783 bool inserted = FunctionsBeingProcessed.insert(FI).second;
784 (void)inserted;
785 assert(inserted && "Recursively being processed?");
786
787 // Compute ABI information.
788 if (CC == llvm::CallingConv::SPIR_KERNEL) {
789 // Force target independent argument handling for the host visible
790 // kernel functions.
791 computeSPIRKernelABIInfo(CGM, *FI);
792 } else if (info.getCC() == CC_Swift || info.getCC() == CC_SwiftAsync) {
793 swiftcall::computeABIInfo(CGM, *FI);
794 } else {
795 getABIInfo().computeInfo(*FI);
796 }
797
798 // Loop over all of the computed argument and return value info. If any of
799 // them are direct or extend without a specified coerce type, specify the
800 // default now.
801 ABIArgInfo &retInfo = FI->getReturnInfo();
802 if (retInfo.canHaveCoerceToType() && retInfo.getCoerceToType() == nullptr)
803 retInfo.setCoerceToType(ConvertType(FI->getReturnType()));
804
805 for (auto &I : FI->arguments())
806 if (I.info.canHaveCoerceToType() && I.info.getCoerceToType() == nullptr)
807 I.info.setCoerceToType(ConvertType(I.type));
808
809 bool erased = FunctionsBeingProcessed.erase(FI); (void)erased;
810 assert(erased && "Not in set?");
811
812 return *FI;
813 }
814
815 CGFunctionInfo *CGFunctionInfo::create(unsigned llvmCC,
816 bool instanceMethod,
817 bool chainCall,
818 const FunctionType::ExtInfo &info,
819 ArrayRef<ExtParameterInfo> paramInfos,
820 CanQualType resultType,
821 ArrayRef<CanQualType> argTypes,
822 RequiredArgs required) {
823 assert(paramInfos.empty() || paramInfos.size() == argTypes.size());
824 assert(!required.allowsOptionalArgs() ||
825 required.getNumRequiredArgs() <= argTypes.size());
826
827 void *buffer =
828 operator new(totalSizeToAlloc<ArgInfo, ExtParameterInfo>(
829 argTypes.size() + 1, paramInfos.size()));
830
831 CGFunctionInfo *FI = new(buffer) CGFunctionInfo();
832 FI->CallingConvention = llvmCC;
833 FI->EffectiveCallingConvention = llvmCC;
834 FI->ASTCallingConvention = info.getCC();
835 FI->InstanceMethod = instanceMethod;
836 FI->ChainCall = chainCall;
837 FI->CmseNSCall = info.getCmseNSCall();
838 FI->NoReturn = info.getNoReturn();
839 FI->ReturnsRetained = info.getProducesResult();
840 FI->NoCallerSavedRegs = info.getNoCallerSavedRegs();
841 FI->NoCfCheck = info.getNoCfCheck();
842 FI->Required = required;
843 FI->HasRegParm = info.getHasRegParm();
844 FI->RegParm = info.getRegParm();
845 FI->ArgStruct = nullptr;
846 FI->ArgStructAlign = 0;
847 FI->NumArgs = argTypes.size();
848 FI->HasExtParameterInfos = !paramInfos.empty();
849 FI->getArgsBuffer()[0].type = resultType;
850 FI->MaxVectorWidth = 0;
851 for (unsigned i = 0, e = argTypes.size(); i != e; ++i)
852 FI->getArgsBuffer()[i + 1].type = argTypes[i];
853 for (unsigned i = 0, e = paramInfos.size(); i != e; ++i)
854 FI->getExtParameterInfosBuffer()[i] = paramInfos[i];
855 return FI;
856 }
857
858 /***/
859
860 namespace {
861 // ABIArgInfo::Expand implementation.
862
863 // Specifies the way QualType passed as ABIArgInfo::Expand is expanded.
864 struct TypeExpansion {
865 enum TypeExpansionKind {
866 // Elements of constant arrays are expanded recursively.
867 TEK_ConstantArray,
868 // Record fields are expanded recursively (but if record is a union, only
869 // the field with the largest size is expanded).
870 TEK_Record,
871 // For complex types, real and imaginary parts are expanded recursively.
872 TEK_Complex,
873 // All other types are not expandable.
874 TEK_None
875 };
876
877 const TypeExpansionKind Kind;
878
879 TypeExpansion(TypeExpansionKind K) : Kind(K) {}
880 virtual ~TypeExpansion() {}
881 };
882
883 struct ConstantArrayExpansion : TypeExpansion {
884 QualType EltTy;
885 uint64_t NumElts;
886
887 ConstantArrayExpansion(QualType EltTy, uint64_t NumElts)
888 : TypeExpansion(TEK_ConstantArray), EltTy(EltTy), NumElts(NumElts) {}
889 static bool classof(const TypeExpansion *TE) {
890 return TE->Kind == TEK_ConstantArray;
891 }
892 };
893
894 struct RecordExpansion : TypeExpansion {
895 SmallVector<const CXXBaseSpecifier *, 1> Bases;
896
897 SmallVector<const FieldDecl *, 1> Fields;
898
899 RecordExpansion(SmallVector<const CXXBaseSpecifier *, 1> &&Bases,
900 SmallVector<const FieldDecl *, 1> &&Fields)
901 : TypeExpansion(TEK_Record), Bases(std::move(Bases)),
902 Fields(std::move(Fields)) {}
903 static bool classof(const TypeExpansion *TE) {
904 return TE->Kind == TEK_Record;
905 }
906 };
907
908 struct ComplexExpansion : TypeExpansion {
909 QualType EltTy;
910
911 ComplexExpansion(QualType EltTy) : TypeExpansion(TEK_Complex), EltTy(EltTy) {}
912 static bool classof(const TypeExpansion *TE) {
913 return TE->Kind == TEK_Complex;
914 }
915 };
916
917 struct NoExpansion : TypeExpansion {
918 NoExpansion() : TypeExpansion(TEK_None) {}
919 static bool classof(const TypeExpansion *TE) {
920 return TE->Kind == TEK_None;
921 }
922 };
923 } // namespace
924
925 static std::unique_ptr<TypeExpansion>
926 getTypeExpansion(QualType Ty, const ASTContext &Context) {
927 if (const ConstantArrayType *AT = Context.getAsConstantArrayType(Ty)) {
928 return std::make_unique<ConstantArrayExpansion>(
929 AT->getElementType(), AT->getSize().getZExtValue());
930 }
931 if (const RecordType *RT = Ty->getAs<RecordType>()) {
932 SmallVector<const CXXBaseSpecifier *, 1> Bases;
933 SmallVector<const FieldDecl *, 1> Fields;
934 const RecordDecl *RD = RT->getDecl();
935 assert(!RD->hasFlexibleArrayMember() &&
936 "Cannot expand structure with flexible array.");
937 if (RD->isUnion()) {
938 // Unions can be here only in degenerative cases - all the fields are same
939 // after flattening. Thus we have to use the "largest" field.
940 const FieldDecl *LargestFD = nullptr;
941 CharUnits UnionSize = CharUnits::Zero();
942
943 for (const auto *FD : RD->fields()) {
944 if (FD->isZeroLengthBitField(Context))
945 continue;
946 assert(!FD->isBitField() &&
947 "Cannot expand structure with bit-field members.");
948 CharUnits FieldSize = Context.getTypeSizeInChars(FD->getType());
949 if (UnionSize < FieldSize) {
950 UnionSize = FieldSize;
951 LargestFD = FD;
952 }
953 }
954 if (LargestFD)
955 Fields.push_back(LargestFD);
956 } else {
957 if (const auto *CXXRD = dyn_cast<CXXRecordDecl>(RD)) {
958 assert(!CXXRD->isDynamicClass() &&
959 "cannot expand vtable pointers in dynamic classes");
960 llvm::append_range(Bases, llvm::make_pointer_range(CXXRD->bases()));
961 }
962
963 for (const auto *FD : RD->fields()) {
964 if (FD->isZeroLengthBitField(Context))
965 continue;
966 assert(!FD->isBitField() &&
967 "Cannot expand structure with bit-field members.");
968 Fields.push_back(FD);
969 }
970 }
971 return std::make_unique<RecordExpansion>(std::move(Bases),
972 std::move(Fields));
973 }
974 if (const ComplexType *CT = Ty->getAs<ComplexType>()) {
975 return std::make_unique<ComplexExpansion>(CT->getElementType());
976 }
977 return std::make_unique<NoExpansion>();
978 }
979
980 static int getExpansionSize(QualType Ty, const ASTContext &Context) {
981 auto Exp = getTypeExpansion(Ty, Context);
982 if (auto CAExp = dyn_cast<ConstantArrayExpansion>(Exp.get())) {
983 return CAExp->NumElts * getExpansionSize(CAExp->EltTy, Context);
984 }
985 if (auto RExp = dyn_cast<RecordExpansion>(Exp.get())) {
986 int Res = 0;
987 for (auto BS : RExp->Bases)
988 Res += getExpansionSize(BS->getType(), Context);
989 for (auto FD : RExp->Fields)
990 Res += getExpansionSize(FD->getType(), Context);
991 return Res;
992 }
993 if (isa<ComplexExpansion>(Exp.get()))
994 return 2;
995 assert(isa<NoExpansion>(Exp.get()));
996 return 1;
997 }
998
999 void
1000 CodeGenTypes::getExpandedTypes(QualType Ty,
1001 SmallVectorImpl<llvm::Type *>::iterator &TI) {
1002 auto Exp = getTypeExpansion(Ty, Context);
1003 if (auto CAExp = dyn_cast<ConstantArrayExpansion>(Exp.get())) {
1004 for (int i = 0, n = CAExp->NumElts; i < n; i++) {
1005 getExpandedTypes(CAExp->EltTy, TI);
1006 }
1007 } else if (auto RExp = dyn_cast<RecordExpansion>(Exp.get())) {
1008 for (auto BS : RExp->Bases)
1009 getExpandedTypes(BS->getType(), TI);
1010 for (auto FD : RExp->Fields)
1011 getExpandedTypes(FD->getType(), TI);
1012 } else if (auto CExp = dyn_cast<ComplexExpansion>(Exp.get())) {
1013 llvm::Type *EltTy = ConvertType(CExp->EltTy);
1014 *TI++ = EltTy;
1015 *TI++ = EltTy;
1016 } else {
1017 assert(isa<NoExpansion>(Exp.get()));
1018 *TI++ = ConvertType(Ty);
1019 }
1020 }
1021
1022 static void forConstantArrayExpansion(CodeGenFunction &CGF,
1023 ConstantArrayExpansion *CAE,
1024 Address BaseAddr,
1025 llvm::function_ref<void(Address)> Fn) {
1026 CharUnits EltSize = CGF.getContext().getTypeSizeInChars(CAE->EltTy);
1027 CharUnits EltAlign =
1028 BaseAddr.getAlignment().alignmentOfArrayElement(EltSize);
1029 llvm::Type *EltTy = CGF.ConvertTypeForMem(CAE->EltTy);
1030
1031 for (int i = 0, n = CAE->NumElts; i < n; i++) {
1032 llvm::Value *EltAddr = CGF.Builder.CreateConstGEP2_32(
1033 BaseAddr.getElementType(), BaseAddr.getPointer(), 0, i);
1034 Fn(Address(EltAddr, EltTy, EltAlign));
1035 }
1036 }
1037
1038 void CodeGenFunction::ExpandTypeFromArgs(QualType Ty, LValue LV,
1039 llvm::Function::arg_iterator &AI) {
1040 assert(LV.isSimple() &&
1041 "Unexpected non-simple lvalue during struct expansion.");
1042
1043 auto Exp = getTypeExpansion(Ty, getContext());
1044 if (auto CAExp = dyn_cast<ConstantArrayExpansion>(Exp.get())) {
1045 forConstantArrayExpansion(
1046 *this, CAExp, LV.getAddress(*this), [&](Address EltAddr) {
1047 LValue LV = MakeAddrLValue(EltAddr, CAExp->EltTy);
1048 ExpandTypeFromArgs(CAExp->EltTy, LV, AI);
1049 });
1050 } else if (auto RExp = dyn_cast<RecordExpansion>(Exp.get())) {
1051 Address This = LV.getAddress(*this);
1052 for (const CXXBaseSpecifier *BS : RExp->Bases) {
1053 // Perform a single step derived-to-base conversion.
1054 Address Base =
1055 GetAddressOfBaseClass(This, Ty->getAsCXXRecordDecl(), &BS, &BS + 1,
1056 /*NullCheckValue=*/false, SourceLocation());
1057 LValue SubLV = MakeAddrLValue(Base, BS->getType());
1058
1059 // Recurse onto bases.
1060 ExpandTypeFromArgs(BS->getType(), SubLV, AI);
1061 }
1062 for (auto FD : RExp->Fields) {
1063 // FIXME: What are the right qualifiers here?
1064 LValue SubLV = EmitLValueForFieldInitialization(LV, FD);
1065 ExpandTypeFromArgs(FD->getType(), SubLV, AI);
1066 }
1067 } else if (isa<ComplexExpansion>(Exp.get())) {
1068 auto realValue = &*AI++;
1069 auto imagValue = &*AI++;
1070 EmitStoreOfComplex(ComplexPairTy(realValue, imagValue), LV, /*init*/ true);
1071 } else {
1072 // Call EmitStoreOfScalar except when the lvalue is a bitfield to emit a
1073 // primitive store.
1074 assert(isa<NoExpansion>(Exp.get()));
1075 llvm::Value *Arg = &*AI++;
1076 if (LV.isBitField()) {
1077 EmitStoreThroughLValue(RValue::get(Arg), LV);
1078 } else {
1079 // TODO: currently there are some places are inconsistent in what LLVM
1080 // pointer type they use (see D118744). Once clang uses opaque pointers
1081 // all LLVM pointer types will be the same and we can remove this check.
1082 if (Arg->getType()->isPointerTy()) {
1083 Address Addr = LV.getAddress(*this);
1084 Arg = Builder.CreateBitCast(Arg, Addr.getElementType());
1085 }
1086 EmitStoreOfScalar(Arg, LV);
1087 }
1088 }
1089 }
1090
1091 void CodeGenFunction::ExpandTypeToArgs(
1092 QualType Ty, CallArg Arg, llvm::FunctionType *IRFuncTy,
1093 SmallVectorImpl<llvm::Value *> &IRCallArgs, unsigned &IRCallArgPos) {
1094 auto Exp = getTypeExpansion(Ty, getContext());
1095 if (auto CAExp = dyn_cast<ConstantArrayExpansion>(Exp.get())) {
1096 Address Addr = Arg.hasLValue() ? Arg.getKnownLValue().getAddress(*this)
1097 : Arg.getKnownRValue().getAggregateAddress();
1098 forConstantArrayExpansion(
1099 *this, CAExp, Addr, [&](Address EltAddr) {
1100 CallArg EltArg = CallArg(
1101 convertTempToRValue(EltAddr, CAExp->EltTy, SourceLocation()),
1102 CAExp->EltTy);
1103 ExpandTypeToArgs(CAExp->EltTy, EltArg, IRFuncTy, IRCallArgs,
1104 IRCallArgPos);
1105 });
1106 } else if (auto RExp = dyn_cast<RecordExpansion>(Exp.get())) {
1107 Address This = Arg.hasLValue() ? Arg.getKnownLValue().getAddress(*this)
1108 : Arg.getKnownRValue().getAggregateAddress();
1109 for (const CXXBaseSpecifier *BS : RExp->Bases) {
1110 // Perform a single step derived-to-base conversion.
1111 Address Base =
1112 GetAddressOfBaseClass(This, Ty->getAsCXXRecordDecl(), &BS, &BS + 1,
1113 /*NullCheckValue=*/false, SourceLocation());
1114 CallArg BaseArg = CallArg(RValue::getAggregate(Base), BS->getType());
1115
1116 // Recurse onto bases.
1117 ExpandTypeToArgs(BS->getType(), BaseArg, IRFuncTy, IRCallArgs,
1118 IRCallArgPos);
1119 }
1120
1121 LValue LV = MakeAddrLValue(This, Ty);
1122 for (auto FD : RExp->Fields) {
1123 CallArg FldArg =
1124 CallArg(EmitRValueForField(LV, FD, SourceLocation()), FD->getType());
1125 ExpandTypeToArgs(FD->getType(), FldArg, IRFuncTy, IRCallArgs,
1126 IRCallArgPos);
1127 }
1128 } else if (isa<ComplexExpansion>(Exp.get())) {
1129 ComplexPairTy CV = Arg.getKnownRValue().getComplexVal();
1130 IRCallArgs[IRCallArgPos++] = CV.first;
1131 IRCallArgs[IRCallArgPos++] = CV.second;
1132 } else {
1133 assert(isa<NoExpansion>(Exp.get()));
1134 auto RV = Arg.getKnownRValue();
1135 assert(RV.isScalar() &&
1136 "Unexpected non-scalar rvalue during struct expansion.");
1137
1138 // Insert a bitcast as needed.
1139 llvm::Value *V = RV.getScalarVal();
1140 if (IRCallArgPos < IRFuncTy->getNumParams() &&
1141 V->getType() != IRFuncTy->getParamType(IRCallArgPos))
1142 V = Builder.CreateBitCast(V, IRFuncTy->getParamType(IRCallArgPos));
1143
1144 IRCallArgs[IRCallArgPos++] = V;
1145 }
1146 }
1147
1148 /// Create a temporary allocation for the purposes of coercion.
1149 static Address CreateTempAllocaForCoercion(CodeGenFunction &CGF, llvm::Type *Ty,
1150 CharUnits MinAlign,
1151 const Twine &Name = "tmp") {
1152 // Don't use an alignment that's worse than what LLVM would prefer.
1153 auto PrefAlign = CGF.CGM.getDataLayout().getPrefTypeAlign(Ty);
1154 CharUnits Align = std::max(MinAlign, CharUnits::fromQuantity(PrefAlign));
1155
1156 return CGF.CreateTempAlloca(Ty, Align, Name + ".coerce");
1157 }
1158
1159 /// EnterStructPointerForCoercedAccess - Given a struct pointer that we are
1160 /// accessing some number of bytes out of it, try to gep into the struct to get
1161 /// at its inner goodness. Dive as deep as possible without entering an element
1162 /// with an in-memory size smaller than DstSize.
1163 static Address
1164 EnterStructPointerForCoercedAccess(Address SrcPtr,
1165 llvm::StructType *SrcSTy,
1166 uint64_t DstSize, CodeGenFunction &CGF) {
1167 // We can't dive into a zero-element struct.
1168 if (SrcSTy->getNumElements() == 0) return SrcPtr;
1169
1170 llvm::Type *FirstElt = SrcSTy->getElementType(0);
1171
1172 // If the first elt is at least as large as what we're looking for, or if the
1173 // first element is the same size as the whole struct, we can enter it. The
1174 // comparison must be made on the store size and not the alloca size. Using
1175 // the alloca size may overstate the size of the load.
1176 uint64_t FirstEltSize =
1177 CGF.CGM.getDataLayout().getTypeStoreSize(FirstElt);
1178 if (FirstEltSize < DstSize &&
1179 FirstEltSize < CGF.CGM.getDataLayout().getTypeStoreSize(SrcSTy))
1180 return SrcPtr;
1181
1182 // GEP into the first element.
1183 SrcPtr = CGF.Builder.CreateStructGEP(SrcPtr, 0, "coerce.dive");
1184
1185 // If the first element is a struct, recurse.
1186 llvm::Type *SrcTy = SrcPtr.getElementType();
1187 if (llvm::StructType *SrcSTy = dyn_cast<llvm::StructType>(SrcTy))
1188 return EnterStructPointerForCoercedAccess(SrcPtr, SrcSTy, DstSize, CGF);
1189
1190 return SrcPtr;
1191 }
1192
1193 /// CoerceIntOrPtrToIntOrPtr - Convert a value Val to the specific Ty where both
1194 /// are either integers or pointers. This does a truncation of the value if it
1195 /// is too large or a zero extension if it is too small.
1196 ///
1197 /// This behaves as if the value were coerced through memory, so on big-endian
1198 /// targets the high bits are preserved in a truncation, while little-endian
1199 /// targets preserve the low bits.
1200 static llvm::Value *CoerceIntOrPtrToIntOrPtr(llvm::Value *Val,
1201 llvm::Type *Ty,
1202 CodeGenFunction &CGF) {
1203 if (Val->getType() == Ty)
1204 return Val;
1205
1206 if (isa<llvm::PointerType>(Val->getType())) {
1207 // If this is Pointer->Pointer avoid conversion to and from int.
1208 if (isa<llvm::PointerType>(Ty))
1209 return CGF.Builder.CreateBitCast(Val, Ty, "coerce.val");
1210
1211 // Convert the pointer to an integer so we can play with its width.
1212 Val = CGF.Builder.CreatePtrToInt(Val, CGF.IntPtrTy, "coerce.val.pi");
1213 }
1214
1215 llvm::Type *DestIntTy = Ty;
1216 if (isa<llvm::PointerType>(DestIntTy))
1217 DestIntTy = CGF.IntPtrTy;
1218
1219 if (Val->getType() != DestIntTy) {
1220 const llvm::DataLayout &DL = CGF.CGM.getDataLayout();
1221 if (DL.isBigEndian()) {
1222 // Preserve the high bits on big-endian targets.
1223 // That is what memory coercion does.
1224 uint64_t SrcSize = DL.getTypeSizeInBits(Val->getType());
1225 uint64_t DstSize = DL.getTypeSizeInBits(DestIntTy);
1226
1227 if (SrcSize > DstSize) {
1228 Val = CGF.Builder.CreateLShr(Val, SrcSize - DstSize, "coerce.highbits");
1229 Val = CGF.Builder.CreateTrunc(Val, DestIntTy, "coerce.val.ii");
1230 } else {
1231 Val = CGF.Builder.CreateZExt(Val, DestIntTy, "coerce.val.ii");
1232 Val = CGF.Builder.CreateShl(Val, DstSize - SrcSize, "coerce.highbits");
1233 }
1234 } else {
1235 // Little-endian targets preserve the low bits. No shifts required.
1236 Val = CGF.Builder.CreateIntCast(Val, DestIntTy, false, "coerce.val.ii");
1237 }
1238 }
1239
1240 if (isa<llvm::PointerType>(Ty))
1241 Val = CGF.Builder.CreateIntToPtr(Val, Ty, "coerce.val.ip");
1242 return Val;
1243 }
1244
1245
1246
1247 /// CreateCoercedLoad - Create a load from \arg SrcPtr interpreted as
1248 /// a pointer to an object of type \arg Ty, known to be aligned to
1249 /// \arg SrcAlign bytes.
1250 ///
1251 /// This safely handles the case when the src type is smaller than the
1252 /// destination type; in this situation the values of bits which not
1253 /// present in the src are undefined.
1254 static llvm::Value *CreateCoercedLoad(Address Src, llvm::Type *Ty,
1255 CodeGenFunction &CGF) {
1256 llvm::Type *SrcTy = Src.getElementType();
1257
1258 // If SrcTy and Ty are the same, just do a load.
1259 if (SrcTy == Ty)
1260 return CGF.Builder.CreateLoad(Src);
1261
1262 llvm::TypeSize DstSize = CGF.CGM.getDataLayout().getTypeAllocSize(Ty);
1263
1264 if (llvm::StructType *SrcSTy = dyn_cast<llvm::StructType>(SrcTy)) {
1265 Src = EnterStructPointerForCoercedAccess(Src, SrcSTy,
1266 DstSize.getFixedValue(), CGF);
1267 SrcTy = Src.getElementType();
1268 }
1269
1270 llvm::TypeSize SrcSize = CGF.CGM.getDataLayout().getTypeAllocSize(SrcTy);
1271
1272 // If the source and destination are integer or pointer types, just do an
1273 // extension or truncation to the desired type.
1274 if ((isa<llvm::IntegerType>(Ty) || isa<llvm::PointerType>(Ty)) &&
1275 (isa<llvm::IntegerType>(SrcTy) || isa<llvm::PointerType>(SrcTy))) {
1276 llvm::Value *Load = CGF.Builder.CreateLoad(Src);
1277 return CoerceIntOrPtrToIntOrPtr(Load, Ty, CGF);
1278 }
1279
1280 // If load is legal, just bitcast the src pointer.
1281 if (!SrcSize.isScalable() && !DstSize.isScalable() &&
1282 SrcSize.getFixedValue() >= DstSize.getFixedValue()) {
1283 // Generally SrcSize is never greater than DstSize, since this means we are
1284 // losing bits. However, this can happen in cases where the structure has
1285 // additional padding, for example due to a user specified alignment.
1286 //
1287 // FIXME: Assert that we aren't truncating non-padding bits when have access
1288 // to that information.
1289 Src = CGF.Builder.CreateElementBitCast(Src, Ty);
1290 return CGF.Builder.CreateLoad(Src);
1291 }
1292
1293 // If coercing a fixed vector to a scalable vector for ABI compatibility, and
1294 // the types match, use the llvm.vector.insert intrinsic to perform the
1295 // conversion.
1296 if (auto *ScalableDst = dyn_cast<llvm::ScalableVectorType>(Ty)) {
1297 if (auto *FixedSrc = dyn_cast<llvm::FixedVectorType>(SrcTy)) {
1298 // If we are casting a fixed i8 vector to a scalable 16 x i1 predicate
1299 // vector, use a vector insert and bitcast the result.
1300 bool NeedsBitcast = false;
1301 auto PredType =
1302 llvm::ScalableVectorType::get(CGF.Builder.getInt1Ty(), 16);
1303 llvm::Type *OrigType = Ty;
1304 if (ScalableDst == PredType &&
1305 FixedSrc->getElementType() == CGF.Builder.getInt8Ty()) {
1306 ScalableDst = llvm::ScalableVectorType::get(CGF.Builder.getInt8Ty(), 2);
1307 NeedsBitcast = true;
1308 }
1309 if (ScalableDst->getElementType() == FixedSrc->getElementType()) {
1310 auto *Load = CGF.Builder.CreateLoad(Src);
1311 auto *UndefVec = llvm::UndefValue::get(ScalableDst);
1312 auto *Zero = llvm::Constant::getNullValue(CGF.CGM.Int64Ty);
1313 llvm::Value *Result = CGF.Builder.CreateInsertVector(
1314 ScalableDst, UndefVec, Load, Zero, "castScalableSve");
1315 if (NeedsBitcast)
1316 Result = CGF.Builder.CreateBitCast(Result, OrigType);
1317 return Result;
1318 }
1319 }
1320 }
1321
1322 // Otherwise do coercion through memory. This is stupid, but simple.
1323 Address Tmp =
1324 CreateTempAllocaForCoercion(CGF, Ty, Src.getAlignment(), Src.getName());
1325 CGF.Builder.CreateMemCpy(
1326 Tmp.getPointer(), Tmp.getAlignment().getAsAlign(), Src.getPointer(),
1327 Src.getAlignment().getAsAlign(),
1328 llvm::ConstantInt::get(CGF.IntPtrTy, SrcSize.getKnownMinValue()));
1329 return CGF.Builder.CreateLoad(Tmp);
1330 }
1331
1332 // Function to store a first-class aggregate into memory. We prefer to
1333 // store the elements rather than the aggregate to be more friendly to
1334 // fast-isel.
1335 // FIXME: Do we need to recurse here?
1336 void CodeGenFunction::EmitAggregateStore(llvm::Value *Val, Address Dest,
1337 bool DestIsVolatile) {
1338 // Prefer scalar stores to first-class aggregate stores.
1339 if (llvm::StructType *STy = dyn_cast<llvm::StructType>(Val->getType())) {
1340 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1341 Address EltPtr = Builder.CreateStructGEP(Dest, i);
1342 llvm::Value *Elt = Builder.CreateExtractValue(Val, i);
1343 Builder.CreateStore(Elt, EltPtr, DestIsVolatile);
1344 }
1345 } else {
1346 Builder.CreateStore(Val, Dest, DestIsVolatile);
1347 }
1348 }
1349
1350 /// CreateCoercedStore - Create a store to \arg DstPtr from \arg Src,
1351 /// where the source and destination may have different types. The
1352 /// destination is known to be aligned to \arg DstAlign bytes.
1353 ///
1354 /// This safely handles the case when the src type is larger than the
1355 /// destination type; the upper bits of the src will be lost.
1356 static void CreateCoercedStore(llvm::Value *Src,
1357 Address Dst,
1358 bool DstIsVolatile,
1359 CodeGenFunction &CGF) {
1360 llvm::Type *SrcTy = Src->getType();
1361 llvm::Type *DstTy = Dst.getElementType();
1362 if (SrcTy == DstTy) {
1363 CGF.Builder.CreateStore(Src, Dst, DstIsVolatile);
1364 return;
1365 }
1366
1367 llvm::TypeSize SrcSize = CGF.CGM.getDataLayout().getTypeAllocSize(SrcTy);
1368
1369 if (llvm::StructType *DstSTy = dyn_cast<llvm::StructType>(DstTy)) {
1370 Dst = EnterStructPointerForCoercedAccess(Dst, DstSTy,
1371 SrcSize.getFixedValue(), CGF);
1372 DstTy = Dst.getElementType();
1373 }
1374
1375 llvm::PointerType *SrcPtrTy = llvm::dyn_cast<llvm::PointerType>(SrcTy);
1376 llvm::PointerType *DstPtrTy = llvm::dyn_cast<llvm::PointerType>(DstTy);
1377 if (SrcPtrTy && DstPtrTy &&
1378 SrcPtrTy->getAddressSpace() != DstPtrTy->getAddressSpace()) {
1379 Src = CGF.Builder.CreatePointerBitCastOrAddrSpaceCast(Src, DstTy);
1380 CGF.Builder.CreateStore(Src, Dst, DstIsVolatile);
1381 return;
1382 }
1383
1384 // If the source and destination are integer or pointer types, just do an
1385 // extension or truncation to the desired type.
1386 if ((isa<llvm::IntegerType>(SrcTy) || isa<llvm::PointerType>(SrcTy)) &&
1387 (isa<llvm::IntegerType>(DstTy) || isa<llvm::PointerType>(DstTy))) {
1388 Src = CoerceIntOrPtrToIntOrPtr(Src, DstTy, CGF);
1389 CGF.Builder.CreateStore(Src, Dst, DstIsVolatile);
1390 return;
1391 }
1392
1393 llvm::TypeSize DstSize = CGF.CGM.getDataLayout().getTypeAllocSize(DstTy);
1394
1395 // If store is legal, just bitcast the src pointer.
1396 if (isa<llvm::ScalableVectorType>(SrcTy) ||
1397 isa<llvm::ScalableVectorType>(DstTy) ||
1398 SrcSize.getFixedValue() <= DstSize.getFixedValue()) {
1399 Dst = CGF.Builder.CreateElementBitCast(Dst, SrcTy);
1400 CGF.EmitAggregateStore(Src, Dst, DstIsVolatile);
1401 } else {
1402 // Otherwise do coercion through memory. This is stupid, but
1403 // simple.
1404
1405 // Generally SrcSize is never greater than DstSize, since this means we are
1406 // losing bits. However, this can happen in cases where the structure has
1407 // additional padding, for example due to a user specified alignment.
1408 //
1409 // FIXME: Assert that we aren't truncating non-padding bits when have access
1410 // to that information.
1411 Address Tmp = CreateTempAllocaForCoercion(CGF, SrcTy, Dst.getAlignment());
1412 CGF.Builder.CreateStore(Src, Tmp);
1413 CGF.Builder.CreateMemCpy(
1414 Dst.getPointer(), Dst.getAlignment().getAsAlign(), Tmp.getPointer(),
1415 Tmp.getAlignment().getAsAlign(),
1416 llvm::ConstantInt::get(CGF.IntPtrTy, DstSize.getFixedValue()));
1417 }
1418 }
1419
1420 static Address emitAddressAtOffset(CodeGenFunction &CGF, Address addr,
1421 const ABIArgInfo &info) {
1422 if (unsigned offset = info.getDirectOffset()) {
1423 addr = CGF.Builder.CreateElementBitCast(addr, CGF.Int8Ty);
1424 addr = CGF.Builder.CreateConstInBoundsByteGEP(addr,
1425 CharUnits::fromQuantity(offset));
1426 addr = CGF.Builder.CreateElementBitCast(addr, info.getCoerceToType());
1427 }
1428 return addr;
1429 }
1430
1431 namespace {
1432
1433 /// Encapsulates information about the way function arguments from
1434 /// CGFunctionInfo should be passed to actual LLVM IR function.
1435 class ClangToLLVMArgMapping {
1436 static const unsigned InvalidIndex = ~0U;
1437 unsigned InallocaArgNo;
1438 unsigned SRetArgNo;
1439 unsigned TotalIRArgs;
1440
1441 /// Arguments of LLVM IR function corresponding to single Clang argument.
1442 struct IRArgs {
1443 unsigned PaddingArgIndex;
1444 // Argument is expanded to IR arguments at positions
1445 // [FirstArgIndex, FirstArgIndex + NumberOfArgs).
1446 unsigned FirstArgIndex;
1447 unsigned NumberOfArgs;
1448
1449 IRArgs()
1450 : PaddingArgIndex(InvalidIndex), FirstArgIndex(InvalidIndex),
1451 NumberOfArgs(0) {}
1452 };
1453
1454 SmallVector<IRArgs, 8> ArgInfo;
1455
1456 public:
1457 ClangToLLVMArgMapping(const ASTContext &Context, const CGFunctionInfo &FI,
1458 bool OnlyRequiredArgs = false)
1459 : InallocaArgNo(InvalidIndex), SRetArgNo(InvalidIndex), TotalIRArgs(0),
1460 ArgInfo(OnlyRequiredArgs ? FI.getNumRequiredArgs() : FI.arg_size()) {
1461 construct(Context, FI, OnlyRequiredArgs);
1462 }
1463
1464 bool hasInallocaArg() const { return InallocaArgNo != InvalidIndex; }
1465 unsigned getInallocaArgNo() const {
1466 assert(hasInallocaArg());
1467 return InallocaArgNo;
1468 }
1469
1470 bool hasSRetArg() const { return SRetArgNo != InvalidIndex; }
1471 unsigned getSRetArgNo() const {
1472 assert(hasSRetArg());
1473 return SRetArgNo;
1474 }
1475
1476 unsigned totalIRArgs() const { return TotalIRArgs; }
1477
1478 bool hasPaddingArg(unsigned ArgNo) const {
1479 assert(ArgNo < ArgInfo.size());
1480 return ArgInfo[ArgNo].PaddingArgIndex != InvalidIndex;
1481 }
1482 unsigned getPaddingArgNo(unsigned ArgNo) const {
1483 assert(hasPaddingArg(ArgNo));
1484 return ArgInfo[ArgNo].PaddingArgIndex;
1485 }
1486
1487 /// Returns index of first IR argument corresponding to ArgNo, and their
1488 /// quantity.
1489 std::pair<unsigned, unsigned> getIRArgs(unsigned ArgNo) const {
1490 assert(ArgNo < ArgInfo.size());
1491 return std::make_pair(ArgInfo[ArgNo].FirstArgIndex,
1492 ArgInfo[ArgNo].NumberOfArgs);
1493 }
1494
1495 private:
1496 void construct(const ASTContext &Context, const CGFunctionInfo &FI,
1497 bool OnlyRequiredArgs);
1498 };
1499
1500 void ClangToLLVMArgMapping::construct(const ASTContext &Context,
1501 const CGFunctionInfo &FI,
1502 bool OnlyRequiredArgs) {
1503 unsigned IRArgNo = 0;
1504 bool SwapThisWithSRet = false;
1505 const ABIArgInfo &RetAI = FI.getReturnInfo();
1506
1507 if (RetAI.getKind() == ABIArgInfo::Indirect) {
1508 SwapThisWithSRet = RetAI.isSRetAfterThis();
1509 SRetArgNo = SwapThisWithSRet ? 1 : IRArgNo++;
1510 }
1511
1512 unsigned ArgNo = 0;
1513 unsigned NumArgs = OnlyRequiredArgs ? FI.getNumRequiredArgs() : FI.arg_size();
1514 for (CGFunctionInfo::const_arg_iterator I = FI.arg_begin(); ArgNo < NumArgs;
1515 ++I, ++ArgNo) {
1516 assert(I != FI.arg_end());
1517 QualType ArgType = I->type;
1518 const ABIArgInfo &AI = I->info;
1519 // Collect data about IR arguments corresponding to Clang argument ArgNo.
1520 auto &IRArgs = ArgInfo[ArgNo];
1521
1522 if (AI.getPaddingType())
1523 IRArgs.PaddingArgIndex = IRArgNo++;
1524
1525 switch (AI.getKind()) {
1526 case ABIArgInfo::Extend:
1527 case ABIArgInfo::Direct: {
1528 // FIXME: handle sseregparm someday...
1529 llvm::StructType *STy = dyn_cast<llvm::StructType>(AI.getCoerceToType());
1530 if (AI.isDirect() && AI.getCanBeFlattened() && STy) {
1531 IRArgs.NumberOfArgs = STy->getNumElements();
1532 } else {
1533 IRArgs.NumberOfArgs = 1;
1534 }
1535 break;
1536 }
1537 case ABIArgInfo::Indirect:
1538 case ABIArgInfo::IndirectAliased:
1539 IRArgs.NumberOfArgs = 1;
1540 break;
1541 case ABIArgInfo::Ignore:
1542 case ABIArgInfo::InAlloca:
1543 // ignore and inalloca doesn't have matching LLVM parameters.
1544 IRArgs.NumberOfArgs = 0;
1545 break;
1546 case ABIArgInfo::CoerceAndExpand:
1547 IRArgs.NumberOfArgs = AI.getCoerceAndExpandTypeSequence().size();
1548 break;
1549 case ABIArgInfo::Expand:
1550 IRArgs.NumberOfArgs = getExpansionSize(ArgType, Context);
1551 break;
1552 }
1553
1554 if (IRArgs.NumberOfArgs > 0) {
1555 IRArgs.FirstArgIndex = IRArgNo;
1556 IRArgNo += IRArgs.NumberOfArgs;
1557 }
1558
1559 // Skip over the sret parameter when it comes second. We already handled it
1560 // above.
1561 if (IRArgNo == 1 && SwapThisWithSRet)
1562 IRArgNo++;
1563 }
1564 assert(ArgNo == ArgInfo.size());
1565
1566 if (FI.usesInAlloca())
1567 InallocaArgNo = IRArgNo++;
1568
1569 TotalIRArgs = IRArgNo;
1570 }
1571 } // namespace
1572
1573 /***/
1574
1575 bool CodeGenModule::ReturnTypeUsesSRet(const CGFunctionInfo &FI) {
1576 const auto &RI = FI.getReturnInfo();
1577 return RI.isIndirect() || (RI.isInAlloca() && RI.getInAllocaSRet());
1578 }
1579
1580 bool CodeGenModule::ReturnSlotInterferesWithArgs(const CGFunctionInfo &FI) {
1581 return ReturnTypeUsesSRet(FI) &&
1582 getTargetCodeGenInfo().doesReturnSlotInterfereWithArgs();
1583 }
1584
1585 bool CodeGenModule::ReturnTypeUsesFPRet(QualType ResultType) {
1586 if (const BuiltinType *BT = ResultType->getAs<BuiltinType>()) {
1587 switch (BT->getKind()) {
1588 default:
1589 return false;
1590 case BuiltinType::Float:
1591 return getTarget().useObjCFPRetForRealType(FloatModeKind::Float);
1592 case BuiltinType::Double:
1593 return getTarget().useObjCFPRetForRealType(FloatModeKind::Double);
1594 case BuiltinType::LongDouble:
1595 return getTarget().useObjCFPRetForRealType(FloatModeKind::LongDouble);
1596 }
1597 }
1598
1599 return false;
1600 }
1601
1602 bool CodeGenModule::ReturnTypeUsesFP2Ret(QualType ResultType) {
1603 if (const ComplexType *CT = ResultType->getAs<ComplexType>()) {
1604 if (const BuiltinType *BT = CT->getElementType()->getAs<BuiltinType>()) {
1605 if (BT->getKind() == BuiltinType::LongDouble)
1606 return getTarget().useObjCFP2RetForComplexLongDouble();
1607 }
1608 }
1609
1610 return false;
1611 }
1612
1613 llvm::FunctionType *CodeGenTypes::GetFunctionType(GlobalDecl GD) {
1614 const CGFunctionInfo &FI = arrangeGlobalDeclaration(GD);
1615 return GetFunctionType(FI);
1616 }
1617
1618 llvm::FunctionType *
1619 CodeGenTypes::GetFunctionType(const CGFunctionInfo &FI) {
1620
1621 bool Inserted = FunctionsBeingProcessed.insert(&FI).second;
1622 (void)Inserted;
1623 assert(Inserted && "Recursively being processed?");
1624
1625 llvm::Type *resultType = nullptr;
1626 const ABIArgInfo &retAI = FI.getReturnInfo();
1627 switch (retAI.getKind()) {
1628 case ABIArgInfo::Expand:
1629 case ABIArgInfo::IndirectAliased:
1630 llvm_unreachable("Invalid ABI kind for return argument");
1631
1632 case ABIArgInfo::Extend:
1633 case ABIArgInfo::Direct:
1634 resultType = retAI.getCoerceToType();
1635 break;
1636
1637 case ABIArgInfo::InAlloca:
1638 if (retAI.getInAllocaSRet()) {
1639 // sret things on win32 aren't void, they return the sret pointer.
1640 QualType ret = FI.getReturnType();
1641 llvm::Type *ty = ConvertType(ret);
1642 unsigned addressSpace = CGM.getTypes().getTargetAddressSpace(ret);
1643 resultType = llvm::PointerType::get(ty, addressSpace);
1644 } else {
1645 resultType = llvm::Type::getVoidTy(getLLVMContext());
1646 }
1647 break;
1648
1649 case ABIArgInfo::Indirect:
1650 case ABIArgInfo::Ignore:
1651 resultType = llvm::Type::getVoidTy(getLLVMContext());
1652 break;
1653
1654 case ABIArgInfo::CoerceAndExpand:
1655 resultType = retAI.getUnpaddedCoerceAndExpandType();
1656 break;
1657 }
1658
1659 ClangToLLVMArgMapping IRFunctionArgs(getContext(), FI, true);
1660 SmallVector<llvm::Type*, 8> ArgTypes(IRFunctionArgs.totalIRArgs());
1661
1662 // Add type for sret argument.
1663 if (IRFunctionArgs.hasSRetArg()) {
1664 QualType Ret = FI.getReturnType();
1665 llvm::Type *Ty = ConvertType(Ret);
1666 unsigned AddressSpace = CGM.getTypes().getTargetAddressSpace(Ret);
1667 ArgTypes[IRFunctionArgs.getSRetArgNo()] =
1668 llvm::PointerType::get(Ty, AddressSpace);
1669 }
1670
1671 // Add type for inalloca argument.
1672 if (IRFunctionArgs.hasInallocaArg()) {
1673 auto ArgStruct = FI.getArgStruct();
1674 assert(ArgStruct);
1675 ArgTypes[IRFunctionArgs.getInallocaArgNo()] = ArgStruct->getPointerTo();
1676 }
1677
1678 // Add in all of the required arguments.
1679 unsigned ArgNo = 0;
1680 CGFunctionInfo::const_arg_iterator it = FI.arg_begin(),
1681 ie = it + FI.getNumRequiredArgs();
1682 for (; it != ie; ++it, ++ArgNo) {
1683 const ABIArgInfo &ArgInfo = it->info;
1684
1685 // Insert a padding type to ensure proper alignment.
1686 if (IRFunctionArgs.hasPaddingArg(ArgNo))
1687 ArgTypes[IRFunctionArgs.getPaddingArgNo(ArgNo)] =
1688 ArgInfo.getPaddingType();
1689
1690 unsigned FirstIRArg, NumIRArgs;
1691 std::tie(FirstIRArg, NumIRArgs) = IRFunctionArgs.getIRArgs(ArgNo);
1692
1693 switch (ArgInfo.getKind()) {
1694 case ABIArgInfo::Ignore:
1695 case ABIArgInfo::InAlloca:
1696 assert(NumIRArgs == 0);
1697 break;
1698
1699 case ABIArgInfo::Indirect: {
1700 assert(NumIRArgs == 1);
1701 // indirect arguments are always on the stack, which is alloca addr space.
1702 llvm::Type *LTy = ConvertTypeForMem(it->type);
1703 ArgTypes[FirstIRArg] = LTy->getPointerTo(
1704 CGM.getDataLayout().getAllocaAddrSpace());
1705 break;
1706 }
1707 case ABIArgInfo::IndirectAliased: {
1708 assert(NumIRArgs == 1);
1709 llvm::Type *LTy = ConvertTypeForMem(it->type);
1710 ArgTypes[FirstIRArg] = LTy->getPointerTo(ArgInfo.getIndirectAddrSpace());
1711 break;
1712 }
1713 case ABIArgInfo::Extend:
1714 case ABIArgInfo::Direct: {
1715 // Fast-isel and the optimizer generally like scalar values better than
1716 // FCAs, so we flatten them if this is safe to do for this argument.
1717 llvm::Type *argType = ArgInfo.getCoerceToType();
1718 llvm::StructType *st = dyn_cast<llvm::StructType>(argType);
1719 if (st && ArgInfo.isDirect() && ArgInfo.getCanBeFlattened()) {
1720 assert(NumIRArgs == st->getNumElements());
1721 for (unsigned i = 0, e = st->getNumElements(); i != e; ++i)
1722 ArgTypes[FirstIRArg + i] = st->getElementType(i);
1723 } else {
1724 assert(NumIRArgs == 1);
1725 ArgTypes[FirstIRArg] = argType;
1726 }
1727 break;
1728 }
1729
1730 case ABIArgInfo::CoerceAndExpand: {
1731 auto ArgTypesIter = ArgTypes.begin() + FirstIRArg;
1732 for (auto *EltTy : ArgInfo.getCoerceAndExpandTypeSequence()) {
1733 *ArgTypesIter++ = EltTy;
1734 }
1735 assert(ArgTypesIter == ArgTypes.begin() + FirstIRArg + NumIRArgs);
1736 break;
1737 }
1738
1739 case ABIArgInfo::Expand:
1740 auto ArgTypesIter = ArgTypes.begin() + FirstIRArg;
1741 getExpandedTypes(it->type, ArgTypesIter);
1742 assert(ArgTypesIter == ArgTypes.begin() + FirstIRArg + NumIRArgs);
1743 break;
1744 }
1745 }
1746
1747 bool Erased = FunctionsBeingProcessed.erase(&FI); (void)Erased;
1748 assert(Erased && "Not in set?");
1749
1750 return llvm::FunctionType::get(resultType, ArgTypes, FI.isVariadic());
1751 }
1752
1753 llvm::Type *CodeGenTypes::GetFunctionTypeForVTable(GlobalDecl GD) {
1754 const CXXMethodDecl *MD = cast<CXXMethodDecl>(GD.getDecl());
1755 const FunctionProtoType *FPT = MD->getType()->getAs<FunctionProtoType>();
1756
1757 if (!isFuncTypeConvertible(FPT))
1758 return llvm::StructType::get(getLLVMContext());
1759
1760 return GetFunctionType(GD);
1761 }
1762
1763 static void AddAttributesFromFunctionProtoType(ASTContext &Ctx,
1764 llvm::AttrBuilder &FuncAttrs,
1765 const FunctionProtoType *FPT) {
1766 if (!FPT)
1767 return;
1768
1769 if (!isUnresolvedExceptionSpec(FPT->getExceptionSpecType()) &&
1770 FPT->isNothrow())
1771 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1772 }
1773
1774 static void AddAttributesFromAssumes(llvm::AttrBuilder &FuncAttrs,
1775 const Decl *Callee) {
1776 if (!Callee)
1777 return;
1778
1779 SmallVector<StringRef, 4> Attrs;
1780
1781 for (const AssumptionAttr *AA : Callee->specific_attrs<AssumptionAttr>())
1782 AA->getAssumption().split(Attrs, ",");
1783
1784 if (!Attrs.empty())
1785 FuncAttrs.addAttribute(llvm::AssumptionAttrKey,
1786 llvm::join(Attrs.begin(), Attrs.end(), ","));
1787 }
1788
1789 bool CodeGenModule::MayDropFunctionReturn(const ASTContext &Context,
1790 QualType ReturnType) const {
1791 // We can't just discard the return value for a record type with a
1792 // complex destructor or a non-trivially copyable type.
1793 if (const RecordType *RT =
1794 ReturnType.getCanonicalType()->getAs<RecordType>()) {
1795 if (const auto *ClassDecl = dyn_cast<CXXRecordDecl>(RT->getDecl()))
1796 return ClassDecl->hasTrivialDestructor();
1797 }
1798 return ReturnType.isTriviallyCopyableType(Context);
1799 }
1800
1801 static bool HasStrictReturn(const CodeGenModule &Module, QualType RetTy,
1802 const Decl *TargetDecl) {
1803 // As-is msan can not tolerate noundef mismatch between caller and
1804 // implementation. Mismatch is possible for e.g. indirect calls from C-caller
1805 // into C++. Such mismatches lead to confusing false reports. To avoid
1806 // expensive workaround on msan we enforce initialization event in uncommon
1807 // cases where it's allowed.
1808 if (Module.getLangOpts().Sanitize.has(SanitizerKind::Memory))
1809 return true;
1810 // C++ explicitly makes returning undefined values UB. C's rule only applies
1811 // to used values, so we never mark them noundef for now.
1812 if (!Module.getLangOpts().CPlusPlus)
1813 return false;
1814 if (TargetDecl) {
1815 if (const FunctionDecl *FDecl = dyn_cast<FunctionDecl>(TargetDecl)) {
1816 if (FDecl->isExternC())
1817 return false;
1818 } else if (const VarDecl *VDecl = dyn_cast<VarDecl>(TargetDecl)) {
1819 // Function pointer.
1820 if (VDecl->isExternC())
1821 return false;
1822 }
1823 }
1824
1825 // We don't want to be too aggressive with the return checking, unless
1826 // it's explicit in the code opts or we're using an appropriate sanitizer.
1827 // Try to respect what the programmer intended.
1828 return Module.getCodeGenOpts().StrictReturn ||
1829 !Module.MayDropFunctionReturn(Module.getContext(), RetTy) ||
1830 Module.getLangOpts().Sanitize.has(SanitizerKind::Return);
1831 }
1832
1833 void CodeGenModule::getDefaultFunctionAttributes(StringRef Name,
1834 bool HasOptnone,
1835 bool AttrOnCallSite,
1836 llvm::AttrBuilder &FuncAttrs) {
1837 // OptimizeNoneAttr takes precedence over -Os or -Oz. No warning needed.
1838 if (!HasOptnone) {
1839 if (CodeGenOpts.OptimizeSize)
1840 FuncAttrs.addAttribute(llvm::Attribute::OptimizeForSize);
1841 if (CodeGenOpts.OptimizeSize == 2)
1842 FuncAttrs.addAttribute(llvm::Attribute::MinSize);
1843 }
1844
1845 if (CodeGenOpts.DisableRedZone)
1846 FuncAttrs.addAttribute(llvm::Attribute::NoRedZone);
1847 if (CodeGenOpts.IndirectTlsSegRefs)
1848 FuncAttrs.addAttribute("indirect-tls-seg-refs");
1849 if (CodeGenOpts.NoImplicitFloat)
1850 FuncAttrs.addAttribute(llvm::Attribute::NoImplicitFloat);
1851
1852 if (AttrOnCallSite) {
1853 // Attributes that should go on the call site only.
1854 // FIXME: Look for 'BuiltinAttr' on the function rather than re-checking
1855 // the -fno-builtin-foo list.
1856 if (!CodeGenOpts.SimplifyLibCalls || LangOpts.isNoBuiltinFunc(Name))
1857 FuncAttrs.addAttribute(llvm::Attribute::NoBuiltin);
1858 if (!CodeGenOpts.TrapFuncName.empty())
1859 FuncAttrs.addAttribute("trap-func-name", CodeGenOpts.TrapFuncName);
1860 } else {
1861 switch (CodeGenOpts.getFramePointer()) {
1862 case CodeGenOptions::FramePointerKind::None:
1863 // This is the default behavior.
1864 break;
1865 case CodeGenOptions::FramePointerKind::NonLeaf:
1866 case CodeGenOptions::FramePointerKind::All:
1867 FuncAttrs.addAttribute("frame-pointer",
1868 CodeGenOptions::getFramePointerKindName(
1869 CodeGenOpts.getFramePointer()));
1870 }
1871
1872 if (CodeGenOpts.LessPreciseFPMAD)
1873 FuncAttrs.addAttribute("less-precise-fpmad", "true");
1874
1875 if (CodeGenOpts.NullPointerIsValid)
1876 FuncAttrs.addAttribute(llvm::Attribute::NullPointerIsValid);
1877
1878 if (CodeGenOpts.FPDenormalMode != llvm::DenormalMode::getIEEE())
1879 FuncAttrs.addAttribute("denormal-fp-math",
1880 CodeGenOpts.FPDenormalMode.str());
1881 if (CodeGenOpts.FP32DenormalMode != CodeGenOpts.FPDenormalMode) {
1882 FuncAttrs.addAttribute(
1883 "denormal-fp-math-f32",
1884 CodeGenOpts.FP32DenormalMode.str());
1885 }
1886
1887 if (LangOpts.getDefaultExceptionMode() == LangOptions::FPE_Ignore)
1888 FuncAttrs.addAttribute("no-trapping-math", "true");
1889
1890 // TODO: Are these all needed?
1891 // unsafe/inf/nan/nsz are handled by instruction-level FastMathFlags.
1892 if (LangOpts.NoHonorInfs)
1893 FuncAttrs.addAttribute("no-infs-fp-math", "true");
1894 if (LangOpts.NoHonorNaNs)
1895 FuncAttrs.addAttribute("no-nans-fp-math", "true");
1896 if (LangOpts.ApproxFunc)
1897 FuncAttrs.addAttribute("approx-func-fp-math", "true");
1898 if (LangOpts.AllowFPReassoc && LangOpts.AllowRecip &&
1899 LangOpts.NoSignedZero && LangOpts.ApproxFunc &&
1900 (LangOpts.getDefaultFPContractMode() ==
1901 LangOptions::FPModeKind::FPM_Fast ||
1902 LangOpts.getDefaultFPContractMode() ==
1903 LangOptions::FPModeKind::FPM_FastHonorPragmas))
1904 FuncAttrs.addAttribute("unsafe-fp-math", "true");
1905 if (CodeGenOpts.SoftFloat)
1906 FuncAttrs.addAttribute("use-soft-float", "true");
1907 FuncAttrs.addAttribute("stack-protector-buffer-size",
1908 llvm::utostr(CodeGenOpts.SSPBufferSize));
1909 if (LangOpts.NoSignedZero)
1910 FuncAttrs.addAttribute("no-signed-zeros-fp-math", "true");
1911
1912 // TODO: Reciprocal estimate codegen options should apply to instructions?
1913 const std::vector<std::string> &Recips = CodeGenOpts.Reciprocals;
1914 if (!Recips.empty())
1915 FuncAttrs.addAttribute("reciprocal-estimates",
1916 llvm::join(Recips, ","));
1917
1918 if (!CodeGenOpts.PreferVectorWidth.empty() &&
1919 CodeGenOpts.PreferVectorWidth != "none")
1920 FuncAttrs.addAttribute("prefer-vector-width",
1921 CodeGenOpts.PreferVectorWidth);
1922
1923 if (CodeGenOpts.StackRealignment)
1924 FuncAttrs.addAttribute("stackrealign");
1925 if (CodeGenOpts.Backchain)
1926 FuncAttrs.addAttribute("backchain");
1927 if (CodeGenOpts.EnableSegmentedStacks)
1928 FuncAttrs.addAttribute("split-stack");
1929
1930 if (CodeGenOpts.SpeculativeLoadHardening)
1931 FuncAttrs.addAttribute(llvm::Attribute::SpeculativeLoadHardening);
1932
1933 // Add zero-call-used-regs attribute.
1934 switch (CodeGenOpts.getZeroCallUsedRegs()) {
1935 case llvm::ZeroCallUsedRegs::ZeroCallUsedRegsKind::Skip:
1936 FuncAttrs.removeAttribute("zero-call-used-regs");
1937 break;
1938 case llvm::ZeroCallUsedRegs::ZeroCallUsedRegsKind::UsedGPRArg:
1939 FuncAttrs.addAttribute("zero-call-used-regs", "used-gpr-arg");
1940 break;
1941 case llvm::ZeroCallUsedRegs::ZeroCallUsedRegsKind::UsedGPR:
1942 FuncAttrs.addAttribute("zero-call-used-regs", "used-gpr");
1943 break;
1944 case llvm::ZeroCallUsedRegs::ZeroCallUsedRegsKind::UsedArg:
1945 FuncAttrs.addAttribute("zero-call-used-regs", "used-arg");
1946 break;
1947 case llvm::ZeroCallUsedRegs::ZeroCallUsedRegsKind::Used:
1948 FuncAttrs.addAttribute("zero-call-used-regs", "used");
1949 break;
1950 case llvm::ZeroCallUsedRegs::ZeroCallUsedRegsKind::AllGPRArg:
1951 FuncAttrs.addAttribute("zero-call-used-regs", "all-gpr-arg");
1952 break;
1953 case llvm::ZeroCallUsedRegs::ZeroCallUsedRegsKind::AllGPR:
1954 FuncAttrs.addAttribute("zero-call-used-regs", "all-gpr");
1955 break;
1956 case llvm::ZeroCallUsedRegs::ZeroCallUsedRegsKind::AllArg:
1957 FuncAttrs.addAttribute("zero-call-used-regs", "all-arg");
1958 break;
1959 case llvm::ZeroCallUsedRegs::ZeroCallUsedRegsKind::All:
1960 FuncAttrs.addAttribute("zero-call-used-regs", "all");
1961 break;
1962 }
1963 }
1964
1965 if (getLangOpts().assumeFunctionsAreConvergent()) {
1966 // Conservatively, mark all functions and calls in CUDA and OpenCL as
1967 // convergent (meaning, they may call an intrinsically convergent op, such
1968 // as __syncthreads() / barrier(), and so can't have certain optimizations
1969 // applied around them). LLVM will remove this attribute where it safely
1970 // can.
1971 FuncAttrs.addAttribute(llvm::Attribute::Convergent);
1972 }
1973
1974 // TODO: NoUnwind attribute should be added for other GPU modes HIP,
1975 // SYCL, OpenMP offload. AFAIK, none of them support exceptions in device
1976 // code.
1977 if ((getLangOpts().CUDA && getLangOpts().CUDAIsDevice) ||
1978 getLangOpts().OpenCL) {
1979 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
1980 }
1981
1982 for (StringRef Attr : CodeGenOpts.DefaultFunctionAttrs) {
1983 StringRef Var, Value;
1984 std::tie(Var, Value) = Attr.split('=');
1985 FuncAttrs.addAttribute(Var, Value);
1986 }
1987 }
1988
1989 void CodeGenModule::addDefaultFunctionDefinitionAttributes(llvm::Function &F) {
1990 llvm::AttrBuilder FuncAttrs(F.getContext());
1991 getDefaultFunctionAttributes(F.getName(), F.hasOptNone(),
1992 /* AttrOnCallSite = */ false, FuncAttrs);
1993 // TODO: call GetCPUAndFeaturesAttributes?
1994 F.addFnAttrs(FuncAttrs);
1995 }
1996
1997 void CodeGenModule::addDefaultFunctionDefinitionAttributes(
1998 llvm::AttrBuilder &attrs) {
1999 getDefaultFunctionAttributes(/*function name*/ "", /*optnone*/ false,
2000 /*for call*/ false, attrs);
2001 GetCPUAndFeaturesAttributes(GlobalDecl(), attrs);
2002 }
2003
2004 static void addNoBuiltinAttributes(llvm::AttrBuilder &FuncAttrs,
2005 const LangOptions &LangOpts,
2006 const NoBuiltinAttr *NBA = nullptr) {
2007 auto AddNoBuiltinAttr = [&FuncAttrs](StringRef BuiltinName) {
2008 SmallString<32> AttributeName;
2009 AttributeName += "no-builtin-";
2010 AttributeName += BuiltinName;
2011 FuncAttrs.addAttribute(AttributeName);
2012 };
2013
2014 // First, handle the language options passed through -fno-builtin.
2015 if (LangOpts.NoBuiltin) {
2016 // -fno-builtin disables them all.
2017 FuncAttrs.addAttribute("no-builtins");
2018 return;
2019 }
2020
2021 // Then, add attributes for builtins specified through -fno-builtin-<name>.
2022 llvm::for_each(LangOpts.NoBuiltinFuncs, AddNoBuiltinAttr);
2023
2024 // Now, let's check the __attribute__((no_builtin("...")) attribute added to
2025 // the source.
2026 if (!NBA)
2027 return;
2028
2029 // If there is a wildcard in the builtin names specified through the
2030 // attribute, disable them all.
2031 if (llvm::is_contained(NBA->builtinNames(), "*")) {
2032 FuncAttrs.addAttribute("no-builtins");
2033 return;
2034 }
2035
2036 // And last, add the rest of the builtin names.
2037 llvm::for_each(NBA->builtinNames(), AddNoBuiltinAttr);
2038 }
2039
2040 static bool DetermineNoUndef(QualType QTy, CodeGenTypes &Types,
2041 const llvm::DataLayout &DL, const ABIArgInfo &AI,
2042 bool CheckCoerce = true) {
2043 llvm::Type *Ty = Types.ConvertTypeForMem(QTy);
2044 if (AI.getKind() == ABIArgInfo::Indirect)
2045 return true;
2046 if (AI.getKind() == ABIArgInfo::Extend)
2047 return true;
2048 if (!DL.typeSizeEqualsStoreSize(Ty))
2049 // TODO: This will result in a modest amount of values not marked noundef
2050 // when they could be. We care about values that *invisibly* contain undef
2051 // bits from the perspective of LLVM IR.
2052 return false;
2053 if (CheckCoerce && AI.canHaveCoerceToType()) {
2054 llvm::Type *CoerceTy = AI.getCoerceToType();
2055 if (llvm::TypeSize::isKnownGT(DL.getTypeSizeInBits(CoerceTy),
2056 DL.getTypeSizeInBits(Ty)))
2057 // If we're coercing to a type with a greater size than the canonical one,
2058 // we're introducing new undef bits.
2059 // Coercing to a type of smaller or equal size is ok, as we know that
2060 // there's no internal padding (typeSizeEqualsStoreSize).
2061 return false;
2062 }
2063 if (QTy->isBitIntType())
2064 return true;
2065 if (QTy->isReferenceType())
2066 return true;
2067 if (QTy->isNullPtrType())
2068 return false;
2069 if (QTy->isMemberPointerType())
2070 // TODO: Some member pointers are `noundef`, but it depends on the ABI. For
2071 // now, never mark them.
2072 return false;
2073 if (QTy->isScalarType()) {
2074 if (const ComplexType *Complex = dyn_cast<ComplexType>(QTy))
2075 return DetermineNoUndef(Complex->getElementType(), Types, DL, AI, false);
2076 return true;
2077 }
2078 if (const VectorType *Vector = dyn_cast<VectorType>(QTy))
2079 return DetermineNoUndef(Vector->getElementType(), Types, DL, AI, false);
2080 if (const MatrixType *Matrix = dyn_cast<MatrixType>(QTy))
2081 return DetermineNoUndef(Matrix->getElementType(), Types, DL, AI, false);
2082 if (const ArrayType *Array = dyn_cast<ArrayType>(QTy))
2083 return DetermineNoUndef(Array->getElementType(), Types, DL, AI, false);
2084
2085 // TODO: Some structs may be `noundef`, in specific situations.
2086 return false;
2087 }
2088
2089 /// Check if the argument of a function has maybe_undef attribute.
2090 static bool IsArgumentMaybeUndef(const Decl *TargetDecl,
2091 unsigned NumRequiredArgs, unsigned ArgNo) {
2092 const auto *FD = dyn_cast_or_null<FunctionDecl>(TargetDecl);
2093 if (!FD)
2094 return false;
2095
2096 // Assume variadic arguments do not have maybe_undef attribute.
2097 if (ArgNo >= NumRequiredArgs)
2098 return false;
2099
2100 // Check if argument has maybe_undef attribute.
2101 if (ArgNo < FD->getNumParams()) {
2102 const ParmVarDecl *Param = FD->getParamDecl(ArgNo);
2103 if (Param && Param->hasAttr<MaybeUndefAttr>())
2104 return true;
2105 }
2106
2107 return false;
2108 }
2109
2110 /// Construct the IR attribute list of a function or call.
2111 ///
2112 /// When adding an attribute, please consider where it should be handled:
2113 ///
2114 /// - getDefaultFunctionAttributes is for attributes that are essentially
2115 /// part of the global target configuration (but perhaps can be
2116 /// overridden on a per-function basis). Adding attributes there
2117 /// will cause them to also be set in frontends that build on Clang's
2118 /// target-configuration logic, as well as for code defined in library
2119 /// modules such as CUDA's libdevice.
2120 ///
2121 /// - ConstructAttributeList builds on top of getDefaultFunctionAttributes
2122 /// and adds declaration-specific, convention-specific, and
2123 /// frontend-specific logic. The last is of particular importance:
2124 /// attributes that restrict how the frontend generates code must be
2125 /// added here rather than getDefaultFunctionAttributes.
2126 ///
2127 void CodeGenModule::ConstructAttributeList(StringRef Name,
2128 const CGFunctionInfo &FI,
2129 CGCalleeInfo CalleeInfo,
2130 llvm::AttributeList &AttrList,
2131 unsigned &CallingConv,
2132 bool AttrOnCallSite, bool IsThunk) {
2133 llvm::AttrBuilder FuncAttrs(getLLVMContext());
2134 llvm::AttrBuilder RetAttrs(getLLVMContext());
2135
2136 // Collect function IR attributes from the CC lowering.
2137 // We'll collect the paramete and result attributes later.
2138 CallingConv = FI.getEffectiveCallingConvention();
2139 if (FI.isNoReturn())
2140 FuncAttrs.addAttribute(llvm::Attribute::NoReturn);
2141 if (FI.isCmseNSCall())
2142 FuncAttrs.addAttribute("cmse_nonsecure_call");
2143
2144 // Collect function IR attributes from the callee prototype if we have one.
2145 AddAttributesFromFunctionProtoType(getContext(), FuncAttrs,
2146 CalleeInfo.getCalleeFunctionProtoType());
2147
2148 const Decl *TargetDecl = CalleeInfo.getCalleeDecl().getDecl();
2149
2150 // Attach assumption attributes to the declaration. If this is a call
2151 // site, attach assumptions from the caller to the call as well.
2152 AddAttributesFromAssumes(FuncAttrs, TargetDecl);
2153
2154 bool HasOptnone = false;
2155 // The NoBuiltinAttr attached to the target FunctionDecl.
2156 const NoBuiltinAttr *NBA = nullptr;
2157
2158 // Some ABIs may result in additional accesses to arguments that may
2159 // otherwise not be present.
2160 auto AddPotentialArgAccess = [&]() {
2161 llvm::Attribute A = FuncAttrs.getAttribute(llvm::Attribute::Memory);
2162 if (A.isValid())
2163 FuncAttrs.addMemoryAttr(A.getMemoryEffects() |
2164 llvm::MemoryEffects::argMemOnly());
2165 };
2166
2167 // Collect function IR attributes based on declaration-specific
2168 // information.
2169 // FIXME: handle sseregparm someday...
2170 if (TargetDecl) {
2171 if (TargetDecl->hasAttr<ReturnsTwiceAttr>())
2172 FuncAttrs.addAttribute(llvm::Attribute::ReturnsTwice);
2173 if (TargetDecl->hasAttr<NoThrowAttr>())
2174 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
2175 if (TargetDecl->hasAttr<NoReturnAttr>())
2176 FuncAttrs.addAttribute(llvm::Attribute::NoReturn);
2177 if (TargetDecl->hasAttr<ColdAttr>())
2178 FuncAttrs.addAttribute(llvm::Attribute::Cold);
2179 if (TargetDecl->hasAttr<HotAttr>())
2180 FuncAttrs.addAttribute(llvm::Attribute::Hot);
2181 if (TargetDecl->hasAttr<NoDuplicateAttr>())
2182 FuncAttrs.addAttribute(llvm::Attribute::NoDuplicate);
2183 if (TargetDecl->hasAttr<ConvergentAttr>())
2184 FuncAttrs.addAttribute(llvm::Attribute::Convergent);
2185
2186 if (const FunctionDecl *Fn = dyn_cast<FunctionDecl>(TargetDecl)) {
2187 AddAttributesFromFunctionProtoType(
2188 getContext(), FuncAttrs, Fn->getType()->getAs<FunctionProtoType>());
2189 if (AttrOnCallSite && Fn->isReplaceableGlobalAllocationFunction()) {
2190 // A sane operator new returns a non-aliasing pointer.
2191 auto Kind = Fn->getDeclName().getCXXOverloadedOperator();
2192 if (getCodeGenOpts().AssumeSaneOperatorNew &&
2193 (Kind == OO_New || Kind == OO_Array_New))
2194 RetAttrs.addAttribute(llvm::Attribute::NoAlias);
2195 }
2196 const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(Fn);
2197 const bool IsVirtualCall = MD && MD->isVirtual();
2198 // Don't use [[noreturn]], _Noreturn or [[no_builtin]] for a call to a
2199 // virtual function. These attributes are not inherited by overloads.
2200 if (!(AttrOnCallSite && IsVirtualCall)) {
2201 if (Fn->isNoReturn())
2202 FuncAttrs.addAttribute(llvm::Attribute::NoReturn);
2203 NBA = Fn->getAttr<NoBuiltinAttr>();
2204 }
2205 // Only place nomerge attribute on call sites, never functions. This
2206 // allows it to work on indirect virtual function calls.
2207 if (AttrOnCallSite && TargetDecl->hasAttr<NoMergeAttr>())
2208 FuncAttrs.addAttribute(llvm::Attribute::NoMerge);
2209 }
2210
2211 // 'const', 'pure' and 'noalias' attributed functions are also nounwind.
2212 if (TargetDecl->hasAttr<ConstAttr>()) {
2213 FuncAttrs.addMemoryAttr(llvm::MemoryEffects::none());
2214 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
2215 // gcc specifies that 'const' functions have greater restrictions than
2216 // 'pure' functions, so they also cannot have infinite loops.
2217 FuncAttrs.addAttribute(llvm::Attribute::WillReturn);
2218 } else if (TargetDecl->hasAttr<PureAttr>()) {
2219 FuncAttrs.addMemoryAttr(llvm::MemoryEffects::readOnly());
2220 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
2221 // gcc specifies that 'pure' functions cannot have infinite loops.
2222 FuncAttrs.addAttribute(llvm::Attribute::WillReturn);
2223 } else if (TargetDecl->hasAttr<NoAliasAttr>()) {
2224 FuncAttrs.addMemoryAttr(llvm::MemoryEffects::argMemOnly());
2225 FuncAttrs.addAttribute(llvm::Attribute::NoUnwind);
2226 }
2227 if (TargetDecl->hasAttr<RestrictAttr>())
2228 RetAttrs.addAttribute(llvm::Attribute::NoAlias);
2229 if (TargetDecl->hasAttr<ReturnsNonNullAttr>() &&
2230 !CodeGenOpts.NullPointerIsValid)
2231 RetAttrs.addAttribute(llvm::Attribute::NonNull);
2232 if (TargetDecl->hasAttr<AnyX86NoCallerSavedRegistersAttr>())
2233 FuncAttrs.addAttribute("no_caller_saved_registers");
2234 if (TargetDecl->hasAttr<AnyX86NoCfCheckAttr>())
2235 FuncAttrs.addAttribute(llvm::Attribute::NoCfCheck);
2236 if (TargetDecl->hasAttr<LeafAttr>())
2237 FuncAttrs.addAttribute(llvm::Attribute::NoCallback);
2238
2239 HasOptnone = TargetDecl->hasAttr<OptimizeNoneAttr>();
2240 if (auto *AllocSize = TargetDecl->getAttr<AllocSizeAttr>()) {
2241 std::optional<unsigned> NumElemsParam;
2242 if (AllocSize->getNumElemsParam().isValid())
2243 NumElemsParam = AllocSize->getNumElemsParam().getLLVMIndex();
2244 FuncAttrs.addAllocSizeAttr(AllocSize->getElemSizeParam().getLLVMIndex(),
2245 NumElemsParam);
2246 }
2247
2248 if (TargetDecl->hasAttr<OpenCLKernelAttr>()) {
2249 if (getLangOpts().OpenCLVersion <= 120) {
2250 // OpenCL v1.2 Work groups are always uniform
2251 FuncAttrs.addAttribute("uniform-work-group-size", "true");
2252 } else {
2253 // OpenCL v2.0 Work groups may be whether uniform or not.
2254 // '-cl-uniform-work-group-size' compile option gets a hint
2255 // to the compiler that the global work-size be a multiple of
2256 // the work-group size specified to clEnqueueNDRangeKernel
2257 // (i.e. work groups are uniform).
2258 FuncAttrs.addAttribute("uniform-work-group-size",
2259 llvm::toStringRef(CodeGenOpts.UniformWGSize));
2260 }
2261 }
2262 }
2263
2264 // Attach "no-builtins" attributes to:
2265 // * call sites: both `nobuiltin` and "no-builtins" or "no-builtin-<name>".
2266 // * definitions: "no-builtins" or "no-builtin-<name>" only.
2267 // The attributes can come from:
2268 // * LangOpts: -ffreestanding, -fno-builtin, -fno-builtin-<name>
2269 // * FunctionDecl attributes: __attribute__((no_builtin(...)))
2270 addNoBuiltinAttributes(FuncAttrs, getLangOpts(), NBA);
2271
2272 // Collect function IR attributes based on global settiings.
2273 getDefaultFunctionAttributes(Name, HasOptnone, AttrOnCallSite, FuncAttrs);
2274
2275 // Override some default IR attributes based on declaration-specific
2276 // information.
2277 if (TargetDecl) {
2278 if (TargetDecl->hasAttr<NoSpeculativeLoadHardeningAttr>())
2279 FuncAttrs.removeAttribute(llvm::Attribute::SpeculativeLoadHardening);
2280 if (TargetDecl->hasAttr<SpeculativeLoadHardeningAttr>())
2281 FuncAttrs.addAttribute(llvm::Attribute::SpeculativeLoadHardening);
2282 if (TargetDecl->hasAttr<NoSplitStackAttr>())
2283 FuncAttrs.removeAttribute("split-stack");
2284 if (TargetDecl->hasAttr<ZeroCallUsedRegsAttr>()) {
2285 // A function "__attribute__((...))" overrides the command-line flag.
2286 auto Kind =
2287 TargetDecl->getAttr<ZeroCallUsedRegsAttr>()->getZeroCallUsedRegs();
2288 FuncAttrs.removeAttribute("zero-call-used-regs");
2289 FuncAttrs.addAttribute(
2290 "zero-call-used-regs",
2291 ZeroCallUsedRegsAttr::ConvertZeroCallUsedRegsKindToStr(Kind));
2292 }
2293
2294 // Add NonLazyBind attribute to function declarations when -fno-plt
2295 // is used.
2296 // FIXME: what if we just haven't processed the function definition
2297 // yet, or if it's an external definition like C99 inline?
2298 if (CodeGenOpts.NoPLT) {
2299 if (auto *Fn = dyn_cast<FunctionDecl>(TargetDecl)) {
2300 if (!Fn->isDefined() && !AttrOnCallSite) {
2301 FuncAttrs.addAttribute(llvm::Attribute::NonLazyBind);
2302 }
2303 }
2304 }
2305 }
2306
2307 // Add "sample-profile-suffix-elision-policy" attribute for internal linkage
2308 // functions with -funique-internal-linkage-names.
2309 if (TargetDecl && CodeGenOpts.UniqueInternalLinkageNames) {
2310 if (const auto *FD = dyn_cast_or_null<FunctionDecl>(TargetDecl)) {
2311 if (!FD->isExternallyVisible())
2312 FuncAttrs.addAttribute("sample-profile-suffix-elision-policy",
2313 "selected");
2314 }
2315 }
2316
2317 // Collect non-call-site function IR attributes from declaration-specific
2318 // information.
2319 if (!AttrOnCallSite) {
2320 if (TargetDecl && TargetDecl->hasAttr<CmseNSEntryAttr>())
2321 FuncAttrs.addAttribute("cmse_nonsecure_entry");
2322
2323 // Whether tail calls are enabled.
2324 auto shouldDisableTailCalls = [&] {
2325 // Should this be honored in getDefaultFunctionAttributes?
2326 if (CodeGenOpts.DisableTailCalls)
2327 return true;
2328
2329 if (!TargetDecl)
2330 return false;
2331
2332 if (TargetDecl->hasAttr<DisableTailCallsAttr>() ||
2333 TargetDecl->hasAttr<AnyX86InterruptAttr>())
2334 return true;
2335
2336 if (CodeGenOpts.NoEscapingBlockTailCalls) {
2337 if (const auto *BD = dyn_cast<BlockDecl>(TargetDecl))
2338 if (!BD->doesNotEscape())
2339 return true;
2340 }
2341
2342 return false;
2343 };
2344 if (shouldDisableTailCalls())
2345 FuncAttrs.addAttribute("disable-tail-calls", "true");
2346
2347 // CPU/feature overrides. addDefaultFunctionDefinitionAttributes
2348 // handles these separately to set them based on the global defaults.
2349 GetCPUAndFeaturesAttributes(CalleeInfo.getCalleeDecl(), FuncAttrs);
2350 }
2351
2352 // Collect attributes from arguments and return values.
2353 ClangToLLVMArgMapping IRFunctionArgs(getContext(), FI);
2354
2355 QualType RetTy = FI.getReturnType();
2356 const ABIArgInfo &RetAI = FI.getReturnInfo();
2357 const llvm::DataLayout &DL = getDataLayout();
2358
2359 // Determine if the return type could be partially undef
2360 if (CodeGenOpts.EnableNoundefAttrs &&
2361 HasStrictReturn(*this, RetTy, TargetDecl)) {
2362 if (!RetTy->isVoidType() && RetAI.getKind() != ABIArgInfo::Indirect &&
2363 DetermineNoUndef(RetTy, getTypes(), DL, RetAI))
2364 RetAttrs.addAttribute(llvm::Attribute::NoUndef);
2365 }
2366
2367 switch (RetAI.getKind()) {
2368 case ABIArgInfo::Extend:
2369 if (RetAI.isSignExt())
2370 RetAttrs.addAttribute(llvm::Attribute::SExt);
2371 else
2372 RetAttrs.addAttribute(llvm::Attribute::ZExt);
2373 [[fallthrough]];
2374 case ABIArgInfo::Direct:
2375 if (RetAI.getInReg())
2376 RetAttrs.addAttribute(llvm::Attribute::InReg);
2377 break;
2378 case ABIArgInfo::Ignore:
2379 break;
2380
2381 case ABIArgInfo::InAlloca:
2382 case ABIArgInfo::Indirect: {
2383 // inalloca and sret disable readnone and readonly
2384 AddPotentialArgAccess();
2385 break;
2386 }
2387
2388 case ABIArgInfo::CoerceAndExpand:
2389 break;
2390
2391 case ABIArgInfo::Expand:
2392 case ABIArgInfo::IndirectAliased:
2393 llvm_unreachable("Invalid ABI kind for return argument");
2394 }
2395
2396 if (!IsThunk) {
2397 // FIXME: fix this properly, https://reviews.llvm.org/D100388
2398 if (const auto *RefTy = RetTy->getAs<ReferenceType>()) {
2399 QualType PTy = RefTy->getPointeeType();
2400 if (!PTy->isIncompleteType() && PTy->isConstantSizeType())
2401 RetAttrs.addDereferenceableAttr(
2402 getMinimumObjectSize(PTy).getQuantity());
2403 if (getTypes().getTargetAddressSpace(PTy) == 0 &&
2404 !CodeGenOpts.NullPointerIsValid)
2405 RetAttrs.addAttribute(llvm::Attribute::NonNull);
2406 if (PTy->isObjectType()) {
2407 llvm::Align Alignment =
2408 getNaturalPointeeTypeAlignment(RetTy).getAsAlign();
2409 RetAttrs.addAlignmentAttr(Alignment);
2410 }
2411 }
2412 }
2413
2414 bool hasUsedSRet = false;
2415 SmallVector<llvm::AttributeSet, 4> ArgAttrs(IRFunctionArgs.totalIRArgs());
2416
2417 // Attach attributes to sret.
2418 if (IRFunctionArgs.hasSRetArg()) {
2419 llvm::AttrBuilder SRETAttrs(getLLVMContext());
2420 SRETAttrs.addStructRetAttr(getTypes().ConvertTypeForMem(RetTy));
2421 hasUsedSRet = true;
2422 if (RetAI.getInReg())
2423 SRETAttrs.addAttribute(llvm::Attribute::InReg);
2424 SRETAttrs.addAlignmentAttr(RetAI.getIndirectAlign().getQuantity());
2425 ArgAttrs[IRFunctionArgs.getSRetArgNo()] =
2426 llvm::AttributeSet::get(getLLVMContext(), SRETAttrs);
2427 }
2428
2429 // Attach attributes to inalloca argument.
2430 if (IRFunctionArgs.hasInallocaArg()) {
2431 llvm::AttrBuilder Attrs(getLLVMContext());
2432 Attrs.addInAllocaAttr(FI.getArgStruct());
2433 ArgAttrs[IRFunctionArgs.getInallocaArgNo()] =
2434 llvm::AttributeSet::get(getLLVMContext(), Attrs);
2435 }
2436
2437 // Apply `nonnull`, `dereferencable(N)` and `align N` to the `this` argument,
2438 // unless this is a thunk function.
2439 // FIXME: fix this properly, https://reviews.llvm.org/D100388
2440 if (FI.isInstanceMethod() && !IRFunctionArgs.hasInallocaArg() &&
2441 !FI.arg_begin()->type->isVoidPointerType() && !IsThunk) {
2442 auto IRArgs = IRFunctionArgs.getIRArgs(0);
2443
2444 assert(IRArgs.second == 1 && "Expected only a single `this` pointer.");
2445
2446 llvm::AttrBuilder Attrs(getLLVMContext());
2447
2448 QualType ThisTy =
2449 FI.arg_begin()->type.castAs<PointerType>()->getPointeeType();
2450
2451 if (!CodeGenOpts.NullPointerIsValid &&
2452 getTypes().getTargetAddressSpace(FI.arg_begin()->type) == 0) {
2453 Attrs.addAttribute(llvm::Attribute::NonNull);
2454 Attrs.addDereferenceableAttr(getMinimumObjectSize(ThisTy).getQuantity());
2455 } else {
2456 // FIXME dereferenceable should be correct here, regardless of
2457 // NullPointerIsValid. However, dereferenceable currently does not always
2458 // respect NullPointerIsValid and may imply nonnull and break the program.
2459 // See https://reviews.llvm.org/D66618 for discussions.
2460 Attrs.addDereferenceableOrNullAttr(
2461 getMinimumObjectSize(
2462 FI.arg_begin()->type.castAs<PointerType>()->getPointeeType())
2463 .getQuantity());
2464 }
2465
2466 llvm::Align Alignment =
2467 getNaturalTypeAlignment(ThisTy, /*BaseInfo=*/nullptr,
2468 /*TBAAInfo=*/nullptr, /*forPointeeType=*/true)
2469 .getAsAlign();
2470 Attrs.addAlignmentAttr(Alignment);
2471
2472 ArgAttrs[IRArgs.first] = llvm::AttributeSet::get(getLLVMContext(), Attrs);
2473 }
2474
2475 unsigned ArgNo = 0;
2476 for (CGFunctionInfo::const_arg_iterator I = FI.arg_begin(),
2477 E = FI.arg_end();
2478 I != E; ++I, ++ArgNo) {
2479 QualType ParamType = I->type;
2480 const ABIArgInfo &AI = I->info;
2481 llvm::AttrBuilder Attrs(getLLVMContext());
2482
2483 // Add attribute for padding argument, if necessary.
2484 if (IRFunctionArgs.hasPaddingArg(ArgNo)) {
2485 if (AI.getPaddingInReg()) {
2486 ArgAttrs[IRFunctionArgs.getPaddingArgNo(ArgNo)] =
2487 llvm::AttributeSet::get(
2488 getLLVMContext(),
2489 llvm::AttrBuilder(getLLVMContext()).addAttribute(llvm::Attribute::InReg));
2490 }
2491 }
2492
2493 // Decide whether the argument we're handling could be partially undef
2494 if (CodeGenOpts.EnableNoundefAttrs &&
2495 DetermineNoUndef(ParamType, getTypes(), DL, AI)) {
2496 Attrs.addAttribute(llvm::Attribute::NoUndef);
2497 }
2498
2499 // 'restrict' -> 'noalias' is done in EmitFunctionProlog when we
2500 // have the corresponding parameter variable. It doesn't make
2501 // sense to do it here because parameters are so messed up.
2502 switch (AI.getKind()) {
2503 case ABIArgInfo::Extend:
2504 if (AI.isSignExt())
2505 Attrs.addAttribute(llvm::Attribute::SExt);
2506 else
2507 Attrs.addAttribute(llvm::Attribute::ZExt);
2508 [[fallthrough]];
2509 case ABIArgInfo::Direct:
2510 if (ArgNo == 0 && FI.isChainCall())
2511 Attrs.addAttribute(llvm::Attribute::Nest);
2512 else if (AI.getInReg())
2513 Attrs.addAttribute(llvm::Attribute::InReg);
2514 Attrs.addStackAlignmentAttr(llvm::MaybeAlign(AI.getDirectAlign()));
2515 break;
2516
2517 case ABIArgInfo::Indirect: {
2518 if (AI.getInReg())
2519 Attrs.addAttribute(llvm::Attribute::InReg);
2520
2521 if (AI.getIndirectByVal())
2522 Attrs.addByValAttr(getTypes().ConvertTypeForMem(ParamType));
2523
2524 auto *Decl = ParamType->getAsRecordDecl();
2525 if (CodeGenOpts.PassByValueIsNoAlias && Decl &&
2526 Decl->getArgPassingRestrictions() == RecordDecl::APK_CanPassInRegs)
2527 // When calling the function, the pointer passed in will be the only
2528 // reference to the underlying object. Mark it accordingly.
2529 Attrs.addAttribute(llvm::Attribute::NoAlias);
2530
2531 // TODO: We could add the byref attribute if not byval, but it would
2532 // require updating many testcases.
2533
2534 CharUnits Align = AI.getIndirectAlign();
2535
2536 // In a byval argument, it is important that the required
2537 // alignment of the type is honored, as LLVM might be creating a
2538 // *new* stack object, and needs to know what alignment to give
2539 // it. (Sometimes it can deduce a sensible alignment on its own,
2540 // but not if clang decides it must emit a packed struct, or the
2541 // user specifies increased alignment requirements.)
2542 //
2543 // This is different from indirect *not* byval, where the object
2544 // exists already, and the align attribute is purely
2545 // informative.
2546 assert(!Align.isZero());
2547
2548 // For now, only add this when we have a byval argument.
2549 // TODO: be less lazy about updating test cases.
2550 if (AI.getIndirectByVal())
2551 Attrs.addAlignmentAttr(Align.getQuantity());
2552
2553 // byval disables readnone and readonly.
2554 AddPotentialArgAccess();
2555 break;
2556 }
2557 case ABIArgInfo::IndirectAliased: {
2558 CharUnits Align = AI.getIndirectAlign();
2559 Attrs.addByRefAttr(getTypes().ConvertTypeForMem(ParamType));
2560 Attrs.addAlignmentAttr(Align.getQuantity());
2561 break;
2562 }
2563 case ABIArgInfo::Ignore:
2564 case ABIArgInfo::Expand:
2565 case ABIArgInfo::CoerceAndExpand:
2566 break;
2567
2568 case ABIArgInfo::InAlloca:
2569 // inalloca disables readnone and readonly.
2570 AddPotentialArgAccess();
2571 continue;
2572 }
2573
2574 if (const auto *RefTy = ParamType->getAs<ReferenceType>()) {
2575 QualType PTy = RefTy->getPointeeType();
2576 if (!PTy->isIncompleteType() && PTy->isConstantSizeType())
2577 Attrs.addDereferenceableAttr(
2578 getMinimumObjectSize(PTy).getQuantity());
2579 if (getTypes().getTargetAddressSpace(PTy) == 0 &&
2580 !CodeGenOpts.NullPointerIsValid)
2581 Attrs.addAttribute(llvm::Attribute::NonNull);
2582 if (PTy->isObjectType()) {
2583 llvm::Align Alignment =
2584 getNaturalPointeeTypeAlignment(ParamType).getAsAlign();
2585 Attrs.addAlignmentAttr(Alignment);
2586 }
2587 }
2588
2589 // From OpenCL spec v3.0.10 section 6.3.5 Alignment of Types:
2590 // > For arguments to a __kernel function declared to be a pointer to a
2591 // > data type, the OpenCL compiler can assume that the pointee is always
2592 // > appropriately aligned as required by the data type.
2593 if (TargetDecl && TargetDecl->hasAttr<OpenCLKernelAttr>() &&
2594 ParamType->isPointerType()) {
2595 QualType PTy = ParamType->getPointeeType();
2596 if (!PTy->isIncompleteType() && PTy->isConstantSizeType()) {
2597 llvm::Align Alignment =
2598 getNaturalPointeeTypeAlignment(ParamType).getAsAlign();
2599 Attrs.addAlignmentAttr(Alignment);
2600 }
2601 }
2602
2603 switch (FI.getExtParameterInfo(ArgNo).getABI()) {
2604 case ParameterABI::Ordinary:
2605 break;
2606
2607 case ParameterABI::SwiftIndirectResult: {
2608 // Add 'sret' if we haven't already used it for something, but
2609 // only if the result is void.
2610 if (!hasUsedSRet && RetTy->isVoidType()) {
2611 Attrs.addStructRetAttr(getTypes().ConvertTypeForMem(ParamType));
2612 hasUsedSRet = true;
2613 }
2614
2615 // Add 'noalias' in either case.
2616 Attrs.addAttribute(llvm::Attribute::NoAlias);
2617
2618 // Add 'dereferenceable' and 'alignment'.
2619 auto PTy = ParamType->getPointeeType();
2620 if (!PTy->isIncompleteType() && PTy->isConstantSizeType()) {
2621 auto info = getContext().getTypeInfoInChars(PTy);
2622 Attrs.addDereferenceableAttr(info.Width.getQuantity());
2623 Attrs.addAlignmentAttr(info.Align.getAsAlign());
2624 }
2625 break;
2626 }
2627
2628 case ParameterABI::SwiftErrorResult:
2629 Attrs.addAttribute(llvm::Attribute::SwiftError);
2630 break;
2631
2632 case ParameterABI::SwiftContext:
2633 Attrs.addAttribute(llvm::Attribute::SwiftSelf);
2634 break;
2635
2636 case ParameterABI::SwiftAsyncContext:
2637 Attrs.addAttribute(llvm::Attribute::SwiftAsync);
2638 break;
2639 }
2640
2641 if (FI.getExtParameterInfo(ArgNo).isNoEscape())
2642 Attrs.addAttribute(llvm::Attribute::NoCapture);
2643
2644 if (Attrs.hasAttributes()) {
2645 unsigned FirstIRArg, NumIRArgs;
2646 std::tie(FirstIRArg, NumIRArgs) = IRFunctionArgs.getIRArgs(ArgNo);
2647 for (unsigned i = 0; i < NumIRArgs; i++)
2648 ArgAttrs[FirstIRArg + i] = ArgAttrs[FirstIRArg + i].addAttributes(
2649 getLLVMContext(), llvm::AttributeSet::get(getLLVMContext(), Attrs));
2650 }
2651 }
2652 assert(ArgNo == FI.arg_size());
2653
2654 AttrList = llvm::AttributeList::get(
2655 getLLVMContext(), llvm::AttributeSet::get(getLLVMContext(), FuncAttrs),
2656 llvm::AttributeSet::get(getLLVMContext(), RetAttrs), ArgAttrs);
2657 }
2658
2659 /// An argument came in as a promoted argument; demote it back to its
2660 /// declared type.
2661 static llvm::Value *emitArgumentDemotion(CodeGenFunction &CGF,
2662 const VarDecl *var,
2663 llvm::Value *value) {
2664 llvm::Type *varType = CGF.ConvertType(var->getType());
2665
2666 // This can happen with promotions that actually don't change the
2667 // underlying type, like the enum promotions.
2668 if (value->getType() == varType) return value;
2669
2670 assert((varType->isIntegerTy() || varType->isFloatingPointTy())
2671 && "unexpected promotion type");
2672
2673 if (isa<llvm::IntegerType>(varType))
2674 return CGF.Builder.CreateTrunc(value, varType, "arg.unpromote");
2675
2676 return CGF.Builder.CreateFPCast(value, varType, "arg.unpromote");
2677 }
2678
2679 /// Returns the attribute (either parameter attribute, or function
2680 /// attribute), which declares argument ArgNo to be non-null.
2681 static const NonNullAttr *getNonNullAttr(const Decl *FD, const ParmVarDecl *PVD,
2682 QualType ArgType, unsigned ArgNo) {
2683 // FIXME: __attribute__((nonnull)) can also be applied to:
2684 // - references to pointers, where the pointee is known to be
2685 // nonnull (apparently a Clang extension)
2686 // - transparent unions containing pointers
2687 // In the former case, LLVM IR cannot represent the constraint. In
2688 // the latter case, we have no guarantee that the transparent union
2689 // is in fact passed as a pointer.
2690 if (!ArgType->isAnyPointerType() && !ArgType->isBlockPointerType())
2691 return nullptr;
2692 // First, check attribute on parameter itself.
2693 if (PVD) {
2694 if (auto ParmNNAttr = PVD->getAttr<NonNullAttr>())
2695 return ParmNNAttr;
2696 }
2697 // Check function attributes.
2698 if (!FD)
2699 return nullptr;
2700 for (const auto *NNAttr : FD->specific_attrs<NonNullAttr>()) {
2701 if (NNAttr->isNonNull(ArgNo))
2702 return NNAttr;
2703 }
2704 return nullptr;
2705 }
2706
2707 namespace {
2708 struct CopyBackSwiftError final : EHScopeStack::Cleanup {
2709 Address Temp;
2710 Address Arg;
2711 CopyBackSwiftError(Address temp, Address arg) : Temp(temp), Arg(arg) {}
2712 void Emit(CodeGenFunction &CGF, Flags flags) override {
2713 llvm::Value *errorValue = CGF.Builder.CreateLoad(Temp);
2714 CGF.Builder.CreateStore(errorValue, Arg);
2715 }
2716 };
2717 }
2718
2719 void CodeGenFunction::EmitFunctionProlog(const CGFunctionInfo &FI,
2720 llvm::Function *Fn,
2721 const FunctionArgList &Args) {
2722 if (CurCodeDecl && CurCodeDecl->hasAttr<NakedAttr>())
2723 // Naked functions don't have prologues.
2724 return;
2725
2726 // If this is an implicit-return-zero function, go ahead and
2727 // initialize the return value. TODO: it might be nice to have
2728 // a more general mechanism for this that didn't require synthesized
2729 // return statements.
2730 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(CurCodeDecl)) {
2731 if (FD->hasImplicitReturnZero()) {
2732 QualType RetTy = FD->getReturnType().getUnqualifiedType();
2733 llvm::Type* LLVMTy = CGM.getTypes().ConvertType(RetTy);
2734 llvm::Constant* Zero = llvm::Constant::getNullValue(LLVMTy);
2735 Builder.CreateStore(Zero, ReturnValue);
2736 }
2737 }
2738
2739 // FIXME: We no longer need the types from FunctionArgList; lift up and
2740 // simplify.
2741
2742 ClangToLLVMArgMapping IRFunctionArgs(CGM.getContext(), FI);
2743 assert(Fn->arg_size() == IRFunctionArgs.totalIRArgs());
2744
2745 // If we're using inalloca, all the memory arguments are GEPs off of the last
2746 // parameter, which is a pointer to the complete memory area.
2747 Address ArgStruct = Address::invalid();
2748 if (IRFunctionArgs.hasInallocaArg()) {
2749 ArgStruct = Address(Fn->getArg(IRFunctionArgs.getInallocaArgNo()),
2750 FI.getArgStruct(), FI.getArgStructAlignment());
2751
2752 assert(ArgStruct.getType() == FI.getArgStruct()->getPointerTo());
2753 }
2754
2755 // Name the struct return parameter.
2756 if (IRFunctionArgs.hasSRetArg()) {
2757 auto AI = Fn->getArg(IRFunctionArgs.getSRetArgNo());
2758 AI->setName("agg.result");
2759 AI->addAttr(llvm::Attribute::NoAlias);
2760 }
2761
2762 // Track if we received the parameter as a pointer (indirect, byval, or
2763 // inalloca). If already have a pointer, EmitParmDecl doesn't need to copy it
2764 // into a local alloca for us.
2765 SmallVector<ParamValue, 16> ArgVals;
2766 ArgVals.reserve(Args.size());
2767
2768 // Create a pointer value for every parameter declaration. This usually
2769 // entails copying one or more LLVM IR arguments into an alloca. Don't push
2770 // any cleanups or do anything that might unwind. We do that separately, so
2771 // we can push the cleanups in the correct order for the ABI.
2772 assert(FI.arg_size() == Args.size() &&
2773 "Mismatch between function signature & arguments.");
2774 unsigned ArgNo = 0;
2775 CGFunctionInfo::const_arg_iterator info_it = FI.arg_begin();
2776 for (FunctionArgList::const_iterator i = Args.begin(), e = Args.end();
2777 i != e; ++i, ++info_it, ++ArgNo) {
2778 const VarDecl *Arg = *i;
2779 const ABIArgInfo &ArgI = info_it->info;
2780
2781 bool isPromoted =
2782 isa<ParmVarDecl>(Arg) && cast<ParmVarDecl>(Arg)->isKNRPromoted();
2783 // We are converting from ABIArgInfo type to VarDecl type directly, unless
2784 // the parameter is promoted. In this case we convert to
2785 // CGFunctionInfo::ArgInfo type with subsequent argument demotion.
2786 QualType Ty = isPromoted ? info_it->type : Arg->getType();
2787 assert(hasScalarEvaluationKind(Ty) ==
2788 hasScalarEvaluationKind(Arg->getType()));
2789
2790 unsigned FirstIRArg, NumIRArgs;
2791 std::tie(FirstIRArg, NumIRArgs) = IRFunctionArgs.getIRArgs(ArgNo);
2792
2793 switch (ArgI.getKind()) {
2794 case ABIArgInfo::InAlloca: {
2795 assert(NumIRArgs == 0);
2796 auto FieldIndex = ArgI.getInAllocaFieldIndex();
2797 Address V =
2798 Builder.CreateStructGEP(ArgStruct, FieldIndex, Arg->getName());
2799 if (ArgI.getInAllocaIndirect())
2800 V = Address(Builder.CreateLoad(V), ConvertTypeForMem(Ty),
2801 getContext().getTypeAlignInChars(Ty));
2802 ArgVals.push_back(ParamValue::forIndirect(V));
2803 break;
2804 }
2805
2806 case ABIArgInfo::Indirect:
2807 case ABIArgInfo::IndirectAliased: {
2808 assert(NumIRArgs == 1);
2809 Address ParamAddr = Address(Fn->getArg(FirstIRArg), ConvertTypeForMem(Ty),
2810 ArgI.getIndirectAlign());
2811
2812 if (!hasScalarEvaluationKind(Ty)) {
2813 // Aggregates and complex variables are accessed by reference. All we
2814 // need to do is realign the value, if requested. Also, if the address
2815 // may be aliased, copy it to ensure that the parameter variable is
2816 // mutable and has a unique adress, as C requires.
2817 Address V = ParamAddr;
2818 if (ArgI.getIndirectRealign() || ArgI.isIndirectAliased()) {
2819 Address AlignedTemp = CreateMemTemp(Ty, "coerce");
2820
2821 // Copy from the incoming argument pointer to the temporary with the
2822 // appropriate alignment.
2823 //
2824 // FIXME: We should have a common utility for generating an aggregate
2825 // copy.
2826 CharUnits Size = getContext().getTypeSizeInChars(Ty);
2827 Builder.CreateMemCpy(
2828 AlignedTemp.getPointer(), AlignedTemp.getAlignment().getAsAlign(),
2829 ParamAddr.getPointer(), ParamAddr.getAlignment().getAsAlign(),
2830 llvm::ConstantInt::get(IntPtrTy, Size.getQuantity()));
2831 V = AlignedTemp;
2832 }
2833 ArgVals.push_back(ParamValue::forIndirect(V));
2834 } else {
2835 // Load scalar value from indirect argument.
2836 llvm::Value *V =
2837 EmitLoadOfScalar(ParamAddr, false, Ty, Arg->getBeginLoc());
2838
2839 if (isPromoted)
2840 V = emitArgumentDemotion(*this, Arg, V);
2841 ArgVals.push_back(ParamValue::forDirect(V));
2842 }
2843 break;
2844 }
2845
2846 case ABIArgInfo::Extend:
2847 case ABIArgInfo::Direct: {
2848 auto AI = Fn->getArg(FirstIRArg);
2849 llvm::Type *LTy = ConvertType(Arg->getType());
2850
2851 // Prepare parameter attributes. So far, only attributes for pointer
2852 // parameters are prepared. See
2853 // http://llvm.org/docs/LangRef.html#paramattrs.
2854 if (ArgI.getDirectOffset() == 0 && LTy->isPointerTy() &&
2855 ArgI.getCoerceToType()->isPointerTy()) {
2856 assert(NumIRArgs == 1);
2857
2858 if (const ParmVarDecl *PVD = dyn_cast<ParmVarDecl>(Arg)) {
2859 // Set `nonnull` attribute if any.
2860 if (getNonNullAttr(CurCodeDecl, PVD, PVD->getType(),
2861 PVD->getFunctionScopeIndex()) &&
2862 !CGM.getCodeGenOpts().NullPointerIsValid)
2863 AI->addAttr(llvm::Attribute::NonNull);
2864
2865 QualType OTy = PVD->getOriginalType();
2866 if (const auto *ArrTy =
2867 getContext().getAsConstantArrayType(OTy)) {
2868 // A C99 array parameter declaration with the static keyword also
2869 // indicates dereferenceability, and if the size is constant we can
2870 // use the dereferenceable attribute (which requires the size in
2871 // bytes).
2872 if (ArrTy->getSizeModifier() == ArrayType::Static) {
2873 QualType ETy = ArrTy->getElementType();
2874 llvm::Align Alignment =
2875 CGM.getNaturalTypeAlignment(ETy).getAsAlign();
2876 AI->addAttrs(llvm::AttrBuilder(getLLVMContext()).addAlignmentAttr(Alignment));
2877 uint64_t ArrSize = ArrTy->getSize().getZExtValue();
2878 if (!ETy->isIncompleteType() && ETy->isConstantSizeType() &&
2879 ArrSize) {
2880 llvm::AttrBuilder Attrs(getLLVMContext());
2881 Attrs.addDereferenceableAttr(
2882 getContext().getTypeSizeInChars(ETy).getQuantity() *
2883 ArrSize);
2884 AI->addAttrs(Attrs);
2885 } else if (getContext().getTargetInfo().getNullPointerValue(
2886 ETy.getAddressSpace()) == 0 &&
2887 !CGM.getCodeGenOpts().NullPointerIsValid) {
2888 AI->addAttr(llvm::Attribute::NonNull);
2889 }
2890 }
2891 } else if (const auto *ArrTy =
2892 getContext().getAsVariableArrayType(OTy)) {
2893 // For C99 VLAs with the static keyword, we don't know the size so
2894 // we can't use the dereferenceable attribute, but in addrspace(0)
2895 // we know that it must be nonnull.
2896 if (ArrTy->getSizeModifier() == VariableArrayType::Static) {
2897 QualType ETy = ArrTy->getElementType();
2898 llvm::Align Alignment =
2899 CGM.getNaturalTypeAlignment(ETy).getAsAlign();
2900 AI->addAttrs(llvm::AttrBuilder(getLLVMContext()).addAlignmentAttr(Alignment));
2901 if (!getTypes().getTargetAddressSpace(ETy) &&
2902 !CGM.getCodeGenOpts().NullPointerIsValid)
2903 AI->addAttr(llvm::Attribute::NonNull);
2904 }
2905 }
2906
2907 // Set `align` attribute if any.
2908 const auto *AVAttr = PVD->getAttr<AlignValueAttr>();
2909 if (!AVAttr)
2910 if (const auto *TOTy = OTy->getAs<TypedefType>())
2911 AVAttr = TOTy->getDecl()->getAttr<AlignValueAttr>();
2912 if (AVAttr && !SanOpts.has(SanitizerKind::Alignment)) {
2913 // If alignment-assumption sanitizer is enabled, we do *not* add
2914 // alignment attribute here, but emit normal alignment assumption,
2915 // so the UBSAN check could function.
2916 llvm::ConstantInt *AlignmentCI =
2917 cast<llvm::ConstantInt>(EmitScalarExpr(AVAttr->getAlignment()));
2918 uint64_t AlignmentInt =
2919 AlignmentCI->getLimitedValue(llvm::Value::MaximumAlignment);
2920 if (AI->getParamAlign().valueOrOne() < AlignmentInt) {
2921 AI->removeAttr(llvm::Attribute::AttrKind::Alignment);
2922 AI->addAttrs(llvm::AttrBuilder(getLLVMContext()).addAlignmentAttr(
2923 llvm::Align(AlignmentInt)));
2924 }
2925 }
2926 }
2927
2928 // Set 'noalias' if an argument type has the `restrict` qualifier.
2929 if (Arg->getType().isRestrictQualified())
2930 AI->addAttr(llvm::Attribute::NoAlias);
2931 }
2932
2933 // Prepare the argument value. If we have the trivial case, handle it
2934 // with no muss and fuss.
2935 if (!isa<llvm::StructType>(ArgI.getCoerceToType()) &&
2936 ArgI.getCoerceToType() == ConvertType(Ty) &&
2937 ArgI.getDirectOffset() == 0) {
2938 assert(NumIRArgs == 1);
2939
2940 // LLVM expects swifterror parameters to be used in very restricted
2941 // ways. Copy the value into a less-restricted temporary.
2942 llvm::Value *V = AI;
2943 if (FI.getExtParameterInfo(ArgNo).getABI()
2944 == ParameterABI::SwiftErrorResult) {
2945 QualType pointeeTy = Ty->getPointeeType();
2946 assert(pointeeTy->isPointerType());
2947 Address temp =
2948 CreateMemTemp(pointeeTy, getPointerAlign(), "swifterror.temp");
2949 Address arg(V, ConvertTypeForMem(pointeeTy),
2950 getContext().getTypeAlignInChars(pointeeTy));
2951 llvm::Value *incomingErrorValue = Builder.CreateLoad(arg);
2952 Builder.CreateStore(incomingErrorValue, temp);
2953 V = temp.getPointer();
2954
2955 // Push a cleanup to copy the value back at the end of the function.
2956 // The convention does not guarantee that the value will be written
2957 // back if the function exits with an unwind exception.
2958 EHStack.pushCleanup<CopyBackSwiftError>(NormalCleanup, temp, arg);
2959 }
2960
2961 // Ensure the argument is the correct type.
2962 if (V->getType() != ArgI.getCoerceToType())
2963 V = Builder.CreateBitCast(V, ArgI.getCoerceToType());
2964
2965 if (isPromoted)
2966 V = emitArgumentDemotion(*this, Arg, V);
2967
2968 // Because of merging of function types from multiple decls it is
2969 // possible for the type of an argument to not match the corresponding
2970 // type in the function type. Since we are codegening the callee
2971 // in here, add a cast to the argument type.
2972 llvm::Type *LTy = ConvertType(Arg->getType());
2973 if (V->getType() != LTy)
2974 V = Builder.CreateBitCast(V, LTy);
2975
2976 ArgVals.push_back(ParamValue::forDirect(V));
2977 break;
2978 }
2979
2980 // VLST arguments are coerced to VLATs at the function boundary for
2981 // ABI consistency. If this is a VLST that was coerced to
2982 // a VLAT at the function boundary and the types match up, use
2983 // llvm.vector.extract to convert back to the original VLST.
2984 if (auto *VecTyTo = dyn_cast<llvm::FixedVectorType>(ConvertType(Ty))) {
2985 llvm::Value *Coerced = Fn->getArg(FirstIRArg);
2986 if (auto *VecTyFrom =
2987 dyn_cast<llvm::ScalableVectorType>(Coerced->getType())) {
2988 // If we are casting a scalable 16 x i1 predicate vector to a fixed i8
2989 // vector, bitcast the source and use a vector extract.
2990 auto PredType =
2991 llvm::ScalableVectorType::get(Builder.getInt1Ty(), 16);
2992 if (VecTyFrom == PredType &&
2993 VecTyTo->getElementType() == Builder.getInt8Ty()) {
2994 VecTyFrom = llvm::ScalableVectorType::get(Builder.getInt8Ty(), 2);
2995 Coerced = Builder.CreateBitCast(Coerced, VecTyFrom);
2996 }
2997 if (VecTyFrom->getElementType() == VecTyTo->getElementType()) {
2998 llvm::Value *Zero = llvm::Constant::getNullValue(CGM.Int64Ty);
2999
3000 assert(NumIRArgs == 1);
3001 Coerced->setName(Arg->getName() + ".coerce");
3002 ArgVals.push_back(ParamValue::forDirect(Builder.CreateExtractVector(
3003 VecTyTo, Coerced, Zero, "castFixedSve")));
3004 break;
3005 }
3006 }
3007 }
3008
3009 Address Alloca = CreateMemTemp(Ty, getContext().getDeclAlign(Arg),
3010 Arg->getName());
3011
3012 // Pointer to store into.
3013 Address Ptr = emitAddressAtOffset(*this, Alloca, ArgI);
3014
3015 // Fast-isel and the optimizer generally like scalar values better than
3016 // FCAs, so we flatten them if this is safe to do for this argument.
3017 llvm::StructType *STy = dyn_cast<llvm::StructType>(ArgI.getCoerceToType());
3018 if (ArgI.isDirect() && ArgI.getCanBeFlattened() && STy &&
3019 STy->getNumElements() > 1) {
3020 uint64_t SrcSize = CGM.getDataLayout().getTypeAllocSize(STy);
3021 llvm::Type *DstTy = Ptr.getElementType();
3022 uint64_t DstSize = CGM.getDataLayout().getTypeAllocSize(DstTy);
3023
3024 Address AddrToStoreInto = Address::invalid();
3025 if (SrcSize <= DstSize) {
3026 AddrToStoreInto = Builder.CreateElementBitCast(Ptr, STy);
3027 } else {
3028 AddrToStoreInto =
3029 CreateTempAlloca(STy, Alloca.getAlignment(), "coerce");
3030 }
3031
3032 assert(STy->getNumElements() == NumIRArgs);
3033 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
3034 auto AI = Fn->getArg(FirstIRArg + i);
3035 AI->setName(Arg->getName() + ".coerce" + Twine(i));
3036 Address EltPtr = Builder.CreateStructGEP(AddrToStoreInto, i);
3037 Builder.CreateStore(AI, EltPtr);
3038 }
3039
3040 if (SrcSize > DstSize) {
3041 Builder.CreateMemCpy(Ptr, AddrToStoreInto, DstSize);
3042 }
3043
3044 } else {
3045 // Simple case, just do a coerced store of the argument into the alloca.
3046 assert(NumIRArgs == 1);
3047 auto AI = Fn->getArg(FirstIRArg);
3048 AI->setName(Arg->getName() + ".coerce");
3049 CreateCoercedStore(AI, Ptr, /*DstIsVolatile=*/false, *this);
3050 }
3051
3052 // Match to what EmitParmDecl is expecting for this type.
3053 if (CodeGenFunction::hasScalarEvaluationKind(Ty)) {
3054 llvm::Value *V =
3055 EmitLoadOfScalar(Alloca, false, Ty, Arg->getBeginLoc());
3056 if (isPromoted)
3057 V = emitArgumentDemotion(*this, Arg, V);
3058 ArgVals.push_back(ParamValue::forDirect(V));
3059 } else {
3060 ArgVals.push_back(ParamValue::forIndirect(Alloca));
3061 }
3062 break;
3063 }
3064
3065 case ABIArgInfo::CoerceAndExpand: {
3066 // Reconstruct into a temporary.
3067 Address alloca = CreateMemTemp(Ty, getContext().getDeclAlign(Arg));
3068 ArgVals.push_back(ParamValue::forIndirect(alloca));
3069
3070 auto coercionType = ArgI.getCoerceAndExpandType();
3071 alloca = Builder.CreateElementBitCast(alloca, coercionType);
3072
3073 unsigned argIndex = FirstIRArg;
3074 for (unsigned i = 0, e = coercionType->getNumElements(); i != e; ++i) {
3075 llvm::Type *eltType = coercionType->getElementType(i);
3076 if (ABIArgInfo::isPaddingForCoerceAndExpand(eltType))
3077 continue;
3078
3079 auto eltAddr = Builder.CreateStructGEP(alloca, i);
3080 auto elt = Fn->getArg(argIndex++);
3081 Builder.CreateStore(elt, eltAddr);
3082 }
3083 assert(argIndex == FirstIRArg + NumIRArgs);
3084 break;
3085 }
3086
3087 case ABIArgInfo::Expand: {
3088 // If this structure was expanded into multiple arguments then
3089 // we need to create a temporary and reconstruct it from the
3090 // arguments.
3091 Address Alloca = CreateMemTemp(Ty, getContext().getDeclAlign(Arg));
3092 LValue LV = MakeAddrLValue(Alloca, Ty);
3093 ArgVals.push_back(ParamValue::forIndirect(Alloca));
3094
3095 auto FnArgIter = Fn->arg_begin() + FirstIRArg;
3096 ExpandTypeFromArgs(Ty, LV, FnArgIter);
3097 assert(FnArgIter == Fn->arg_begin() + FirstIRArg + NumIRArgs);
3098 for (unsigned i = 0, e = NumIRArgs; i != e; ++i) {
3099 auto AI = Fn->getArg(FirstIRArg + i);
3100 AI->setName(Arg->getName() + "." + Twine(i));
3101 }
3102 break;
3103 }
3104
3105 case ABIArgInfo::Ignore:
3106 assert(NumIRArgs == 0);
3107 // Initialize the local variable appropriately.
3108 if (!hasScalarEvaluationKind(Ty)) {
3109 ArgVals.push_back(ParamValue::forIndirect(CreateMemTemp(Ty)));
3110 } else {
3111 llvm::Value *U = llvm::UndefValue::get(ConvertType(Arg->getType()));
3112 ArgVals.push_back(ParamValue::forDirect(U));
3113 }
3114 break;
3115 }
3116 }
3117
3118 if (getTarget().getCXXABI().areArgsDestroyedLeftToRightInCallee()) {
3119 for (int I = Args.size() - 1; I >= 0; --I)
3120 EmitParmDecl(*Args[I], ArgVals[I], I + 1);
3121 } else {
3122 for (unsigned I = 0, E = Args.size(); I != E; ++I)
3123 EmitParmDecl(*Args[I], ArgVals[I], I + 1);
3124 }
3125 }
3126
3127 static void eraseUnusedBitCasts(llvm::Instruction *insn) {
3128 while (insn->use_empty()) {
3129 llvm::BitCastInst *bitcast = dyn_cast<llvm::BitCastInst>(insn);
3130 if (!bitcast) return;
3131
3132 // This is "safe" because we would have used a ConstantExpr otherwise.
3133 insn = cast<llvm::Instruction>(bitcast->getOperand(0));
3134 bitcast->eraseFromParent();
3135 }
3136 }
3137
3138 /// Try to emit a fused autorelease of a return result.
3139 static llvm::Value *tryEmitFusedAutoreleaseOfResult(CodeGenFunction &CGF,
3140 llvm::Value *result) {
3141 // We must be immediately followed the cast.
3142 llvm::BasicBlock *BB = CGF.Builder.GetInsertBlock();
3143 if (BB->empty()) return nullptr;
3144 if (&BB->back() != result) return nullptr;
3145
3146 llvm::Type *resultType = result->getType();
3147
3148 // result is in a BasicBlock and is therefore an Instruction.
3149 llvm::Instruction *generator = cast<llvm::Instruction>(result);
3150
3151 SmallVector<llvm::Instruction *, 4> InstsToKill;
3152
3153 // Look for:
3154 // %generator = bitcast %type1* %generator2 to %type2*
3155 while (llvm::BitCastInst *bitcast = dyn_cast<llvm::BitCastInst>(generator)) {
3156 // We would have emitted this as a constant if the operand weren't
3157 // an Instruction.
3158 generator = cast<llvm::Instruction>(bitcast->getOperand(0));
3159
3160 // Require the generator to be immediately followed by the cast.
3161 if (generator->getNextNode() != bitcast)
3162 return nullptr;
3163
3164 InstsToKill.push_back(bitcast);
3165 }
3166
3167 // Look for:
3168 // %generator = call i8* @objc_retain(i8* %originalResult)
3169 // or
3170 // %generator = call i8* @objc_retainAutoreleasedReturnValue(i8* %originalResult)
3171 llvm::CallInst *call = dyn_cast<llvm::CallInst>(generator);
3172 if (!call) return nullptr;
3173
3174 bool doRetainAutorelease;
3175
3176 if (call->getCalledOperand() == CGF.CGM.getObjCEntrypoints().objc_retain) {
3177 doRetainAutorelease = true;
3178 } else if (call->getCalledOperand() ==
3179 CGF.CGM.getObjCEntrypoints().objc_retainAutoreleasedReturnValue) {
3180 doRetainAutorelease = false;
3181
3182 // If we emitted an assembly marker for this call (and the
3183 // ARCEntrypoints field should have been set if so), go looking
3184 // for that call. If we can't find it, we can't do this
3185 // optimization. But it should always be the immediately previous
3186 // instruction, unless we needed bitcasts around the call.
3187 if (CGF.CGM.getObjCEntrypoints().retainAutoreleasedReturnValueMarker) {
3188 llvm::Instruction *prev = call->getPrevNode();
3189 assert(prev);
3190 if (isa<llvm::BitCastInst>(prev)) {
3191 prev = prev->getPrevNode();
3192 assert(prev);
3193 }
3194 assert(isa<llvm::CallInst>(prev));
3195 assert(cast<llvm::CallInst>(prev)->getCalledOperand() ==
3196 CGF.CGM.getObjCEntrypoints().retainAutoreleasedReturnValueMarker);
3197 InstsToKill.push_back(prev);
3198 }
3199 } else {
3200 return nullptr;
3201 }
3202
3203 result = call->getArgOperand(0);
3204 InstsToKill.push_back(call);
3205
3206 // Keep killing bitcasts, for sanity. Note that we no longer care
3207 // about precise ordering as long as there's exactly one use.
3208 while (llvm::BitCastInst *bitcast = dyn_cast<llvm::BitCastInst>(result)) {
3209 if (!bitcast->hasOneUse()) break;
3210 InstsToKill.push_back(bitcast);
3211 result = bitcast->getOperand(0);
3212 }
3213
3214 // Delete all the unnecessary instructions, from latest to earliest.
3215 for (auto *I : InstsToKill)
3216 I->eraseFromParent();
3217
3218 // Do the fused retain/autorelease if we were asked to.
3219 if (doRetainAutorelease)
3220 result = CGF.EmitARCRetainAutoreleaseReturnValue(result);
3221
3222 // Cast back to the result type.
3223 return CGF.Builder.CreateBitCast(result, resultType);
3224 }
3225
3226 /// If this is a +1 of the value of an immutable 'self', remove it.
3227 static llvm::Value *tryRemoveRetainOfSelf(CodeGenFunction &CGF,
3228 llvm::Value *result) {
3229 // This is only applicable to a method with an immutable 'self'.
3230 const ObjCMethodDecl *method =
3231 dyn_cast_or_null<ObjCMethodDecl>(CGF.CurCodeDecl);
3232 if (!method) return nullptr;
3233 const VarDecl *self = method->getSelfDecl();
3234 if (!self->getType().isConstQualified()) return nullptr;
3235
3236 // Look for a retain call.
3237 llvm::CallInst *retainCall =
3238 dyn_cast<llvm::CallInst>(result->stripPointerCasts());
3239 if (!retainCall || retainCall->getCalledOperand() !=
3240 CGF.CGM.getObjCEntrypoints().objc_retain)
3241 return nullptr;
3242
3243 // Look for an ordinary load of 'self'.
3244 llvm::Value *retainedValue = retainCall->getArgOperand(0);
3245 llvm::LoadInst *load =
3246 dyn_cast<llvm::LoadInst>(retainedValue->stripPointerCasts());
3247 if (!load || load->isAtomic() || load->isVolatile() ||
3248 load->getPointerOperand() != CGF.GetAddrOfLocalVar(self).getPointer())
3249 return nullptr;
3250
3251 // Okay! Burn it all down. This relies for correctness on the
3252 // assumption that the retain is emitted as part of the return and
3253 // that thereafter everything is used "linearly".
3254 llvm::Type *resultType = result->getType();
3255 eraseUnusedBitCasts(cast<llvm::Instruction>(result));
3256 assert(retainCall->use_empty());
3257 retainCall->eraseFromParent();
3258 eraseUnusedBitCasts(cast<llvm::Instruction>(retainedValue));
3259
3260 return CGF.Builder.CreateBitCast(load, resultType);
3261 }
3262
3263 /// Emit an ARC autorelease of the result of a function.
3264 ///
3265 /// \return the value to actually return from the function
3266 static llvm::Value *emitAutoreleaseOfResult(CodeGenFunction &CGF,
3267 llvm::Value *result) {
3268 // If we're returning 'self', kill the initial retain. This is a
3269 // heuristic attempt to "encourage correctness" in the really unfortunate
3270 // case where we have a return of self during a dealloc and we desperately
3271 // need to avoid the possible autorelease.
3272 if (llvm::Value *self = tryRemoveRetainOfSelf(CGF, result))
3273 return self;
3274
3275 // At -O0, try to emit a fused retain/autorelease.
3276 if (CGF.shouldUseFusedARCCalls())
3277 if (llvm::Value *fused = tryEmitFusedAutoreleaseOfResult(CGF, result))
3278 return fused;
3279
3280 return CGF.EmitARCAutoreleaseReturnValue(result);
3281 }
3282
3283 /// Heuristically search for a dominating store to the return-value slot.
3284 static llvm::StoreInst *findDominatingStoreToReturnValue(CodeGenFunction &CGF) {
3285 // Check if a User is a store which pointerOperand is the ReturnValue.
3286 // We are looking for stores to the ReturnValue, not for stores of the
3287 // ReturnValue to some other location.
3288 auto GetStoreIfValid = [&CGF](llvm::User *U) -> llvm::StoreInst * {
3289 auto *SI = dyn_cast<llvm::StoreInst>(U);
3290 if (!SI || SI->getPointerOperand() != CGF.ReturnValue.getPointer() ||
3291 SI->getValueOperand()->getType() != CGF.ReturnValue.getElementType())
3292 return nullptr;
3293 // These aren't actually possible for non-coerced returns, and we
3294 // only care about non-coerced returns on this code path.
3295 assert(!SI->isAtomic() && !SI->isVolatile());
3296 return SI;
3297 };
3298 // If there are multiple uses of the return-value slot, just check
3299 // for something immediately preceding the IP. Sometimes this can
3300 // happen with how we generate implicit-returns; it can also happen
3301 // with noreturn cleanups.
3302 if (!CGF.ReturnValue.getPointer()->hasOneUse()) {
3303 llvm::BasicBlock *IP = CGF.Builder.GetInsertBlock();
3304 if (IP->empty()) return nullptr;
3305
3306 // Look at directly preceding instruction, skipping bitcasts and lifetime
3307 // markers.
3308 for (llvm::Instruction &I : make_range(IP->rbegin(), IP->rend())) {
3309 if (isa<llvm::BitCastInst>(&I))
3310 continue;
3311 if (auto *II = dyn_cast<llvm::IntrinsicInst>(&I))
3312 if (II->getIntrinsicID() == llvm::Intrinsic::lifetime_end)
3313 continue;
3314
3315 return GetStoreIfValid(&I);
3316 }
3317 return nullptr;
3318 }
3319
3320 llvm::StoreInst *store =
3321 GetStoreIfValid(CGF.ReturnValue.getPointer()->user_back());
3322 if (!store) return nullptr;
3323
3324 // Now do a first-and-dirty dominance check: just walk up the
3325 // single-predecessors chain from the current insertion point.
3326 llvm::BasicBlock *StoreBB = store->getParent();
3327 llvm::BasicBlock *IP = CGF.Builder.GetInsertBlock();
3328 while (IP != StoreBB) {
3329 if (!(IP = IP->getSinglePredecessor()))
3330 return nullptr;
3331 }
3332
3333 // Okay, the store's basic block dominates the insertion point; we
3334 // can do our thing.
3335 return store;
3336 }
3337
3338 // Helper functions for EmitCMSEClearRecord
3339
3340 // Set the bits corresponding to a field having width `BitWidth` and located at
3341 // offset `BitOffset` (from the least significant bit) within a storage unit of
3342 // `Bits.size()` bytes. Each element of `Bits` corresponds to one target byte.
3343 // Use little-endian layout, i.e.`Bits[0]` is the LSB.
3344 static void setBitRange(SmallVectorImpl<uint64_t> &Bits, int BitOffset,
3345 int BitWidth, int CharWidth) {
3346 assert(CharWidth <= 64);
3347 assert(static_cast<unsigned>(BitWidth) <= Bits.size() * CharWidth);
3348
3349 int Pos = 0;
3350 if (BitOffset >= CharWidth) {
3351 Pos += BitOffset / CharWidth;
3352 BitOffset = BitOffset % CharWidth;
3353 }
3354
3355 const uint64_t Used = (uint64_t(1) << CharWidth) - 1;
3356 if (BitOffset + BitWidth >= CharWidth) {
3357 Bits[Pos++] |= (Used << BitOffset) & Used;
3358 BitWidth -= CharWidth - BitOffset;
3359 BitOffset = 0;
3360 }
3361
3362 while (BitWidth >= CharWidth) {
3363 Bits[Pos++] = Used;
3364 BitWidth -= CharWidth;
3365 }
3366
3367 if (BitWidth > 0)
3368 Bits[Pos++] |= (Used >> (CharWidth - BitWidth)) << BitOffset;
3369 }
3370
3371 // Set the bits corresponding to a field having width `BitWidth` and located at
3372 // offset `BitOffset` (from the least significant bit) within a storage unit of
3373 // `StorageSize` bytes, located at `StorageOffset` in `Bits`. Each element of
3374 // `Bits` corresponds to one target byte. Use target endian layout.
3375 static void setBitRange(SmallVectorImpl<uint64_t> &Bits, int StorageOffset,
3376 int StorageSize, int BitOffset, int BitWidth,
3377 int CharWidth, bool BigEndian) {
3378
3379 SmallVector<uint64_t, 8> TmpBits(StorageSize);
3380 setBitRange(TmpBits, BitOffset, BitWidth, CharWidth);
3381
3382 if (BigEndian)
3383 std::reverse(TmpBits.begin(), TmpBits.end());
3384
3385 for (uint64_t V : TmpBits)
3386 Bits[StorageOffset++] |= V;
3387 }
3388
3389 static void setUsedBits(CodeGenModule &, QualType, int,
3390 SmallVectorImpl<uint64_t> &);
3391
3392 // Set the bits in `Bits`, which correspond to the value representations of
3393 // the actual members of the record type `RTy`. Note that this function does
3394 // not handle base classes, virtual tables, etc, since they cannot happen in
3395 // CMSE function arguments or return. The bit mask corresponds to the target
3396 // memory layout, i.e. it's endian dependent.
3397 static void setUsedBits(CodeGenModule &CGM, const RecordType *RTy, int Offset,
3398 SmallVectorImpl<uint64_t> &Bits) {
3399 ASTContext &Context = CGM.getContext();
3400 int CharWidth = Context.getCharWidth();
3401 const RecordDecl *RD = RTy->getDecl()->getDefinition();
3402 const ASTRecordLayout &ASTLayout = Context.getASTRecordLayout(RD);
3403 const CGRecordLayout &Layout = CGM.getTypes().getCGRecordLayout(RD);
3404
3405 int Idx = 0;
3406 for (auto I = RD->field_begin(), E = RD->field_end(); I != E; ++I, ++Idx) {
3407 const FieldDecl *F = *I;
3408
3409 if (F->isUnnamedBitfield() || F->isZeroLengthBitField(Context) ||
3410 F->getType()->isIncompleteArrayType())
3411 continue;
3412
3413 if (F->isBitField()) {
3414 const CGBitFieldInfo &BFI = Layout.getBitFieldInfo(F);
3415 setBitRange(Bits, Offset + BFI.StorageOffset.getQuantity(),
3416 BFI.StorageSize / CharWidth, BFI.Offset,
3417 BFI.Size, CharWidth,
3418 CGM.getDataLayout().isBigEndian());
3419 continue;
3420 }
3421
3422 setUsedBits(CGM, F->getType(),
3423 Offset + ASTLayout.getFieldOffset(Idx) / CharWidth, Bits);
3424 }
3425 }
3426
3427 // Set the bits in `Bits`, which correspond to the value representations of
3428 // the elements of an array type `ATy`.
3429 static void setUsedBits(CodeGenModule &CGM, const ConstantArrayType *ATy,
3430 int Offset, SmallVectorImpl<uint64_t> &Bits) {
3431 const ASTContext &Context = CGM.getContext();
3432
3433 QualType ETy = Context.getBaseElementType(ATy);
3434 int Size = Context.getTypeSizeInChars(ETy).getQuantity();
3435 SmallVector<uint64_t, 4> TmpBits(Size);
3436 setUsedBits(CGM, ETy, 0, TmpBits);
3437
3438 for (int I = 0, N = Context.getConstantArrayElementCount(ATy); I < N; ++I) {
3439 auto Src = TmpBits.begin();
3440 auto Dst = Bits.begin() + Offset + I * Size;
3441 for (int J = 0; J < Size; ++J)
3442 *Dst++ |= *Src++;
3443 }
3444 }
3445
3446 // Set the bits in `Bits`, which correspond to the value representations of
3447 // the type `QTy`.
3448 static void setUsedBits(CodeGenModule &CGM, QualType QTy, int Offset,
3449 SmallVectorImpl<uint64_t> &Bits) {
3450 if (const auto *RTy = QTy->getAs<RecordType>())
3451 return setUsedBits(CGM, RTy, Offset, Bits);
3452
3453 ASTContext &Context = CGM.getContext();
3454 if (const auto *ATy = Context.getAsConstantArrayType(QTy))
3455 return setUsedBits(CGM, ATy, Offset, Bits);
3456
3457 int Size = Context.getTypeSizeInChars(QTy).getQuantity();
3458 if (Size <= 0)
3459 return;
3460
3461 std::fill_n(Bits.begin() + Offset, Size,
3462 (uint64_t(1) << Context.getCharWidth()) - 1);
3463 }
3464
3465 static uint64_t buildMultiCharMask(const SmallVectorImpl<uint64_t> &Bits,
3466 int Pos, int Size, int CharWidth,
3467 bool BigEndian) {
3468 assert(Size > 0);
3469 uint64_t Mask = 0;
3470 if (BigEndian) {
3471 for (auto P = Bits.begin() + Pos, E = Bits.begin() + Pos + Size; P != E;
3472 ++P)
3473 Mask = (Mask << CharWidth) | *P;
3474 } else {
3475 auto P = Bits.begin() + Pos + Size, End = Bits.begin() + Pos;
3476 do
3477 Mask = (Mask << CharWidth) | *--P;
3478 while (P != End);
3479 }
3480 return Mask;
3481 }
3482
3483 // Emit code to clear the bits in a record, which aren't a part of any user
3484 // declared member, when the record is a function return.
3485 llvm::Value *CodeGenFunction::EmitCMSEClearRecord(llvm::Value *Src,
3486 llvm::IntegerType *ITy,
3487 QualType QTy) {
3488 assert(Src->getType() == ITy);
3489 assert(ITy->getScalarSizeInBits() <= 64);
3490
3491 const llvm::DataLayout &DataLayout = CGM.getDataLayout();
3492 int Size = DataLayout.getTypeStoreSize(ITy);
3493 SmallVector<uint64_t, 4> Bits(Size);
3494 setUsedBits(CGM, QTy->castAs<RecordType>(), 0, Bits);
3495
3496 int CharWidth = CGM.getContext().getCharWidth();
3497 uint64_t Mask =
3498 buildMultiCharMask(Bits, 0, Size, CharWidth, DataLayout.isBigEndian());
3499
3500 return Builder.CreateAnd(Src, Mask, "cmse.clear");
3501 }
3502
3503 // Emit code to clear the bits in a record, which aren't a part of any user
3504 // declared member, when the record is a function argument.
3505 llvm::Value *CodeGenFunction::EmitCMSEClearRecord(llvm::Value *Src,
3506 llvm::ArrayType *ATy,
3507 QualType QTy) {
3508 const llvm::DataLayout &DataLayout = CGM.getDataLayout();
3509 int Size = DataLayout.getTypeStoreSize(ATy);
3510 SmallVector<uint64_t, 16> Bits(Size);
3511 setUsedBits(CGM, QTy->castAs<RecordType>(), 0, Bits);
3512
3513 // Clear each element of the LLVM array.
3514 int CharWidth = CGM.getContext().getCharWidth();
3515 int CharsPerElt =
3516 ATy->getArrayElementType()->getScalarSizeInBits() / CharWidth;
3517 int MaskIndex = 0;
3518 llvm::Value *R = llvm::PoisonValue::get(ATy);
3519 for (int I = 0, N = ATy->getArrayNumElements(); I != N; ++I) {
3520 uint64_t Mask = buildMultiCharMask(Bits, MaskIndex, CharsPerElt, CharWidth,
3521 DataLayout.isBigEndian());
3522 MaskIndex += CharsPerElt;
3523 llvm::Value *T0 = Builder.CreateExtractValue(Src, I);
3524 llvm::Value *T1 = Builder.CreateAnd(T0, Mask, "cmse.clear");
3525 R = Builder.CreateInsertValue(R, T1, I);
3526 }
3527
3528 return R;
3529 }
3530
3531 void CodeGenFunction::EmitFunctionEpilog(const CGFunctionInfo &FI,
3532 bool EmitRetDbgLoc,
3533 SourceLocation EndLoc) {
3534 if (FI.isNoReturn()) {
3535 // Noreturn functions don't return.
3536 EmitUnreachable(EndLoc);
3537 return;
3538 }
3539
3540 if (CurCodeDecl && CurCodeDecl->hasAttr<NakedAttr>()) {
3541 // Naked functions don't have epilogues.
3542 Builder.CreateUnreachable();
3543 return;
3544 }
3545
3546 // Functions with no result always return void.
3547 if (!ReturnValue.isValid()) {
3548 Builder.CreateRetVoid();
3549 return;
3550 }
3551
3552 llvm::DebugLoc RetDbgLoc;
3553 llvm::Value *RV = nullptr;
3554 QualType RetTy = FI.getReturnType();
3555 const ABIArgInfo &RetAI = FI.getReturnInfo();
3556
3557 switch (RetAI.getKind()) {
3558 case ABIArgInfo::InAlloca:
3559 // Aggregates get evaluated directly into the destination. Sometimes we
3560 // need to return the sret value in a register, though.
3561 assert(hasAggregateEvaluationKind(RetTy));
3562 if (RetAI.getInAllocaSRet()) {
3563 llvm::Function::arg_iterator EI = CurFn->arg_end();
3564 --EI;
3565 llvm::Value *ArgStruct = &*EI;
3566 llvm::Value *SRet = Builder.CreateStructGEP(
3567 FI.getArgStruct(), ArgStruct, RetAI.getInAllocaFieldIndex());
3568 llvm::Type *Ty =
3569 cast<llvm::GetElementPtrInst>(SRet)->getResultElementType();
3570 RV = Builder.CreateAlignedLoad(Ty, SRet, getPointerAlign(), "sret");
3571 }
3572 break;
3573
3574 case ABIArgInfo::Indirect: {
3575 auto AI = CurFn->arg_begin();
3576 if (RetAI.isSRetAfterThis())
3577 ++AI;
3578 switch (getEvaluationKind(RetTy)) {
3579 case TEK_Complex: {
3580 ComplexPairTy RT =
3581 EmitLoadOfComplex(MakeAddrLValue(ReturnValue, RetTy), EndLoc);
3582 EmitStoreOfComplex(RT, MakeNaturalAlignAddrLValue(&*AI, RetTy),
3583 /*isInit*/ true);
3584 break;
3585 }
3586 case TEK_Aggregate:
3587 // Do nothing; aggregates get evaluated directly into the destination.
3588 break;
3589 case TEK_Scalar: {
3590 LValueBaseInfo BaseInfo;
3591 TBAAAccessInfo TBAAInfo;
3592 CharUnits Alignment =
3593 CGM.getNaturalTypeAlignment(RetTy, &BaseInfo, &TBAAInfo);
3594 Address ArgAddr(&*AI, ConvertType(RetTy), Alignment);
3595 LValue ArgVal =
3596 LValue::MakeAddr(ArgAddr, RetTy, getContext(), BaseInfo, TBAAInfo);
3597 EmitStoreOfScalar(
3598 Builder.CreateLoad(ReturnValue), ArgVal, /*isInit*/ true);
3599 break;
3600 }
3601 }
3602 break;
3603 }
3604
3605 case ABIArgInfo::Extend:
3606 case ABIArgInfo::Direct:
3607 if (RetAI.getCoerceToType() == ConvertType(RetTy) &&
3608 RetAI.getDirectOffset() == 0) {
3609 // The internal return value temp always will have pointer-to-return-type
3610 // type, just do a load.
3611
3612 // If there is a dominating store to ReturnValue, we can elide
3613 // the load, zap the store, and usually zap the alloca.
3614 if (llvm::StoreInst *SI =
3615 findDominatingStoreToReturnValue(*this)) {
3616 // Reuse the debug location from the store unless there is
3617 // cleanup code to be emitted between the store and return
3618 // instruction.
3619 if (EmitRetDbgLoc && !AutoreleaseResult)
3620 RetDbgLoc = SI->getDebugLoc();
3621 // Get the stored value and nuke the now-dead store.
3622 RV = SI->getValueOperand();
3623 SI->eraseFromParent();
3624
3625 // Otherwise, we have to do a simple load.
3626 } else {
3627 RV = Builder.CreateLoad(ReturnValue);
3628 }
3629 } else {
3630 // If the value is offset in memory, apply the offset now.
3631 Address V = emitAddressAtOffset(*this, ReturnValue, RetAI);
3632
3633 RV = CreateCoercedLoad(V, RetAI.getCoerceToType(), *this);
3634 }
3635
3636 // In ARC, end functions that return a retainable type with a call
3637 // to objc_autoreleaseReturnValue.
3638 if (AutoreleaseResult) {
3639 #ifndef NDEBUG
3640 // Type::isObjCRetainabletype has to be called on a QualType that hasn't
3641 // been stripped of the typedefs, so we cannot use RetTy here. Get the
3642 // original return type of FunctionDecl, CurCodeDecl, and BlockDecl from
3643 // CurCodeDecl or BlockInfo.
3644 QualType RT;
3645
3646 if (auto *FD = dyn_cast<FunctionDecl>(CurCodeDecl))
3647 RT = FD->getReturnType();
3648 else if (auto *MD = dyn_cast<ObjCMethodDecl>(CurCodeDecl))
3649 RT = MD->getReturnType();
3650 else if (isa<BlockDecl>(CurCodeDecl))
3651 RT = BlockInfo->BlockExpression->getFunctionType()->getReturnType();
3652 else
3653 llvm_unreachable("Unexpected function/method type");
3654
3655 assert(getLangOpts().ObjCAutoRefCount &&
3656 !FI.isReturnsRetained() &&
3657 RT->isObjCRetainableType());
3658 #endif
3659 RV = emitAutoreleaseOfResult(*this, RV);
3660 }
3661
3662 break;
3663
3664 case ABIArgInfo::Ignore:
3665 break;
3666
3667 case ABIArgInfo::CoerceAndExpand: {
3668 auto coercionType = RetAI.getCoerceAndExpandType();
3669
3670 // Load all of the coerced elements out into results.
3671 llvm::SmallVector<llvm::Value*, 4> results;
3672 Address addr = Builder.CreateElementBitCast(ReturnValue, coercionType);
3673 for (unsigned i = 0, e = coercionType->getNumElements(); i != e; ++i) {
3674 auto coercedEltType = coercionType->getElementType(i);
3675 if (ABIArgInfo::isPaddingForCoerceAndExpand(coercedEltType))
3676 continue;
3677
3678 auto eltAddr = Builder.CreateStructGEP(addr, i);
3679 auto elt = Builder.CreateLoad(eltAddr);
3680 results.push_back(elt);
3681 }
3682
3683 // If we have one result, it's the single direct result type.
3684 if (results.size() == 1) {
3685 RV = results[0];
3686
3687 // Otherwise, we need to make a first-class aggregate.
3688 } else {
3689 // Construct a return type that lacks padding elements.
3690 llvm::Type *returnType = RetAI.getUnpaddedCoerceAndExpandType();
3691
3692 RV = llvm::PoisonValue::get(returnType);
3693 for (unsigned i = 0, e = results.size(); i != e; ++i) {
3694 RV = Builder.CreateInsertValue(RV, results[i], i);
3695 }
3696 }
3697 break;
3698 }
3699 case ABIArgInfo::Expand:
3700 case ABIArgInfo::IndirectAliased:
3701 llvm_unreachable("Invalid ABI kind for return argument");
3702 }
3703
3704 llvm::Instruction *Ret;
3705 if (RV) {
3706 if (CurFuncDecl && CurFuncDecl->hasAttr<CmseNSEntryAttr>()) {
3707 // For certain return types, clear padding bits, as they may reveal
3708 // sensitive information.
3709 // Small struct/union types are passed as integers.
3710 auto *ITy = dyn_cast<llvm::IntegerType>(RV->getType());
3711 if (ITy != nullptr && isa<RecordType>(RetTy.getCanonicalType()))
3712 RV = EmitCMSEClearRecord(RV, ITy, RetTy);
3713 }
3714 EmitReturnValueCheck(RV);
3715 Ret = Builder.CreateRet(RV);
3716 } else {
3717 Ret = Builder.CreateRetVoid();
3718 }
3719
3720 if (RetDbgLoc)
3721 Ret->setDebugLoc(std::move(RetDbgLoc));
3722 }
3723
3724 void CodeGenFunction::EmitReturnValueCheck(llvm::Value *RV) {
3725 // A current decl may not be available when emitting vtable thunks.
3726 if (!CurCodeDecl)
3727 return;
3728
3729 // If the return block isn't reachable, neither is this check, so don't emit
3730 // it.
3731 if (ReturnBlock.isValid() && ReturnBlock.getBlock()->use_empty())
3732 return;
3733
3734 ReturnsNonNullAttr *RetNNAttr = nullptr;
3735 if (SanOpts.has(SanitizerKind::ReturnsNonnullAttribute))
3736 RetNNAttr = CurCodeDecl->getAttr<ReturnsNonNullAttr>();
3737
3738 if (!RetNNAttr && !requiresReturnValueNullabilityCheck())
3739 return;
3740
3741 // Prefer the returns_nonnull attribute if it's present.
3742 SourceLocation AttrLoc;
3743 SanitizerMask CheckKind;
3744 SanitizerHandler Handler;
3745 if (RetNNAttr) {
3746 assert(!requiresReturnValueNullabilityCheck() &&
3747 "Cannot check nullability and the nonnull attribute");
3748 AttrLoc = RetNNAttr->getLocation();
3749 CheckKind = SanitizerKind::ReturnsNonnullAttribute;
3750 Handler = SanitizerHandler::NonnullReturn;
3751 } else {
3752 if (auto *DD = dyn_cast<DeclaratorDecl>(CurCodeDecl))
3753 if (auto *TSI = DD->getTypeSourceInfo())
3754 if (auto FTL = TSI->getTypeLoc().getAsAdjusted<FunctionTypeLoc>())
3755 AttrLoc = FTL.getReturnLoc().findNullabilityLoc();
3756 CheckKind = SanitizerKind::NullabilityReturn;
3757 Handler = SanitizerHandler::NullabilityReturn;
3758 }
3759
3760 SanitizerScope SanScope(this);
3761
3762 // Make sure the "return" source location is valid. If we're checking a
3763 // nullability annotation, make sure the preconditions for the check are met.
3764 llvm::BasicBlock *Check = createBasicBlock("nullcheck");
3765 llvm::BasicBlock *NoCheck = createBasicBlock("no.nullcheck");
3766 llvm::Value *SLocPtr = Builder.CreateLoad(ReturnLocation, "return.sloc.load");
3767 llvm::Value *CanNullCheck = Builder.CreateIsNotNull(SLocPtr);
3768 if (requiresReturnValueNullabilityCheck())
3769 CanNullCheck =
3770 Builder.CreateAnd(CanNullCheck, RetValNullabilityPrecondition);
3771 Builder.CreateCondBr(CanNullCheck, Check, NoCheck);
3772 EmitBlock(Check);
3773
3774 // Now do the null check.
3775 llvm::Value *Cond = Builder.CreateIsNotNull(RV);
3776 llvm::Constant *StaticData[] = {EmitCheckSourceLocation(AttrLoc)};
3777 llvm::Value *DynamicData[] = {SLocPtr};
3778 EmitCheck(std::make_pair(Cond, CheckKind), Handler, StaticData, DynamicData);
3779
3780 EmitBlock(NoCheck);
3781
3782 #ifndef NDEBUG
3783 // The return location should not be used after the check has been emitted.
3784 ReturnLocation = Address::invalid();
3785 #endif
3786 }
3787
3788 static bool isInAllocaArgument(CGCXXABI &ABI, QualType type) {
3789 const CXXRecordDecl *RD = type->getAsCXXRecordDecl();
3790 return RD && ABI.getRecordArgABI(RD) == CGCXXABI::RAA_DirectInMemory;
3791 }
3792
3793 static AggValueSlot createPlaceholderSlot(CodeGenFunction &CGF,
3794 QualType Ty) {
3795 // FIXME: Generate IR in one pass, rather than going back and fixing up these
3796 // placeholders.
3797 llvm::Type *IRTy = CGF.ConvertTypeForMem(Ty);
3798 llvm::Type *IRPtrTy = IRTy->getPointerTo();
3799 llvm::Value *Placeholder = llvm::PoisonValue::get(IRPtrTy->getPointerTo());
3800
3801 // FIXME: When we generate this IR in one pass, we shouldn't need
3802 // this win32-specific alignment hack.
3803 CharUnits Align = CharUnits::fromQuantity(4);
3804 Placeholder = CGF.Builder.CreateAlignedLoad(IRPtrTy, Placeholder, Align);
3805
3806 return AggValueSlot::forAddr(Address(Placeholder, IRTy, Align),
3807 Ty.getQualifiers(),
3808 AggValueSlot::IsNotDestructed,
3809 AggValueSlot::DoesNotNeedGCBarriers,
3810 AggValueSlot::IsNotAliased,
3811 AggValueSlot::DoesNotOverlap);
3812 }
3813
3814 void CodeGenFunction::EmitDelegateCallArg(CallArgList &args,
3815 const VarDecl *param,
3816 SourceLocation loc) {
3817 // StartFunction converted the ABI-lowered parameter(s) into a
3818 // local alloca. We need to turn that into an r-value suitable
3819 // for EmitCall.
3820 Address local = GetAddrOfLocalVar(param);
3821
3822 QualType type = param->getType();
3823
3824 if (isInAllocaArgument(CGM.getCXXABI(), type)) {
3825 CGM.ErrorUnsupported(param, "forwarded non-trivially copyable parameter");
3826 }
3827
3828 // GetAddrOfLocalVar returns a pointer-to-pointer for references,
3829 // but the argument needs to be the original pointer.
3830 if (type->isReferenceType()) {
3831 args.add(RValue::get(Builder.CreateLoad(local)), type);
3832
3833 // In ARC, move out of consumed arguments so that the release cleanup
3834 // entered by StartFunction doesn't cause an over-release. This isn't
3835 // optimal -O0 code generation, but it should get cleaned up when
3836 // optimization is enabled. This also assumes that delegate calls are
3837 // performed exactly once for a set of arguments, but that should be safe.
3838 } else if (getLangOpts().ObjCAutoRefCount &&
3839 param->hasAttr<NSConsumedAttr>() &&
3840 type->isObjCRetainableType()) {
3841 llvm::Value *ptr = Builder.CreateLoad(local);
3842 auto null =
3843 llvm::ConstantPointerNull::get(cast<llvm::PointerType>(ptr->getType()));
3844 Builder.CreateStore(null, local);
3845 args.add(RValue::get(ptr), type);
3846
3847 // For the most part, we just need to load the alloca, except that
3848 // aggregate r-values are actually pointers to temporaries.
3849 } else {
3850 args.add(convertTempToRValue(local, type, loc), type);
3851 }
3852
3853 // Deactivate the cleanup for the callee-destructed param that was pushed.
3854 if (type->isRecordType() && !CurFuncIsThunk &&
3855 type->castAs<RecordType>()->getDecl()->isParamDestroyedInCallee() &&
3856 param->needsDestruction(getContext())) {
3857 EHScopeStack::stable_iterator cleanup =
3858 CalleeDestructedParamCleanups.lookup(cast<ParmVarDecl>(param));
3859 assert(cleanup.isValid() &&
3860 "cleanup for callee-destructed param not recorded");
3861 // This unreachable is a temporary marker which will be removed later.
3862 llvm::Instruction *isActive = Builder.CreateUnreachable();
3863 args.addArgCleanupDeactivation(cleanup, isActive);
3864 }
3865 }
3866
3867 static bool isProvablyNull(llvm::Value *addr) {
3868 return isa<llvm::ConstantPointerNull>(addr);
3869 }
3870
3871 /// Emit the actual writing-back of a writeback.
3872 static void emitWriteback(CodeGenFunction &CGF,
3873 const CallArgList::Writeback &writeback) {
3874 const LValue &srcLV = writeback.Source;
3875 Address srcAddr = srcLV.getAddress(CGF);
3876 assert(!isProvablyNull(srcAddr.getPointer()) &&
3877 "shouldn't have writeback for provably null argument");
3878
3879 llvm::BasicBlock *contBB = nullptr;
3880
3881 // If the argument wasn't provably non-null, we need to null check
3882 // before doing the store.
3883 bool provablyNonNull = llvm::isKnownNonZero(srcAddr.getPointer(),
3884 CGF.CGM.getDataLayout());
3885 if (!provablyNonNull) {
3886 llvm::BasicBlock *writebackBB = CGF.createBasicBlock("icr.writeback");
3887 contBB = CGF.createBasicBlock("icr.done");
3888
3889 llvm::Value *isNull =
3890 CGF.Builder.CreateIsNull(srcAddr.getPointer(), "icr.isnull");
3891 CGF.Builder.CreateCondBr(isNull, contBB, writebackBB);
3892 CGF.EmitBlock(writebackBB);
3893 }
3894
3895 // Load the value to writeback.
3896 llvm::Value *value = CGF.Builder.CreateLoad(writeback.Temporary);
3897
3898 // Cast it back, in case we're writing an id to a Foo* or something.
3899 value = CGF.Builder.CreateBitCast(value, srcAddr.getElementType(),
3900 "icr.writeback-cast");
3901
3902 // Perform the writeback.
3903
3904 // If we have a "to use" value, it's something we need to emit a use
3905 // of. This has to be carefully threaded in: if it's done after the
3906 // release it's potentially undefined behavior (and the optimizer
3907 // will ignore it), and if it happens before the retain then the
3908 // optimizer could move the release there.
3909 if (writeback.ToUse) {
3910 assert(srcLV.getObjCLifetime() == Qualifiers::OCL_Strong);
3911
3912 // Retain the new value. No need to block-copy here: the block's
3913 // being passed up the stack.
3914 value = CGF.EmitARCRetainNonBlock(value);
3915
3916 // Emit the intrinsic use here.
3917 CGF.EmitARCIntrinsicUse(writeback.ToUse);
3918
3919 // Load the old value (primitively).
3920 llvm::Value *oldValue = CGF.EmitLoadOfScalar(srcLV, SourceLocation());
3921
3922 // Put the new value in place (primitively).
3923 CGF.EmitStoreOfScalar(value, srcLV, /*init*/ false);
3924
3925 // Release the old value.
3926 CGF.EmitARCRelease(oldValue, srcLV.isARCPreciseLifetime());
3927
3928 // Otherwise, we can just do a normal lvalue store.
3929 } else {
3930 CGF.EmitStoreThroughLValue(RValue::get(value), srcLV);
3931 }
3932
3933 // Jump to the continuation block.
3934 if (!provablyNonNull)
3935 CGF.EmitBlock(contBB);
3936 }
3937
3938 static void emitWritebacks(CodeGenFunction &CGF,
3939 const CallArgList &args) {
3940 for (const auto &I : args.writebacks())
3941 emitWriteback(CGF, I);
3942 }
3943
3944 static void deactivateArgCleanupsBeforeCall(CodeGenFunction &CGF,
3945 const CallArgList &CallArgs) {
3946 ArrayRef<CallArgList::CallArgCleanup> Cleanups =
3947 CallArgs.getCleanupsToDeactivate();
3948 // Iterate in reverse to increase the likelihood of popping the cleanup.
3949 for (const auto &I : llvm::reverse(Cleanups)) {
3950 CGF.DeactivateCleanupBlock(I.Cleanup, I.IsActiveIP);
3951 I.IsActiveIP->eraseFromParent();
3952 }
3953 }
3954
3955 static const Expr *maybeGetUnaryAddrOfOperand(const Expr *E) {
3956 if (const UnaryOperator *uop = dyn_cast<UnaryOperator>(E->IgnoreParens()))
3957 if (uop->getOpcode() == UO_AddrOf)
3958 return uop->getSubExpr();
3959 return nullptr;
3960 }
3961
3962 /// Emit an argument that's being passed call-by-writeback. That is,
3963 /// we are passing the address of an __autoreleased temporary; it
3964 /// might be copy-initialized with the current value of the given
3965 /// address, but it will definitely be copied out of after the call.
3966 static void emitWritebackArg(CodeGenFunction &CGF, CallArgList &args,
3967 const ObjCIndirectCopyRestoreExpr *CRE) {
3968 LValue srcLV;
3969
3970 // Make an optimistic effort to emit the address as an l-value.
3971 // This can fail if the argument expression is more complicated.
3972 if (const Expr *lvExpr = maybeGetUnaryAddrOfOperand(CRE->getSubExpr())) {
3973 srcLV = CGF.EmitLValue(lvExpr);
3974
3975 // Otherwise, just emit it as a scalar.
3976 } else {
3977 Address srcAddr = CGF.EmitPointerWithAlignment(CRE->getSubExpr());
3978
3979 QualType srcAddrType =
3980 CRE->getSubExpr()->getType()->castAs<PointerType>()->getPointeeType();
3981 srcLV = CGF.MakeAddrLValue(srcAddr, srcAddrType);
3982 }
3983 Address srcAddr = srcLV.getAddress(CGF);
3984
3985 // The dest and src types don't necessarily match in LLVM terms
3986 // because of the crazy ObjC compatibility rules.
3987
3988 llvm::PointerType *destType =
3989 cast<llvm::PointerType>(CGF.ConvertType(CRE->getType()));
3990 llvm::Type *destElemType =
3991 CGF.ConvertTypeForMem(CRE->getType()->getPointeeType());
3992
3993 // If the address is a constant null, just pass the appropriate null.
3994 if (isProvablyNull(srcAddr.getPointer())) {
3995 args.add(RValue::get(llvm::ConstantPointerNull::get(destType)),
3996 CRE->getType());
3997 return;
3998 }
3999
4000 // Create the temporary.
4001 Address temp =
4002 CGF.CreateTempAlloca(destElemType, CGF.getPointerAlign(), "icr.temp");
4003 // Loading an l-value can introduce a cleanup if the l-value is __weak,
4004 // and that cleanup will be conditional if we can't prove that the l-value
4005 // isn't null, so we need to register a dominating point so that the cleanups
4006 // system will make valid IR.
4007 CodeGenFunction::ConditionalEvaluation condEval(CGF);
4008
4009 // Zero-initialize it if we're not doing a copy-initialization.
4010 bool shouldCopy = CRE->shouldCopy();
4011 if (!shouldCopy) {
4012 llvm::Value *null =
4013 llvm::ConstantPointerNull::get(cast<llvm::PointerType>(destElemType));
4014 CGF.Builder.CreateStore(null, temp);
4015 }
4016
4017 llvm::BasicBlock *contBB = nullptr;
4018 llvm::BasicBlock *originBB = nullptr;
4019
4020 // If the address is *not* known to be non-null, we need to switch.
4021 llvm::Value *finalArgument;
4022
4023 bool provablyNonNull = llvm::isKnownNonZero(srcAddr.getPointer(),
4024 CGF.CGM.getDataLayout());
4025 if (provablyNonNull) {
4026 finalArgument = temp.getPointer();
4027 } else {
4028 llvm::Value *isNull =
4029 CGF.Builder.CreateIsNull(srcAddr.getPointer(), "icr.isnull");
4030
4031 finalArgument = CGF.Builder.CreateSelect(isNull,
4032 llvm::ConstantPointerNull::get(destType),
4033 temp.getPointer(), "icr.argument");
4034
4035 // If we need to copy, then the load has to be conditional, which
4036 // means we need control flow.
4037 if (shouldCopy) {
4038 originBB = CGF.Builder.GetInsertBlock();
4039 contBB = CGF.createBasicBlock("icr.cont");
4040 llvm::BasicBlock *copyBB = CGF.createBasicBlock("icr.copy");
4041 CGF.Builder.CreateCondBr(isNull, contBB, copyBB);
4042 CGF.EmitBlock(copyBB);
4043 condEval.begin(CGF);
4044 }
4045 }
4046
4047 llvm::Value *valueToUse = nullptr;
4048
4049 // Perform a copy if necessary.
4050 if (shouldCopy) {
4051 RValue srcRV = CGF.EmitLoadOfLValue(srcLV, SourceLocation());
4052 assert(srcRV.isScalar());
4053
4054 llvm::Value *src = srcRV.getScalarVal();
4055 src = CGF.Builder.CreateBitCast(src, destElemType, "icr.cast");
4056
4057 // Use an ordinary store, not a store-to-lvalue.
4058 CGF.Builder.CreateStore(src, temp);
4059
4060 // If optimization is enabled, and the value was held in a
4061 // __strong variable, we need to tell the optimizer that this
4062 // value has to stay alive until we're doing the store back.
4063 // This is because the temporary is effectively unretained,
4064 // and so otherwise we can violate the high-level semantics.
4065 if (CGF.CGM.getCodeGenOpts().OptimizationLevel != 0 &&
4066 srcLV.getObjCLifetime() == Qualifiers::OCL_Strong) {
4067 valueToUse = src;
4068 }
4069 }
4070
4071 // Finish the control flow if we needed it.
4072 if (shouldCopy && !provablyNonNull) {
4073 llvm::BasicBlock *copyBB = CGF.Builder.GetInsertBlock();
4074 CGF.EmitBlock(contBB);
4075
4076 // Make a phi for the value to intrinsically use.
4077 if (valueToUse) {
4078 llvm::PHINode *phiToUse = CGF.Builder.CreatePHI(valueToUse->getType(), 2,
4079 "icr.to-use");
4080 phiToUse->addIncoming(valueToUse, copyBB);
4081 phiToUse->addIncoming(llvm::UndefValue::get(valueToUse->getType()),
4082 originBB);
4083 valueToUse = phiToUse;
4084 }
4085
4086 condEval.end(CGF);
4087 }
4088
4089 args.addWriteback(srcLV, temp, valueToUse);
4090 args.add(RValue::get(finalArgument), CRE->getType());
4091 }
4092
4093 void CallArgList::allocateArgumentMemory(CodeGenFunction &CGF) {
4094 assert(!StackBase);
4095
4096 // Save the stack.
4097 llvm::Function *F = CGF.CGM.getIntrinsic(llvm::Intrinsic::stacksave);
4098 StackBase = CGF.Builder.CreateCall(F, {}, "inalloca.save");
4099 }
4100
4101 void CallArgList::freeArgumentMemory(CodeGenFunction &CGF) const {
4102 if (StackBase) {
4103 // Restore the stack after the call.
4104 llvm::Function *F = CGF.CGM.getIntrinsic(llvm::Intrinsic::stackrestore);
4105 CGF.Builder.CreateCall(F, StackBase);
4106 }
4107 }
4108
4109 void CodeGenFunction::EmitNonNullArgCheck(RValue RV, QualType ArgType,
4110 SourceLocation ArgLoc,
4111 AbstractCallee AC,
4112 unsigned ParmNum) {
4113 if (!AC.getDecl() || !(SanOpts.has(SanitizerKind::NonnullAttribute) ||
4114 SanOpts.has(SanitizerKind::NullabilityArg)))
4115 return;
4116
4117 // The param decl may be missing in a variadic function.
4118 auto PVD = ParmNum < AC.getNumParams() ? AC.getParamDecl(ParmNum) : nullptr;
4119 unsigned ArgNo = PVD ? PVD->getFunctionScopeIndex() : ParmNum;
4120
4121 // Prefer the nonnull attribute if it's present.
4122 const NonNullAttr *NNAttr = nullptr;
4123 if (SanOpts.has(SanitizerKind::NonnullAttribute))
4124 NNAttr = getNonNullAttr(AC.getDecl(), PVD, ArgType, ArgNo);
4125
4126 bool CanCheckNullability = false;
4127 if (SanOpts.has(SanitizerKind::NullabilityArg) && !NNAttr && PVD) {
4128 auto Nullability = PVD->getType()->getNullability();
4129 CanCheckNullability = Nullability &&
4130 *Nullability == NullabilityKind::NonNull &&
4131 PVD->getTypeSourceInfo();
4132 }
4133
4134 if (!NNAttr && !CanCheckNullability)
4135 return;
4136
4137 SourceLocation AttrLoc;
4138 SanitizerMask CheckKind;
4139 SanitizerHandler Handler;
4140 if (NNAttr) {
4141 AttrLoc = NNAttr->getLocation();
4142 CheckKind = SanitizerKind::NonnullAttribute;
4143 Handler = SanitizerHandler::NonnullArg;
4144 } else {
4145 AttrLoc = PVD->getTypeSourceInfo()->getTypeLoc().findNullabilityLoc();
4146 CheckKind = SanitizerKind::NullabilityArg;
4147 Handler = SanitizerHandler::NullabilityArg;
4148 }
4149
4150 SanitizerScope SanScope(this);
4151 llvm::Value *Cond = EmitNonNullRValueCheck(RV, ArgType);
4152 llvm::Constant *StaticData[] = {
4153 EmitCheckSourceLocation(ArgLoc), EmitCheckSourceLocation(AttrLoc),
4154 llvm::ConstantInt::get(Int32Ty, ArgNo + 1),
4155 };
4156 EmitCheck(std::make_pair(Cond, CheckKind), Handler, StaticData, std::nullopt);
4157 }
4158
4159 // Check if the call is going to use the inalloca convention. This needs to
4160 // agree with CGFunctionInfo::usesInAlloca. The CGFunctionInfo is arranged
4161 // later, so we can't check it directly.
4162 static bool hasInAllocaArgs(CodeGenModule &CGM, CallingConv ExplicitCC,
4163 ArrayRef<QualType> ArgTypes) {
4164 // The Swift calling conventions don't go through the target-specific
4165 // argument classification, they never use inalloca.
4166 // TODO: Consider limiting inalloca use to only calling conventions supported
4167 // by MSVC.
4168 if (ExplicitCC == CC_Swift || ExplicitCC == CC_SwiftAsync)
4169 return false;
4170 if (!CGM.getTarget().getCXXABI().isMicrosoft())
4171 return false;
4172 return llvm::any_of(ArgTypes, [&](QualType Ty) {
4173 return isInAllocaArgument(CGM.getCXXABI(), Ty);
4174 });
4175 }
4176
4177 #ifndef NDEBUG
4178 // Determine whether the given argument is an Objective-C method
4179 // that may have type parameters in its signature.
4180 static bool isObjCMethodWithTypeParams(const ObjCMethodDecl *method) {
4181 const DeclContext *dc = method->getDeclContext();
4182 if (const ObjCInterfaceDecl *classDecl = dyn_cast<ObjCInterfaceDecl>(dc)) {
4183 return classDecl->getTypeParamListAsWritten();
4184 }
4185
4186 if (const ObjCCategoryDecl *catDecl = dyn_cast<ObjCCategoryDecl>(dc)) {
4187 return catDecl->getTypeParamList();
4188 }
4189
4190 return false;
4191 }
4192 #endif
4193
4194 /// EmitCallArgs - Emit call arguments for a function.
4195 void CodeGenFunction::EmitCallArgs(
4196 CallArgList &Args, PrototypeWrapper Prototype,
4197 llvm::iterator_range<CallExpr::const_arg_iterator> ArgRange,
4198 AbstractCallee AC, unsigned ParamsToSkip, EvaluationOrder Order) {
4199 SmallVector<QualType, 16> ArgTypes;
4200
4201 assert((ParamsToSkip == 0 || Prototype.P) &&
4202 "Can't skip parameters if type info is not provided");
4203
4204 // This variable only captures *explicitly* written conventions, not those
4205 // applied by default via command line flags or target defaults, such as
4206 // thiscall, aapcs, stdcall via -mrtd, etc. Computing that correctly would
4207 // require knowing if this is a C++ instance method or being able to see
4208 // unprototyped FunctionTypes.
4209 CallingConv ExplicitCC = CC_C;
4210
4211 // First, if a prototype was provided, use those argument types.
4212 bool IsVariadic = false;
4213 if (Prototype.P) {
4214 const auto *MD = Prototype.P.dyn_cast<const ObjCMethodDecl *>();
4215 if (MD) {
4216 IsVariadic = MD->isVariadic();
4217 ExplicitCC = getCallingConventionForDecl(
4218 MD, CGM.getTarget().getTriple().isOSWindows());
4219 ArgTypes.assign(MD->param_type_begin() + ParamsToSkip,
4220 MD->param_type_end());
4221 } else {
4222 const auto *FPT = Prototype.P.get<const FunctionProtoType *>();
4223 IsVariadic = FPT->isVariadic();
4224 ExplicitCC = FPT->getExtInfo().getCC();
4225 ArgTypes.assign(FPT->param_type_begin() + ParamsToSkip,
4226 FPT->param_type_end());
4227 }
4228
4229 #ifndef NDEBUG
4230 // Check that the prototyped types match the argument expression types.
4231 bool isGenericMethod = MD && isObjCMethodWithTypeParams(MD);
4232 CallExpr::const_arg_iterator Arg = ArgRange.begin();
4233 for (QualType Ty : ArgTypes) {
4234 assert(Arg != ArgRange.end() && "Running over edge of argument list!");
4235 assert(
4236 (isGenericMethod || Ty->isVariablyModifiedType() ||
4237 Ty.getNonReferenceType()->isObjCRetainableType() ||
4238 getContext()
4239 .getCanonicalType(Ty.getNonReferenceType())
4240 .getTypePtr() ==
4241 getContext().getCanonicalType((*Arg)->getType()).getTypePtr()) &&
4242 "type mismatch in call argument!");
4243 ++Arg;
4244 }
4245
4246 // Either we've emitted all the call args, or we have a call to variadic
4247 // function.
4248 assert((Arg == ArgRange.end() || IsVariadic) &&
4249 "Extra arguments in non-variadic function!");
4250 #endif
4251 }
4252
4253 // If we still have any arguments, emit them using the type of the argument.
4254 for (auto *A : llvm::drop_begin(ArgRange, ArgTypes.size()))
4255 ArgTypes.push_back(IsVariadic ? getVarArgType(A) : A->getType());
4256 assert((int)ArgTypes.size() == (ArgRange.end() - ArgRange.begin()));
4257
4258 // We must evaluate arguments from right to left in the MS C++ ABI,
4259 // because arguments are destroyed left to right in the callee. As a special
4260 // case, there are certain language constructs that require left-to-right
4261 // evaluation, and in those cases we consider the evaluation order requirement
4262 // to trump the "destruction order is reverse construction order" guarantee.
4263 bool LeftToRight =
4264 CGM.getTarget().getCXXABI().areArgsDestroyedLeftToRightInCallee()
4265 ? Order == EvaluationOrder::ForceLeftToRight
4266 : Order != EvaluationOrder::ForceRightToLeft;
4267
4268 auto MaybeEmitImplicitObjectSize = [&](unsigned I, const Expr *Arg,
4269 RValue EmittedArg) {
4270 if (!AC.hasFunctionDecl() || I >= AC.getNumParams())
4271 return;
4272 auto *PS = AC.getParamDecl(I)->getAttr<PassObjectSizeAttr>();
4273 if (PS == nullptr)
4274 return;
4275
4276 const auto &Context = getContext();
4277 auto SizeTy = Context.getSizeType();
4278 auto T = Builder.getIntNTy(Context.getTypeSize(SizeTy));
4279 assert(EmittedArg.getScalarVal() && "We emitted nothing for the arg?");
4280 llvm::Value *V = evaluateOrEmitBuiltinObjectSize(Arg, PS->getType(), T,
4281 EmittedArg.getScalarVal(),
4282 PS->isDynamic());
4283 Args.add(RValue::get(V), SizeTy);
4284 // If we're emitting args in reverse, be sure to do so with
4285 // pass_object_size, as well.
4286 if (!LeftToRight)
4287 std::swap(Args.back(), *(&Args.back() - 1));
4288 };
4289
4290 // Insert a stack save if we're going to need any inalloca args.
4291 if (hasInAllocaArgs(CGM, ExplicitCC, ArgTypes)) {
4292 assert(getTarget().getTriple().getArch() == llvm::Triple::x86 &&
4293 "inalloca only supported on x86");
4294 Args.allocateArgumentMemory(*this);
4295 }
4296
4297 // Evaluate each argument in the appropriate order.
4298 size_t CallArgsStart = Args.size();
4299 for (unsigned I = 0, E = ArgTypes.size(); I != E; ++I) {
4300 unsigned Idx = LeftToRight ? I : E - I - 1;
4301 CallExpr::const_arg_iterator Arg = ArgRange.begin() + Idx;
4302 unsigned InitialArgSize = Args.size();
4303 // If *Arg is an ObjCIndirectCopyRestoreExpr, check that either the types of
4304 // the argument and parameter match or the objc method is parameterized.
4305 assert((!isa<ObjCIndirectCopyRestoreExpr>(*Arg) ||
4306 getContext().hasSameUnqualifiedType((*Arg)->getType(),
4307 ArgTypes[Idx]) ||
4308 (isa<ObjCMethodDecl>(AC.getDecl()) &&
4309 isObjCMethodWithTypeParams(cast<ObjCMethodDecl>(AC.getDecl())))) &&
4310 "Argument and parameter types don't match");
4311 EmitCallArg(Args, *Arg, ArgTypes[Idx]);
4312 // In particular, we depend on it being the last arg in Args, and the
4313 // objectsize bits depend on there only being one arg if !LeftToRight.
4314 assert(InitialArgSize + 1 == Args.size() &&
4315 "The code below depends on only adding one arg per EmitCallArg");
4316 (void)InitialArgSize;
4317 // Since pointer argument are never emitted as LValue, it is safe to emit
4318 // non-null argument check for r-value only.
4319 if (!Args.back().hasLValue()) {
4320 RValue RVArg = Args.back().getKnownRValue();
4321 EmitNonNullArgCheck(RVArg, ArgTypes[Idx], (*Arg)->getExprLoc(), AC,
4322 ParamsToSkip + Idx);
4323 // @llvm.objectsize should never have side-effects and shouldn't need
4324 // destruction/cleanups, so we can safely "emit" it after its arg,
4325 // regardless of right-to-leftness
4326 MaybeEmitImplicitObjectSize(Idx, *Arg, RVArg);
4327 }
4328 }
4329
4330 if (!LeftToRight) {
4331 // Un-reverse the arguments we just evaluated so they match up with the LLVM
4332 // IR function.
4333 std::reverse(Args.begin() + CallArgsStart, Args.end());
4334 }
4335 }
4336
4337 namespace {
4338
4339 struct DestroyUnpassedArg final : EHScopeStack::Cleanup {
4340 DestroyUnpassedArg(Address Addr, QualType Ty)
4341 : Addr(Addr), Ty(Ty) {}
4342
4343 Address Addr;
4344 QualType Ty;
4345
4346 void Emit(CodeGenFunction &CGF, Flags flags) override {
4347 QualType::DestructionKind DtorKind = Ty.isDestructedType();
4348 if (DtorKind == QualType::DK_cxx_destructor) {
4349 const CXXDestructorDecl *Dtor = Ty->getAsCXXRecordDecl()->getDestructor();
4350 assert(!Dtor->isTrivial());
4351 CGF.EmitCXXDestructorCall(Dtor, Dtor_Complete, /*for vbase*/ false,
4352 /*Delegating=*/false, Addr, Ty);
4353 } else {
4354 CGF.callCStructDestructor(CGF.MakeAddrLValue(Addr, Ty));
4355 }
4356 }
4357 };
4358
4359 struct DisableDebugLocationUpdates {
4360 CodeGenFunction &CGF;
4361 bool disabledDebugInfo;
4362 DisableDebugLocationUpdates(CodeGenFunction &CGF, const Expr *E) : CGF(CGF) {
4363 if ((disabledDebugInfo = isa<CXXDefaultArgExpr>(E) && CGF.getDebugInfo()))
4364 CGF.disableDebugInfo();
4365 }
4366 ~DisableDebugLocationUpdates() {
4367 if (disabledDebugInfo)
4368 CGF.enableDebugInfo();
4369 }
4370 };
4371
4372 } // end anonymous namespace
4373
4374 RValue CallArg::getRValue(CodeGenFunction &CGF) const {
4375 if (!HasLV)
4376 return RV;
4377 LValue Copy = CGF.MakeAddrLValue(CGF.CreateMemTemp(Ty), Ty);
4378 CGF.EmitAggregateCopy(Copy, LV, Ty, AggValueSlot::DoesNotOverlap,
4379 LV.isVolatile());
4380 IsUsed = true;
4381 return RValue::getAggregate(Copy.getAddress(CGF));
4382 }
4383
4384 void CallArg::copyInto(CodeGenFunction &CGF, Address Addr) const {
4385 LValue Dst = CGF.MakeAddrLValue(Addr, Ty);
4386 if (!HasLV && RV.isScalar())
4387 CGF.EmitStoreOfScalar(RV.getScalarVal(), Dst, /*isInit=*/true);
4388 else if (!HasLV && RV.isComplex())
4389 CGF.EmitStoreOfComplex(RV.getComplexVal(), Dst, /*init=*/true);
4390 else {
4391 auto Addr = HasLV ? LV.getAddress(CGF) : RV.getAggregateAddress();
4392 LValue SrcLV = CGF.MakeAddrLValue(Addr, Ty);
4393 // We assume that call args are never copied into subobjects.
4394 CGF.EmitAggregateCopy(Dst, SrcLV, Ty, AggValueSlot::DoesNotOverlap,
4395 HasLV ? LV.isVolatileQualified()
4396 : RV.isVolatileQualified());
4397 }
4398 IsUsed = true;
4399 }
4400
4401 void CodeGenFunction::EmitCallArg(CallArgList &args, const Expr *E,
4402 QualType type) {
4403 DisableDebugLocationUpdates Dis(*this, E);
4404 if (const ObjCIndirectCopyRestoreExpr *CRE
4405 = dyn_cast<ObjCIndirectCopyRestoreExpr>(E)) {
4406 assert(getLangOpts().ObjCAutoRefCount);
4407 return emitWritebackArg(*this, args, CRE);
4408 }
4409
4410 assert(type->isReferenceType() == E->isGLValue() &&
4411 "reference binding to unmaterialized r-value!");
4412
4413 if (E->isGLValue()) {
4414 assert(E->getObjectKind() == OK_Ordinary);
4415 return args.add(EmitReferenceBindingToExpr(E), type);
4416 }
4417
4418 bool HasAggregateEvalKind = hasAggregateEvaluationKind(type);
4419
4420 // In the Microsoft C++ ABI, aggregate arguments are destructed by the callee.
4421 // However, we still have to push an EH-only cleanup in case we unwind before
4422 // we make it to the call.
4423 if (type->isRecordType() &&
4424 type->castAs<RecordType>()->getDecl()->isParamDestroyedInCallee()) {
4425 // If we're using inalloca, use the argument memory. Otherwise, use a
4426 // temporary.
4427 AggValueSlot Slot = args.isUsingInAlloca()
4428 ? createPlaceholderSlot(*this, type) : CreateAggTemp(type, "agg.tmp");
4429
4430 bool DestroyedInCallee = true, NeedsEHCleanup = true;
4431 if (const auto *RD = type->getAsCXXRecordDecl())
4432 DestroyedInCallee = RD->hasNonTrivialDestructor();
4433 else
4434 NeedsEHCleanup = needsEHCleanup(type.isDestructedType());
4435
4436 if (DestroyedInCallee)
4437 Slot.setExternallyDestructed();
4438
4439 EmitAggExpr(E, Slot);
4440 RValue RV = Slot.asRValue();
4441 args.add(RV, type);
4442
4443 if (DestroyedInCallee && NeedsEHCleanup) {
4444 // Create a no-op GEP between the placeholder and the cleanup so we can
4445 // RAUW it successfully. It also serves as a marker of the first
4446 // instruction where the cleanup is active.
4447 pushFullExprCleanup<DestroyUnpassedArg>(EHCleanup, Slot.getAddress(),
4448 type);
4449 // This unreachable is a temporary marker which will be removed later.
4450 llvm::Instruction *IsActive = Builder.CreateUnreachable();
4451 args.addArgCleanupDeactivation(EHStack.stable_begin(), IsActive);
4452 }
4453 return;
4454 }
4455
4456 if (HasAggregateEvalKind && isa<ImplicitCastExpr>(E) &&
4457 cast<CastExpr>(E)->getCastKind() == CK_LValueToRValue) {
4458 LValue L = EmitLValue(cast<CastExpr>(E)->getSubExpr());
4459 assert(L.isSimple());
4460 args.addUncopiedAggregate(L, type);
4461 return;
4462 }
4463
4464 args.add(EmitAnyExprToTemp(E), type);
4465 }
4466
4467 QualType CodeGenFunction::getVarArgType(const Expr *Arg) {
4468 // System headers on Windows define NULL to 0 instead of 0LL on Win64. MSVC
4469 // implicitly widens null pointer constants that are arguments to varargs
4470 // functions to pointer-sized ints.
4471 if (!getTarget().getTriple().isOSWindows())
4472 return Arg->getType();
4473
4474 if (Arg->getType()->isIntegerType() &&
4475 getContext().getTypeSize(Arg->getType()) <
4476 getContext().getTargetInfo().getPointerWidth(LangAS::Default) &&
4477 Arg->isNullPointerConstant(getContext(),
4478 Expr::NPC_ValueDependentIsNotNull)) {
4479 return getContext().getIntPtrType();
4480 }
4481
4482 return Arg->getType();
4483 }
4484
4485 // In ObjC ARC mode with no ObjC ARC exception safety, tell the ARC
4486 // optimizer it can aggressively ignore unwind edges.
4487 void
4488 CodeGenFunction::AddObjCARCExceptionMetadata(llvm::Instruction *Inst) {
4489 if (CGM.getCodeGenOpts().OptimizationLevel != 0 &&
4490 !CGM.getCodeGenOpts().ObjCAutoRefCountExceptions)
4491 Inst->setMetadata("clang.arc.no_objc_arc_exceptions",
4492 CGM.getNoObjCARCExceptionsMetadata());
4493 }
4494
4495 /// Emits a call to the given no-arguments nounwind runtime function.
4496 llvm::CallInst *
4497 CodeGenFunction::EmitNounwindRuntimeCall(llvm::FunctionCallee callee,
4498 const llvm::Twine &name) {
4499 return EmitNounwindRuntimeCall(callee, std::nullopt, name);
4500 }
4501
4502 /// Emits a call to the given nounwind runtime function.
4503 llvm::CallInst *
4504 CodeGenFunction::EmitNounwindRuntimeCall(llvm::FunctionCallee callee,
4505 ArrayRef<llvm::Value *> args,
4506 const llvm::Twine &name) {
4507 llvm::CallInst *call = EmitRuntimeCall(callee, args, name);
4508 call->setDoesNotThrow();
4509 return call;
4510 }
4511
4512 /// Emits a simple call (never an invoke) to the given no-arguments
4513 /// runtime function.
4514 llvm::CallInst *CodeGenFunction::EmitRuntimeCall(llvm::FunctionCallee callee,
4515 const llvm::Twine &name) {
4516 return EmitRuntimeCall(callee, std::nullopt, name);
4517 }
4518
4519 // Calls which may throw must have operand bundles indicating which funclet
4520 // they are nested within.
4521 SmallVector<llvm::OperandBundleDef, 1>
4522 CodeGenFunction::getBundlesForFunclet(llvm::Value *Callee) {
4523 // There is no need for a funclet operand bundle if we aren't inside a
4524 // funclet.
4525 if (!CurrentFuncletPad)
4526 return (SmallVector<llvm::OperandBundleDef, 1>());
4527
4528 // Skip intrinsics which cannot throw (as long as they don't lower into
4529 // regular function calls in the course of IR transformations).
4530 if (auto *CalleeFn = dyn_cast<llvm::Function>(Callee->stripPointerCasts())) {
4531 if (CalleeFn->isIntrinsic() && CalleeFn->doesNotThrow()) {
4532 auto IID = CalleeFn->getIntrinsicID();
4533 if (!llvm::IntrinsicInst::mayLowerToFunctionCall(IID))
4534 return (SmallVector<llvm::OperandBundleDef, 1>());
4535 }
4536 }
4537
4538 SmallVector<llvm::OperandBundleDef, 1> BundleList;
4539 BundleList.emplace_back("funclet", CurrentFuncletPad);
4540 return BundleList;
4541 }
4542
4543 /// Emits a simple call (never an invoke) to the given runtime function.
4544 llvm::CallInst *CodeGenFunction::EmitRuntimeCall(llvm::FunctionCallee callee,
4545 ArrayRef<llvm::Value *> args,
4546 const llvm::Twine &name) {
4547 llvm::CallInst *call = Builder.CreateCall(
4548 callee, args, getBundlesForFunclet(callee.getCallee()), name);
4549 call->setCallingConv(getRuntimeCC());
4550 return call;
4551 }
4552
4553 /// Emits a call or invoke to the given noreturn runtime function.
4554 void CodeGenFunction::EmitNoreturnRuntimeCallOrInvoke(
4555 llvm::FunctionCallee callee, ArrayRef<llvm::Value *> args) {
4556 SmallVector<llvm::OperandBundleDef, 1> BundleList =
4557 getBundlesForFunclet(callee.getCallee());
4558
4559 if (getInvokeDest()) {
4560 llvm::InvokeInst *invoke =
4561 Builder.CreateInvoke(callee,
4562 getUnreachableBlock(),
4563 getInvokeDest(),
4564 args,
4565 BundleList);
4566 invoke->setDoesNotReturn();
4567 invoke->setCallingConv(getRuntimeCC());
4568 } else {
4569 llvm::CallInst *call = Builder.CreateCall(callee, args, BundleList);
4570 call->setDoesNotReturn();
4571 call->setCallingConv(getRuntimeCC());
4572 Builder.CreateUnreachable();
4573 }
4574 }
4575
4576 /// Emits a call or invoke instruction to the given nullary runtime function.
4577 llvm::CallBase *
4578 CodeGenFunction::EmitRuntimeCallOrInvoke(llvm::FunctionCallee callee,
4579 const Twine &name) {
4580 return EmitRuntimeCallOrInvoke(callee, std::nullopt, name);
4581 }
4582
4583 /// Emits a call or invoke instruction to the given runtime function.
4584 llvm::CallBase *
4585 CodeGenFunction::EmitRuntimeCallOrInvoke(llvm::FunctionCallee callee,
4586 ArrayRef<llvm::Value *> args,
4587 const Twine &name) {
4588 llvm::CallBase *call = EmitCallOrInvoke(callee, args, name);
4589 call->setCallingConv(getRuntimeCC());
4590 return call;
4591 }
4592
4593 /// Emits a call or invoke instruction to the given function, depending
4594 /// on the current state of the EH stack.
4595 llvm::CallBase *CodeGenFunction::EmitCallOrInvoke(llvm::FunctionCallee Callee,
4596 ArrayRef<llvm::Value *> Args,
4597 const Twine &Name) {
4598 llvm::BasicBlock *InvokeDest = getInvokeDest();
4599 SmallVector<llvm::OperandBundleDef, 1> BundleList =
4600 getBundlesForFunclet(Callee.getCallee());
4601
4602 llvm::CallBase *Inst;
4603 if (!InvokeDest)
4604 Inst = Builder.CreateCall(Callee, Args, BundleList, Name);
4605 else {
4606 llvm::BasicBlock *ContBB = createBasicBlock("invoke.cont");
4607 Inst = Builder.CreateInvoke(Callee, ContBB, InvokeDest, Args, BundleList,
4608 Name);
4609 EmitBlock(ContBB);
4610 }
4611
4612 // In ObjC ARC mode with no ObjC ARC exception safety, tell the ARC
4613 // optimizer it can aggressively ignore unwind edges.
4614 if (CGM.getLangOpts().ObjCAutoRefCount)
4615 AddObjCARCExceptionMetadata(Inst);
4616
4617 return Inst;
4618 }
4619
4620 void CodeGenFunction::deferPlaceholderReplacement(llvm::Instruction *Old,
4621 llvm::Value *New) {
4622 DeferredReplacements.push_back(
4623 std::make_pair(llvm::WeakTrackingVH(Old), New));
4624 }
4625
4626 namespace {
4627
4628 /// Specify given \p NewAlign as the alignment of return value attribute. If
4629 /// such attribute already exists, re-set it to the maximal one of two options.
4630 [[nodiscard]] llvm::AttributeList
4631 maybeRaiseRetAlignmentAttribute(llvm::LLVMContext &Ctx,
4632 const llvm::AttributeList &Attrs,
4633 llvm::Align NewAlign) {
4634 llvm::Align CurAlign = Attrs.getRetAlignment().valueOrOne();
4635 if (CurAlign >= NewAlign)
4636 return Attrs;
4637 llvm::Attribute AlignAttr = llvm::Attribute::getWithAlignment(Ctx, NewAlign);
4638 return Attrs.removeRetAttribute(Ctx, llvm::Attribute::AttrKind::Alignment)
4639 .addRetAttribute(Ctx, AlignAttr);
4640 }
4641
4642 template <typename AlignedAttrTy> class AbstractAssumeAlignedAttrEmitter {
4643 protected:
4644 CodeGenFunction &CGF;
4645
4646 /// We do nothing if this is, or becomes, nullptr.
4647 const AlignedAttrTy *AA = nullptr;
4648
4649 llvm::Value *Alignment = nullptr; // May or may not be a constant.
4650 llvm::ConstantInt *OffsetCI = nullptr; // Constant, hopefully zero.
4651
4652 AbstractAssumeAlignedAttrEmitter(CodeGenFunction &CGF_, const Decl *FuncDecl)
4653 : CGF(CGF_) {
4654 if (!FuncDecl)
4655 return;
4656 AA = FuncDecl->getAttr<AlignedAttrTy>();
4657 }
4658
4659 public:
4660 /// If we can, materialize the alignment as an attribute on return value.
4661 [[nodiscard]] llvm::AttributeList
4662 TryEmitAsCallSiteAttribute(const llvm::AttributeList &Attrs) {
4663 if (!AA || OffsetCI || CGF.SanOpts.has(SanitizerKind::Alignment))
4664 return Attrs;
4665 const auto *AlignmentCI = dyn_cast<llvm::ConstantInt>(Alignment);
4666 if (!AlignmentCI)
4667 return Attrs;
4668 // We may legitimately have non-power-of-2 alignment here.
4669 // If so, this is UB land, emit it via `@llvm.assume` instead.
4670 if (!AlignmentCI->getValue().isPowerOf2())
4671 return Attrs;
4672 llvm::AttributeList NewAttrs = maybeRaiseRetAlignmentAttribute(
4673 CGF.getLLVMContext(), Attrs,
4674 llvm::Align(
4675 AlignmentCI->getLimitedValue(llvm::Value::MaximumAlignment)));
4676 AA = nullptr; // We're done. Disallow doing anything else.
4677 return NewAttrs;
4678 }
4679
4680 /// Emit alignment assumption.
4681 /// This is a general fallback that we take if either there is an offset,
4682 /// or the alignment is variable or we are sanitizing for alignment.
4683 void EmitAsAnAssumption(SourceLocation Loc, QualType RetTy, RValue &Ret) {
4684 if (!AA)
4685 return;
4686 CGF.emitAlignmentAssumption(Ret.getScalarVal(), RetTy, Loc,
4687 AA->getLocation(), Alignment, OffsetCI);
4688 AA = nullptr; // We're done. Disallow doing anything else.
4689 }
4690 };
4691
4692 /// Helper data structure to emit `AssumeAlignedAttr`.
4693 class AssumeAlignedAttrEmitter final
4694 : public AbstractAssumeAlignedAttrEmitter<AssumeAlignedAttr> {
4695 public:
4696 AssumeAlignedAttrEmitter(CodeGenFunction &CGF_, const Decl *FuncDecl)
4697 : AbstractAssumeAlignedAttrEmitter(CGF_, FuncDecl) {
4698 if (!AA)
4699 return;
4700 // It is guaranteed that the alignment/offset are constants.
4701 Alignment = cast<llvm::ConstantInt>(CGF.EmitScalarExpr(AA->getAlignment()));
4702 if (Expr *Offset = AA->getOffset()) {
4703 OffsetCI = cast<llvm::ConstantInt>(CGF.EmitScalarExpr(Offset));
4704 if (OffsetCI->isNullValue()) // Canonicalize zero offset to no offset.
4705 OffsetCI = nullptr;
4706 }
4707 }
4708 };
4709
4710 /// Helper data structure to emit `AllocAlignAttr`.
4711 class AllocAlignAttrEmitter final
4712 : public AbstractAssumeAlignedAttrEmitter<AllocAlignAttr> {
4713 public:
4714 AllocAlignAttrEmitter(CodeGenFunction &CGF_, const Decl *FuncDecl,
4715 const CallArgList &CallArgs)
4716 : AbstractAssumeAlignedAttrEmitter(CGF_, FuncDecl) {
4717 if (!AA)
4718 return;
4719 // Alignment may or may not be a constant, and that is okay.
4720 Alignment = CallArgs[AA->getParamIndex().getLLVMIndex()]
4721 .getRValue(CGF)
4722 .getScalarVal();
4723 }
4724 };
4725
4726 } // namespace
4727
4728 static unsigned getMaxVectorWidth(const llvm::Type *Ty) {
4729 if (auto *VT = dyn_cast<llvm::VectorType>(Ty))
4730 return VT->getPrimitiveSizeInBits().getKnownMinValue();
4731 if (auto *AT = dyn_cast<llvm::ArrayType>(Ty))
4732 return getMaxVectorWidth(AT->getElementType());
4733
4734 unsigned MaxVectorWidth = 0;
4735 if (auto *ST = dyn_cast<llvm::StructType>(Ty))
4736 for (auto *I : ST->elements())
4737 MaxVectorWidth = std::max(MaxVectorWidth, getMaxVectorWidth(I));
4738 return MaxVectorWidth;
4739 }
4740
4741 RValue CodeGenFunction::EmitCall(const CGFunctionInfo &CallInfo,
4742 const CGCallee &Callee,
4743 ReturnValueSlot ReturnValue,
4744 const CallArgList &CallArgs,
4745 llvm::CallBase **callOrInvoke, bool IsMustTail,
4746 SourceLocation Loc) {
4747 // FIXME: We no longer need the types from CallArgs; lift up and simplify.
4748
4749 assert(Callee.isOrdinary() || Callee.isVirtual());
4750
4751 // Handle struct-return functions by passing a pointer to the
4752 // location that we would like to return into.
4753 QualType RetTy = CallInfo.getReturnType();
4754 const ABIArgInfo &RetAI = CallInfo.getReturnInfo();
4755
4756 llvm::FunctionType *IRFuncTy = getTypes().GetFunctionType(CallInfo);
4757
4758 const Decl *TargetDecl = Callee.getAbstractInfo().getCalleeDecl().getDecl();
4759 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(TargetDecl)) {
4760 // We can only guarantee that a function is called from the correct
4761 // context/function based on the appropriate target attributes,
4762 // so only check in the case where we have both always_inline and target
4763 // since otherwise we could be making a conditional call after a check for
4764 // the proper cpu features (and it won't cause code generation issues due to
4765 // function based code generation).
4766 if (TargetDecl->hasAttr<AlwaysInlineAttr>() &&
4767 TargetDecl->hasAttr<TargetAttr>())
4768 checkTargetFeatures(Loc, FD);
4769
4770 // Some architectures (such as x86-64) have the ABI changed based on
4771 // attribute-target/features. Give them a chance to diagnose.
4772 CGM.getTargetCodeGenInfo().checkFunctionCallABI(
4773 CGM, Loc, dyn_cast_or_null<FunctionDecl>(CurCodeDecl), FD, CallArgs);
4774 }
4775
4776 #ifndef NDEBUG
4777 if (!(CallInfo.isVariadic() && CallInfo.getArgStruct())) {
4778 // For an inalloca varargs function, we don't expect CallInfo to match the
4779 // function pointer's type, because the inalloca struct a will have extra
4780 // fields in it for the varargs parameters. Code later in this function
4781 // bitcasts the function pointer to the type derived from CallInfo.
4782 //
4783 // In other cases, we assert that the types match up (until pointers stop
4784 // having pointee types).
4785 if (Callee.isVirtual())
4786 assert(IRFuncTy == Callee.getVirtualFunctionType());
4787 else {
4788 llvm::PointerType *PtrTy =
4789 llvm::cast<llvm::PointerType>(Callee.getFunctionPointer()->getType());
4790 assert(PtrTy->isOpaqueOrPointeeTypeMatches(IRFuncTy));
4791 }
4792 }
4793 #endif
4794
4795 // 1. Set up the arguments.
4796
4797 // If we're using inalloca, insert the allocation after the stack save.
4798 // FIXME: Do this earlier rather than hacking it in here!
4799 Address ArgMemory = Address::invalid();
4800 if (llvm::StructType *ArgStruct = CallInfo.getArgStruct()) {
4801 const llvm::DataLayout &DL = CGM.getDataLayout();
4802 llvm::Instruction *IP = CallArgs.getStackBase();
4803 llvm::AllocaInst *AI;
4804 if (IP) {
4805 IP = IP->getNextNode();
4806 AI = new llvm::AllocaInst(ArgStruct, DL.getAllocaAddrSpace(),
4807 "argmem", IP);
4808 } else {
4809 AI = CreateTempAlloca(ArgStruct, "argmem");
4810 }
4811 auto Align = CallInfo.getArgStructAlignment();
4812 AI->setAlignment(Align.getAsAlign());
4813 AI->setUsedWithInAlloca(true);
4814 assert(AI->isUsedWithInAlloca() && !AI->isStaticAlloca());
4815 ArgMemory = Address(AI, ArgStruct, Align);
4816 }
4817
4818 ClangToLLVMArgMapping IRFunctionArgs(CGM.getContext(), CallInfo);
4819 SmallVector<llvm::Value *, 16> IRCallArgs(IRFunctionArgs.totalIRArgs());
4820
4821 // If the call returns a temporary with struct return, create a temporary
4822 // alloca to hold the result, unless one is given to us.
4823 Address SRetPtr = Address::invalid();
4824 Address SRetAlloca = Address::invalid();
4825 llvm::Value *UnusedReturnSizePtr = nullptr;
4826 if (RetAI.isIndirect() || RetAI.isInAlloca() || RetAI.isCoerceAndExpand()) {
4827 if (!ReturnValue.isNull()) {
4828 SRetPtr = ReturnValue.getValue();
4829 } else {
4830 SRetPtr = CreateMemTemp(RetTy, "tmp", &SRetAlloca);
4831 if (HaveInsertPoint() && ReturnValue.isUnused()) {
4832 llvm::TypeSize size =
4833 CGM.getDataLayout().getTypeAllocSize(ConvertTypeForMem(RetTy));
4834 UnusedReturnSizePtr = EmitLifetimeStart(size, SRetAlloca.getPointer());
4835 }
4836 }
4837 if (IRFunctionArgs.hasSRetArg()) {
4838 IRCallArgs[IRFunctionArgs.getSRetArgNo()] = SRetPtr.getPointer();
4839 } else if (RetAI.isInAlloca()) {
4840 Address Addr =
4841 Builder.CreateStructGEP(ArgMemory, RetAI.getInAllocaFieldIndex());
4842 Builder.CreateStore(SRetPtr.getPointer(), Addr);
4843 }
4844 }
4845
4846 Address swiftErrorTemp = Address::invalid();
4847 Address swiftErrorArg = Address::invalid();
4848
4849 // When passing arguments using temporary allocas, we need to add the
4850 // appropriate lifetime markers. This vector keeps track of all the lifetime
4851 // markers that need to be ended right after the call.
4852 SmallVector<CallLifetimeEnd, 2> CallLifetimeEndAfterCall;
4853
4854 // Translate all of the arguments as necessary to match the IR lowering.
4855 assert(CallInfo.arg_size() == CallArgs.size() &&
4856 "Mismatch between function signature & arguments.");
4857 unsigned ArgNo = 0;
4858 CGFunctionInfo::const_arg_iterator info_it = CallInfo.arg_begin();
4859 for (CallArgList::const_iterator I = CallArgs.begin(), E = CallArgs.end();
4860 I != E; ++I, ++info_it, ++ArgNo) {
4861 const ABIArgInfo &ArgInfo = info_it->info;
4862
4863 // Insert a padding argument to ensure proper alignment.
4864 if (IRFunctionArgs.hasPaddingArg(ArgNo))
4865 IRCallArgs[IRFunctionArgs.getPaddingArgNo(ArgNo)] =
4866 llvm::UndefValue::get(ArgInfo.getPaddingType());
4867
4868 unsigned FirstIRArg, NumIRArgs;
4869 std::tie(FirstIRArg, NumIRArgs) = IRFunctionArgs.getIRArgs(ArgNo);
4870
4871 bool ArgHasMaybeUndefAttr =
4872 IsArgumentMaybeUndef(TargetDecl, CallInfo.getNumRequiredArgs(), ArgNo);
4873
4874 switch (ArgInfo.getKind()) {
4875 case ABIArgInfo::InAlloca: {
4876 assert(NumIRArgs == 0);
4877 assert(getTarget().getTriple().getArch() == llvm::Triple::x86);
4878 if (I->isAggregate()) {
4879 Address Addr = I->hasLValue()
4880 ? I->getKnownLValue().getAddress(*this)
4881 : I->getKnownRValue().getAggregateAddress();
4882 llvm::Instruction *Placeholder =
4883 cast<llvm::Instruction>(Addr.getPointer());
4884
4885 if (!ArgInfo.getInAllocaIndirect()) {
4886 // Replace the placeholder with the appropriate argument slot GEP.
4887 CGBuilderTy::InsertPoint IP = Builder.saveIP();
4888 Builder.SetInsertPoint(Placeholder);
4889 Addr = Builder.CreateStructGEP(ArgMemory,
4890 ArgInfo.getInAllocaFieldIndex());
4891 Builder.restoreIP(IP);
4892 } else {
4893 // For indirect things such as overaligned structs, replace the
4894 // placeholder with a regular aggregate temporary alloca. Store the
4895 // address of this alloca into the struct.
4896 Addr = CreateMemTemp(info_it->type, "inalloca.indirect.tmp");
4897 Address ArgSlot = Builder.CreateStructGEP(
4898 ArgMemory, ArgInfo.getInAllocaFieldIndex());
4899 Builder.CreateStore(Addr.getPointer(), ArgSlot);
4900 }
4901 deferPlaceholderReplacement(Placeholder, Addr.getPointer());
4902 } else if (ArgInfo.getInAllocaIndirect()) {
4903 // Make a temporary alloca and store the address of it into the argument
4904 // struct.
4905 Address Addr = CreateMemTempWithoutCast(
4906 I->Ty, getContext().getTypeAlignInChars(I->Ty),
4907 "indirect-arg-temp");
4908 I->copyInto(*this, Addr);
4909 Address ArgSlot =
4910 Builder.CreateStructGEP(ArgMemory, ArgInfo.getInAllocaFieldIndex());
4911 Builder.CreateStore(Addr.getPointer(), ArgSlot);
4912 } else {
4913 // Store the RValue into the argument struct.
4914 Address Addr =
4915 Builder.CreateStructGEP(ArgMemory, ArgInfo.getInAllocaFieldIndex());
4916 // There are some cases where a trivial bitcast is not avoidable. The
4917 // definition of a type later in a translation unit may change it's type
4918 // from {}* to (%struct.foo*)*.
4919 Addr = Builder.CreateElementBitCast(Addr, ConvertTypeForMem(I->Ty));
4920 I->copyInto(*this, Addr);
4921 }
4922 break;
4923 }
4924
4925 case ABIArgInfo::Indirect:
4926 case ABIArgInfo::IndirectAliased: {
4927 assert(NumIRArgs == 1);
4928 if (!I->isAggregate()) {
4929 // Make a temporary alloca to pass the argument.
4930 Address Addr = CreateMemTempWithoutCast(
4931 I->Ty, ArgInfo.getIndirectAlign(), "indirect-arg-temp");
4932
4933 llvm::Value *Val = Addr.getPointer();
4934 if (ArgHasMaybeUndefAttr)
4935 Val = Builder.CreateFreeze(Addr.getPointer());
4936 IRCallArgs[FirstIRArg] = Val;
4937
4938 I->copyInto(*this, Addr);
4939 } else {
4940 // We want to avoid creating an unnecessary temporary+copy here;
4941 // however, we need one in three cases:
4942 // 1. If the argument is not byval, and we are required to copy the
4943 // source. (This case doesn't occur on any common architecture.)
4944 // 2. If the argument is byval, RV is not sufficiently aligned, and
4945 // we cannot force it to be sufficiently aligned.
4946 // 3. If the argument is byval, but RV is not located in default
4947 // or alloca address space.
4948 Address Addr = I->hasLValue()
4949 ? I->getKnownLValue().getAddress(*this)
4950 : I->getKnownRValue().getAggregateAddress();
4951 llvm::Value *V = Addr.getPointer();
4952 CharUnits Align = ArgInfo.getIndirectAlign();
4953 const llvm::DataLayout *TD = &CGM.getDataLayout();
4954
4955 assert((FirstIRArg >= IRFuncTy->getNumParams() ||
4956 IRFuncTy->getParamType(FirstIRArg)->getPointerAddressSpace() ==
4957 TD->getAllocaAddrSpace()) &&
4958 "indirect argument must be in alloca address space");
4959
4960 bool NeedCopy = false;
4961
4962 if (Addr.getAlignment() < Align &&
4963 llvm::getOrEnforceKnownAlignment(V, Align.getAsAlign(), *TD) <
4964 Align.getAsAlign()) {
4965 NeedCopy = true;
4966 } else if (I->hasLValue()) {
4967 auto LV = I->getKnownLValue();
4968 auto AS = LV.getAddressSpace();
4969
4970 if (!ArgInfo.getIndirectByVal() ||
4971 (LV.getAlignment() < getContext().getTypeAlignInChars(I->Ty))) {
4972 NeedCopy = true;
4973 }
4974 if (!getLangOpts().OpenCL) {
4975 if ((ArgInfo.getIndirectByVal() &&
4976 (AS != LangAS::Default &&
4977 AS != CGM.getASTAllocaAddressSpace()))) {
4978 NeedCopy = true;
4979 }
4980 }
4981 // For OpenCL even if RV is located in default or alloca address space
4982 // we don't want to perform address space cast for it.
4983 else if ((ArgInfo.getIndirectByVal() &&
4984 Addr.getType()->getAddressSpace() != IRFuncTy->
4985 getParamType(FirstIRArg)->getPointerAddressSpace())) {
4986 NeedCopy = true;
4987 }
4988 }
4989
4990 if (NeedCopy) {
4991 // Create an aligned temporary, and copy to it.
4992 Address AI = CreateMemTempWithoutCast(
4993 I->Ty, ArgInfo.getIndirectAlign(), "byval-temp");
4994 llvm::Value *Val = AI.getPointer();
4995 if (ArgHasMaybeUndefAttr)
4996 Val = Builder.CreateFreeze(AI.getPointer());
4997 IRCallArgs[FirstIRArg] = Val;
4998
4999 // Emit lifetime markers for the temporary alloca.
5000 llvm::TypeSize ByvalTempElementSize =
5001 CGM.getDataLayout().getTypeAllocSize(AI.getElementType());
5002 llvm::Value *LifetimeSize =
5003 EmitLifetimeStart(ByvalTempElementSize, AI.getPointer());
5004
5005 // Add cleanup code to emit the end lifetime marker after the call.
5006 if (LifetimeSize) // In case we disabled lifetime markers.
5007 CallLifetimeEndAfterCall.emplace_back(AI, LifetimeSize);
5008
5009 // Generate the copy.
5010 I->copyInto(*this, AI);
5011 } else {
5012 // Skip the extra memcpy call.
5013 auto *T = llvm::PointerType::getWithSamePointeeType(
5014 cast<llvm::PointerType>(V->getType()),
5015 CGM.getDataLayout().getAllocaAddrSpace());
5016
5017 llvm::Value *Val = getTargetHooks().performAddrSpaceCast(
5018 *this, V, LangAS::Default, CGM.getASTAllocaAddressSpace(), T,
5019 true);
5020 if (ArgHasMaybeUndefAttr)
5021 Val = Builder.CreateFreeze(Val);
5022 IRCallArgs[FirstIRArg] = Val;
5023 }
5024 }
5025 break;
5026 }
5027
5028 case ABIArgInfo::Ignore:
5029 assert(NumIRArgs == 0);
5030 break;
5031
5032 case ABIArgInfo::Extend:
5033 case ABIArgInfo::Direct: {
5034 if (!isa<llvm::StructType>(ArgInfo.getCoerceToType()) &&
5035 ArgInfo.getCoerceToType() == ConvertType(info_it->type) &&
5036 ArgInfo.getDirectOffset() == 0) {
5037 assert(NumIRArgs == 1);
5038 llvm::Value *V;
5039 if (!I->isAggregate())
5040 V = I->getKnownRValue().getScalarVal();
5041 else
5042 V = Builder.CreateLoad(
5043 I->hasLValue() ? I->getKnownLValue().getAddress(*this)
5044 : I->getKnownRValue().getAggregateAddress());
5045
5046 // Implement swifterror by copying into a new swifterror argument.
5047 // We'll write back in the normal path out of the call.
5048 if (CallInfo.getExtParameterInfo(ArgNo).getABI()
5049 == ParameterABI::SwiftErrorResult) {
5050 assert(!swiftErrorTemp.isValid() && "multiple swifterror args");
5051
5052 QualType pointeeTy = I->Ty->getPointeeType();
5053 swiftErrorArg = Address(V, ConvertTypeForMem(pointeeTy),
5054 getContext().getTypeAlignInChars(pointeeTy));
5055
5056 swiftErrorTemp =
5057 CreateMemTemp(pointeeTy, getPointerAlign(), "swifterror.temp");
5058 V = swiftErrorTemp.getPointer();
5059 cast<llvm::AllocaInst>(V)->setSwiftError(true);
5060
5061 llvm::Value *errorValue = Builder.CreateLoad(swiftErrorArg);
5062 Builder.CreateStore(errorValue, swiftErrorTemp);
5063 }
5064
5065 // We might have to widen integers, but we should never truncate.
5066 if (ArgInfo.getCoerceToType() != V->getType() &&
5067 V->getType()->isIntegerTy())
5068 V = Builder.CreateZExt(V, ArgInfo.getCoerceToType());
5069
5070 // If the argument doesn't match, perform a bitcast to coerce it. This
5071 // can happen due to trivial type mismatches.
5072 if (FirstIRArg < IRFuncTy->getNumParams() &&
5073 V->getType() != IRFuncTy->getParamType(FirstIRArg))
5074 V = Builder.CreateBitCast(V, IRFuncTy->getParamType(FirstIRArg));
5075
5076 if (ArgHasMaybeUndefAttr)
5077 V = Builder.CreateFreeze(V);
5078 IRCallArgs[FirstIRArg] = V;
5079 break;
5080 }
5081
5082 // FIXME: Avoid the conversion through memory if possible.
5083 Address Src = Address::invalid();
5084 if (!I->isAggregate()) {
5085 Src = CreateMemTemp(I->Ty, "coerce");
5086 I->copyInto(*this, Src);
5087 } else {
5088 Src = I->hasLValue() ? I->getKnownLValue().getAddress(*this)
5089 : I->getKnownRValue().getAggregateAddress();
5090 }
5091
5092 // If the value is offset in memory, apply the offset now.
5093 Src = emitAddressAtOffset(*this, Src, ArgInfo);
5094
5095 // Fast-isel and the optimizer generally like scalar values better than
5096 // FCAs, so we flatten them if this is safe to do for this argument.
5097 llvm::StructType *STy =
5098 dyn_cast<llvm::StructType>(ArgInfo.getCoerceToType());
5099 if (STy && ArgInfo.isDirect() && ArgInfo.getCanBeFlattened()) {
5100 llvm::Type *SrcTy = Src.getElementType();
5101 uint64_t SrcSize = CGM.getDataLayout().getTypeAllocSize(SrcTy);
5102 uint64_t DstSize = CGM.getDataLayout().getTypeAllocSize(STy);
5103
5104 // If the source type is smaller than the destination type of the
5105 // coerce-to logic, copy the source value into a temp alloca the size
5106 // of the destination type to allow loading all of it. The bits past
5107 // the source value are left undef.
5108 if (SrcSize < DstSize) {
5109 Address TempAlloca
5110 = CreateTempAlloca(STy, Src.getAlignment(),
5111 Src.getName() + ".coerce");
5112 Builder.CreateMemCpy(TempAlloca, Src, SrcSize);
5113 Src = TempAlloca;
5114 } else {
5115 Src = Builder.CreateElementBitCast(Src, STy);
5116 }
5117
5118 assert(NumIRArgs == STy->getNumElements());
5119 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
5120 Address EltPtr = Builder.CreateStructGEP(Src, i);
5121 llvm::Value *LI = Builder.CreateLoad(EltPtr);
5122 if (ArgHasMaybeUndefAttr)
5123 LI = Builder.CreateFreeze(LI);
5124 IRCallArgs[FirstIRArg + i] = LI;
5125 }
5126 } else {
5127 // In the simple case, just pass the coerced loaded value.
5128 assert(NumIRArgs == 1);
5129 llvm::Value *Load =
5130 CreateCoercedLoad(Src, ArgInfo.getCoerceToType(), *this);
5131
5132 if (CallInfo.isCmseNSCall()) {
5133 // For certain parameter types, clear padding bits, as they may reveal
5134 // sensitive information.
5135 // Small struct/union types are passed as integer arrays.
5136 auto *ATy = dyn_cast<llvm::ArrayType>(Load->getType());
5137 if (ATy != nullptr && isa<RecordType>(I->Ty.getCanonicalType()))
5138 Load = EmitCMSEClearRecord(Load, ATy, I->Ty);
5139 }
5140
5141 if (ArgHasMaybeUndefAttr)
5142 Load = Builder.CreateFreeze(Load);
5143 IRCallArgs[FirstIRArg] = Load;
5144 }
5145
5146 break;
5147 }
5148
5149 case ABIArgInfo::CoerceAndExpand: {
5150 auto coercionType = ArgInfo.getCoerceAndExpandType();
5151 auto layout = CGM.getDataLayout().getStructLayout(coercionType);
5152
5153 llvm::Value *tempSize = nullptr;
5154 Address addr = Address::invalid();
5155 Address AllocaAddr = Address::invalid();
5156 if (I->isAggregate()) {
5157 addr = I->hasLValue() ? I->getKnownLValue().getAddress(*this)
5158 : I->getKnownRValue().getAggregateAddress();
5159
5160 } else {
5161 RValue RV = I->getKnownRValue();
5162 assert(RV.isScalar()); // complex should always just be direct
5163
5164 llvm::Type *scalarType = RV.getScalarVal()->getType();
5165 auto scalarSize = CGM.getDataLayout().getTypeAllocSize(scalarType);
5166 auto scalarAlign = CGM.getDataLayout().getPrefTypeAlign(scalarType);
5167
5168 // Materialize to a temporary.
5169 addr = CreateTempAlloca(
5170 RV.getScalarVal()->getType(),
5171 CharUnits::fromQuantity(std::max(layout->getAlignment(), scalarAlign)),
5172 "tmp",
5173 /*ArraySize=*/nullptr, &AllocaAddr);
5174 tempSize = EmitLifetimeStart(scalarSize, AllocaAddr.getPointer());
5175
5176 Builder.CreateStore(RV.getScalarVal(), addr);
5177 }
5178
5179 addr = Builder.CreateElementBitCast(addr, coercionType);
5180
5181 unsigned IRArgPos = FirstIRArg;
5182 for (unsigned i = 0, e = coercionType->getNumElements(); i != e; ++i) {
5183 llvm::Type *eltType = coercionType->getElementType(i);
5184 if (ABIArgInfo::isPaddingForCoerceAndExpand(eltType)) continue;
5185 Address eltAddr = Builder.CreateStructGEP(addr, i);
5186 llvm::Value *elt = Builder.CreateLoad(eltAddr);
5187 if (ArgHasMaybeUndefAttr)
5188 elt = Builder.CreateFreeze(elt);
5189 IRCallArgs[IRArgPos++] = elt;
5190 }
5191 assert(IRArgPos == FirstIRArg + NumIRArgs);
5192
5193 if (tempSize) {
5194 EmitLifetimeEnd(tempSize, AllocaAddr.getPointer());
5195 }
5196
5197 break;
5198 }
5199
5200 case ABIArgInfo::Expand: {
5201 unsigned IRArgPos = FirstIRArg;
5202 ExpandTypeToArgs(I->Ty, *I, IRFuncTy, IRCallArgs, IRArgPos);
5203 assert(IRArgPos == FirstIRArg + NumIRArgs);
5204 break;
5205 }
5206 }
5207 }
5208
5209 const CGCallee &ConcreteCallee = Callee.prepareConcreteCallee(*this);
5210 llvm::Value *CalleePtr = ConcreteCallee.getFunctionPointer();
5211
5212 // If we're using inalloca, set up that argument.
5213 if (ArgMemory.isValid()) {
5214 llvm::Value *Arg = ArgMemory.getPointer();
5215 if (CallInfo.isVariadic()) {
5216 // When passing non-POD arguments by value to variadic functions, we will
5217 // end up with a variadic prototype and an inalloca call site. In such
5218 // cases, we can't do any parameter mismatch checks. Give up and bitcast
5219 // the callee.
5220 unsigned CalleeAS = CalleePtr->getType()->getPointerAddressSpace();
5221 CalleePtr =
5222 Builder.CreateBitCast(CalleePtr, IRFuncTy->getPointerTo(CalleeAS));
5223 } else {
5224 llvm::Type *LastParamTy =
5225 IRFuncTy->getParamType(IRFuncTy->getNumParams() - 1);
5226 if (Arg->getType() != LastParamTy) {
5227 #ifndef NDEBUG
5228 // Assert that these structs have equivalent element types.
5229 llvm::StructType *FullTy = CallInfo.getArgStruct();
5230 if (!LastParamTy->isOpaquePointerTy()) {
5231 llvm::StructType *DeclaredTy = cast<llvm::StructType>(
5232 LastParamTy->getNonOpaquePointerElementType());
5233 assert(DeclaredTy->getNumElements() == FullTy->getNumElements());
5234 for (auto DI = DeclaredTy->element_begin(),
5235 DE = DeclaredTy->element_end(),
5236 FI = FullTy->element_begin();
5237 DI != DE; ++DI, ++FI)
5238 assert(*DI == *FI);
5239 }
5240 #endif
5241 Arg = Builder.CreateBitCast(Arg, LastParamTy);
5242 }
5243 }
5244 assert(IRFunctionArgs.hasInallocaArg());
5245 IRCallArgs[IRFunctionArgs.getInallocaArgNo()] = Arg;
5246 }
5247
5248 // 2. Prepare the function pointer.
5249
5250 // If the callee is a bitcast of a non-variadic function to have a
5251 // variadic function pointer type, check to see if we can remove the
5252 // bitcast. This comes up with unprototyped functions.
5253 //
5254 // This makes the IR nicer, but more importantly it ensures that we
5255 // can inline the function at -O0 if it is marked always_inline.
5256 auto simplifyVariadicCallee = [](llvm::FunctionType *CalleeFT,
5257 llvm::Value *Ptr) -> llvm::Function * {
5258 if (!CalleeFT->isVarArg())
5259 return nullptr;
5260
5261 // Get underlying value if it's a bitcast
5262 if (llvm::ConstantExpr *CE = dyn_cast<llvm::ConstantExpr>(Ptr)) {
5263 if (CE->getOpcode() == llvm::Instruction::BitCast)
5264 Ptr = CE->getOperand(0);
5265 }
5266
5267 llvm::Function *OrigFn = dyn_cast<llvm::Function>(Ptr);
5268 if (!OrigFn)
5269 return nullptr;
5270
5271 llvm::FunctionType *OrigFT = OrigFn->getFunctionType();
5272
5273 // If the original type is variadic, or if any of the component types
5274 // disagree, we cannot remove the cast.
5275 if (OrigFT->isVarArg() ||
5276 OrigFT->getNumParams() != CalleeFT->getNumParams() ||
5277 OrigFT->getReturnType() != CalleeFT->getReturnType())
5278 return nullptr;
5279
5280 for (unsigned i = 0, e = OrigFT->getNumParams(); i != e; ++i)
5281 if (OrigFT->getParamType(i) != CalleeFT->getParamType(i))
5282 return nullptr;
5283
5284 return OrigFn;
5285 };
5286
5287 if (llvm::Function *OrigFn = simplifyVariadicCallee(IRFuncTy, CalleePtr)) {
5288 CalleePtr = OrigFn;
5289 IRFuncTy = OrigFn->getFunctionType();
5290 }
5291
5292 // 3. Perform the actual call.
5293
5294 // Deactivate any cleanups that we're supposed to do immediately before
5295 // the call.
5296 if (!CallArgs.getCleanupsToDeactivate().empty())
5297 deactivateArgCleanupsBeforeCall(*this, CallArgs);
5298
5299 // Assert that the arguments we computed match up. The IR verifier
5300 // will catch this, but this is a common enough source of problems
5301 // during IRGen changes that it's way better for debugging to catch
5302 // it ourselves here.
5303 #ifndef NDEBUG
5304 assert(IRCallArgs.size() == IRFuncTy->getNumParams() || IRFuncTy->isVarArg());
5305 for (unsigned i = 0; i < IRCallArgs.size(); ++i) {
5306 // Inalloca argument can have different type.
5307 if (IRFunctionArgs.hasInallocaArg() &&
5308 i == IRFunctionArgs.getInallocaArgNo())
5309 continue;
5310 if (i < IRFuncTy->getNumParams())
5311 assert(IRCallArgs[i]->getType() == IRFuncTy->getParamType(i));
5312 }
5313 #endif
5314
5315 // Update the largest vector width if any arguments have vector types.
5316 for (unsigned i = 0; i < IRCallArgs.size(); ++i)
5317 LargestVectorWidth = std::max(LargestVectorWidth,
5318 getMaxVectorWidth(IRCallArgs[i]->getType()));
5319
5320 // Compute the calling convention and attributes.
5321 unsigned CallingConv;
5322 llvm::AttributeList Attrs;
5323 CGM.ConstructAttributeList(CalleePtr->getName(), CallInfo,
5324 Callee.getAbstractInfo(), Attrs, CallingConv,
5325 /*AttrOnCallSite=*/true,
5326 /*IsThunk=*/false);
5327
5328 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(CurFuncDecl))
5329 if (FD->hasAttr<StrictFPAttr>())
5330 // All calls within a strictfp function are marked strictfp
5331 Attrs = Attrs.addFnAttribute(getLLVMContext(), llvm::Attribute::StrictFP);
5332
5333 // Add call-site nomerge attribute if exists.
5334 if (InNoMergeAttributedStmt)
5335 Attrs = Attrs.addFnAttribute(getLLVMContext(), llvm::Attribute::NoMerge);
5336
5337 // Add call-site noinline attribute if exists.
5338 if (InNoInlineAttributedStmt)
5339 Attrs = Attrs.addFnAttribute(getLLVMContext(), llvm::Attribute::NoInline);
5340
5341 // Add call-site always_inline attribute if exists.
5342 if (InAlwaysInlineAttributedStmt)
5343 Attrs =
5344 Attrs.addFnAttribute(getLLVMContext(), llvm::Attribute::AlwaysInline);
5345
5346 // Apply some call-site-specific attributes.
5347 // TODO: work this into building the attribute set.
5348
5349 // Apply always_inline to all calls within flatten functions.
5350 // FIXME: should this really take priority over __try, below?
5351 if (CurCodeDecl && CurCodeDecl->hasAttr<FlattenAttr>() &&
5352 !InNoInlineAttributedStmt &&
5353 !(TargetDecl && TargetDecl->hasAttr<NoInlineAttr>())) {
5354 Attrs =
5355 Attrs.addFnAttribute(getLLVMContext(), llvm::Attribute::AlwaysInline);
5356 }
5357
5358 // Disable inlining inside SEH __try blocks.
5359 if (isSEHTryScope()) {
5360 Attrs = Attrs.addFnAttribute(getLLVMContext(), llvm::Attribute::NoInline);
5361 }
5362
5363 // Decide whether to use a call or an invoke.
5364 bool CannotThrow;
5365 if (currentFunctionUsesSEHTry()) {
5366 // SEH cares about asynchronous exceptions, so everything can "throw."
5367 CannotThrow = false;
5368 } else if (isCleanupPadScope() &&
5369 EHPersonality::get(*this).isMSVCXXPersonality()) {
5370 // The MSVC++ personality will implicitly terminate the program if an
5371 // exception is thrown during a cleanup outside of a try/catch.
5372 // We don't need to model anything in IR to get this behavior.
5373 CannotThrow = true;
5374 } else {
5375 // Otherwise, nounwind call sites will never throw.
5376 CannotThrow = Attrs.hasFnAttr(llvm::Attribute::NoUnwind);
5377
5378 if (auto *FPtr = dyn_cast<llvm::Function>(CalleePtr))
5379 if (FPtr->hasFnAttribute(llvm::Attribute::NoUnwind))
5380 CannotThrow = true;
5381 }
5382
5383 // If we made a temporary, be sure to clean up after ourselves. Note that we
5384 // can't depend on being inside of an ExprWithCleanups, so we need to manually
5385 // pop this cleanup later on. Being eager about this is OK, since this
5386 // temporary is 'invisible' outside of the callee.
5387 if (UnusedReturnSizePtr)
5388 pushFullExprCleanup<CallLifetimeEnd>(NormalEHLifetimeMarker, SRetAlloca,
5389 UnusedReturnSizePtr);
5390
5391 llvm::BasicBlock *InvokeDest = CannotThrow ? nullptr : getInvokeDest();
5392
5393 SmallVector<llvm::OperandBundleDef, 1> BundleList =
5394 getBundlesForFunclet(CalleePtr);
5395
5396 if (SanOpts.has(SanitizerKind::KCFI) &&
5397 !isa_and_nonnull<FunctionDecl>(TargetDecl))
5398 EmitKCFIOperandBundle(ConcreteCallee, BundleList);
5399
5400 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(CurFuncDecl))
5401 if (FD->hasAttr<StrictFPAttr>())
5402 // All calls within a strictfp function are marked strictfp
5403 Attrs = Attrs.addFnAttribute(getLLVMContext(), llvm::Attribute::StrictFP);
5404
5405 AssumeAlignedAttrEmitter AssumeAlignedAttrEmitter(*this, TargetDecl);
5406 Attrs = AssumeAlignedAttrEmitter.TryEmitAsCallSiteAttribute(Attrs);
5407
5408 AllocAlignAttrEmitter AllocAlignAttrEmitter(*this, TargetDecl, CallArgs);
5409 Attrs = AllocAlignAttrEmitter.TryEmitAsCallSiteAttribute(Attrs);
5410
5411 // Emit the actual call/invoke instruction.
5412 llvm::CallBase *CI;
5413 if (!InvokeDest) {
5414 CI = Builder.CreateCall(IRFuncTy, CalleePtr, IRCallArgs, BundleList);
5415 } else {
5416 llvm::BasicBlock *Cont = createBasicBlock("invoke.cont");
5417 CI = Builder.CreateInvoke(IRFuncTy, CalleePtr, Cont, InvokeDest, IRCallArgs,
5418 BundleList);
5419 EmitBlock(Cont);
5420 }
5421 if (callOrInvoke)
5422 *callOrInvoke = CI;
5423
5424 // If this is within a function that has the guard(nocf) attribute and is an
5425 // indirect call, add the "guard_nocf" attribute to this call to indicate that
5426 // Control Flow Guard checks should not be added, even if the call is inlined.
5427 if (const auto *FD = dyn_cast_or_null<FunctionDecl>(CurFuncDecl)) {
5428 if (const auto *A = FD->getAttr<CFGuardAttr>()) {
5429 if (A->getGuard() == CFGuardAttr::GuardArg::nocf && !CI->getCalledFunction())
5430 Attrs = Attrs.addFnAttribute(getLLVMContext(), "guard_nocf");
5431 }
5432 }
5433
5434 // Apply the attributes and calling convention.
5435 CI->setAttributes(Attrs);
5436 CI->setCallingConv(static_cast<llvm::CallingConv::ID>(CallingConv));
5437
5438 // Apply various metadata.
5439
5440 if (!CI->getType()->isVoidTy())
5441 CI->setName("call");
5442
5443 // Update largest vector width from the return type.
5444 LargestVectorWidth =
5445 std::max(LargestVectorWidth, getMaxVectorWidth(CI->getType()));
5446
5447 // Insert instrumentation or attach profile metadata at indirect call sites.
5448 // For more details, see the comment before the definition of
5449 // IPVK_IndirectCallTarget in InstrProfData.inc.
5450 if (!CI->getCalledFunction())
5451 PGO.valueProfile(Builder, llvm::IPVK_IndirectCallTarget,
5452 CI, CalleePtr);
5453
5454 // In ObjC ARC mode with no ObjC ARC exception safety, tell the ARC
5455 // optimizer it can aggressively ignore unwind edges.
5456 if (CGM.getLangOpts().ObjCAutoRefCount)
5457 AddObjCARCExceptionMetadata(CI);
5458
5459 // Set tail call kind if necessary.
5460 if (llvm::CallInst *Call = dyn_cast<llvm::CallInst>(CI)) {
5461 if (TargetDecl && TargetDecl->hasAttr<NotTailCalledAttr>())
5462 Call->setTailCallKind(llvm::CallInst::TCK_NoTail);
5463 else if (IsMustTail)
5464 Call->setTailCallKind(llvm::CallInst::TCK_MustTail);
5465 }
5466
5467 // Add metadata for calls to MSAllocator functions
5468 if (getDebugInfo() && TargetDecl &&
5469 TargetDecl->hasAttr<MSAllocatorAttr>())
5470 getDebugInfo()->addHeapAllocSiteMetadata(CI, RetTy->getPointeeType(), Loc);
5471
5472 // Add metadata if calling an __attribute__((error(""))) or warning fn.
5473 if (TargetDecl && TargetDecl->hasAttr<ErrorAttr>()) {
5474 llvm::ConstantInt *Line =
5475 llvm::ConstantInt::get(Int32Ty, Loc.getRawEncoding());
5476 llvm::ConstantAsMetadata *MD = llvm::ConstantAsMetadata::get(Line);
5477 llvm::MDTuple *MDT = llvm::MDNode::get(getLLVMContext(), {MD});
5478 CI->setMetadata("srcloc", MDT);
5479 }
5480
5481 // 4. Finish the call.
5482
5483 // If the call doesn't return, finish the basic block and clear the
5484 // insertion point; this allows the rest of IRGen to discard
5485 // unreachable code.
5486 if (CI->doesNotReturn()) {
5487 if (UnusedReturnSizePtr)
5488 PopCleanupBlock();
5489
5490 // Strip away the noreturn attribute to better diagnose unreachable UB.
5491 if (SanOpts.has(SanitizerKind::Unreachable)) {
5492 // Also remove from function since CallBase::hasFnAttr additionally checks
5493 // attributes of the called function.
5494 if (auto *F = CI->getCalledFunction())
5495 F->removeFnAttr(llvm::Attribute::NoReturn);
5496 CI->removeFnAttr(llvm::Attribute::NoReturn);
5497
5498 // Avoid incompatibility with ASan which relies on the `noreturn`
5499 // attribute to insert handler calls.
5500 if (SanOpts.hasOneOf(SanitizerKind::Address |
5501 SanitizerKind::KernelAddress)) {
5502 SanitizerScope SanScope(this);
5503 llvm::IRBuilder<>::InsertPointGuard IPGuard(Builder);
5504 Builder.SetInsertPoint(CI);
5505 auto *FnType = llvm::FunctionType::get(CGM.VoidTy, /*isVarArg=*/false);
5506 llvm::FunctionCallee Fn =
5507 CGM.CreateRuntimeFunction(FnType, "__asan_handle_no_return");
5508 EmitNounwindRuntimeCall(Fn);
5509 }
5510 }
5511
5512 EmitUnreachable(Loc);
5513 Builder.ClearInsertionPoint();
5514
5515 // FIXME: For now, emit a dummy basic block because expr emitters in
5516 // generally are not ready to handle emitting expressions at unreachable
5517 // points.
5518 EnsureInsertPoint();
5519
5520 // Return a reasonable RValue.
5521 return GetUndefRValue(RetTy);
5522 }
5523
5524 // If this is a musttail call, return immediately. We do not branch to the
5525 // epilogue in this case.
5526 if (IsMustTail) {
5527 for (auto it = EHStack.find(CurrentCleanupScopeDepth); it != EHStack.end();
5528 ++it) {
5529 EHCleanupScope *Cleanup = dyn_cast<EHCleanupScope>(&*it);
5530 if (!(Cleanup && Cleanup->getCleanup()->isRedundantBeforeReturn()))
5531 CGM.ErrorUnsupported(MustTailCall, "tail call skipping over cleanups");
5532 }
5533 if (CI->getType()->isVoidTy())
5534 Builder.CreateRetVoid();
5535 else
5536 Builder.CreateRet(CI);
5537 Builder.ClearInsertionPoint();
5538 EnsureInsertPoint();
5539 return GetUndefRValue(RetTy);
5540 }
5541
5542 // Perform the swifterror writeback.
5543 if (swiftErrorTemp.isValid()) {
5544 llvm::Value *errorResult = Builder.CreateLoad(swiftErrorTemp);
5545 Builder.CreateStore(errorResult, swiftErrorArg);
5546 }
5547
5548 // Emit any call-associated writebacks immediately. Arguably this
5549 // should happen after any return-value munging.
5550 if (CallArgs.hasWritebacks())
5551 emitWritebacks(*this, CallArgs);
5552
5553 // The stack cleanup for inalloca arguments has to run out of the normal
5554 // lexical order, so deactivate it and run it manually here.
5555 CallArgs.freeArgumentMemory(*this);
5556
5557 // Extract the return value.
5558 RValue Ret = [&] {
5559 switch (RetAI.getKind()) {
5560 case ABIArgInfo::CoerceAndExpand: {
5561 auto coercionType = RetAI.getCoerceAndExpandType();
5562
5563 Address addr = SRetPtr;
5564 addr = Builder.CreateElementBitCast(addr, coercionType);
5565
5566 assert(CI->getType() == RetAI.getUnpaddedCoerceAndExpandType());
5567 bool requiresExtract = isa<llvm::StructType>(CI->getType());
5568
5569 unsigned unpaddedIndex = 0;
5570 for (unsigned i = 0, e = coercionType->getNumElements(); i != e; ++i) {
5571 llvm::Type *eltType = coercionType->getElementType(i);
5572 if (ABIArgInfo::isPaddingForCoerceAndExpand(eltType)) continue;
5573 Address eltAddr = Builder.CreateStructGEP(addr, i);
5574 llvm::Value *elt = CI;
5575 if (requiresExtract)
5576 elt = Builder.CreateExtractValue(elt, unpaddedIndex++);
5577 else
5578 assert(unpaddedIndex == 0);
5579 Builder.CreateStore(elt, eltAddr);
5580 }
5581 // FALLTHROUGH
5582 [[fallthrough]];
5583 }
5584
5585 case ABIArgInfo::InAlloca:
5586 case ABIArgInfo::Indirect: {
5587 RValue ret = convertTempToRValue(SRetPtr, RetTy, SourceLocation());
5588 if (UnusedReturnSizePtr)
5589 PopCleanupBlock();
5590 return ret;
5591 }
5592
5593 case ABIArgInfo::Ignore:
5594 // If we are ignoring an argument that had a result, make sure to
5595 // construct the appropriate return value for our caller.
5596 return GetUndefRValue(RetTy);
5597
5598 case ABIArgInfo::Extend:
5599 case ABIArgInfo::Direct: {
5600 llvm::Type *RetIRTy = ConvertType(RetTy);
5601 if (RetAI.getCoerceToType() == RetIRTy && RetAI.getDirectOffset() == 0) {
5602 switch (getEvaluationKind(RetTy)) {
5603 case TEK_Complex: {
5604 llvm::Value *Real = Builder.CreateExtractValue(CI, 0);
5605 llvm::Value *Imag = Builder.CreateExtractValue(CI, 1);
5606 return RValue::getComplex(std::make_pair(Real, Imag));
5607 }
5608 case TEK_Aggregate: {
5609 Address DestPtr = ReturnValue.getValue();
5610 bool DestIsVolatile = ReturnValue.isVolatile();
5611
5612 if (!DestPtr.isValid()) {
5613 DestPtr = CreateMemTemp(RetTy, "agg.tmp");
5614 DestIsVolatile = false;
5615 }
5616 EmitAggregateStore(CI, DestPtr, DestIsVolatile);
5617 return RValue::getAggregate(DestPtr);
5618 }
5619 case TEK_Scalar: {
5620 // If the argument doesn't match, perform a bitcast to coerce it. This
5621 // can happen due to trivial type mismatches.
5622 llvm::Value *V = CI;
5623 if (V->getType() != RetIRTy)
5624 V = Builder.CreateBitCast(V, RetIRTy);
5625 return RValue::get(V);
5626 }
5627 }
5628 llvm_unreachable("bad evaluation kind");
5629 }
5630
5631 Address DestPtr = ReturnValue.getValue();
5632 bool DestIsVolatile = ReturnValue.isVolatile();
5633
5634 if (!DestPtr.isValid()) {
5635 DestPtr = CreateMemTemp(RetTy, "coerce");
5636 DestIsVolatile = false;
5637 }
5638
5639 // If the value is offset in memory, apply the offset now.
5640 Address StorePtr = emitAddressAtOffset(*this, DestPtr, RetAI);
5641 CreateCoercedStore(CI, StorePtr, DestIsVolatile, *this);
5642
5643 return convertTempToRValue(DestPtr, RetTy, SourceLocation());
5644 }
5645
5646 case ABIArgInfo::Expand:
5647 case ABIArgInfo::IndirectAliased:
5648 llvm_unreachable("Invalid ABI kind for return argument");
5649 }
5650
5651 llvm_unreachable("Unhandled ABIArgInfo::Kind");
5652 } ();
5653
5654 // Emit the assume_aligned check on the return value.
5655 if (Ret.isScalar() && TargetDecl) {
5656 AssumeAlignedAttrEmitter.EmitAsAnAssumption(Loc, RetTy, Ret);
5657 AllocAlignAttrEmitter.EmitAsAnAssumption(Loc, RetTy, Ret);
5658 }
5659
5660 // Explicitly call CallLifetimeEnd::Emit just to re-use the code even though
5661 // we can't use the full cleanup mechanism.
5662 for (CallLifetimeEnd &LifetimeEnd : CallLifetimeEndAfterCall)
5663 LifetimeEnd.Emit(*this, /*Flags=*/{});
5664
5665 if (!ReturnValue.isExternallyDestructed() &&
5666 RetTy.isDestructedType() == QualType::DK_nontrivial_c_struct)
5667 pushDestroy(QualType::DK_nontrivial_c_struct, Ret.getAggregateAddress(),
5668 RetTy);
5669
5670 return Ret;
5671 }
5672
5673 CGCallee CGCallee::prepareConcreteCallee(CodeGenFunction &CGF) const {
5674 if (isVirtual()) {
5675 const CallExpr *CE = getVirtualCallExpr();
5676 return CGF.CGM.getCXXABI().getVirtualFunctionPointer(
5677 CGF, getVirtualMethodDecl(), getThisAddress(), getVirtualFunctionType(),
5678 CE ? CE->getBeginLoc() : SourceLocation());
5679 }
5680
5681 return *this;
5682 }
5683
5684 /* VarArg handling */
5685
5686 Address CodeGenFunction::EmitVAArg(VAArgExpr *VE, Address &VAListAddr) {
5687 VAListAddr = VE->isMicrosoftABI()
5688 ? EmitMSVAListRef(VE->getSubExpr())
5689 : EmitVAListRef(VE->getSubExpr());
5690 QualType Ty = VE->getType();
5691 if (VE->isMicrosoftABI())
5692 return CGM.getTypes().getABIInfo().EmitMSVAArg(*this, VAListAddr, Ty);
5693 return CGM.getTypes().getABIInfo().EmitVAArg(*this, VAListAddr, Ty);
5694 }