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 "ABIInfoImpl.h" |
17 | #include "CGBlocks.h" |
18 | #include "CGCXXABI.h" |
19 | #include "CGCleanup.h" |
20 | #include "CGRecordLayout.h" |
21 | #include "CodeGenFunction.h" |
22 | #include "CodeGenModule.h" |
23 | #include "TargetInfo.h" |
24 | #include "clang/AST/Attr.h" |
25 | #include "clang/AST/Decl.h" |
26 | #include "clang/AST/DeclCXX.h" |
27 | #include "clang/AST/DeclObjC.h" |
28 | #include "clang/Basic/CodeGenOptions.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/AttributeMask.h" |
36 | #include "llvm/IR/Attributes.h" |
37 | #include "llvm/IR/CallingConv.h" |
38 | #include "llvm/IR/DataLayout.h" |
39 | #include "llvm/IR/InlineAsm.h" |
40 | #include "llvm/IR/IntrinsicInst.h" |
41 | #include "llvm/IR/Intrinsics.h" |
42 | #include "llvm/IR/Type.h" |
43 | #include "llvm/Transforms/Utils/Local.h" |
44 | #include <optional> |
45 | using namespace clang; |
46 | using namespace CodeGen; |
47 | |
48 | /***/ |
49 | |
50 | unsigned CodeGenTypes::ClangCallConvToLLVMCallConv(CallingConv CC) { |
51 | switch (CC) { |
52 | default: return llvm::CallingConv::C; |
53 | case CC_X86StdCall: return llvm::CallingConv::X86_StdCall; |
54 | case CC_X86FastCall: return llvm::CallingConv::X86_FastCall; |
55 | case CC_X86RegCall: return llvm::CallingConv::X86_RegCall; |
56 | case CC_X86ThisCall: return llvm::CallingConv::X86_ThisCall; |
57 | case CC_Win64: return llvm::CallingConv::Win64; |
58 | case CC_X86_64SysV: return llvm::CallingConv::X86_64_SysV; |
59 | case CC_AAPCS: return llvm::CallingConv::ARM_AAPCS; |
60 | case CC_AAPCS_VFP: return llvm::CallingConv::ARM_AAPCS_VFP; |
61 | case CC_IntelOclBicc: return llvm::CallingConv::Intel_OCL_BI; |
62 | // TODO: Add support for __pascal to LLVM. |
63 | case CC_X86Pascal: return llvm::CallingConv::C; |
64 | // TODO: Add support for __vectorcall to LLVM. |
65 | case CC_X86VectorCall: return llvm::CallingConv::X86_VectorCall; |
66 | case CC_AArch64VectorCall: return llvm::CallingConv::AArch64_VectorCall; |
67 | case CC_AArch64SVEPCS: return llvm::CallingConv::AArch64_SVE_VectorCall; |
68 | case CC_AMDGPUKernelCall: return llvm::CallingConv::AMDGPU_KERNEL; |
69 | case CC_SpirFunction: return llvm::CallingConv::SPIR_FUNC; |
70 | case CC_OpenCLKernel: return CGM.getTargetCodeGenInfo().getOpenCLKernelCallingConv(); |
71 | case CC_PreserveMost: return llvm::CallingConv::PreserveMost; |
72 | case CC_PreserveAll: return llvm::CallingConv::PreserveAll; |
73 | case CC_Swift: return llvm::CallingConv::Swift; |
74 | case CC_SwiftAsync: return llvm::CallingConv::SwiftTail; |
75 | case CC_M68kRTD: return llvm::CallingConv::M68k_RTD; |
76 | case CC_PreserveNone: return llvm::CallingConv::PreserveNone; |
77 | } |
78 | } |
79 | |
80 | /// Derives the 'this' type for codegen purposes, i.e. ignoring method CVR |
81 | /// qualification. Either or both of RD and MD may be null. A null RD indicates |
82 | /// that there is no meaningful 'this' type, and a null MD can occur when |
83 | /// calling a method pointer. |
84 | CanQualType CodeGenTypes::DeriveThisType(const CXXRecordDecl *RD, |
85 | const CXXMethodDecl *MD) { |
86 | QualType RecTy; |
87 | if (RD) |
88 | RecTy = Context.getTagDeclType(RD)->getCanonicalTypeInternal(); |
89 | else |
90 | RecTy = Context.VoidTy; |
91 | |
92 | if (MD) |
93 | RecTy = Context.getAddrSpaceQualType(T: RecTy, AddressSpace: MD->getMethodQualifiers().getAddressSpace()); |
94 | return Context.getPointerType(T: CanQualType::CreateUnsafe(Other: RecTy)); |
95 | } |
96 | |
97 | /// Returns the canonical formal type of the given C++ method. |
98 | static CanQual<FunctionProtoType> GetFormalType(const CXXMethodDecl *MD) { |
99 | return MD->getType()->getCanonicalTypeUnqualified() |
100 | .getAs<FunctionProtoType>(); |
101 | } |
102 | |
103 | /// Returns the "extra-canonicalized" return type, which discards |
104 | /// qualifiers on the return type. Codegen doesn't care about them, |
105 | /// and it makes ABI code a little easier to be able to assume that |
106 | /// all parameter and return types are top-level unqualified. |
107 | static CanQualType GetReturnType(QualType RetTy) { |
108 | return RetTy->getCanonicalTypeUnqualified().getUnqualifiedType(); |
109 | } |
110 | |
111 | /// Arrange the argument and result information for a value of the given |
112 | /// unprototyped freestanding function type. |
113 | const CGFunctionInfo & |
114 | CodeGenTypes::arrangeFreeFunctionType(CanQual<FunctionNoProtoType> FTNP) { |
115 | // When translating an unprototyped function type, always use a |
116 | // variadic type. |
117 | return arrangeLLVMFunctionInfo(FTNP->getReturnType().getUnqualifiedType(), |
118 | FnInfoOpts::None, std::nullopt, |
119 | FTNP->getExtInfo(), {}, RequiredArgs(0)); |
120 | } |
121 | |
122 | static void addExtParameterInfosForCall( |
123 | llvm::SmallVectorImpl<FunctionProtoType::ExtParameterInfo> ¶mInfos, |
124 | const FunctionProtoType *proto, |
125 | unsigned prefixArgs, |
126 | unsigned totalArgs) { |
127 | assert(proto->hasExtParameterInfos()); |
128 | assert(paramInfos.size() <= prefixArgs); |
129 | assert(proto->getNumParams() + prefixArgs <= totalArgs); |
130 | |
131 | paramInfos.reserve(N: totalArgs); |
132 | |
133 | // Add default infos for any prefix args that don't already have infos. |
134 | paramInfos.resize(N: prefixArgs); |
135 | |
136 | // Add infos for the prototype. |
137 | for (const auto &ParamInfo : proto->getExtParameterInfos()) { |
138 | paramInfos.push_back(Elt: ParamInfo); |
139 | // pass_object_size params have no parameter info. |
140 | if (ParamInfo.hasPassObjectSize()) |
141 | paramInfos.emplace_back(); |
142 | } |
143 | |
144 | assert(paramInfos.size() <= totalArgs && |
145 | "Did we forget to insert pass_object_size args?" ); |
146 | // Add default infos for the variadic and/or suffix arguments. |
147 | paramInfos.resize(N: totalArgs); |
148 | } |
149 | |
150 | /// Adds the formal parameters in FPT to the given prefix. If any parameter in |
151 | /// FPT has pass_object_size attrs, then we'll add parameters for those, too. |
152 | static void appendParameterTypes(const CodeGenTypes &CGT, |
153 | SmallVectorImpl<CanQualType> &prefix, |
154 | SmallVectorImpl<FunctionProtoType::ExtParameterInfo> ¶mInfos, |
155 | CanQual<FunctionProtoType> FPT) { |
156 | // Fast path: don't touch param info if we don't need to. |
157 | if (!FPT->hasExtParameterInfos()) { |
158 | assert(paramInfos.empty() && |
159 | "We have paramInfos, but the prototype doesn't?" ); |
160 | prefix.append(FPT->param_type_begin(), FPT->param_type_end()); |
161 | return; |
162 | } |
163 | |
164 | unsigned PrefixSize = prefix.size(); |
165 | // In the vast majority of cases, we'll have precisely FPT->getNumParams() |
166 | // parameters; the only thing that can change this is the presence of |
167 | // pass_object_size. So, we preallocate for the common case. |
168 | prefix.reserve(prefix.size() + FPT->getNumParams()); |
169 | |
170 | auto ExtInfos = FPT->getExtParameterInfos(); |
171 | assert(ExtInfos.size() == FPT->getNumParams()); |
172 | for (unsigned I = 0, E = FPT->getNumParams(); I != E; ++I) { |
173 | prefix.push_back(FPT->getParamType(I)); |
174 | if (ExtInfos[I].hasPassObjectSize()) |
175 | prefix.push_back(Elt: CGT.getContext().getSizeType()); |
176 | } |
177 | |
178 | addExtParameterInfosForCall(paramInfos, proto: FPT.getTypePtr(), prefixArgs: PrefixSize, |
179 | totalArgs: prefix.size()); |
180 | } |
181 | |
182 | /// Arrange the LLVM function layout for a value of the given function |
183 | /// type, on top of any implicit parameters already stored. |
184 | static const CGFunctionInfo & |
185 | arrangeLLVMFunctionInfo(CodeGenTypes &CGT, bool instanceMethod, |
186 | SmallVectorImpl<CanQualType> &prefix, |
187 | CanQual<FunctionProtoType> FTP) { |
188 | SmallVector<FunctionProtoType::ExtParameterInfo, 16> paramInfos; |
189 | RequiredArgs Required = RequiredArgs::forPrototypePlus(prototype: FTP, additional: prefix.size()); |
190 | // FIXME: Kill copy. |
191 | appendParameterTypes(CGT, prefix, paramInfos, FPT: FTP); |
192 | CanQualType resultType = FTP->getReturnType().getUnqualifiedType(); |
193 | |
194 | FnInfoOpts opts = |
195 | instanceMethod ? FnInfoOpts::IsInstanceMethod : FnInfoOpts::None; |
196 | return CGT.arrangeLLVMFunctionInfo(resultType, opts, prefix, |
197 | FTP->getExtInfo(), paramInfos, Required); |
198 | } |
199 | |
200 | /// Arrange the argument and result information for a value of the |
201 | /// given freestanding function type. |
202 | const CGFunctionInfo & |
203 | CodeGenTypes::arrangeFreeFunctionType(CanQual<FunctionProtoType> FTP) { |
204 | SmallVector<CanQualType, 16> argTypes; |
205 | return ::arrangeLLVMFunctionInfo(CGT&: *this, /*instanceMethod=*/false, prefix&: argTypes, |
206 | FTP); |
207 | } |
208 | |
209 | static CallingConv getCallingConventionForDecl(const ObjCMethodDecl *D, |
210 | bool IsWindows) { |
211 | // Set the appropriate calling convention for the Function. |
212 | if (D->hasAttr<StdCallAttr>()) |
213 | return CC_X86StdCall; |
214 | |
215 | if (D->hasAttr<FastCallAttr>()) |
216 | return CC_X86FastCall; |
217 | |
218 | if (D->hasAttr<RegCallAttr>()) |
219 | return CC_X86RegCall; |
220 | |
221 | if (D->hasAttr<ThisCallAttr>()) |
222 | return CC_X86ThisCall; |
223 | |
224 | if (D->hasAttr<VectorCallAttr>()) |
225 | return CC_X86VectorCall; |
226 | |
227 | if (D->hasAttr<PascalAttr>()) |
228 | return CC_X86Pascal; |
229 | |
230 | if (PcsAttr *PCS = D->getAttr<PcsAttr>()) |
231 | return (PCS->getPCS() == PcsAttr::AAPCS ? CC_AAPCS : CC_AAPCS_VFP); |
232 | |
233 | if (D->hasAttr<AArch64VectorPcsAttr>()) |
234 | return CC_AArch64VectorCall; |
235 | |
236 | if (D->hasAttr<AArch64SVEPcsAttr>()) |
237 | return CC_AArch64SVEPCS; |
238 | |
239 | if (D->hasAttr<AMDGPUKernelCallAttr>()) |
240 | return CC_AMDGPUKernelCall; |
241 | |
242 | if (D->hasAttr<IntelOclBiccAttr>()) |
243 | return CC_IntelOclBicc; |
244 | |
245 | if (D->hasAttr<MSABIAttr>()) |
246 | return IsWindows ? CC_C : CC_Win64; |
247 | |
248 | if (D->hasAttr<SysVABIAttr>()) |
249 | return IsWindows ? CC_X86_64SysV : CC_C; |
250 | |
251 | if (D->hasAttr<PreserveMostAttr>()) |
252 | return CC_PreserveMost; |
253 | |
254 | if (D->hasAttr<PreserveAllAttr>()) |
255 | return CC_PreserveAll; |
256 | |
257 | if (D->hasAttr<M68kRTDAttr>()) |
258 | return CC_M68kRTD; |
259 | |
260 | if (D->hasAttr<PreserveNoneAttr>()) |
261 | return CC_PreserveNone; |
262 | |
263 | return CC_C; |
264 | } |
265 | |
266 | /// Arrange the argument and result information for a call to an |
267 | /// unknown C++ non-static member function of the given abstract type. |
268 | /// (A null RD means we don't have any meaningful "this" argument type, |
269 | /// so fall back to a generic pointer type). |
270 | /// The member function must be an ordinary function, i.e. not a |
271 | /// constructor or destructor. |
272 | const CGFunctionInfo & |
273 | CodeGenTypes::arrangeCXXMethodType(const CXXRecordDecl *RD, |
274 | const FunctionProtoType *FTP, |
275 | const CXXMethodDecl *MD) { |
276 | SmallVector<CanQualType, 16> argTypes; |
277 | |
278 | // Add the 'this' pointer. |
279 | argTypes.push_back(Elt: DeriveThisType(RD, MD)); |
280 | |
281 | return ::arrangeLLVMFunctionInfo( |
282 | CGT&: *this, /*instanceMethod=*/true, prefix&: argTypes, |
283 | FTP: FTP->getCanonicalTypeUnqualified().getAs<FunctionProtoType>()); |
284 | } |
285 | |
286 | /// Set calling convention for CUDA/HIP kernel. |
287 | static void setCUDAKernelCallingConvention(CanQualType &FTy, CodeGenModule &CGM, |
288 | const FunctionDecl *FD) { |
289 | if (FD->hasAttr<CUDAGlobalAttr>()) { |
290 | const FunctionType *FT = FTy->getAs<FunctionType>(); |
291 | CGM.getTargetCodeGenInfo().setCUDAKernelCallingConvention(FT); |
292 | FTy = FT->getCanonicalTypeUnqualified(); |
293 | } |
294 | } |
295 | |
296 | /// Arrange the argument and result information for a declaration or |
297 | /// definition of the given C++ non-static member function. The |
298 | /// member function must be an ordinary function, i.e. not a |
299 | /// constructor or destructor. |
300 | const CGFunctionInfo & |
301 | CodeGenTypes::arrangeCXXMethodDeclaration(const CXXMethodDecl *MD) { |
302 | assert(!isa<CXXConstructorDecl>(MD) && "wrong method for constructors!" ); |
303 | assert(!isa<CXXDestructorDecl>(MD) && "wrong method for destructors!" ); |
304 | |
305 | CanQualType FT = GetFormalType(MD).getAs<Type>(); |
306 | setCUDAKernelCallingConvention(FT, CGM, MD); |
307 | auto prototype = FT.getAs<FunctionProtoType>(); |
308 | |
309 | if (MD->isImplicitObjectMemberFunction()) { |
310 | // The abstract case is perfectly fine. |
311 | const CXXRecordDecl *ThisType = TheCXXABI.getThisArgumentTypeForMethod(MD); |
312 | return arrangeCXXMethodType(RD: ThisType, FTP: prototype.getTypePtr(), MD); |
313 | } |
314 | |
315 | return arrangeFreeFunctionType(FTP: prototype); |
316 | } |
317 | |
318 | bool CodeGenTypes::inheritingCtorHasParams( |
319 | const InheritedConstructor &Inherited, CXXCtorType Type) { |
320 | // Parameters are unnecessary if we're constructing a base class subobject |
321 | // and the inherited constructor lives in a virtual base. |
322 | return Type == Ctor_Complete || |
323 | !Inherited.getShadowDecl()->constructsVirtualBase() || |
324 | !Target.getCXXABI().hasConstructorVariants(); |
325 | } |
326 | |
327 | const CGFunctionInfo & |
328 | CodeGenTypes::arrangeCXXStructorDeclaration(GlobalDecl GD) { |
329 | auto *MD = cast<CXXMethodDecl>(Val: GD.getDecl()); |
330 | |
331 | SmallVector<CanQualType, 16> argTypes; |
332 | SmallVector<FunctionProtoType::ExtParameterInfo, 16> paramInfos; |
333 | |
334 | const CXXRecordDecl *ThisType = TheCXXABI.getThisArgumentTypeForMethod(GD); |
335 | argTypes.push_back(Elt: DeriveThisType(RD: ThisType, MD)); |
336 | |
337 | bool PassParams = true; |
338 | |
339 | if (auto *CD = dyn_cast<CXXConstructorDecl>(Val: MD)) { |
340 | // A base class inheriting constructor doesn't get forwarded arguments |
341 | // needed to construct a virtual base (or base class thereof). |
342 | if (auto Inherited = CD->getInheritedConstructor()) |
343 | PassParams = inheritingCtorHasParams(Inherited, Type: GD.getCtorType()); |
344 | } |
345 | |
346 | CanQual<FunctionProtoType> FTP = GetFormalType(MD); |
347 | |
348 | // Add the formal parameters. |
349 | if (PassParams) |
350 | appendParameterTypes(CGT: *this, prefix&: argTypes, paramInfos, FPT: FTP); |
351 | |
352 | CGCXXABI::AddedStructorArgCounts AddedArgs = |
353 | TheCXXABI.buildStructorSignature(GD, ArgTys&: argTypes); |
354 | if (!paramInfos.empty()) { |
355 | // Note: prefix implies after the first param. |
356 | if (AddedArgs.Prefix) |
357 | paramInfos.insert(I: paramInfos.begin() + 1, NumToInsert: AddedArgs.Prefix, |
358 | Elt: FunctionProtoType::ExtParameterInfo{}); |
359 | if (AddedArgs.Suffix) |
360 | paramInfos.append(NumInputs: AddedArgs.Suffix, |
361 | Elt: FunctionProtoType::ExtParameterInfo{}); |
362 | } |
363 | |
364 | RequiredArgs required = |
365 | (PassParams && MD->isVariadic() ? RequiredArgs(argTypes.size()) |
366 | : RequiredArgs::All); |
367 | |
368 | FunctionType::ExtInfo extInfo = FTP->getExtInfo(); |
369 | CanQualType resultType = TheCXXABI.HasThisReturn(GD) |
370 | ? argTypes.front() |
371 | : TheCXXABI.hasMostDerivedReturn(GD) |
372 | ? CGM.getContext().VoidPtrTy |
373 | : Context.VoidTy; |
374 | return arrangeLLVMFunctionInfo(returnType: resultType, opts: FnInfoOpts::IsInstanceMethod, |
375 | argTypes, info: extInfo, paramInfos, args: required); |
376 | } |
377 | |
378 | static SmallVector<CanQualType, 16> |
379 | getArgTypesForCall(ASTContext &ctx, const CallArgList &args) { |
380 | SmallVector<CanQualType, 16> argTypes; |
381 | for (auto &arg : args) |
382 | argTypes.push_back(Elt: ctx.getCanonicalParamType(T: arg.Ty)); |
383 | return argTypes; |
384 | } |
385 | |
386 | static SmallVector<CanQualType, 16> |
387 | getArgTypesForDeclaration(ASTContext &ctx, const FunctionArgList &args) { |
388 | SmallVector<CanQualType, 16> argTypes; |
389 | for (auto &arg : args) |
390 | argTypes.push_back(Elt: ctx.getCanonicalParamType(T: arg->getType())); |
391 | return argTypes; |
392 | } |
393 | |
394 | static llvm::SmallVector<FunctionProtoType::ExtParameterInfo, 16> |
395 | getExtParameterInfosForCall(const FunctionProtoType *proto, |
396 | unsigned prefixArgs, unsigned totalArgs) { |
397 | llvm::SmallVector<FunctionProtoType::ExtParameterInfo, 16> result; |
398 | if (proto->hasExtParameterInfos()) { |
399 | addExtParameterInfosForCall(paramInfos&: result, proto, prefixArgs, totalArgs); |
400 | } |
401 | return result; |
402 | } |
403 | |
404 | /// Arrange a call to a C++ method, passing the given arguments. |
405 | /// |
406 | /// ExtraPrefixArgs is the number of ABI-specific args passed after the `this` |
407 | /// parameter. |
408 | /// ExtraSuffixArgs is the number of ABI-specific args passed at the end of |
409 | /// args. |
410 | /// PassProtoArgs indicates whether `args` has args for the parameters in the |
411 | /// given CXXConstructorDecl. |
412 | const CGFunctionInfo & |
413 | CodeGenTypes::arrangeCXXConstructorCall(const CallArgList &args, |
414 | const CXXConstructorDecl *D, |
415 | CXXCtorType CtorKind, |
416 | unsigned , |
417 | unsigned , |
418 | bool PassProtoArgs) { |
419 | // FIXME: Kill copy. |
420 | SmallVector<CanQualType, 16> ArgTypes; |
421 | for (const auto &Arg : args) |
422 | ArgTypes.push_back(Elt: Context.getCanonicalParamType(T: Arg.Ty)); |
423 | |
424 | // +1 for implicit this, which should always be args[0]. |
425 | unsigned TotalPrefixArgs = 1 + ExtraPrefixArgs; |
426 | |
427 | CanQual<FunctionProtoType> FPT = GetFormalType(D); |
428 | RequiredArgs Required = PassProtoArgs |
429 | ? RequiredArgs::forPrototypePlus( |
430 | prototype: FPT, additional: TotalPrefixArgs + ExtraSuffixArgs) |
431 | : RequiredArgs::All; |
432 | |
433 | GlobalDecl GD(D, CtorKind); |
434 | CanQualType ResultType = TheCXXABI.HasThisReturn(GD) |
435 | ? ArgTypes.front() |
436 | : TheCXXABI.hasMostDerivedReturn(GD) |
437 | ? CGM.getContext().VoidPtrTy |
438 | : Context.VoidTy; |
439 | |
440 | FunctionType::ExtInfo Info = FPT->getExtInfo(); |
441 | llvm::SmallVector<FunctionProtoType::ExtParameterInfo, 16> ParamInfos; |
442 | // If the prototype args are elided, we should only have ABI-specific args, |
443 | // which never have param info. |
444 | if (PassProtoArgs && FPT->hasExtParameterInfos()) { |
445 | // ABI-specific suffix arguments are treated the same as variadic arguments. |
446 | addExtParameterInfosForCall(paramInfos&: ParamInfos, proto: FPT.getTypePtr(), prefixArgs: TotalPrefixArgs, |
447 | totalArgs: ArgTypes.size()); |
448 | } |
449 | |
450 | return arrangeLLVMFunctionInfo(returnType: ResultType, opts: FnInfoOpts::IsInstanceMethod, |
451 | argTypes: ArgTypes, info: Info, paramInfos: ParamInfos, args: Required); |
452 | } |
453 | |
454 | /// Arrange the argument and result information for the declaration or |
455 | /// definition of the given function. |
456 | const CGFunctionInfo & |
457 | CodeGenTypes::arrangeFunctionDeclaration(const FunctionDecl *FD) { |
458 | if (const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(Val: FD)) |
459 | if (MD->isImplicitObjectMemberFunction()) |
460 | return arrangeCXXMethodDeclaration(MD); |
461 | |
462 | CanQualType FTy = FD->getType()->getCanonicalTypeUnqualified(); |
463 | |
464 | assert(isa<FunctionType>(FTy)); |
465 | setCUDAKernelCallingConvention(FTy, CGM, FD); |
466 | |
467 | // When declaring a function without a prototype, always use a |
468 | // non-variadic type. |
469 | if (CanQual<FunctionNoProtoType> noProto = FTy.getAs<FunctionNoProtoType>()) { |
470 | return arrangeLLVMFunctionInfo(noProto->getReturnType(), FnInfoOpts::None, |
471 | std::nullopt, noProto->getExtInfo(), {}, |
472 | RequiredArgs::All); |
473 | } |
474 | |
475 | return arrangeFreeFunctionType(FTP: FTy.castAs<FunctionProtoType>()); |
476 | } |
477 | |
478 | /// Arrange the argument and result information for the declaration or |
479 | /// definition of an Objective-C method. |
480 | const CGFunctionInfo & |
481 | CodeGenTypes::arrangeObjCMethodDeclaration(const ObjCMethodDecl *MD) { |
482 | // It happens that this is the same as a call with no optional |
483 | // arguments, except also using the formal 'self' type. |
484 | return arrangeObjCMessageSendSignature(MD, receiverType: MD->getSelfDecl()->getType()); |
485 | } |
486 | |
487 | /// Arrange the argument and result information for the function type |
488 | /// through which to perform a send to the given Objective-C method, |
489 | /// using the given receiver type. The receiver type is not always |
490 | /// the 'self' type of the method or even an Objective-C pointer type. |
491 | /// This is *not* the right method for actually performing such a |
492 | /// message send, due to the possibility of optional arguments. |
493 | const CGFunctionInfo & |
494 | CodeGenTypes::arrangeObjCMessageSendSignature(const ObjCMethodDecl *MD, |
495 | QualType receiverType) { |
496 | SmallVector<CanQualType, 16> argTys; |
497 | SmallVector<FunctionProtoType::ExtParameterInfo, 4> extParamInfos( |
498 | MD->isDirectMethod() ? 1 : 2); |
499 | argTys.push_back(Elt: Context.getCanonicalParamType(T: receiverType)); |
500 | if (!MD->isDirectMethod()) |
501 | argTys.push_back(Elt: Context.getCanonicalParamType(T: Context.getObjCSelType())); |
502 | // FIXME: Kill copy? |
503 | for (const auto *I : MD->parameters()) { |
504 | argTys.push_back(Elt: Context.getCanonicalParamType(T: I->getType())); |
505 | auto extParamInfo = FunctionProtoType::ExtParameterInfo().withIsNoEscape( |
506 | I->hasAttr<NoEscapeAttr>()); |
507 | extParamInfos.push_back(Elt: extParamInfo); |
508 | } |
509 | |
510 | FunctionType::ExtInfo einfo; |
511 | bool IsWindows = getContext().getTargetInfo().getTriple().isOSWindows(); |
512 | einfo = einfo.withCallingConv(cc: getCallingConventionForDecl(D: MD, IsWindows)); |
513 | |
514 | if (getContext().getLangOpts().ObjCAutoRefCount && |
515 | MD->hasAttr<NSReturnsRetainedAttr>()) |
516 | einfo = einfo.withProducesResult(producesResult: true); |
517 | |
518 | RequiredArgs required = |
519 | (MD->isVariadic() ? RequiredArgs(argTys.size()) : RequiredArgs::All); |
520 | |
521 | return arrangeLLVMFunctionInfo(returnType: GetReturnType(RetTy: MD->getReturnType()), |
522 | opts: FnInfoOpts::None, argTypes: argTys, info: einfo, paramInfos: extParamInfos, |
523 | args: required); |
524 | } |
525 | |
526 | const CGFunctionInfo & |
527 | CodeGenTypes::arrangeUnprototypedObjCMessageSend(QualType returnType, |
528 | const CallArgList &args) { |
529 | auto argTypes = getArgTypesForCall(ctx&: Context, args); |
530 | FunctionType::ExtInfo einfo; |
531 | |
532 | return arrangeLLVMFunctionInfo(returnType: GetReturnType(RetTy: returnType), opts: FnInfoOpts::None, |
533 | argTypes, info: einfo, paramInfos: {}, args: RequiredArgs::All); |
534 | } |
535 | |
536 | const CGFunctionInfo & |
537 | CodeGenTypes::arrangeGlobalDeclaration(GlobalDecl GD) { |
538 | // FIXME: Do we need to handle ObjCMethodDecl? |
539 | const FunctionDecl *FD = cast<FunctionDecl>(Val: GD.getDecl()); |
540 | |
541 | if (isa<CXXConstructorDecl>(Val: GD.getDecl()) || |
542 | isa<CXXDestructorDecl>(Val: GD.getDecl())) |
543 | return arrangeCXXStructorDeclaration(GD); |
544 | |
545 | return arrangeFunctionDeclaration(FD); |
546 | } |
547 | |
548 | /// Arrange a thunk that takes 'this' as the first parameter followed by |
549 | /// varargs. Return a void pointer, regardless of the actual return type. |
550 | /// The body of the thunk will end in a musttail call to a function of the |
551 | /// correct type, and the caller will bitcast the function to the correct |
552 | /// prototype. |
553 | const CGFunctionInfo & |
554 | CodeGenTypes::arrangeUnprototypedMustTailThunk(const CXXMethodDecl *MD) { |
555 | assert(MD->isVirtual() && "only methods have thunks" ); |
556 | CanQual<FunctionProtoType> FTP = GetFormalType(MD); |
557 | CanQualType ArgTys[] = {DeriveThisType(RD: MD->getParent(), MD)}; |
558 | return arrangeLLVMFunctionInfo(Context.VoidTy, FnInfoOpts::None, ArgTys, |
559 | FTP->getExtInfo(), {}, RequiredArgs(1)); |
560 | } |
561 | |
562 | const CGFunctionInfo & |
563 | CodeGenTypes::arrangeMSCtorClosure(const CXXConstructorDecl *CD, |
564 | CXXCtorType CT) { |
565 | assert(CT == Ctor_CopyingClosure || CT == Ctor_DefaultClosure); |
566 | |
567 | CanQual<FunctionProtoType> FTP = GetFormalType(CD); |
568 | SmallVector<CanQualType, 2> ArgTys; |
569 | const CXXRecordDecl *RD = CD->getParent(); |
570 | ArgTys.push_back(Elt: DeriveThisType(RD, CD)); |
571 | if (CT == Ctor_CopyingClosure) |
572 | ArgTys.push_back(*FTP->param_type_begin()); |
573 | if (RD->getNumVBases() > 0) |
574 | ArgTys.push_back(Elt: Context.IntTy); |
575 | CallingConv CC = Context.getDefaultCallingConvention( |
576 | /*IsVariadic=*/false, /*IsCXXMethod=*/true); |
577 | return arrangeLLVMFunctionInfo(returnType: Context.VoidTy, opts: FnInfoOpts::IsInstanceMethod, |
578 | argTypes: ArgTys, info: FunctionType::ExtInfo(CC), paramInfos: {}, |
579 | args: RequiredArgs::All); |
580 | } |
581 | |
582 | /// Arrange a call as unto a free function, except possibly with an |
583 | /// additional number of formal parameters considered required. |
584 | static const CGFunctionInfo & |
585 | arrangeFreeFunctionLikeCall(CodeGenTypes &CGT, |
586 | CodeGenModule &CGM, |
587 | const CallArgList &args, |
588 | const FunctionType *fnType, |
589 | unsigned , |
590 | bool chainCall) { |
591 | assert(args.size() >= numExtraRequiredArgs); |
592 | |
593 | llvm::SmallVector<FunctionProtoType::ExtParameterInfo, 16> paramInfos; |
594 | |
595 | // In most cases, there are no optional arguments. |
596 | RequiredArgs required = RequiredArgs::All; |
597 | |
598 | // If we have a variadic prototype, the required arguments are the |
599 | // extra prefix plus the arguments in the prototype. |
600 | if (const FunctionProtoType *proto = dyn_cast<FunctionProtoType>(Val: fnType)) { |
601 | if (proto->isVariadic()) |
602 | required = RequiredArgs::forPrototypePlus(prototype: proto, additional: numExtraRequiredArgs); |
603 | |
604 | if (proto->hasExtParameterInfos()) |
605 | addExtParameterInfosForCall(paramInfos, proto, prefixArgs: numExtraRequiredArgs, |
606 | totalArgs: args.size()); |
607 | |
608 | // If we don't have a prototype at all, but we're supposed to |
609 | // explicitly use the variadic convention for unprototyped calls, |
610 | // treat all of the arguments as required but preserve the nominal |
611 | // possibility of variadics. |
612 | } else if (CGM.getTargetCodeGenInfo() |
613 | .isNoProtoCallVariadic(args, |
614 | fnType: cast<FunctionNoProtoType>(Val: fnType))) { |
615 | required = RequiredArgs(args.size()); |
616 | } |
617 | |
618 | // FIXME: Kill copy. |
619 | SmallVector<CanQualType, 16> argTypes; |
620 | for (const auto &arg : args) |
621 | argTypes.push_back(Elt: CGT.getContext().getCanonicalParamType(T: arg.Ty)); |
622 | FnInfoOpts opts = chainCall ? FnInfoOpts::IsChainCall : FnInfoOpts::None; |
623 | return CGT.arrangeLLVMFunctionInfo(returnType: GetReturnType(RetTy: fnType->getReturnType()), |
624 | opts, argTypes, info: fnType->getExtInfo(), |
625 | paramInfos, args: required); |
626 | } |
627 | |
628 | /// Figure out the rules for calling a function with the given formal |
629 | /// type using the given arguments. The arguments are necessary |
630 | /// because the function might be unprototyped, in which case it's |
631 | /// target-dependent in crazy ways. |
632 | const CGFunctionInfo & |
633 | CodeGenTypes::arrangeFreeFunctionCall(const CallArgList &args, |
634 | const FunctionType *fnType, |
635 | bool chainCall) { |
636 | return arrangeFreeFunctionLikeCall(CGT&: *this, CGM, args, fnType, |
637 | numExtraRequiredArgs: chainCall ? 1 : 0, chainCall); |
638 | } |
639 | |
640 | /// A block function is essentially a free function with an |
641 | /// extra implicit argument. |
642 | const CGFunctionInfo & |
643 | CodeGenTypes::arrangeBlockFunctionCall(const CallArgList &args, |
644 | const FunctionType *fnType) { |
645 | return arrangeFreeFunctionLikeCall(CGT&: *this, CGM, args, fnType, numExtraRequiredArgs: 1, |
646 | /*chainCall=*/false); |
647 | } |
648 | |
649 | const CGFunctionInfo & |
650 | CodeGenTypes::arrangeBlockFunctionDeclaration(const FunctionProtoType *proto, |
651 | const FunctionArgList ¶ms) { |
652 | auto paramInfos = getExtParameterInfosForCall(proto, prefixArgs: 1, totalArgs: params.size()); |
653 | auto argTypes = getArgTypesForDeclaration(ctx&: Context, args: params); |
654 | |
655 | return arrangeLLVMFunctionInfo(returnType: GetReturnType(proto->getReturnType()), |
656 | opts: FnInfoOpts::None, argTypes, |
657 | info: proto->getExtInfo(), paramInfos, |
658 | args: RequiredArgs::forPrototypePlus(prototype: proto, additional: 1)); |
659 | } |
660 | |
661 | const CGFunctionInfo & |
662 | CodeGenTypes::arrangeBuiltinFunctionCall(QualType resultType, |
663 | const CallArgList &args) { |
664 | // FIXME: Kill copy. |
665 | SmallVector<CanQualType, 16> argTypes; |
666 | for (const auto &Arg : args) |
667 | argTypes.push_back(Elt: Context.getCanonicalParamType(T: Arg.Ty)); |
668 | return arrangeLLVMFunctionInfo(returnType: GetReturnType(RetTy: resultType), opts: FnInfoOpts::None, |
669 | argTypes, info: FunctionType::ExtInfo(), |
670 | /*paramInfos=*/{}, args: RequiredArgs::All); |
671 | } |
672 | |
673 | const CGFunctionInfo & |
674 | CodeGenTypes::arrangeBuiltinFunctionDeclaration(QualType resultType, |
675 | const FunctionArgList &args) { |
676 | auto argTypes = getArgTypesForDeclaration(ctx&: Context, args); |
677 | |
678 | return arrangeLLVMFunctionInfo(returnType: GetReturnType(RetTy: resultType), opts: FnInfoOpts::None, |
679 | argTypes, info: FunctionType::ExtInfo(), paramInfos: {}, |
680 | args: RequiredArgs::All); |
681 | } |
682 | |
683 | const CGFunctionInfo & |
684 | CodeGenTypes::arrangeBuiltinFunctionDeclaration(CanQualType resultType, |
685 | ArrayRef<CanQualType> argTypes) { |
686 | return arrangeLLVMFunctionInfo(returnType: resultType, opts: FnInfoOpts::None, argTypes, |
687 | info: FunctionType::ExtInfo(), paramInfos: {}, |
688 | args: RequiredArgs::All); |
689 | } |
690 | |
691 | /// Arrange a call to a C++ method, passing the given arguments. |
692 | /// |
693 | /// numPrefixArgs is the number of ABI-specific prefix arguments we have. It |
694 | /// does not count `this`. |
695 | const CGFunctionInfo & |
696 | CodeGenTypes::arrangeCXXMethodCall(const CallArgList &args, |
697 | const FunctionProtoType *proto, |
698 | RequiredArgs required, |
699 | unsigned numPrefixArgs) { |
700 | assert(numPrefixArgs + 1 <= args.size() && |
701 | "Emitting a call with less args than the required prefix?" ); |
702 | // Add one to account for `this`. It's a bit awkward here, but we don't count |
703 | // `this` in similar places elsewhere. |
704 | auto paramInfos = |
705 | getExtParameterInfosForCall(proto, prefixArgs: numPrefixArgs + 1, totalArgs: args.size()); |
706 | |
707 | // FIXME: Kill copy. |
708 | auto argTypes = getArgTypesForCall(ctx&: Context, args); |
709 | |
710 | FunctionType::ExtInfo info = proto->getExtInfo(); |
711 | return arrangeLLVMFunctionInfo(returnType: GetReturnType(proto->getReturnType()), |
712 | opts: FnInfoOpts::IsInstanceMethod, argTypes, info, |
713 | paramInfos, args: required); |
714 | } |
715 | |
716 | const CGFunctionInfo &CodeGenTypes::arrangeNullaryFunction() { |
717 | return arrangeLLVMFunctionInfo(returnType: getContext().VoidTy, opts: FnInfoOpts::None, |
718 | argTypes: std::nullopt, info: FunctionType::ExtInfo(), paramInfos: {}, |
719 | args: RequiredArgs::All); |
720 | } |
721 | |
722 | const CGFunctionInfo & |
723 | CodeGenTypes::arrangeCall(const CGFunctionInfo &signature, |
724 | const CallArgList &args) { |
725 | assert(signature.arg_size() <= args.size()); |
726 | if (signature.arg_size() == args.size()) |
727 | return signature; |
728 | |
729 | SmallVector<FunctionProtoType::ExtParameterInfo, 16> paramInfos; |
730 | auto sigParamInfos = signature.getExtParameterInfos(); |
731 | if (!sigParamInfos.empty()) { |
732 | paramInfos.append(in_start: sigParamInfos.begin(), in_end: sigParamInfos.end()); |
733 | paramInfos.resize(N: args.size()); |
734 | } |
735 | |
736 | auto argTypes = getArgTypesForCall(ctx&: Context, args); |
737 | |
738 | assert(signature.getRequiredArgs().allowsOptionalArgs()); |
739 | FnInfoOpts opts = FnInfoOpts::None; |
740 | if (signature.isInstanceMethod()) |
741 | opts |= FnInfoOpts::IsInstanceMethod; |
742 | if (signature.isChainCall()) |
743 | opts |= FnInfoOpts::IsChainCall; |
744 | if (signature.isDelegateCall()) |
745 | opts |= FnInfoOpts::IsDelegateCall; |
746 | return arrangeLLVMFunctionInfo(returnType: signature.getReturnType(), opts, argTypes, |
747 | info: signature.getExtInfo(), paramInfos, |
748 | args: signature.getRequiredArgs()); |
749 | } |
750 | |
751 | namespace clang { |
752 | namespace CodeGen { |
753 | void computeSPIRKernelABIInfo(CodeGenModule &CGM, CGFunctionInfo &FI); |
754 | } |
755 | } |
756 | |
757 | /// Arrange the argument and result information for an abstract value |
758 | /// of a given function type. This is the method which all of the |
759 | /// above functions ultimately defer to. |
760 | const CGFunctionInfo &CodeGenTypes::arrangeLLVMFunctionInfo( |
761 | CanQualType resultType, FnInfoOpts opts, ArrayRef<CanQualType> argTypes, |
762 | FunctionType::ExtInfo info, |
763 | ArrayRef<FunctionProtoType::ExtParameterInfo> paramInfos, |
764 | RequiredArgs required) { |
765 | assert(llvm::all_of(argTypes, |
766 | [](CanQualType T) { return T.isCanonicalAsParam(); })); |
767 | |
768 | // Lookup or create unique function info. |
769 | llvm::FoldingSetNodeID ID; |
770 | bool isInstanceMethod = |
771 | (opts & FnInfoOpts::IsInstanceMethod) == FnInfoOpts::IsInstanceMethod; |
772 | bool isChainCall = |
773 | (opts & FnInfoOpts::IsChainCall) == FnInfoOpts::IsChainCall; |
774 | bool isDelegateCall = |
775 | (opts & FnInfoOpts::IsDelegateCall) == FnInfoOpts::IsDelegateCall; |
776 | CGFunctionInfo::Profile(ID, InstanceMethod: isInstanceMethod, ChainCall: isChainCall, IsDelegateCall: isDelegateCall, |
777 | info, paramInfos, required, resultType, argTypes); |
778 | |
779 | void *insertPos = nullptr; |
780 | CGFunctionInfo *FI = FunctionInfos.FindNodeOrInsertPos(ID, InsertPos&: insertPos); |
781 | if (FI) |
782 | return *FI; |
783 | |
784 | unsigned CC = ClangCallConvToLLVMCallConv(CC: info.getCC()); |
785 | |
786 | // Construct the function info. We co-allocate the ArgInfos. |
787 | FI = CGFunctionInfo::create(llvmCC: CC, instanceMethod: isInstanceMethod, chainCall: isChainCall, delegateCall: isDelegateCall, |
788 | extInfo: info, paramInfos, resultType, argTypes, required); |
789 | FunctionInfos.InsertNode(N: FI, InsertPos: insertPos); |
790 | |
791 | bool inserted = FunctionsBeingProcessed.insert(Ptr: FI).second; |
792 | (void)inserted; |
793 | assert(inserted && "Recursively being processed?" ); |
794 | |
795 | // Compute ABI information. |
796 | if (CC == llvm::CallingConv::SPIR_KERNEL) { |
797 | // Force target independent argument handling for the host visible |
798 | // kernel functions. |
799 | computeSPIRKernelABIInfo(CGM, FI&: *FI); |
800 | } else if (info.getCC() == CC_Swift || info.getCC() == CC_SwiftAsync) { |
801 | swiftcall::computeABIInfo(CGM, FI&: *FI); |
802 | } else { |
803 | getABIInfo().computeInfo(FI&: *FI); |
804 | } |
805 | |
806 | // Loop over all of the computed argument and return value info. If any of |
807 | // them are direct or extend without a specified coerce type, specify the |
808 | // default now. |
809 | ABIArgInfo &retInfo = FI->getReturnInfo(); |
810 | if (retInfo.canHaveCoerceToType() && retInfo.getCoerceToType() == nullptr) |
811 | retInfo.setCoerceToType(ConvertType(T: FI->getReturnType())); |
812 | |
813 | for (auto &I : FI->arguments()) |
814 | if (I.info.canHaveCoerceToType() && I.info.getCoerceToType() == nullptr) |
815 | I.info.setCoerceToType(ConvertType(T: I.type)); |
816 | |
817 | bool erased = FunctionsBeingProcessed.erase(Ptr: FI); (void)erased; |
818 | assert(erased && "Not in set?" ); |
819 | |
820 | return *FI; |
821 | } |
822 | |
823 | CGFunctionInfo *CGFunctionInfo::create(unsigned llvmCC, bool instanceMethod, |
824 | bool chainCall, bool delegateCall, |
825 | const FunctionType::ExtInfo &info, |
826 | ArrayRef<ExtParameterInfo> paramInfos, |
827 | CanQualType resultType, |
828 | ArrayRef<CanQualType> argTypes, |
829 | RequiredArgs required) { |
830 | assert(paramInfos.empty() || paramInfos.size() == argTypes.size()); |
831 | assert(!required.allowsOptionalArgs() || |
832 | required.getNumRequiredArgs() <= argTypes.size()); |
833 | |
834 | void *buffer = |
835 | operator new(totalSizeToAlloc<ArgInfo, ExtParameterInfo>( |
836 | Counts: argTypes.size() + 1, Counts: paramInfos.size())); |
837 | |
838 | CGFunctionInfo *FI = new(buffer) CGFunctionInfo(); |
839 | FI->CallingConvention = llvmCC; |
840 | FI->EffectiveCallingConvention = llvmCC; |
841 | FI->ASTCallingConvention = info.getCC(); |
842 | FI->InstanceMethod = instanceMethod; |
843 | FI->ChainCall = chainCall; |
844 | FI->DelegateCall = delegateCall; |
845 | FI->CmseNSCall = info.getCmseNSCall(); |
846 | FI->NoReturn = info.getNoReturn(); |
847 | FI->ReturnsRetained = info.getProducesResult(); |
848 | FI->NoCallerSavedRegs = info.getNoCallerSavedRegs(); |
849 | FI->NoCfCheck = info.getNoCfCheck(); |
850 | FI->Required = required; |
851 | FI->HasRegParm = info.getHasRegParm(); |
852 | FI->RegParm = info.getRegParm(); |
853 | FI->ArgStruct = nullptr; |
854 | FI->ArgStructAlign = 0; |
855 | FI->NumArgs = argTypes.size(); |
856 | FI->HasExtParameterInfos = !paramInfos.empty(); |
857 | FI->getArgsBuffer()[0].type = resultType; |
858 | FI->MaxVectorWidth = 0; |
859 | for (unsigned i = 0, e = argTypes.size(); i != e; ++i) |
860 | FI->getArgsBuffer()[i + 1].type = argTypes[i]; |
861 | for (unsigned i = 0, e = paramInfos.size(); i != e; ++i) |
862 | FI->getExtParameterInfosBuffer()[i] = paramInfos[i]; |
863 | return FI; |
864 | } |
865 | |
866 | /***/ |
867 | |
868 | namespace { |
869 | // ABIArgInfo::Expand implementation. |
870 | |
871 | // Specifies the way QualType passed as ABIArgInfo::Expand is expanded. |
872 | struct TypeExpansion { |
873 | enum TypeExpansionKind { |
874 | // Elements of constant arrays are expanded recursively. |
875 | TEK_ConstantArray, |
876 | // Record fields are expanded recursively (but if record is a union, only |
877 | // the field with the largest size is expanded). |
878 | TEK_Record, |
879 | // For complex types, real and imaginary parts are expanded recursively. |
880 | TEK_Complex, |
881 | // All other types are not expandable. |
882 | TEK_None |
883 | }; |
884 | |
885 | const TypeExpansionKind Kind; |
886 | |
887 | TypeExpansion(TypeExpansionKind K) : Kind(K) {} |
888 | virtual ~TypeExpansion() {} |
889 | }; |
890 | |
891 | struct ConstantArrayExpansion : TypeExpansion { |
892 | QualType EltTy; |
893 | uint64_t NumElts; |
894 | |
895 | ConstantArrayExpansion(QualType EltTy, uint64_t NumElts) |
896 | : TypeExpansion(TEK_ConstantArray), EltTy(EltTy), NumElts(NumElts) {} |
897 | static bool classof(const TypeExpansion *TE) { |
898 | return TE->Kind == TEK_ConstantArray; |
899 | } |
900 | }; |
901 | |
902 | struct RecordExpansion : TypeExpansion { |
903 | SmallVector<const CXXBaseSpecifier *, 1> Bases; |
904 | |
905 | SmallVector<const FieldDecl *, 1> Fields; |
906 | |
907 | RecordExpansion(SmallVector<const CXXBaseSpecifier *, 1> &&Bases, |
908 | SmallVector<const FieldDecl *, 1> &&Fields) |
909 | : TypeExpansion(TEK_Record), Bases(std::move(Bases)), |
910 | Fields(std::move(Fields)) {} |
911 | static bool classof(const TypeExpansion *TE) { |
912 | return TE->Kind == TEK_Record; |
913 | } |
914 | }; |
915 | |
916 | struct ComplexExpansion : TypeExpansion { |
917 | QualType EltTy; |
918 | |
919 | ComplexExpansion(QualType EltTy) : TypeExpansion(TEK_Complex), EltTy(EltTy) {} |
920 | static bool classof(const TypeExpansion *TE) { |
921 | return TE->Kind == TEK_Complex; |
922 | } |
923 | }; |
924 | |
925 | struct NoExpansion : TypeExpansion { |
926 | NoExpansion() : TypeExpansion(TEK_None) {} |
927 | static bool classof(const TypeExpansion *TE) { |
928 | return TE->Kind == TEK_None; |
929 | } |
930 | }; |
931 | } // namespace |
932 | |
933 | static std::unique_ptr<TypeExpansion> |
934 | getTypeExpansion(QualType Ty, const ASTContext &Context) { |
935 | if (const ConstantArrayType *AT = Context.getAsConstantArrayType(T: Ty)) { |
936 | return std::make_unique<ConstantArrayExpansion>( |
937 | AT->getElementType(), AT->getSize().getZExtValue()); |
938 | } |
939 | if (const RecordType *RT = Ty->getAs<RecordType>()) { |
940 | SmallVector<const CXXBaseSpecifier *, 1> Bases; |
941 | SmallVector<const FieldDecl *, 1> Fields; |
942 | const RecordDecl *RD = RT->getDecl(); |
943 | assert(!RD->hasFlexibleArrayMember() && |
944 | "Cannot expand structure with flexible array." ); |
945 | if (RD->isUnion()) { |
946 | // Unions can be here only in degenerative cases - all the fields are same |
947 | // after flattening. Thus we have to use the "largest" field. |
948 | const FieldDecl *LargestFD = nullptr; |
949 | CharUnits UnionSize = CharUnits::Zero(); |
950 | |
951 | for (const auto *FD : RD->fields()) { |
952 | if (FD->isZeroLengthBitField(Ctx: Context)) |
953 | continue; |
954 | assert(!FD->isBitField() && |
955 | "Cannot expand structure with bit-field members." ); |
956 | CharUnits FieldSize = Context.getTypeSizeInChars(FD->getType()); |
957 | if (UnionSize < FieldSize) { |
958 | UnionSize = FieldSize; |
959 | LargestFD = FD; |
960 | } |
961 | } |
962 | if (LargestFD) |
963 | Fields.push_back(Elt: LargestFD); |
964 | } else { |
965 | if (const auto *CXXRD = dyn_cast<CXXRecordDecl>(Val: RD)) { |
966 | assert(!CXXRD->isDynamicClass() && |
967 | "cannot expand vtable pointers in dynamic classes" ); |
968 | llvm::append_range(C&: Bases, R: llvm::make_pointer_range(Range: CXXRD->bases())); |
969 | } |
970 | |
971 | for (const auto *FD : RD->fields()) { |
972 | if (FD->isZeroLengthBitField(Ctx: Context)) |
973 | continue; |
974 | assert(!FD->isBitField() && |
975 | "Cannot expand structure with bit-field members." ); |
976 | Fields.push_back(Elt: FD); |
977 | } |
978 | } |
979 | return std::make_unique<RecordExpansion>(args: std::move(Bases), |
980 | args: std::move(Fields)); |
981 | } |
982 | if (const ComplexType *CT = Ty->getAs<ComplexType>()) { |
983 | return std::make_unique<ComplexExpansion>(args: CT->getElementType()); |
984 | } |
985 | return std::make_unique<NoExpansion>(); |
986 | } |
987 | |
988 | static int getExpansionSize(QualType Ty, const ASTContext &Context) { |
989 | auto Exp = getTypeExpansion(Ty, Context); |
990 | if (auto CAExp = dyn_cast<ConstantArrayExpansion>(Val: Exp.get())) { |
991 | return CAExp->NumElts * getExpansionSize(CAExp->EltTy, Context); |
992 | } |
993 | if (auto RExp = dyn_cast<RecordExpansion>(Val: Exp.get())) { |
994 | int Res = 0; |
995 | for (auto BS : RExp->Bases) |
996 | Res += getExpansionSize(Ty: BS->getType(), Context); |
997 | for (auto FD : RExp->Fields) |
998 | Res += getExpansionSize(FD->getType(), Context); |
999 | return Res; |
1000 | } |
1001 | if (isa<ComplexExpansion>(Val: Exp.get())) |
1002 | return 2; |
1003 | assert(isa<NoExpansion>(Exp.get())); |
1004 | return 1; |
1005 | } |
1006 | |
1007 | void |
1008 | CodeGenTypes::getExpandedTypes(QualType Ty, |
1009 | SmallVectorImpl<llvm::Type *>::iterator &TI) { |
1010 | auto Exp = getTypeExpansion(Ty, Context); |
1011 | if (auto CAExp = dyn_cast<ConstantArrayExpansion>(Val: Exp.get())) { |
1012 | for (int i = 0, n = CAExp->NumElts; i < n; i++) { |
1013 | getExpandedTypes(Ty: CAExp->EltTy, TI); |
1014 | } |
1015 | } else if (auto RExp = dyn_cast<RecordExpansion>(Val: Exp.get())) { |
1016 | for (auto BS : RExp->Bases) |
1017 | getExpandedTypes(Ty: BS->getType(), TI); |
1018 | for (auto FD : RExp->Fields) |
1019 | getExpandedTypes(Ty: FD->getType(), TI); |
1020 | } else if (auto CExp = dyn_cast<ComplexExpansion>(Val: Exp.get())) { |
1021 | llvm::Type *EltTy = ConvertType(T: CExp->EltTy); |
1022 | *TI++ = EltTy; |
1023 | *TI++ = EltTy; |
1024 | } else { |
1025 | assert(isa<NoExpansion>(Exp.get())); |
1026 | *TI++ = ConvertType(T: Ty); |
1027 | } |
1028 | } |
1029 | |
1030 | static void forConstantArrayExpansion(CodeGenFunction &CGF, |
1031 | ConstantArrayExpansion *CAE, |
1032 | Address BaseAddr, |
1033 | llvm::function_ref<void(Address)> Fn) { |
1034 | CharUnits EltSize = CGF.getContext().getTypeSizeInChars(CAE->EltTy); |
1035 | CharUnits EltAlign = |
1036 | BaseAddr.getAlignment().alignmentOfArrayElement(elementSize: EltSize); |
1037 | llvm::Type *EltTy = CGF.ConvertTypeForMem(T: CAE->EltTy); |
1038 | |
1039 | for (int i = 0, n = CAE->NumElts; i < n; i++) { |
1040 | llvm::Value *EltAddr = CGF.Builder.CreateConstGEP2_32( |
1041 | Ty: BaseAddr.getElementType(), Ptr: BaseAddr.getPointer(), Idx0: 0, Idx1: i); |
1042 | Fn(Address(EltAddr, EltTy, EltAlign)); |
1043 | } |
1044 | } |
1045 | |
1046 | void CodeGenFunction::ExpandTypeFromArgs(QualType Ty, LValue LV, |
1047 | llvm::Function::arg_iterator &AI) { |
1048 | assert(LV.isSimple() && |
1049 | "Unexpected non-simple lvalue during struct expansion." ); |
1050 | |
1051 | auto Exp = getTypeExpansion(Ty, Context: getContext()); |
1052 | if (auto CAExp = dyn_cast<ConstantArrayExpansion>(Val: Exp.get())) { |
1053 | forConstantArrayExpansion( |
1054 | CGF&: *this, CAE: CAExp, BaseAddr: LV.getAddress(CGF&: *this), Fn: [&](Address EltAddr) { |
1055 | LValue LV = MakeAddrLValue(EltAddr, CAExp->EltTy); |
1056 | ExpandTypeFromArgs(Ty: CAExp->EltTy, LV, AI); |
1057 | }); |
1058 | } else if (auto RExp = dyn_cast<RecordExpansion>(Val: Exp.get())) { |
1059 | Address This = LV.getAddress(CGF&: *this); |
1060 | for (const CXXBaseSpecifier *BS : RExp->Bases) { |
1061 | // Perform a single step derived-to-base conversion. |
1062 | Address Base = |
1063 | GetAddressOfBaseClass(Value: This, Derived: Ty->getAsCXXRecordDecl(), PathBegin: &BS, PathEnd: &BS + 1, |
1064 | /*NullCheckValue=*/false, Loc: SourceLocation()); |
1065 | LValue SubLV = MakeAddrLValue(Addr: Base, T: BS->getType()); |
1066 | |
1067 | // Recurse onto bases. |
1068 | ExpandTypeFromArgs(Ty: BS->getType(), LV: SubLV, AI); |
1069 | } |
1070 | for (auto FD : RExp->Fields) { |
1071 | // FIXME: What are the right qualifiers here? |
1072 | LValue SubLV = EmitLValueForFieldInitialization(Base: LV, Field: FD); |
1073 | ExpandTypeFromArgs(Ty: FD->getType(), LV: SubLV, AI); |
1074 | } |
1075 | } else if (isa<ComplexExpansion>(Val: Exp.get())) { |
1076 | auto realValue = &*AI++; |
1077 | auto imagValue = &*AI++; |
1078 | EmitStoreOfComplex(V: ComplexPairTy(realValue, imagValue), dest: LV, /*init*/ isInit: true); |
1079 | } else { |
1080 | // Call EmitStoreOfScalar except when the lvalue is a bitfield to emit a |
1081 | // primitive store. |
1082 | assert(isa<NoExpansion>(Exp.get())); |
1083 | llvm::Value *Arg = &*AI++; |
1084 | if (LV.isBitField()) { |
1085 | EmitStoreThroughLValue(Src: RValue::get(V: Arg), Dst: LV); |
1086 | } else { |
1087 | // TODO: currently there are some places are inconsistent in what LLVM |
1088 | // pointer type they use (see D118744). Once clang uses opaque pointers |
1089 | // all LLVM pointer types will be the same and we can remove this check. |
1090 | if (Arg->getType()->isPointerTy()) { |
1091 | Address Addr = LV.getAddress(CGF&: *this); |
1092 | Arg = Builder.CreateBitCast(V: Arg, DestTy: Addr.getElementType()); |
1093 | } |
1094 | EmitStoreOfScalar(value: Arg, lvalue: LV); |
1095 | } |
1096 | } |
1097 | } |
1098 | |
1099 | void CodeGenFunction::ExpandTypeToArgs( |
1100 | QualType Ty, CallArg Arg, llvm::FunctionType *IRFuncTy, |
1101 | SmallVectorImpl<llvm::Value *> &IRCallArgs, unsigned &IRCallArgPos) { |
1102 | auto Exp = getTypeExpansion(Ty, Context: getContext()); |
1103 | if (auto CAExp = dyn_cast<ConstantArrayExpansion>(Val: Exp.get())) { |
1104 | Address Addr = Arg.hasLValue() ? Arg.getKnownLValue().getAddress(CGF&: *this) |
1105 | : Arg.getKnownRValue().getAggregateAddress(); |
1106 | forConstantArrayExpansion( |
1107 | CGF&: *this, CAE: CAExp, BaseAddr: Addr, Fn: [&](Address EltAddr) { |
1108 | CallArg EltArg = CallArg( |
1109 | convertTempToRValue(addr: EltAddr, type: CAExp->EltTy, Loc: SourceLocation()), |
1110 | CAExp->EltTy); |
1111 | ExpandTypeToArgs(Ty: CAExp->EltTy, Arg: EltArg, IRFuncTy, IRCallArgs, |
1112 | IRCallArgPos); |
1113 | }); |
1114 | } else if (auto RExp = dyn_cast<RecordExpansion>(Val: Exp.get())) { |
1115 | Address This = Arg.hasLValue() ? Arg.getKnownLValue().getAddress(CGF&: *this) |
1116 | : Arg.getKnownRValue().getAggregateAddress(); |
1117 | for (const CXXBaseSpecifier *BS : RExp->Bases) { |
1118 | // Perform a single step derived-to-base conversion. |
1119 | Address Base = |
1120 | GetAddressOfBaseClass(Value: This, Derived: Ty->getAsCXXRecordDecl(), PathBegin: &BS, PathEnd: &BS + 1, |
1121 | /*NullCheckValue=*/false, Loc: SourceLocation()); |
1122 | CallArg BaseArg = CallArg(RValue::getAggregate(addr: Base), BS->getType()); |
1123 | |
1124 | // Recurse onto bases. |
1125 | ExpandTypeToArgs(Ty: BS->getType(), Arg: BaseArg, IRFuncTy, IRCallArgs, |
1126 | IRCallArgPos); |
1127 | } |
1128 | |
1129 | LValue LV = MakeAddrLValue(Addr: This, T: Ty); |
1130 | for (auto FD : RExp->Fields) { |
1131 | CallArg FldArg = |
1132 | CallArg(EmitRValueForField(LV, FD, Loc: SourceLocation()), FD->getType()); |
1133 | ExpandTypeToArgs(Ty: FD->getType(), Arg: FldArg, IRFuncTy, IRCallArgs, |
1134 | IRCallArgPos); |
1135 | } |
1136 | } else if (isa<ComplexExpansion>(Val: Exp.get())) { |
1137 | ComplexPairTy CV = Arg.getKnownRValue().getComplexVal(); |
1138 | IRCallArgs[IRCallArgPos++] = CV.first; |
1139 | IRCallArgs[IRCallArgPos++] = CV.second; |
1140 | } else { |
1141 | assert(isa<NoExpansion>(Exp.get())); |
1142 | auto RV = Arg.getKnownRValue(); |
1143 | assert(RV.isScalar() && |
1144 | "Unexpected non-scalar rvalue during struct expansion." ); |
1145 | |
1146 | // Insert a bitcast as needed. |
1147 | llvm::Value *V = RV.getScalarVal(); |
1148 | if (IRCallArgPos < IRFuncTy->getNumParams() && |
1149 | V->getType() != IRFuncTy->getParamType(i: IRCallArgPos)) |
1150 | V = Builder.CreateBitCast(V, DestTy: IRFuncTy->getParamType(i: IRCallArgPos)); |
1151 | |
1152 | IRCallArgs[IRCallArgPos++] = V; |
1153 | } |
1154 | } |
1155 | |
1156 | /// Create a temporary allocation for the purposes of coercion. |
1157 | static Address CreateTempAllocaForCoercion(CodeGenFunction &CGF, llvm::Type *Ty, |
1158 | CharUnits MinAlign, |
1159 | const Twine &Name = "tmp" ) { |
1160 | // Don't use an alignment that's worse than what LLVM would prefer. |
1161 | auto PrefAlign = CGF.CGM.getDataLayout().getPrefTypeAlign(Ty); |
1162 | CharUnits Align = std::max(a: MinAlign, b: CharUnits::fromQuantity(Quantity: PrefAlign)); |
1163 | |
1164 | return CGF.CreateTempAlloca(Ty, align: Align, Name: Name + ".coerce" ); |
1165 | } |
1166 | |
1167 | /// EnterStructPointerForCoercedAccess - Given a struct pointer that we are |
1168 | /// accessing some number of bytes out of it, try to gep into the struct to get |
1169 | /// at its inner goodness. Dive as deep as possible without entering an element |
1170 | /// with an in-memory size smaller than DstSize. |
1171 | static Address |
1172 | EnterStructPointerForCoercedAccess(Address SrcPtr, |
1173 | llvm::StructType *SrcSTy, |
1174 | uint64_t DstSize, CodeGenFunction &CGF) { |
1175 | // We can't dive into a zero-element struct. |
1176 | if (SrcSTy->getNumElements() == 0) return SrcPtr; |
1177 | |
1178 | llvm::Type *FirstElt = SrcSTy->getElementType(N: 0); |
1179 | |
1180 | // If the first elt is at least as large as what we're looking for, or if the |
1181 | // first element is the same size as the whole struct, we can enter it. The |
1182 | // comparison must be made on the store size and not the alloca size. Using |
1183 | // the alloca size may overstate the size of the load. |
1184 | uint64_t FirstEltSize = |
1185 | CGF.CGM.getDataLayout().getTypeStoreSize(Ty: FirstElt); |
1186 | if (FirstEltSize < DstSize && |
1187 | FirstEltSize < CGF.CGM.getDataLayout().getTypeStoreSize(Ty: SrcSTy)) |
1188 | return SrcPtr; |
1189 | |
1190 | // GEP into the first element. |
1191 | SrcPtr = CGF.Builder.CreateStructGEP(Addr: SrcPtr, Index: 0, Name: "coerce.dive" ); |
1192 | |
1193 | // If the first element is a struct, recurse. |
1194 | llvm::Type *SrcTy = SrcPtr.getElementType(); |
1195 | if (llvm::StructType *SrcSTy = dyn_cast<llvm::StructType>(Val: SrcTy)) |
1196 | return EnterStructPointerForCoercedAccess(SrcPtr, SrcSTy, DstSize, CGF); |
1197 | |
1198 | return SrcPtr; |
1199 | } |
1200 | |
1201 | /// CoerceIntOrPtrToIntOrPtr - Convert a value Val to the specific Ty where both |
1202 | /// are either integers or pointers. This does a truncation of the value if it |
1203 | /// is too large or a zero extension if it is too small. |
1204 | /// |
1205 | /// This behaves as if the value were coerced through memory, so on big-endian |
1206 | /// targets the high bits are preserved in a truncation, while little-endian |
1207 | /// targets preserve the low bits. |
1208 | static llvm::Value *CoerceIntOrPtrToIntOrPtr(llvm::Value *Val, |
1209 | llvm::Type *Ty, |
1210 | CodeGenFunction &CGF) { |
1211 | if (Val->getType() == Ty) |
1212 | return Val; |
1213 | |
1214 | if (isa<llvm::PointerType>(Val: Val->getType())) { |
1215 | // If this is Pointer->Pointer avoid conversion to and from int. |
1216 | if (isa<llvm::PointerType>(Val: Ty)) |
1217 | return CGF.Builder.CreateBitCast(V: Val, DestTy: Ty, Name: "coerce.val" ); |
1218 | |
1219 | // Convert the pointer to an integer so we can play with its width. |
1220 | Val = CGF.Builder.CreatePtrToInt(V: Val, DestTy: CGF.IntPtrTy, Name: "coerce.val.pi" ); |
1221 | } |
1222 | |
1223 | llvm::Type *DestIntTy = Ty; |
1224 | if (isa<llvm::PointerType>(Val: DestIntTy)) |
1225 | DestIntTy = CGF.IntPtrTy; |
1226 | |
1227 | if (Val->getType() != DestIntTy) { |
1228 | const llvm::DataLayout &DL = CGF.CGM.getDataLayout(); |
1229 | if (DL.isBigEndian()) { |
1230 | // Preserve the high bits on big-endian targets. |
1231 | // That is what memory coercion does. |
1232 | uint64_t SrcSize = DL.getTypeSizeInBits(Ty: Val->getType()); |
1233 | uint64_t DstSize = DL.getTypeSizeInBits(Ty: DestIntTy); |
1234 | |
1235 | if (SrcSize > DstSize) { |
1236 | Val = CGF.Builder.CreateLShr(LHS: Val, RHS: SrcSize - DstSize, Name: "coerce.highbits" ); |
1237 | Val = CGF.Builder.CreateTrunc(V: Val, DestTy: DestIntTy, Name: "coerce.val.ii" ); |
1238 | } else { |
1239 | Val = CGF.Builder.CreateZExt(V: Val, DestTy: DestIntTy, Name: "coerce.val.ii" ); |
1240 | Val = CGF.Builder.CreateShl(LHS: Val, RHS: DstSize - SrcSize, Name: "coerce.highbits" ); |
1241 | } |
1242 | } else { |
1243 | // Little-endian targets preserve the low bits. No shifts required. |
1244 | Val = CGF.Builder.CreateIntCast(V: Val, DestTy: DestIntTy, isSigned: false, Name: "coerce.val.ii" ); |
1245 | } |
1246 | } |
1247 | |
1248 | if (isa<llvm::PointerType>(Val: Ty)) |
1249 | Val = CGF.Builder.CreateIntToPtr(V: Val, DestTy: Ty, Name: "coerce.val.ip" ); |
1250 | return Val; |
1251 | } |
1252 | |
1253 | |
1254 | |
1255 | /// CreateCoercedLoad - Create a load from \arg SrcPtr interpreted as |
1256 | /// a pointer to an object of type \arg Ty, known to be aligned to |
1257 | /// \arg SrcAlign bytes. |
1258 | /// |
1259 | /// This safely handles the case when the src type is smaller than the |
1260 | /// destination type; in this situation the values of bits which not |
1261 | /// present in the src are undefined. |
1262 | static llvm::Value *CreateCoercedLoad(Address Src, llvm::Type *Ty, |
1263 | CodeGenFunction &CGF) { |
1264 | llvm::Type *SrcTy = Src.getElementType(); |
1265 | |
1266 | // If SrcTy and Ty are the same, just do a load. |
1267 | if (SrcTy == Ty) |
1268 | return CGF.Builder.CreateLoad(Addr: Src); |
1269 | |
1270 | llvm::TypeSize DstSize = CGF.CGM.getDataLayout().getTypeAllocSize(Ty); |
1271 | |
1272 | if (llvm::StructType *SrcSTy = dyn_cast<llvm::StructType>(Val: SrcTy)) { |
1273 | Src = EnterStructPointerForCoercedAccess(SrcPtr: Src, SrcSTy, |
1274 | DstSize: DstSize.getFixedValue(), CGF); |
1275 | SrcTy = Src.getElementType(); |
1276 | } |
1277 | |
1278 | llvm::TypeSize SrcSize = CGF.CGM.getDataLayout().getTypeAllocSize(Ty: SrcTy); |
1279 | |
1280 | // If the source and destination are integer or pointer types, just do an |
1281 | // extension or truncation to the desired type. |
1282 | if ((isa<llvm::IntegerType>(Val: Ty) || isa<llvm::PointerType>(Val: Ty)) && |
1283 | (isa<llvm::IntegerType>(Val: SrcTy) || isa<llvm::PointerType>(Val: SrcTy))) { |
1284 | llvm::Value *Load = CGF.Builder.CreateLoad(Addr: Src); |
1285 | return CoerceIntOrPtrToIntOrPtr(Val: Load, Ty, CGF); |
1286 | } |
1287 | |
1288 | // If load is legal, just bitcast the src pointer. |
1289 | if (!SrcSize.isScalable() && !DstSize.isScalable() && |
1290 | SrcSize.getFixedValue() >= DstSize.getFixedValue()) { |
1291 | // Generally SrcSize is never greater than DstSize, since this means we are |
1292 | // losing bits. However, this can happen in cases where the structure has |
1293 | // additional padding, for example due to a user specified alignment. |
1294 | // |
1295 | // FIXME: Assert that we aren't truncating non-padding bits when have access |
1296 | // to that information. |
1297 | Src = Src.withElementType(ElemTy: Ty); |
1298 | return CGF.Builder.CreateLoad(Addr: Src); |
1299 | } |
1300 | |
1301 | // If coercing a fixed vector to a scalable vector for ABI compatibility, and |
1302 | // the types match, use the llvm.vector.insert intrinsic to perform the |
1303 | // conversion. |
1304 | if (auto *ScalableDstTy = dyn_cast<llvm::ScalableVectorType>(Val: Ty)) { |
1305 | if (auto *FixedSrcTy = dyn_cast<llvm::FixedVectorType>(Val: SrcTy)) { |
1306 | // If we are casting a fixed i8 vector to a scalable i1 predicate |
1307 | // vector, use a vector insert and bitcast the result. |
1308 | if (ScalableDstTy->getElementType()->isIntegerTy(Bitwidth: 1) && |
1309 | ScalableDstTy->getElementCount().isKnownMultipleOf(RHS: 8) && |
1310 | FixedSrcTy->getElementType()->isIntegerTy(Bitwidth: 8)) { |
1311 | ScalableDstTy = llvm::ScalableVectorType::get( |
1312 | ElementType: FixedSrcTy->getElementType(), |
1313 | MinNumElts: ScalableDstTy->getElementCount().getKnownMinValue() / 8); |
1314 | } |
1315 | if (ScalableDstTy->getElementType() == FixedSrcTy->getElementType()) { |
1316 | auto *Load = CGF.Builder.CreateLoad(Addr: Src); |
1317 | auto *UndefVec = llvm::UndefValue::get(T: ScalableDstTy); |
1318 | auto *Zero = llvm::Constant::getNullValue(Ty: CGF.CGM.Int64Ty); |
1319 | llvm::Value *Result = CGF.Builder.CreateInsertVector( |
1320 | DstType: ScalableDstTy, SrcVec: UndefVec, SubVec: Load, Idx: Zero, Name: "cast.scalable" ); |
1321 | if (ScalableDstTy != Ty) |
1322 | Result = CGF.Builder.CreateBitCast(V: Result, DestTy: Ty); |
1323 | return Result; |
1324 | } |
1325 | } |
1326 | } |
1327 | |
1328 | // Otherwise do coercion through memory. This is stupid, but simple. |
1329 | Address Tmp = |
1330 | CreateTempAllocaForCoercion(CGF, Ty, MinAlign: Src.getAlignment(), Name: Src.getName()); |
1331 | CGF.Builder.CreateMemCpy( |
1332 | Dst: Tmp.getPointer(), DstAlign: Tmp.getAlignment().getAsAlign(), Src: Src.getPointer(), |
1333 | SrcAlign: Src.getAlignment().getAsAlign(), |
1334 | Size: llvm::ConstantInt::get(Ty: CGF.IntPtrTy, V: SrcSize.getKnownMinValue())); |
1335 | return CGF.Builder.CreateLoad(Addr: Tmp); |
1336 | } |
1337 | |
1338 | // Function to store a first-class aggregate into memory. We prefer to |
1339 | // store the elements rather than the aggregate to be more friendly to |
1340 | // fast-isel. |
1341 | // FIXME: Do we need to recurse here? |
1342 | void CodeGenFunction::EmitAggregateStore(llvm::Value *Val, Address Dest, |
1343 | bool DestIsVolatile) { |
1344 | // Prefer scalar stores to first-class aggregate stores. |
1345 | if (llvm::StructType *STy = dyn_cast<llvm::StructType>(Val: Val->getType())) { |
1346 | for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { |
1347 | Address EltPtr = Builder.CreateStructGEP(Addr: Dest, Index: i); |
1348 | llvm::Value *Elt = Builder.CreateExtractValue(Agg: Val, Idxs: i); |
1349 | Builder.CreateStore(Val: Elt, Addr: EltPtr, IsVolatile: DestIsVolatile); |
1350 | } |
1351 | } else { |
1352 | Builder.CreateStore(Val, Addr: Dest, IsVolatile: DestIsVolatile); |
1353 | } |
1354 | } |
1355 | |
1356 | /// CreateCoercedStore - Create a store to \arg DstPtr from \arg Src, |
1357 | /// where the source and destination may have different types. The |
1358 | /// destination is known to be aligned to \arg DstAlign bytes. |
1359 | /// |
1360 | /// This safely handles the case when the src type is larger than the |
1361 | /// destination type; the upper bits of the src will be lost. |
1362 | static void CreateCoercedStore(llvm::Value *Src, |
1363 | Address Dst, |
1364 | bool DstIsVolatile, |
1365 | CodeGenFunction &CGF) { |
1366 | llvm::Type *SrcTy = Src->getType(); |
1367 | llvm::Type *DstTy = Dst.getElementType(); |
1368 | if (SrcTy == DstTy) { |
1369 | CGF.Builder.CreateStore(Val: Src, Addr: Dst, IsVolatile: DstIsVolatile); |
1370 | return; |
1371 | } |
1372 | |
1373 | llvm::TypeSize SrcSize = CGF.CGM.getDataLayout().getTypeAllocSize(Ty: SrcTy); |
1374 | |
1375 | if (llvm::StructType *DstSTy = dyn_cast<llvm::StructType>(Val: DstTy)) { |
1376 | Dst = EnterStructPointerForCoercedAccess(SrcPtr: Dst, SrcSTy: DstSTy, |
1377 | DstSize: SrcSize.getFixedValue(), CGF); |
1378 | DstTy = Dst.getElementType(); |
1379 | } |
1380 | |
1381 | llvm::PointerType *SrcPtrTy = llvm::dyn_cast<llvm::PointerType>(Val: SrcTy); |
1382 | llvm::PointerType *DstPtrTy = llvm::dyn_cast<llvm::PointerType>(Val: DstTy); |
1383 | if (SrcPtrTy && DstPtrTy && |
1384 | SrcPtrTy->getAddressSpace() != DstPtrTy->getAddressSpace()) { |
1385 | Src = CGF.Builder.CreateAddrSpaceCast(V: Src, DestTy: DstTy); |
1386 | CGF.Builder.CreateStore(Val: Src, Addr: Dst, IsVolatile: DstIsVolatile); |
1387 | return; |
1388 | } |
1389 | |
1390 | // If the source and destination are integer or pointer types, just do an |
1391 | // extension or truncation to the desired type. |
1392 | if ((isa<llvm::IntegerType>(Val: SrcTy) || isa<llvm::PointerType>(Val: SrcTy)) && |
1393 | (isa<llvm::IntegerType>(Val: DstTy) || isa<llvm::PointerType>(Val: DstTy))) { |
1394 | Src = CoerceIntOrPtrToIntOrPtr(Val: Src, Ty: DstTy, CGF); |
1395 | CGF.Builder.CreateStore(Val: Src, Addr: Dst, IsVolatile: DstIsVolatile); |
1396 | return; |
1397 | } |
1398 | |
1399 | llvm::TypeSize DstSize = CGF.CGM.getDataLayout().getTypeAllocSize(Ty: DstTy); |
1400 | |
1401 | // If store is legal, just bitcast the src pointer. |
1402 | if (isa<llvm::ScalableVectorType>(Val: SrcTy) || |
1403 | isa<llvm::ScalableVectorType>(Val: DstTy) || |
1404 | SrcSize.getFixedValue() <= DstSize.getFixedValue()) { |
1405 | Dst = Dst.withElementType(ElemTy: SrcTy); |
1406 | CGF.EmitAggregateStore(Val: Src, Dest: Dst, DestIsVolatile: DstIsVolatile); |
1407 | } else { |
1408 | // Otherwise do coercion through memory. This is stupid, but |
1409 | // simple. |
1410 | |
1411 | // Generally SrcSize is never greater than DstSize, since this means we are |
1412 | // losing bits. However, this can happen in cases where the structure has |
1413 | // additional padding, for example due to a user specified alignment. |
1414 | // |
1415 | // FIXME: Assert that we aren't truncating non-padding bits when have access |
1416 | // to that information. |
1417 | Address Tmp = CreateTempAllocaForCoercion(CGF, Ty: SrcTy, MinAlign: Dst.getAlignment()); |
1418 | CGF.Builder.CreateStore(Val: Src, Addr: Tmp); |
1419 | CGF.Builder.CreateMemCpy( |
1420 | Dst: Dst.getPointer(), DstAlign: Dst.getAlignment().getAsAlign(), Src: Tmp.getPointer(), |
1421 | SrcAlign: Tmp.getAlignment().getAsAlign(), |
1422 | Size: llvm::ConstantInt::get(Ty: CGF.IntPtrTy, V: DstSize.getFixedValue())); |
1423 | } |
1424 | } |
1425 | |
1426 | static Address emitAddressAtOffset(CodeGenFunction &CGF, Address addr, |
1427 | const ABIArgInfo &info) { |
1428 | if (unsigned offset = info.getDirectOffset()) { |
1429 | addr = addr.withElementType(ElemTy: CGF.Int8Ty); |
1430 | addr = CGF.Builder.CreateConstInBoundsByteGEP(Addr: addr, |
1431 | Offset: CharUnits::fromQuantity(Quantity: offset)); |
1432 | addr = addr.withElementType(ElemTy: info.getCoerceToType()); |
1433 | } |
1434 | return addr; |
1435 | } |
1436 | |
1437 | namespace { |
1438 | |
1439 | /// Encapsulates information about the way function arguments from |
1440 | /// CGFunctionInfo should be passed to actual LLVM IR function. |
1441 | class ClangToLLVMArgMapping { |
1442 | static const unsigned InvalidIndex = ~0U; |
1443 | unsigned InallocaArgNo; |
1444 | unsigned SRetArgNo; |
1445 | unsigned TotalIRArgs; |
1446 | |
1447 | /// Arguments of LLVM IR function corresponding to single Clang argument. |
1448 | struct IRArgs { |
1449 | unsigned PaddingArgIndex; |
1450 | // Argument is expanded to IR arguments at positions |
1451 | // [FirstArgIndex, FirstArgIndex + NumberOfArgs). |
1452 | unsigned FirstArgIndex; |
1453 | unsigned NumberOfArgs; |
1454 | |
1455 | IRArgs() |
1456 | : PaddingArgIndex(InvalidIndex), FirstArgIndex(InvalidIndex), |
1457 | NumberOfArgs(0) {} |
1458 | }; |
1459 | |
1460 | SmallVector<IRArgs, 8> ArgInfo; |
1461 | |
1462 | public: |
1463 | ClangToLLVMArgMapping(const ASTContext &Context, const CGFunctionInfo &FI, |
1464 | bool OnlyRequiredArgs = false) |
1465 | : InallocaArgNo(InvalidIndex), SRetArgNo(InvalidIndex), TotalIRArgs(0), |
1466 | ArgInfo(OnlyRequiredArgs ? FI.getNumRequiredArgs() : FI.arg_size()) { |
1467 | construct(Context, FI, OnlyRequiredArgs); |
1468 | } |
1469 | |
1470 | bool hasInallocaArg() const { return InallocaArgNo != InvalidIndex; } |
1471 | unsigned getInallocaArgNo() const { |
1472 | assert(hasInallocaArg()); |
1473 | return InallocaArgNo; |
1474 | } |
1475 | |
1476 | bool hasSRetArg() const { return SRetArgNo != InvalidIndex; } |
1477 | unsigned getSRetArgNo() const { |
1478 | assert(hasSRetArg()); |
1479 | return SRetArgNo; |
1480 | } |
1481 | |
1482 | unsigned totalIRArgs() const { return TotalIRArgs; } |
1483 | |
1484 | bool hasPaddingArg(unsigned ArgNo) const { |
1485 | assert(ArgNo < ArgInfo.size()); |
1486 | return ArgInfo[ArgNo].PaddingArgIndex != InvalidIndex; |
1487 | } |
1488 | unsigned getPaddingArgNo(unsigned ArgNo) const { |
1489 | assert(hasPaddingArg(ArgNo)); |
1490 | return ArgInfo[ArgNo].PaddingArgIndex; |
1491 | } |
1492 | |
1493 | /// Returns index of first IR argument corresponding to ArgNo, and their |
1494 | /// quantity. |
1495 | std::pair<unsigned, unsigned> getIRArgs(unsigned ArgNo) const { |
1496 | assert(ArgNo < ArgInfo.size()); |
1497 | return std::make_pair(x: ArgInfo[ArgNo].FirstArgIndex, |
1498 | y: ArgInfo[ArgNo].NumberOfArgs); |
1499 | } |
1500 | |
1501 | private: |
1502 | void construct(const ASTContext &Context, const CGFunctionInfo &FI, |
1503 | bool OnlyRequiredArgs); |
1504 | }; |
1505 | |
1506 | void ClangToLLVMArgMapping::construct(const ASTContext &Context, |
1507 | const CGFunctionInfo &FI, |
1508 | bool OnlyRequiredArgs) { |
1509 | unsigned IRArgNo = 0; |
1510 | bool SwapThisWithSRet = false; |
1511 | const ABIArgInfo &RetAI = FI.getReturnInfo(); |
1512 | |
1513 | if (RetAI.getKind() == ABIArgInfo::Indirect) { |
1514 | SwapThisWithSRet = RetAI.isSRetAfterThis(); |
1515 | SRetArgNo = SwapThisWithSRet ? 1 : IRArgNo++; |
1516 | } |
1517 | |
1518 | unsigned ArgNo = 0; |
1519 | unsigned NumArgs = OnlyRequiredArgs ? FI.getNumRequiredArgs() : FI.arg_size(); |
1520 | for (CGFunctionInfo::const_arg_iterator I = FI.arg_begin(); ArgNo < NumArgs; |
1521 | ++I, ++ArgNo) { |
1522 | assert(I != FI.arg_end()); |
1523 | QualType ArgType = I->type; |
1524 | const ABIArgInfo &AI = I->info; |
1525 | // Collect data about IR arguments corresponding to Clang argument ArgNo. |
1526 | auto &IRArgs = ArgInfo[ArgNo]; |
1527 | |
1528 | if (AI.getPaddingType()) |
1529 | IRArgs.PaddingArgIndex = IRArgNo++; |
1530 | |
1531 | switch (AI.getKind()) { |
1532 | case ABIArgInfo::Extend: |
1533 | case ABIArgInfo::Direct: { |
1534 | // FIXME: handle sseregparm someday... |
1535 | llvm::StructType *STy = dyn_cast<llvm::StructType>(Val: AI.getCoerceToType()); |
1536 | if (AI.isDirect() && AI.getCanBeFlattened() && STy) { |
1537 | IRArgs.NumberOfArgs = STy->getNumElements(); |
1538 | } else { |
1539 | IRArgs.NumberOfArgs = 1; |
1540 | } |
1541 | break; |
1542 | } |
1543 | case ABIArgInfo::Indirect: |
1544 | case ABIArgInfo::IndirectAliased: |
1545 | IRArgs.NumberOfArgs = 1; |
1546 | break; |
1547 | case ABIArgInfo::Ignore: |
1548 | case ABIArgInfo::InAlloca: |
1549 | // ignore and inalloca doesn't have matching LLVM parameters. |
1550 | IRArgs.NumberOfArgs = 0; |
1551 | break; |
1552 | case ABIArgInfo::CoerceAndExpand: |
1553 | IRArgs.NumberOfArgs = AI.getCoerceAndExpandTypeSequence().size(); |
1554 | break; |
1555 | case ABIArgInfo::Expand: |
1556 | IRArgs.NumberOfArgs = getExpansionSize(Ty: ArgType, Context); |
1557 | break; |
1558 | } |
1559 | |
1560 | if (IRArgs.NumberOfArgs > 0) { |
1561 | IRArgs.FirstArgIndex = IRArgNo; |
1562 | IRArgNo += IRArgs.NumberOfArgs; |
1563 | } |
1564 | |
1565 | // Skip over the sret parameter when it comes second. We already handled it |
1566 | // above. |
1567 | if (IRArgNo == 1 && SwapThisWithSRet) |
1568 | IRArgNo++; |
1569 | } |
1570 | assert(ArgNo == ArgInfo.size()); |
1571 | |
1572 | if (FI.usesInAlloca()) |
1573 | InallocaArgNo = IRArgNo++; |
1574 | |
1575 | TotalIRArgs = IRArgNo; |
1576 | } |
1577 | } // namespace |
1578 | |
1579 | /***/ |
1580 | |
1581 | bool CodeGenModule::ReturnTypeUsesSRet(const CGFunctionInfo &FI) { |
1582 | const auto &RI = FI.getReturnInfo(); |
1583 | return RI.isIndirect() || (RI.isInAlloca() && RI.getInAllocaSRet()); |
1584 | } |
1585 | |
1586 | bool CodeGenModule::ReturnSlotInterferesWithArgs(const CGFunctionInfo &FI) { |
1587 | return ReturnTypeUsesSRet(FI) && |
1588 | getTargetCodeGenInfo().doesReturnSlotInterfereWithArgs(); |
1589 | } |
1590 | |
1591 | bool CodeGenModule::ReturnTypeUsesFPRet(QualType ResultType) { |
1592 | if (const BuiltinType *BT = ResultType->getAs<BuiltinType>()) { |
1593 | switch (BT->getKind()) { |
1594 | default: |
1595 | return false; |
1596 | case BuiltinType::Float: |
1597 | return getTarget().useObjCFPRetForRealType(T: FloatModeKind::Float); |
1598 | case BuiltinType::Double: |
1599 | return getTarget().useObjCFPRetForRealType(T: FloatModeKind::Double); |
1600 | case BuiltinType::LongDouble: |
1601 | return getTarget().useObjCFPRetForRealType(T: FloatModeKind::LongDouble); |
1602 | } |
1603 | } |
1604 | |
1605 | return false; |
1606 | } |
1607 | |
1608 | bool CodeGenModule::ReturnTypeUsesFP2Ret(QualType ResultType) { |
1609 | if (const ComplexType *CT = ResultType->getAs<ComplexType>()) { |
1610 | if (const BuiltinType *BT = CT->getElementType()->getAs<BuiltinType>()) { |
1611 | if (BT->getKind() == BuiltinType::LongDouble) |
1612 | return getTarget().useObjCFP2RetForComplexLongDouble(); |
1613 | } |
1614 | } |
1615 | |
1616 | return false; |
1617 | } |
1618 | |
1619 | llvm::FunctionType *CodeGenTypes::GetFunctionType(GlobalDecl GD) { |
1620 | const CGFunctionInfo &FI = arrangeGlobalDeclaration(GD); |
1621 | return GetFunctionType(Info: FI); |
1622 | } |
1623 | |
1624 | llvm::FunctionType * |
1625 | CodeGenTypes::GetFunctionType(const CGFunctionInfo &FI) { |
1626 | |
1627 | bool Inserted = FunctionsBeingProcessed.insert(Ptr: &FI).second; |
1628 | (void)Inserted; |
1629 | assert(Inserted && "Recursively being processed?" ); |
1630 | |
1631 | llvm::Type *resultType = nullptr; |
1632 | const ABIArgInfo &retAI = FI.getReturnInfo(); |
1633 | switch (retAI.getKind()) { |
1634 | case ABIArgInfo::Expand: |
1635 | case ABIArgInfo::IndirectAliased: |
1636 | llvm_unreachable("Invalid ABI kind for return argument" ); |
1637 | |
1638 | case ABIArgInfo::Extend: |
1639 | case ABIArgInfo::Direct: |
1640 | resultType = retAI.getCoerceToType(); |
1641 | break; |
1642 | |
1643 | case ABIArgInfo::InAlloca: |
1644 | if (retAI.getInAllocaSRet()) { |
1645 | // sret things on win32 aren't void, they return the sret pointer. |
1646 | QualType ret = FI.getReturnType(); |
1647 | unsigned addressSpace = CGM.getTypes().getTargetAddressSpace(T: ret); |
1648 | resultType = llvm::PointerType::get(C&: getLLVMContext(), AddressSpace: addressSpace); |
1649 | } else { |
1650 | resultType = llvm::Type::getVoidTy(C&: getLLVMContext()); |
1651 | } |
1652 | break; |
1653 | |
1654 | case ABIArgInfo::Indirect: |
1655 | case ABIArgInfo::Ignore: |
1656 | resultType = llvm::Type::getVoidTy(C&: getLLVMContext()); |
1657 | break; |
1658 | |
1659 | case ABIArgInfo::CoerceAndExpand: |
1660 | resultType = retAI.getUnpaddedCoerceAndExpandType(); |
1661 | break; |
1662 | } |
1663 | |
1664 | ClangToLLVMArgMapping IRFunctionArgs(getContext(), FI, true); |
1665 | SmallVector<llvm::Type*, 8> ArgTypes(IRFunctionArgs.totalIRArgs()); |
1666 | |
1667 | // Add type for sret argument. |
1668 | if (IRFunctionArgs.hasSRetArg()) { |
1669 | QualType Ret = FI.getReturnType(); |
1670 | unsigned AddressSpace = CGM.getTypes().getTargetAddressSpace(T: Ret); |
1671 | ArgTypes[IRFunctionArgs.getSRetArgNo()] = |
1672 | llvm::PointerType::get(C&: getLLVMContext(), AddressSpace); |
1673 | } |
1674 | |
1675 | // Add type for inalloca argument. |
1676 | if (IRFunctionArgs.hasInallocaArg()) |
1677 | ArgTypes[IRFunctionArgs.getInallocaArgNo()] = |
1678 | llvm::PointerType::getUnqual(C&: getLLVMContext()); |
1679 | |
1680 | // Add in all of the required arguments. |
1681 | unsigned ArgNo = 0; |
1682 | CGFunctionInfo::const_arg_iterator it = FI.arg_begin(), |
1683 | ie = it + FI.getNumRequiredArgs(); |
1684 | for (; it != ie; ++it, ++ArgNo) { |
1685 | const ABIArgInfo &ArgInfo = it->info; |
1686 | |
1687 | // Insert a padding type to ensure proper alignment. |
1688 | if (IRFunctionArgs.hasPaddingArg(ArgNo)) |
1689 | ArgTypes[IRFunctionArgs.getPaddingArgNo(ArgNo)] = |
1690 | ArgInfo.getPaddingType(); |
1691 | |
1692 | unsigned FirstIRArg, NumIRArgs; |
1693 | std::tie(args&: FirstIRArg, args&: NumIRArgs) = IRFunctionArgs.getIRArgs(ArgNo); |
1694 | |
1695 | switch (ArgInfo.getKind()) { |
1696 | case ABIArgInfo::Ignore: |
1697 | case ABIArgInfo::InAlloca: |
1698 | assert(NumIRArgs == 0); |
1699 | break; |
1700 | |
1701 | case ABIArgInfo::Indirect: |
1702 | assert(NumIRArgs == 1); |
1703 | // indirect arguments are always on the stack, which is alloca addr space. |
1704 | ArgTypes[FirstIRArg] = llvm::PointerType::get( |
1705 | C&: getLLVMContext(), AddressSpace: CGM.getDataLayout().getAllocaAddrSpace()); |
1706 | break; |
1707 | case ABIArgInfo::IndirectAliased: |
1708 | assert(NumIRArgs == 1); |
1709 | ArgTypes[FirstIRArg] = llvm::PointerType::get( |
1710 | C&: getLLVMContext(), AddressSpace: ArgInfo.getIndirectAddrSpace()); |
1711 | break; |
1712 | case ABIArgInfo::Extend: |
1713 | case ABIArgInfo::Direct: { |
1714 | // Fast-isel and the optimizer generally like scalar values better than |
1715 | // FCAs, so we flatten them if this is safe to do for this argument. |
1716 | llvm::Type *argType = ArgInfo.getCoerceToType(); |
1717 | llvm::StructType *st = dyn_cast<llvm::StructType>(Val: argType); |
1718 | if (st && ArgInfo.isDirect() && ArgInfo.getCanBeFlattened()) { |
1719 | assert(NumIRArgs == st->getNumElements()); |
1720 | for (unsigned i = 0, e = st->getNumElements(); i != e; ++i) |
1721 | ArgTypes[FirstIRArg + i] = st->getElementType(N: i); |
1722 | } else { |
1723 | assert(NumIRArgs == 1); |
1724 | ArgTypes[FirstIRArg] = argType; |
1725 | } |
1726 | break; |
1727 | } |
1728 | |
1729 | case ABIArgInfo::CoerceAndExpand: { |
1730 | auto ArgTypesIter = ArgTypes.begin() + FirstIRArg; |
1731 | for (auto *EltTy : ArgInfo.getCoerceAndExpandTypeSequence()) { |
1732 | *ArgTypesIter++ = EltTy; |
1733 | } |
1734 | assert(ArgTypesIter == ArgTypes.begin() + FirstIRArg + NumIRArgs); |
1735 | break; |
1736 | } |
1737 | |
1738 | case ABIArgInfo::Expand: |
1739 | auto ArgTypesIter = ArgTypes.begin() + FirstIRArg; |
1740 | getExpandedTypes(Ty: it->type, TI&: ArgTypesIter); |
1741 | assert(ArgTypesIter == ArgTypes.begin() + FirstIRArg + NumIRArgs); |
1742 | break; |
1743 | } |
1744 | } |
1745 | |
1746 | bool Erased = FunctionsBeingProcessed.erase(Ptr: &FI); (void)Erased; |
1747 | assert(Erased && "Not in set?" ); |
1748 | |
1749 | return llvm::FunctionType::get(Result: resultType, Params: ArgTypes, isVarArg: FI.isVariadic()); |
1750 | } |
1751 | |
1752 | llvm::Type *CodeGenTypes::GetFunctionTypeForVTable(GlobalDecl GD) { |
1753 | const CXXMethodDecl *MD = cast<CXXMethodDecl>(Val: GD.getDecl()); |
1754 | const FunctionProtoType *FPT = MD->getType()->castAs<FunctionProtoType>(); |
1755 | |
1756 | if (!isFuncTypeConvertible(FPT)) |
1757 | return llvm::StructType::get(Context&: getLLVMContext()); |
1758 | |
1759 | return GetFunctionType(GD); |
1760 | } |
1761 | |
1762 | static void AddAttributesFromFunctionProtoType(ASTContext &Ctx, |
1763 | llvm::AttrBuilder &FuncAttrs, |
1764 | const FunctionProtoType *FPT) { |
1765 | if (!FPT) |
1766 | return; |
1767 | |
1768 | if (!isUnresolvedExceptionSpec(FPT->getExceptionSpecType()) && |
1769 | FPT->isNothrow()) |
1770 | FuncAttrs.addAttribute(llvm::Attribute::NoUnwind); |
1771 | |
1772 | unsigned SMEBits = FPT->getAArch64SMEAttributes(); |
1773 | if (SMEBits & FunctionType::SME_PStateSMEnabledMask) |
1774 | FuncAttrs.addAttribute(A: "aarch64_pstate_sm_enabled" ); |
1775 | if (SMEBits & FunctionType::SME_PStateSMCompatibleMask) |
1776 | FuncAttrs.addAttribute(A: "aarch64_pstate_sm_compatible" ); |
1777 | |
1778 | // ZA |
1779 | if (FunctionType::getArmZAState(AttrBits: SMEBits) == FunctionType::ARM_Preserves) |
1780 | FuncAttrs.addAttribute(A: "aarch64_preserves_za" ); |
1781 | if (FunctionType::getArmZAState(AttrBits: SMEBits) == FunctionType::ARM_In) |
1782 | FuncAttrs.addAttribute(A: "aarch64_in_za" ); |
1783 | if (FunctionType::getArmZAState(AttrBits: SMEBits) == FunctionType::ARM_Out) |
1784 | FuncAttrs.addAttribute(A: "aarch64_out_za" ); |
1785 | if (FunctionType::getArmZAState(AttrBits: SMEBits) == FunctionType::ARM_InOut) |
1786 | FuncAttrs.addAttribute(A: "aarch64_inout_za" ); |
1787 | |
1788 | // ZT0 |
1789 | if (FunctionType::getArmZT0State(AttrBits: SMEBits) == FunctionType::ARM_Preserves) |
1790 | FuncAttrs.addAttribute(A: "aarch64_preserves_zt0" ); |
1791 | if (FunctionType::getArmZT0State(AttrBits: SMEBits) == FunctionType::ARM_In) |
1792 | FuncAttrs.addAttribute(A: "aarch64_in_zt0" ); |
1793 | if (FunctionType::getArmZT0State(AttrBits: SMEBits) == FunctionType::ARM_Out) |
1794 | FuncAttrs.addAttribute(A: "aarch64_out_zt0" ); |
1795 | if (FunctionType::getArmZT0State(AttrBits: SMEBits) == FunctionType::ARM_InOut) |
1796 | FuncAttrs.addAttribute(A: "aarch64_inout_zt0" ); |
1797 | } |
1798 | |
1799 | static void AddAttributesFromAssumes(llvm::AttrBuilder &FuncAttrs, |
1800 | const Decl *Callee) { |
1801 | if (!Callee) |
1802 | return; |
1803 | |
1804 | SmallVector<StringRef, 4> Attrs; |
1805 | |
1806 | for (const AssumptionAttr *AA : Callee->specific_attrs<AssumptionAttr>()) |
1807 | AA->getAssumption().split(Attrs, "," ); |
1808 | |
1809 | if (!Attrs.empty()) |
1810 | FuncAttrs.addAttribute(A: llvm::AssumptionAttrKey, |
1811 | V: llvm::join(Begin: Attrs.begin(), End: Attrs.end(), Separator: "," )); |
1812 | } |
1813 | |
1814 | bool CodeGenModule::MayDropFunctionReturn(const ASTContext &Context, |
1815 | QualType ReturnType) const { |
1816 | // We can't just discard the return value for a record type with a |
1817 | // complex destructor or a non-trivially copyable type. |
1818 | if (const RecordType *RT = |
1819 | ReturnType.getCanonicalType()->getAs<RecordType>()) { |
1820 | if (const auto *ClassDecl = dyn_cast<CXXRecordDecl>(Val: RT->getDecl())) |
1821 | return ClassDecl->hasTrivialDestructor(); |
1822 | } |
1823 | return ReturnType.isTriviallyCopyableType(Context); |
1824 | } |
1825 | |
1826 | static bool HasStrictReturn(const CodeGenModule &Module, QualType RetTy, |
1827 | const Decl *TargetDecl) { |
1828 | // As-is msan can not tolerate noundef mismatch between caller and |
1829 | // implementation. Mismatch is possible for e.g. indirect calls from C-caller |
1830 | // into C++. Such mismatches lead to confusing false reports. To avoid |
1831 | // expensive workaround on msan we enforce initialization event in uncommon |
1832 | // cases where it's allowed. |
1833 | if (Module.getLangOpts().Sanitize.has(K: SanitizerKind::Memory)) |
1834 | return true; |
1835 | // C++ explicitly makes returning undefined values UB. C's rule only applies |
1836 | // to used values, so we never mark them noundef for now. |
1837 | if (!Module.getLangOpts().CPlusPlus) |
1838 | return false; |
1839 | if (TargetDecl) { |
1840 | if (const FunctionDecl *FDecl = dyn_cast<FunctionDecl>(Val: TargetDecl)) { |
1841 | if (FDecl->isExternC()) |
1842 | return false; |
1843 | } else if (const VarDecl *VDecl = dyn_cast<VarDecl>(Val: TargetDecl)) { |
1844 | // Function pointer. |
1845 | if (VDecl->isExternC()) |
1846 | return false; |
1847 | } |
1848 | } |
1849 | |
1850 | // We don't want to be too aggressive with the return checking, unless |
1851 | // it's explicit in the code opts or we're using an appropriate sanitizer. |
1852 | // Try to respect what the programmer intended. |
1853 | return Module.getCodeGenOpts().StrictReturn || |
1854 | !Module.MayDropFunctionReturn(Context: Module.getContext(), ReturnType: RetTy) || |
1855 | Module.getLangOpts().Sanitize.has(K: SanitizerKind::Return); |
1856 | } |
1857 | |
1858 | /// Add denormal-fp-math and denormal-fp-math-f32 as appropriate for the |
1859 | /// requested denormal behavior, accounting for the overriding behavior of the |
1860 | /// -f32 case. |
1861 | static void addDenormalModeAttrs(llvm::DenormalMode FPDenormalMode, |
1862 | llvm::DenormalMode FP32DenormalMode, |
1863 | llvm::AttrBuilder &FuncAttrs) { |
1864 | if (FPDenormalMode != llvm::DenormalMode::getDefault()) |
1865 | FuncAttrs.addAttribute(A: "denormal-fp-math" , V: FPDenormalMode.str()); |
1866 | |
1867 | if (FP32DenormalMode != FPDenormalMode && FP32DenormalMode.isValid()) |
1868 | FuncAttrs.addAttribute(A: "denormal-fp-math-f32" , V: FP32DenormalMode.str()); |
1869 | } |
1870 | |
1871 | /// Add default attributes to a function, which have merge semantics under |
1872 | /// -mlink-builtin-bitcode and should not simply overwrite any existing |
1873 | /// attributes in the linked library. |
1874 | static void |
1875 | addMergableDefaultFunctionAttributes(const CodeGenOptions &CodeGenOpts, |
1876 | llvm::AttrBuilder &FuncAttrs) { |
1877 | addDenormalModeAttrs(FPDenormalMode: CodeGenOpts.FPDenormalMode, FP32DenormalMode: CodeGenOpts.FP32DenormalMode, |
1878 | FuncAttrs); |
1879 | } |
1880 | |
1881 | static void getTrivialDefaultFunctionAttributes( |
1882 | StringRef Name, bool HasOptnone, const CodeGenOptions &CodeGenOpts, |
1883 | const LangOptions &LangOpts, bool AttrOnCallSite, |
1884 | llvm::AttrBuilder &FuncAttrs) { |
1885 | // OptimizeNoneAttr takes precedence over -Os or -Oz. No warning needed. |
1886 | if (!HasOptnone) { |
1887 | if (CodeGenOpts.OptimizeSize) |
1888 | FuncAttrs.addAttribute(llvm::Attribute::OptimizeForSize); |
1889 | if (CodeGenOpts.OptimizeSize == 2) |
1890 | FuncAttrs.addAttribute(llvm::Attribute::MinSize); |
1891 | } |
1892 | |
1893 | if (CodeGenOpts.DisableRedZone) |
1894 | FuncAttrs.addAttribute(llvm::Attribute::NoRedZone); |
1895 | if (CodeGenOpts.IndirectTlsSegRefs) |
1896 | FuncAttrs.addAttribute(A: "indirect-tls-seg-refs" ); |
1897 | if (CodeGenOpts.NoImplicitFloat) |
1898 | FuncAttrs.addAttribute(llvm::Attribute::NoImplicitFloat); |
1899 | |
1900 | if (AttrOnCallSite) { |
1901 | // Attributes that should go on the call site only. |
1902 | // FIXME: Look for 'BuiltinAttr' on the function rather than re-checking |
1903 | // the -fno-builtin-foo list. |
1904 | if (!CodeGenOpts.SimplifyLibCalls || LangOpts.isNoBuiltinFunc(Name)) |
1905 | FuncAttrs.addAttribute(llvm::Attribute::NoBuiltin); |
1906 | if (!CodeGenOpts.TrapFuncName.empty()) |
1907 | FuncAttrs.addAttribute(A: "trap-func-name" , V: CodeGenOpts.TrapFuncName); |
1908 | } else { |
1909 | switch (CodeGenOpts.getFramePointer()) { |
1910 | case CodeGenOptions::FramePointerKind::None: |
1911 | // This is the default behavior. |
1912 | break; |
1913 | case CodeGenOptions::FramePointerKind::NonLeaf: |
1914 | case CodeGenOptions::FramePointerKind::All: |
1915 | FuncAttrs.addAttribute(A: "frame-pointer" , |
1916 | V: CodeGenOptions::getFramePointerKindName( |
1917 | Kind: CodeGenOpts.getFramePointer())); |
1918 | } |
1919 | |
1920 | if (CodeGenOpts.LessPreciseFPMAD) |
1921 | FuncAttrs.addAttribute(A: "less-precise-fpmad" , V: "true" ); |
1922 | |
1923 | if (CodeGenOpts.NullPointerIsValid) |
1924 | FuncAttrs.addAttribute(llvm::Attribute::NullPointerIsValid); |
1925 | |
1926 | if (LangOpts.getDefaultExceptionMode() == LangOptions::FPE_Ignore) |
1927 | FuncAttrs.addAttribute(A: "no-trapping-math" , V: "true" ); |
1928 | |
1929 | // TODO: Are these all needed? |
1930 | // unsafe/inf/nan/nsz are handled by instruction-level FastMathFlags. |
1931 | if (LangOpts.NoHonorInfs) |
1932 | FuncAttrs.addAttribute(A: "no-infs-fp-math" , V: "true" ); |
1933 | if (LangOpts.NoHonorNaNs) |
1934 | FuncAttrs.addAttribute(A: "no-nans-fp-math" , V: "true" ); |
1935 | if (LangOpts.ApproxFunc) |
1936 | FuncAttrs.addAttribute(A: "approx-func-fp-math" , V: "true" ); |
1937 | if (LangOpts.AllowFPReassoc && LangOpts.AllowRecip && |
1938 | LangOpts.NoSignedZero && LangOpts.ApproxFunc && |
1939 | (LangOpts.getDefaultFPContractMode() == |
1940 | LangOptions::FPModeKind::FPM_Fast || |
1941 | LangOpts.getDefaultFPContractMode() == |
1942 | LangOptions::FPModeKind::FPM_FastHonorPragmas)) |
1943 | FuncAttrs.addAttribute(A: "unsafe-fp-math" , V: "true" ); |
1944 | if (CodeGenOpts.SoftFloat) |
1945 | FuncAttrs.addAttribute(A: "use-soft-float" , V: "true" ); |
1946 | FuncAttrs.addAttribute(A: "stack-protector-buffer-size" , |
1947 | V: llvm::utostr(X: CodeGenOpts.SSPBufferSize)); |
1948 | if (LangOpts.NoSignedZero) |
1949 | FuncAttrs.addAttribute(A: "no-signed-zeros-fp-math" , V: "true" ); |
1950 | |
1951 | // TODO: Reciprocal estimate codegen options should apply to instructions? |
1952 | const std::vector<std::string> &Recips = CodeGenOpts.Reciprocals; |
1953 | if (!Recips.empty()) |
1954 | FuncAttrs.addAttribute(A: "reciprocal-estimates" , |
1955 | V: llvm::join(R: Recips, Separator: "," )); |
1956 | |
1957 | if (!CodeGenOpts.PreferVectorWidth.empty() && |
1958 | CodeGenOpts.PreferVectorWidth != "none" ) |
1959 | FuncAttrs.addAttribute(A: "prefer-vector-width" , |
1960 | V: CodeGenOpts.PreferVectorWidth); |
1961 | |
1962 | if (CodeGenOpts.StackRealignment) |
1963 | FuncAttrs.addAttribute(A: "stackrealign" ); |
1964 | if (CodeGenOpts.Backchain) |
1965 | FuncAttrs.addAttribute(A: "backchain" ); |
1966 | if (CodeGenOpts.EnableSegmentedStacks) |
1967 | FuncAttrs.addAttribute(A: "split-stack" ); |
1968 | |
1969 | if (CodeGenOpts.SpeculativeLoadHardening) |
1970 | FuncAttrs.addAttribute(llvm::Attribute::SpeculativeLoadHardening); |
1971 | |
1972 | // Add zero-call-used-regs attribute. |
1973 | switch (CodeGenOpts.getZeroCallUsedRegs()) { |
1974 | case llvm::ZeroCallUsedRegs::ZeroCallUsedRegsKind::Skip: |
1975 | FuncAttrs.removeAttribute(A: "zero-call-used-regs" ); |
1976 | break; |
1977 | case llvm::ZeroCallUsedRegs::ZeroCallUsedRegsKind::UsedGPRArg: |
1978 | FuncAttrs.addAttribute(A: "zero-call-used-regs" , V: "used-gpr-arg" ); |
1979 | break; |
1980 | case llvm::ZeroCallUsedRegs::ZeroCallUsedRegsKind::UsedGPR: |
1981 | FuncAttrs.addAttribute(A: "zero-call-used-regs" , V: "used-gpr" ); |
1982 | break; |
1983 | case llvm::ZeroCallUsedRegs::ZeroCallUsedRegsKind::UsedArg: |
1984 | FuncAttrs.addAttribute(A: "zero-call-used-regs" , V: "used-arg" ); |
1985 | break; |
1986 | case llvm::ZeroCallUsedRegs::ZeroCallUsedRegsKind::Used: |
1987 | FuncAttrs.addAttribute(A: "zero-call-used-regs" , V: "used" ); |
1988 | break; |
1989 | case llvm::ZeroCallUsedRegs::ZeroCallUsedRegsKind::AllGPRArg: |
1990 | FuncAttrs.addAttribute(A: "zero-call-used-regs" , V: "all-gpr-arg" ); |
1991 | break; |
1992 | case llvm::ZeroCallUsedRegs::ZeroCallUsedRegsKind::AllGPR: |
1993 | FuncAttrs.addAttribute(A: "zero-call-used-regs" , V: "all-gpr" ); |
1994 | break; |
1995 | case llvm::ZeroCallUsedRegs::ZeroCallUsedRegsKind::AllArg: |
1996 | FuncAttrs.addAttribute(A: "zero-call-used-regs" , V: "all-arg" ); |
1997 | break; |
1998 | case llvm::ZeroCallUsedRegs::ZeroCallUsedRegsKind::All: |
1999 | FuncAttrs.addAttribute(A: "zero-call-used-regs" , V: "all" ); |
2000 | break; |
2001 | } |
2002 | } |
2003 | |
2004 | if (LangOpts.assumeFunctionsAreConvergent()) { |
2005 | // Conservatively, mark all functions and calls in CUDA and OpenCL as |
2006 | // convergent (meaning, they may call an intrinsically convergent op, such |
2007 | // as __syncthreads() / barrier(), and so can't have certain optimizations |
2008 | // applied around them). LLVM will remove this attribute where it safely |
2009 | // can. |
2010 | FuncAttrs.addAttribute(llvm::Attribute::Convergent); |
2011 | } |
2012 | |
2013 | // TODO: NoUnwind attribute should be added for other GPU modes HIP, |
2014 | // OpenMP offload. AFAIK, neither of them support exceptions in device code. |
2015 | if ((LangOpts.CUDA && LangOpts.CUDAIsDevice) || LangOpts.OpenCL || |
2016 | LangOpts.SYCLIsDevice) { |
2017 | FuncAttrs.addAttribute(llvm::Attribute::NoUnwind); |
2018 | } |
2019 | |
2020 | for (StringRef Attr : CodeGenOpts.DefaultFunctionAttrs) { |
2021 | StringRef Var, Value; |
2022 | std::tie(args&: Var, args&: Value) = Attr.split(Separator: '='); |
2023 | FuncAttrs.addAttribute(A: Var, V: Value); |
2024 | } |
2025 | } |
2026 | |
2027 | /// Merges `target-features` from \TargetOpts and \F, and sets the result in |
2028 | /// \FuncAttr |
2029 | /// * features from \F are always kept |
2030 | /// * a feature from \TargetOpts is kept if itself and its opposite are absent |
2031 | /// from \F |
2032 | static void |
2033 | overrideFunctionFeaturesWithTargetFeatures(llvm::AttrBuilder &FuncAttr, |
2034 | const llvm::Function &F, |
2035 | const TargetOptions &TargetOpts) { |
2036 | auto FFeatures = F.getFnAttribute(Kind: "target-features" ); |
2037 | |
2038 | llvm::StringSet<> MergedNames; |
2039 | SmallVector<StringRef> MergedFeatures; |
2040 | MergedFeatures.reserve(N: TargetOpts.Features.size()); |
2041 | |
2042 | auto AddUnmergedFeatures = [&](auto &&FeatureRange) { |
2043 | for (StringRef Feature : FeatureRange) { |
2044 | if (Feature.empty()) |
2045 | continue; |
2046 | assert(Feature[0] == '+' || Feature[0] == '-'); |
2047 | StringRef Name = Feature.drop_front(N: 1); |
2048 | bool Merged = !MergedNames.insert(key: Name).second; |
2049 | if (!Merged) |
2050 | MergedFeatures.push_back(Elt: Feature); |
2051 | } |
2052 | }; |
2053 | |
2054 | if (FFeatures.isValid()) |
2055 | AddUnmergedFeatures(llvm::split(Str: FFeatures.getValueAsString(), Separator: ',')); |
2056 | AddUnmergedFeatures(TargetOpts.Features); |
2057 | |
2058 | if (!MergedFeatures.empty()) { |
2059 | llvm::sort(C&: MergedFeatures); |
2060 | FuncAttr.addAttribute(A: "target-features" , V: llvm::join(R&: MergedFeatures, Separator: "," )); |
2061 | } |
2062 | } |
2063 | |
2064 | void CodeGen::mergeDefaultFunctionDefinitionAttributes( |
2065 | llvm::Function &F, const CodeGenOptions &CodeGenOpts, |
2066 | const LangOptions &LangOpts, const TargetOptions &TargetOpts, |
2067 | bool WillInternalize) { |
2068 | |
2069 | llvm::AttrBuilder FuncAttrs(F.getContext()); |
2070 | // Here we only extract the options that are relevant compared to the version |
2071 | // from GetCPUAndFeaturesAttributes. |
2072 | if (!TargetOpts.CPU.empty()) |
2073 | FuncAttrs.addAttribute(A: "target-cpu" , V: TargetOpts.CPU); |
2074 | if (!TargetOpts.TuneCPU.empty()) |
2075 | FuncAttrs.addAttribute(A: "tune-cpu" , V: TargetOpts.TuneCPU); |
2076 | |
2077 | ::getTrivialDefaultFunctionAttributes(Name: F.getName(), HasOptnone: F.hasOptNone(), |
2078 | CodeGenOpts, LangOpts, |
2079 | /*AttrOnCallSite=*/false, FuncAttrs); |
2080 | |
2081 | if (!WillInternalize && F.isInterposable()) { |
2082 | // Do not promote "dynamic" denormal-fp-math to this translation unit's |
2083 | // setting for weak functions that won't be internalized. The user has no |
2084 | // real control for how builtin bitcode is linked, so we shouldn't assume |
2085 | // later copies will use a consistent mode. |
2086 | F.addFnAttrs(Attrs: FuncAttrs); |
2087 | return; |
2088 | } |
2089 | |
2090 | llvm::AttributeMask AttrsToRemove; |
2091 | |
2092 | llvm::DenormalMode DenormModeToMerge = F.getDenormalModeRaw(); |
2093 | llvm::DenormalMode DenormModeToMergeF32 = F.getDenormalModeF32Raw(); |
2094 | llvm::DenormalMode Merged = |
2095 | CodeGenOpts.FPDenormalMode.mergeCalleeMode(Callee: DenormModeToMerge); |
2096 | llvm::DenormalMode MergedF32 = CodeGenOpts.FP32DenormalMode; |
2097 | |
2098 | if (DenormModeToMergeF32.isValid()) { |
2099 | MergedF32 = |
2100 | CodeGenOpts.FP32DenormalMode.mergeCalleeMode(Callee: DenormModeToMergeF32); |
2101 | } |
2102 | |
2103 | if (Merged == llvm::DenormalMode::getDefault()) { |
2104 | AttrsToRemove.addAttribute(A: "denormal-fp-math" ); |
2105 | } else if (Merged != DenormModeToMerge) { |
2106 | // Overwrite existing attribute |
2107 | FuncAttrs.addAttribute(A: "denormal-fp-math" , |
2108 | V: CodeGenOpts.FPDenormalMode.str()); |
2109 | } |
2110 | |
2111 | if (MergedF32 == llvm::DenormalMode::getDefault()) { |
2112 | AttrsToRemove.addAttribute(A: "denormal-fp-math-f32" ); |
2113 | } else if (MergedF32 != DenormModeToMergeF32) { |
2114 | // Overwrite existing attribute |
2115 | FuncAttrs.addAttribute(A: "denormal-fp-math-f32" , |
2116 | V: CodeGenOpts.FP32DenormalMode.str()); |
2117 | } |
2118 | |
2119 | F.removeFnAttrs(Attrs: AttrsToRemove); |
2120 | addDenormalModeAttrs(FPDenormalMode: Merged, FP32DenormalMode: MergedF32, FuncAttrs); |
2121 | |
2122 | overrideFunctionFeaturesWithTargetFeatures(FuncAttr&: FuncAttrs, F, TargetOpts); |
2123 | |
2124 | F.addFnAttrs(Attrs: FuncAttrs); |
2125 | } |
2126 | |
2127 | void CodeGenModule::getTrivialDefaultFunctionAttributes( |
2128 | StringRef Name, bool HasOptnone, bool AttrOnCallSite, |
2129 | llvm::AttrBuilder &FuncAttrs) { |
2130 | ::getTrivialDefaultFunctionAttributes(Name, HasOptnone, CodeGenOpts: getCodeGenOpts(), |
2131 | LangOpts: getLangOpts(), AttrOnCallSite, |
2132 | FuncAttrs); |
2133 | } |
2134 | |
2135 | void CodeGenModule::getDefaultFunctionAttributes(StringRef Name, |
2136 | bool HasOptnone, |
2137 | bool AttrOnCallSite, |
2138 | llvm::AttrBuilder &FuncAttrs) { |
2139 | getTrivialDefaultFunctionAttributes(Name, HasOptnone, AttrOnCallSite, |
2140 | FuncAttrs); |
2141 | // If we're just getting the default, get the default values for mergeable |
2142 | // attributes. |
2143 | if (!AttrOnCallSite) |
2144 | addMergableDefaultFunctionAttributes(CodeGenOpts, FuncAttrs); |
2145 | } |
2146 | |
2147 | void CodeGenModule::addDefaultFunctionDefinitionAttributes( |
2148 | llvm::AttrBuilder &attrs) { |
2149 | getDefaultFunctionAttributes(/*function name*/ Name: "" , /*optnone*/ HasOptnone: false, |
2150 | /*for call*/ AttrOnCallSite: false, FuncAttrs&: attrs); |
2151 | GetCPUAndFeaturesAttributes(GD: GlobalDecl(), AttrBuilder&: attrs); |
2152 | } |
2153 | |
2154 | static void addNoBuiltinAttributes(llvm::AttrBuilder &FuncAttrs, |
2155 | const LangOptions &LangOpts, |
2156 | const NoBuiltinAttr *NBA = nullptr) { |
2157 | auto AddNoBuiltinAttr = [&FuncAttrs](StringRef BuiltinName) { |
2158 | SmallString<32> AttributeName; |
2159 | AttributeName += "no-builtin-" ; |
2160 | AttributeName += BuiltinName; |
2161 | FuncAttrs.addAttribute(A: AttributeName); |
2162 | }; |
2163 | |
2164 | // First, handle the language options passed through -fno-builtin. |
2165 | if (LangOpts.NoBuiltin) { |
2166 | // -fno-builtin disables them all. |
2167 | FuncAttrs.addAttribute(A: "no-builtins" ); |
2168 | return; |
2169 | } |
2170 | |
2171 | // Then, add attributes for builtins specified through -fno-builtin-<name>. |
2172 | llvm::for_each(Range: LangOpts.NoBuiltinFuncs, F: AddNoBuiltinAttr); |
2173 | |
2174 | // Now, let's check the __attribute__((no_builtin("...")) attribute added to |
2175 | // the source. |
2176 | if (!NBA) |
2177 | return; |
2178 | |
2179 | // If there is a wildcard in the builtin names specified through the |
2180 | // attribute, disable them all. |
2181 | if (llvm::is_contained(NBA->builtinNames(), "*" )) { |
2182 | FuncAttrs.addAttribute(A: "no-builtins" ); |
2183 | return; |
2184 | } |
2185 | |
2186 | // And last, add the rest of the builtin names. |
2187 | llvm::for_each(NBA->builtinNames(), AddNoBuiltinAttr); |
2188 | } |
2189 | |
2190 | static bool DetermineNoUndef(QualType QTy, CodeGenTypes &Types, |
2191 | const llvm::DataLayout &DL, const ABIArgInfo &AI, |
2192 | bool CheckCoerce = true) { |
2193 | llvm::Type *Ty = Types.ConvertTypeForMem(T: QTy); |
2194 | if (AI.getKind() == ABIArgInfo::Indirect || |
2195 | AI.getKind() == ABIArgInfo::IndirectAliased) |
2196 | return true; |
2197 | if (AI.getKind() == ABIArgInfo::Extend) |
2198 | return true; |
2199 | if (!DL.typeSizeEqualsStoreSize(Ty)) |
2200 | // TODO: This will result in a modest amount of values not marked noundef |
2201 | // when they could be. We care about values that *invisibly* contain undef |
2202 | // bits from the perspective of LLVM IR. |
2203 | return false; |
2204 | if (CheckCoerce && AI.canHaveCoerceToType()) { |
2205 | llvm::Type *CoerceTy = AI.getCoerceToType(); |
2206 | if (llvm::TypeSize::isKnownGT(LHS: DL.getTypeSizeInBits(Ty: CoerceTy), |
2207 | RHS: DL.getTypeSizeInBits(Ty))) |
2208 | // If we're coercing to a type with a greater size than the canonical one, |
2209 | // we're introducing new undef bits. |
2210 | // Coercing to a type of smaller or equal size is ok, as we know that |
2211 | // there's no internal padding (typeSizeEqualsStoreSize). |
2212 | return false; |
2213 | } |
2214 | if (QTy->isBitIntType()) |
2215 | return true; |
2216 | if (QTy->isReferenceType()) |
2217 | return true; |
2218 | if (QTy->isNullPtrType()) |
2219 | return false; |
2220 | if (QTy->isMemberPointerType()) |
2221 | // TODO: Some member pointers are `noundef`, but it depends on the ABI. For |
2222 | // now, never mark them. |
2223 | return false; |
2224 | if (QTy->isScalarType()) { |
2225 | if (const ComplexType *Complex = dyn_cast<ComplexType>(Val&: QTy)) |
2226 | return DetermineNoUndef(QTy: Complex->getElementType(), Types, DL, AI, CheckCoerce: false); |
2227 | return true; |
2228 | } |
2229 | if (const VectorType *Vector = dyn_cast<VectorType>(Val&: QTy)) |
2230 | return DetermineNoUndef(QTy: Vector->getElementType(), Types, DL, AI, CheckCoerce: false); |
2231 | if (const MatrixType *Matrix = dyn_cast<MatrixType>(Val&: QTy)) |
2232 | return DetermineNoUndef(QTy: Matrix->getElementType(), Types, DL, AI, CheckCoerce: false); |
2233 | if (const ArrayType *Array = dyn_cast<ArrayType>(Val&: QTy)) |
2234 | return DetermineNoUndef(QTy: Array->getElementType(), Types, DL, AI, CheckCoerce: false); |
2235 | |
2236 | // TODO: Some structs may be `noundef`, in specific situations. |
2237 | return false; |
2238 | } |
2239 | |
2240 | /// Check if the argument of a function has maybe_undef attribute. |
2241 | static bool IsArgumentMaybeUndef(const Decl *TargetDecl, |
2242 | unsigned NumRequiredArgs, unsigned ArgNo) { |
2243 | const auto *FD = dyn_cast_or_null<FunctionDecl>(Val: TargetDecl); |
2244 | if (!FD) |
2245 | return false; |
2246 | |
2247 | // Assume variadic arguments do not have maybe_undef attribute. |
2248 | if (ArgNo >= NumRequiredArgs) |
2249 | return false; |
2250 | |
2251 | // Check if argument has maybe_undef attribute. |
2252 | if (ArgNo < FD->getNumParams()) { |
2253 | const ParmVarDecl *Param = FD->getParamDecl(i: ArgNo); |
2254 | if (Param && Param->hasAttr<MaybeUndefAttr>()) |
2255 | return true; |
2256 | } |
2257 | |
2258 | return false; |
2259 | } |
2260 | |
2261 | /// Test if it's legal to apply nofpclass for the given parameter type and it's |
2262 | /// lowered IR type. |
2263 | static bool canApplyNoFPClass(const ABIArgInfo &AI, QualType ParamType, |
2264 | bool IsReturn) { |
2265 | // Should only apply to FP types in the source, not ABI promoted. |
2266 | if (!ParamType->hasFloatingRepresentation()) |
2267 | return false; |
2268 | |
2269 | // The promoted-to IR type also needs to support nofpclass. |
2270 | llvm::Type *IRTy = AI.getCoerceToType(); |
2271 | if (llvm::AttributeFuncs::isNoFPClassCompatibleType(Ty: IRTy)) |
2272 | return true; |
2273 | |
2274 | if (llvm::StructType *ST = dyn_cast<llvm::StructType>(Val: IRTy)) { |
2275 | return !IsReturn && AI.getCanBeFlattened() && |
2276 | llvm::all_of(Range: ST->elements(), P: [](llvm::Type *Ty) { |
2277 | return llvm::AttributeFuncs::isNoFPClassCompatibleType(Ty); |
2278 | }); |
2279 | } |
2280 | |
2281 | return false; |
2282 | } |
2283 | |
2284 | /// Return the nofpclass mask that can be applied to floating-point parameters. |
2285 | static llvm::FPClassTest getNoFPClassTestMask(const LangOptions &LangOpts) { |
2286 | llvm::FPClassTest Mask = llvm::fcNone; |
2287 | if (LangOpts.NoHonorInfs) |
2288 | Mask |= llvm::fcInf; |
2289 | if (LangOpts.NoHonorNaNs) |
2290 | Mask |= llvm::fcNan; |
2291 | return Mask; |
2292 | } |
2293 | |
2294 | void CodeGenModule::AdjustMemoryAttribute(StringRef Name, |
2295 | CGCalleeInfo CalleeInfo, |
2296 | llvm::AttributeList &Attrs) { |
2297 | if (Attrs.getMemoryEffects().getModRef() == llvm::ModRefInfo::NoModRef) { |
2298 | Attrs = Attrs.removeFnAttribute(getLLVMContext(), llvm::Attribute::Memory); |
2299 | llvm::Attribute MemoryAttr = llvm::Attribute::getWithMemoryEffects( |
2300 | Context&: getLLVMContext(), ME: llvm::MemoryEffects::writeOnly()); |
2301 | Attrs = Attrs.addFnAttribute(C&: getLLVMContext(), Attr: MemoryAttr); |
2302 | } |
2303 | } |
2304 | |
2305 | /// Construct the IR attribute list of a function or call. |
2306 | /// |
2307 | /// When adding an attribute, please consider where it should be handled: |
2308 | /// |
2309 | /// - getDefaultFunctionAttributes is for attributes that are essentially |
2310 | /// part of the global target configuration (but perhaps can be |
2311 | /// overridden on a per-function basis). Adding attributes there |
2312 | /// will cause them to also be set in frontends that build on Clang's |
2313 | /// target-configuration logic, as well as for code defined in library |
2314 | /// modules such as CUDA's libdevice. |
2315 | /// |
2316 | /// - ConstructAttributeList builds on top of getDefaultFunctionAttributes |
2317 | /// and adds declaration-specific, convention-specific, and |
2318 | /// frontend-specific logic. The last is of particular importance: |
2319 | /// attributes that restrict how the frontend generates code must be |
2320 | /// added here rather than getDefaultFunctionAttributes. |
2321 | /// |
2322 | void CodeGenModule::ConstructAttributeList(StringRef Name, |
2323 | const CGFunctionInfo &FI, |
2324 | CGCalleeInfo CalleeInfo, |
2325 | llvm::AttributeList &AttrList, |
2326 | unsigned &CallingConv, |
2327 | bool AttrOnCallSite, bool IsThunk) { |
2328 | llvm::AttrBuilder FuncAttrs(getLLVMContext()); |
2329 | llvm::AttrBuilder RetAttrs(getLLVMContext()); |
2330 | |
2331 | // Collect function IR attributes from the CC lowering. |
2332 | // We'll collect the paramete and result attributes later. |
2333 | CallingConv = FI.getEffectiveCallingConvention(); |
2334 | if (FI.isNoReturn()) |
2335 | FuncAttrs.addAttribute(llvm::Attribute::NoReturn); |
2336 | if (FI.isCmseNSCall()) |
2337 | FuncAttrs.addAttribute(A: "cmse_nonsecure_call" ); |
2338 | |
2339 | // Collect function IR attributes from the callee prototype if we have one. |
2340 | AddAttributesFromFunctionProtoType(Ctx&: getContext(), FuncAttrs, |
2341 | FPT: CalleeInfo.getCalleeFunctionProtoType()); |
2342 | |
2343 | const Decl *TargetDecl = CalleeInfo.getCalleeDecl().getDecl(); |
2344 | |
2345 | // Attach assumption attributes to the declaration. If this is a call |
2346 | // site, attach assumptions from the caller to the call as well. |
2347 | AddAttributesFromAssumes(FuncAttrs, Callee: TargetDecl); |
2348 | |
2349 | bool HasOptnone = false; |
2350 | // The NoBuiltinAttr attached to the target FunctionDecl. |
2351 | const NoBuiltinAttr *NBA = nullptr; |
2352 | |
2353 | // Some ABIs may result in additional accesses to arguments that may |
2354 | // otherwise not be present. |
2355 | auto AddPotentialArgAccess = [&]() { |
2356 | llvm::Attribute A = FuncAttrs.getAttribute(llvm::Attribute::Memory); |
2357 | if (A.isValid()) |
2358 | FuncAttrs.addMemoryAttr(ME: A.getMemoryEffects() | |
2359 | llvm::MemoryEffects::argMemOnly()); |
2360 | }; |
2361 | |
2362 | // Collect function IR attributes based on declaration-specific |
2363 | // information. |
2364 | // FIXME: handle sseregparm someday... |
2365 | if (TargetDecl) { |
2366 | if (TargetDecl->hasAttr<ReturnsTwiceAttr>()) |
2367 | FuncAttrs.addAttribute(llvm::Attribute::ReturnsTwice); |
2368 | if (TargetDecl->hasAttr<NoThrowAttr>()) |
2369 | FuncAttrs.addAttribute(llvm::Attribute::NoUnwind); |
2370 | if (TargetDecl->hasAttr<NoReturnAttr>()) |
2371 | FuncAttrs.addAttribute(llvm::Attribute::NoReturn); |
2372 | if (TargetDecl->hasAttr<ColdAttr>()) |
2373 | FuncAttrs.addAttribute(llvm::Attribute::Cold); |
2374 | if (TargetDecl->hasAttr<HotAttr>()) |
2375 | FuncAttrs.addAttribute(llvm::Attribute::Hot); |
2376 | if (TargetDecl->hasAttr<NoDuplicateAttr>()) |
2377 | FuncAttrs.addAttribute(llvm::Attribute::NoDuplicate); |
2378 | if (TargetDecl->hasAttr<ConvergentAttr>()) |
2379 | FuncAttrs.addAttribute(llvm::Attribute::Convergent); |
2380 | |
2381 | if (const FunctionDecl *Fn = dyn_cast<FunctionDecl>(Val: TargetDecl)) { |
2382 | AddAttributesFromFunctionProtoType( |
2383 | getContext(), FuncAttrs, Fn->getType()->getAs<FunctionProtoType>()); |
2384 | if (AttrOnCallSite && Fn->isReplaceableGlobalAllocationFunction()) { |
2385 | // A sane operator new returns a non-aliasing pointer. |
2386 | auto Kind = Fn->getDeclName().getCXXOverloadedOperator(); |
2387 | if (getCodeGenOpts().AssumeSaneOperatorNew && |
2388 | (Kind == OO_New || Kind == OO_Array_New)) |
2389 | RetAttrs.addAttribute(llvm::Attribute::NoAlias); |
2390 | } |
2391 | const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(Val: Fn); |
2392 | const bool IsVirtualCall = MD && MD->isVirtual(); |
2393 | // Don't use [[noreturn]], _Noreturn or [[no_builtin]] for a call to a |
2394 | // virtual function. These attributes are not inherited by overloads. |
2395 | if (!(AttrOnCallSite && IsVirtualCall)) { |
2396 | if (Fn->isNoReturn()) |
2397 | FuncAttrs.addAttribute(llvm::Attribute::NoReturn); |
2398 | NBA = Fn->getAttr<NoBuiltinAttr>(); |
2399 | } |
2400 | } |
2401 | |
2402 | if (isa<FunctionDecl>(Val: TargetDecl) || isa<VarDecl>(Val: TargetDecl)) { |
2403 | // Only place nomerge attribute on call sites, never functions. This |
2404 | // allows it to work on indirect virtual function calls. |
2405 | if (AttrOnCallSite && TargetDecl->hasAttr<NoMergeAttr>()) |
2406 | FuncAttrs.addAttribute(llvm::Attribute::NoMerge); |
2407 | } |
2408 | |
2409 | // 'const', 'pure' and 'noalias' attributed functions are also nounwind. |
2410 | if (TargetDecl->hasAttr<ConstAttr>()) { |
2411 | FuncAttrs.addMemoryAttr(ME: llvm::MemoryEffects::none()); |
2412 | FuncAttrs.addAttribute(llvm::Attribute::NoUnwind); |
2413 | // gcc specifies that 'const' functions have greater restrictions than |
2414 | // 'pure' functions, so they also cannot have infinite loops. |
2415 | FuncAttrs.addAttribute(llvm::Attribute::WillReturn); |
2416 | } else if (TargetDecl->hasAttr<PureAttr>()) { |
2417 | FuncAttrs.addMemoryAttr(ME: llvm::MemoryEffects::readOnly()); |
2418 | FuncAttrs.addAttribute(llvm::Attribute::NoUnwind); |
2419 | // gcc specifies that 'pure' functions cannot have infinite loops. |
2420 | FuncAttrs.addAttribute(llvm::Attribute::WillReturn); |
2421 | } else if (TargetDecl->hasAttr<NoAliasAttr>()) { |
2422 | FuncAttrs.addMemoryAttr(ME: llvm::MemoryEffects::inaccessibleOrArgMemOnly()); |
2423 | FuncAttrs.addAttribute(llvm::Attribute::NoUnwind); |
2424 | } |
2425 | if (TargetDecl->hasAttr<RestrictAttr>()) |
2426 | RetAttrs.addAttribute(llvm::Attribute::NoAlias); |
2427 | if (TargetDecl->hasAttr<ReturnsNonNullAttr>() && |
2428 | !CodeGenOpts.NullPointerIsValid) |
2429 | RetAttrs.addAttribute(llvm::Attribute::NonNull); |
2430 | if (TargetDecl->hasAttr<AnyX86NoCallerSavedRegistersAttr>()) |
2431 | FuncAttrs.addAttribute(A: "no_caller_saved_registers" ); |
2432 | if (TargetDecl->hasAttr<AnyX86NoCfCheckAttr>()) |
2433 | FuncAttrs.addAttribute(llvm::Attribute::NoCfCheck); |
2434 | if (TargetDecl->hasAttr<LeafAttr>()) |
2435 | FuncAttrs.addAttribute(llvm::Attribute::NoCallback); |
2436 | |
2437 | HasOptnone = TargetDecl->hasAttr<OptimizeNoneAttr>(); |
2438 | if (auto *AllocSize = TargetDecl->getAttr<AllocSizeAttr>()) { |
2439 | std::optional<unsigned> NumElemsParam; |
2440 | if (AllocSize->getNumElemsParam().isValid()) |
2441 | NumElemsParam = AllocSize->getNumElemsParam().getLLVMIndex(); |
2442 | FuncAttrs.addAllocSizeAttr(ElemSizeArg: AllocSize->getElemSizeParam().getLLVMIndex(), |
2443 | NumElemsArg: NumElemsParam); |
2444 | } |
2445 | |
2446 | if (TargetDecl->hasAttr<OpenCLKernelAttr>()) { |
2447 | if (getLangOpts().OpenCLVersion <= 120) { |
2448 | // OpenCL v1.2 Work groups are always uniform |
2449 | FuncAttrs.addAttribute(A: "uniform-work-group-size" , V: "true" ); |
2450 | } else { |
2451 | // OpenCL v2.0 Work groups may be whether uniform or not. |
2452 | // '-cl-uniform-work-group-size' compile option gets a hint |
2453 | // to the compiler that the global work-size be a multiple of |
2454 | // the work-group size specified to clEnqueueNDRangeKernel |
2455 | // (i.e. work groups are uniform). |
2456 | FuncAttrs.addAttribute( |
2457 | A: "uniform-work-group-size" , |
2458 | V: llvm::toStringRef(B: getLangOpts().OffloadUniformBlock)); |
2459 | } |
2460 | } |
2461 | |
2462 | if (TargetDecl->hasAttr<CUDAGlobalAttr>() && |
2463 | getLangOpts().OffloadUniformBlock) |
2464 | FuncAttrs.addAttribute(A: "uniform-work-group-size" , V: "true" ); |
2465 | |
2466 | if (TargetDecl->hasAttr<ArmLocallyStreamingAttr>()) |
2467 | FuncAttrs.addAttribute(A: "aarch64_pstate_sm_body" ); |
2468 | } |
2469 | |
2470 | // Attach "no-builtins" attributes to: |
2471 | // * call sites: both `nobuiltin` and "no-builtins" or "no-builtin-<name>". |
2472 | // * definitions: "no-builtins" or "no-builtin-<name>" only. |
2473 | // The attributes can come from: |
2474 | // * LangOpts: -ffreestanding, -fno-builtin, -fno-builtin-<name> |
2475 | // * FunctionDecl attributes: __attribute__((no_builtin(...))) |
2476 | addNoBuiltinAttributes(FuncAttrs, getLangOpts(), NBA); |
2477 | |
2478 | // Collect function IR attributes based on global settiings. |
2479 | getDefaultFunctionAttributes(Name, HasOptnone, AttrOnCallSite, FuncAttrs); |
2480 | |
2481 | // Override some default IR attributes based on declaration-specific |
2482 | // information. |
2483 | if (TargetDecl) { |
2484 | if (TargetDecl->hasAttr<NoSpeculativeLoadHardeningAttr>()) |
2485 | FuncAttrs.removeAttribute(llvm::Attribute::SpeculativeLoadHardening); |
2486 | if (TargetDecl->hasAttr<SpeculativeLoadHardeningAttr>()) |
2487 | FuncAttrs.addAttribute(llvm::Attribute::SpeculativeLoadHardening); |
2488 | if (TargetDecl->hasAttr<NoSplitStackAttr>()) |
2489 | FuncAttrs.removeAttribute(A: "split-stack" ); |
2490 | if (TargetDecl->hasAttr<ZeroCallUsedRegsAttr>()) { |
2491 | // A function "__attribute__((...))" overrides the command-line flag. |
2492 | auto Kind = |
2493 | TargetDecl->getAttr<ZeroCallUsedRegsAttr>()->getZeroCallUsedRegs(); |
2494 | FuncAttrs.removeAttribute(A: "zero-call-used-regs" ); |
2495 | FuncAttrs.addAttribute( |
2496 | "zero-call-used-regs" , |
2497 | ZeroCallUsedRegsAttr::ConvertZeroCallUsedRegsKindToStr(Kind)); |
2498 | } |
2499 | |
2500 | // Add NonLazyBind attribute to function declarations when -fno-plt |
2501 | // is used. |
2502 | // FIXME: what if we just haven't processed the function definition |
2503 | // yet, or if it's an external definition like C99 inline? |
2504 | if (CodeGenOpts.NoPLT) { |
2505 | if (auto *Fn = dyn_cast<FunctionDecl>(Val: TargetDecl)) { |
2506 | if (!Fn->isDefined() && !AttrOnCallSite) { |
2507 | FuncAttrs.addAttribute(llvm::Attribute::NonLazyBind); |
2508 | } |
2509 | } |
2510 | } |
2511 | } |
2512 | |
2513 | // Add "sample-profile-suffix-elision-policy" attribute for internal linkage |
2514 | // functions with -funique-internal-linkage-names. |
2515 | if (TargetDecl && CodeGenOpts.UniqueInternalLinkageNames) { |
2516 | if (const auto *FD = dyn_cast_or_null<FunctionDecl>(Val: TargetDecl)) { |
2517 | if (!FD->isExternallyVisible()) |
2518 | FuncAttrs.addAttribute(A: "sample-profile-suffix-elision-policy" , |
2519 | V: "selected" ); |
2520 | } |
2521 | } |
2522 | |
2523 | // Collect non-call-site function IR attributes from declaration-specific |
2524 | // information. |
2525 | if (!AttrOnCallSite) { |
2526 | if (TargetDecl && TargetDecl->hasAttr<CmseNSEntryAttr>()) |
2527 | FuncAttrs.addAttribute(A: "cmse_nonsecure_entry" ); |
2528 | |
2529 | // Whether tail calls are enabled. |
2530 | auto shouldDisableTailCalls = [&] { |
2531 | // Should this be honored in getDefaultFunctionAttributes? |
2532 | if (CodeGenOpts.DisableTailCalls) |
2533 | return true; |
2534 | |
2535 | if (!TargetDecl) |
2536 | return false; |
2537 | |
2538 | if (TargetDecl->hasAttr<DisableTailCallsAttr>() || |
2539 | TargetDecl->hasAttr<AnyX86InterruptAttr>()) |
2540 | return true; |
2541 | |
2542 | if (CodeGenOpts.NoEscapingBlockTailCalls) { |
2543 | if (const auto *BD = dyn_cast<BlockDecl>(Val: TargetDecl)) |
2544 | if (!BD->doesNotEscape()) |
2545 | return true; |
2546 | } |
2547 | |
2548 | return false; |
2549 | }; |
2550 | if (shouldDisableTailCalls()) |
2551 | FuncAttrs.addAttribute(A: "disable-tail-calls" , V: "true" ); |
2552 | |
2553 | // CPU/feature overrides. addDefaultFunctionDefinitionAttributes |
2554 | // handles these separately to set them based on the global defaults. |
2555 | GetCPUAndFeaturesAttributes(GD: CalleeInfo.getCalleeDecl(), AttrBuilder&: FuncAttrs); |
2556 | } |
2557 | |
2558 | // Collect attributes from arguments and return values. |
2559 | ClangToLLVMArgMapping IRFunctionArgs(getContext(), FI); |
2560 | |
2561 | QualType RetTy = FI.getReturnType(); |
2562 | const ABIArgInfo &RetAI = FI.getReturnInfo(); |
2563 | const llvm::DataLayout &DL = getDataLayout(); |
2564 | |
2565 | // Determine if the return type could be partially undef |
2566 | if (CodeGenOpts.EnableNoundefAttrs && |
2567 | HasStrictReturn(Module: *this, RetTy, TargetDecl)) { |
2568 | if (!RetTy->isVoidType() && RetAI.getKind() != ABIArgInfo::Indirect && |
2569 | DetermineNoUndef(RetTy, getTypes(), DL, RetAI)) |
2570 | RetAttrs.addAttribute(llvm::Attribute::NoUndef); |
2571 | } |
2572 | |
2573 | switch (RetAI.getKind()) { |
2574 | case ABIArgInfo::Extend: |
2575 | if (RetAI.isSignExt()) |
2576 | RetAttrs.addAttribute(llvm::Attribute::SExt); |
2577 | else |
2578 | RetAttrs.addAttribute(llvm::Attribute::ZExt); |
2579 | [[fallthrough]]; |
2580 | case ABIArgInfo::Direct: |
2581 | if (RetAI.getInReg()) |
2582 | RetAttrs.addAttribute(llvm::Attribute::InReg); |
2583 | |
2584 | if (canApplyNoFPClass(AI: RetAI, ParamType: RetTy, IsReturn: true)) |
2585 | RetAttrs.addNoFPClassAttr(NoFPClassMask: getNoFPClassTestMask(LangOpts: getLangOpts())); |
2586 | |
2587 | break; |
2588 | case ABIArgInfo::Ignore: |
2589 | break; |
2590 | |
2591 | case ABIArgInfo::InAlloca: |
2592 | case ABIArgInfo::Indirect: { |
2593 | // inalloca and sret disable readnone and readonly |
2594 | AddPotentialArgAccess(); |
2595 | break; |
2596 | } |
2597 | |
2598 | case ABIArgInfo::CoerceAndExpand: |
2599 | break; |
2600 | |
2601 | case ABIArgInfo::Expand: |
2602 | case ABIArgInfo::IndirectAliased: |
2603 | llvm_unreachable("Invalid ABI kind for return argument" ); |
2604 | } |
2605 | |
2606 | if (!IsThunk) { |
2607 | // FIXME: fix this properly, https://reviews.llvm.org/D100388 |
2608 | if (const auto *RefTy = RetTy->getAs<ReferenceType>()) { |
2609 | QualType PTy = RefTy->getPointeeType(); |
2610 | if (!PTy->isIncompleteType() && PTy->isConstantSizeType()) |
2611 | RetAttrs.addDereferenceableAttr( |
2612 | Bytes: getMinimumObjectSize(Ty: PTy).getQuantity()); |
2613 | if (getTypes().getTargetAddressSpace(PTy) == 0 && |
2614 | !CodeGenOpts.NullPointerIsValid) |
2615 | RetAttrs.addAttribute(llvm::Attribute::NonNull); |
2616 | if (PTy->isObjectType()) { |
2617 | llvm::Align Alignment = |
2618 | getNaturalPointeeTypeAlignment(T: RetTy).getAsAlign(); |
2619 | RetAttrs.addAlignmentAttr(Align: Alignment); |
2620 | } |
2621 | } |
2622 | } |
2623 | |
2624 | bool hasUsedSRet = false; |
2625 | SmallVector<llvm::AttributeSet, 4> ArgAttrs(IRFunctionArgs.totalIRArgs()); |
2626 | |
2627 | // Attach attributes to sret. |
2628 | if (IRFunctionArgs.hasSRetArg()) { |
2629 | llvm::AttrBuilder SRETAttrs(getLLVMContext()); |
2630 | SRETAttrs.addStructRetAttr(Ty: getTypes().ConvertTypeForMem(T: RetTy)); |
2631 | SRETAttrs.addAttribute(llvm::Attribute::Writable); |
2632 | SRETAttrs.addAttribute(llvm::Attribute::DeadOnUnwind); |
2633 | hasUsedSRet = true; |
2634 | if (RetAI.getInReg()) |
2635 | SRETAttrs.addAttribute(llvm::Attribute::InReg); |
2636 | SRETAttrs.addAlignmentAttr(Align: RetAI.getIndirectAlign().getQuantity()); |
2637 | ArgAttrs[IRFunctionArgs.getSRetArgNo()] = |
2638 | llvm::AttributeSet::get(C&: getLLVMContext(), B: SRETAttrs); |
2639 | } |
2640 | |
2641 | // Attach attributes to inalloca argument. |
2642 | if (IRFunctionArgs.hasInallocaArg()) { |
2643 | llvm::AttrBuilder Attrs(getLLVMContext()); |
2644 | Attrs.addInAllocaAttr(Ty: FI.getArgStruct()); |
2645 | ArgAttrs[IRFunctionArgs.getInallocaArgNo()] = |
2646 | llvm::AttributeSet::get(C&: getLLVMContext(), B: Attrs); |
2647 | } |
2648 | |
2649 | // Apply `nonnull`, `dereferencable(N)` and `align N` to the `this` argument, |
2650 | // unless this is a thunk function. |
2651 | // FIXME: fix this properly, https://reviews.llvm.org/D100388 |
2652 | if (FI.isInstanceMethod() && !IRFunctionArgs.hasInallocaArg() && |
2653 | !FI.arg_begin()->type->isVoidPointerType() && !IsThunk) { |
2654 | auto IRArgs = IRFunctionArgs.getIRArgs(ArgNo: 0); |
2655 | |
2656 | assert(IRArgs.second == 1 && "Expected only a single `this` pointer." ); |
2657 | |
2658 | llvm::AttrBuilder Attrs(getLLVMContext()); |
2659 | |
2660 | QualType ThisTy = |
2661 | FI.arg_begin()->type.getTypePtr()->getPointeeType(); |
2662 | |
2663 | if (!CodeGenOpts.NullPointerIsValid && |
2664 | getTypes().getTargetAddressSpace(T: FI.arg_begin()->type) == 0) { |
2665 | Attrs.addAttribute(llvm::Attribute::NonNull); |
2666 | Attrs.addDereferenceableAttr(Bytes: getMinimumObjectSize(Ty: ThisTy).getQuantity()); |
2667 | } else { |
2668 | // FIXME dereferenceable should be correct here, regardless of |
2669 | // NullPointerIsValid. However, dereferenceable currently does not always |
2670 | // respect NullPointerIsValid and may imply nonnull and break the program. |
2671 | // See https://reviews.llvm.org/D66618 for discussions. |
2672 | Attrs.addDereferenceableOrNullAttr( |
2673 | Bytes: getMinimumObjectSize( |
2674 | Ty: FI.arg_begin()->type.castAs<PointerType>()->getPointeeType()) |
2675 | .getQuantity()); |
2676 | } |
2677 | |
2678 | llvm::Align Alignment = |
2679 | getNaturalTypeAlignment(T: ThisTy, /*BaseInfo=*/nullptr, |
2680 | /*TBAAInfo=*/nullptr, /*forPointeeType=*/true) |
2681 | .getAsAlign(); |
2682 | Attrs.addAlignmentAttr(Align: Alignment); |
2683 | |
2684 | ArgAttrs[IRArgs.first] = llvm::AttributeSet::get(C&: getLLVMContext(), B: Attrs); |
2685 | } |
2686 | |
2687 | unsigned ArgNo = 0; |
2688 | for (CGFunctionInfo::const_arg_iterator I = FI.arg_begin(), |
2689 | E = FI.arg_end(); |
2690 | I != E; ++I, ++ArgNo) { |
2691 | QualType ParamType = I->type; |
2692 | const ABIArgInfo &AI = I->info; |
2693 | llvm::AttrBuilder Attrs(getLLVMContext()); |
2694 | |
2695 | // Add attribute for padding argument, if necessary. |
2696 | if (IRFunctionArgs.hasPaddingArg(ArgNo)) { |
2697 | if (AI.getPaddingInReg()) { |
2698 | ArgAttrs[IRFunctionArgs.getPaddingArgNo(ArgNo)] = |
2699 | llvm::AttributeSet::get( |
2700 | getLLVMContext(), |
2701 | llvm::AttrBuilder(getLLVMContext()).addAttribute(llvm::Attribute::InReg)); |
2702 | } |
2703 | } |
2704 | |
2705 | // Decide whether the argument we're handling could be partially undef |
2706 | if (CodeGenOpts.EnableNoundefAttrs && |
2707 | DetermineNoUndef(QTy: ParamType, Types&: getTypes(), DL, AI)) { |
2708 | Attrs.addAttribute(llvm::Attribute::NoUndef); |
2709 | } |
2710 | |
2711 | // 'restrict' -> 'noalias' is done in EmitFunctionProlog when we |
2712 | // have the corresponding parameter variable. It doesn't make |
2713 | // sense to do it here because parameters are so messed up. |
2714 | switch (AI.getKind()) { |
2715 | case ABIArgInfo::Extend: |
2716 | if (AI.isSignExt()) |
2717 | Attrs.addAttribute(llvm::Attribute::SExt); |
2718 | else |
2719 | Attrs.addAttribute(llvm::Attribute::ZExt); |
2720 | [[fallthrough]]; |
2721 | case ABIArgInfo::Direct: |
2722 | if (ArgNo == 0 && FI.isChainCall()) |
2723 | Attrs.addAttribute(llvm::Attribute::Nest); |
2724 | else if (AI.getInReg()) |
2725 | Attrs.addAttribute(llvm::Attribute::InReg); |
2726 | Attrs.addStackAlignmentAttr(Align: llvm::MaybeAlign(AI.getDirectAlign())); |
2727 | |
2728 | if (canApplyNoFPClass(AI, ParamType, IsReturn: false)) |
2729 | Attrs.addNoFPClassAttr(NoFPClassMask: getNoFPClassTestMask(LangOpts: getLangOpts())); |
2730 | break; |
2731 | case ABIArgInfo::Indirect: { |
2732 | if (AI.getInReg()) |
2733 | Attrs.addAttribute(llvm::Attribute::InReg); |
2734 | |
2735 | if (AI.getIndirectByVal()) |
2736 | Attrs.addByValAttr(Ty: getTypes().ConvertTypeForMem(T: ParamType)); |
2737 | |
2738 | auto *Decl = ParamType->getAsRecordDecl(); |
2739 | if (CodeGenOpts.PassByValueIsNoAlias && Decl && |
2740 | Decl->getArgPassingRestrictions() == |
2741 | RecordArgPassingKind::CanPassInRegs) |
2742 | // When calling the function, the pointer passed in will be the only |
2743 | // reference to the underlying object. Mark it accordingly. |
2744 | Attrs.addAttribute(llvm::Attribute::NoAlias); |
2745 | |
2746 | // TODO: We could add the byref attribute if not byval, but it would |
2747 | // require updating many testcases. |
2748 | |
2749 | CharUnits Align = AI.getIndirectAlign(); |
2750 | |
2751 | // In a byval argument, it is important that the required |
2752 | // alignment of the type is honored, as LLVM might be creating a |
2753 | // *new* stack object, and needs to know what alignment to give |
2754 | // it. (Sometimes it can deduce a sensible alignment on its own, |
2755 | // but not if clang decides it must emit a packed struct, or the |
2756 | // user specifies increased alignment requirements.) |
2757 | // |
2758 | // This is different from indirect *not* byval, where the object |
2759 | // exists already, and the align attribute is purely |
2760 | // informative. |
2761 | assert(!Align.isZero()); |
2762 | |
2763 | // For now, only add this when we have a byval argument. |
2764 | // TODO: be less lazy about updating test cases. |
2765 | if (AI.getIndirectByVal()) |
2766 | Attrs.addAlignmentAttr(Align: Align.getQuantity()); |
2767 | |
2768 | // byval disables readnone and readonly. |
2769 | AddPotentialArgAccess(); |
2770 | break; |
2771 | } |
2772 | case ABIArgInfo::IndirectAliased: { |
2773 | CharUnits Align = AI.getIndirectAlign(); |
2774 | Attrs.addByRefAttr(Ty: getTypes().ConvertTypeForMem(T: ParamType)); |
2775 | Attrs.addAlignmentAttr(Align: Align.getQuantity()); |
2776 | break; |
2777 | } |
2778 | case ABIArgInfo::Ignore: |
2779 | case ABIArgInfo::Expand: |
2780 | case ABIArgInfo::CoerceAndExpand: |
2781 | break; |
2782 | |
2783 | case ABIArgInfo::InAlloca: |
2784 | // inalloca disables readnone and readonly. |
2785 | AddPotentialArgAccess(); |
2786 | continue; |
2787 | } |
2788 | |
2789 | if (const auto *RefTy = ParamType->getAs<ReferenceType>()) { |
2790 | QualType PTy = RefTy->getPointeeType(); |
2791 | if (!PTy->isIncompleteType() && PTy->isConstantSizeType()) |
2792 | Attrs.addDereferenceableAttr( |
2793 | Bytes: getMinimumObjectSize(Ty: PTy).getQuantity()); |
2794 | if (getTypes().getTargetAddressSpace(PTy) == 0 && |
2795 | !CodeGenOpts.NullPointerIsValid) |
2796 | Attrs.addAttribute(llvm::Attribute::NonNull); |
2797 | if (PTy->isObjectType()) { |
2798 | llvm::Align Alignment = |
2799 | getNaturalPointeeTypeAlignment(T: ParamType).getAsAlign(); |
2800 | Attrs.addAlignmentAttr(Align: Alignment); |
2801 | } |
2802 | } |
2803 | |
2804 | // From OpenCL spec v3.0.10 section 6.3.5 Alignment of Types: |
2805 | // > For arguments to a __kernel function declared to be a pointer to a |
2806 | // > data type, the OpenCL compiler can assume that the pointee is always |
2807 | // > appropriately aligned as required by the data type. |
2808 | if (TargetDecl && TargetDecl->hasAttr<OpenCLKernelAttr>() && |
2809 | ParamType->isPointerType()) { |
2810 | QualType PTy = ParamType->getPointeeType(); |
2811 | if (!PTy->isIncompleteType() && PTy->isConstantSizeType()) { |
2812 | llvm::Align Alignment = |
2813 | getNaturalPointeeTypeAlignment(T: ParamType).getAsAlign(); |
2814 | Attrs.addAlignmentAttr(Align: Alignment); |
2815 | } |
2816 | } |
2817 | |
2818 | switch (FI.getExtParameterInfo(argIndex: ArgNo).getABI()) { |
2819 | case ParameterABI::Ordinary: |
2820 | break; |
2821 | |
2822 | case ParameterABI::SwiftIndirectResult: { |
2823 | // Add 'sret' if we haven't already used it for something, but |
2824 | // only if the result is void. |
2825 | if (!hasUsedSRet && RetTy->isVoidType()) { |
2826 | Attrs.addStructRetAttr(Ty: getTypes().ConvertTypeForMem(T: ParamType)); |
2827 | hasUsedSRet = true; |
2828 | } |
2829 | |
2830 | // Add 'noalias' in either case. |
2831 | Attrs.addAttribute(llvm::Attribute::NoAlias); |
2832 | |
2833 | // Add 'dereferenceable' and 'alignment'. |
2834 | auto PTy = ParamType->getPointeeType(); |
2835 | if (!PTy->isIncompleteType() && PTy->isConstantSizeType()) { |
2836 | auto info = getContext().getTypeInfoInChars(PTy); |
2837 | Attrs.addDereferenceableAttr(Bytes: info.Width.getQuantity()); |
2838 | Attrs.addAlignmentAttr(info.Align.getAsAlign()); |
2839 | } |
2840 | break; |
2841 | } |
2842 | |
2843 | case ParameterABI::SwiftErrorResult: |
2844 | Attrs.addAttribute(llvm::Attribute::SwiftError); |
2845 | break; |
2846 | |
2847 | case ParameterABI::SwiftContext: |
2848 | Attrs.addAttribute(llvm::Attribute::SwiftSelf); |
2849 | break; |
2850 | |
2851 | case ParameterABI::SwiftAsyncContext: |
2852 | Attrs.addAttribute(llvm::Attribute::SwiftAsync); |
2853 | break; |
2854 | } |
2855 | |
2856 | if (FI.getExtParameterInfo(ArgNo).isNoEscape()) |
2857 | Attrs.addAttribute(llvm::Attribute::NoCapture); |
2858 | |
2859 | if (Attrs.hasAttributes()) { |
2860 | unsigned FirstIRArg, NumIRArgs; |
2861 | std::tie(args&: FirstIRArg, args&: NumIRArgs) = IRFunctionArgs.getIRArgs(ArgNo); |
2862 | for (unsigned i = 0; i < NumIRArgs; i++) |
2863 | ArgAttrs[FirstIRArg + i] = ArgAttrs[FirstIRArg + i].addAttributes( |
2864 | C&: getLLVMContext(), AS: llvm::AttributeSet::get(C&: getLLVMContext(), B: Attrs)); |
2865 | } |
2866 | } |
2867 | assert(ArgNo == FI.arg_size()); |
2868 | |
2869 | AttrList = llvm::AttributeList::get( |
2870 | C&: getLLVMContext(), FnAttrs: llvm::AttributeSet::get(C&: getLLVMContext(), B: FuncAttrs), |
2871 | RetAttrs: llvm::AttributeSet::get(C&: getLLVMContext(), B: RetAttrs), ArgAttrs); |
2872 | } |
2873 | |
2874 | /// An argument came in as a promoted argument; demote it back to its |
2875 | /// declared type. |
2876 | static llvm::Value *emitArgumentDemotion(CodeGenFunction &CGF, |
2877 | const VarDecl *var, |
2878 | llvm::Value *value) { |
2879 | llvm::Type *varType = CGF.ConvertType(var->getType()); |
2880 | |
2881 | // This can happen with promotions that actually don't change the |
2882 | // underlying type, like the enum promotions. |
2883 | if (value->getType() == varType) return value; |
2884 | |
2885 | assert((varType->isIntegerTy() || varType->isFloatingPointTy()) |
2886 | && "unexpected promotion type" ); |
2887 | |
2888 | if (isa<llvm::IntegerType>(Val: varType)) |
2889 | return CGF.Builder.CreateTrunc(V: value, DestTy: varType, Name: "arg.unpromote" ); |
2890 | |
2891 | return CGF.Builder.CreateFPCast(V: value, DestTy: varType, Name: "arg.unpromote" ); |
2892 | } |
2893 | |
2894 | /// Returns the attribute (either parameter attribute, or function |
2895 | /// attribute), which declares argument ArgNo to be non-null. |
2896 | static const NonNullAttr *getNonNullAttr(const Decl *FD, const ParmVarDecl *PVD, |
2897 | QualType ArgType, unsigned ArgNo) { |
2898 | // FIXME: __attribute__((nonnull)) can also be applied to: |
2899 | // - references to pointers, where the pointee is known to be |
2900 | // nonnull (apparently a Clang extension) |
2901 | // - transparent unions containing pointers |
2902 | // In the former case, LLVM IR cannot represent the constraint. In |
2903 | // the latter case, we have no guarantee that the transparent union |
2904 | // is in fact passed as a pointer. |
2905 | if (!ArgType->isAnyPointerType() && !ArgType->isBlockPointerType()) |
2906 | return nullptr; |
2907 | // First, check attribute on parameter itself. |
2908 | if (PVD) { |
2909 | if (auto ParmNNAttr = PVD->getAttr<NonNullAttr>()) |
2910 | return ParmNNAttr; |
2911 | } |
2912 | // Check function attributes. |
2913 | if (!FD) |
2914 | return nullptr; |
2915 | for (const auto *NNAttr : FD->specific_attrs<NonNullAttr>()) { |
2916 | if (NNAttr->isNonNull(ArgNo)) |
2917 | return NNAttr; |
2918 | } |
2919 | return nullptr; |
2920 | } |
2921 | |
2922 | namespace { |
2923 | struct CopyBackSwiftError final : EHScopeStack::Cleanup { |
2924 | Address Temp; |
2925 | Address Arg; |
2926 | CopyBackSwiftError(Address temp, Address arg) : Temp(temp), Arg(arg) {} |
2927 | void Emit(CodeGenFunction &CGF, Flags flags) override { |
2928 | llvm::Value *errorValue = CGF.Builder.CreateLoad(Addr: Temp); |
2929 | CGF.Builder.CreateStore(Val: errorValue, Addr: Arg); |
2930 | } |
2931 | }; |
2932 | } |
2933 | |
2934 | void CodeGenFunction::EmitFunctionProlog(const CGFunctionInfo &FI, |
2935 | llvm::Function *Fn, |
2936 | const FunctionArgList &Args) { |
2937 | if (CurCodeDecl && CurCodeDecl->hasAttr<NakedAttr>()) |
2938 | // Naked functions don't have prologues. |
2939 | return; |
2940 | |
2941 | // If this is an implicit-return-zero function, go ahead and |
2942 | // initialize the return value. TODO: it might be nice to have |
2943 | // a more general mechanism for this that didn't require synthesized |
2944 | // return statements. |
2945 | if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(Val: CurCodeDecl)) { |
2946 | if (FD->hasImplicitReturnZero()) { |
2947 | QualType RetTy = FD->getReturnType().getUnqualifiedType(); |
2948 | llvm::Type* LLVMTy = CGM.getTypes().ConvertType(T: RetTy); |
2949 | llvm::Constant* Zero = llvm::Constant::getNullValue(Ty: LLVMTy); |
2950 | Builder.CreateStore(Val: Zero, Addr: ReturnValue); |
2951 | } |
2952 | } |
2953 | |
2954 | // FIXME: We no longer need the types from FunctionArgList; lift up and |
2955 | // simplify. |
2956 | |
2957 | ClangToLLVMArgMapping IRFunctionArgs(CGM.getContext(), FI); |
2958 | assert(Fn->arg_size() == IRFunctionArgs.totalIRArgs()); |
2959 | |
2960 | // If we're using inalloca, all the memory arguments are GEPs off of the last |
2961 | // parameter, which is a pointer to the complete memory area. |
2962 | Address ArgStruct = Address::invalid(); |
2963 | if (IRFunctionArgs.hasInallocaArg()) |
2964 | ArgStruct = Address(Fn->getArg(i: IRFunctionArgs.getInallocaArgNo()), |
2965 | FI.getArgStruct(), FI.getArgStructAlignment()); |
2966 | |
2967 | // Name the struct return parameter. |
2968 | if (IRFunctionArgs.hasSRetArg()) { |
2969 | auto AI = Fn->getArg(i: IRFunctionArgs.getSRetArgNo()); |
2970 | AI->setName("agg.result" ); |
2971 | AI->addAttr(llvm::Attribute::NoAlias); |
2972 | } |
2973 | |
2974 | // Track if we received the parameter as a pointer (indirect, byval, or |
2975 | // inalloca). If already have a pointer, EmitParmDecl doesn't need to copy it |
2976 | // into a local alloca for us. |
2977 | SmallVector<ParamValue, 16> ArgVals; |
2978 | ArgVals.reserve(N: Args.size()); |
2979 | |
2980 | // Create a pointer value for every parameter declaration. This usually |
2981 | // entails copying one or more LLVM IR arguments into an alloca. Don't push |
2982 | // any cleanups or do anything that might unwind. We do that separately, so |
2983 | // we can push the cleanups in the correct order for the ABI. |
2984 | assert(FI.arg_size() == Args.size() && |
2985 | "Mismatch between function signature & arguments." ); |
2986 | unsigned ArgNo = 0; |
2987 | CGFunctionInfo::const_arg_iterator info_it = FI.arg_begin(); |
2988 | for (FunctionArgList::const_iterator i = Args.begin(), e = Args.end(); |
2989 | i != e; ++i, ++info_it, ++ArgNo) { |
2990 | const VarDecl *Arg = *i; |
2991 | const ABIArgInfo &ArgI = info_it->info; |
2992 | |
2993 | bool isPromoted = |
2994 | isa<ParmVarDecl>(Val: Arg) && cast<ParmVarDecl>(Val: Arg)->isKNRPromoted(); |
2995 | // We are converting from ABIArgInfo type to VarDecl type directly, unless |
2996 | // the parameter is promoted. In this case we convert to |
2997 | // CGFunctionInfo::ArgInfo type with subsequent argument demotion. |
2998 | QualType Ty = isPromoted ? info_it->type : Arg->getType(); |
2999 | assert(hasScalarEvaluationKind(Ty) == |
3000 | hasScalarEvaluationKind(Arg->getType())); |
3001 | |
3002 | unsigned FirstIRArg, NumIRArgs; |
3003 | std::tie(args&: FirstIRArg, args&: NumIRArgs) = IRFunctionArgs.getIRArgs(ArgNo); |
3004 | |
3005 | switch (ArgI.getKind()) { |
3006 | case ABIArgInfo::InAlloca: { |
3007 | assert(NumIRArgs == 0); |
3008 | auto FieldIndex = ArgI.getInAllocaFieldIndex(); |
3009 | Address V = |
3010 | Builder.CreateStructGEP(ArgStruct, FieldIndex, Arg->getName()); |
3011 | if (ArgI.getInAllocaIndirect()) |
3012 | V = Address(Builder.CreateLoad(Addr: V), ConvertTypeForMem(T: Ty), |
3013 | getContext().getTypeAlignInChars(T: Ty)); |
3014 | ArgVals.push_back(Elt: ParamValue::forIndirect(addr: V)); |
3015 | break; |
3016 | } |
3017 | |
3018 | case ABIArgInfo::Indirect: |
3019 | case ABIArgInfo::IndirectAliased: { |
3020 | assert(NumIRArgs == 1); |
3021 | Address ParamAddr = Address(Fn->getArg(i: FirstIRArg), ConvertTypeForMem(T: Ty), |
3022 | ArgI.getIndirectAlign(), KnownNonNull); |
3023 | |
3024 | if (!hasScalarEvaluationKind(T: Ty)) { |
3025 | // Aggregates and complex variables are accessed by reference. All we |
3026 | // need to do is realign the value, if requested. Also, if the address |
3027 | // may be aliased, copy it to ensure that the parameter variable is |
3028 | // mutable and has a unique adress, as C requires. |
3029 | Address V = ParamAddr; |
3030 | if (ArgI.getIndirectRealign() || ArgI.isIndirectAliased()) { |
3031 | Address AlignedTemp = CreateMemTemp(T: Ty, Name: "coerce" ); |
3032 | |
3033 | // Copy from the incoming argument pointer to the temporary with the |
3034 | // appropriate alignment. |
3035 | // |
3036 | // FIXME: We should have a common utility for generating an aggregate |
3037 | // copy. |
3038 | CharUnits Size = getContext().getTypeSizeInChars(T: Ty); |
3039 | Builder.CreateMemCpy( |
3040 | Dst: AlignedTemp.getPointer(), DstAlign: AlignedTemp.getAlignment().getAsAlign(), |
3041 | Src: ParamAddr.getPointer(), SrcAlign: ParamAddr.getAlignment().getAsAlign(), |
3042 | Size: llvm::ConstantInt::get(Ty: IntPtrTy, V: Size.getQuantity())); |
3043 | V = AlignedTemp; |
3044 | } |
3045 | ArgVals.push_back(Elt: ParamValue::forIndirect(addr: V)); |
3046 | } else { |
3047 | // Load scalar value from indirect argument. |
3048 | llvm::Value *V = |
3049 | EmitLoadOfScalar(ParamAddr, false, Ty, Arg->getBeginLoc()); |
3050 | |
3051 | if (isPromoted) |
3052 | V = emitArgumentDemotion(CGF&: *this, var: Arg, value: V); |
3053 | ArgVals.push_back(Elt: ParamValue::forDirect(value: V)); |
3054 | } |
3055 | break; |
3056 | } |
3057 | |
3058 | case ABIArgInfo::Extend: |
3059 | case ABIArgInfo::Direct: { |
3060 | auto AI = Fn->getArg(i: FirstIRArg); |
3061 | llvm::Type *LTy = ConvertType(Arg->getType()); |
3062 | |
3063 | // Prepare parameter attributes. So far, only attributes for pointer |
3064 | // parameters are prepared. See |
3065 | // http://llvm.org/docs/LangRef.html#paramattrs. |
3066 | if (ArgI.getDirectOffset() == 0 && LTy->isPointerTy() && |
3067 | ArgI.getCoerceToType()->isPointerTy()) { |
3068 | assert(NumIRArgs == 1); |
3069 | |
3070 | if (const ParmVarDecl *PVD = dyn_cast<ParmVarDecl>(Val: Arg)) { |
3071 | // Set `nonnull` attribute if any. |
3072 | if (getNonNullAttr(CurCodeDecl, PVD, PVD->getType(), |
3073 | PVD->getFunctionScopeIndex()) && |
3074 | !CGM.getCodeGenOpts().NullPointerIsValid) |
3075 | AI->addAttr(llvm::Attribute::NonNull); |
3076 | |
3077 | QualType OTy = PVD->getOriginalType(); |
3078 | if (const auto *ArrTy = |
3079 | getContext().getAsConstantArrayType(T: OTy)) { |
3080 | // A C99 array parameter declaration with the static keyword also |
3081 | // indicates dereferenceability, and if the size is constant we can |
3082 | // use the dereferenceable attribute (which requires the size in |
3083 | // bytes). |
3084 | if (ArrTy->getSizeModifier() == ArraySizeModifier::Static) { |
3085 | QualType ETy = ArrTy->getElementType(); |
3086 | llvm::Align Alignment = |
3087 | CGM.getNaturalTypeAlignment(T: ETy).getAsAlign(); |
3088 | AI->addAttrs(B&: llvm::AttrBuilder(getLLVMContext()).addAlignmentAttr(Align: Alignment)); |
3089 | uint64_t ArrSize = ArrTy->getSize().getZExtValue(); |
3090 | if (!ETy->isIncompleteType() && ETy->isConstantSizeType() && |
3091 | ArrSize) { |
3092 | llvm::AttrBuilder Attrs(getLLVMContext()); |
3093 | Attrs.addDereferenceableAttr( |
3094 | Bytes: getContext().getTypeSizeInChars(T: ETy).getQuantity() * |
3095 | ArrSize); |
3096 | AI->addAttrs(B&: Attrs); |
3097 | } else if (getContext().getTargetInfo().getNullPointerValue( |
3098 | AddrSpace: ETy.getAddressSpace()) == 0 && |
3099 | !CGM.getCodeGenOpts().NullPointerIsValid) { |
3100 | AI->addAttr(llvm::Attribute::NonNull); |
3101 | } |
3102 | } |
3103 | } else if (const auto *ArrTy = |
3104 | getContext().getAsVariableArrayType(T: OTy)) { |
3105 | // For C99 VLAs with the static keyword, we don't know the size so |
3106 | // we can't use the dereferenceable attribute, but in addrspace(0) |
3107 | // we know that it must be nonnull. |
3108 | if (ArrTy->getSizeModifier() == ArraySizeModifier::Static) { |
3109 | QualType ETy = ArrTy->getElementType(); |
3110 | llvm::Align Alignment = |
3111 | CGM.getNaturalTypeAlignment(T: ETy).getAsAlign(); |
3112 | AI->addAttrs(B&: llvm::AttrBuilder(getLLVMContext()).addAlignmentAttr(Align: Alignment)); |
3113 | if (!getTypes().getTargetAddressSpace(ETy) && |
3114 | !CGM.getCodeGenOpts().NullPointerIsValid) |
3115 | AI->addAttr(llvm::Attribute::NonNull); |
3116 | } |
3117 | } |
3118 | |
3119 | // Set `align` attribute if any. |
3120 | const auto *AVAttr = PVD->getAttr<AlignValueAttr>(); |
3121 | if (!AVAttr) |
3122 | if (const auto *TOTy = OTy->getAs<TypedefType>()) |
3123 | AVAttr = TOTy->getDecl()->getAttr<AlignValueAttr>(); |
3124 | if (AVAttr && !SanOpts.has(K: SanitizerKind::Alignment)) { |
3125 | // If alignment-assumption sanitizer is enabled, we do *not* add |
3126 | // alignment attribute here, but emit normal alignment assumption, |
3127 | // so the UBSAN check could function. |
3128 | llvm::ConstantInt *AlignmentCI = |
3129 | cast<llvm::ConstantInt>(EmitScalarExpr(E: AVAttr->getAlignment())); |
3130 | uint64_t AlignmentInt = |
3131 | AlignmentCI->getLimitedValue(Limit: llvm::Value::MaximumAlignment); |
3132 | if (AI->getParamAlign().valueOrOne() < AlignmentInt) { |
3133 | AI->removeAttr(llvm::Attribute::AttrKind::Alignment); |
3134 | AI->addAttrs(B&: llvm::AttrBuilder(getLLVMContext()).addAlignmentAttr( |
3135 | Align: llvm::Align(AlignmentInt))); |
3136 | } |
3137 | } |
3138 | } |
3139 | |
3140 | // Set 'noalias' if an argument type has the `restrict` qualifier. |
3141 | if (Arg->getType().isRestrictQualified()) |
3142 | AI->addAttr(llvm::Attribute::NoAlias); |
3143 | } |
3144 | |
3145 | // Prepare the argument value. If we have the trivial case, handle it |
3146 | // with no muss and fuss. |
3147 | if (!isa<llvm::StructType>(Val: ArgI.getCoerceToType()) && |
3148 | ArgI.getCoerceToType() == ConvertType(T: Ty) && |
3149 | ArgI.getDirectOffset() == 0) { |
3150 | assert(NumIRArgs == 1); |
3151 | |
3152 | // LLVM expects swifterror parameters to be used in very restricted |
3153 | // ways. Copy the value into a less-restricted temporary. |
3154 | llvm::Value *V = AI; |
3155 | if (FI.getExtParameterInfo(argIndex: ArgNo).getABI() |
3156 | == ParameterABI::SwiftErrorResult) { |
3157 | QualType pointeeTy = Ty->getPointeeType(); |
3158 | assert(pointeeTy->isPointerType()); |
3159 | Address temp = |
3160 | CreateMemTemp(pointeeTy, getPointerAlign(), "swifterror.temp" ); |
3161 | Address arg(V, ConvertTypeForMem(T: pointeeTy), |
3162 | getContext().getTypeAlignInChars(T: pointeeTy)); |
3163 | llvm::Value *incomingErrorValue = Builder.CreateLoad(Addr: arg); |
3164 | Builder.CreateStore(Val: incomingErrorValue, Addr: temp); |
3165 | V = temp.getPointer(); |
3166 | |
3167 | // Push a cleanup to copy the value back at the end of the function. |
3168 | // The convention does not guarantee that the value will be written |
3169 | // back if the function exits with an unwind exception. |
3170 | EHStack.pushCleanup<CopyBackSwiftError>(Kind: NormalCleanup, A: temp, A: arg); |
3171 | } |
3172 | |
3173 | // Ensure the argument is the correct type. |
3174 | if (V->getType() != ArgI.getCoerceToType()) |
3175 | V = Builder.CreateBitCast(V, DestTy: ArgI.getCoerceToType()); |
3176 | |
3177 | if (isPromoted) |
3178 | V = emitArgumentDemotion(CGF&: *this, var: Arg, value: V); |
3179 | |
3180 | // Because of merging of function types from multiple decls it is |
3181 | // possible for the type of an argument to not match the corresponding |
3182 | // type in the function type. Since we are codegening the callee |
3183 | // in here, add a cast to the argument type. |
3184 | llvm::Type *LTy = ConvertType(Arg->getType()); |
3185 | if (V->getType() != LTy) |
3186 | V = Builder.CreateBitCast(V, DestTy: LTy); |
3187 | |
3188 | ArgVals.push_back(Elt: ParamValue::forDirect(value: V)); |
3189 | break; |
3190 | } |
3191 | |
3192 | // VLST arguments are coerced to VLATs at the function boundary for |
3193 | // ABI consistency. If this is a VLST that was coerced to |
3194 | // a VLAT at the function boundary and the types match up, use |
3195 | // llvm.vector.extract to convert back to the original VLST. |
3196 | if (auto *VecTyTo = dyn_cast<llvm::FixedVectorType>(ConvertType(Ty))) { |
3197 | llvm::Value *Coerced = Fn->getArg(i: FirstIRArg); |
3198 | if (auto *VecTyFrom = |
3199 | dyn_cast<llvm::ScalableVectorType>(Val: Coerced->getType())) { |
3200 | // If we are casting a scalable i1 predicate vector to a fixed i8 |
3201 | // vector, bitcast the source and use a vector extract. |
3202 | if (VecTyFrom->getElementType()->isIntegerTy(Bitwidth: 1) && |
3203 | VecTyFrom->getElementCount().isKnownMultipleOf(RHS: 8) && |
3204 | VecTyTo->getElementType() == Builder.getInt8Ty()) { |
3205 | VecTyFrom = llvm::ScalableVectorType::get( |
3206 | VecTyTo->getElementType(), |
3207 | VecTyFrom->getElementCount().getKnownMinValue() / 8); |
3208 | Coerced = Builder.CreateBitCast(V: Coerced, DestTy: VecTyFrom); |
3209 | } |
3210 | if (VecTyFrom->getElementType() == VecTyTo->getElementType()) { |
3211 | llvm::Value *Zero = llvm::Constant::getNullValue(Ty: CGM.Int64Ty); |
3212 | |
3213 | assert(NumIRArgs == 1); |
3214 | Coerced->setName(Arg->getName() + ".coerce" ); |
3215 | ArgVals.push_back(Elt: ParamValue::forDirect(value: Builder.CreateExtractVector( |
3216 | DstType: VecTyTo, SrcVec: Coerced, Idx: Zero, Name: "cast.fixed" ))); |
3217 | break; |
3218 | } |
3219 | } |
3220 | } |
3221 | |
3222 | llvm::StructType *STy = |
3223 | dyn_cast<llvm::StructType>(Val: ArgI.getCoerceToType()); |
3224 | llvm::TypeSize StructSize; |
3225 | llvm::TypeSize PtrElementSize; |
3226 | if (ArgI.isDirect() && !ArgI.getCanBeFlattened() && STy && |
3227 | STy->getNumElements() > 1) { |
3228 | StructSize = CGM.getDataLayout().getTypeAllocSize(Ty: STy); |
3229 | PtrElementSize = |
3230 | CGM.getDataLayout().getTypeAllocSize(Ty: ConvertTypeForMem(T: Ty)); |
3231 | if (STy->containsHomogeneousScalableVectorTypes()) { |
3232 | assert(StructSize == PtrElementSize && |
3233 | "Only allow non-fractional movement of structure with" |
3234 | "homogeneous scalable vector type" ); |
3235 | |
3236 | ArgVals.push_back(Elt: ParamValue::forDirect(value: AI)); |
3237 | break; |
3238 | } |
3239 | } |
3240 | |
3241 | Address Alloca = CreateMemTemp(Ty, getContext().getDeclAlign(Arg), |
3242 | Arg->getName()); |
3243 | |
3244 | // Pointer to store into. |
3245 | Address Ptr = emitAddressAtOffset(CGF&: *this, addr: Alloca, info: ArgI); |
3246 | |
3247 | // Fast-isel and the optimizer generally like scalar values better than |
3248 | // FCAs, so we flatten them if this is safe to do for this argument. |
3249 | if (ArgI.isDirect() && ArgI.getCanBeFlattened() && STy && |
3250 | STy->getNumElements() > 1) { |
3251 | llvm::TypeSize StructSize = CGM.getDataLayout().getTypeAllocSize(Ty: STy); |
3252 | llvm::TypeSize PtrElementSize = |
3253 | CGM.getDataLayout().getTypeAllocSize(Ty: Ptr.getElementType()); |
3254 | if (StructSize.isScalable()) { |
3255 | assert(STy->containsHomogeneousScalableVectorTypes() && |
3256 | "ABI only supports structure with homogeneous scalable vector " |
3257 | "type" ); |
3258 | assert(StructSize == PtrElementSize && |
3259 | "Only allow non-fractional movement of structure with" |
3260 | "homogeneous scalable vector type" ); |
3261 | assert(STy->getNumElements() == NumIRArgs); |
3262 | |
3263 | llvm::Value *LoadedStructValue = llvm::PoisonValue::get(T: STy); |
3264 | for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { |
3265 | auto *AI = Fn->getArg(i: FirstIRArg + i); |
3266 | AI->setName(Arg->getName() + ".coerce" + Twine(i)); |
3267 | LoadedStructValue = |
3268 | Builder.CreateInsertValue(Agg: LoadedStructValue, Val: AI, Idxs: i); |
3269 | } |
3270 | |
3271 | Builder.CreateStore(Val: LoadedStructValue, Addr: Ptr); |
3272 | } else { |
3273 | uint64_t SrcSize = StructSize.getFixedValue(); |
3274 | uint64_t DstSize = PtrElementSize.getFixedValue(); |
3275 | |
3276 | Address AddrToStoreInto = Address::invalid(); |
3277 | if (SrcSize <= DstSize) { |
3278 | AddrToStoreInto = Ptr.withElementType(ElemTy: STy); |
3279 | } else { |
3280 | AddrToStoreInto = |
3281 | CreateTempAlloca(Ty: STy, align: Alloca.getAlignment(), Name: "coerce" ); |
3282 | } |
3283 | |
3284 | assert(STy->getNumElements() == NumIRArgs); |
3285 | for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { |
3286 | auto AI = Fn->getArg(i: FirstIRArg + i); |
3287 | AI->setName(Arg->getName() + ".coerce" + Twine(i)); |
3288 | Address EltPtr = Builder.CreateStructGEP(Addr: AddrToStoreInto, Index: i); |
3289 | Builder.CreateStore(Val: AI, Addr: EltPtr); |
3290 | } |
3291 | |
3292 | if (SrcSize > DstSize) { |
3293 | Builder.CreateMemCpy(Dest: Ptr, Src: AddrToStoreInto, Size: DstSize); |
3294 | } |
3295 | } |
3296 | } else { |
3297 | // Simple case, just do a coerced store of the argument into the alloca. |
3298 | assert(NumIRArgs == 1); |
3299 | auto AI = Fn->getArg(i: FirstIRArg); |
3300 | AI->setName(Arg->getName() + ".coerce" ); |
3301 | CreateCoercedStore(Src: AI, Dst: Ptr, /*DstIsVolatile=*/false, CGF&: *this); |
3302 | } |
3303 | |
3304 | // Match to what EmitParmDecl is expecting for this type. |
3305 | if (CodeGenFunction::hasScalarEvaluationKind(T: Ty)) { |
3306 | llvm::Value *V = |
3307 | EmitLoadOfScalar(Alloca, false, Ty, Arg->getBeginLoc()); |
3308 | if (isPromoted) |
3309 | V = emitArgumentDemotion(CGF&: *this, var: Arg, value: V); |
3310 | ArgVals.push_back(Elt: ParamValue::forDirect(value: V)); |
3311 | } else { |
3312 | ArgVals.push_back(Elt: ParamValue::forIndirect(addr: Alloca)); |
3313 | } |
3314 | break; |
3315 | } |
3316 | |
3317 | case ABIArgInfo::CoerceAndExpand: { |
3318 | // Reconstruct into a temporary. |
3319 | Address alloca = CreateMemTemp(T: Ty, Align: getContext().getDeclAlign(Arg)); |
3320 | ArgVals.push_back(Elt: ParamValue::forIndirect(addr: alloca)); |
3321 | |
3322 | auto coercionType = ArgI.getCoerceAndExpandType(); |
3323 | alloca = alloca.withElementType(ElemTy: coercionType); |
3324 | |
3325 | unsigned argIndex = FirstIRArg; |
3326 | for (unsigned i = 0, e = coercionType->getNumElements(); i != e; ++i) { |
3327 | llvm::Type *eltType = coercionType->getElementType(N: i); |
3328 | if (ABIArgInfo::isPaddingForCoerceAndExpand(eltType)) |
3329 | continue; |
3330 | |
3331 | auto eltAddr = Builder.CreateStructGEP(Addr: alloca, Index: i); |
3332 | auto elt = Fn->getArg(i: argIndex++); |
3333 | Builder.CreateStore(Val: elt, Addr: eltAddr); |
3334 | } |
3335 | assert(argIndex == FirstIRArg + NumIRArgs); |
3336 | break; |
3337 | } |
3338 | |
3339 | case ABIArgInfo::Expand: { |
3340 | // If this structure was expanded into multiple arguments then |
3341 | // we need to create a temporary and reconstruct it from the |
3342 | // arguments. |
3343 | Address Alloca = CreateMemTemp(T: Ty, Align: getContext().getDeclAlign(Arg)); |
3344 | LValue LV = MakeAddrLValue(Addr: Alloca, T: Ty); |
3345 | ArgVals.push_back(Elt: ParamValue::forIndirect(addr: Alloca)); |
3346 | |
3347 | auto FnArgIter = Fn->arg_begin() + FirstIRArg; |
3348 | ExpandTypeFromArgs(Ty, LV, AI&: FnArgIter); |
3349 | assert(FnArgIter == Fn->arg_begin() + FirstIRArg + NumIRArgs); |
3350 | for (unsigned i = 0, e = NumIRArgs; i != e; ++i) { |
3351 | auto AI = Fn->getArg(i: FirstIRArg + i); |
3352 | AI->setName(Arg->getName() + "." + Twine(i)); |
3353 | } |
3354 | break; |
3355 | } |
3356 | |
3357 | case ABIArgInfo::Ignore: |
3358 | assert(NumIRArgs == 0); |
3359 | // Initialize the local variable appropriately. |
3360 | if (!hasScalarEvaluationKind(T: Ty)) { |
3361 | ArgVals.push_back(Elt: ParamValue::forIndirect(addr: CreateMemTemp(T: Ty))); |
3362 | } else { |
3363 | llvm::Value *U = llvm::UndefValue::get(T: ConvertType(Arg->getType())); |
3364 | ArgVals.push_back(Elt: ParamValue::forDirect(value: U)); |
3365 | } |
3366 | break; |
3367 | } |
3368 | } |
3369 | |
3370 | if (getTarget().getCXXABI().areArgsDestroyedLeftToRightInCallee()) { |
3371 | for (int I = Args.size() - 1; I >= 0; --I) |
3372 | EmitParmDecl(D: *Args[I], Arg: ArgVals[I], ArgNo: I + 1); |
3373 | } else { |
3374 | for (unsigned I = 0, E = Args.size(); I != E; ++I) |
3375 | EmitParmDecl(D: *Args[I], Arg: ArgVals[I], ArgNo: I + 1); |
3376 | } |
3377 | } |
3378 | |
3379 | static void eraseUnusedBitCasts(llvm::Instruction *insn) { |
3380 | while (insn->use_empty()) { |
3381 | llvm::BitCastInst *bitcast = dyn_cast<llvm::BitCastInst>(Val: insn); |
3382 | if (!bitcast) return; |
3383 | |
3384 | // This is "safe" because we would have used a ConstantExpr otherwise. |
3385 | insn = cast<llvm::Instruction>(Val: bitcast->getOperand(i_nocapture: 0)); |
3386 | bitcast->eraseFromParent(); |
3387 | } |
3388 | } |
3389 | |
3390 | /// Try to emit a fused autorelease of a return result. |
3391 | static llvm::Value *tryEmitFusedAutoreleaseOfResult(CodeGenFunction &CGF, |
3392 | llvm::Value *result) { |
3393 | // We must be immediately followed the cast. |
3394 | llvm::BasicBlock *BB = CGF.Builder.GetInsertBlock(); |
3395 | if (BB->empty()) return nullptr; |
3396 | if (&BB->back() != result) return nullptr; |
3397 | |
3398 | llvm::Type *resultType = result->getType(); |
3399 | |
3400 | // result is in a BasicBlock and is therefore an Instruction. |
3401 | llvm::Instruction *generator = cast<llvm::Instruction>(Val: result); |
3402 | |
3403 | SmallVector<llvm::Instruction *, 4> InstsToKill; |
3404 | |
3405 | // Look for: |
3406 | // %generator = bitcast %type1* %generator2 to %type2* |
3407 | while (llvm::BitCastInst *bitcast = dyn_cast<llvm::BitCastInst>(Val: generator)) { |
3408 | // We would have emitted this as a constant if the operand weren't |
3409 | // an Instruction. |
3410 | generator = cast<llvm::Instruction>(Val: bitcast->getOperand(i_nocapture: 0)); |
3411 | |
3412 | // Require the generator to be immediately followed by the cast. |
3413 | if (generator->getNextNode() != bitcast) |
3414 | return nullptr; |
3415 | |
3416 | InstsToKill.push_back(Elt: bitcast); |
3417 | } |
3418 | |
3419 | // Look for: |
3420 | // %generator = call i8* @objc_retain(i8* %originalResult) |
3421 | // or |
3422 | // %generator = call i8* @objc_retainAutoreleasedReturnValue(i8* %originalResult) |
3423 | llvm::CallInst *call = dyn_cast<llvm::CallInst>(Val: generator); |
3424 | if (!call) return nullptr; |
3425 | |
3426 | bool doRetainAutorelease; |
3427 | |
3428 | if (call->getCalledOperand() == CGF.CGM.getObjCEntrypoints().objc_retain) { |
3429 | doRetainAutorelease = true; |
3430 | } else if (call->getCalledOperand() == |
3431 | CGF.CGM.getObjCEntrypoints().objc_retainAutoreleasedReturnValue) { |
3432 | doRetainAutorelease = false; |
3433 | |
3434 | // If we emitted an assembly marker for this call (and the |
3435 | // ARCEntrypoints field should have been set if so), go looking |
3436 | // for that call. If we can't find it, we can't do this |
3437 | // optimization. But it should always be the immediately previous |
3438 | // instruction, unless we needed bitcasts around the call. |
3439 | if (CGF.CGM.getObjCEntrypoints().retainAutoreleasedReturnValueMarker) { |
3440 | llvm::Instruction *prev = call->getPrevNode(); |
3441 | assert(prev); |
3442 | if (isa<llvm::BitCastInst>(Val: prev)) { |
3443 | prev = prev->getPrevNode(); |
3444 | assert(prev); |
3445 | } |
3446 | assert(isa<llvm::CallInst>(prev)); |
3447 | assert(cast<llvm::CallInst>(prev)->getCalledOperand() == |
3448 | CGF.CGM.getObjCEntrypoints().retainAutoreleasedReturnValueMarker); |
3449 | InstsToKill.push_back(Elt: prev); |
3450 | } |
3451 | } else { |
3452 | return nullptr; |
3453 | } |
3454 | |
3455 | result = call->getArgOperand(i: 0); |
3456 | InstsToKill.push_back(Elt: call); |
3457 | |
3458 | // Keep killing bitcasts, for sanity. Note that we no longer care |
3459 | // about precise ordering as long as there's exactly one use. |
3460 | while (llvm::BitCastInst *bitcast = dyn_cast<llvm::BitCastInst>(Val: result)) { |
3461 | if (!bitcast->hasOneUse()) break; |
3462 | InstsToKill.push_back(Elt: bitcast); |
3463 | result = bitcast->getOperand(i_nocapture: 0); |
3464 | } |
3465 | |
3466 | // Delete all the unnecessary instructions, from latest to earliest. |
3467 | for (auto *I : InstsToKill) |
3468 | I->eraseFromParent(); |
3469 | |
3470 | // Do the fused retain/autorelease if we were asked to. |
3471 | if (doRetainAutorelease) |
3472 | result = CGF.EmitARCRetainAutoreleaseReturnValue(value: result); |
3473 | |
3474 | // Cast back to the result type. |
3475 | return CGF.Builder.CreateBitCast(V: result, DestTy: resultType); |
3476 | } |
3477 | |
3478 | /// If this is a +1 of the value of an immutable 'self', remove it. |
3479 | static llvm::Value *tryRemoveRetainOfSelf(CodeGenFunction &CGF, |
3480 | llvm::Value *result) { |
3481 | // This is only applicable to a method with an immutable 'self'. |
3482 | const ObjCMethodDecl *method = |
3483 | dyn_cast_or_null<ObjCMethodDecl>(Val: CGF.CurCodeDecl); |
3484 | if (!method) return nullptr; |
3485 | const VarDecl *self = method->getSelfDecl(); |
3486 | if (!self->getType().isConstQualified()) return nullptr; |
3487 | |
3488 | // Look for a retain call. Note: stripPointerCasts looks through returned arg |
3489 | // functions, which would cause us to miss the retain. |
3490 | llvm::CallInst *retainCall = dyn_cast<llvm::CallInst>(Val: result); |
3491 | if (!retainCall || retainCall->getCalledOperand() != |
3492 | CGF.CGM.getObjCEntrypoints().objc_retain) |
3493 | return nullptr; |
3494 | |
3495 | // Look for an ordinary load of 'self'. |
3496 | llvm::Value *retainedValue = retainCall->getArgOperand(i: 0); |
3497 | llvm::LoadInst *load = |
3498 | dyn_cast<llvm::LoadInst>(Val: retainedValue->stripPointerCasts()); |
3499 | if (!load || load->isAtomic() || load->isVolatile() || |
3500 | load->getPointerOperand() != CGF.GetAddrOfLocalVar(VD: self).getPointer()) |
3501 | return nullptr; |
3502 | |
3503 | // Okay! Burn it all down. This relies for correctness on the |
3504 | // assumption that the retain is emitted as part of the return and |
3505 | // that thereafter everything is used "linearly". |
3506 | llvm::Type *resultType = result->getType(); |
3507 | eraseUnusedBitCasts(insn: cast<llvm::Instruction>(Val: result)); |
3508 | assert(retainCall->use_empty()); |
3509 | retainCall->eraseFromParent(); |
3510 | eraseUnusedBitCasts(insn: cast<llvm::Instruction>(Val: retainedValue)); |
3511 | |
3512 | return CGF.Builder.CreateBitCast(V: load, DestTy: resultType); |
3513 | } |
3514 | |
3515 | /// Emit an ARC autorelease of the result of a function. |
3516 | /// |
3517 | /// \return the value to actually return from the function |
3518 | static llvm::Value *emitAutoreleaseOfResult(CodeGenFunction &CGF, |
3519 | llvm::Value *result) { |
3520 | // If we're returning 'self', kill the initial retain. This is a |
3521 | // heuristic attempt to "encourage correctness" in the really unfortunate |
3522 | // case where we have a return of self during a dealloc and we desperately |
3523 | // need to avoid the possible autorelease. |
3524 | if (llvm::Value *self = tryRemoveRetainOfSelf(CGF, result)) |
3525 | return self; |
3526 | |
3527 | // At -O0, try to emit a fused retain/autorelease. |
3528 | if (CGF.shouldUseFusedARCCalls()) |
3529 | if (llvm::Value *fused = tryEmitFusedAutoreleaseOfResult(CGF, result)) |
3530 | return fused; |
3531 | |
3532 | return CGF.EmitARCAutoreleaseReturnValue(value: result); |
3533 | } |
3534 | |
3535 | /// Heuristically search for a dominating store to the return-value slot. |
3536 | static llvm::StoreInst *findDominatingStoreToReturnValue(CodeGenFunction &CGF) { |
3537 | // Check if a User is a store which pointerOperand is the ReturnValue. |
3538 | // We are looking for stores to the ReturnValue, not for stores of the |
3539 | // ReturnValue to some other location. |
3540 | auto GetStoreIfValid = [&CGF](llvm::User *U) -> llvm::StoreInst * { |
3541 | auto *SI = dyn_cast<llvm::StoreInst>(Val: U); |
3542 | if (!SI || SI->getPointerOperand() != CGF.ReturnValue.getPointer() || |
3543 | SI->getValueOperand()->getType() != CGF.ReturnValue.getElementType()) |
3544 | return nullptr; |
3545 | // These aren't actually possible for non-coerced returns, and we |
3546 | // only care about non-coerced returns on this code path. |
3547 | // All memory instructions inside __try block are volatile. |
3548 | assert(!SI->isAtomic() && |
3549 | (!SI->isVolatile() || CGF.currentFunctionUsesSEHTry())); |
3550 | return SI; |
3551 | }; |
3552 | // If there are multiple uses of the return-value slot, just check |
3553 | // for something immediately preceding the IP. Sometimes this can |
3554 | // happen with how we generate implicit-returns; it can also happen |
3555 | // with noreturn cleanups. |
3556 | if (!CGF.ReturnValue.getPointer()->hasOneUse()) { |
3557 | llvm::BasicBlock *IP = CGF.Builder.GetInsertBlock(); |
3558 | if (IP->empty()) return nullptr; |
3559 | |
3560 | // Look at directly preceding instruction, skipping bitcasts and lifetime |
3561 | // markers. |
3562 | for (llvm::Instruction &I : make_range(x: IP->rbegin(), y: IP->rend())) { |
3563 | if (isa<llvm::BitCastInst>(Val: &I)) |
3564 | continue; |
3565 | if (auto *II = dyn_cast<llvm::IntrinsicInst>(Val: &I)) |
3566 | if (II->getIntrinsicID() == llvm::Intrinsic::lifetime_end) |
3567 | continue; |
3568 | |
3569 | return GetStoreIfValid(&I); |
3570 | } |
3571 | return nullptr; |
3572 | } |
3573 | |
3574 | llvm::StoreInst *store = |
3575 | GetStoreIfValid(CGF.ReturnValue.getPointer()->user_back()); |
3576 | if (!store) return nullptr; |
3577 | |
3578 | // Now do a first-and-dirty dominance check: just walk up the |
3579 | // single-predecessors chain from the current insertion point. |
3580 | llvm::BasicBlock *StoreBB = store->getParent(); |
3581 | llvm::BasicBlock *IP = CGF.Builder.GetInsertBlock(); |
3582 | llvm::SmallPtrSet<llvm::BasicBlock *, 4> SeenBBs; |
3583 | while (IP != StoreBB) { |
3584 | if (!SeenBBs.insert(Ptr: IP).second || !(IP = IP->getSinglePredecessor())) |
3585 | return nullptr; |
3586 | } |
3587 | |
3588 | // Okay, the store's basic block dominates the insertion point; we |
3589 | // can do our thing. |
3590 | return store; |
3591 | } |
3592 | |
3593 | // Helper functions for EmitCMSEClearRecord |
3594 | |
3595 | // Set the bits corresponding to a field having width `BitWidth` and located at |
3596 | // offset `BitOffset` (from the least significant bit) within a storage unit of |
3597 | // `Bits.size()` bytes. Each element of `Bits` corresponds to one target byte. |
3598 | // Use little-endian layout, i.e.`Bits[0]` is the LSB. |
3599 | static void setBitRange(SmallVectorImpl<uint64_t> &Bits, int BitOffset, |
3600 | int BitWidth, int CharWidth) { |
3601 | assert(CharWidth <= 64); |
3602 | assert(static_cast<unsigned>(BitWidth) <= Bits.size() * CharWidth); |
3603 | |
3604 | int Pos = 0; |
3605 | if (BitOffset >= CharWidth) { |
3606 | Pos += BitOffset / CharWidth; |
3607 | BitOffset = BitOffset % CharWidth; |
3608 | } |
3609 | |
3610 | const uint64_t Used = (uint64_t(1) << CharWidth) - 1; |
3611 | if (BitOffset + BitWidth >= CharWidth) { |
3612 | Bits[Pos++] |= (Used << BitOffset) & Used; |
3613 | BitWidth -= CharWidth - BitOffset; |
3614 | BitOffset = 0; |
3615 | } |
3616 | |
3617 | while (BitWidth >= CharWidth) { |
3618 | Bits[Pos++] = Used; |
3619 | BitWidth -= CharWidth; |
3620 | } |
3621 | |
3622 | if (BitWidth > 0) |
3623 | Bits[Pos++] |= (Used >> (CharWidth - BitWidth)) << BitOffset; |
3624 | } |
3625 | |
3626 | // Set the bits corresponding to a field having width `BitWidth` and located at |
3627 | // offset `BitOffset` (from the least significant bit) within a storage unit of |
3628 | // `StorageSize` bytes, located at `StorageOffset` in `Bits`. Each element of |
3629 | // `Bits` corresponds to one target byte. Use target endian layout. |
3630 | static void setBitRange(SmallVectorImpl<uint64_t> &Bits, int StorageOffset, |
3631 | int StorageSize, int BitOffset, int BitWidth, |
3632 | int CharWidth, bool BigEndian) { |
3633 | |
3634 | SmallVector<uint64_t, 8> TmpBits(StorageSize); |
3635 | setBitRange(Bits&: TmpBits, BitOffset, BitWidth, CharWidth); |
3636 | |
3637 | if (BigEndian) |
3638 | std::reverse(first: TmpBits.begin(), last: TmpBits.end()); |
3639 | |
3640 | for (uint64_t V : TmpBits) |
3641 | Bits[StorageOffset++] |= V; |
3642 | } |
3643 | |
3644 | static void setUsedBits(CodeGenModule &, QualType, int, |
3645 | SmallVectorImpl<uint64_t> &); |
3646 | |
3647 | // Set the bits in `Bits`, which correspond to the value representations of |
3648 | // the actual members of the record type `RTy`. Note that this function does |
3649 | // not handle base classes, virtual tables, etc, since they cannot happen in |
3650 | // CMSE function arguments or return. The bit mask corresponds to the target |
3651 | // memory layout, i.e. it's endian dependent. |
3652 | static void setUsedBits(CodeGenModule &CGM, const RecordType *RTy, int Offset, |
3653 | SmallVectorImpl<uint64_t> &Bits) { |
3654 | ASTContext &Context = CGM.getContext(); |
3655 | int CharWidth = Context.getCharWidth(); |
3656 | const RecordDecl *RD = RTy->getDecl()->getDefinition(); |
3657 | const ASTRecordLayout &ASTLayout = Context.getASTRecordLayout(D: RD); |
3658 | const CGRecordLayout &Layout = CGM.getTypes().getCGRecordLayout(RD); |
3659 | |
3660 | int Idx = 0; |
3661 | for (auto I = RD->field_begin(), E = RD->field_end(); I != E; ++I, ++Idx) { |
3662 | const FieldDecl *F = *I; |
3663 | |
3664 | if (F->isUnnamedBitfield() || F->isZeroLengthBitField(Ctx: Context) || |
3665 | F->getType()->isIncompleteArrayType()) |
3666 | continue; |
3667 | |
3668 | if (F->isBitField()) { |
3669 | const CGBitFieldInfo &BFI = Layout.getBitFieldInfo(FD: F); |
3670 | setBitRange(Bits, StorageOffset: Offset + BFI.StorageOffset.getQuantity(), |
3671 | StorageSize: BFI.StorageSize / CharWidth, BitOffset: BFI.Offset, |
3672 | BitWidth: BFI.Size, CharWidth, |
3673 | BigEndian: CGM.getDataLayout().isBigEndian()); |
3674 | continue; |
3675 | } |
3676 | |
3677 | setUsedBits(CGM, F->getType(), |
3678 | Offset + ASTLayout.getFieldOffset(FieldNo: Idx) / CharWidth, Bits); |
3679 | } |
3680 | } |
3681 | |
3682 | // Set the bits in `Bits`, which correspond to the value representations of |
3683 | // the elements of an array type `ATy`. |
3684 | static void setUsedBits(CodeGenModule &CGM, const ConstantArrayType *ATy, |
3685 | int Offset, SmallVectorImpl<uint64_t> &Bits) { |
3686 | const ASTContext &Context = CGM.getContext(); |
3687 | |
3688 | QualType ETy = Context.getBaseElementType(ATy); |
3689 | int Size = Context.getTypeSizeInChars(T: ETy).getQuantity(); |
3690 | SmallVector<uint64_t, 4> TmpBits(Size); |
3691 | setUsedBits(CGM, ETy, 0, TmpBits); |
3692 | |
3693 | for (int I = 0, N = Context.getConstantArrayElementCount(CA: ATy); I < N; ++I) { |
3694 | auto Src = TmpBits.begin(); |
3695 | auto Dst = Bits.begin() + Offset + I * Size; |
3696 | for (int J = 0; J < Size; ++J) |
3697 | *Dst++ |= *Src++; |
3698 | } |
3699 | } |
3700 | |
3701 | // Set the bits in `Bits`, which correspond to the value representations of |
3702 | // the type `QTy`. |
3703 | static void setUsedBits(CodeGenModule &CGM, QualType QTy, int Offset, |
3704 | SmallVectorImpl<uint64_t> &Bits) { |
3705 | if (const auto *RTy = QTy->getAs<RecordType>()) |
3706 | return setUsedBits(CGM, RTy, Offset, Bits); |
3707 | |
3708 | ASTContext &Context = CGM.getContext(); |
3709 | if (const auto *ATy = Context.getAsConstantArrayType(T: QTy)) |
3710 | return setUsedBits(CGM, ATy, Offset, Bits); |
3711 | |
3712 | int Size = Context.getTypeSizeInChars(T: QTy).getQuantity(); |
3713 | if (Size <= 0) |
3714 | return; |
3715 | |
3716 | std::fill_n(first: Bits.begin() + Offset, n: Size, |
3717 | value: (uint64_t(1) << Context.getCharWidth()) - 1); |
3718 | } |
3719 | |
3720 | static uint64_t buildMultiCharMask(const SmallVectorImpl<uint64_t> &Bits, |
3721 | int Pos, int Size, int CharWidth, |
3722 | bool BigEndian) { |
3723 | assert(Size > 0); |
3724 | uint64_t Mask = 0; |
3725 | if (BigEndian) { |
3726 | for (auto P = Bits.begin() + Pos, E = Bits.begin() + Pos + Size; P != E; |
3727 | ++P) |
3728 | Mask = (Mask << CharWidth) | *P; |
3729 | } else { |
3730 | auto P = Bits.begin() + Pos + Size, End = Bits.begin() + Pos; |
3731 | do |
3732 | Mask = (Mask << CharWidth) | *--P; |
3733 | while (P != End); |
3734 | } |
3735 | return Mask; |
3736 | } |
3737 | |
3738 | // Emit code to clear the bits in a record, which aren't a part of any user |
3739 | // declared member, when the record is a function return. |
3740 | llvm::Value *CodeGenFunction::EmitCMSEClearRecord(llvm::Value *Src, |
3741 | llvm::IntegerType *ITy, |
3742 | QualType QTy) { |
3743 | assert(Src->getType() == ITy); |
3744 | assert(ITy->getScalarSizeInBits() <= 64); |
3745 | |
3746 | const llvm::DataLayout &DataLayout = CGM.getDataLayout(); |
3747 | int Size = DataLayout.getTypeStoreSize(Ty: ITy); |
3748 | SmallVector<uint64_t, 4> Bits(Size); |
3749 | setUsedBits(CGM, RTy: QTy->castAs<RecordType>(), Offset: 0, Bits); |
3750 | |
3751 | int CharWidth = CGM.getContext().getCharWidth(); |
3752 | uint64_t Mask = |
3753 | buildMultiCharMask(Bits, Pos: 0, Size, CharWidth, BigEndian: DataLayout.isBigEndian()); |
3754 | |
3755 | return Builder.CreateAnd(LHS: Src, RHS: Mask, Name: "cmse.clear" ); |
3756 | } |
3757 | |
3758 | // Emit code to clear the bits in a record, which aren't a part of any user |
3759 | // declared member, when the record is a function argument. |
3760 | llvm::Value *CodeGenFunction::EmitCMSEClearRecord(llvm::Value *Src, |
3761 | llvm::ArrayType *ATy, |
3762 | QualType QTy) { |
3763 | const llvm::DataLayout &DataLayout = CGM.getDataLayout(); |
3764 | int Size = DataLayout.getTypeStoreSize(Ty: ATy); |
3765 | SmallVector<uint64_t, 16> Bits(Size); |
3766 | setUsedBits(CGM, RTy: QTy->castAs<RecordType>(), Offset: 0, Bits); |
3767 | |
3768 | // Clear each element of the LLVM array. |
3769 | int CharWidth = CGM.getContext().getCharWidth(); |
3770 | int CharsPerElt = |
3771 | ATy->getArrayElementType()->getScalarSizeInBits() / CharWidth; |
3772 | int MaskIndex = 0; |
3773 | llvm::Value *R = llvm::PoisonValue::get(T: ATy); |
3774 | for (int I = 0, N = ATy->getArrayNumElements(); I != N; ++I) { |
3775 | uint64_t Mask = buildMultiCharMask(Bits, Pos: MaskIndex, Size: CharsPerElt, CharWidth, |
3776 | BigEndian: DataLayout.isBigEndian()); |
3777 | MaskIndex += CharsPerElt; |
3778 | llvm::Value *T0 = Builder.CreateExtractValue(Agg: Src, Idxs: I); |
3779 | llvm::Value *T1 = Builder.CreateAnd(LHS: T0, RHS: Mask, Name: "cmse.clear" ); |
3780 | R = Builder.CreateInsertValue(Agg: R, Val: T1, Idxs: I); |
3781 | } |
3782 | |
3783 | return R; |
3784 | } |
3785 | |
3786 | void CodeGenFunction::EmitFunctionEpilog(const CGFunctionInfo &FI, |
3787 | bool EmitRetDbgLoc, |
3788 | SourceLocation EndLoc) { |
3789 | if (FI.isNoReturn()) { |
3790 | // Noreturn functions don't return. |
3791 | EmitUnreachable(Loc: EndLoc); |
3792 | return; |
3793 | } |
3794 | |
3795 | if (CurCodeDecl && CurCodeDecl->hasAttr<NakedAttr>()) { |
3796 | // Naked functions don't have epilogues. |
3797 | Builder.CreateUnreachable(); |
3798 | return; |
3799 | } |
3800 | |
3801 | // Functions with no result always return void. |
3802 | if (!ReturnValue.isValid()) { |
3803 | Builder.CreateRetVoid(); |
3804 | return; |
3805 | } |
3806 | |
3807 | llvm::DebugLoc RetDbgLoc; |
3808 | llvm::Value *RV = nullptr; |
3809 | QualType RetTy = FI.getReturnType(); |
3810 | const ABIArgInfo &RetAI = FI.getReturnInfo(); |
3811 | |
3812 | switch (RetAI.getKind()) { |
3813 | case ABIArgInfo::InAlloca: |
3814 | // Aggregates get evaluated directly into the destination. Sometimes we |
3815 | // need to return the sret value in a register, though. |
3816 | assert(hasAggregateEvaluationKind(RetTy)); |
3817 | if (RetAI.getInAllocaSRet()) { |
3818 | llvm::Function::arg_iterator EI = CurFn->arg_end(); |
3819 | --EI; |
3820 | llvm::Value *ArgStruct = &*EI; |
3821 | llvm::Value *SRet = Builder.CreateStructGEP( |
3822 | Ty: FI.getArgStruct(), Ptr: ArgStruct, Idx: RetAI.getInAllocaFieldIndex()); |
3823 | llvm::Type *Ty = |
3824 | cast<llvm::GetElementPtrInst>(Val: SRet)->getResultElementType(); |
3825 | RV = Builder.CreateAlignedLoad(Ty, SRet, getPointerAlign(), "sret" ); |
3826 | } |
3827 | break; |
3828 | |
3829 | case ABIArgInfo::Indirect: { |
3830 | auto AI = CurFn->arg_begin(); |
3831 | if (RetAI.isSRetAfterThis()) |
3832 | ++AI; |
3833 | switch (getEvaluationKind(T: RetTy)) { |
3834 | case TEK_Complex: { |
3835 | ComplexPairTy RT = |
3836 | EmitLoadOfComplex(src: MakeAddrLValue(Addr: ReturnValue, T: RetTy), loc: EndLoc); |
3837 | EmitStoreOfComplex(V: RT, dest: MakeNaturalAlignAddrLValue(V: &*AI, T: RetTy), |
3838 | /*isInit*/ true); |
3839 | break; |
3840 | } |
3841 | case TEK_Aggregate: |
3842 | // Do nothing; aggregates get evaluated directly into the destination. |
3843 | break; |
3844 | case TEK_Scalar: { |
3845 | LValueBaseInfo BaseInfo; |
3846 | TBAAAccessInfo TBAAInfo; |
3847 | CharUnits Alignment = |
3848 | CGM.getNaturalTypeAlignment(T: RetTy, BaseInfo: &BaseInfo, TBAAInfo: &TBAAInfo); |
3849 | Address ArgAddr(&*AI, ConvertType(T: RetTy), Alignment); |
3850 | LValue ArgVal = |
3851 | LValue::MakeAddr(address: ArgAddr, type: RetTy, Context&: getContext(), BaseInfo, TBAAInfo); |
3852 | EmitStoreOfScalar( |
3853 | value: Builder.CreateLoad(Addr: ReturnValue), lvalue: ArgVal, /*isInit*/ true); |
3854 | break; |
3855 | } |
3856 | } |
3857 | break; |
3858 | } |
3859 | |
3860 | case ABIArgInfo::Extend: |
3861 | case ABIArgInfo::Direct: |
3862 | if (RetAI.getCoerceToType() == ConvertType(T: RetTy) && |
3863 | RetAI.getDirectOffset() == 0) { |
3864 | // The internal return value temp always will have pointer-to-return-type |
3865 | // type, just do a load. |
3866 | |
3867 | // If there is a dominating store to ReturnValue, we can elide |
3868 | // the load, zap the store, and usually zap the alloca. |
3869 | if (llvm::StoreInst *SI = |
3870 | findDominatingStoreToReturnValue(CGF&: *this)) { |
3871 | // Reuse the debug location from the store unless there is |
3872 | // cleanup code to be emitted between the store and return |
3873 | // instruction. |
3874 | if (EmitRetDbgLoc && !AutoreleaseResult) |
3875 | RetDbgLoc = SI->getDebugLoc(); |
3876 | // Get the stored value and nuke the now-dead store. |
3877 | RV = SI->getValueOperand(); |
3878 | SI->eraseFromParent(); |
3879 | |
3880 | // Otherwise, we have to do a simple load. |
3881 | } else { |
3882 | RV = Builder.CreateLoad(Addr: ReturnValue); |
3883 | } |
3884 | } else { |
3885 | // If the value is offset in memory, apply the offset now. |
3886 | Address V = emitAddressAtOffset(CGF&: *this, addr: ReturnValue, info: RetAI); |
3887 | |
3888 | RV = CreateCoercedLoad(Src: V, Ty: RetAI.getCoerceToType(), CGF&: *this); |
3889 | } |
3890 | |
3891 | // In ARC, end functions that return a retainable type with a call |
3892 | // to objc_autoreleaseReturnValue. |
3893 | if (AutoreleaseResult) { |
3894 | #ifndef NDEBUG |
3895 | // Type::isObjCRetainabletype has to be called on a QualType that hasn't |
3896 | // been stripped of the typedefs, so we cannot use RetTy here. Get the |
3897 | // original return type of FunctionDecl, CurCodeDecl, and BlockDecl from |
3898 | // CurCodeDecl or BlockInfo. |
3899 | QualType RT; |
3900 | |
3901 | if (auto *FD = dyn_cast<FunctionDecl>(Val: CurCodeDecl)) |
3902 | RT = FD->getReturnType(); |
3903 | else if (auto *MD = dyn_cast<ObjCMethodDecl>(Val: CurCodeDecl)) |
3904 | RT = MD->getReturnType(); |
3905 | else if (isa<BlockDecl>(Val: CurCodeDecl)) |
3906 | RT = BlockInfo->BlockExpression->getFunctionType()->getReturnType(); |
3907 | else |
3908 | llvm_unreachable("Unexpected function/method type" ); |
3909 | |
3910 | assert(getLangOpts().ObjCAutoRefCount && |
3911 | !FI.isReturnsRetained() && |
3912 | RT->isObjCRetainableType()); |
3913 | #endif |
3914 | RV = emitAutoreleaseOfResult(CGF&: *this, result: RV); |
3915 | } |
3916 | |
3917 | break; |
3918 | |
3919 | case ABIArgInfo::Ignore: |
3920 | break; |
3921 | |
3922 | case ABIArgInfo::CoerceAndExpand: { |
3923 | auto coercionType = RetAI.getCoerceAndExpandType(); |
3924 | |
3925 | // Load all of the coerced elements out into results. |
3926 | llvm::SmallVector<llvm::Value*, 4> results; |
3927 | Address addr = ReturnValue.withElementType(ElemTy: coercionType); |
3928 | for (unsigned i = 0, e = coercionType->getNumElements(); i != e; ++i) { |
3929 | auto coercedEltType = coercionType->getElementType(N: i); |
3930 | if (ABIArgInfo::isPaddingForCoerceAndExpand(eltType: coercedEltType)) |
3931 | continue; |
3932 | |
3933 | auto eltAddr = Builder.CreateStructGEP(Addr: addr, Index: i); |
3934 | auto elt = Builder.CreateLoad(Addr: eltAddr); |
3935 | results.push_back(Elt: elt); |
3936 | } |
3937 | |
3938 | // If we have one result, it's the single direct result type. |
3939 | if (results.size() == 1) { |
3940 | RV = results[0]; |
3941 | |
3942 | // Otherwise, we need to make a first-class aggregate. |
3943 | } else { |
3944 | // Construct a return type that lacks padding elements. |
3945 | llvm::Type *returnType = RetAI.getUnpaddedCoerceAndExpandType(); |
3946 | |
3947 | RV = llvm::PoisonValue::get(T: returnType); |
3948 | for (unsigned i = 0, e = results.size(); i != e; ++i) { |
3949 | RV = Builder.CreateInsertValue(Agg: RV, Val: results[i], Idxs: i); |
3950 | } |
3951 | } |
3952 | break; |
3953 | } |
3954 | case ABIArgInfo::Expand: |
3955 | case ABIArgInfo::IndirectAliased: |
3956 | llvm_unreachable("Invalid ABI kind for return argument" ); |
3957 | } |
3958 | |
3959 | llvm::Instruction *Ret; |
3960 | if (RV) { |
3961 | if (CurFuncDecl && CurFuncDecl->hasAttr<CmseNSEntryAttr>()) { |
3962 | // For certain return types, clear padding bits, as they may reveal |
3963 | // sensitive information. |
3964 | // Small struct/union types are passed as integers. |
3965 | auto *ITy = dyn_cast<llvm::IntegerType>(Val: RV->getType()); |
3966 | if (ITy != nullptr && isa<RecordType>(Val: RetTy.getCanonicalType())) |
3967 | RV = EmitCMSEClearRecord(Src: RV, ITy, QTy: RetTy); |
3968 | } |
3969 | EmitReturnValueCheck(RV); |
3970 | Ret = Builder.CreateRet(V: RV); |
3971 | } else { |
3972 | Ret = Builder.CreateRetVoid(); |
3973 | } |
3974 | |
3975 | if (RetDbgLoc) |
3976 | Ret->setDebugLoc(std::move(RetDbgLoc)); |
3977 | } |
3978 | |
3979 | void CodeGenFunction::EmitReturnValueCheck(llvm::Value *RV) { |
3980 | // A current decl may not be available when emitting vtable thunks. |
3981 | if (!CurCodeDecl) |
3982 | return; |
3983 | |
3984 | // If the return block isn't reachable, neither is this check, so don't emit |
3985 | // it. |
3986 | if (ReturnBlock.isValid() && ReturnBlock.getBlock()->use_empty()) |
3987 | return; |
3988 | |
3989 | ReturnsNonNullAttr *RetNNAttr = nullptr; |
3990 | if (SanOpts.has(SanitizerKind::ReturnsNonnullAttribute)) |
3991 | RetNNAttr = CurCodeDecl->getAttr<ReturnsNonNullAttr>(); |
3992 | |
3993 | if (!RetNNAttr && !requiresReturnValueNullabilityCheck()) |
3994 | return; |
3995 | |
3996 | // Prefer the returns_nonnull attribute if it's present. |
3997 | SourceLocation AttrLoc; |
3998 | SanitizerMask CheckKind; |
3999 | SanitizerHandler Handler; |
4000 | if (RetNNAttr) { |
4001 | assert(!requiresReturnValueNullabilityCheck() && |
4002 | "Cannot check nullability and the nonnull attribute" ); |
4003 | AttrLoc = RetNNAttr->getLocation(); |
4004 | CheckKind = SanitizerKind::ReturnsNonnullAttribute; |
4005 | Handler = SanitizerHandler::NonnullReturn; |
4006 | } else { |
4007 | if (auto *DD = dyn_cast<DeclaratorDecl>(Val: CurCodeDecl)) |
4008 | if (auto *TSI = DD->getTypeSourceInfo()) |
4009 | if (auto FTL = TSI->getTypeLoc().getAsAdjusted<FunctionTypeLoc>()) |
4010 | AttrLoc = FTL.getReturnLoc().findNullabilityLoc(); |
4011 | CheckKind = SanitizerKind::NullabilityReturn; |
4012 | Handler = SanitizerHandler::NullabilityReturn; |
4013 | } |
4014 | |
4015 | SanitizerScope SanScope(this); |
4016 | |
4017 | // Make sure the "return" source location is valid. If we're checking a |
4018 | // nullability annotation, make sure the preconditions for the check are met. |
4019 | llvm::BasicBlock *Check = createBasicBlock(name: "nullcheck" ); |
4020 | llvm::BasicBlock *NoCheck = createBasicBlock(name: "no.nullcheck" ); |
4021 | llvm::Value *SLocPtr = Builder.CreateLoad(Addr: ReturnLocation, Name: "return.sloc.load" ); |
4022 | llvm::Value *CanNullCheck = Builder.CreateIsNotNull(Arg: SLocPtr); |
4023 | if (requiresReturnValueNullabilityCheck()) |
4024 | CanNullCheck = |
4025 | Builder.CreateAnd(LHS: CanNullCheck, RHS: RetValNullabilityPrecondition); |
4026 | Builder.CreateCondBr(Cond: CanNullCheck, True: Check, False: NoCheck); |
4027 | EmitBlock(BB: Check); |
4028 | |
4029 | // Now do the null check. |
4030 | llvm::Value *Cond = Builder.CreateIsNotNull(Arg: RV); |
4031 | llvm::Constant *StaticData[] = {EmitCheckSourceLocation(Loc: AttrLoc)}; |
4032 | llvm::Value *DynamicData[] = {SLocPtr}; |
4033 | EmitCheck(Checked: std::make_pair(x&: Cond, y&: CheckKind), Check: Handler, StaticArgs: StaticData, DynamicArgs: DynamicData); |
4034 | |
4035 | EmitBlock(BB: NoCheck); |
4036 | |
4037 | #ifndef NDEBUG |
4038 | // The return location should not be used after the check has been emitted. |
4039 | ReturnLocation = Address::invalid(); |
4040 | #endif |
4041 | } |
4042 | |
4043 | static bool isInAllocaArgument(CGCXXABI &ABI, QualType type) { |
4044 | const CXXRecordDecl *RD = type->getAsCXXRecordDecl(); |
4045 | return RD && ABI.getRecordArgABI(RD) == CGCXXABI::RAA_DirectInMemory; |
4046 | } |
4047 | |
4048 | static AggValueSlot createPlaceholderSlot(CodeGenFunction &CGF, |
4049 | QualType Ty) { |
4050 | // FIXME: Generate IR in one pass, rather than going back and fixing up these |
4051 | // placeholders. |
4052 | llvm::Type *IRTy = CGF.ConvertTypeForMem(T: Ty); |
4053 | llvm::Type *IRPtrTy = llvm::PointerType::getUnqual(C&: CGF.getLLVMContext()); |
4054 | llvm::Value *Placeholder = llvm::PoisonValue::get(T: IRPtrTy); |
4055 | |
4056 | // FIXME: When we generate this IR in one pass, we shouldn't need |
4057 | // this win32-specific alignment hack. |
4058 | CharUnits Align = CharUnits::fromQuantity(Quantity: 4); |
4059 | Placeholder = CGF.Builder.CreateAlignedLoad(Ty: IRPtrTy, Addr: Placeholder, Align); |
4060 | |
4061 | return AggValueSlot::forAddr(addr: Address(Placeholder, IRTy, Align), |
4062 | quals: Ty.getQualifiers(), |
4063 | isDestructed: AggValueSlot::IsNotDestructed, |
4064 | needsGC: AggValueSlot::DoesNotNeedGCBarriers, |
4065 | isAliased: AggValueSlot::IsNotAliased, |
4066 | mayOverlap: AggValueSlot::DoesNotOverlap); |
4067 | } |
4068 | |
4069 | void CodeGenFunction::EmitDelegateCallArg(CallArgList &args, |
4070 | const VarDecl *param, |
4071 | SourceLocation loc) { |
4072 | // StartFunction converted the ABI-lowered parameter(s) into a |
4073 | // local alloca. We need to turn that into an r-value suitable |
4074 | // for EmitCall. |
4075 | Address local = GetAddrOfLocalVar(VD: param); |
4076 | |
4077 | QualType type = param->getType(); |
4078 | |
4079 | // GetAddrOfLocalVar returns a pointer-to-pointer for references, |
4080 | // but the argument needs to be the original pointer. |
4081 | if (type->isReferenceType()) { |
4082 | args.add(rvalue: RValue::get(V: Builder.CreateLoad(Addr: local)), type); |
4083 | |
4084 | // In ARC, move out of consumed arguments so that the release cleanup |
4085 | // entered by StartFunction doesn't cause an over-release. This isn't |
4086 | // optimal -O0 code generation, but it should get cleaned up when |
4087 | // optimization is enabled. This also assumes that delegate calls are |
4088 | // performed exactly once for a set of arguments, but that should be safe. |
4089 | } else if (getLangOpts().ObjCAutoRefCount && |
4090 | param->hasAttr<NSConsumedAttr>() && |
4091 | type->isObjCRetainableType()) { |
4092 | llvm::Value *ptr = Builder.CreateLoad(Addr: local); |
4093 | auto null = |
4094 | llvm::ConstantPointerNull::get(T: cast<llvm::PointerType>(Val: ptr->getType())); |
4095 | Builder.CreateStore(Val: null, Addr: local); |
4096 | args.add(rvalue: RValue::get(V: ptr), type); |
4097 | |
4098 | // For the most part, we just need to load the alloca, except that |
4099 | // aggregate r-values are actually pointers to temporaries. |
4100 | } else { |
4101 | args.add(rvalue: convertTempToRValue(addr: local, type, Loc: loc), type); |
4102 | } |
4103 | |
4104 | // Deactivate the cleanup for the callee-destructed param that was pushed. |
4105 | if (type->isRecordType() && !CurFuncIsThunk && |
4106 | type->castAs<RecordType>()->getDecl()->isParamDestroyedInCallee() && |
4107 | param->needsDestruction(Ctx: getContext())) { |
4108 | EHScopeStack::stable_iterator cleanup = |
4109 | CalleeDestructedParamCleanups.lookup(Val: cast<ParmVarDecl>(Val: param)); |
4110 | assert(cleanup.isValid() && |
4111 | "cleanup for callee-destructed param not recorded" ); |
4112 | // This unreachable is a temporary marker which will be removed later. |
4113 | llvm::Instruction *isActive = Builder.CreateUnreachable(); |
4114 | args.addArgCleanupDeactivation(Cleanup: cleanup, IsActiveIP: isActive); |
4115 | } |
4116 | } |
4117 | |
4118 | static bool isProvablyNull(llvm::Value *addr) { |
4119 | return isa<llvm::ConstantPointerNull>(Val: addr); |
4120 | } |
4121 | |
4122 | /// Emit the actual writing-back of a writeback. |
4123 | static void emitWriteback(CodeGenFunction &CGF, |
4124 | const CallArgList::Writeback &writeback) { |
4125 | const LValue &srcLV = writeback.Source; |
4126 | Address srcAddr = srcLV.getAddress(CGF); |
4127 | assert(!isProvablyNull(srcAddr.getPointer()) && |
4128 | "shouldn't have writeback for provably null argument" ); |
4129 | |
4130 | llvm::BasicBlock *contBB = nullptr; |
4131 | |
4132 | // If the argument wasn't provably non-null, we need to null check |
4133 | // before doing the store. |
4134 | bool provablyNonNull = llvm::isKnownNonZero(V: srcAddr.getPointer(), |
4135 | DL: CGF.CGM.getDataLayout()); |
4136 | if (!provablyNonNull) { |
4137 | llvm::BasicBlock *writebackBB = CGF.createBasicBlock(name: "icr.writeback" ); |
4138 | contBB = CGF.createBasicBlock(name: "icr.done" ); |
4139 | |
4140 | llvm::Value *isNull = |
4141 | CGF.Builder.CreateIsNull(Arg: srcAddr.getPointer(), Name: "icr.isnull" ); |
4142 | CGF.Builder.CreateCondBr(Cond: isNull, True: contBB, False: writebackBB); |
4143 | CGF.EmitBlock(BB: writebackBB); |
4144 | } |
4145 | |
4146 | // Load the value to writeback. |
4147 | llvm::Value *value = CGF.Builder.CreateLoad(Addr: writeback.Temporary); |
4148 | |
4149 | // Cast it back, in case we're writing an id to a Foo* or something. |
4150 | value = CGF.Builder.CreateBitCast(V: value, DestTy: srcAddr.getElementType(), |
4151 | Name: "icr.writeback-cast" ); |
4152 | |
4153 | // Perform the writeback. |
4154 | |
4155 | // If we have a "to use" value, it's something we need to emit a use |
4156 | // of. This has to be carefully threaded in: if it's done after the |
4157 | // release it's potentially undefined behavior (and the optimizer |
4158 | // will ignore it), and if it happens before the retain then the |
4159 | // optimizer could move the release there. |
4160 | if (writeback.ToUse) { |
4161 | assert(srcLV.getObjCLifetime() == Qualifiers::OCL_Strong); |
4162 | |
4163 | // Retain the new value. No need to block-copy here: the block's |
4164 | // being passed up the stack. |
4165 | value = CGF.EmitARCRetainNonBlock(value); |
4166 | |
4167 | // Emit the intrinsic use here. |
4168 | CGF.EmitARCIntrinsicUse(values: writeback.ToUse); |
4169 | |
4170 | // Load the old value (primitively). |
4171 | llvm::Value *oldValue = CGF.EmitLoadOfScalar(lvalue: srcLV, Loc: SourceLocation()); |
4172 | |
4173 | // Put the new value in place (primitively). |
4174 | CGF.EmitStoreOfScalar(value, lvalue: srcLV, /*init*/ isInit: false); |
4175 | |
4176 | // Release the old value. |
4177 | CGF.EmitARCRelease(value: oldValue, precise: srcLV.isARCPreciseLifetime()); |
4178 | |
4179 | // Otherwise, we can just do a normal lvalue store. |
4180 | } else { |
4181 | CGF.EmitStoreThroughLValue(Src: RValue::get(V: value), Dst: srcLV); |
4182 | } |
4183 | |
4184 | // Jump to the continuation block. |
4185 | if (!provablyNonNull) |
4186 | CGF.EmitBlock(BB: contBB); |
4187 | } |
4188 | |
4189 | static void emitWritebacks(CodeGenFunction &CGF, |
4190 | const CallArgList &args) { |
4191 | for (const auto &I : args.writebacks()) |
4192 | emitWriteback(CGF, writeback: I); |
4193 | } |
4194 | |
4195 | static void deactivateArgCleanupsBeforeCall(CodeGenFunction &CGF, |
4196 | const CallArgList &CallArgs) { |
4197 | ArrayRef<CallArgList::CallArgCleanup> Cleanups = |
4198 | CallArgs.getCleanupsToDeactivate(); |
4199 | // Iterate in reverse to increase the likelihood of popping the cleanup. |
4200 | for (const auto &I : llvm::reverse(C&: Cleanups)) { |
4201 | CGF.DeactivateCleanupBlock(Cleanup: I.Cleanup, DominatingIP: I.IsActiveIP); |
4202 | I.IsActiveIP->eraseFromParent(); |
4203 | } |
4204 | } |
4205 | |
4206 | static const Expr *maybeGetUnaryAddrOfOperand(const Expr *E) { |
4207 | if (const UnaryOperator *uop = dyn_cast<UnaryOperator>(Val: E->IgnoreParens())) |
4208 | if (uop->getOpcode() == UO_AddrOf) |
4209 | return uop->getSubExpr(); |
4210 | return nullptr; |
4211 | } |
4212 | |
4213 | /// Emit an argument that's being passed call-by-writeback. That is, |
4214 | /// we are passing the address of an __autoreleased temporary; it |
4215 | /// might be copy-initialized with the current value of the given |
4216 | /// address, but it will definitely be copied out of after the call. |
4217 | static void emitWritebackArg(CodeGenFunction &CGF, CallArgList &args, |
4218 | const ObjCIndirectCopyRestoreExpr *CRE) { |
4219 | LValue srcLV; |
4220 | |
4221 | // Make an optimistic effort to emit the address as an l-value. |
4222 | // This can fail if the argument expression is more complicated. |
4223 | if (const Expr *lvExpr = maybeGetUnaryAddrOfOperand(E: CRE->getSubExpr())) { |
4224 | srcLV = CGF.EmitLValue(E: lvExpr); |
4225 | |
4226 | // Otherwise, just emit it as a scalar. |
4227 | } else { |
4228 | Address srcAddr = CGF.EmitPointerWithAlignment(Addr: CRE->getSubExpr()); |
4229 | |
4230 | QualType srcAddrType = |
4231 | CRE->getSubExpr()->getType()->castAs<PointerType>()->getPointeeType(); |
4232 | srcLV = CGF.MakeAddrLValue(Addr: srcAddr, T: srcAddrType); |
4233 | } |
4234 | Address srcAddr = srcLV.getAddress(CGF); |
4235 | |
4236 | // The dest and src types don't necessarily match in LLVM terms |
4237 | // because of the crazy ObjC compatibility rules. |
4238 | |
4239 | llvm::PointerType *destType = |
4240 | cast<llvm::PointerType>(CGF.ConvertType(CRE->getType())); |
4241 | llvm::Type *destElemType = |
4242 | CGF.ConvertTypeForMem(T: CRE->getType()->getPointeeType()); |
4243 | |
4244 | // If the address is a constant null, just pass the appropriate null. |
4245 | if (isProvablyNull(addr: srcAddr.getPointer())) { |
4246 | args.add(rvalue: RValue::get(V: llvm::ConstantPointerNull::get(T: destType)), |
4247 | type: CRE->getType()); |
4248 | return; |
4249 | } |
4250 | |
4251 | // Create the temporary. |
4252 | Address temp = |
4253 | CGF.CreateTempAlloca(destElemType, CGF.getPointerAlign(), "icr.temp" ); |
4254 | // Loading an l-value can introduce a cleanup if the l-value is __weak, |
4255 | // and that cleanup will be conditional if we can't prove that the l-value |
4256 | // isn't null, so we need to register a dominating point so that the cleanups |
4257 | // system will make valid IR. |
4258 | CodeGenFunction::ConditionalEvaluation condEval(CGF); |
4259 | |
4260 | // Zero-initialize it if we're not doing a copy-initialization. |
4261 | bool shouldCopy = CRE->shouldCopy(); |
4262 | if (!shouldCopy) { |
4263 | llvm::Value *null = |
4264 | llvm::ConstantPointerNull::get(T: cast<llvm::PointerType>(Val: destElemType)); |
4265 | CGF.Builder.CreateStore(Val: null, Addr: temp); |
4266 | } |
4267 | |
4268 | llvm::BasicBlock *contBB = nullptr; |
4269 | llvm::BasicBlock *originBB = nullptr; |
4270 | |
4271 | // If the address is *not* known to be non-null, we need to switch. |
4272 | llvm::Value *finalArgument; |
4273 | |
4274 | bool provablyNonNull = llvm::isKnownNonZero(V: srcAddr.getPointer(), |
4275 | DL: CGF.CGM.getDataLayout()); |
4276 | if (provablyNonNull) { |
4277 | finalArgument = temp.getPointer(); |
4278 | } else { |
4279 | llvm::Value *isNull = |
4280 | CGF.Builder.CreateIsNull(Arg: srcAddr.getPointer(), Name: "icr.isnull" ); |
4281 | |
4282 | finalArgument = CGF.Builder.CreateSelect(C: isNull, |
4283 | True: llvm::ConstantPointerNull::get(T: destType), |
4284 | False: temp.getPointer(), Name: "icr.argument" ); |
4285 | |
4286 | // If we need to copy, then the load has to be conditional, which |
4287 | // means we need control flow. |
4288 | if (shouldCopy) { |
4289 | originBB = CGF.Builder.GetInsertBlock(); |
4290 | contBB = CGF.createBasicBlock(name: "icr.cont" ); |
4291 | llvm::BasicBlock *copyBB = CGF.createBasicBlock(name: "icr.copy" ); |
4292 | CGF.Builder.CreateCondBr(Cond: isNull, True: contBB, False: copyBB); |
4293 | CGF.EmitBlock(BB: copyBB); |
4294 | condEval.begin(CGF); |
4295 | } |
4296 | } |
4297 | |
4298 | llvm::Value *valueToUse = nullptr; |
4299 | |
4300 | // Perform a copy if necessary. |
4301 | if (shouldCopy) { |
4302 | RValue srcRV = CGF.EmitLoadOfLValue(V: srcLV, Loc: SourceLocation()); |
4303 | assert(srcRV.isScalar()); |
4304 | |
4305 | llvm::Value *src = srcRV.getScalarVal(); |
4306 | src = CGF.Builder.CreateBitCast(V: src, DestTy: destElemType, Name: "icr.cast" ); |
4307 | |
4308 | // Use an ordinary store, not a store-to-lvalue. |
4309 | CGF.Builder.CreateStore(Val: src, Addr: temp); |
4310 | |
4311 | // If optimization is enabled, and the value was held in a |
4312 | // __strong variable, we need to tell the optimizer that this |
4313 | // value has to stay alive until we're doing the store back. |
4314 | // This is because the temporary is effectively unretained, |
4315 | // and so otherwise we can violate the high-level semantics. |
4316 | if (CGF.CGM.getCodeGenOpts().OptimizationLevel != 0 && |
4317 | srcLV.getObjCLifetime() == Qualifiers::OCL_Strong) { |
4318 | valueToUse = src; |
4319 | } |
4320 | } |
4321 | |
4322 | // Finish the control flow if we needed it. |
4323 | if (shouldCopy && !provablyNonNull) { |
4324 | llvm::BasicBlock *copyBB = CGF.Builder.GetInsertBlock(); |
4325 | CGF.EmitBlock(BB: contBB); |
4326 | |
4327 | // Make a phi for the value to intrinsically use. |
4328 | if (valueToUse) { |
4329 | llvm::PHINode *phiToUse = CGF.Builder.CreatePHI(Ty: valueToUse->getType(), NumReservedValues: 2, |
4330 | Name: "icr.to-use" ); |
4331 | phiToUse->addIncoming(V: valueToUse, BB: copyBB); |
4332 | phiToUse->addIncoming(V: llvm::UndefValue::get(T: valueToUse->getType()), |
4333 | BB: originBB); |
4334 | valueToUse = phiToUse; |
4335 | } |
4336 | |
4337 | condEval.end(CGF); |
4338 | } |
4339 | |
4340 | args.addWriteback(srcLV, temporary: temp, toUse: valueToUse); |
4341 | args.add(rvalue: RValue::get(V: finalArgument), type: CRE->getType()); |
4342 | } |
4343 | |
4344 | void CallArgList::allocateArgumentMemory(CodeGenFunction &CGF) { |
4345 | assert(!StackBase); |
4346 | |
4347 | // Save the stack. |
4348 | StackBase = CGF.Builder.CreateStackSave(Name: "inalloca.save" ); |
4349 | } |
4350 | |
4351 | void CallArgList::freeArgumentMemory(CodeGenFunction &CGF) const { |
4352 | if (StackBase) { |
4353 | // Restore the stack after the call. |
4354 | CGF.Builder.CreateStackRestore(Ptr: StackBase); |
4355 | } |
4356 | } |
4357 | |
4358 | void CodeGenFunction::EmitNonNullArgCheck(RValue RV, QualType ArgType, |
4359 | SourceLocation ArgLoc, |
4360 | AbstractCallee AC, |
4361 | unsigned ParmNum) { |
4362 | if (!AC.getDecl() || !(SanOpts.has(K: SanitizerKind::NonnullAttribute) || |
4363 | SanOpts.has(K: SanitizerKind::NullabilityArg))) |
4364 | return; |
4365 | |
4366 | // The param decl may be missing in a variadic function. |
4367 | auto PVD = ParmNum < AC.getNumParams() ? AC.getParamDecl(I: ParmNum) : nullptr; |
4368 | unsigned ArgNo = PVD ? PVD->getFunctionScopeIndex() : ParmNum; |
4369 | |
4370 | // Prefer the nonnull attribute if it's present. |
4371 | const NonNullAttr *NNAttr = nullptr; |
4372 | if (SanOpts.has(SanitizerKind::NonnullAttribute)) |
4373 | NNAttr = getNonNullAttr(AC.getDecl(), PVD, ArgType, ArgNo); |
4374 | |
4375 | bool CanCheckNullability = false; |
4376 | if (SanOpts.has(K: SanitizerKind::NullabilityArg) && !NNAttr && PVD) { |
4377 | auto Nullability = PVD->getType()->getNullability(); |
4378 | CanCheckNullability = Nullability && |
4379 | *Nullability == NullabilityKind::NonNull && |
4380 | PVD->getTypeSourceInfo(); |
4381 | } |
4382 | |
4383 | if (!NNAttr && !CanCheckNullability) |
4384 | return; |
4385 | |
4386 | SourceLocation AttrLoc; |
4387 | SanitizerMask CheckKind; |
4388 | SanitizerHandler Handler; |
4389 | if (NNAttr) { |
4390 | AttrLoc = NNAttr->getLocation(); |
4391 | CheckKind = SanitizerKind::NonnullAttribute; |
4392 | Handler = SanitizerHandler::NonnullArg; |
4393 | } else { |
4394 | AttrLoc = PVD->getTypeSourceInfo()->getTypeLoc().findNullabilityLoc(); |
4395 | CheckKind = SanitizerKind::NullabilityArg; |
4396 | Handler = SanitizerHandler::NullabilityArg; |
4397 | } |
4398 | |
4399 | SanitizerScope SanScope(this); |
4400 | llvm::Value *Cond = EmitNonNullRValueCheck(RV, T: ArgType); |
4401 | llvm::Constant *StaticData[] = { |
4402 | EmitCheckSourceLocation(Loc: ArgLoc), EmitCheckSourceLocation(Loc: AttrLoc), |
4403 | llvm::ConstantInt::get(Ty: Int32Ty, V: ArgNo + 1), |
4404 | }; |
4405 | EmitCheck(Checked: std::make_pair(x&: Cond, y&: CheckKind), Check: Handler, StaticArgs: StaticData, DynamicArgs: std::nullopt); |
4406 | } |
4407 | |
4408 | // Check if the call is going to use the inalloca convention. This needs to |
4409 | // agree with CGFunctionInfo::usesInAlloca. The CGFunctionInfo is arranged |
4410 | // later, so we can't check it directly. |
4411 | static bool hasInAllocaArgs(CodeGenModule &CGM, CallingConv ExplicitCC, |
4412 | ArrayRef<QualType> ArgTypes) { |
4413 | // The Swift calling conventions don't go through the target-specific |
4414 | // argument classification, they never use inalloca. |
4415 | // TODO: Consider limiting inalloca use to only calling conventions supported |
4416 | // by MSVC. |
4417 | if (ExplicitCC == CC_Swift || ExplicitCC == CC_SwiftAsync) |
4418 | return false; |
4419 | if (!CGM.getTarget().getCXXABI().isMicrosoft()) |
4420 | return false; |
4421 | return llvm::any_of(Range&: ArgTypes, P: [&](QualType Ty) { |
4422 | return isInAllocaArgument(ABI&: CGM.getCXXABI(), type: Ty); |
4423 | }); |
4424 | } |
4425 | |
4426 | #ifndef NDEBUG |
4427 | // Determine whether the given argument is an Objective-C method |
4428 | // that may have type parameters in its signature. |
4429 | static bool isObjCMethodWithTypeParams(const ObjCMethodDecl *method) { |
4430 | const DeclContext *dc = method->getDeclContext(); |
4431 | if (const ObjCInterfaceDecl *classDecl = dyn_cast<ObjCInterfaceDecl>(Val: dc)) { |
4432 | return classDecl->getTypeParamListAsWritten(); |
4433 | } |
4434 | |
4435 | if (const ObjCCategoryDecl *catDecl = dyn_cast<ObjCCategoryDecl>(Val: dc)) { |
4436 | return catDecl->getTypeParamList(); |
4437 | } |
4438 | |
4439 | return false; |
4440 | } |
4441 | #endif |
4442 | |
4443 | /// EmitCallArgs - Emit call arguments for a function. |
4444 | void CodeGenFunction::EmitCallArgs( |
4445 | CallArgList &Args, PrototypeWrapper Prototype, |
4446 | llvm::iterator_range<CallExpr::const_arg_iterator> ArgRange, |
4447 | AbstractCallee AC, unsigned ParamsToSkip, EvaluationOrder Order) { |
4448 | SmallVector<QualType, 16> ArgTypes; |
4449 | |
4450 | assert((ParamsToSkip == 0 || Prototype.P) && |
4451 | "Can't skip parameters if type info is not provided" ); |
4452 | |
4453 | // This variable only captures *explicitly* written conventions, not those |
4454 | // applied by default via command line flags or target defaults, such as |
4455 | // thiscall, aapcs, stdcall via -mrtd, etc. Computing that correctly would |
4456 | // require knowing if this is a C++ instance method or being able to see |
4457 | // unprototyped FunctionTypes. |
4458 | CallingConv ExplicitCC = CC_C; |
4459 | |
4460 | // First, if a prototype was provided, use those argument types. |
4461 | bool IsVariadic = false; |
4462 | if (Prototype.P) { |
4463 | const auto *MD = Prototype.P.dyn_cast<const ObjCMethodDecl *>(); |
4464 | if (MD) { |
4465 | IsVariadic = MD->isVariadic(); |
4466 | ExplicitCC = getCallingConventionForDecl( |
4467 | D: MD, IsWindows: CGM.getTarget().getTriple().isOSWindows()); |
4468 | ArgTypes.assign(in_start: MD->param_type_begin() + ParamsToSkip, |
4469 | in_end: MD->param_type_end()); |
4470 | } else { |
4471 | const auto *FPT = Prototype.P.get<const FunctionProtoType *>(); |
4472 | IsVariadic = FPT->isVariadic(); |
4473 | ExplicitCC = FPT->getExtInfo().getCC(); |
4474 | ArgTypes.assign(in_start: FPT->param_type_begin() + ParamsToSkip, |
4475 | in_end: FPT->param_type_end()); |
4476 | } |
4477 | |
4478 | #ifndef NDEBUG |
4479 | // Check that the prototyped types match the argument expression types. |
4480 | bool isGenericMethod = MD && isObjCMethodWithTypeParams(method: MD); |
4481 | CallExpr::const_arg_iterator Arg = ArgRange.begin(); |
4482 | for (QualType Ty : ArgTypes) { |
4483 | assert(Arg != ArgRange.end() && "Running over edge of argument list!" ); |
4484 | assert( |
4485 | (isGenericMethod || Ty->isVariablyModifiedType() || |
4486 | Ty.getNonReferenceType()->isObjCRetainableType() || |
4487 | getContext() |
4488 | .getCanonicalType(Ty.getNonReferenceType()) |
4489 | .getTypePtr() == |
4490 | getContext().getCanonicalType((*Arg)->getType()).getTypePtr()) && |
4491 | "type mismatch in call argument!" ); |
4492 | ++Arg; |
4493 | } |
4494 | |
4495 | // Either we've emitted all the call args, or we have a call to variadic |
4496 | // function. |
4497 | assert((Arg == ArgRange.end() || IsVariadic) && |
4498 | "Extra arguments in non-variadic function!" ); |
4499 | #endif |
4500 | } |
4501 | |
4502 | // If we still have any arguments, emit them using the type of the argument. |
4503 | for (auto *A : llvm::drop_begin(RangeOrContainer&: ArgRange, N: ArgTypes.size())) |
4504 | ArgTypes.push_back(Elt: IsVariadic ? getVarArgType(A) : A->getType()); |
4505 | assert((int)ArgTypes.size() == (ArgRange.end() - ArgRange.begin())); |
4506 | |
4507 | // We must evaluate arguments from right to left in the MS C++ ABI, |
4508 | // because arguments are destroyed left to right in the callee. As a special |
4509 | // case, there are certain language constructs that require left-to-right |
4510 | // evaluation, and in those cases we consider the evaluation order requirement |
4511 | // to trump the "destruction order is reverse construction order" guarantee. |
4512 | bool LeftToRight = |
4513 | CGM.getTarget().getCXXABI().areArgsDestroyedLeftToRightInCallee() |
4514 | ? Order == EvaluationOrder::ForceLeftToRight |
4515 | : Order != EvaluationOrder::ForceRightToLeft; |
4516 | |
4517 | auto MaybeEmitImplicitObjectSize = [&](unsigned I, const Expr *Arg, |
4518 | RValue EmittedArg) { |
4519 | if (!AC.hasFunctionDecl() || I >= AC.getNumParams()) |
4520 | return; |
4521 | auto *PS = AC.getParamDecl(I)->getAttr<PassObjectSizeAttr>(); |
4522 | if (PS == nullptr) |
4523 | return; |
4524 | |
4525 | const auto &Context = getContext(); |
4526 | auto SizeTy = Context.getSizeType(); |
4527 | auto T = Builder.getIntNTy(N: Context.getTypeSize(T: SizeTy)); |
4528 | assert(EmittedArg.getScalarVal() && "We emitted nothing for the arg?" ); |
4529 | llvm::Value *V = evaluateOrEmitBuiltinObjectSize(E: Arg, Type: PS->getType(), ResType: T, |
4530 | EmittedE: EmittedArg.getScalarVal(), |
4531 | IsDynamic: PS->isDynamic()); |
4532 | Args.add(rvalue: RValue::get(V), type: SizeTy); |
4533 | // If we're emitting args in reverse, be sure to do so with |
4534 | // pass_object_size, as well. |
4535 | if (!LeftToRight) |
4536 | std::swap(a&: Args.back(), b&: *(&Args.back() - 1)); |
4537 | }; |
4538 | |
4539 | // Insert a stack save if we're going to need any inalloca args. |
4540 | if (hasInAllocaArgs(CGM, ExplicitCC, ArgTypes)) { |
4541 | assert(getTarget().getTriple().getArch() == llvm::Triple::x86 && |
4542 | "inalloca only supported on x86" ); |
4543 | Args.allocateArgumentMemory(CGF&: *this); |
4544 | } |
4545 | |
4546 | // Evaluate each argument in the appropriate order. |
4547 | size_t CallArgsStart = Args.size(); |
4548 | for (unsigned I = 0, E = ArgTypes.size(); I != E; ++I) { |
4549 | unsigned Idx = LeftToRight ? I : E - I - 1; |
4550 | CallExpr::const_arg_iterator Arg = ArgRange.begin() + Idx; |
4551 | unsigned InitialArgSize = Args.size(); |
4552 | // If *Arg is an ObjCIndirectCopyRestoreExpr, check that either the types of |
4553 | // the argument and parameter match or the objc method is parameterized. |
4554 | assert((!isa<ObjCIndirectCopyRestoreExpr>(*Arg) || |
4555 | getContext().hasSameUnqualifiedType((*Arg)->getType(), |
4556 | ArgTypes[Idx]) || |
4557 | (isa<ObjCMethodDecl>(AC.getDecl()) && |
4558 | isObjCMethodWithTypeParams(cast<ObjCMethodDecl>(AC.getDecl())))) && |
4559 | "Argument and parameter types don't match" ); |
4560 | EmitCallArg(Args, *Arg, ArgTypes[Idx]); |
4561 | // In particular, we depend on it being the last arg in Args, and the |
4562 | // objectsize bits depend on there only being one arg if !LeftToRight. |
4563 | assert(InitialArgSize + 1 == Args.size() && |
4564 | "The code below depends on only adding one arg per EmitCallArg" ); |
4565 | (void)InitialArgSize; |
4566 | // Since pointer argument are never emitted as LValue, it is safe to emit |
4567 | // non-null argument check for r-value only. |
4568 | if (!Args.back().hasLValue()) { |
4569 | RValue RVArg = Args.back().getKnownRValue(); |
4570 | EmitNonNullArgCheck(RV: RVArg, ArgType: ArgTypes[Idx], ArgLoc: (*Arg)->getExprLoc(), AC, |
4571 | ParmNum: ParamsToSkip + Idx); |
4572 | // @llvm.objectsize should never have side-effects and shouldn't need |
4573 | // destruction/cleanups, so we can safely "emit" it after its arg, |
4574 | // regardless of right-to-leftness |
4575 | MaybeEmitImplicitObjectSize(Idx, *Arg, RVArg); |
4576 | } |
4577 | } |
4578 | |
4579 | if (!LeftToRight) { |
4580 | // Un-reverse the arguments we just evaluated so they match up with the LLVM |
4581 | // IR function. |
4582 | std::reverse(first: Args.begin() + CallArgsStart, last: Args.end()); |
4583 | } |
4584 | } |
4585 | |
4586 | namespace { |
4587 | |
4588 | struct DestroyUnpassedArg final : EHScopeStack::Cleanup { |
4589 | DestroyUnpassedArg(Address Addr, QualType Ty) |
4590 | : Addr(Addr), Ty(Ty) {} |
4591 | |
4592 | Address Addr; |
4593 | QualType Ty; |
4594 | |
4595 | void Emit(CodeGenFunction &CGF, Flags flags) override { |
4596 | QualType::DestructionKind DtorKind = Ty.isDestructedType(); |
4597 | if (DtorKind == QualType::DK_cxx_destructor) { |
4598 | const CXXDestructorDecl *Dtor = Ty->getAsCXXRecordDecl()->getDestructor(); |
4599 | assert(!Dtor->isTrivial()); |
4600 | CGF.EmitCXXDestructorCall(Dtor, Dtor_Complete, /*for vbase*/ false, |
4601 | /*Delegating=*/false, Addr, Ty); |
4602 | } else { |
4603 | CGF.callCStructDestructor(CGF.MakeAddrLValue(Addr, Ty)); |
4604 | } |
4605 | } |
4606 | }; |
4607 | |
4608 | struct DisableDebugLocationUpdates { |
4609 | CodeGenFunction &CGF; |
4610 | bool disabledDebugInfo; |
4611 | DisableDebugLocationUpdates(CodeGenFunction &CGF, const Expr *E) : CGF(CGF) { |
4612 | if ((disabledDebugInfo = isa<CXXDefaultArgExpr>(Val: E) && CGF.getDebugInfo())) |
4613 | CGF.disableDebugInfo(); |
4614 | } |
4615 | ~DisableDebugLocationUpdates() { |
4616 | if (disabledDebugInfo) |
4617 | CGF.enableDebugInfo(); |
4618 | } |
4619 | }; |
4620 | |
4621 | } // end anonymous namespace |
4622 | |
4623 | RValue CallArg::getRValue(CodeGenFunction &CGF) const { |
4624 | if (!HasLV) |
4625 | return RV; |
4626 | LValue Copy = CGF.MakeAddrLValue(CGF.CreateMemTemp(Ty), Ty); |
4627 | CGF.EmitAggregateCopy(Copy, LV, Ty, AggValueSlot::DoesNotOverlap, |
4628 | LV.isVolatile()); |
4629 | IsUsed = true; |
4630 | return RValue::getAggregate(addr: Copy.getAddress(CGF)); |
4631 | } |
4632 | |
4633 | void CallArg::copyInto(CodeGenFunction &CGF, Address Addr) const { |
4634 | LValue Dst = CGF.MakeAddrLValue(Addr, Ty); |
4635 | if (!HasLV && RV.isScalar()) |
4636 | CGF.EmitStoreOfScalar(RV.getScalarVal(), Dst, /*isInit=*/true); |
4637 | else if (!HasLV && RV.isComplex()) |
4638 | CGF.EmitStoreOfComplex(RV.getComplexVal(), Dst, /*init=*/true); |
4639 | else { |
4640 | auto Addr = HasLV ? LV.getAddress(CGF) : RV.getAggregateAddress(); |
4641 | LValue SrcLV = CGF.MakeAddrLValue(Addr, Ty); |
4642 | // We assume that call args are never copied into subobjects. |
4643 | CGF.EmitAggregateCopy(Dst, SrcLV, Ty, AggValueSlot::DoesNotOverlap, |
4644 | HasLV ? LV.isVolatileQualified() |
4645 | : RV.isVolatileQualified()); |
4646 | } |
4647 | IsUsed = true; |
4648 | } |
4649 | |
4650 | void CodeGenFunction::EmitCallArg(CallArgList &args, const Expr *E, |
4651 | QualType type) { |
4652 | DisableDebugLocationUpdates Dis(*this, E); |
4653 | if (const ObjCIndirectCopyRestoreExpr *CRE |
4654 | = dyn_cast<ObjCIndirectCopyRestoreExpr>(Val: E)) { |
4655 | assert(getLangOpts().ObjCAutoRefCount); |
4656 | return emitWritebackArg(CGF&: *this, args, CRE); |
4657 | } |
4658 | |
4659 | assert(type->isReferenceType() == E->isGLValue() && |
4660 | "reference binding to unmaterialized r-value!" ); |
4661 | |
4662 | if (E->isGLValue()) { |
4663 | assert(E->getObjectKind() == OK_Ordinary); |
4664 | return args.add(rvalue: EmitReferenceBindingToExpr(E), type); |
4665 | } |
4666 | |
4667 | bool HasAggregateEvalKind = hasAggregateEvaluationKind(T: type); |
4668 | |
4669 | // In the Microsoft C++ ABI, aggregate arguments are destructed by the callee. |
4670 | // However, we still have to push an EH-only cleanup in case we unwind before |
4671 | // we make it to the call. |
4672 | if (type->isRecordType() && |
4673 | type->castAs<RecordType>()->getDecl()->isParamDestroyedInCallee()) { |
4674 | // If we're using inalloca, use the argument memory. Otherwise, use a |
4675 | // temporary. |
4676 | AggValueSlot Slot = args.isUsingInAlloca() |
4677 | ? createPlaceholderSlot(CGF&: *this, Ty: type) : CreateAggTemp(T: type, Name: "agg.tmp" ); |
4678 | |
4679 | bool DestroyedInCallee = true, NeedsEHCleanup = true; |
4680 | if (const auto *RD = type->getAsCXXRecordDecl()) |
4681 | DestroyedInCallee = RD->hasNonTrivialDestructor(); |
4682 | else |
4683 | NeedsEHCleanup = needsEHCleanup(kind: type.isDestructedType()); |
4684 | |
4685 | if (DestroyedInCallee) |
4686 | Slot.setExternallyDestructed(); |
4687 | |
4688 | EmitAggExpr(E, AS: Slot); |
4689 | RValue RV = Slot.asRValue(); |
4690 | args.add(rvalue: RV, type); |
4691 | |
4692 | if (DestroyedInCallee && NeedsEHCleanup) { |
4693 | // Create a no-op GEP between the placeholder and the cleanup so we can |
4694 | // RAUW it successfully. It also serves as a marker of the first |
4695 | // instruction where the cleanup is active. |
4696 | pushFullExprCleanup<DestroyUnpassedArg>(kind: EHCleanup, A: Slot.getAddress(), |
4697 | A: type); |
4698 | // This unreachable is a temporary marker which will be removed later. |
4699 | llvm::Instruction *IsActive = Builder.CreateUnreachable(); |
4700 | args.addArgCleanupDeactivation(Cleanup: EHStack.stable_begin(), IsActiveIP: IsActive); |
4701 | } |
4702 | return; |
4703 | } |
4704 | |
4705 | if (HasAggregateEvalKind && isa<ImplicitCastExpr>(Val: E) && |
4706 | cast<CastExpr>(Val: E)->getCastKind() == CK_LValueToRValue) { |
4707 | LValue L = EmitLValue(E: cast<CastExpr>(Val: E)->getSubExpr()); |
4708 | assert(L.isSimple()); |
4709 | args.addUncopiedAggregate(LV: L, type); |
4710 | return; |
4711 | } |
4712 | |
4713 | args.add(rvalue: EmitAnyExprToTemp(E), type); |
4714 | } |
4715 | |
4716 | QualType CodeGenFunction::getVarArgType(const Expr *Arg) { |
4717 | // System headers on Windows define NULL to 0 instead of 0LL on Win64. MSVC |
4718 | // implicitly widens null pointer constants that are arguments to varargs |
4719 | // functions to pointer-sized ints. |
4720 | if (!getTarget().getTriple().isOSWindows()) |
4721 | return Arg->getType(); |
4722 | |
4723 | if (Arg->getType()->isIntegerType() && |
4724 | getContext().getTypeSize(T: Arg->getType()) < |
4725 | getContext().getTargetInfo().getPointerWidth(AddrSpace: LangAS::Default) && |
4726 | Arg->isNullPointerConstant(Ctx&: getContext(), |
4727 | NPC: Expr::NPC_ValueDependentIsNotNull)) { |
4728 | return getContext().getIntPtrType(); |
4729 | } |
4730 | |
4731 | return Arg->getType(); |
4732 | } |
4733 | |
4734 | // In ObjC ARC mode with no ObjC ARC exception safety, tell the ARC |
4735 | // optimizer it can aggressively ignore unwind edges. |
4736 | void |
4737 | CodeGenFunction::AddObjCARCExceptionMetadata(llvm::Instruction *Inst) { |
4738 | if (CGM.getCodeGenOpts().OptimizationLevel != 0 && |
4739 | !CGM.getCodeGenOpts().ObjCAutoRefCountExceptions) |
4740 | Inst->setMetadata(Kind: "clang.arc.no_objc_arc_exceptions" , |
4741 | Node: CGM.getNoObjCARCExceptionsMetadata()); |
4742 | } |
4743 | |
4744 | /// Emits a call to the given no-arguments nounwind runtime function. |
4745 | llvm::CallInst * |
4746 | CodeGenFunction::EmitNounwindRuntimeCall(llvm::FunctionCallee callee, |
4747 | const llvm::Twine &name) { |
4748 | return EmitNounwindRuntimeCall(callee, args: std::nullopt, name); |
4749 | } |
4750 | |
4751 | /// Emits a call to the given nounwind runtime function. |
4752 | llvm::CallInst * |
4753 | CodeGenFunction::EmitNounwindRuntimeCall(llvm::FunctionCallee callee, |
4754 | ArrayRef<llvm::Value *> args, |
4755 | const llvm::Twine &name) { |
4756 | llvm::CallInst *call = EmitRuntimeCall(callee, args, name); |
4757 | call->setDoesNotThrow(); |
4758 | return call; |
4759 | } |
4760 | |
4761 | /// Emits a simple call (never an invoke) to the given no-arguments |
4762 | /// runtime function. |
4763 | llvm::CallInst *CodeGenFunction::EmitRuntimeCall(llvm::FunctionCallee callee, |
4764 | const llvm::Twine &name) { |
4765 | return EmitRuntimeCall(callee, args: std::nullopt, name); |
4766 | } |
4767 | |
4768 | // Calls which may throw must have operand bundles indicating which funclet |
4769 | // they are nested within. |
4770 | SmallVector<llvm::OperandBundleDef, 1> |
4771 | CodeGenFunction::getBundlesForFunclet(llvm::Value *Callee) { |
4772 | // There is no need for a funclet operand bundle if we aren't inside a |
4773 | // funclet. |
4774 | if (!CurrentFuncletPad) |
4775 | return (SmallVector<llvm::OperandBundleDef, 1>()); |
4776 | |
4777 | // Skip intrinsics which cannot throw (as long as they don't lower into |
4778 | // regular function calls in the course of IR transformations). |
4779 | if (auto *CalleeFn = dyn_cast<llvm::Function>(Val: Callee->stripPointerCasts())) { |
4780 | if (CalleeFn->isIntrinsic() && CalleeFn->doesNotThrow()) { |
4781 | auto IID = CalleeFn->getIntrinsicID(); |
4782 | if (!llvm::IntrinsicInst::mayLowerToFunctionCall(IID)) |
4783 | return (SmallVector<llvm::OperandBundleDef, 1>()); |
4784 | } |
4785 | } |
4786 | |
4787 | SmallVector<llvm::OperandBundleDef, 1> BundleList; |
4788 | BundleList.emplace_back(Args: "funclet" , Args&: CurrentFuncletPad); |
4789 | return BundleList; |
4790 | } |
4791 | |
4792 | /// Emits a simple call (never an invoke) to the given runtime function. |
4793 | llvm::CallInst *CodeGenFunction::EmitRuntimeCall(llvm::FunctionCallee callee, |
4794 | ArrayRef<llvm::Value *> args, |
4795 | const llvm::Twine &name) { |
4796 | llvm::CallInst *call = Builder.CreateCall( |
4797 | Callee: callee, Args: args, OpBundles: getBundlesForFunclet(Callee: callee.getCallee()), Name: name); |
4798 | call->setCallingConv(getRuntimeCC()); |
4799 | return call; |
4800 | } |
4801 | |
4802 | /// Emits a call or invoke to the given noreturn runtime function. |
4803 | void CodeGenFunction::EmitNoreturnRuntimeCallOrInvoke( |
4804 | llvm::FunctionCallee callee, ArrayRef<llvm::Value *> args) { |
4805 | SmallVector<llvm::OperandBundleDef, 1> BundleList = |
4806 | getBundlesForFunclet(Callee: callee.getCallee()); |
4807 | |
4808 | if (getInvokeDest()) { |
4809 | llvm::InvokeInst *invoke = |
4810 | Builder.CreateInvoke(Callee: callee, |
4811 | NormalDest: getUnreachableBlock(), |
4812 | UnwindDest: getInvokeDest(), |
4813 | Args: args, |
4814 | OpBundles: BundleList); |
4815 | invoke->setDoesNotReturn(); |
4816 | invoke->setCallingConv(getRuntimeCC()); |
4817 | } else { |
4818 | llvm::CallInst *call = Builder.CreateCall(Callee: callee, Args: args, OpBundles: BundleList); |
4819 | call->setDoesNotReturn(); |
4820 | call->setCallingConv(getRuntimeCC()); |
4821 | Builder.CreateUnreachable(); |
4822 | } |
4823 | } |
4824 | |
4825 | /// Emits a call or invoke instruction to the given nullary runtime function. |
4826 | llvm::CallBase * |
4827 | CodeGenFunction::EmitRuntimeCallOrInvoke(llvm::FunctionCallee callee, |
4828 | const Twine &name) { |
4829 | return EmitRuntimeCallOrInvoke(callee, args: std::nullopt, name); |
4830 | } |
4831 | |
4832 | /// Emits a call or invoke instruction to the given runtime function. |
4833 | llvm::CallBase * |
4834 | CodeGenFunction::EmitRuntimeCallOrInvoke(llvm::FunctionCallee callee, |
4835 | ArrayRef<llvm::Value *> args, |
4836 | const Twine &name) { |
4837 | llvm::CallBase *call = EmitCallOrInvoke(Callee: callee, Args: args, Name: name); |
4838 | call->setCallingConv(getRuntimeCC()); |
4839 | return call; |
4840 | } |
4841 | |
4842 | /// Emits a call or invoke instruction to the given function, depending |
4843 | /// on the current state of the EH stack. |
4844 | llvm::CallBase *CodeGenFunction::EmitCallOrInvoke(llvm::FunctionCallee Callee, |
4845 | ArrayRef<llvm::Value *> Args, |
4846 | const Twine &Name) { |
4847 | llvm::BasicBlock *InvokeDest = getInvokeDest(); |
4848 | SmallVector<llvm::OperandBundleDef, 1> BundleList = |
4849 | getBundlesForFunclet(Callee: Callee.getCallee()); |
4850 | |
4851 | llvm::CallBase *Inst; |
4852 | if (!InvokeDest) |
4853 | Inst = Builder.CreateCall(Callee, Args, OpBundles: BundleList, Name); |
4854 | else { |
4855 | llvm::BasicBlock *ContBB = createBasicBlock(name: "invoke.cont" ); |
4856 | Inst = Builder.CreateInvoke(Callee, NormalDest: ContBB, UnwindDest: InvokeDest, Args, OpBundles: BundleList, |
4857 | Name); |
4858 | EmitBlock(BB: ContBB); |
4859 | } |
4860 | |
4861 | // In ObjC ARC mode with no ObjC ARC exception safety, tell the ARC |
4862 | // optimizer it can aggressively ignore unwind edges. |
4863 | if (CGM.getLangOpts().ObjCAutoRefCount) |
4864 | AddObjCARCExceptionMetadata(Inst); |
4865 | |
4866 | return Inst; |
4867 | } |
4868 | |
4869 | void CodeGenFunction::deferPlaceholderReplacement(llvm::Instruction *Old, |
4870 | llvm::Value *New) { |
4871 | DeferredReplacements.push_back( |
4872 | Elt: std::make_pair(x: llvm::WeakTrackingVH(Old), y&: New)); |
4873 | } |
4874 | |
4875 | namespace { |
4876 | |
4877 | /// Specify given \p NewAlign as the alignment of return value attribute. If |
4878 | /// such attribute already exists, re-set it to the maximal one of two options. |
4879 | [[nodiscard]] llvm::AttributeList |
4880 | maybeRaiseRetAlignmentAttribute(llvm::LLVMContext &Ctx, |
4881 | const llvm::AttributeList &Attrs, |
4882 | llvm::Align NewAlign) { |
4883 | llvm::Align CurAlign = Attrs.getRetAlignment().valueOrOne(); |
4884 | if (CurAlign >= NewAlign) |
4885 | return Attrs; |
4886 | llvm::Attribute AlignAttr = llvm::Attribute::getWithAlignment(Context&: Ctx, Alignment: NewAlign); |
4887 | return Attrs.removeRetAttribute(Ctx, llvm::Attribute::AttrKind::Alignment) |
4888 | .addRetAttribute(Ctx, AlignAttr); |
4889 | } |
4890 | |
4891 | template <typename AlignedAttrTy> class AbstractAssumeAlignedAttrEmitter { |
4892 | protected: |
4893 | CodeGenFunction &CGF; |
4894 | |
4895 | /// We do nothing if this is, or becomes, nullptr. |
4896 | const AlignedAttrTy *AA = nullptr; |
4897 | |
4898 | llvm::Value *Alignment = nullptr; // May or may not be a constant. |
4899 | llvm::ConstantInt *OffsetCI = nullptr; // Constant, hopefully zero. |
4900 | |
4901 | AbstractAssumeAlignedAttrEmitter(CodeGenFunction &CGF_, const Decl *FuncDecl) |
4902 | : CGF(CGF_) { |
4903 | if (!FuncDecl) |
4904 | return; |
4905 | AA = FuncDecl->getAttr<AlignedAttrTy>(); |
4906 | } |
4907 | |
4908 | public: |
4909 | /// If we can, materialize the alignment as an attribute on return value. |
4910 | [[nodiscard]] llvm::AttributeList |
4911 | TryEmitAsCallSiteAttribute(const llvm::AttributeList &Attrs) { |
4912 | if (!AA || OffsetCI || CGF.SanOpts.has(K: SanitizerKind::Alignment)) |
4913 | return Attrs; |
4914 | const auto *AlignmentCI = dyn_cast<llvm::ConstantInt>(Val: Alignment); |
4915 | if (!AlignmentCI) |
4916 | return Attrs; |
4917 | // We may legitimately have non-power-of-2 alignment here. |
4918 | // If so, this is UB land, emit it via `@llvm.assume` instead. |
4919 | if (!AlignmentCI->getValue().isPowerOf2()) |
4920 | return Attrs; |
4921 | llvm::AttributeList NewAttrs = maybeRaiseRetAlignmentAttribute( |
4922 | Ctx&: CGF.getLLVMContext(), Attrs, |
4923 | NewAlign: llvm::Align( |
4924 | AlignmentCI->getLimitedValue(Limit: llvm::Value::MaximumAlignment))); |
4925 | AA = nullptr; // We're done. Disallow doing anything else. |
4926 | return NewAttrs; |
4927 | } |
4928 | |
4929 | /// Emit alignment assumption. |
4930 | /// This is a general fallback that we take if either there is an offset, |
4931 | /// or the alignment is variable or we are sanitizing for alignment. |
4932 | void EmitAsAnAssumption(SourceLocation Loc, QualType RetTy, RValue &Ret) { |
4933 | if (!AA) |
4934 | return; |
4935 | CGF.emitAlignmentAssumption(Ret.getScalarVal(), RetTy, Loc, |
4936 | AA->getLocation(), Alignment, OffsetCI); |
4937 | AA = nullptr; // We're done. Disallow doing anything else. |
4938 | } |
4939 | }; |
4940 | |
4941 | /// Helper data structure to emit `AssumeAlignedAttr`. |
4942 | class AssumeAlignedAttrEmitter final |
4943 | : public AbstractAssumeAlignedAttrEmitter<AssumeAlignedAttr> { |
4944 | public: |
4945 | AssumeAlignedAttrEmitter(CodeGenFunction &CGF_, const Decl *FuncDecl) |
4946 | : AbstractAssumeAlignedAttrEmitter(CGF_, FuncDecl) { |
4947 | if (!AA) |
4948 | return; |
4949 | // It is guaranteed that the alignment/offset are constants. |
4950 | Alignment = cast<llvm::ConstantInt>(CGF.EmitScalarExpr(AA->getAlignment())); |
4951 | if (Expr *Offset = AA->getOffset()) { |
4952 | OffsetCI = cast<llvm::ConstantInt>(CGF.EmitScalarExpr(Offset)); |
4953 | if (OffsetCI->isNullValue()) // Canonicalize zero offset to no offset. |
4954 | OffsetCI = nullptr; |
4955 | } |
4956 | } |
4957 | }; |
4958 | |
4959 | /// Helper data structure to emit `AllocAlignAttr`. |
4960 | class AllocAlignAttrEmitter final |
4961 | : public AbstractAssumeAlignedAttrEmitter<AllocAlignAttr> { |
4962 | public: |
4963 | AllocAlignAttrEmitter(CodeGenFunction &CGF_, const Decl *FuncDecl, |
4964 | const CallArgList &CallArgs) |
4965 | : AbstractAssumeAlignedAttrEmitter(CGF_, FuncDecl) { |
4966 | if (!AA) |
4967 | return; |
4968 | // Alignment may or may not be a constant, and that is okay. |
4969 | Alignment = CallArgs[AA->getParamIndex().getLLVMIndex()] |
4970 | .getRValue(CGF) |
4971 | .getScalarVal(); |
4972 | } |
4973 | }; |
4974 | |
4975 | } // namespace |
4976 | |
4977 | static unsigned getMaxVectorWidth(const llvm::Type *Ty) { |
4978 | if (auto *VT = dyn_cast<llvm::VectorType>(Val: Ty)) |
4979 | return VT->getPrimitiveSizeInBits().getKnownMinValue(); |
4980 | if (auto *AT = dyn_cast<llvm::ArrayType>(Val: Ty)) |
4981 | return getMaxVectorWidth(Ty: AT->getElementType()); |
4982 | |
4983 | unsigned MaxVectorWidth = 0; |
4984 | if (auto *ST = dyn_cast<llvm::StructType>(Val: Ty)) |
4985 | for (auto *I : ST->elements()) |
4986 | MaxVectorWidth = std::max(a: MaxVectorWidth, b: getMaxVectorWidth(Ty: I)); |
4987 | return MaxVectorWidth; |
4988 | } |
4989 | |
4990 | RValue CodeGenFunction::EmitCall(const CGFunctionInfo &CallInfo, |
4991 | const CGCallee &Callee, |
4992 | ReturnValueSlot ReturnValue, |
4993 | const CallArgList &CallArgs, |
4994 | llvm::CallBase **callOrInvoke, bool IsMustTail, |
4995 | SourceLocation Loc) { |
4996 | // FIXME: We no longer need the types from CallArgs; lift up and simplify. |
4997 | |
4998 | assert(Callee.isOrdinary() || Callee.isVirtual()); |
4999 | |
5000 | // Handle struct-return functions by passing a pointer to the |
5001 | // location that we would like to return into. |
5002 | QualType RetTy = CallInfo.getReturnType(); |
5003 | const ABIArgInfo &RetAI = CallInfo.getReturnInfo(); |
5004 | |
5005 | llvm::FunctionType *IRFuncTy = getTypes().GetFunctionType(FI: CallInfo); |
5006 | |
5007 | const Decl *TargetDecl = Callee.getAbstractInfo().getCalleeDecl().getDecl(); |
5008 | if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(Val: TargetDecl)) { |
5009 | // We can only guarantee that a function is called from the correct |
5010 | // context/function based on the appropriate target attributes, |
5011 | // so only check in the case where we have both always_inline and target |
5012 | // since otherwise we could be making a conditional call after a check for |
5013 | // the proper cpu features (and it won't cause code generation issues due to |
5014 | // function based code generation). |
5015 | if (TargetDecl->hasAttr<AlwaysInlineAttr>() && |
5016 | (TargetDecl->hasAttr<TargetAttr>() || |
5017 | (CurFuncDecl && CurFuncDecl->hasAttr<TargetAttr>()))) |
5018 | checkTargetFeatures(Loc, TargetDecl: FD); |
5019 | |
5020 | // Some architectures (such as x86-64) have the ABI changed based on |
5021 | // attribute-target/features. Give them a chance to diagnose. |
5022 | CGM.getTargetCodeGenInfo().checkFunctionCallABI( |
5023 | CGM, CallLoc: Loc, Caller: dyn_cast_or_null<FunctionDecl>(Val: CurCodeDecl), Callee: FD, Args: CallArgs); |
5024 | } |
5025 | |
5026 | // 1. Set up the arguments. |
5027 | |
5028 | // If we're using inalloca, insert the allocation after the stack save. |
5029 | // FIXME: Do this earlier rather than hacking it in here! |
5030 | Address ArgMemory = Address::invalid(); |
5031 | if (llvm::StructType *ArgStruct = CallInfo.getArgStruct()) { |
5032 | const llvm::DataLayout &DL = CGM.getDataLayout(); |
5033 | llvm::Instruction *IP = CallArgs.getStackBase(); |
5034 | llvm::AllocaInst *AI; |
5035 | if (IP) { |
5036 | IP = IP->getNextNode(); |
5037 | AI = new llvm::AllocaInst(ArgStruct, DL.getAllocaAddrSpace(), |
5038 | "argmem" , IP); |
5039 | } else { |
5040 | AI = CreateTempAlloca(Ty: ArgStruct, Name: "argmem" ); |
5041 | } |
5042 | auto Align = CallInfo.getArgStructAlignment(); |
5043 | AI->setAlignment(Align.getAsAlign()); |
5044 | AI->setUsedWithInAlloca(true); |
5045 | assert(AI->isUsedWithInAlloca() && !AI->isStaticAlloca()); |
5046 | ArgMemory = Address(AI, ArgStruct, Align); |
5047 | } |
5048 | |
5049 | ClangToLLVMArgMapping IRFunctionArgs(CGM.getContext(), CallInfo); |
5050 | SmallVector<llvm::Value *, 16> IRCallArgs(IRFunctionArgs.totalIRArgs()); |
5051 | |
5052 | // If the call returns a temporary with struct return, create a temporary |
5053 | // alloca to hold the result, unless one is given to us. |
5054 | Address SRetPtr = Address::invalid(); |
5055 | Address SRetAlloca = Address::invalid(); |
5056 | llvm::Value *UnusedReturnSizePtr = nullptr; |
5057 | if (RetAI.isIndirect() || RetAI.isInAlloca() || RetAI.isCoerceAndExpand()) { |
5058 | if (!ReturnValue.isNull()) { |
5059 | SRetPtr = ReturnValue.getValue(); |
5060 | } else { |
5061 | SRetPtr = CreateMemTemp(T: RetTy, Name: "tmp" , Alloca: &SRetAlloca); |
5062 | if (HaveInsertPoint() && ReturnValue.isUnused()) { |
5063 | llvm::TypeSize size = |
5064 | CGM.getDataLayout().getTypeAllocSize(Ty: ConvertTypeForMem(T: RetTy)); |
5065 | UnusedReturnSizePtr = EmitLifetimeStart(Size: size, Addr: SRetAlloca.getPointer()); |
5066 | } |
5067 | } |
5068 | if (IRFunctionArgs.hasSRetArg()) { |
5069 | IRCallArgs[IRFunctionArgs.getSRetArgNo()] = SRetPtr.getPointer(); |
5070 | } else if (RetAI.isInAlloca()) { |
5071 | Address Addr = |
5072 | Builder.CreateStructGEP(Addr: ArgMemory, Index: RetAI.getInAllocaFieldIndex()); |
5073 | Builder.CreateStore(Val: SRetPtr.getPointer(), Addr); |
5074 | } |
5075 | } |
5076 | |
5077 | Address swiftErrorTemp = Address::invalid(); |
5078 | Address swiftErrorArg = Address::invalid(); |
5079 | |
5080 | // When passing arguments using temporary allocas, we need to add the |
5081 | // appropriate lifetime markers. This vector keeps track of all the lifetime |
5082 | // markers that need to be ended right after the call. |
5083 | SmallVector<CallLifetimeEnd, 2> CallLifetimeEndAfterCall; |
5084 | |
5085 | // Translate all of the arguments as necessary to match the IR lowering. |
5086 | assert(CallInfo.arg_size() == CallArgs.size() && |
5087 | "Mismatch between function signature & arguments." ); |
5088 | unsigned ArgNo = 0; |
5089 | CGFunctionInfo::const_arg_iterator info_it = CallInfo.arg_begin(); |
5090 | for (CallArgList::const_iterator I = CallArgs.begin(), E = CallArgs.end(); |
5091 | I != E; ++I, ++info_it, ++ArgNo) { |
5092 | const ABIArgInfo &ArgInfo = info_it->info; |
5093 | |
5094 | // Insert a padding argument to ensure proper alignment. |
5095 | if (IRFunctionArgs.hasPaddingArg(ArgNo)) |
5096 | IRCallArgs[IRFunctionArgs.getPaddingArgNo(ArgNo)] = |
5097 | llvm::UndefValue::get(T: ArgInfo.getPaddingType()); |
5098 | |
5099 | unsigned FirstIRArg, NumIRArgs; |
5100 | std::tie(args&: FirstIRArg, args&: NumIRArgs) = IRFunctionArgs.getIRArgs(ArgNo); |
5101 | |
5102 | bool ArgHasMaybeUndefAttr = |
5103 | IsArgumentMaybeUndef(TargetDecl, NumRequiredArgs: CallInfo.getNumRequiredArgs(), ArgNo); |
5104 | |
5105 | switch (ArgInfo.getKind()) { |
5106 | case ABIArgInfo::InAlloca: { |
5107 | assert(NumIRArgs == 0); |
5108 | assert(getTarget().getTriple().getArch() == llvm::Triple::x86); |
5109 | if (I->isAggregate()) { |
5110 | Address Addr = I->hasLValue() |
5111 | ? I->getKnownLValue().getAddress(CGF&: *this) |
5112 | : I->getKnownRValue().getAggregateAddress(); |
5113 | llvm::Instruction *Placeholder = |
5114 | cast<llvm::Instruction>(Val: Addr.getPointer()); |
5115 | |
5116 | if (!ArgInfo.getInAllocaIndirect()) { |
5117 | // Replace the placeholder with the appropriate argument slot GEP. |
5118 | CGBuilderTy::InsertPoint IP = Builder.saveIP(); |
5119 | Builder.SetInsertPoint(Placeholder); |
5120 | Addr = Builder.CreateStructGEP(Addr: ArgMemory, |
5121 | Index: ArgInfo.getInAllocaFieldIndex()); |
5122 | Builder.restoreIP(IP); |
5123 | } else { |
5124 | // For indirect things such as overaligned structs, replace the |
5125 | // placeholder with a regular aggregate temporary alloca. Store the |
5126 | // address of this alloca into the struct. |
5127 | Addr = CreateMemTemp(info_it->type, "inalloca.indirect.tmp" ); |
5128 | Address ArgSlot = Builder.CreateStructGEP( |
5129 | Addr: ArgMemory, Index: ArgInfo.getInAllocaFieldIndex()); |
5130 | Builder.CreateStore(Val: Addr.getPointer(), Addr: ArgSlot); |
5131 | } |
5132 | deferPlaceholderReplacement(Old: Placeholder, New: Addr.getPointer()); |
5133 | } else if (ArgInfo.getInAllocaIndirect()) { |
5134 | // Make a temporary alloca and store the address of it into the argument |
5135 | // struct. |
5136 | Address Addr = CreateMemTempWithoutCast( |
5137 | I->Ty, getContext().getTypeAlignInChars(I->Ty), |
5138 | "indirect-arg-temp" ); |
5139 | I->copyInto(CGF&: *this, Addr); |
5140 | Address ArgSlot = |
5141 | Builder.CreateStructGEP(Addr: ArgMemory, Index: ArgInfo.getInAllocaFieldIndex()); |
5142 | Builder.CreateStore(Val: Addr.getPointer(), Addr: ArgSlot); |
5143 | } else { |
5144 | // Store the RValue into the argument struct. |
5145 | Address Addr = |
5146 | Builder.CreateStructGEP(Addr: ArgMemory, Index: ArgInfo.getInAllocaFieldIndex()); |
5147 | Addr = Addr.withElementType(ElemTy: ConvertTypeForMem(T: I->Ty)); |
5148 | I->copyInto(CGF&: *this, Addr); |
5149 | } |
5150 | break; |
5151 | } |
5152 | |
5153 | case ABIArgInfo::Indirect: |
5154 | case ABIArgInfo::IndirectAliased: { |
5155 | assert(NumIRArgs == 1); |
5156 | if (!I->isAggregate()) { |
5157 | // Make a temporary alloca to pass the argument. |
5158 | Address Addr = CreateMemTempWithoutCast( |
5159 | I->Ty, ArgInfo.getIndirectAlign(), "indirect-arg-temp" ); |
5160 | |
5161 | llvm::Value *Val = Addr.getPointer(); |
5162 | if (ArgHasMaybeUndefAttr) |
5163 | Val = Builder.CreateFreeze(V: Addr.getPointer()); |
5164 | IRCallArgs[FirstIRArg] = Val; |
5165 | |
5166 | I->copyInto(CGF&: *this, Addr); |
5167 | } else { |
5168 | // We want to avoid creating an unnecessary temporary+copy here; |
5169 | // however, we need one in three cases: |
5170 | // 1. If the argument is not byval, and we are required to copy the |
5171 | // source. (This case doesn't occur on any common architecture.) |
5172 | // 2. If the argument is byval, RV is not sufficiently aligned, and |
5173 | // we cannot force it to be sufficiently aligned. |
5174 | // 3. If the argument is byval, but RV is not located in default |
5175 | // or alloca address space. |
5176 | Address Addr = I->hasLValue() |
5177 | ? I->getKnownLValue().getAddress(CGF&: *this) |
5178 | : I->getKnownRValue().getAggregateAddress(); |
5179 | llvm::Value *V = Addr.getPointer(); |
5180 | CharUnits Align = ArgInfo.getIndirectAlign(); |
5181 | const llvm::DataLayout *TD = &CGM.getDataLayout(); |
5182 | |
5183 | assert((FirstIRArg >= IRFuncTy->getNumParams() || |
5184 | IRFuncTy->getParamType(FirstIRArg)->getPointerAddressSpace() == |
5185 | TD->getAllocaAddrSpace()) && |
5186 | "indirect argument must be in alloca address space" ); |
5187 | |
5188 | bool NeedCopy = false; |
5189 | if (Addr.getAlignment() < Align && |
5190 | llvm::getOrEnforceKnownAlignment(V, PrefAlign: Align.getAsAlign(), DL: *TD) < |
5191 | Align.getAsAlign()) { |
5192 | NeedCopy = true; |
5193 | } else if (I->hasLValue()) { |
5194 | auto LV = I->getKnownLValue(); |
5195 | auto AS = LV.getAddressSpace(); |
5196 | |
5197 | bool isByValOrRef = |
5198 | ArgInfo.isIndirectAliased() || ArgInfo.getIndirectByVal(); |
5199 | |
5200 | if (!isByValOrRef || |
5201 | (LV.getAlignment() < getContext().getTypeAlignInChars(I->Ty))) { |
5202 | NeedCopy = true; |
5203 | } |
5204 | if (!getLangOpts().OpenCL) { |
5205 | if ((isByValOrRef && |
5206 | (AS != LangAS::Default && |
5207 | AS != CGM.getASTAllocaAddressSpace()))) { |
5208 | NeedCopy = true; |
5209 | } |
5210 | } |
5211 | // For OpenCL even if RV is located in default or alloca address space |
5212 | // we don't want to perform address space cast for it. |
5213 | else if ((isByValOrRef && |
5214 | Addr.getType()->getAddressSpace() != IRFuncTy-> |
5215 | getParamType(i: FirstIRArg)->getPointerAddressSpace())) { |
5216 | NeedCopy = true; |
5217 | } |
5218 | } |
5219 | |
5220 | if (NeedCopy) { |
5221 | // Create an aligned temporary, and copy to it. |
5222 | Address AI = CreateMemTempWithoutCast( |
5223 | I->Ty, ArgInfo.getIndirectAlign(), "byval-temp" ); |
5224 | llvm::Value *Val = AI.getPointer(); |
5225 | if (ArgHasMaybeUndefAttr) |
5226 | Val = Builder.CreateFreeze(V: AI.getPointer()); |
5227 | IRCallArgs[FirstIRArg] = Val; |
5228 | |
5229 | // Emit lifetime markers for the temporary alloca. |
5230 | llvm::TypeSize ByvalTempElementSize = |
5231 | CGM.getDataLayout().getTypeAllocSize(Ty: AI.getElementType()); |
5232 | llvm::Value *LifetimeSize = |
5233 | EmitLifetimeStart(Size: ByvalTempElementSize, Addr: AI.getPointer()); |
5234 | |
5235 | // Add cleanup code to emit the end lifetime marker after the call. |
5236 | if (LifetimeSize) // In case we disabled lifetime markers. |
5237 | CallLifetimeEndAfterCall.emplace_back(Args&: AI, Args&: LifetimeSize); |
5238 | |
5239 | // Generate the copy. |
5240 | I->copyInto(CGF&: *this, Addr: AI); |
5241 | } else { |
5242 | // Skip the extra memcpy call. |
5243 | auto *T = llvm::PointerType::get( |
5244 | C&: CGM.getLLVMContext(), AddressSpace: CGM.getDataLayout().getAllocaAddrSpace()); |
5245 | |
5246 | llvm::Value *Val = getTargetHooks().performAddrSpaceCast( |
5247 | *this, V, LangAS::Default, CGM.getASTAllocaAddressSpace(), T, |
5248 | true); |
5249 | if (ArgHasMaybeUndefAttr) |
5250 | Val = Builder.CreateFreeze(V: Val); |
5251 | IRCallArgs[FirstIRArg] = Val; |
5252 | } |
5253 | } |
5254 | break; |
5255 | } |
5256 | |
5257 | case ABIArgInfo::Ignore: |
5258 | assert(NumIRArgs == 0); |
5259 | break; |
5260 | |
5261 | case ABIArgInfo::Extend: |
5262 | case ABIArgInfo::Direct: { |
5263 | if (!isa<llvm::StructType>(Val: ArgInfo.getCoerceToType()) && |
5264 | ArgInfo.getCoerceToType() == ConvertType(info_it->type) && |
5265 | ArgInfo.getDirectOffset() == 0) { |
5266 | assert(NumIRArgs == 1); |
5267 | llvm::Value *V; |
5268 | if (!I->isAggregate()) |
5269 | V = I->getKnownRValue().getScalarVal(); |
5270 | else |
5271 | V = Builder.CreateLoad( |
5272 | Addr: I->hasLValue() ? I->getKnownLValue().getAddress(CGF&: *this) |
5273 | : I->getKnownRValue().getAggregateAddress()); |
5274 | |
5275 | // Implement swifterror by copying into a new swifterror argument. |
5276 | // We'll write back in the normal path out of the call. |
5277 | if (CallInfo.getExtParameterInfo(argIndex: ArgNo).getABI() |
5278 | == ParameterABI::SwiftErrorResult) { |
5279 | assert(!swiftErrorTemp.isValid() && "multiple swifterror args" ); |
5280 | |
5281 | QualType pointeeTy = I->Ty->getPointeeType(); |
5282 | swiftErrorArg = Address(V, ConvertTypeForMem(T: pointeeTy), |
5283 | getContext().getTypeAlignInChars(T: pointeeTy)); |
5284 | |
5285 | swiftErrorTemp = |
5286 | CreateMemTemp(pointeeTy, getPointerAlign(), "swifterror.temp" ); |
5287 | V = swiftErrorTemp.getPointer(); |
5288 | cast<llvm::AllocaInst>(Val: V)->setSwiftError(true); |
5289 | |
5290 | llvm::Value *errorValue = Builder.CreateLoad(Addr: swiftErrorArg); |
5291 | Builder.CreateStore(Val: errorValue, Addr: swiftErrorTemp); |
5292 | } |
5293 | |
5294 | // We might have to widen integers, but we should never truncate. |
5295 | if (ArgInfo.getCoerceToType() != V->getType() && |
5296 | V->getType()->isIntegerTy()) |
5297 | V = Builder.CreateZExt(V, DestTy: ArgInfo.getCoerceToType()); |
5298 | |
5299 | // If the argument doesn't match, perform a bitcast to coerce it. This |
5300 | // can happen due to trivial type mismatches. |
5301 | if (FirstIRArg < IRFuncTy->getNumParams() && |
5302 | V->getType() != IRFuncTy->getParamType(i: FirstIRArg)) |
5303 | V = Builder.CreateBitCast(V, DestTy: IRFuncTy->getParamType(i: FirstIRArg)); |
5304 | |
5305 | if (ArgHasMaybeUndefAttr) |
5306 | V = Builder.CreateFreeze(V); |
5307 | IRCallArgs[FirstIRArg] = V; |
5308 | break; |
5309 | } |
5310 | |
5311 | llvm::StructType *STy = |
5312 | dyn_cast<llvm::StructType>(Val: ArgInfo.getCoerceToType()); |
5313 | llvm::Type *SrcTy = ConvertTypeForMem(T: I->Ty); |
5314 | llvm::TypeSize SrcTypeSize; |
5315 | llvm::TypeSize DstTypeSize; |
5316 | if (STy && ArgInfo.isDirect() && !ArgInfo.getCanBeFlattened()) { |
5317 | SrcTypeSize = CGM.getDataLayout().getTypeAllocSize(Ty: SrcTy); |
5318 | DstTypeSize = CGM.getDataLayout().getTypeAllocSize(Ty: STy); |
5319 | if (STy->containsHomogeneousScalableVectorTypes()) { |
5320 | assert(SrcTypeSize == DstTypeSize && |
5321 | "Only allow non-fractional movement of structure with " |
5322 | "homogeneous scalable vector type" ); |
5323 | |
5324 | IRCallArgs[FirstIRArg] = I->getKnownRValue().getScalarVal(); |
5325 | break; |
5326 | } |
5327 | } |
5328 | |
5329 | // FIXME: Avoid the conversion through memory if possible. |
5330 | Address Src = Address::invalid(); |
5331 | if (!I->isAggregate()) { |
5332 | Src = CreateMemTemp(I->Ty, "coerce" ); |
5333 | I->copyInto(CGF&: *this, Addr: Src); |
5334 | } else { |
5335 | Src = I->hasLValue() ? I->getKnownLValue().getAddress(CGF&: *this) |
5336 | : I->getKnownRValue().getAggregateAddress(); |
5337 | } |
5338 | |
5339 | // If the value is offset in memory, apply the offset now. |
5340 | Src = emitAddressAtOffset(CGF&: *this, addr: Src, info: ArgInfo); |
5341 | |
5342 | // Fast-isel and the optimizer generally like scalar values better than |
5343 | // FCAs, so we flatten them if this is safe to do for this argument. |
5344 | if (STy && ArgInfo.isDirect() && ArgInfo.getCanBeFlattened()) { |
5345 | llvm::Type *SrcTy = Src.getElementType(); |
5346 | llvm::TypeSize SrcTypeSize = |
5347 | CGM.getDataLayout().getTypeAllocSize(Ty: SrcTy); |
5348 | llvm::TypeSize DstTypeSize = CGM.getDataLayout().getTypeAllocSize(Ty: STy); |
5349 | if (SrcTypeSize.isScalable()) { |
5350 | assert(STy->containsHomogeneousScalableVectorTypes() && |
5351 | "ABI only supports structure with homogeneous scalable vector " |
5352 | "type" ); |
5353 | assert(SrcTypeSize == DstTypeSize && |
5354 | "Only allow non-fractional movement of structure with " |
5355 | "homogeneous scalable vector type" ); |
5356 | assert(NumIRArgs == STy->getNumElements()); |
5357 | |
5358 | llvm::Value *StoredStructValue = |
5359 | Builder.CreateLoad(Addr: Src, Name: Src.getName() + ".tuple" ); |
5360 | for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { |
5361 | llvm::Value * = Builder.CreateExtractValue( |
5362 | Agg: StoredStructValue, Idxs: i, Name: Src.getName() + ".extract" + Twine(i)); |
5363 | IRCallArgs[FirstIRArg + i] = Extract; |
5364 | } |
5365 | } else { |
5366 | uint64_t SrcSize = SrcTypeSize.getFixedValue(); |
5367 | uint64_t DstSize = DstTypeSize.getFixedValue(); |
5368 | |
5369 | // If the source type is smaller than the destination type of the |
5370 | // coerce-to logic, copy the source value into a temp alloca the size |
5371 | // of the destination type to allow loading all of it. The bits past |
5372 | // the source value are left undef. |
5373 | if (SrcSize < DstSize) { |
5374 | Address TempAlloca = CreateTempAlloca(Ty: STy, align: Src.getAlignment(), |
5375 | Name: Src.getName() + ".coerce" ); |
5376 | Builder.CreateMemCpy(Dest: TempAlloca, Src, Size: SrcSize); |
5377 | Src = TempAlloca; |
5378 | } else { |
5379 | Src = Src.withElementType(ElemTy: STy); |
5380 | } |
5381 | |
5382 | assert(NumIRArgs == STy->getNumElements()); |
5383 | for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) { |
5384 | Address EltPtr = Builder.CreateStructGEP(Addr: Src, Index: i); |
5385 | llvm::Value *LI = Builder.CreateLoad(Addr: EltPtr); |
5386 | if (ArgHasMaybeUndefAttr) |
5387 | LI = Builder.CreateFreeze(V: LI); |
5388 | IRCallArgs[FirstIRArg + i] = LI; |
5389 | } |
5390 | } |
5391 | } else { |
5392 | // In the simple case, just pass the coerced loaded value. |
5393 | assert(NumIRArgs == 1); |
5394 | llvm::Value *Load = |
5395 | CreateCoercedLoad(Src, Ty: ArgInfo.getCoerceToType(), CGF&: *this); |
5396 | |
5397 | if (CallInfo.isCmseNSCall()) { |
5398 | // For certain parameter types, clear padding bits, as they may reveal |
5399 | // sensitive information. |
5400 | // Small struct/union types are passed as integer arrays. |
5401 | auto *ATy = dyn_cast<llvm::ArrayType>(Val: Load->getType()); |
5402 | if (ATy != nullptr && isa<RecordType>(I->Ty.getCanonicalType())) |
5403 | Load = EmitCMSEClearRecord(Load, ATy, I->Ty); |
5404 | } |
5405 | |
5406 | if (ArgHasMaybeUndefAttr) |
5407 | Load = Builder.CreateFreeze(V: Load); |
5408 | IRCallArgs[FirstIRArg] = Load; |
5409 | } |
5410 | |
5411 | break; |
5412 | } |
5413 | |
5414 | case ABIArgInfo::CoerceAndExpand: { |
5415 | auto coercionType = ArgInfo.getCoerceAndExpandType(); |
5416 | auto layout = CGM.getDataLayout().getStructLayout(Ty: coercionType); |
5417 | |
5418 | llvm::Value *tempSize = nullptr; |
5419 | Address addr = Address::invalid(); |
5420 | Address AllocaAddr = Address::invalid(); |
5421 | if (I->isAggregate()) { |
5422 | addr = I->hasLValue() ? I->getKnownLValue().getAddress(CGF&: *this) |
5423 | : I->getKnownRValue().getAggregateAddress(); |
5424 | |
5425 | } else { |
5426 | RValue RV = I->getKnownRValue(); |
5427 | assert(RV.isScalar()); // complex should always just be direct |
5428 | |
5429 | llvm::Type *scalarType = RV.getScalarVal()->getType(); |
5430 | auto scalarSize = CGM.getDataLayout().getTypeAllocSize(Ty: scalarType); |
5431 | auto scalarAlign = CGM.getDataLayout().getPrefTypeAlign(Ty: scalarType); |
5432 | |
5433 | // Materialize to a temporary. |
5434 | addr = CreateTempAlloca( |
5435 | Ty: RV.getScalarVal()->getType(), |
5436 | align: CharUnits::fromQuantity(Quantity: std::max(a: layout->getAlignment(), b: scalarAlign)), |
5437 | Name: "tmp" , |
5438 | /*ArraySize=*/nullptr, Alloca: &AllocaAddr); |
5439 | tempSize = EmitLifetimeStart(Size: scalarSize, Addr: AllocaAddr.getPointer()); |
5440 | |
5441 | Builder.CreateStore(Val: RV.getScalarVal(), Addr: addr); |
5442 | } |
5443 | |
5444 | addr = addr.withElementType(ElemTy: coercionType); |
5445 | |
5446 | unsigned IRArgPos = FirstIRArg; |
5447 | for (unsigned i = 0, e = coercionType->getNumElements(); i != e; ++i) { |
5448 | llvm::Type *eltType = coercionType->getElementType(N: i); |
5449 | if (ABIArgInfo::isPaddingForCoerceAndExpand(eltType)) continue; |
5450 | Address eltAddr = Builder.CreateStructGEP(Addr: addr, Index: i); |
5451 | llvm::Value *elt = Builder.CreateLoad(Addr: eltAddr); |
5452 | if (ArgHasMaybeUndefAttr) |
5453 | elt = Builder.CreateFreeze(V: elt); |
5454 | IRCallArgs[IRArgPos++] = elt; |
5455 | } |
5456 | assert(IRArgPos == FirstIRArg + NumIRArgs); |
5457 | |
5458 | if (tempSize) { |
5459 | EmitLifetimeEnd(Size: tempSize, Addr: AllocaAddr.getPointer()); |
5460 | } |
5461 | |
5462 | break; |
5463 | } |
5464 | |
5465 | case ABIArgInfo::Expand: { |
5466 | unsigned IRArgPos = FirstIRArg; |
5467 | ExpandTypeToArgs(Ty: I->Ty, Arg: *I, IRFuncTy, IRCallArgs, IRCallArgPos&: IRArgPos); |
5468 | assert(IRArgPos == FirstIRArg + NumIRArgs); |
5469 | break; |
5470 | } |
5471 | } |
5472 | } |
5473 | |
5474 | const CGCallee &ConcreteCallee = Callee.prepareConcreteCallee(CGF&: *this); |
5475 | llvm::Value *CalleePtr = ConcreteCallee.getFunctionPointer(); |
5476 | |
5477 | // If we're using inalloca, set up that argument. |
5478 | if (ArgMemory.isValid()) { |
5479 | llvm::Value *Arg = ArgMemory.getPointer(); |
5480 | assert(IRFunctionArgs.hasInallocaArg()); |
5481 | IRCallArgs[IRFunctionArgs.getInallocaArgNo()] = Arg; |
5482 | } |
5483 | |
5484 | // 2. Prepare the function pointer. |
5485 | |
5486 | // If the callee is a bitcast of a non-variadic function to have a |
5487 | // variadic function pointer type, check to see if we can remove the |
5488 | // bitcast. This comes up with unprototyped functions. |
5489 | // |
5490 | // This makes the IR nicer, but more importantly it ensures that we |
5491 | // can inline the function at -O0 if it is marked always_inline. |
5492 | auto simplifyVariadicCallee = [](llvm::FunctionType *CalleeFT, |
5493 | llvm::Value *Ptr) -> llvm::Function * { |
5494 | if (!CalleeFT->isVarArg()) |
5495 | return nullptr; |
5496 | |
5497 | // Get underlying value if it's a bitcast |
5498 | if (llvm::ConstantExpr *CE = dyn_cast<llvm::ConstantExpr>(Val: Ptr)) { |
5499 | if (CE->getOpcode() == llvm::Instruction::BitCast) |
5500 | Ptr = CE->getOperand(i_nocapture: 0); |
5501 | } |
5502 | |
5503 | llvm::Function *OrigFn = dyn_cast<llvm::Function>(Val: Ptr); |
5504 | if (!OrigFn) |
5505 | return nullptr; |
5506 | |
5507 | llvm::FunctionType *OrigFT = OrigFn->getFunctionType(); |
5508 | |
5509 | // If the original type is variadic, or if any of the component types |
5510 | // disagree, we cannot remove the cast. |
5511 | if (OrigFT->isVarArg() || |
5512 | OrigFT->getNumParams() != CalleeFT->getNumParams() || |
5513 | OrigFT->getReturnType() != CalleeFT->getReturnType()) |
5514 | return nullptr; |
5515 | |
5516 | for (unsigned i = 0, e = OrigFT->getNumParams(); i != e; ++i) |
5517 | if (OrigFT->getParamType(i) != CalleeFT->getParamType(i)) |
5518 | return nullptr; |
5519 | |
5520 | return OrigFn; |
5521 | }; |
5522 | |
5523 | if (llvm::Function *OrigFn = simplifyVariadicCallee(IRFuncTy, CalleePtr)) { |
5524 | CalleePtr = OrigFn; |
5525 | IRFuncTy = OrigFn->getFunctionType(); |
5526 | } |
5527 | |
5528 | // 3. Perform the actual call. |
5529 | |
5530 | // Deactivate any cleanups that we're supposed to do immediately before |
5531 | // the call. |
5532 | if (!CallArgs.getCleanupsToDeactivate().empty()) |
5533 | deactivateArgCleanupsBeforeCall(CGF&: *this, CallArgs); |
5534 | |
5535 | // Assert that the arguments we computed match up. The IR verifier |
5536 | // will catch this, but this is a common enough source of problems |
5537 | // during IRGen changes that it's way better for debugging to catch |
5538 | // it ourselves here. |
5539 | #ifndef NDEBUG |
5540 | assert(IRCallArgs.size() == IRFuncTy->getNumParams() || IRFuncTy->isVarArg()); |
5541 | for (unsigned i = 0; i < IRCallArgs.size(); ++i) { |
5542 | // Inalloca argument can have different type. |
5543 | if (IRFunctionArgs.hasInallocaArg() && |
5544 | i == IRFunctionArgs.getInallocaArgNo()) |
5545 | continue; |
5546 | if (i < IRFuncTy->getNumParams()) |
5547 | assert(IRCallArgs[i]->getType() == IRFuncTy->getParamType(i)); |
5548 | } |
5549 | #endif |
5550 | |
5551 | // Update the largest vector width if any arguments have vector types. |
5552 | for (unsigned i = 0; i < IRCallArgs.size(); ++i) |
5553 | LargestVectorWidth = std::max(a: LargestVectorWidth, |
5554 | b: getMaxVectorWidth(Ty: IRCallArgs[i]->getType())); |
5555 | |
5556 | // Compute the calling convention and attributes. |
5557 | unsigned CallingConv; |
5558 | llvm::AttributeList Attrs; |
5559 | CGM.ConstructAttributeList(Name: CalleePtr->getName(), FI: CallInfo, |
5560 | CalleeInfo: Callee.getAbstractInfo(), AttrList&: Attrs, CallingConv, |
5561 | /*AttrOnCallSite=*/true, |
5562 | /*IsThunk=*/false); |
5563 | |
5564 | if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(Val: CurFuncDecl)) { |
5565 | if (FD->hasAttr<StrictFPAttr>()) |
5566 | // All calls within a strictfp function are marked strictfp |
5567 | Attrs = Attrs.addFnAttribute(getLLVMContext(), llvm::Attribute::StrictFP); |
5568 | |
5569 | // If -ffast-math is enabled and the function is guarded by an |
5570 | // '__attribute__((optnone)) adjust the memory attribute so the BE emits the |
5571 | // library call instead of the intrinsic. |
5572 | if (FD->hasAttr<OptimizeNoneAttr>() && getLangOpts().FastMath) |
5573 | CGM.AdjustMemoryAttribute(Name: CalleePtr->getName(), CalleeInfo: Callee.getAbstractInfo(), |
5574 | Attrs); |
5575 | } |
5576 | // Add call-site nomerge attribute if exists. |
5577 | if (InNoMergeAttributedStmt) |
5578 | Attrs = Attrs.addFnAttribute(getLLVMContext(), llvm::Attribute::NoMerge); |
5579 | |
5580 | // Add call-site noinline attribute if exists. |
5581 | if (InNoInlineAttributedStmt) |
5582 | Attrs = Attrs.addFnAttribute(getLLVMContext(), llvm::Attribute::NoInline); |
5583 | |
5584 | // Add call-site always_inline attribute if exists. |
5585 | if (InAlwaysInlineAttributedStmt) |
5586 | Attrs = |
5587 | Attrs.addFnAttribute(getLLVMContext(), llvm::Attribute::AlwaysInline); |
5588 | |
5589 | // Apply some call-site-specific attributes. |
5590 | // TODO: work this into building the attribute set. |
5591 | |
5592 | // Apply always_inline to all calls within flatten functions. |
5593 | // FIXME: should this really take priority over __try, below? |
5594 | if (CurCodeDecl && CurCodeDecl->hasAttr<FlattenAttr>() && |
5595 | !InNoInlineAttributedStmt && |
5596 | !(TargetDecl && TargetDecl->hasAttr<NoInlineAttr>())) { |
5597 | Attrs = |
5598 | Attrs.addFnAttribute(getLLVMContext(), llvm::Attribute::AlwaysInline); |
5599 | } |
5600 | |
5601 | // Disable inlining inside SEH __try blocks. |
5602 | if (isSEHTryScope()) { |
5603 | Attrs = Attrs.addFnAttribute(getLLVMContext(), llvm::Attribute::NoInline); |
5604 | } |
5605 | |
5606 | // Decide whether to use a call or an invoke. |
5607 | bool CannotThrow; |
5608 | if (currentFunctionUsesSEHTry()) { |
5609 | // SEH cares about asynchronous exceptions, so everything can "throw." |
5610 | CannotThrow = false; |
5611 | } else if (isCleanupPadScope() && |
5612 | EHPersonality::get(CGF&: *this).isMSVCXXPersonality()) { |
5613 | // The MSVC++ personality will implicitly terminate the program if an |
5614 | // exception is thrown during a cleanup outside of a try/catch. |
5615 | // We don't need to model anything in IR to get this behavior. |
5616 | CannotThrow = true; |
5617 | } else { |
5618 | // Otherwise, nounwind call sites will never throw. |
5619 | CannotThrow = Attrs.hasFnAttr(llvm::Attribute::NoUnwind); |
5620 | |
5621 | if (auto *FPtr = dyn_cast<llvm::Function>(Val: CalleePtr)) |
5622 | if (FPtr->hasFnAttribute(llvm::Attribute::NoUnwind)) |
5623 | CannotThrow = true; |
5624 | } |
5625 | |
5626 | // If we made a temporary, be sure to clean up after ourselves. Note that we |
5627 | // can't depend on being inside of an ExprWithCleanups, so we need to manually |
5628 | // pop this cleanup later on. Being eager about this is OK, since this |
5629 | // temporary is 'invisible' outside of the callee. |
5630 | if (UnusedReturnSizePtr) |
5631 | pushFullExprCleanup<CallLifetimeEnd>(kind: NormalEHLifetimeMarker, A: SRetAlloca, |
5632 | A: UnusedReturnSizePtr); |
5633 | |
5634 | llvm::BasicBlock *InvokeDest = CannotThrow ? nullptr : getInvokeDest(); |
5635 | |
5636 | SmallVector<llvm::OperandBundleDef, 1> BundleList = |
5637 | getBundlesForFunclet(Callee: CalleePtr); |
5638 | |
5639 | if (SanOpts.has(K: SanitizerKind::KCFI) && |
5640 | !isa_and_nonnull<FunctionDecl>(Val: TargetDecl)) |
5641 | EmitKCFIOperandBundle(Callee: ConcreteCallee, Bundles&: BundleList); |
5642 | |
5643 | if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(CurFuncDecl)) |
5644 | if (FD->hasAttr<StrictFPAttr>()) |
5645 | // All calls within a strictfp function are marked strictfp |
5646 | Attrs = Attrs.addFnAttribute(getLLVMContext(), llvm::Attribute::StrictFP); |
5647 | |
5648 | AssumeAlignedAttrEmitter AssumeAlignedAttrEmitter(*this, TargetDecl); |
5649 | Attrs = AssumeAlignedAttrEmitter.TryEmitAsCallSiteAttribute(Attrs); |
5650 | |
5651 | AllocAlignAttrEmitter AllocAlignAttrEmitter(*this, TargetDecl, CallArgs); |
5652 | Attrs = AllocAlignAttrEmitter.TryEmitAsCallSiteAttribute(Attrs); |
5653 | |
5654 | // Emit the actual call/invoke instruction. |
5655 | llvm::CallBase *CI; |
5656 | if (!InvokeDest) { |
5657 | CI = Builder.CreateCall(FTy: IRFuncTy, Callee: CalleePtr, Args: IRCallArgs, OpBundles: BundleList); |
5658 | } else { |
5659 | llvm::BasicBlock *Cont = createBasicBlock(name: "invoke.cont" ); |
5660 | CI = Builder.CreateInvoke(Ty: IRFuncTy, Callee: CalleePtr, NormalDest: Cont, UnwindDest: InvokeDest, Args: IRCallArgs, |
5661 | OpBundles: BundleList); |
5662 | EmitBlock(BB: Cont); |
5663 | } |
5664 | if (CI->getCalledFunction() && CI->getCalledFunction()->hasName() && |
5665 | CI->getCalledFunction()->getName().starts_with(Prefix: "_Z4sqrt" )) { |
5666 | SetSqrtFPAccuracy(CI); |
5667 | } |
5668 | if (callOrInvoke) |
5669 | *callOrInvoke = CI; |
5670 | |
5671 | // If this is within a function that has the guard(nocf) attribute and is an |
5672 | // indirect call, add the "guard_nocf" attribute to this call to indicate that |
5673 | // Control Flow Guard checks should not be added, even if the call is inlined. |
5674 | if (const auto *FD = dyn_cast_or_null<FunctionDecl>(Val: CurFuncDecl)) { |
5675 | if (const auto *A = FD->getAttr<CFGuardAttr>()) { |
5676 | if (A->getGuard() == CFGuardAttr::GuardArg::nocf && !CI->getCalledFunction()) |
5677 | Attrs = Attrs.addFnAttribute(C&: getLLVMContext(), Kind: "guard_nocf" ); |
5678 | } |
5679 | } |
5680 | |
5681 | // Apply the attributes and calling convention. |
5682 | CI->setAttributes(Attrs); |
5683 | CI->setCallingConv(static_cast<llvm::CallingConv::ID>(CallingConv)); |
5684 | |
5685 | // Apply various metadata. |
5686 | |
5687 | if (!CI->getType()->isVoidTy()) |
5688 | CI->setName("call" ); |
5689 | |
5690 | // Update largest vector width from the return type. |
5691 | LargestVectorWidth = |
5692 | std::max(a: LargestVectorWidth, b: getMaxVectorWidth(Ty: CI->getType())); |
5693 | |
5694 | // Insert instrumentation or attach profile metadata at indirect call sites. |
5695 | // For more details, see the comment before the definition of |
5696 | // IPVK_IndirectCallTarget in InstrProfData.inc. |
5697 | if (!CI->getCalledFunction()) |
5698 | PGO.valueProfile(Builder, ValueKind: llvm::IPVK_IndirectCallTarget, |
5699 | ValueSite: CI, ValuePtr: CalleePtr); |
5700 | |
5701 | // In ObjC ARC mode with no ObjC ARC exception safety, tell the ARC |
5702 | // optimizer it can aggressively ignore unwind edges. |
5703 | if (CGM.getLangOpts().ObjCAutoRefCount) |
5704 | AddObjCARCExceptionMetadata(Inst: CI); |
5705 | |
5706 | // Set tail call kind if necessary. |
5707 | if (llvm::CallInst *Call = dyn_cast<llvm::CallInst>(Val: CI)) { |
5708 | if (TargetDecl && TargetDecl->hasAttr<NotTailCalledAttr>()) |
5709 | Call->setTailCallKind(llvm::CallInst::TCK_NoTail); |
5710 | else if (IsMustTail) |
5711 | Call->setTailCallKind(llvm::CallInst::TCK_MustTail); |
5712 | } |
5713 | |
5714 | // Add metadata for calls to MSAllocator functions |
5715 | if (getDebugInfo() && TargetDecl && |
5716 | TargetDecl->hasAttr<MSAllocatorAttr>()) |
5717 | getDebugInfo()->addHeapAllocSiteMetadata(CallSite: CI, AllocatedTy: RetTy->getPointeeType(), Loc); |
5718 | |
5719 | // Add metadata if calling an __attribute__((error(""))) or warning fn. |
5720 | if (TargetDecl && TargetDecl->hasAttr<ErrorAttr>()) { |
5721 | llvm::ConstantInt *Line = |
5722 | llvm::ConstantInt::get(Ty: Int32Ty, V: Loc.getRawEncoding()); |
5723 | llvm::ConstantAsMetadata *MD = llvm::ConstantAsMetadata::get(C: Line); |
5724 | llvm::MDTuple *MDT = llvm::MDNode::get(Context&: getLLVMContext(), MDs: {MD}); |
5725 | CI->setMetadata(Kind: "srcloc" , Node: MDT); |
5726 | } |
5727 | |
5728 | // 4. Finish the call. |
5729 | |
5730 | // If the call doesn't return, finish the basic block and clear the |
5731 | // insertion point; this allows the rest of IRGen to discard |
5732 | // unreachable code. |
5733 | if (CI->doesNotReturn()) { |
5734 | if (UnusedReturnSizePtr) |
5735 | PopCleanupBlock(); |
5736 | |
5737 | // Strip away the noreturn attribute to better diagnose unreachable UB. |
5738 | if (SanOpts.has(K: SanitizerKind::Unreachable)) { |
5739 | // Also remove from function since CallBase::hasFnAttr additionally checks |
5740 | // attributes of the called function. |
5741 | if (auto *F = CI->getCalledFunction()) |
5742 | F->removeFnAttr(llvm::Attribute::NoReturn); |
5743 | CI->removeFnAttr(llvm::Attribute::NoReturn); |
5744 | |
5745 | // Avoid incompatibility with ASan which relies on the `noreturn` |
5746 | // attribute to insert handler calls. |
5747 | if (SanOpts.hasOneOf(K: SanitizerKind::Address | |
5748 | SanitizerKind::KernelAddress)) { |
5749 | SanitizerScope SanScope(this); |
5750 | llvm::IRBuilder<>::InsertPointGuard IPGuard(Builder); |
5751 | Builder.SetInsertPoint(CI); |
5752 | auto *FnType = llvm::FunctionType::get(Result: CGM.VoidTy, /*isVarArg=*/false); |
5753 | llvm::FunctionCallee Fn = |
5754 | CGM.CreateRuntimeFunction(Ty: FnType, Name: "__asan_handle_no_return" ); |
5755 | EmitNounwindRuntimeCall(callee: Fn); |
5756 | } |
5757 | } |
5758 | |
5759 | EmitUnreachable(Loc); |
5760 | Builder.ClearInsertionPoint(); |
5761 | |
5762 | // FIXME: For now, emit a dummy basic block because expr emitters in |
5763 | // generally are not ready to handle emitting expressions at unreachable |
5764 | // points. |
5765 | EnsureInsertPoint(); |
5766 | |
5767 | // Return a reasonable RValue. |
5768 | return GetUndefRValue(Ty: RetTy); |
5769 | } |
5770 | |
5771 | // If this is a musttail call, return immediately. We do not branch to the |
5772 | // epilogue in this case. |
5773 | if (IsMustTail) { |
5774 | for (auto it = EHStack.find(sp: CurrentCleanupScopeDepth); it != EHStack.end(); |
5775 | ++it) { |
5776 | EHCleanupScope *Cleanup = dyn_cast<EHCleanupScope>(Val: &*it); |
5777 | if (!(Cleanup && Cleanup->getCleanup()->isRedundantBeforeReturn())) |
5778 | CGM.ErrorUnsupported(MustTailCall, "tail call skipping over cleanups" ); |
5779 | } |
5780 | if (CI->getType()->isVoidTy()) |
5781 | Builder.CreateRetVoid(); |
5782 | else |
5783 | Builder.CreateRet(V: CI); |
5784 | Builder.ClearInsertionPoint(); |
5785 | EnsureInsertPoint(); |
5786 | return GetUndefRValue(Ty: RetTy); |
5787 | } |
5788 | |
5789 | // Perform the swifterror writeback. |
5790 | if (swiftErrorTemp.isValid()) { |
5791 | llvm::Value *errorResult = Builder.CreateLoad(Addr: swiftErrorTemp); |
5792 | Builder.CreateStore(Val: errorResult, Addr: swiftErrorArg); |
5793 | } |
5794 | |
5795 | // Emit any call-associated writebacks immediately. Arguably this |
5796 | // should happen after any return-value munging. |
5797 | if (CallArgs.hasWritebacks()) |
5798 | emitWritebacks(CGF&: *this, args: CallArgs); |
5799 | |
5800 | // The stack cleanup for inalloca arguments has to run out of the normal |
5801 | // lexical order, so deactivate it and run it manually here. |
5802 | CallArgs.freeArgumentMemory(CGF&: *this); |
5803 | |
5804 | // Extract the return value. |
5805 | RValue Ret = [&] { |
5806 | switch (RetAI.getKind()) { |
5807 | case ABIArgInfo::CoerceAndExpand: { |
5808 | auto coercionType = RetAI.getCoerceAndExpandType(); |
5809 | |
5810 | Address addr = SRetPtr.withElementType(ElemTy: coercionType); |
5811 | |
5812 | assert(CI->getType() == RetAI.getUnpaddedCoerceAndExpandType()); |
5813 | bool = isa<llvm::StructType>(Val: CI->getType()); |
5814 | |
5815 | unsigned unpaddedIndex = 0; |
5816 | for (unsigned i = 0, e = coercionType->getNumElements(); i != e; ++i) { |
5817 | llvm::Type *eltType = coercionType->getElementType(N: i); |
5818 | if (ABIArgInfo::isPaddingForCoerceAndExpand(eltType)) continue; |
5819 | Address eltAddr = Builder.CreateStructGEP(Addr: addr, Index: i); |
5820 | llvm::Value *elt = CI; |
5821 | if (requiresExtract) |
5822 | elt = Builder.CreateExtractValue(Agg: elt, Idxs: unpaddedIndex++); |
5823 | else |
5824 | assert(unpaddedIndex == 0); |
5825 | Builder.CreateStore(Val: elt, Addr: eltAddr); |
5826 | } |
5827 | [[fallthrough]]; |
5828 | } |
5829 | |
5830 | case ABIArgInfo::InAlloca: |
5831 | case ABIArgInfo::Indirect: { |
5832 | RValue ret = convertTempToRValue(addr: SRetPtr, type: RetTy, Loc: SourceLocation()); |
5833 | if (UnusedReturnSizePtr) |
5834 | PopCleanupBlock(); |
5835 | return ret; |
5836 | } |
5837 | |
5838 | case ABIArgInfo::Ignore: |
5839 | // If we are ignoring an argument that had a result, make sure to |
5840 | // construct the appropriate return value for our caller. |
5841 | return GetUndefRValue(Ty: RetTy); |
5842 | |
5843 | case ABIArgInfo::Extend: |
5844 | case ABIArgInfo::Direct: { |
5845 | llvm::Type *RetIRTy = ConvertType(T: RetTy); |
5846 | if (RetAI.getCoerceToType() == RetIRTy && RetAI.getDirectOffset() == 0) { |
5847 | switch (getEvaluationKind(T: RetTy)) { |
5848 | case TEK_Complex: { |
5849 | llvm::Value *Real = Builder.CreateExtractValue(Agg: CI, Idxs: 0); |
5850 | llvm::Value *Imag = Builder.CreateExtractValue(Agg: CI, Idxs: 1); |
5851 | return RValue::getComplex(C: std::make_pair(x&: Real, y&: Imag)); |
5852 | } |
5853 | case TEK_Aggregate: { |
5854 | Address DestPtr = ReturnValue.getValue(); |
5855 | bool DestIsVolatile = ReturnValue.isVolatile(); |
5856 | |
5857 | if (!DestPtr.isValid()) { |
5858 | DestPtr = CreateMemTemp(T: RetTy, Name: "agg.tmp" ); |
5859 | DestIsVolatile = false; |
5860 | } |
5861 | EmitAggregateStore(Val: CI, Dest: DestPtr, DestIsVolatile); |
5862 | return RValue::getAggregate(addr: DestPtr); |
5863 | } |
5864 | case TEK_Scalar: { |
5865 | // If the argument doesn't match, perform a bitcast to coerce it. This |
5866 | // can happen due to trivial type mismatches. |
5867 | llvm::Value *V = CI; |
5868 | if (V->getType() != RetIRTy) |
5869 | V = Builder.CreateBitCast(V, DestTy: RetIRTy); |
5870 | return RValue::get(V); |
5871 | } |
5872 | } |
5873 | llvm_unreachable("bad evaluation kind" ); |
5874 | } |
5875 | |
5876 | // If coercing a fixed vector from a scalable vector for ABI |
5877 | // compatibility, and the types match, use the llvm.vector.extract |
5878 | // intrinsic to perform the conversion. |
5879 | if (auto *FixedDstTy = dyn_cast<llvm::FixedVectorType>(Val: RetIRTy)) { |
5880 | llvm::Value *V = CI; |
5881 | if (auto *ScalableSrcTy = |
5882 | dyn_cast<llvm::ScalableVectorType>(Val: V->getType())) { |
5883 | if (FixedDstTy->getElementType() == ScalableSrcTy->getElementType()) { |
5884 | llvm::Value *Zero = llvm::Constant::getNullValue(Ty: CGM.Int64Ty); |
5885 | V = Builder.CreateExtractVector(DstType: FixedDstTy, SrcVec: V, Idx: Zero, Name: "cast.fixed" ); |
5886 | return RValue::get(V); |
5887 | } |
5888 | } |
5889 | } |
5890 | |
5891 | Address DestPtr = ReturnValue.getValue(); |
5892 | bool DestIsVolatile = ReturnValue.isVolatile(); |
5893 | |
5894 | if (!DestPtr.isValid()) { |
5895 | DestPtr = CreateMemTemp(T: RetTy, Name: "coerce" ); |
5896 | DestIsVolatile = false; |
5897 | } |
5898 | |
5899 | // An empty record can overlap other data (if declared with |
5900 | // no_unique_address); omit the store for such types - as there is no |
5901 | // actual data to store. |
5902 | if (!isEmptyRecord(Context&: getContext(), T: RetTy, AllowArrays: true)) { |
5903 | // If the value is offset in memory, apply the offset now. |
5904 | Address StorePtr = emitAddressAtOffset(CGF&: *this, addr: DestPtr, info: RetAI); |
5905 | CreateCoercedStore(Src: CI, Dst: StorePtr, DstIsVolatile: DestIsVolatile, CGF&: *this); |
5906 | } |
5907 | |
5908 | return convertTempToRValue(addr: DestPtr, type: RetTy, Loc: SourceLocation()); |
5909 | } |
5910 | |
5911 | case ABIArgInfo::Expand: |
5912 | case ABIArgInfo::IndirectAliased: |
5913 | llvm_unreachable("Invalid ABI kind for return argument" ); |
5914 | } |
5915 | |
5916 | llvm_unreachable("Unhandled ABIArgInfo::Kind" ); |
5917 | } (); |
5918 | |
5919 | // Emit the assume_aligned check on the return value. |
5920 | if (Ret.isScalar() && TargetDecl) { |
5921 | AssumeAlignedAttrEmitter.EmitAsAnAssumption(Loc, RetTy, Ret); |
5922 | AllocAlignAttrEmitter.EmitAsAnAssumption(Loc, RetTy, Ret); |
5923 | } |
5924 | |
5925 | // Explicitly call CallLifetimeEnd::Emit just to re-use the code even though |
5926 | // we can't use the full cleanup mechanism. |
5927 | for (CallLifetimeEnd &LifetimeEnd : CallLifetimeEndAfterCall) |
5928 | LifetimeEnd.Emit(CGF&: *this, /*Flags=*/flags: {}); |
5929 | |
5930 | if (!ReturnValue.isExternallyDestructed() && |
5931 | RetTy.isDestructedType() == QualType::DK_nontrivial_c_struct) |
5932 | pushDestroy(dtorKind: QualType::DK_nontrivial_c_struct, addr: Ret.getAggregateAddress(), |
5933 | type: RetTy); |
5934 | |
5935 | return Ret; |
5936 | } |
5937 | |
5938 | CGCallee CGCallee::prepareConcreteCallee(CodeGenFunction &CGF) const { |
5939 | if (isVirtual()) { |
5940 | const CallExpr *CE = getVirtualCallExpr(); |
5941 | return CGF.CGM.getCXXABI().getVirtualFunctionPointer( |
5942 | CGF, GD: getVirtualMethodDecl(), This: getThisAddress(), Ty: getVirtualFunctionType(), |
5943 | Loc: CE ? CE->getBeginLoc() : SourceLocation()); |
5944 | } |
5945 | |
5946 | return *this; |
5947 | } |
5948 | |
5949 | /* VarArg handling */ |
5950 | |
5951 | Address CodeGenFunction::EmitVAArg(VAArgExpr *VE, Address &VAListAddr) { |
5952 | VAListAddr = VE->isMicrosoftABI() |
5953 | ? EmitMSVAListRef(E: VE->getSubExpr()) |
5954 | : EmitVAListRef(E: VE->getSubExpr()); |
5955 | QualType Ty = VE->getType(); |
5956 | if (VE->isMicrosoftABI()) |
5957 | return CGM.getTypes().getABIInfo().EmitMSVAArg(CGF&: *this, VAListAddr, Ty); |
5958 | return CGM.getTypes().getABIInfo().EmitVAArg(CGF&: *this, VAListAddr, Ty); |
5959 | } |
5960 | |