1 | //===--- CGDecl.cpp - Emit LLVM Code for declarations ---------------------===// |
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 | // This contains code to emit Decl nodes as LLVM code. |
10 | // |
11 | //===----------------------------------------------------------------------===// |
12 | |
13 | #include "CGBlocks.h" |
14 | #include "CGCXXABI.h" |
15 | #include "CGCleanup.h" |
16 | #include "CGDebugInfo.h" |
17 | #include "CGOpenCLRuntime.h" |
18 | #include "CGOpenMPRuntime.h" |
19 | #include "CodeGenFunction.h" |
20 | #include "CodeGenModule.h" |
21 | #include "ConstantEmitter.h" |
22 | #include "PatternInit.h" |
23 | #include "TargetInfo.h" |
24 | #include "clang/AST/ASTContext.h" |
25 | #include "clang/AST/Attr.h" |
26 | #include "clang/AST/CharUnits.h" |
27 | #include "clang/AST/Decl.h" |
28 | #include "clang/AST/DeclObjC.h" |
29 | #include "clang/AST/DeclOpenMP.h" |
30 | #include "clang/Basic/CodeGenOptions.h" |
31 | #include "clang/Basic/SourceManager.h" |
32 | #include "clang/Basic/TargetInfo.h" |
33 | #include "clang/CodeGen/CGFunctionInfo.h" |
34 | #include "clang/Sema/Sema.h" |
35 | #include "llvm/Analysis/ValueTracking.h" |
36 | #include "llvm/IR/DataLayout.h" |
37 | #include "llvm/IR/GlobalVariable.h" |
38 | #include "llvm/IR/Intrinsics.h" |
39 | #include "llvm/IR/Type.h" |
40 | #include <optional> |
41 | |
42 | using namespace clang; |
43 | using namespace CodeGen; |
44 | |
45 | static_assert(clang::Sema::MaximumAlignment <= llvm::Value::MaximumAlignment, |
46 | "Clang max alignment greater than what LLVM supports?" ); |
47 | |
48 | void CodeGenFunction::EmitDecl(const Decl &D) { |
49 | switch (D.getKind()) { |
50 | case Decl::BuiltinTemplate: |
51 | case Decl::TranslationUnit: |
52 | case Decl::ExternCContext: |
53 | case Decl::Namespace: |
54 | case Decl::UnresolvedUsingTypename: |
55 | case Decl::ClassTemplateSpecialization: |
56 | case Decl::ClassTemplatePartialSpecialization: |
57 | case Decl::VarTemplateSpecialization: |
58 | case Decl::VarTemplatePartialSpecialization: |
59 | case Decl::TemplateTypeParm: |
60 | case Decl::UnresolvedUsingValue: |
61 | case Decl::NonTypeTemplateParm: |
62 | case Decl::CXXDeductionGuide: |
63 | case Decl::CXXMethod: |
64 | case Decl::CXXConstructor: |
65 | case Decl::CXXDestructor: |
66 | case Decl::CXXConversion: |
67 | case Decl::Field: |
68 | case Decl::MSProperty: |
69 | case Decl::IndirectField: |
70 | case Decl::ObjCIvar: |
71 | case Decl::ObjCAtDefsField: |
72 | case Decl::ParmVar: |
73 | case Decl::ImplicitParam: |
74 | case Decl::ClassTemplate: |
75 | case Decl::VarTemplate: |
76 | case Decl::FunctionTemplate: |
77 | case Decl::TypeAliasTemplate: |
78 | case Decl::TemplateTemplateParm: |
79 | case Decl::ObjCMethod: |
80 | case Decl::ObjCCategory: |
81 | case Decl::ObjCProtocol: |
82 | case Decl::ObjCInterface: |
83 | case Decl::ObjCCategoryImpl: |
84 | case Decl::ObjCImplementation: |
85 | case Decl::ObjCProperty: |
86 | case Decl::ObjCCompatibleAlias: |
87 | case Decl::PragmaComment: |
88 | case Decl::PragmaDetectMismatch: |
89 | case Decl::AccessSpec: |
90 | case Decl::LinkageSpec: |
91 | case Decl::Export: |
92 | case Decl::ObjCPropertyImpl: |
93 | case Decl::FileScopeAsm: |
94 | case Decl::TopLevelStmt: |
95 | case Decl::Friend: |
96 | case Decl::FriendTemplate: |
97 | case Decl::Block: |
98 | case Decl::Captured: |
99 | case Decl::UsingShadow: |
100 | case Decl::ConstructorUsingShadow: |
101 | case Decl::ObjCTypeParam: |
102 | case Decl::Binding: |
103 | case Decl::UnresolvedUsingIfExists: |
104 | case Decl::HLSLBuffer: |
105 | llvm_unreachable("Declaration should not be in declstmts!" ); |
106 | case Decl::Record: // struct/union/class X; |
107 | case Decl::CXXRecord: // struct/union/class X; [C++] |
108 | if (CGDebugInfo *DI = getDebugInfo()) |
109 | if (cast<RecordDecl>(Val: D).getDefinition()) |
110 | DI->EmitAndRetainType(Ty: getContext().getRecordType(Decl: cast<RecordDecl>(Val: &D))); |
111 | return; |
112 | case Decl::Enum: // enum X; |
113 | if (CGDebugInfo *DI = getDebugInfo()) |
114 | if (cast<EnumDecl>(Val: D).getDefinition()) |
115 | DI->EmitAndRetainType(Ty: getContext().getEnumType(Decl: cast<EnumDecl>(Val: &D))); |
116 | return; |
117 | case Decl::Function: // void X(); |
118 | case Decl::EnumConstant: // enum ? { X = ? } |
119 | case Decl::StaticAssert: // static_assert(X, ""); [C++0x] |
120 | case Decl::Label: // __label__ x; |
121 | case Decl::Import: |
122 | case Decl::MSGuid: // __declspec(uuid("...")) |
123 | case Decl::UnnamedGlobalConstant: |
124 | case Decl::TemplateParamObject: |
125 | case Decl::OMPThreadPrivate: |
126 | case Decl::OMPAllocate: |
127 | case Decl::OMPCapturedExpr: |
128 | case Decl::OMPRequires: |
129 | case Decl::Empty: |
130 | case Decl::Concept: |
131 | case Decl::ImplicitConceptSpecialization: |
132 | case Decl::LifetimeExtendedTemporary: |
133 | case Decl::RequiresExprBody: |
134 | // None of these decls require codegen support. |
135 | return; |
136 | |
137 | case Decl::NamespaceAlias: |
138 | if (CGDebugInfo *DI = getDebugInfo()) |
139 | DI->EmitNamespaceAlias(NA: cast<NamespaceAliasDecl>(Val: D)); |
140 | return; |
141 | case Decl::Using: // using X; [C++] |
142 | if (CGDebugInfo *DI = getDebugInfo()) |
143 | DI->EmitUsingDecl(UD: cast<UsingDecl>(Val: D)); |
144 | return; |
145 | case Decl::UsingEnum: // using enum X; [C++] |
146 | if (CGDebugInfo *DI = getDebugInfo()) |
147 | DI->EmitUsingEnumDecl(UD: cast<UsingEnumDecl>(Val: D)); |
148 | return; |
149 | case Decl::UsingPack: |
150 | for (auto *Using : cast<UsingPackDecl>(Val: D).expansions()) |
151 | EmitDecl(*Using); |
152 | return; |
153 | case Decl::UsingDirective: // using namespace X; [C++] |
154 | if (CGDebugInfo *DI = getDebugInfo()) |
155 | DI->EmitUsingDirective(UD: cast<UsingDirectiveDecl>(Val: D)); |
156 | return; |
157 | case Decl::Var: |
158 | case Decl::Decomposition: { |
159 | const VarDecl &VD = cast<VarDecl>(Val: D); |
160 | assert(VD.isLocalVarDecl() && |
161 | "Should not see file-scope variables inside a function!" ); |
162 | EmitVarDecl(D: VD); |
163 | if (auto *DD = dyn_cast<DecompositionDecl>(Val: &VD)) |
164 | for (auto *B : DD->bindings()) |
165 | if (auto *HD = B->getHoldingVar()) |
166 | EmitVarDecl(D: *HD); |
167 | return; |
168 | } |
169 | |
170 | case Decl::OMPDeclareReduction: |
171 | return CGM.EmitOMPDeclareReduction(D: cast<OMPDeclareReductionDecl>(Val: &D), CGF: this); |
172 | |
173 | case Decl::OMPDeclareMapper: |
174 | return CGM.EmitOMPDeclareMapper(D: cast<OMPDeclareMapperDecl>(Val: &D), CGF: this); |
175 | |
176 | case Decl::Typedef: // typedef int X; |
177 | case Decl::TypeAlias: { // using X = int; [C++0x] |
178 | QualType Ty = cast<TypedefNameDecl>(Val: D).getUnderlyingType(); |
179 | if (CGDebugInfo *DI = getDebugInfo()) |
180 | DI->EmitAndRetainType(Ty); |
181 | if (Ty->isVariablyModifiedType()) |
182 | EmitVariablyModifiedType(Ty); |
183 | return; |
184 | } |
185 | } |
186 | } |
187 | |
188 | /// EmitVarDecl - This method handles emission of any variable declaration |
189 | /// inside a function, including static vars etc. |
190 | void CodeGenFunction::EmitVarDecl(const VarDecl &D) { |
191 | if (D.hasExternalStorage()) |
192 | // Don't emit it now, allow it to be emitted lazily on its first use. |
193 | return; |
194 | |
195 | // Some function-scope variable does not have static storage but still |
196 | // needs to be emitted like a static variable, e.g. a function-scope |
197 | // variable in constant address space in OpenCL. |
198 | if (D.getStorageDuration() != SD_Automatic) { |
199 | // Static sampler variables translated to function calls. |
200 | if (D.getType()->isSamplerT()) |
201 | return; |
202 | |
203 | llvm::GlobalValue::LinkageTypes Linkage = |
204 | CGM.getLLVMLinkageVarDefinition(VD: &D); |
205 | |
206 | // FIXME: We need to force the emission/use of a guard variable for |
207 | // some variables even if we can constant-evaluate them because |
208 | // we can't guarantee every translation unit will constant-evaluate them. |
209 | |
210 | return EmitStaticVarDecl(D, Linkage); |
211 | } |
212 | |
213 | if (D.getType().getAddressSpace() == LangAS::opencl_local) |
214 | return CGM.getOpenCLRuntime().EmitWorkGroupLocalVarDecl(CGF&: *this, D); |
215 | |
216 | assert(D.hasLocalStorage()); |
217 | return EmitAutoVarDecl(D); |
218 | } |
219 | |
220 | static std::string getStaticDeclName(CodeGenModule &CGM, const VarDecl &D) { |
221 | if (CGM.getLangOpts().CPlusPlus) |
222 | return CGM.getMangledName(GD: &D).str(); |
223 | |
224 | // If this isn't C++, we don't need a mangled name, just a pretty one. |
225 | assert(!D.isExternallyVisible() && "name shouldn't matter" ); |
226 | std::string ContextName; |
227 | const DeclContext *DC = D.getDeclContext(); |
228 | if (auto *CD = dyn_cast<CapturedDecl>(DC)) |
229 | DC = cast<DeclContext>(CD->getNonClosureContext()); |
230 | if (const auto *FD = dyn_cast<FunctionDecl>(DC)) |
231 | ContextName = std::string(CGM.getMangledName(GD: FD)); |
232 | else if (const auto *BD = dyn_cast<BlockDecl>(DC)) |
233 | ContextName = std::string(CGM.getBlockMangledName(GD: GlobalDecl(), BD: BD)); |
234 | else if (const auto *OMD = dyn_cast<ObjCMethodDecl>(DC)) |
235 | ContextName = OMD->getSelector().getAsString(); |
236 | else |
237 | llvm_unreachable("Unknown context for static var decl" ); |
238 | |
239 | ContextName += "." + D.getNameAsString(); |
240 | return ContextName; |
241 | } |
242 | |
243 | llvm::Constant *CodeGenModule::getOrCreateStaticVarDecl( |
244 | const VarDecl &D, llvm::GlobalValue::LinkageTypes Linkage) { |
245 | // In general, we don't always emit static var decls once before we reference |
246 | // them. It is possible to reference them before emitting the function that |
247 | // contains them, and it is possible to emit the containing function multiple |
248 | // times. |
249 | if (llvm::Constant *ExistingGV = StaticLocalDeclMap[&D]) |
250 | return ExistingGV; |
251 | |
252 | QualType Ty = D.getType(); |
253 | assert(Ty->isConstantSizeType() && "VLAs can't be static" ); |
254 | |
255 | // Use the label if the variable is renamed with the asm-label extension. |
256 | std::string Name; |
257 | if (D.hasAttr<AsmLabelAttr>()) |
258 | Name = std::string(getMangledName(GD: &D)); |
259 | else |
260 | Name = getStaticDeclName(CGM&: *this, D); |
261 | |
262 | llvm::Type *LTy = getTypes().ConvertTypeForMem(T: Ty); |
263 | LangAS AS = GetGlobalVarAddressSpace(D: &D); |
264 | unsigned TargetAS = getContext().getTargetAddressSpace(AS); |
265 | |
266 | // OpenCL variables in local address space and CUDA shared |
267 | // variables cannot have an initializer. |
268 | llvm::Constant *Init = nullptr; |
269 | if (Ty.getAddressSpace() == LangAS::opencl_local || |
270 | D.hasAttr<CUDASharedAttr>() || D.hasAttr<LoaderUninitializedAttr>()) |
271 | Init = llvm::UndefValue::get(T: LTy); |
272 | else |
273 | Init = EmitNullConstant(T: Ty); |
274 | |
275 | llvm::GlobalVariable *GV = new llvm::GlobalVariable( |
276 | getModule(), LTy, Ty.isConstant(Ctx: getContext()), Linkage, Init, Name, |
277 | nullptr, llvm::GlobalVariable::NotThreadLocal, TargetAS); |
278 | GV->setAlignment(getContext().getDeclAlign(&D).getAsAlign()); |
279 | |
280 | if (supportsCOMDAT() && GV->isWeakForLinker()) |
281 | GV->setComdat(TheModule.getOrInsertComdat(Name: GV->getName())); |
282 | |
283 | if (D.getTLSKind()) |
284 | setTLSMode(GV, D); |
285 | |
286 | setGVProperties(GV, GD: &D); |
287 | getTargetCodeGenInfo().setTargetAttributes(D: cast<Decl>(Val: &D), GV, M&: *this); |
288 | |
289 | // Make sure the result is of the correct type. |
290 | LangAS ExpectedAS = Ty.getAddressSpace(); |
291 | llvm::Constant *Addr = GV; |
292 | if (AS != ExpectedAS) { |
293 | Addr = getTargetCodeGenInfo().performAddrSpaceCast( |
294 | CGM&: *this, V: GV, SrcAddr: AS, DestAddr: ExpectedAS, |
295 | DestTy: llvm::PointerType::get(C&: getLLVMContext(), |
296 | AddressSpace: getContext().getTargetAddressSpace(AS: ExpectedAS))); |
297 | } |
298 | |
299 | setStaticLocalDeclAddress(D: &D, C: Addr); |
300 | |
301 | // Ensure that the static local gets initialized by making sure the parent |
302 | // function gets emitted eventually. |
303 | const Decl *DC = cast<Decl>(D.getDeclContext()); |
304 | |
305 | // We can't name blocks or captured statements directly, so try to emit their |
306 | // parents. |
307 | if (isa<BlockDecl>(Val: DC) || isa<CapturedDecl>(Val: DC)) { |
308 | DC = DC->getNonClosureContext(); |
309 | // FIXME: Ensure that global blocks get emitted. |
310 | if (!DC) |
311 | return Addr; |
312 | } |
313 | |
314 | GlobalDecl GD; |
315 | if (const auto *CD = dyn_cast<CXXConstructorDecl>(DC)) |
316 | GD = GlobalDecl(CD, Ctor_Base); |
317 | else if (const auto *DD = dyn_cast<CXXDestructorDecl>(DC)) |
318 | GD = GlobalDecl(DD, Dtor_Base); |
319 | else if (const auto *FD = dyn_cast<FunctionDecl>(DC)) |
320 | GD = GlobalDecl(FD); |
321 | else { |
322 | // Don't do anything for Obj-C method decls or global closures. We should |
323 | // never defer them. |
324 | assert(isa<ObjCMethodDecl>(DC) && "unexpected parent code decl" ); |
325 | } |
326 | if (GD.getDecl()) { |
327 | // Disable emission of the parent function for the OpenMP device codegen. |
328 | CGOpenMPRuntime::DisableAutoDeclareTargetRAII NoDeclTarget(*this); |
329 | (void)GetAddrOfGlobal(GD); |
330 | } |
331 | |
332 | return Addr; |
333 | } |
334 | |
335 | /// AddInitializerToStaticVarDecl - Add the initializer for 'D' to the |
336 | /// global variable that has already been created for it. If the initializer |
337 | /// has a different type than GV does, this may free GV and return a different |
338 | /// one. Otherwise it just returns GV. |
339 | llvm::GlobalVariable * |
340 | CodeGenFunction::AddInitializerToStaticVarDecl(const VarDecl &D, |
341 | llvm::GlobalVariable *GV) { |
342 | ConstantEmitter emitter(*this); |
343 | llvm::Constant *Init = emitter.tryEmitForInitializer(D); |
344 | |
345 | // If constant emission failed, then this should be a C++ static |
346 | // initializer. |
347 | if (!Init) { |
348 | if (!getLangOpts().CPlusPlus) |
349 | CGM.ErrorUnsupported(D.getInit(), "constant l-value expression" ); |
350 | else if (D.hasFlexibleArrayInit(Ctx: getContext())) |
351 | CGM.ErrorUnsupported(D.getInit(), "flexible array initializer" ); |
352 | else if (HaveInsertPoint()) { |
353 | // Since we have a static initializer, this global variable can't |
354 | // be constant. |
355 | GV->setConstant(false); |
356 | |
357 | EmitCXXGuardedInit(D, DeclPtr: GV, /*PerformInit*/true); |
358 | } |
359 | return GV; |
360 | } |
361 | |
362 | #ifndef NDEBUG |
363 | CharUnits VarSize = CGM.getContext().getTypeSizeInChars(D.getType()) + |
364 | D.getFlexibleArrayInitChars(Ctx: getContext()); |
365 | CharUnits CstSize = CharUnits::fromQuantity( |
366 | Quantity: CGM.getDataLayout().getTypeAllocSize(Ty: Init->getType())); |
367 | assert(VarSize == CstSize && "Emitted constant has unexpected size" ); |
368 | #endif |
369 | |
370 | // The initializer may differ in type from the global. Rewrite |
371 | // the global to match the initializer. (We have to do this |
372 | // because some types, like unions, can't be completely represented |
373 | // in the LLVM type system.) |
374 | if (GV->getValueType() != Init->getType()) { |
375 | llvm::GlobalVariable *OldGV = GV; |
376 | |
377 | GV = new llvm::GlobalVariable( |
378 | CGM.getModule(), Init->getType(), OldGV->isConstant(), |
379 | OldGV->getLinkage(), Init, "" , |
380 | /*InsertBefore*/ OldGV, OldGV->getThreadLocalMode(), |
381 | OldGV->getType()->getPointerAddressSpace()); |
382 | GV->setVisibility(OldGV->getVisibility()); |
383 | GV->setDSOLocal(OldGV->isDSOLocal()); |
384 | GV->setComdat(OldGV->getComdat()); |
385 | |
386 | // Steal the name of the old global |
387 | GV->takeName(V: OldGV); |
388 | |
389 | // Replace all uses of the old global with the new global |
390 | OldGV->replaceAllUsesWith(V: GV); |
391 | |
392 | // Erase the old global, since it is no longer used. |
393 | OldGV->eraseFromParent(); |
394 | } |
395 | |
396 | bool NeedsDtor = |
397 | D.needsDestruction(Ctx: getContext()) == QualType::DK_cxx_destructor; |
398 | |
399 | GV->setConstant( |
400 | D.getType().isConstantStorage(getContext(), true, !NeedsDtor)); |
401 | GV->setInitializer(Init); |
402 | |
403 | emitter.finalize(global: GV); |
404 | |
405 | if (NeedsDtor && HaveInsertPoint()) { |
406 | // We have a constant initializer, but a nontrivial destructor. We still |
407 | // need to perform a guarded "initialization" in order to register the |
408 | // destructor. |
409 | EmitCXXGuardedInit(D, DeclPtr: GV, /*PerformInit*/false); |
410 | } |
411 | |
412 | return GV; |
413 | } |
414 | |
415 | void CodeGenFunction::EmitStaticVarDecl(const VarDecl &D, |
416 | llvm::GlobalValue::LinkageTypes Linkage) { |
417 | // Check to see if we already have a global variable for this |
418 | // declaration. This can happen when double-emitting function |
419 | // bodies, e.g. with complete and base constructors. |
420 | llvm::Constant *addr = CGM.getOrCreateStaticVarDecl(D, Linkage); |
421 | CharUnits alignment = getContext().getDeclAlign(&D); |
422 | |
423 | // Store into LocalDeclMap before generating initializer to handle |
424 | // circular references. |
425 | llvm::Type *elemTy = ConvertTypeForMem(T: D.getType()); |
426 | setAddrOfLocalVar(VD: &D, Addr: Address(addr, elemTy, alignment)); |
427 | |
428 | // We can't have a VLA here, but we can have a pointer to a VLA, |
429 | // even though that doesn't really make any sense. |
430 | // Make sure to evaluate VLA bounds now so that we have them for later. |
431 | if (D.getType()->isVariablyModifiedType()) |
432 | EmitVariablyModifiedType(Ty: D.getType()); |
433 | |
434 | // Save the type in case adding the initializer forces a type change. |
435 | llvm::Type *expectedType = addr->getType(); |
436 | |
437 | llvm::GlobalVariable *var = |
438 | cast<llvm::GlobalVariable>(Val: addr->stripPointerCasts()); |
439 | |
440 | // CUDA's local and local static __shared__ variables should not |
441 | // have any non-empty initializers. This is ensured by Sema. |
442 | // Whatever initializer such variable may have when it gets here is |
443 | // a no-op and should not be emitted. |
444 | bool isCudaSharedVar = getLangOpts().CUDA && getLangOpts().CUDAIsDevice && |
445 | D.hasAttr<CUDASharedAttr>(); |
446 | // If this value has an initializer, emit it. |
447 | if (D.getInit() && !isCudaSharedVar) |
448 | var = AddInitializerToStaticVarDecl(D, GV: var); |
449 | |
450 | var->setAlignment(alignment.getAsAlign()); |
451 | |
452 | if (D.hasAttr<AnnotateAttr>()) |
453 | CGM.AddGlobalAnnotations(&D, var); |
454 | |
455 | if (auto *SA = D.getAttr<PragmaClangBSSSectionAttr>()) |
456 | var->addAttribute("bss-section" , SA->getName()); |
457 | if (auto *SA = D.getAttr<PragmaClangDataSectionAttr>()) |
458 | var->addAttribute("data-section" , SA->getName()); |
459 | if (auto *SA = D.getAttr<PragmaClangRodataSectionAttr>()) |
460 | var->addAttribute("rodata-section" , SA->getName()); |
461 | if (auto *SA = D.getAttr<PragmaClangRelroSectionAttr>()) |
462 | var->addAttribute("relro-section" , SA->getName()); |
463 | |
464 | if (const SectionAttr *SA = D.getAttr<SectionAttr>()) |
465 | var->setSection(SA->getName()); |
466 | |
467 | if (D.hasAttr<RetainAttr>()) |
468 | CGM.addUsedGlobal(GV: var); |
469 | else if (D.hasAttr<UsedAttr>()) |
470 | CGM.addUsedOrCompilerUsedGlobal(GV: var); |
471 | |
472 | if (CGM.getCodeGenOpts().KeepPersistentStorageVariables) |
473 | CGM.addUsedOrCompilerUsedGlobal(GV: var); |
474 | |
475 | // We may have to cast the constant because of the initializer |
476 | // mismatch above. |
477 | // |
478 | // FIXME: It is really dangerous to store this in the map; if anyone |
479 | // RAUW's the GV uses of this constant will be invalid. |
480 | llvm::Constant *castedAddr = |
481 | llvm::ConstantExpr::getPointerBitCastOrAddrSpaceCast(C: var, Ty: expectedType); |
482 | LocalDeclMap.find(&D)->second = Address(castedAddr, elemTy, alignment); |
483 | CGM.setStaticLocalDeclAddress(D: &D, C: castedAddr); |
484 | |
485 | CGM.getSanitizerMetadata()->reportGlobal(GV: var, D); |
486 | |
487 | // Emit global variable debug descriptor for static vars. |
488 | CGDebugInfo *DI = getDebugInfo(); |
489 | if (DI && CGM.getCodeGenOpts().hasReducedDebugInfo()) { |
490 | DI->setLocation(D.getLocation()); |
491 | DI->EmitGlobalVariable(GV: var, Decl: &D); |
492 | } |
493 | } |
494 | |
495 | namespace { |
496 | struct DestroyObject final : EHScopeStack::Cleanup { |
497 | DestroyObject(Address addr, QualType type, |
498 | CodeGenFunction::Destroyer *destroyer, |
499 | bool useEHCleanupForArray) |
500 | : addr(addr), type(type), destroyer(destroyer), |
501 | useEHCleanupForArray(useEHCleanupForArray) {} |
502 | |
503 | Address addr; |
504 | QualType type; |
505 | CodeGenFunction::Destroyer *destroyer; |
506 | bool useEHCleanupForArray; |
507 | |
508 | void Emit(CodeGenFunction &CGF, Flags flags) override { |
509 | // Don't use an EH cleanup recursively from an EH cleanup. |
510 | bool useEHCleanupForArray = |
511 | flags.isForNormalCleanup() && this->useEHCleanupForArray; |
512 | |
513 | CGF.emitDestroy(addr, type, destroyer, useEHCleanupForArray); |
514 | } |
515 | }; |
516 | |
517 | template <class Derived> |
518 | struct DestroyNRVOVariable : EHScopeStack::Cleanup { |
519 | DestroyNRVOVariable(Address addr, QualType type, llvm::Value *NRVOFlag) |
520 | : NRVOFlag(NRVOFlag), Loc(addr), Ty(type) {} |
521 | |
522 | llvm::Value *NRVOFlag; |
523 | Address Loc; |
524 | QualType Ty; |
525 | |
526 | void Emit(CodeGenFunction &CGF, Flags flags) override { |
527 | // Along the exceptions path we always execute the dtor. |
528 | bool NRVO = flags.isForNormalCleanup() && NRVOFlag; |
529 | |
530 | llvm::BasicBlock *SkipDtorBB = nullptr; |
531 | if (NRVO) { |
532 | // If we exited via NRVO, we skip the destructor call. |
533 | llvm::BasicBlock *RunDtorBB = CGF.createBasicBlock(name: "nrvo.unused" ); |
534 | SkipDtorBB = CGF.createBasicBlock(name: "nrvo.skipdtor" ); |
535 | llvm::Value *DidNRVO = |
536 | CGF.Builder.CreateFlagLoad(Addr: NRVOFlag, Name: "nrvo.val" ); |
537 | CGF.Builder.CreateCondBr(Cond: DidNRVO, True: SkipDtorBB, False: RunDtorBB); |
538 | CGF.EmitBlock(BB: RunDtorBB); |
539 | } |
540 | |
541 | static_cast<Derived *>(this)->emitDestructorCall(CGF); |
542 | |
543 | if (NRVO) CGF.EmitBlock(BB: SkipDtorBB); |
544 | } |
545 | |
546 | virtual ~DestroyNRVOVariable() = default; |
547 | }; |
548 | |
549 | struct DestroyNRVOVariableCXX final |
550 | : DestroyNRVOVariable<DestroyNRVOVariableCXX> { |
551 | DestroyNRVOVariableCXX(Address addr, QualType type, |
552 | const CXXDestructorDecl *Dtor, llvm::Value *NRVOFlag) |
553 | : DestroyNRVOVariable<DestroyNRVOVariableCXX>(addr, type, NRVOFlag), |
554 | Dtor(Dtor) {} |
555 | |
556 | const CXXDestructorDecl *Dtor; |
557 | |
558 | void emitDestructorCall(CodeGenFunction &CGF) { |
559 | CGF.EmitCXXDestructorCall(Dtor, Dtor_Complete, |
560 | /*ForVirtualBase=*/false, |
561 | /*Delegating=*/false, Loc, Ty); |
562 | } |
563 | }; |
564 | |
565 | struct DestroyNRVOVariableC final |
566 | : DestroyNRVOVariable<DestroyNRVOVariableC> { |
567 | DestroyNRVOVariableC(Address addr, llvm::Value *NRVOFlag, QualType Ty) |
568 | : DestroyNRVOVariable<DestroyNRVOVariableC>(addr, Ty, NRVOFlag) {} |
569 | |
570 | void emitDestructorCall(CodeGenFunction &CGF) { |
571 | CGF.destroyNonTrivialCStruct(CGF, Loc, Ty); |
572 | } |
573 | }; |
574 | |
575 | struct CallStackRestore final : EHScopeStack::Cleanup { |
576 | Address Stack; |
577 | CallStackRestore(Address Stack) : Stack(Stack) {} |
578 | bool isRedundantBeforeReturn() override { return true; } |
579 | void Emit(CodeGenFunction &CGF, Flags flags) override { |
580 | llvm::Value *V = CGF.Builder.CreateLoad(Addr: Stack); |
581 | CGF.Builder.CreateStackRestore(Ptr: V); |
582 | } |
583 | }; |
584 | |
585 | struct KmpcAllocFree final : EHScopeStack::Cleanup { |
586 | std::pair<llvm::Value *, llvm::Value *> AddrSizePair; |
587 | KmpcAllocFree(const std::pair<llvm::Value *, llvm::Value *> &AddrSizePair) |
588 | : AddrSizePair(AddrSizePair) {} |
589 | void Emit(CodeGenFunction &CGF, Flags EmissionFlags) override { |
590 | auto &RT = CGF.CGM.getOpenMPRuntime(); |
591 | RT.getKmpcFreeShared(CGF, AddrSizePair); |
592 | } |
593 | }; |
594 | |
595 | struct ExtendGCLifetime final : EHScopeStack::Cleanup { |
596 | const VarDecl &Var; |
597 | ExtendGCLifetime(const VarDecl *var) : Var(*var) {} |
598 | |
599 | void Emit(CodeGenFunction &CGF, Flags flags) override { |
600 | // Compute the address of the local variable, in case it's a |
601 | // byref or something. |
602 | DeclRefExpr DRE(CGF.getContext(), const_cast<VarDecl *>(&Var), false, |
603 | Var.getType(), VK_LValue, SourceLocation()); |
604 | llvm::Value *value = CGF.EmitLoadOfScalar(lvalue: CGF.EmitDeclRefLValue(E: &DRE), |
605 | Loc: SourceLocation()); |
606 | CGF.EmitExtendGCLifetime(object: value); |
607 | } |
608 | }; |
609 | |
610 | struct CallCleanupFunction final : EHScopeStack::Cleanup { |
611 | llvm::Constant *CleanupFn; |
612 | const CGFunctionInfo &FnInfo; |
613 | const VarDecl &Var; |
614 | |
615 | CallCleanupFunction(llvm::Constant *CleanupFn, const CGFunctionInfo *Info, |
616 | const VarDecl *Var) |
617 | : CleanupFn(CleanupFn), FnInfo(*Info), Var(*Var) {} |
618 | |
619 | void Emit(CodeGenFunction &CGF, Flags flags) override { |
620 | DeclRefExpr DRE(CGF.getContext(), const_cast<VarDecl *>(&Var), false, |
621 | Var.getType(), VK_LValue, SourceLocation()); |
622 | // Compute the address of the local variable, in case it's a byref |
623 | // or something. |
624 | llvm::Value *Addr = CGF.EmitDeclRefLValue(E: &DRE).getPointer(CGF); |
625 | |
626 | // In some cases, the type of the function argument will be different from |
627 | // the type of the pointer. An example of this is |
628 | // void f(void* arg); |
629 | // __attribute__((cleanup(f))) void *g; |
630 | // |
631 | // To fix this we insert a bitcast here. |
632 | QualType ArgTy = FnInfo.arg_begin()->type; |
633 | llvm::Value *Arg = |
634 | CGF.Builder.CreateBitCast(V: Addr, DestTy: CGF.ConvertType(T: ArgTy)); |
635 | |
636 | CallArgList Args; |
637 | Args.add(rvalue: RValue::get(V: Arg), |
638 | type: CGF.getContext().getPointerType(Var.getType())); |
639 | auto Callee = CGCallee::forDirect(functionPtr: CleanupFn); |
640 | CGF.EmitCall(CallInfo: FnInfo, Callee, ReturnValue: ReturnValueSlot(), Args); |
641 | } |
642 | }; |
643 | } // end anonymous namespace |
644 | |
645 | /// EmitAutoVarWithLifetime - Does the setup required for an automatic |
646 | /// variable with lifetime. |
647 | static void EmitAutoVarWithLifetime(CodeGenFunction &CGF, const VarDecl &var, |
648 | Address addr, |
649 | Qualifiers::ObjCLifetime lifetime) { |
650 | switch (lifetime) { |
651 | case Qualifiers::OCL_None: |
652 | llvm_unreachable("present but none" ); |
653 | |
654 | case Qualifiers::OCL_ExplicitNone: |
655 | // nothing to do |
656 | break; |
657 | |
658 | case Qualifiers::OCL_Strong: { |
659 | CodeGenFunction::Destroyer *destroyer = |
660 | (var.hasAttr<ObjCPreciseLifetimeAttr>() |
661 | ? CodeGenFunction::destroyARCStrongPrecise |
662 | : CodeGenFunction::destroyARCStrongImprecise); |
663 | |
664 | CleanupKind cleanupKind = CGF.getARCCleanupKind(); |
665 | CGF.pushDestroy(cleanupKind, addr, var.getType(), destroyer, |
666 | cleanupKind & EHCleanup); |
667 | break; |
668 | } |
669 | case Qualifiers::OCL_Autoreleasing: |
670 | // nothing to do |
671 | break; |
672 | |
673 | case Qualifiers::OCL_Weak: |
674 | // __weak objects always get EH cleanups; otherwise, exceptions |
675 | // could cause really nasty crashes instead of mere leaks. |
676 | CGF.pushDestroy(NormalAndEHCleanup, addr, var.getType(), |
677 | CodeGenFunction::destroyARCWeak, |
678 | /*useEHCleanup*/ true); |
679 | break; |
680 | } |
681 | } |
682 | |
683 | static bool isAccessedBy(const VarDecl &var, const Stmt *s) { |
684 | if (const Expr *e = dyn_cast<Expr>(Val: s)) { |
685 | // Skip the most common kinds of expressions that make |
686 | // hierarchy-walking expensive. |
687 | s = e = e->IgnoreParenCasts(); |
688 | |
689 | if (const DeclRefExpr *ref = dyn_cast<DeclRefExpr>(Val: e)) |
690 | return (ref->getDecl() == &var); |
691 | if (const BlockExpr *be = dyn_cast<BlockExpr>(Val: e)) { |
692 | const BlockDecl *block = be->getBlockDecl(); |
693 | for (const auto &I : block->captures()) { |
694 | if (I.getVariable() == &var) |
695 | return true; |
696 | } |
697 | } |
698 | } |
699 | |
700 | for (const Stmt *SubStmt : s->children()) |
701 | // SubStmt might be null; as in missing decl or conditional of an if-stmt. |
702 | if (SubStmt && isAccessedBy(var, s: SubStmt)) |
703 | return true; |
704 | |
705 | return false; |
706 | } |
707 | |
708 | static bool isAccessedBy(const ValueDecl *decl, const Expr *e) { |
709 | if (!decl) return false; |
710 | if (!isa<VarDecl>(Val: decl)) return false; |
711 | const VarDecl *var = cast<VarDecl>(Val: decl); |
712 | return isAccessedBy(*var, e); |
713 | } |
714 | |
715 | static bool tryEmitARCCopyWeakInit(CodeGenFunction &CGF, |
716 | const LValue &destLV, const Expr *init) { |
717 | bool needsCast = false; |
718 | |
719 | while (auto castExpr = dyn_cast<CastExpr>(Val: init->IgnoreParens())) { |
720 | switch (castExpr->getCastKind()) { |
721 | // Look through casts that don't require representation changes. |
722 | case CK_NoOp: |
723 | case CK_BitCast: |
724 | case CK_BlockPointerToObjCPointerCast: |
725 | needsCast = true; |
726 | break; |
727 | |
728 | // If we find an l-value to r-value cast from a __weak variable, |
729 | // emit this operation as a copy or move. |
730 | case CK_LValueToRValue: { |
731 | const Expr *srcExpr = castExpr->getSubExpr(); |
732 | if (srcExpr->getType().getObjCLifetime() != Qualifiers::OCL_Weak) |
733 | return false; |
734 | |
735 | // Emit the source l-value. |
736 | LValue srcLV = CGF.EmitLValue(E: srcExpr); |
737 | |
738 | // Handle a formal type change to avoid asserting. |
739 | auto srcAddr = srcLV.getAddress(CGF); |
740 | if (needsCast) { |
741 | srcAddr = |
742 | srcAddr.withElementType(ElemTy: destLV.getAddress(CGF).getElementType()); |
743 | } |
744 | |
745 | // If it was an l-value, use objc_copyWeak. |
746 | if (srcExpr->isLValue()) { |
747 | CGF.EmitARCCopyWeak(dst: destLV.getAddress(CGF), src: srcAddr); |
748 | } else { |
749 | assert(srcExpr->isXValue()); |
750 | CGF.EmitARCMoveWeak(dst: destLV.getAddress(CGF), src: srcAddr); |
751 | } |
752 | return true; |
753 | } |
754 | |
755 | // Stop at anything else. |
756 | default: |
757 | return false; |
758 | } |
759 | |
760 | init = castExpr->getSubExpr(); |
761 | } |
762 | return false; |
763 | } |
764 | |
765 | static void drillIntoBlockVariable(CodeGenFunction &CGF, |
766 | LValue &lvalue, |
767 | const VarDecl *var) { |
768 | lvalue.setAddress(CGF.emitBlockByrefAddress(baseAddr: lvalue.getAddress(CGF), V: var)); |
769 | } |
770 | |
771 | void CodeGenFunction::EmitNullabilityCheck(LValue LHS, llvm::Value *RHS, |
772 | SourceLocation Loc) { |
773 | if (!SanOpts.has(K: SanitizerKind::NullabilityAssign)) |
774 | return; |
775 | |
776 | auto Nullability = LHS.getType()->getNullability(); |
777 | if (!Nullability || *Nullability != NullabilityKind::NonNull) |
778 | return; |
779 | |
780 | // Check if the right hand side of the assignment is nonnull, if the left |
781 | // hand side must be nonnull. |
782 | SanitizerScope SanScope(this); |
783 | llvm::Value *IsNotNull = Builder.CreateIsNotNull(Arg: RHS); |
784 | llvm::Constant *StaticData[] = { |
785 | EmitCheckSourceLocation(Loc), EmitCheckTypeDescriptor(T: LHS.getType()), |
786 | llvm::ConstantInt::get(Ty: Int8Ty, V: 0), // The LogAlignment info is unused. |
787 | llvm::ConstantInt::get(Ty: Int8Ty, V: TCK_NonnullAssign)}; |
788 | EmitCheck(Checked: {{IsNotNull, SanitizerKind::NullabilityAssign}}, |
789 | Check: SanitizerHandler::TypeMismatch, StaticArgs: StaticData, DynamicArgs: RHS); |
790 | } |
791 | |
792 | void CodeGenFunction::EmitScalarInit(const Expr *init, const ValueDecl *D, |
793 | LValue lvalue, bool capturedByInit) { |
794 | Qualifiers::ObjCLifetime lifetime = lvalue.getObjCLifetime(); |
795 | if (!lifetime) { |
796 | llvm::Value *value = EmitScalarExpr(E: init); |
797 | if (capturedByInit) |
798 | drillIntoBlockVariable(CGF&: *this, lvalue, var: cast<VarDecl>(Val: D)); |
799 | EmitNullabilityCheck(LHS: lvalue, RHS: value, Loc: init->getExprLoc()); |
800 | EmitStoreThroughLValue(Src: RValue::get(V: value), Dst: lvalue, isInit: true); |
801 | return; |
802 | } |
803 | |
804 | if (const CXXDefaultInitExpr *DIE = dyn_cast<CXXDefaultInitExpr>(Val: init)) |
805 | init = DIE->getExpr(); |
806 | |
807 | // If we're emitting a value with lifetime, we have to do the |
808 | // initialization *before* we leave the cleanup scopes. |
809 | if (auto *EWC = dyn_cast<ExprWithCleanups>(Val: init)) { |
810 | CodeGenFunction::RunCleanupsScope Scope(*this); |
811 | return EmitScalarInit(init: EWC->getSubExpr(), D, lvalue, capturedByInit); |
812 | } |
813 | |
814 | // We have to maintain the illusion that the variable is |
815 | // zero-initialized. If the variable might be accessed in its |
816 | // initializer, zero-initialize before running the initializer, then |
817 | // actually perform the initialization with an assign. |
818 | bool accessedByInit = false; |
819 | if (lifetime != Qualifiers::OCL_ExplicitNone) |
820 | accessedByInit = (capturedByInit || isAccessedBy(decl: D, e: init)); |
821 | if (accessedByInit) { |
822 | LValue tempLV = lvalue; |
823 | // Drill down to the __block object if necessary. |
824 | if (capturedByInit) { |
825 | // We can use a simple GEP for this because it can't have been |
826 | // moved yet. |
827 | tempLV.setAddress(emitBlockByrefAddress(baseAddr: tempLV.getAddress(CGF&: *this), |
828 | V: cast<VarDecl>(Val: D), |
829 | /*follow*/ followForward: false)); |
830 | } |
831 | |
832 | auto ty = |
833 | cast<llvm::PointerType>(Val: tempLV.getAddress(CGF&: *this).getElementType()); |
834 | llvm::Value *zero = CGM.getNullPointer(T: ty, QT: tempLV.getType()); |
835 | |
836 | // If __weak, we want to use a barrier under certain conditions. |
837 | if (lifetime == Qualifiers::OCL_Weak) |
838 | EmitARCInitWeak(addr: tempLV.getAddress(CGF&: *this), value: zero); |
839 | |
840 | // Otherwise just do a simple store. |
841 | else |
842 | EmitStoreOfScalar(value: zero, lvalue: tempLV, /* isInitialization */ isInit: true); |
843 | } |
844 | |
845 | // Emit the initializer. |
846 | llvm::Value *value = nullptr; |
847 | |
848 | switch (lifetime) { |
849 | case Qualifiers::OCL_None: |
850 | llvm_unreachable("present but none" ); |
851 | |
852 | case Qualifiers::OCL_Strong: { |
853 | if (!D || !isa<VarDecl>(Val: D) || !cast<VarDecl>(Val: D)->isARCPseudoStrong()) { |
854 | value = EmitARCRetainScalarExpr(expr: init); |
855 | break; |
856 | } |
857 | // If D is pseudo-strong, treat it like __unsafe_unretained here. This means |
858 | // that we omit the retain, and causes non-autoreleased return values to be |
859 | // immediately released. |
860 | [[fallthrough]]; |
861 | } |
862 | |
863 | case Qualifiers::OCL_ExplicitNone: |
864 | value = EmitARCUnsafeUnretainedScalarExpr(expr: init); |
865 | break; |
866 | |
867 | case Qualifiers::OCL_Weak: { |
868 | // If it's not accessed by the initializer, try to emit the |
869 | // initialization with a copy or move. |
870 | if (!accessedByInit && tryEmitARCCopyWeakInit(CGF&: *this, destLV: lvalue, init)) { |
871 | return; |
872 | } |
873 | |
874 | // No way to optimize a producing initializer into this. It's not |
875 | // worth optimizing for, because the value will immediately |
876 | // disappear in the common case. |
877 | value = EmitScalarExpr(E: init); |
878 | |
879 | if (capturedByInit) drillIntoBlockVariable(CGF&: *this, lvalue, var: cast<VarDecl>(Val: D)); |
880 | if (accessedByInit) |
881 | EmitARCStoreWeak(addr: lvalue.getAddress(CGF&: *this), value, /*ignored*/ true); |
882 | else |
883 | EmitARCInitWeak(addr: lvalue.getAddress(CGF&: *this), value); |
884 | return; |
885 | } |
886 | |
887 | case Qualifiers::OCL_Autoreleasing: |
888 | value = EmitARCRetainAutoreleaseScalarExpr(expr: init); |
889 | break; |
890 | } |
891 | |
892 | if (capturedByInit) drillIntoBlockVariable(CGF&: *this, lvalue, var: cast<VarDecl>(Val: D)); |
893 | |
894 | EmitNullabilityCheck(LHS: lvalue, RHS: value, Loc: init->getExprLoc()); |
895 | |
896 | // If the variable might have been accessed by its initializer, we |
897 | // might have to initialize with a barrier. We have to do this for |
898 | // both __weak and __strong, but __weak got filtered out above. |
899 | if (accessedByInit && lifetime == Qualifiers::OCL_Strong) { |
900 | llvm::Value *oldValue = EmitLoadOfScalar(lvalue, Loc: init->getExprLoc()); |
901 | EmitStoreOfScalar(value, lvalue, /* isInitialization */ isInit: true); |
902 | EmitARCRelease(value: oldValue, precise: ARCImpreciseLifetime); |
903 | return; |
904 | } |
905 | |
906 | EmitStoreOfScalar(value, lvalue, /* isInitialization */ isInit: true); |
907 | } |
908 | |
909 | /// Decide whether we can emit the non-zero parts of the specified initializer |
910 | /// with equal or fewer than NumStores scalar stores. |
911 | static bool canEmitInitWithFewStoresAfterBZero(llvm::Constant *Init, |
912 | unsigned &NumStores) { |
913 | // Zero and Undef never requires any extra stores. |
914 | if (isa<llvm::ConstantAggregateZero>(Val: Init) || |
915 | isa<llvm::ConstantPointerNull>(Val: Init) || |
916 | isa<llvm::UndefValue>(Val: Init)) |
917 | return true; |
918 | if (isa<llvm::ConstantInt>(Val: Init) || isa<llvm::ConstantFP>(Val: Init) || |
919 | isa<llvm::ConstantVector>(Val: Init) || isa<llvm::BlockAddress>(Val: Init) || |
920 | isa<llvm::ConstantExpr>(Val: Init)) |
921 | return Init->isNullValue() || NumStores--; |
922 | |
923 | // See if we can emit each element. |
924 | if (isa<llvm::ConstantArray>(Val: Init) || isa<llvm::ConstantStruct>(Val: Init)) { |
925 | for (unsigned i = 0, e = Init->getNumOperands(); i != e; ++i) { |
926 | llvm::Constant *Elt = cast<llvm::Constant>(Val: Init->getOperand(i)); |
927 | if (!canEmitInitWithFewStoresAfterBZero(Init: Elt, NumStores)) |
928 | return false; |
929 | } |
930 | return true; |
931 | } |
932 | |
933 | if (llvm::ConstantDataSequential *CDS = |
934 | dyn_cast<llvm::ConstantDataSequential>(Val: Init)) { |
935 | for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) { |
936 | llvm::Constant *Elt = CDS->getElementAsConstant(i); |
937 | if (!canEmitInitWithFewStoresAfterBZero(Init: Elt, NumStores)) |
938 | return false; |
939 | } |
940 | return true; |
941 | } |
942 | |
943 | // Anything else is hard and scary. |
944 | return false; |
945 | } |
946 | |
947 | /// For inits that canEmitInitWithFewStoresAfterBZero returned true for, emit |
948 | /// the scalar stores that would be required. |
949 | static void emitStoresForInitAfterBZero(CodeGenModule &CGM, |
950 | llvm::Constant *Init, Address Loc, |
951 | bool isVolatile, CGBuilderTy &Builder, |
952 | bool IsAutoInit) { |
953 | assert(!Init->isNullValue() && !isa<llvm::UndefValue>(Init) && |
954 | "called emitStoresForInitAfterBZero for zero or undef value." ); |
955 | |
956 | if (isa<llvm::ConstantInt>(Val: Init) || isa<llvm::ConstantFP>(Val: Init) || |
957 | isa<llvm::ConstantVector>(Val: Init) || isa<llvm::BlockAddress>(Val: Init) || |
958 | isa<llvm::ConstantExpr>(Val: Init)) { |
959 | auto *I = Builder.CreateStore(Val: Init, Addr: Loc, IsVolatile: isVolatile); |
960 | if (IsAutoInit) |
961 | I->addAnnotationMetadata(Annotation: "auto-init" ); |
962 | return; |
963 | } |
964 | |
965 | if (llvm::ConstantDataSequential *CDS = |
966 | dyn_cast<llvm::ConstantDataSequential>(Val: Init)) { |
967 | for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) { |
968 | llvm::Constant *Elt = CDS->getElementAsConstant(i); |
969 | |
970 | // If necessary, get a pointer to the element and emit it. |
971 | if (!Elt->isNullValue() && !isa<llvm::UndefValue>(Val: Elt)) |
972 | emitStoresForInitAfterBZero( |
973 | CGM, Init: Elt, Loc: Builder.CreateConstInBoundsGEP2_32(Addr: Loc, Idx0: 0, Idx1: i), isVolatile, |
974 | Builder, IsAutoInit); |
975 | } |
976 | return; |
977 | } |
978 | |
979 | assert((isa<llvm::ConstantStruct>(Init) || isa<llvm::ConstantArray>(Init)) && |
980 | "Unknown value type!" ); |
981 | |
982 | for (unsigned i = 0, e = Init->getNumOperands(); i != e; ++i) { |
983 | llvm::Constant *Elt = cast<llvm::Constant>(Val: Init->getOperand(i)); |
984 | |
985 | // If necessary, get a pointer to the element and emit it. |
986 | if (!Elt->isNullValue() && !isa<llvm::UndefValue>(Val: Elt)) |
987 | emitStoresForInitAfterBZero(CGM, Init: Elt, |
988 | Loc: Builder.CreateConstInBoundsGEP2_32(Addr: Loc, Idx0: 0, Idx1: i), |
989 | isVolatile, Builder, IsAutoInit); |
990 | } |
991 | } |
992 | |
993 | /// Decide whether we should use bzero plus some stores to initialize a local |
994 | /// variable instead of using a memcpy from a constant global. It is beneficial |
995 | /// to use bzero if the global is all zeros, or mostly zeros and large. |
996 | static bool shouldUseBZeroPlusStoresToInitialize(llvm::Constant *Init, |
997 | uint64_t GlobalSize) { |
998 | // If a global is all zeros, always use a bzero. |
999 | if (isa<llvm::ConstantAggregateZero>(Val: Init)) return true; |
1000 | |
1001 | // If a non-zero global is <= 32 bytes, always use a memcpy. If it is large, |
1002 | // do it if it will require 6 or fewer scalar stores. |
1003 | // TODO: Should budget depends on the size? Avoiding a large global warrants |
1004 | // plopping in more stores. |
1005 | unsigned StoreBudget = 6; |
1006 | uint64_t SizeLimit = 32; |
1007 | |
1008 | return GlobalSize > SizeLimit && |
1009 | canEmitInitWithFewStoresAfterBZero(Init, NumStores&: StoreBudget); |
1010 | } |
1011 | |
1012 | /// Decide whether we should use memset to initialize a local variable instead |
1013 | /// of using a memcpy from a constant global. Assumes we've already decided to |
1014 | /// not user bzero. |
1015 | /// FIXME We could be more clever, as we are for bzero above, and generate |
1016 | /// memset followed by stores. It's unclear that's worth the effort. |
1017 | static llvm::Value *shouldUseMemSetToInitialize(llvm::Constant *Init, |
1018 | uint64_t GlobalSize, |
1019 | const llvm::DataLayout &DL) { |
1020 | uint64_t SizeLimit = 32; |
1021 | if (GlobalSize <= SizeLimit) |
1022 | return nullptr; |
1023 | return llvm::isBytewiseValue(V: Init, DL); |
1024 | } |
1025 | |
1026 | /// Decide whether we want to split a constant structure or array store into a |
1027 | /// sequence of its fields' stores. This may cost us code size and compilation |
1028 | /// speed, but plays better with store optimizations. |
1029 | static bool shouldSplitConstantStore(CodeGenModule &CGM, |
1030 | uint64_t GlobalByteSize) { |
1031 | // Don't break things that occupy more than one cacheline. |
1032 | uint64_t ByteSizeLimit = 64; |
1033 | if (CGM.getCodeGenOpts().OptimizationLevel == 0) |
1034 | return false; |
1035 | if (GlobalByteSize <= ByteSizeLimit) |
1036 | return true; |
1037 | return false; |
1038 | } |
1039 | |
1040 | enum class IsPattern { No, Yes }; |
1041 | |
1042 | /// Generate a constant filled with either a pattern or zeroes. |
1043 | static llvm::Constant *patternOrZeroFor(CodeGenModule &CGM, IsPattern isPattern, |
1044 | llvm::Type *Ty) { |
1045 | if (isPattern == IsPattern::Yes) |
1046 | return initializationPatternFor(CGM, Ty); |
1047 | else |
1048 | return llvm::Constant::getNullValue(Ty); |
1049 | } |
1050 | |
1051 | static llvm::Constant *constWithPadding(CodeGenModule &CGM, IsPattern isPattern, |
1052 | llvm::Constant *constant); |
1053 | |
1054 | /// Helper function for constWithPadding() to deal with padding in structures. |
1055 | static llvm::Constant *constStructWithPadding(CodeGenModule &CGM, |
1056 | IsPattern isPattern, |
1057 | llvm::StructType *STy, |
1058 | llvm::Constant *constant) { |
1059 | const llvm::DataLayout &DL = CGM.getDataLayout(); |
1060 | const llvm::StructLayout *Layout = DL.getStructLayout(Ty: STy); |
1061 | llvm::Type *Int8Ty = llvm::IntegerType::getInt8Ty(C&: CGM.getLLVMContext()); |
1062 | unsigned SizeSoFar = 0; |
1063 | SmallVector<llvm::Constant *, 8> Values; |
1064 | bool NestedIntact = true; |
1065 | for (unsigned i = 0, e = STy->getNumElements(); i != e; i++) { |
1066 | unsigned CurOff = Layout->getElementOffset(Idx: i); |
1067 | if (SizeSoFar < CurOff) { |
1068 | assert(!STy->isPacked()); |
1069 | auto *PadTy = llvm::ArrayType::get(ElementType: Int8Ty, NumElements: CurOff - SizeSoFar); |
1070 | Values.push_back(Elt: patternOrZeroFor(CGM, isPattern, Ty: PadTy)); |
1071 | } |
1072 | llvm::Constant *CurOp; |
1073 | if (constant->isZeroValue()) |
1074 | CurOp = llvm::Constant::getNullValue(Ty: STy->getElementType(N: i)); |
1075 | else |
1076 | CurOp = cast<llvm::Constant>(Val: constant->getAggregateElement(Elt: i)); |
1077 | auto *NewOp = constWithPadding(CGM, isPattern, constant: CurOp); |
1078 | if (CurOp != NewOp) |
1079 | NestedIntact = false; |
1080 | Values.push_back(Elt: NewOp); |
1081 | SizeSoFar = CurOff + DL.getTypeAllocSize(Ty: CurOp->getType()); |
1082 | } |
1083 | unsigned TotalSize = Layout->getSizeInBytes(); |
1084 | if (SizeSoFar < TotalSize) { |
1085 | auto *PadTy = llvm::ArrayType::get(ElementType: Int8Ty, NumElements: TotalSize - SizeSoFar); |
1086 | Values.push_back(Elt: patternOrZeroFor(CGM, isPattern, Ty: PadTy)); |
1087 | } |
1088 | if (NestedIntact && Values.size() == STy->getNumElements()) |
1089 | return constant; |
1090 | return llvm::ConstantStruct::getAnon(V: Values, Packed: STy->isPacked()); |
1091 | } |
1092 | |
1093 | /// Replace all padding bytes in a given constant with either a pattern byte or |
1094 | /// 0x00. |
1095 | static llvm::Constant *constWithPadding(CodeGenModule &CGM, IsPattern isPattern, |
1096 | llvm::Constant *constant) { |
1097 | llvm::Type *OrigTy = constant->getType(); |
1098 | if (const auto STy = dyn_cast<llvm::StructType>(Val: OrigTy)) |
1099 | return constStructWithPadding(CGM, isPattern, STy, constant); |
1100 | if (auto *ArrayTy = dyn_cast<llvm::ArrayType>(Val: OrigTy)) { |
1101 | llvm::SmallVector<llvm::Constant *, 8> Values; |
1102 | uint64_t Size = ArrayTy->getNumElements(); |
1103 | if (!Size) |
1104 | return constant; |
1105 | llvm::Type *ElemTy = ArrayTy->getElementType(); |
1106 | bool ZeroInitializer = constant->isNullValue(); |
1107 | llvm::Constant *OpValue, *PaddedOp; |
1108 | if (ZeroInitializer) { |
1109 | OpValue = llvm::Constant::getNullValue(Ty: ElemTy); |
1110 | PaddedOp = constWithPadding(CGM, isPattern, constant: OpValue); |
1111 | } |
1112 | for (unsigned Op = 0; Op != Size; ++Op) { |
1113 | if (!ZeroInitializer) { |
1114 | OpValue = constant->getAggregateElement(Elt: Op); |
1115 | PaddedOp = constWithPadding(CGM, isPattern, constant: OpValue); |
1116 | } |
1117 | Values.push_back(Elt: PaddedOp); |
1118 | } |
1119 | auto *NewElemTy = Values[0]->getType(); |
1120 | if (NewElemTy == ElemTy) |
1121 | return constant; |
1122 | auto *NewArrayTy = llvm::ArrayType::get(ElementType: NewElemTy, NumElements: Size); |
1123 | return llvm::ConstantArray::get(T: NewArrayTy, V: Values); |
1124 | } |
1125 | // FIXME: Add handling for tail padding in vectors. Vectors don't |
1126 | // have padding between or inside elements, but the total amount of |
1127 | // data can be less than the allocated size. |
1128 | return constant; |
1129 | } |
1130 | |
1131 | Address CodeGenModule::createUnnamedGlobalFrom(const VarDecl &D, |
1132 | llvm::Constant *Constant, |
1133 | CharUnits Align) { |
1134 | auto FunctionName = [&](const DeclContext *DC) -> std::string { |
1135 | if (const auto *FD = dyn_cast<FunctionDecl>(Val: DC)) { |
1136 | if (const auto *CC = dyn_cast<CXXConstructorDecl>(Val: FD)) |
1137 | return CC->getNameAsString(); |
1138 | if (const auto *CD = dyn_cast<CXXDestructorDecl>(Val: FD)) |
1139 | return CD->getNameAsString(); |
1140 | return std::string(getMangledName(GD: FD)); |
1141 | } else if (const auto *OM = dyn_cast<ObjCMethodDecl>(Val: DC)) { |
1142 | return OM->getNameAsString(); |
1143 | } else if (isa<BlockDecl>(Val: DC)) { |
1144 | return "<block>" ; |
1145 | } else if (isa<CapturedDecl>(Val: DC)) { |
1146 | return "<captured>" ; |
1147 | } else { |
1148 | llvm_unreachable("expected a function or method" ); |
1149 | } |
1150 | }; |
1151 | |
1152 | // Form a simple per-variable cache of these values in case we find we |
1153 | // want to reuse them. |
1154 | llvm::GlobalVariable *&CacheEntry = InitializerConstants[&D]; |
1155 | if (!CacheEntry || CacheEntry->getInitializer() != Constant) { |
1156 | auto *Ty = Constant->getType(); |
1157 | bool isConstant = true; |
1158 | llvm::GlobalVariable *InsertBefore = nullptr; |
1159 | unsigned AS = |
1160 | getContext().getTargetAddressSpace(AS: GetGlobalConstantAddressSpace()); |
1161 | std::string Name; |
1162 | if (D.hasGlobalStorage()) |
1163 | Name = getMangledName(GD: &D).str() + ".const" ; |
1164 | else if (const DeclContext *DC = D.getParentFunctionOrMethod()) |
1165 | Name = ("__const." + FunctionName(DC) + "." + D.getName()).str(); |
1166 | else |
1167 | llvm_unreachable("local variable has no parent function or method" ); |
1168 | llvm::GlobalVariable *GV = new llvm::GlobalVariable( |
1169 | getModule(), Ty, isConstant, llvm::GlobalValue::PrivateLinkage, |
1170 | Constant, Name, InsertBefore, llvm::GlobalValue::NotThreadLocal, AS); |
1171 | GV->setAlignment(Align.getAsAlign()); |
1172 | GV->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::Global); |
1173 | CacheEntry = GV; |
1174 | } else if (CacheEntry->getAlignment() < uint64_t(Align.getQuantity())) { |
1175 | CacheEntry->setAlignment(Align.getAsAlign()); |
1176 | } |
1177 | |
1178 | return Address(CacheEntry, CacheEntry->getValueType(), Align); |
1179 | } |
1180 | |
1181 | static Address createUnnamedGlobalForMemcpyFrom(CodeGenModule &CGM, |
1182 | const VarDecl &D, |
1183 | CGBuilderTy &Builder, |
1184 | llvm::Constant *Constant, |
1185 | CharUnits Align) { |
1186 | Address SrcPtr = CGM.createUnnamedGlobalFrom(D, Constant, Align); |
1187 | return SrcPtr.withElementType(ElemTy: CGM.Int8Ty); |
1188 | } |
1189 | |
1190 | static void emitStoresForConstant(CodeGenModule &CGM, const VarDecl &D, |
1191 | Address Loc, bool isVolatile, |
1192 | CGBuilderTy &Builder, |
1193 | llvm::Constant *constant, bool IsAutoInit) { |
1194 | auto *Ty = constant->getType(); |
1195 | uint64_t ConstantSize = CGM.getDataLayout().getTypeAllocSize(Ty); |
1196 | if (!ConstantSize) |
1197 | return; |
1198 | |
1199 | bool canDoSingleStore = Ty->isIntOrIntVectorTy() || |
1200 | Ty->isPtrOrPtrVectorTy() || Ty->isFPOrFPVectorTy(); |
1201 | if (canDoSingleStore) { |
1202 | auto *I = Builder.CreateStore(Val: constant, Addr: Loc, IsVolatile: isVolatile); |
1203 | if (IsAutoInit) |
1204 | I->addAnnotationMetadata(Annotation: "auto-init" ); |
1205 | return; |
1206 | } |
1207 | |
1208 | auto *SizeVal = llvm::ConstantInt::get(Ty: CGM.IntPtrTy, V: ConstantSize); |
1209 | |
1210 | // If the initializer is all or mostly the same, codegen with bzero / memset |
1211 | // then do a few stores afterward. |
1212 | if (shouldUseBZeroPlusStoresToInitialize(Init: constant, GlobalSize: ConstantSize)) { |
1213 | auto *I = Builder.CreateMemSet(Dest: Loc, Value: llvm::ConstantInt::get(Ty: CGM.Int8Ty, V: 0), |
1214 | Size: SizeVal, IsVolatile: isVolatile); |
1215 | if (IsAutoInit) |
1216 | I->addAnnotationMetadata(Annotation: "auto-init" ); |
1217 | |
1218 | bool valueAlreadyCorrect = |
1219 | constant->isNullValue() || isa<llvm::UndefValue>(Val: constant); |
1220 | if (!valueAlreadyCorrect) { |
1221 | Loc = Loc.withElementType(ElemTy: Ty); |
1222 | emitStoresForInitAfterBZero(CGM, Init: constant, Loc, isVolatile, Builder, |
1223 | IsAutoInit); |
1224 | } |
1225 | return; |
1226 | } |
1227 | |
1228 | // If the initializer is a repeated byte pattern, use memset. |
1229 | llvm::Value *Pattern = |
1230 | shouldUseMemSetToInitialize(Init: constant, GlobalSize: ConstantSize, DL: CGM.getDataLayout()); |
1231 | if (Pattern) { |
1232 | uint64_t Value = 0x00; |
1233 | if (!isa<llvm::UndefValue>(Val: Pattern)) { |
1234 | const llvm::APInt &AP = cast<llvm::ConstantInt>(Val: Pattern)->getValue(); |
1235 | assert(AP.getBitWidth() <= 8); |
1236 | Value = AP.getLimitedValue(); |
1237 | } |
1238 | auto *I = Builder.CreateMemSet( |
1239 | Dest: Loc, Value: llvm::ConstantInt::get(Ty: CGM.Int8Ty, V: Value), Size: SizeVal, IsVolatile: isVolatile); |
1240 | if (IsAutoInit) |
1241 | I->addAnnotationMetadata(Annotation: "auto-init" ); |
1242 | return; |
1243 | } |
1244 | |
1245 | // If the initializer is small or trivialAutoVarInit is set, use a handful of |
1246 | // stores. |
1247 | bool IsTrivialAutoVarInitPattern = |
1248 | CGM.getContext().getLangOpts().getTrivialAutoVarInit() == |
1249 | LangOptions::TrivialAutoVarInitKind::Pattern; |
1250 | if (shouldSplitConstantStore(CGM, GlobalByteSize: ConstantSize)) { |
1251 | if (auto *STy = dyn_cast<llvm::StructType>(Val: Ty)) { |
1252 | if (STy == Loc.getElementType() || |
1253 | (STy != Loc.getElementType() && IsTrivialAutoVarInitPattern)) { |
1254 | const llvm::StructLayout *Layout = |
1255 | CGM.getDataLayout().getStructLayout(Ty: STy); |
1256 | for (unsigned i = 0; i != constant->getNumOperands(); i++) { |
1257 | CharUnits CurOff = |
1258 | CharUnits::fromQuantity(Quantity: Layout->getElementOffset(Idx: i)); |
1259 | Address EltPtr = Builder.CreateConstInBoundsByteGEP( |
1260 | Addr: Loc.withElementType(ElemTy: CGM.Int8Ty), Offset: CurOff); |
1261 | emitStoresForConstant(CGM, D, Loc: EltPtr, isVolatile, Builder, |
1262 | constant: constant->getAggregateElement(Elt: i), IsAutoInit); |
1263 | } |
1264 | return; |
1265 | } |
1266 | } else if (auto *ATy = dyn_cast<llvm::ArrayType>(Val: Ty)) { |
1267 | if (ATy == Loc.getElementType() || |
1268 | (ATy != Loc.getElementType() && IsTrivialAutoVarInitPattern)) { |
1269 | for (unsigned i = 0; i != ATy->getNumElements(); i++) { |
1270 | Address EltPtr = Builder.CreateConstGEP( |
1271 | Addr: Loc.withElementType(ElemTy: ATy->getElementType()), Index: i); |
1272 | emitStoresForConstant(CGM, D, Loc: EltPtr, isVolatile, Builder, |
1273 | constant: constant->getAggregateElement(Elt: i), IsAutoInit); |
1274 | } |
1275 | return; |
1276 | } |
1277 | } |
1278 | } |
1279 | |
1280 | // Copy from a global. |
1281 | auto *I = |
1282 | Builder.CreateMemCpy(Dest: Loc, |
1283 | Src: createUnnamedGlobalForMemcpyFrom( |
1284 | CGM, D, Builder, Constant: constant, Align: Loc.getAlignment()), |
1285 | Size: SizeVal, IsVolatile: isVolatile); |
1286 | if (IsAutoInit) |
1287 | I->addAnnotationMetadata(Annotation: "auto-init" ); |
1288 | } |
1289 | |
1290 | static void emitStoresForZeroInit(CodeGenModule &CGM, const VarDecl &D, |
1291 | Address Loc, bool isVolatile, |
1292 | CGBuilderTy &Builder) { |
1293 | llvm::Type *ElTy = Loc.getElementType(); |
1294 | llvm::Constant *constant = |
1295 | constWithPadding(CGM, isPattern: IsPattern::No, constant: llvm::Constant::getNullValue(Ty: ElTy)); |
1296 | emitStoresForConstant(CGM, D, Loc, isVolatile, Builder, constant, |
1297 | /*IsAutoInit=*/true); |
1298 | } |
1299 | |
1300 | static void emitStoresForPatternInit(CodeGenModule &CGM, const VarDecl &D, |
1301 | Address Loc, bool isVolatile, |
1302 | CGBuilderTy &Builder) { |
1303 | llvm::Type *ElTy = Loc.getElementType(); |
1304 | llvm::Constant *constant = constWithPadding( |
1305 | CGM, isPattern: IsPattern::Yes, constant: initializationPatternFor(CGM, ElTy)); |
1306 | assert(!isa<llvm::UndefValue>(constant)); |
1307 | emitStoresForConstant(CGM, D, Loc, isVolatile, Builder, constant, |
1308 | /*IsAutoInit=*/true); |
1309 | } |
1310 | |
1311 | static bool containsUndef(llvm::Constant *constant) { |
1312 | auto *Ty = constant->getType(); |
1313 | if (isa<llvm::UndefValue>(Val: constant)) |
1314 | return true; |
1315 | if (Ty->isStructTy() || Ty->isArrayTy() || Ty->isVectorTy()) |
1316 | for (llvm::Use &Op : constant->operands()) |
1317 | if (containsUndef(constant: cast<llvm::Constant>(Val&: Op))) |
1318 | return true; |
1319 | return false; |
1320 | } |
1321 | |
1322 | static llvm::Constant *replaceUndef(CodeGenModule &CGM, IsPattern isPattern, |
1323 | llvm::Constant *constant) { |
1324 | auto *Ty = constant->getType(); |
1325 | if (isa<llvm::UndefValue>(Val: constant)) |
1326 | return patternOrZeroFor(CGM, isPattern, Ty); |
1327 | if (!(Ty->isStructTy() || Ty->isArrayTy() || Ty->isVectorTy())) |
1328 | return constant; |
1329 | if (!containsUndef(constant)) |
1330 | return constant; |
1331 | llvm::SmallVector<llvm::Constant *, 8> Values(constant->getNumOperands()); |
1332 | for (unsigned Op = 0, NumOp = constant->getNumOperands(); Op != NumOp; ++Op) { |
1333 | auto *OpValue = cast<llvm::Constant>(Val: constant->getOperand(i: Op)); |
1334 | Values[Op] = replaceUndef(CGM, isPattern, constant: OpValue); |
1335 | } |
1336 | if (Ty->isStructTy()) |
1337 | return llvm::ConstantStruct::get(T: cast<llvm::StructType>(Val: Ty), V: Values); |
1338 | if (Ty->isArrayTy()) |
1339 | return llvm::ConstantArray::get(T: cast<llvm::ArrayType>(Val: Ty), V: Values); |
1340 | assert(Ty->isVectorTy()); |
1341 | return llvm::ConstantVector::get(V: Values); |
1342 | } |
1343 | |
1344 | /// EmitAutoVarDecl - Emit code and set up an entry in LocalDeclMap for a |
1345 | /// variable declaration with auto, register, or no storage class specifier. |
1346 | /// These turn into simple stack objects, or GlobalValues depending on target. |
1347 | void CodeGenFunction::EmitAutoVarDecl(const VarDecl &D) { |
1348 | AutoVarEmission emission = EmitAutoVarAlloca(var: D); |
1349 | EmitAutoVarInit(emission); |
1350 | EmitAutoVarCleanups(emission); |
1351 | } |
1352 | |
1353 | /// Emit a lifetime.begin marker if some criteria are satisfied. |
1354 | /// \return a pointer to the temporary size Value if a marker was emitted, null |
1355 | /// otherwise |
1356 | llvm::Value *CodeGenFunction::EmitLifetimeStart(llvm::TypeSize Size, |
1357 | llvm::Value *Addr) { |
1358 | if (!ShouldEmitLifetimeMarkers) |
1359 | return nullptr; |
1360 | |
1361 | assert(Addr->getType()->getPointerAddressSpace() == |
1362 | CGM.getDataLayout().getAllocaAddrSpace() && |
1363 | "Pointer should be in alloca address space" ); |
1364 | llvm::Value *SizeV = llvm::ConstantInt::get( |
1365 | Ty: Int64Ty, V: Size.isScalable() ? -1 : Size.getFixedValue()); |
1366 | llvm::CallInst *C = |
1367 | Builder.CreateCall(Callee: CGM.getLLVMLifetimeStartFn(), Args: {SizeV, Addr}); |
1368 | C->setDoesNotThrow(); |
1369 | return SizeV; |
1370 | } |
1371 | |
1372 | void CodeGenFunction::EmitLifetimeEnd(llvm::Value *Size, llvm::Value *Addr) { |
1373 | assert(Addr->getType()->getPointerAddressSpace() == |
1374 | CGM.getDataLayout().getAllocaAddrSpace() && |
1375 | "Pointer should be in alloca address space" ); |
1376 | llvm::CallInst *C = |
1377 | Builder.CreateCall(Callee: CGM.getLLVMLifetimeEndFn(), Args: {Size, Addr}); |
1378 | C->setDoesNotThrow(); |
1379 | } |
1380 | |
1381 | void CodeGenFunction::EmitAndRegisterVariableArrayDimensions( |
1382 | CGDebugInfo *DI, const VarDecl &D, bool EmitDebugInfo) { |
1383 | // For each dimension stores its QualType and corresponding |
1384 | // size-expression Value. |
1385 | SmallVector<CodeGenFunction::VlaSizePair, 4> Dimensions; |
1386 | SmallVector<const IdentifierInfo *, 4> VLAExprNames; |
1387 | |
1388 | // Break down the array into individual dimensions. |
1389 | QualType Type1D = D.getType(); |
1390 | while (getContext().getAsVariableArrayType(T: Type1D)) { |
1391 | auto VlaSize = getVLAElements1D(vla: Type1D); |
1392 | if (auto *C = dyn_cast<llvm::ConstantInt>(VlaSize.NumElts)) |
1393 | Dimensions.emplace_back(C, Type1D.getUnqualifiedType()); |
1394 | else { |
1395 | // Generate a locally unique name for the size expression. |
1396 | Twine Name = Twine("__vla_expr" ) + Twine(VLAExprCounter++); |
1397 | SmallString<12> Buffer; |
1398 | StringRef NameRef = Name.toStringRef(Out&: Buffer); |
1399 | auto &Ident = getContext().Idents.getOwn(Name: NameRef); |
1400 | VLAExprNames.push_back(Elt: &Ident); |
1401 | auto SizeExprAddr = |
1402 | CreateDefaultAlignTempAlloca(Ty: VlaSize.NumElts->getType(), Name: NameRef); |
1403 | Builder.CreateStore(Val: VlaSize.NumElts, Addr: SizeExprAddr); |
1404 | Dimensions.emplace_back(SizeExprAddr.getPointer(), |
1405 | Type1D.getUnqualifiedType()); |
1406 | } |
1407 | Type1D = VlaSize.Type; |
1408 | } |
1409 | |
1410 | if (!EmitDebugInfo) |
1411 | return; |
1412 | |
1413 | // Register each dimension's size-expression with a DILocalVariable, |
1414 | // so that it can be used by CGDebugInfo when instantiating a DISubrange |
1415 | // to describe this array. |
1416 | unsigned NameIdx = 0; |
1417 | for (auto &VlaSize : Dimensions) { |
1418 | llvm::Metadata *MD; |
1419 | if (auto *C = dyn_cast<llvm::ConstantInt>(Val: VlaSize.NumElts)) |
1420 | MD = llvm::ConstantAsMetadata::get(C); |
1421 | else { |
1422 | // Create an artificial VarDecl to generate debug info for. |
1423 | const IdentifierInfo *NameIdent = VLAExprNames[NameIdx++]; |
1424 | auto QT = getContext().getIntTypeForBitwidth( |
1425 | DestWidth: SizeTy->getScalarSizeInBits(), Signed: false); |
1426 | auto *ArtificialDecl = VarDecl::Create( |
1427 | C&: getContext(), DC: const_cast<DeclContext *>(D.getDeclContext()), |
1428 | StartLoc: D.getLocation(), IdLoc: D.getLocation(), Id: NameIdent, T: QT, |
1429 | TInfo: getContext().CreateTypeSourceInfo(T: QT), S: SC_Auto); |
1430 | ArtificialDecl->setImplicit(); |
1431 | |
1432 | MD = DI->EmitDeclareOfAutoVariable(Decl: ArtificialDecl, AI: VlaSize.NumElts, |
1433 | Builder); |
1434 | } |
1435 | assert(MD && "No Size expression debug node created" ); |
1436 | DI->registerVLASizeExpression(Ty: VlaSize.Type, SizeExpr: MD); |
1437 | } |
1438 | } |
1439 | |
1440 | /// EmitAutoVarAlloca - Emit the alloca and debug information for a |
1441 | /// local variable. Does not emit initialization or destruction. |
1442 | CodeGenFunction::AutoVarEmission |
1443 | CodeGenFunction::EmitAutoVarAlloca(const VarDecl &D) { |
1444 | QualType Ty = D.getType(); |
1445 | assert( |
1446 | Ty.getAddressSpace() == LangAS::Default || |
1447 | (Ty.getAddressSpace() == LangAS::opencl_private && getLangOpts().OpenCL)); |
1448 | |
1449 | AutoVarEmission emission(D); |
1450 | |
1451 | bool isEscapingByRef = D.isEscapingByref(); |
1452 | emission.IsEscapingByRef = isEscapingByRef; |
1453 | |
1454 | CharUnits alignment = getContext().getDeclAlign(&D); |
1455 | |
1456 | // If the type is variably-modified, emit all the VLA sizes for it. |
1457 | if (Ty->isVariablyModifiedType()) |
1458 | EmitVariablyModifiedType(Ty); |
1459 | |
1460 | auto *DI = getDebugInfo(); |
1461 | bool EmitDebugInfo = DI && CGM.getCodeGenOpts().hasReducedDebugInfo(); |
1462 | |
1463 | Address address = Address::invalid(); |
1464 | RawAddress AllocaAddr = RawAddress::invalid(); |
1465 | Address OpenMPLocalAddr = Address::invalid(); |
1466 | if (CGM.getLangOpts().OpenMPIRBuilder) |
1467 | OpenMPLocalAddr = OMPBuilderCBHelpers::getAddressOfLocalVariable(CGF&: *this, VD: &D); |
1468 | else |
1469 | OpenMPLocalAddr = |
1470 | getLangOpts().OpenMP |
1471 | ? CGM.getOpenMPRuntime().getAddressOfLocalVariable(CGF&: *this, VD: &D) |
1472 | : Address::invalid(); |
1473 | |
1474 | bool NRVO = getLangOpts().ElideConstructors && D.isNRVOVariable(); |
1475 | |
1476 | if (getLangOpts().OpenMP && OpenMPLocalAddr.isValid()) { |
1477 | address = OpenMPLocalAddr; |
1478 | AllocaAddr = OpenMPLocalAddr; |
1479 | } else if (Ty->isConstantSizeType()) { |
1480 | // If this value is an array or struct with a statically determinable |
1481 | // constant initializer, there are optimizations we can do. |
1482 | // |
1483 | // TODO: We should constant-evaluate the initializer of any variable, |
1484 | // as long as it is initialized by a constant expression. Currently, |
1485 | // isConstantInitializer produces wrong answers for structs with |
1486 | // reference or bitfield members, and a few other cases, and checking |
1487 | // for POD-ness protects us from some of these. |
1488 | if (D.getInit() && (Ty->isArrayType() || Ty->isRecordType()) && |
1489 | (D.isConstexpr() || |
1490 | ((Ty.isPODType(Context: getContext()) || |
1491 | getContext().getBaseElementType(QT: Ty)->isObjCObjectPointerType()) && |
1492 | D.getInit()->isConstantInitializer(Ctx&: getContext(), ForRef: false)))) { |
1493 | |
1494 | // If the variable's a const type, and it's neither an NRVO |
1495 | // candidate nor a __block variable and has no mutable members, |
1496 | // emit it as a global instead. |
1497 | // Exception is if a variable is located in non-constant address space |
1498 | // in OpenCL. |
1499 | bool NeedsDtor = |
1500 | D.needsDestruction(Ctx: getContext()) == QualType::DK_cxx_destructor; |
1501 | if ((!getLangOpts().OpenCL || |
1502 | Ty.getAddressSpace() == LangAS::opencl_constant) && |
1503 | (CGM.getCodeGenOpts().MergeAllConstants && !NRVO && |
1504 | !isEscapingByRef && |
1505 | Ty.isConstantStorage(Ctx: getContext(), ExcludeCtor: true, ExcludeDtor: !NeedsDtor))) { |
1506 | EmitStaticVarDecl(D, Linkage: llvm::GlobalValue::InternalLinkage); |
1507 | |
1508 | // Signal this condition to later callbacks. |
1509 | emission.Addr = Address::invalid(); |
1510 | assert(emission.wasEmittedAsGlobal()); |
1511 | return emission; |
1512 | } |
1513 | |
1514 | // Otherwise, tell the initialization code that we're in this case. |
1515 | emission.IsConstantAggregate = true; |
1516 | } |
1517 | |
1518 | // A normal fixed sized variable becomes an alloca in the entry block, |
1519 | // unless: |
1520 | // - it's an NRVO variable. |
1521 | // - we are compiling OpenMP and it's an OpenMP local variable. |
1522 | if (NRVO) { |
1523 | // The named return value optimization: allocate this variable in the |
1524 | // return slot, so that we can elide the copy when returning this |
1525 | // variable (C++0x [class.copy]p34). |
1526 | address = ReturnValue; |
1527 | AllocaAddr = |
1528 | RawAddress(ReturnValue.emitRawPointer(CGF&: *this), |
1529 | ReturnValue.getElementType(), ReturnValue.getAlignment()); |
1530 | ; |
1531 | |
1532 | if (const RecordType *RecordTy = Ty->getAs<RecordType>()) { |
1533 | const auto *RD = RecordTy->getDecl(); |
1534 | const auto *CXXRD = dyn_cast<CXXRecordDecl>(RD); |
1535 | if ((CXXRD && !CXXRD->hasTrivialDestructor()) || |
1536 | RD->isNonTrivialToPrimitiveDestroy()) { |
1537 | // Create a flag that is used to indicate when the NRVO was applied |
1538 | // to this variable. Set it to zero to indicate that NRVO was not |
1539 | // applied. |
1540 | llvm::Value *Zero = Builder.getFalse(); |
1541 | RawAddress NRVOFlag = |
1542 | CreateTempAlloca(Ty: Zero->getType(), align: CharUnits::One(), Name: "nrvo" ); |
1543 | EnsureInsertPoint(); |
1544 | Builder.CreateStore(Val: Zero, Addr: NRVOFlag); |
1545 | |
1546 | // Record the NRVO flag for this variable. |
1547 | NRVOFlags[&D] = NRVOFlag.getPointer(); |
1548 | emission.NRVOFlag = NRVOFlag.getPointer(); |
1549 | } |
1550 | } |
1551 | } else { |
1552 | CharUnits allocaAlignment; |
1553 | llvm::Type *allocaTy; |
1554 | if (isEscapingByRef) { |
1555 | auto &byrefInfo = getBlockByrefInfo(var: &D); |
1556 | allocaTy = byrefInfo.Type; |
1557 | allocaAlignment = byrefInfo.ByrefAlignment; |
1558 | } else { |
1559 | allocaTy = ConvertTypeForMem(T: Ty); |
1560 | allocaAlignment = alignment; |
1561 | } |
1562 | |
1563 | // Create the alloca. Note that we set the name separately from |
1564 | // building the instruction so that it's there even in no-asserts |
1565 | // builds. |
1566 | address = CreateTempAlloca(allocaTy, allocaAlignment, D.getName(), |
1567 | /*ArraySize=*/nullptr, &AllocaAddr); |
1568 | |
1569 | // Don't emit lifetime markers for MSVC catch parameters. The lifetime of |
1570 | // the catch parameter starts in the catchpad instruction, and we can't |
1571 | // insert code in those basic blocks. |
1572 | bool IsMSCatchParam = |
1573 | D.isExceptionVariable() && getTarget().getCXXABI().isMicrosoft(); |
1574 | |
1575 | // Emit a lifetime intrinsic if meaningful. There's no point in doing this |
1576 | // if we don't have a valid insertion point (?). |
1577 | if (HaveInsertPoint() && !IsMSCatchParam) { |
1578 | // If there's a jump into the lifetime of this variable, its lifetime |
1579 | // gets broken up into several regions in IR, which requires more work |
1580 | // to handle correctly. For now, just omit the intrinsics; this is a |
1581 | // rare case, and it's better to just be conservatively correct. |
1582 | // PR28267. |
1583 | // |
1584 | // We have to do this in all language modes if there's a jump past the |
1585 | // declaration. We also have to do it in C if there's a jump to an |
1586 | // earlier point in the current block because non-VLA lifetimes begin as |
1587 | // soon as the containing block is entered, not when its variables |
1588 | // actually come into scope; suppressing the lifetime annotations |
1589 | // completely in this case is unnecessarily pessimistic, but again, this |
1590 | // is rare. |
1591 | if (!Bypasses.IsBypassed(D: &D) && |
1592 | !(!getLangOpts().CPlusPlus && hasLabelBeenSeenInCurrentScope())) { |
1593 | llvm::TypeSize Size = CGM.getDataLayout().getTypeAllocSize(Ty: allocaTy); |
1594 | emission.SizeForLifetimeMarkers = |
1595 | EmitLifetimeStart(Size, Addr: AllocaAddr.getPointer()); |
1596 | } |
1597 | } else { |
1598 | assert(!emission.useLifetimeMarkers()); |
1599 | } |
1600 | } |
1601 | } else { |
1602 | EnsureInsertPoint(); |
1603 | |
1604 | // Delayed globalization for variable length declarations. This ensures that |
1605 | // the expression representing the length has been emitted and can be used |
1606 | // by the definition of the VLA. Since this is an escaped declaration, in |
1607 | // OpenMP we have to use a call to __kmpc_alloc_shared(). The matching |
1608 | // deallocation call to __kmpc_free_shared() is emitted later. |
1609 | bool VarAllocated = false; |
1610 | if (getLangOpts().OpenMPIsTargetDevice) { |
1611 | auto &RT = CGM.getOpenMPRuntime(); |
1612 | if (RT.isDelayedVariableLengthDecl(CGF&: *this, VD: &D)) { |
1613 | // Emit call to __kmpc_alloc_shared() instead of the alloca. |
1614 | std::pair<llvm::Value *, llvm::Value *> AddrSizePair = |
1615 | RT.getKmpcAllocShared(CGF&: *this, VD: &D); |
1616 | |
1617 | // Save the address of the allocation: |
1618 | LValue Base = MakeAddrLValue(AddrSizePair.first, D.getType(), |
1619 | CGM.getContext().getDeclAlign(&D), |
1620 | AlignmentSource::Decl); |
1621 | address = Base.getAddress(CGF&: *this); |
1622 | |
1623 | // Push a cleanup block to emit the call to __kmpc_free_shared in the |
1624 | // appropriate location at the end of the scope of the |
1625 | // __kmpc_alloc_shared functions: |
1626 | pushKmpcAllocFree(Kind: NormalCleanup, AddrSizePair); |
1627 | |
1628 | // Mark variable as allocated: |
1629 | VarAllocated = true; |
1630 | } |
1631 | } |
1632 | |
1633 | if (!VarAllocated) { |
1634 | if (!DidCallStackSave) { |
1635 | // Save the stack. |
1636 | Address Stack = |
1637 | CreateDefaultAlignTempAlloca(Ty: AllocaInt8PtrTy, Name: "saved_stack" ); |
1638 | |
1639 | llvm::Value *V = Builder.CreateStackSave(); |
1640 | assert(V->getType() == AllocaInt8PtrTy); |
1641 | Builder.CreateStore(Val: V, Addr: Stack); |
1642 | |
1643 | DidCallStackSave = true; |
1644 | |
1645 | // Push a cleanup block and restore the stack there. |
1646 | // FIXME: in general circumstances, this should be an EH cleanup. |
1647 | pushStackRestore(kind: NormalCleanup, SPMem: Stack); |
1648 | } |
1649 | |
1650 | auto VlaSize = getVLASize(vla: Ty); |
1651 | llvm::Type *llvmTy = ConvertTypeForMem(T: VlaSize.Type); |
1652 | |
1653 | // Allocate memory for the array. |
1654 | address = CreateTempAlloca(llvmTy, alignment, "vla" , VlaSize.NumElts, |
1655 | &AllocaAddr); |
1656 | } |
1657 | |
1658 | // If we have debug info enabled, properly describe the VLA dimensions for |
1659 | // this type by registering the vla size expression for each of the |
1660 | // dimensions. |
1661 | EmitAndRegisterVariableArrayDimensions(DI, D, EmitDebugInfo); |
1662 | } |
1663 | |
1664 | setAddrOfLocalVar(VD: &D, Addr: address); |
1665 | emission.Addr = address; |
1666 | emission.AllocaAddr = AllocaAddr; |
1667 | |
1668 | // Emit debug info for local var declaration. |
1669 | if (EmitDebugInfo && HaveInsertPoint()) { |
1670 | Address DebugAddr = address; |
1671 | bool UsePointerValue = NRVO && ReturnValuePointer.isValid(); |
1672 | DI->setLocation(D.getLocation()); |
1673 | |
1674 | // If NRVO, use a pointer to the return address. |
1675 | if (UsePointerValue) { |
1676 | DebugAddr = ReturnValuePointer; |
1677 | AllocaAddr = ReturnValuePointer; |
1678 | } |
1679 | (void)DI->EmitDeclareOfAutoVariable(Decl: &D, AI: AllocaAddr.getPointer(), Builder, |
1680 | UsePointerValue); |
1681 | } |
1682 | |
1683 | if (D.hasAttr<AnnotateAttr>() && HaveInsertPoint()) |
1684 | EmitVarAnnotations(D: &D, V: address.emitRawPointer(CGF&: *this)); |
1685 | |
1686 | // Make sure we call @llvm.lifetime.end. |
1687 | if (emission.useLifetimeMarkers()) |
1688 | EHStack.pushCleanup<CallLifetimeEnd>(Kind: NormalEHLifetimeMarker, |
1689 | A: emission.getOriginalAllocatedAddress(), |
1690 | A: emission.getSizeForLifetimeMarkers()); |
1691 | |
1692 | return emission; |
1693 | } |
1694 | |
1695 | static bool isCapturedBy(const VarDecl &, const Expr *); |
1696 | |
1697 | /// Determines whether the given __block variable is potentially |
1698 | /// captured by the given statement. |
1699 | static bool isCapturedBy(const VarDecl &Var, const Stmt *S) { |
1700 | if (const Expr *E = dyn_cast<Expr>(Val: S)) |
1701 | return isCapturedBy(Var, E); |
1702 | for (const Stmt *SubStmt : S->children()) |
1703 | if (isCapturedBy(Var, S: SubStmt)) |
1704 | return true; |
1705 | return false; |
1706 | } |
1707 | |
1708 | /// Determines whether the given __block variable is potentially |
1709 | /// captured by the given expression. |
1710 | static bool isCapturedBy(const VarDecl &Var, const Expr *E) { |
1711 | // Skip the most common kinds of expressions that make |
1712 | // hierarchy-walking expensive. |
1713 | E = E->IgnoreParenCasts(); |
1714 | |
1715 | if (const BlockExpr *BE = dyn_cast<BlockExpr>(Val: E)) { |
1716 | const BlockDecl *Block = BE->getBlockDecl(); |
1717 | for (const auto &I : Block->captures()) { |
1718 | if (I.getVariable() == &Var) |
1719 | return true; |
1720 | } |
1721 | |
1722 | // No need to walk into the subexpressions. |
1723 | return false; |
1724 | } |
1725 | |
1726 | if (const StmtExpr *SE = dyn_cast<StmtExpr>(Val: E)) { |
1727 | const CompoundStmt *CS = SE->getSubStmt(); |
1728 | for (const auto *BI : CS->body()) |
1729 | if (const auto *BIE = dyn_cast<Expr>(Val: BI)) { |
1730 | if (isCapturedBy(Var, E: BIE)) |
1731 | return true; |
1732 | } |
1733 | else if (const auto *DS = dyn_cast<DeclStmt>(Val: BI)) { |
1734 | // special case declarations |
1735 | for (const auto *I : DS->decls()) { |
1736 | if (const auto *VD = dyn_cast<VarDecl>(Val: (I))) { |
1737 | const Expr *Init = VD->getInit(); |
1738 | if (Init && isCapturedBy(Var, E: Init)) |
1739 | return true; |
1740 | } |
1741 | } |
1742 | } |
1743 | else |
1744 | // FIXME. Make safe assumption assuming arbitrary statements cause capturing. |
1745 | // Later, provide code to poke into statements for capture analysis. |
1746 | return true; |
1747 | return false; |
1748 | } |
1749 | |
1750 | for (const Stmt *SubStmt : E->children()) |
1751 | if (isCapturedBy(Var, SubStmt)) |
1752 | return true; |
1753 | |
1754 | return false; |
1755 | } |
1756 | |
1757 | /// Determine whether the given initializer is trivial in the sense |
1758 | /// that it requires no code to be generated. |
1759 | bool CodeGenFunction::isTrivialInitializer(const Expr *Init) { |
1760 | if (!Init) |
1761 | return true; |
1762 | |
1763 | if (const CXXConstructExpr *Construct = dyn_cast<CXXConstructExpr>(Val: Init)) |
1764 | if (CXXConstructorDecl *Constructor = Construct->getConstructor()) |
1765 | if (Constructor->isTrivial() && |
1766 | Constructor->isDefaultConstructor() && |
1767 | !Construct->requiresZeroInitialization()) |
1768 | return true; |
1769 | |
1770 | return false; |
1771 | } |
1772 | |
1773 | void CodeGenFunction::emitZeroOrPatternForAutoVarInit(QualType type, |
1774 | const VarDecl &D, |
1775 | Address Loc) { |
1776 | auto trivialAutoVarInit = getContext().getLangOpts().getTrivialAutoVarInit(); |
1777 | auto trivialAutoVarInitMaxSize = |
1778 | getContext().getLangOpts().TrivialAutoVarInitMaxSize; |
1779 | CharUnits Size = getContext().getTypeSizeInChars(T: type); |
1780 | bool isVolatile = type.isVolatileQualified(); |
1781 | if (!Size.isZero()) { |
1782 | // We skip auto-init variables by their alloc size. Take this as an example: |
1783 | // "struct Foo {int x; char buff[1024];}" Assume the max-size flag is 1023. |
1784 | // All Foo type variables will be skipped. Ideally, we only skip the buff |
1785 | // array and still auto-init X in this example. |
1786 | // TODO: Improve the size filtering to by member size. |
1787 | auto allocSize = CGM.getDataLayout().getTypeAllocSize(Ty: Loc.getElementType()); |
1788 | switch (trivialAutoVarInit) { |
1789 | case LangOptions::TrivialAutoVarInitKind::Uninitialized: |
1790 | llvm_unreachable("Uninitialized handled by caller" ); |
1791 | case LangOptions::TrivialAutoVarInitKind::Zero: |
1792 | if (CGM.stopAutoInit()) |
1793 | return; |
1794 | if (trivialAutoVarInitMaxSize > 0 && |
1795 | allocSize > trivialAutoVarInitMaxSize) |
1796 | return; |
1797 | emitStoresForZeroInit(CGM, D, Loc, isVolatile, Builder); |
1798 | break; |
1799 | case LangOptions::TrivialAutoVarInitKind::Pattern: |
1800 | if (CGM.stopAutoInit()) |
1801 | return; |
1802 | if (trivialAutoVarInitMaxSize > 0 && |
1803 | allocSize > trivialAutoVarInitMaxSize) |
1804 | return; |
1805 | emitStoresForPatternInit(CGM, D, Loc, isVolatile, Builder); |
1806 | break; |
1807 | } |
1808 | return; |
1809 | } |
1810 | |
1811 | // VLAs look zero-sized to getTypeInfo. We can't emit constant stores to |
1812 | // them, so emit a memcpy with the VLA size to initialize each element. |
1813 | // Technically zero-sized or negative-sized VLAs are undefined, and UBSan |
1814 | // will catch that code, but there exists code which generates zero-sized |
1815 | // VLAs. Be nice and initialize whatever they requested. |
1816 | const auto *VlaType = getContext().getAsVariableArrayType(T: type); |
1817 | if (!VlaType) |
1818 | return; |
1819 | auto VlaSize = getVLASize(vla: VlaType); |
1820 | auto SizeVal = VlaSize.NumElts; |
1821 | CharUnits EltSize = getContext().getTypeSizeInChars(VlaSize.Type); |
1822 | switch (trivialAutoVarInit) { |
1823 | case LangOptions::TrivialAutoVarInitKind::Uninitialized: |
1824 | llvm_unreachable("Uninitialized handled by caller" ); |
1825 | |
1826 | case LangOptions::TrivialAutoVarInitKind::Zero: { |
1827 | if (CGM.stopAutoInit()) |
1828 | return; |
1829 | if (!EltSize.isOne()) |
1830 | SizeVal = Builder.CreateNUWMul(LHS: SizeVal, RHS: CGM.getSize(numChars: EltSize)); |
1831 | auto *I = Builder.CreateMemSet(Dest: Loc, Value: llvm::ConstantInt::get(Ty: Int8Ty, V: 0), |
1832 | Size: SizeVal, IsVolatile: isVolatile); |
1833 | I->addAnnotationMetadata(Annotation: "auto-init" ); |
1834 | break; |
1835 | } |
1836 | |
1837 | case LangOptions::TrivialAutoVarInitKind::Pattern: { |
1838 | if (CGM.stopAutoInit()) |
1839 | return; |
1840 | llvm::Type *ElTy = Loc.getElementType(); |
1841 | llvm::Constant *Constant = constWithPadding( |
1842 | CGM, isPattern: IsPattern::Yes, constant: initializationPatternFor(CGM, ElTy)); |
1843 | CharUnits ConstantAlign = getContext().getTypeAlignInChars(VlaSize.Type); |
1844 | llvm::BasicBlock *SetupBB = createBasicBlock(name: "vla-setup.loop" ); |
1845 | llvm::BasicBlock *LoopBB = createBasicBlock(name: "vla-init.loop" ); |
1846 | llvm::BasicBlock *ContBB = createBasicBlock(name: "vla-init.cont" ); |
1847 | llvm::Value *IsZeroSizedVLA = Builder.CreateICmpEQ( |
1848 | LHS: SizeVal, RHS: llvm::ConstantInt::get(Ty: SizeVal->getType(), V: 0), |
1849 | Name: "vla.iszerosized" ); |
1850 | Builder.CreateCondBr(Cond: IsZeroSizedVLA, True: ContBB, False: SetupBB); |
1851 | EmitBlock(BB: SetupBB); |
1852 | if (!EltSize.isOne()) |
1853 | SizeVal = Builder.CreateNUWMul(LHS: SizeVal, RHS: CGM.getSize(numChars: EltSize)); |
1854 | llvm::Value *BaseSizeInChars = |
1855 | llvm::ConstantInt::get(Ty: IntPtrTy, V: EltSize.getQuantity()); |
1856 | Address Begin = Loc.withElementType(ElemTy: Int8Ty); |
1857 | llvm::Value *End = Builder.CreateInBoundsGEP(Ty: Begin.getElementType(), |
1858 | Ptr: Begin.emitRawPointer(CGF&: *this), |
1859 | IdxList: SizeVal, Name: "vla.end" ); |
1860 | llvm::BasicBlock *OriginBB = Builder.GetInsertBlock(); |
1861 | EmitBlock(BB: LoopBB); |
1862 | llvm::PHINode *Cur = Builder.CreatePHI(Ty: Begin.getType(), NumReservedValues: 2, Name: "vla.cur" ); |
1863 | Cur->addIncoming(V: Begin.emitRawPointer(CGF&: *this), BB: OriginBB); |
1864 | CharUnits CurAlign = Loc.getAlignment().alignmentOfArrayElement(elementSize: EltSize); |
1865 | auto *I = |
1866 | Builder.CreateMemCpy(Dest: Address(Cur, Int8Ty, CurAlign), |
1867 | Src: createUnnamedGlobalForMemcpyFrom( |
1868 | CGM, D, Builder, Constant, Align: ConstantAlign), |
1869 | Size: BaseSizeInChars, IsVolatile: isVolatile); |
1870 | I->addAnnotationMetadata("auto-init" ); |
1871 | llvm::Value *Next = |
1872 | Builder.CreateInBoundsGEP(Ty: Int8Ty, Ptr: Cur, IdxList: BaseSizeInChars, Name: "vla.next" ); |
1873 | llvm::Value *Done = Builder.CreateICmpEQ(LHS: Next, RHS: End, Name: "vla-init.isdone" ); |
1874 | Builder.CreateCondBr(Cond: Done, True: ContBB, False: LoopBB); |
1875 | Cur->addIncoming(V: Next, BB: LoopBB); |
1876 | EmitBlock(BB: ContBB); |
1877 | } break; |
1878 | } |
1879 | } |
1880 | |
1881 | void CodeGenFunction::EmitAutoVarInit(const AutoVarEmission &emission) { |
1882 | assert(emission.Variable && "emission was not valid!" ); |
1883 | |
1884 | // If this was emitted as a global constant, we're done. |
1885 | if (emission.wasEmittedAsGlobal()) return; |
1886 | |
1887 | const VarDecl &D = *emission.Variable; |
1888 | auto DL = ApplyDebugLocation::CreateDefaultArtificial(CGF&: *this, TemporaryLocation: D.getLocation()); |
1889 | QualType type = D.getType(); |
1890 | |
1891 | // If this local has an initializer, emit it now. |
1892 | const Expr *Init = D.getInit(); |
1893 | |
1894 | // If we are at an unreachable point, we don't need to emit the initializer |
1895 | // unless it contains a label. |
1896 | if (!HaveInsertPoint()) { |
1897 | if (!Init || !ContainsLabel(Init)) return; |
1898 | EnsureInsertPoint(); |
1899 | } |
1900 | |
1901 | // Initialize the structure of a __block variable. |
1902 | if (emission.IsEscapingByRef) |
1903 | emitByrefStructureInit(emission); |
1904 | |
1905 | // Initialize the variable here if it doesn't have a initializer and it is a |
1906 | // C struct that is non-trivial to initialize or an array containing such a |
1907 | // struct. |
1908 | if (!Init && |
1909 | type.isNonTrivialToPrimitiveDefaultInitialize() == |
1910 | QualType::PDIK_Struct) { |
1911 | LValue Dst = MakeAddrLValue(Addr: emission.getAllocatedAddress(), T: type); |
1912 | if (emission.IsEscapingByRef) |
1913 | drillIntoBlockVariable(CGF&: *this, lvalue&: Dst, var: &D); |
1914 | defaultInitNonTrivialCStructVar(Dst); |
1915 | return; |
1916 | } |
1917 | |
1918 | // Check whether this is a byref variable that's potentially |
1919 | // captured and moved by its own initializer. If so, we'll need to |
1920 | // emit the initializer first, then copy into the variable. |
1921 | bool capturedByInit = |
1922 | Init && emission.IsEscapingByRef && isCapturedBy(Var: D, E: Init); |
1923 | |
1924 | bool = !capturedByInit; |
1925 | const Address Loc = |
1926 | locIsByrefHeader ? emission.getObjectAddress(CGF&: *this) : emission.Addr; |
1927 | |
1928 | // Note: constexpr already initializes everything correctly. |
1929 | LangOptions::TrivialAutoVarInitKind trivialAutoVarInit = |
1930 | (D.isConstexpr() |
1931 | ? LangOptions::TrivialAutoVarInitKind::Uninitialized |
1932 | : (D.getAttr<UninitializedAttr>() |
1933 | ? LangOptions::TrivialAutoVarInitKind::Uninitialized |
1934 | : getContext().getLangOpts().getTrivialAutoVarInit())); |
1935 | |
1936 | auto initializeWhatIsTechnicallyUninitialized = [&](Address Loc) { |
1937 | if (trivialAutoVarInit == |
1938 | LangOptions::TrivialAutoVarInitKind::Uninitialized) |
1939 | return; |
1940 | |
1941 | // Only initialize a __block's storage: we always initialize the header. |
1942 | if (emission.IsEscapingByRef && !locIsByrefHeader) |
1943 | Loc = emitBlockByrefAddress(baseAddr: Loc, V: &D, /*follow=*/followForward: false); |
1944 | |
1945 | return emitZeroOrPatternForAutoVarInit(type, D, Loc); |
1946 | }; |
1947 | |
1948 | if (isTrivialInitializer(Init)) |
1949 | return initializeWhatIsTechnicallyUninitialized(Loc); |
1950 | |
1951 | llvm::Constant *constant = nullptr; |
1952 | if (emission.IsConstantAggregate || |
1953 | D.mightBeUsableInConstantExpressions(C: getContext())) { |
1954 | assert(!capturedByInit && "constant init contains a capturing block?" ); |
1955 | constant = ConstantEmitter(*this).tryEmitAbstractForInitializer(D); |
1956 | if (constant && !constant->isZeroValue() && |
1957 | (trivialAutoVarInit != |
1958 | LangOptions::TrivialAutoVarInitKind::Uninitialized)) { |
1959 | IsPattern isPattern = |
1960 | (trivialAutoVarInit == LangOptions::TrivialAutoVarInitKind::Pattern) |
1961 | ? IsPattern::Yes |
1962 | : IsPattern::No; |
1963 | // C guarantees that brace-init with fewer initializers than members in |
1964 | // the aggregate will initialize the rest of the aggregate as-if it were |
1965 | // static initialization. In turn static initialization guarantees that |
1966 | // padding is initialized to zero bits. We could instead pattern-init if D |
1967 | // has any ImplicitValueInitExpr, but that seems to be unintuitive |
1968 | // behavior. |
1969 | constant = constWithPadding(CGM, isPattern: IsPattern::No, |
1970 | constant: replaceUndef(CGM, isPattern, constant)); |
1971 | } |
1972 | } |
1973 | |
1974 | if (!constant) { |
1975 | initializeWhatIsTechnicallyUninitialized(Loc); |
1976 | LValue lv = MakeAddrLValue(Addr: Loc, T: type); |
1977 | lv.setNonGC(true); |
1978 | return EmitExprAsInit(Init, &D, lv, capturedByInit); |
1979 | } |
1980 | |
1981 | if (!emission.IsConstantAggregate) { |
1982 | // For simple scalar/complex initialization, store the value directly. |
1983 | LValue lv = MakeAddrLValue(Addr: Loc, T: type); |
1984 | lv.setNonGC(true); |
1985 | return EmitStoreThroughLValue(Src: RValue::get(V: constant), Dst: lv, isInit: true); |
1986 | } |
1987 | |
1988 | emitStoresForConstant(CGM, D, Loc: Loc.withElementType(ElemTy: CGM.Int8Ty), |
1989 | isVolatile: type.isVolatileQualified(), Builder, constant, |
1990 | /*IsAutoInit=*/false); |
1991 | } |
1992 | |
1993 | /// Emit an expression as an initializer for an object (variable, field, etc.) |
1994 | /// at the given location. The expression is not necessarily the normal |
1995 | /// initializer for the object, and the address is not necessarily |
1996 | /// its normal location. |
1997 | /// |
1998 | /// \param init the initializing expression |
1999 | /// \param D the object to act as if we're initializing |
2000 | /// \param lvalue the lvalue to initialize |
2001 | /// \param capturedByInit true if \p D is a __block variable |
2002 | /// whose address is potentially changed by the initializer |
2003 | void CodeGenFunction::EmitExprAsInit(const Expr *init, const ValueDecl *D, |
2004 | LValue lvalue, bool capturedByInit) { |
2005 | QualType type = D->getType(); |
2006 | |
2007 | if (type->isReferenceType()) { |
2008 | RValue rvalue = EmitReferenceBindingToExpr(E: init); |
2009 | if (capturedByInit) |
2010 | drillIntoBlockVariable(CGF&: *this, lvalue, var: cast<VarDecl>(Val: D)); |
2011 | EmitStoreThroughLValue(Src: rvalue, Dst: lvalue, isInit: true); |
2012 | return; |
2013 | } |
2014 | switch (getEvaluationKind(T: type)) { |
2015 | case TEK_Scalar: |
2016 | EmitScalarInit(init, D, lvalue, capturedByInit); |
2017 | return; |
2018 | case TEK_Complex: { |
2019 | ComplexPairTy complex = EmitComplexExpr(E: init); |
2020 | if (capturedByInit) |
2021 | drillIntoBlockVariable(CGF&: *this, lvalue, var: cast<VarDecl>(Val: D)); |
2022 | EmitStoreOfComplex(V: complex, dest: lvalue, /*init*/ isInit: true); |
2023 | return; |
2024 | } |
2025 | case TEK_Aggregate: |
2026 | if (type->isAtomicType()) { |
2027 | EmitAtomicInit(E: const_cast<Expr*>(init), lvalue); |
2028 | } else { |
2029 | AggValueSlot::Overlap_t Overlap = AggValueSlot::MayOverlap; |
2030 | if (isa<VarDecl>(Val: D)) |
2031 | Overlap = AggValueSlot::DoesNotOverlap; |
2032 | else if (auto *FD = dyn_cast<FieldDecl>(Val: D)) |
2033 | Overlap = getOverlapForFieldInit(FD); |
2034 | // TODO: how can we delay here if D is captured by its initializer? |
2035 | EmitAggExpr(E: init, AS: AggValueSlot::forLValue( |
2036 | LV: lvalue, CGF&: *this, isDestructed: AggValueSlot::IsDestructed, |
2037 | needsGC: AggValueSlot::DoesNotNeedGCBarriers, |
2038 | isAliased: AggValueSlot::IsNotAliased, mayOverlap: Overlap)); |
2039 | } |
2040 | return; |
2041 | } |
2042 | llvm_unreachable("bad evaluation kind" ); |
2043 | } |
2044 | |
2045 | /// Enter a destroy cleanup for the given local variable. |
2046 | void CodeGenFunction::emitAutoVarTypeCleanup( |
2047 | const CodeGenFunction::AutoVarEmission &emission, |
2048 | QualType::DestructionKind dtorKind) { |
2049 | assert(dtorKind != QualType::DK_none); |
2050 | |
2051 | // Note that for __block variables, we want to destroy the |
2052 | // original stack object, not the possibly forwarded object. |
2053 | Address addr = emission.getObjectAddress(CGF&: *this); |
2054 | |
2055 | const VarDecl *var = emission.Variable; |
2056 | QualType type = var->getType(); |
2057 | |
2058 | CleanupKind cleanupKind = NormalAndEHCleanup; |
2059 | CodeGenFunction::Destroyer *destroyer = nullptr; |
2060 | |
2061 | switch (dtorKind) { |
2062 | case QualType::DK_none: |
2063 | llvm_unreachable("no cleanup for trivially-destructible variable" ); |
2064 | |
2065 | case QualType::DK_cxx_destructor: |
2066 | // If there's an NRVO flag on the emission, we need a different |
2067 | // cleanup. |
2068 | if (emission.NRVOFlag) { |
2069 | assert(!type->isArrayType()); |
2070 | CXXDestructorDecl *dtor = type->getAsCXXRecordDecl()->getDestructor(); |
2071 | EHStack.pushCleanup<DestroyNRVOVariableCXX>(Kind: cleanupKind, A: addr, A: type, A: dtor, |
2072 | A: emission.NRVOFlag); |
2073 | return; |
2074 | } |
2075 | break; |
2076 | |
2077 | case QualType::DK_objc_strong_lifetime: |
2078 | // Suppress cleanups for pseudo-strong variables. |
2079 | if (var->isARCPseudoStrong()) return; |
2080 | |
2081 | // Otherwise, consider whether to use an EH cleanup or not. |
2082 | cleanupKind = getARCCleanupKind(); |
2083 | |
2084 | // Use the imprecise destroyer by default. |
2085 | if (!var->hasAttr<ObjCPreciseLifetimeAttr>()) |
2086 | destroyer = CodeGenFunction::destroyARCStrongImprecise; |
2087 | break; |
2088 | |
2089 | case QualType::DK_objc_weak_lifetime: |
2090 | break; |
2091 | |
2092 | case QualType::DK_nontrivial_c_struct: |
2093 | destroyer = CodeGenFunction::destroyNonTrivialCStruct; |
2094 | if (emission.NRVOFlag) { |
2095 | assert(!type->isArrayType()); |
2096 | EHStack.pushCleanup<DestroyNRVOVariableC>(Kind: cleanupKind, A: addr, |
2097 | A: emission.NRVOFlag, A: type); |
2098 | return; |
2099 | } |
2100 | break; |
2101 | } |
2102 | |
2103 | // If we haven't chosen a more specific destroyer, use the default. |
2104 | if (!destroyer) destroyer = getDestroyer(destructionKind: dtorKind); |
2105 | |
2106 | // Use an EH cleanup in array destructors iff the destructor itself |
2107 | // is being pushed as an EH cleanup. |
2108 | bool useEHCleanup = (cleanupKind & EHCleanup); |
2109 | EHStack.pushCleanup<DestroyObject>(Kind: cleanupKind, A: addr, A: type, A: destroyer, |
2110 | A: useEHCleanup); |
2111 | } |
2112 | |
2113 | void CodeGenFunction::EmitAutoVarCleanups(const AutoVarEmission &emission) { |
2114 | assert(emission.Variable && "emission was not valid!" ); |
2115 | |
2116 | // If this was emitted as a global constant, we're done. |
2117 | if (emission.wasEmittedAsGlobal()) return; |
2118 | |
2119 | // If we don't have an insertion point, we're done. Sema prevents |
2120 | // us from jumping into any of these scopes anyway. |
2121 | if (!HaveInsertPoint()) return; |
2122 | |
2123 | const VarDecl &D = *emission.Variable; |
2124 | |
2125 | // Check the type for a cleanup. |
2126 | if (QualType::DestructionKind dtorKind = D.needsDestruction(Ctx: getContext())) |
2127 | emitAutoVarTypeCleanup(emission, dtorKind); |
2128 | |
2129 | // In GC mode, honor objc_precise_lifetime. |
2130 | if (getLangOpts().getGC() != LangOptions::NonGC && |
2131 | D.hasAttr<ObjCPreciseLifetimeAttr>()) { |
2132 | EHStack.pushCleanup<ExtendGCLifetime>(Kind: NormalCleanup, A: &D); |
2133 | } |
2134 | |
2135 | // Handle the cleanup attribute. |
2136 | if (const CleanupAttr *CA = D.getAttr<CleanupAttr>()) { |
2137 | const FunctionDecl *FD = CA->getFunctionDecl(); |
2138 | |
2139 | llvm::Constant *F = CGM.GetAddrOfFunction(GD: FD); |
2140 | assert(F && "Could not find function!" ); |
2141 | |
2142 | const CGFunctionInfo &Info = CGM.getTypes().arrangeFunctionDeclaration(FD); |
2143 | EHStack.pushCleanup<CallCleanupFunction>(Kind: NormalAndEHCleanup, A: F, A: &Info, A: &D); |
2144 | } |
2145 | |
2146 | // If this is a block variable, call _Block_object_destroy |
2147 | // (on the unforwarded address). Don't enter this cleanup if we're in pure-GC |
2148 | // mode. |
2149 | if (emission.IsEscapingByRef && |
2150 | CGM.getLangOpts().getGC() != LangOptions::GCOnly) { |
2151 | BlockFieldFlags Flags = BLOCK_FIELD_IS_BYREF; |
2152 | if (emission.Variable->getType().isObjCGCWeak()) |
2153 | Flags |= BLOCK_FIELD_IS_WEAK; |
2154 | enterByrefCleanup(Kind: NormalAndEHCleanup, Addr: emission.Addr, Flags, |
2155 | /*LoadBlockVarAddr*/ false, |
2156 | CanThrow: cxxDestructorCanThrow(T: emission.Variable->getType())); |
2157 | } |
2158 | } |
2159 | |
2160 | CodeGenFunction::Destroyer * |
2161 | CodeGenFunction::getDestroyer(QualType::DestructionKind kind) { |
2162 | switch (kind) { |
2163 | case QualType::DK_none: llvm_unreachable("no destroyer for trivial dtor" ); |
2164 | case QualType::DK_cxx_destructor: |
2165 | return destroyCXXObject; |
2166 | case QualType::DK_objc_strong_lifetime: |
2167 | return destroyARCStrongPrecise; |
2168 | case QualType::DK_objc_weak_lifetime: |
2169 | return destroyARCWeak; |
2170 | case QualType::DK_nontrivial_c_struct: |
2171 | return destroyNonTrivialCStruct; |
2172 | } |
2173 | llvm_unreachable("Unknown DestructionKind" ); |
2174 | } |
2175 | |
2176 | /// pushEHDestroy - Push the standard destructor for the given type as |
2177 | /// an EH-only cleanup. |
2178 | void CodeGenFunction::pushEHDestroy(QualType::DestructionKind dtorKind, |
2179 | Address addr, QualType type) { |
2180 | assert(dtorKind && "cannot push destructor for trivial type" ); |
2181 | assert(needsEHCleanup(dtorKind)); |
2182 | |
2183 | pushDestroy(kind: EHCleanup, addr, type, destroyer: getDestroyer(kind: dtorKind), useEHCleanupForArray: true); |
2184 | } |
2185 | |
2186 | /// pushDestroy - Push the standard destructor for the given type as |
2187 | /// at least a normal cleanup. |
2188 | void CodeGenFunction::pushDestroy(QualType::DestructionKind dtorKind, |
2189 | Address addr, QualType type) { |
2190 | assert(dtorKind && "cannot push destructor for trivial type" ); |
2191 | |
2192 | CleanupKind cleanupKind = getCleanupKind(kind: dtorKind); |
2193 | pushDestroy(kind: cleanupKind, addr, type, destroyer: getDestroyer(kind: dtorKind), |
2194 | useEHCleanupForArray: cleanupKind & EHCleanup); |
2195 | } |
2196 | |
2197 | void CodeGenFunction::pushDestroy(CleanupKind cleanupKind, Address addr, |
2198 | QualType type, Destroyer *destroyer, |
2199 | bool useEHCleanupForArray) { |
2200 | pushFullExprCleanup<DestroyObject>(kind: cleanupKind, A: addr, A: type, |
2201 | A: destroyer, A: useEHCleanupForArray); |
2202 | } |
2203 | |
2204 | void CodeGenFunction::pushStackRestore(CleanupKind Kind, Address SPMem) { |
2205 | EHStack.pushCleanup<CallStackRestore>(Kind, A: SPMem); |
2206 | } |
2207 | |
2208 | void CodeGenFunction::pushKmpcAllocFree( |
2209 | CleanupKind Kind, std::pair<llvm::Value *, llvm::Value *> AddrSizePair) { |
2210 | EHStack.pushCleanup<KmpcAllocFree>(Kind, A: AddrSizePair); |
2211 | } |
2212 | |
2213 | void CodeGenFunction::pushLifetimeExtendedDestroy(CleanupKind cleanupKind, |
2214 | Address addr, QualType type, |
2215 | Destroyer *destroyer, |
2216 | bool useEHCleanupForArray) { |
2217 | // If we're not in a conditional branch, we don't need to bother generating a |
2218 | // conditional cleanup. |
2219 | if (!isInConditionalBranch()) { |
2220 | // Push an EH-only cleanup for the object now. |
2221 | // FIXME: When popping normal cleanups, we need to keep this EH cleanup |
2222 | // around in case a temporary's destructor throws an exception. |
2223 | if (cleanupKind & EHCleanup) |
2224 | EHStack.pushCleanup<DestroyObject>( |
2225 | Kind: static_cast<CleanupKind>(cleanupKind & ~NormalCleanup), A: addr, A: type, |
2226 | A: destroyer, A: useEHCleanupForArray); |
2227 | |
2228 | return pushCleanupAfterFullExprWithActiveFlag<DestroyObject>( |
2229 | Kind: cleanupKind, ActiveFlag: Address::invalid(), A: addr, A: type, A: destroyer, A: useEHCleanupForArray); |
2230 | } |
2231 | |
2232 | // Otherwise, we should only destroy the object if it's been initialized. |
2233 | // Re-use the active flag and saved address across both the EH and end of |
2234 | // scope cleanups. |
2235 | |
2236 | using SavedType = typename DominatingValue<Address>::saved_type; |
2237 | using ConditionalCleanupType = |
2238 | EHScopeStack::ConditionalCleanup<DestroyObject, Address, QualType, |
2239 | Destroyer *, bool>; |
2240 | |
2241 | Address ActiveFlag = createCleanupActiveFlag(); |
2242 | SavedType SavedAddr = saveValueInCond(value: addr); |
2243 | |
2244 | if (cleanupKind & EHCleanup) { |
2245 | EHStack.pushCleanup<ConditionalCleanupType>( |
2246 | Kind: static_cast<CleanupKind>(cleanupKind & ~NormalCleanup), A: SavedAddr, A: type, |
2247 | A: destroyer, A: useEHCleanupForArray); |
2248 | initFullExprCleanupWithFlag(ActiveFlag); |
2249 | } |
2250 | |
2251 | pushCleanupAfterFullExprWithActiveFlag<ConditionalCleanupType>( |
2252 | Kind: cleanupKind, ActiveFlag, A: SavedAddr, A: type, A: destroyer, |
2253 | A: useEHCleanupForArray); |
2254 | } |
2255 | |
2256 | /// emitDestroy - Immediately perform the destruction of the given |
2257 | /// object. |
2258 | /// |
2259 | /// \param addr - the address of the object; a type* |
2260 | /// \param type - the type of the object; if an array type, all |
2261 | /// objects are destroyed in reverse order |
2262 | /// \param destroyer - the function to call to destroy individual |
2263 | /// elements |
2264 | /// \param useEHCleanupForArray - whether an EH cleanup should be |
2265 | /// used when destroying array elements, in case one of the |
2266 | /// destructions throws an exception |
2267 | void CodeGenFunction::emitDestroy(Address addr, QualType type, |
2268 | Destroyer *destroyer, |
2269 | bool useEHCleanupForArray) { |
2270 | const ArrayType *arrayType = getContext().getAsArrayType(T: type); |
2271 | if (!arrayType) |
2272 | return destroyer(*this, addr, type); |
2273 | |
2274 | llvm::Value *length = emitArrayLength(arrayType, baseType&: type, addr); |
2275 | |
2276 | CharUnits elementAlign = |
2277 | addr.getAlignment() |
2278 | .alignmentOfArrayElement(elementSize: getContext().getTypeSizeInChars(T: type)); |
2279 | |
2280 | // Normally we have to check whether the array is zero-length. |
2281 | bool checkZeroLength = true; |
2282 | |
2283 | // But if the array length is constant, we can suppress that. |
2284 | if (llvm::ConstantInt *constLength = dyn_cast<llvm::ConstantInt>(Val: length)) { |
2285 | // ...and if it's constant zero, we can just skip the entire thing. |
2286 | if (constLength->isZero()) return; |
2287 | checkZeroLength = false; |
2288 | } |
2289 | |
2290 | llvm::Value *begin = addr.emitRawPointer(CGF&: *this); |
2291 | llvm::Value *end = |
2292 | Builder.CreateInBoundsGEP(Ty: addr.getElementType(), Ptr: begin, IdxList: length); |
2293 | emitArrayDestroy(begin, end, elementType: type, elementAlign, destroyer, |
2294 | checkZeroLength, useEHCleanup: useEHCleanupForArray); |
2295 | } |
2296 | |
2297 | /// emitArrayDestroy - Destroys all the elements of the given array, |
2298 | /// beginning from last to first. The array cannot be zero-length. |
2299 | /// |
2300 | /// \param begin - a type* denoting the first element of the array |
2301 | /// \param end - a type* denoting one past the end of the array |
2302 | /// \param elementType - the element type of the array |
2303 | /// \param destroyer - the function to call to destroy elements |
2304 | /// \param useEHCleanup - whether to push an EH cleanup to destroy |
2305 | /// the remaining elements in case the destruction of a single |
2306 | /// element throws |
2307 | void CodeGenFunction::emitArrayDestroy(llvm::Value *begin, |
2308 | llvm::Value *end, |
2309 | QualType elementType, |
2310 | CharUnits elementAlign, |
2311 | Destroyer *destroyer, |
2312 | bool checkZeroLength, |
2313 | bool useEHCleanup) { |
2314 | assert(!elementType->isArrayType()); |
2315 | |
2316 | // The basic structure here is a do-while loop, because we don't |
2317 | // need to check for the zero-element case. |
2318 | llvm::BasicBlock *bodyBB = createBasicBlock(name: "arraydestroy.body" ); |
2319 | llvm::BasicBlock *doneBB = createBasicBlock(name: "arraydestroy.done" ); |
2320 | |
2321 | if (checkZeroLength) { |
2322 | llvm::Value *isEmpty = Builder.CreateICmpEQ(LHS: begin, RHS: end, |
2323 | Name: "arraydestroy.isempty" ); |
2324 | Builder.CreateCondBr(Cond: isEmpty, True: doneBB, False: bodyBB); |
2325 | } |
2326 | |
2327 | // Enter the loop body, making that address the current address. |
2328 | llvm::BasicBlock *entryBB = Builder.GetInsertBlock(); |
2329 | EmitBlock(BB: bodyBB); |
2330 | llvm::PHINode *elementPast = |
2331 | Builder.CreatePHI(Ty: begin->getType(), NumReservedValues: 2, Name: "arraydestroy.elementPast" ); |
2332 | elementPast->addIncoming(V: end, BB: entryBB); |
2333 | |
2334 | // Shift the address back by one element. |
2335 | llvm::Value *negativeOne = llvm::ConstantInt::get(Ty: SizeTy, V: -1, IsSigned: true); |
2336 | llvm::Type *llvmElementType = ConvertTypeForMem(T: elementType); |
2337 | llvm::Value *element = Builder.CreateInBoundsGEP( |
2338 | Ty: llvmElementType, Ptr: elementPast, IdxList: negativeOne, Name: "arraydestroy.element" ); |
2339 | |
2340 | if (useEHCleanup) |
2341 | pushRegularPartialArrayCleanup(arrayBegin: begin, arrayEnd: element, elementType, elementAlignment: elementAlign, |
2342 | destroyer); |
2343 | |
2344 | // Perform the actual destruction there. |
2345 | destroyer(*this, Address(element, llvmElementType, elementAlign), |
2346 | elementType); |
2347 | |
2348 | if (useEHCleanup) |
2349 | PopCleanupBlock(); |
2350 | |
2351 | // Check whether we've reached the end. |
2352 | llvm::Value *done = Builder.CreateICmpEQ(LHS: element, RHS: begin, Name: "arraydestroy.done" ); |
2353 | Builder.CreateCondBr(Cond: done, True: doneBB, False: bodyBB); |
2354 | elementPast->addIncoming(V: element, BB: Builder.GetInsertBlock()); |
2355 | |
2356 | // Done. |
2357 | EmitBlock(BB: doneBB); |
2358 | } |
2359 | |
2360 | /// Perform partial array destruction as if in an EH cleanup. Unlike |
2361 | /// emitArrayDestroy, the element type here may still be an array type. |
2362 | static void emitPartialArrayDestroy(CodeGenFunction &CGF, |
2363 | llvm::Value *begin, llvm::Value *end, |
2364 | QualType type, CharUnits elementAlign, |
2365 | CodeGenFunction::Destroyer *destroyer) { |
2366 | llvm::Type *elemTy = CGF.ConvertTypeForMem(T: type); |
2367 | |
2368 | // If the element type is itself an array, drill down. |
2369 | unsigned arrayDepth = 0; |
2370 | while (const ArrayType *arrayType = CGF.getContext().getAsArrayType(T: type)) { |
2371 | // VLAs don't require a GEP index to walk into. |
2372 | if (!isa<VariableArrayType>(Val: arrayType)) |
2373 | arrayDepth++; |
2374 | type = arrayType->getElementType(); |
2375 | } |
2376 | |
2377 | if (arrayDepth) { |
2378 | llvm::Value *zero = llvm::ConstantInt::get(Ty: CGF.SizeTy, V: 0); |
2379 | |
2380 | SmallVector<llvm::Value*,4> gepIndices(arrayDepth+1, zero); |
2381 | begin = CGF.Builder.CreateInBoundsGEP( |
2382 | Ty: elemTy, Ptr: begin, IdxList: gepIndices, Name: "pad.arraybegin" ); |
2383 | end = CGF.Builder.CreateInBoundsGEP( |
2384 | Ty: elemTy, Ptr: end, IdxList: gepIndices, Name: "pad.arrayend" ); |
2385 | } |
2386 | |
2387 | // Destroy the array. We don't ever need an EH cleanup because we |
2388 | // assume that we're in an EH cleanup ourselves, so a throwing |
2389 | // destructor causes an immediate terminate. |
2390 | CGF.emitArrayDestroy(begin, end, elementType: type, elementAlign, destroyer, |
2391 | /*checkZeroLength*/ true, /*useEHCleanup*/ false); |
2392 | } |
2393 | |
2394 | namespace { |
2395 | /// RegularPartialArrayDestroy - a cleanup which performs a partial |
2396 | /// array destroy where the end pointer is regularly determined and |
2397 | /// does not need to be loaded from a local. |
2398 | class RegularPartialArrayDestroy final : public EHScopeStack::Cleanup { |
2399 | llvm::Value *ArrayBegin; |
2400 | llvm::Value *ArrayEnd; |
2401 | QualType ElementType; |
2402 | CodeGenFunction::Destroyer *Destroyer; |
2403 | CharUnits ElementAlign; |
2404 | public: |
2405 | RegularPartialArrayDestroy(llvm::Value *arrayBegin, llvm::Value *arrayEnd, |
2406 | QualType elementType, CharUnits elementAlign, |
2407 | CodeGenFunction::Destroyer *destroyer) |
2408 | : ArrayBegin(arrayBegin), ArrayEnd(arrayEnd), |
2409 | ElementType(elementType), Destroyer(destroyer), |
2410 | ElementAlign(elementAlign) {} |
2411 | |
2412 | void Emit(CodeGenFunction &CGF, Flags flags) override { |
2413 | emitPartialArrayDestroy(CGF, ArrayBegin, ArrayEnd, |
2414 | ElementType, ElementAlign, Destroyer); |
2415 | } |
2416 | }; |
2417 | |
2418 | /// IrregularPartialArrayDestroy - a cleanup which performs a |
2419 | /// partial array destroy where the end pointer is irregularly |
2420 | /// determined and must be loaded from a local. |
2421 | class IrregularPartialArrayDestroy final : public EHScopeStack::Cleanup { |
2422 | llvm::Value *ArrayBegin; |
2423 | Address ArrayEndPointer; |
2424 | QualType ElementType; |
2425 | CodeGenFunction::Destroyer *Destroyer; |
2426 | CharUnits ElementAlign; |
2427 | public: |
2428 | IrregularPartialArrayDestroy(llvm::Value *arrayBegin, |
2429 | Address arrayEndPointer, |
2430 | QualType elementType, |
2431 | CharUnits elementAlign, |
2432 | CodeGenFunction::Destroyer *destroyer) |
2433 | : ArrayBegin(arrayBegin), ArrayEndPointer(arrayEndPointer), |
2434 | ElementType(elementType), Destroyer(destroyer), |
2435 | ElementAlign(elementAlign) {} |
2436 | |
2437 | void Emit(CodeGenFunction &CGF, Flags flags) override { |
2438 | llvm::Value *arrayEnd = CGF.Builder.CreateLoad(Addr: ArrayEndPointer); |
2439 | emitPartialArrayDestroy(CGF, ArrayBegin, arrayEnd, |
2440 | ElementType, ElementAlign, Destroyer); |
2441 | } |
2442 | }; |
2443 | } // end anonymous namespace |
2444 | |
2445 | /// pushIrregularPartialArrayCleanup - Push an EH cleanup to destroy |
2446 | /// already-constructed elements of the given array. The cleanup |
2447 | /// may be popped with DeactivateCleanupBlock or PopCleanupBlock. |
2448 | /// |
2449 | /// \param elementType - the immediate element type of the array; |
2450 | /// possibly still an array type |
2451 | void CodeGenFunction::pushIrregularPartialArrayCleanup(llvm::Value *arrayBegin, |
2452 | Address arrayEndPointer, |
2453 | QualType elementType, |
2454 | CharUnits elementAlign, |
2455 | Destroyer *destroyer) { |
2456 | pushFullExprCleanup<IrregularPartialArrayDestroy>(kind: EHCleanup, |
2457 | A: arrayBegin, A: arrayEndPointer, |
2458 | A: elementType, A: elementAlign, |
2459 | A: destroyer); |
2460 | } |
2461 | |
2462 | /// pushRegularPartialArrayCleanup - Push an EH cleanup to destroy |
2463 | /// already-constructed elements of the given array. The cleanup |
2464 | /// may be popped with DeactivateCleanupBlock or PopCleanupBlock. |
2465 | /// |
2466 | /// \param elementType - the immediate element type of the array; |
2467 | /// possibly still an array type |
2468 | void CodeGenFunction::pushRegularPartialArrayCleanup(llvm::Value *arrayBegin, |
2469 | llvm::Value *arrayEnd, |
2470 | QualType elementType, |
2471 | CharUnits elementAlign, |
2472 | Destroyer *destroyer) { |
2473 | pushFullExprCleanup<RegularPartialArrayDestroy>(kind: EHCleanup, |
2474 | A: arrayBegin, A: arrayEnd, |
2475 | A: elementType, A: elementAlign, |
2476 | A: destroyer); |
2477 | } |
2478 | |
2479 | /// Lazily declare the @llvm.lifetime.start intrinsic. |
2480 | llvm::Function *CodeGenModule::getLLVMLifetimeStartFn() { |
2481 | if (LifetimeStartFn) |
2482 | return LifetimeStartFn; |
2483 | LifetimeStartFn = llvm::Intrinsic::getDeclaration(&getModule(), |
2484 | llvm::Intrinsic::lifetime_start, AllocaInt8PtrTy); |
2485 | return LifetimeStartFn; |
2486 | } |
2487 | |
2488 | /// Lazily declare the @llvm.lifetime.end intrinsic. |
2489 | llvm::Function *CodeGenModule::getLLVMLifetimeEndFn() { |
2490 | if (LifetimeEndFn) |
2491 | return LifetimeEndFn; |
2492 | LifetimeEndFn = llvm::Intrinsic::getDeclaration(&getModule(), |
2493 | llvm::Intrinsic::lifetime_end, AllocaInt8PtrTy); |
2494 | return LifetimeEndFn; |
2495 | } |
2496 | |
2497 | namespace { |
2498 | /// A cleanup to perform a release of an object at the end of a |
2499 | /// function. This is used to balance out the incoming +1 of a |
2500 | /// ns_consumed argument when we can't reasonably do that just by |
2501 | /// not doing the initial retain for a __block argument. |
2502 | struct ConsumeARCParameter final : EHScopeStack::Cleanup { |
2503 | ConsumeARCParameter(llvm::Value *param, |
2504 | ARCPreciseLifetime_t precise) |
2505 | : Param(param), Precise(precise) {} |
2506 | |
2507 | llvm::Value *Param; |
2508 | ARCPreciseLifetime_t Precise; |
2509 | |
2510 | void Emit(CodeGenFunction &CGF, Flags flags) override { |
2511 | CGF.EmitARCRelease(value: Param, precise: Precise); |
2512 | } |
2513 | }; |
2514 | } // end anonymous namespace |
2515 | |
2516 | /// Emit an alloca (or GlobalValue depending on target) |
2517 | /// for the specified parameter and set up LocalDeclMap. |
2518 | void CodeGenFunction::EmitParmDecl(const VarDecl &D, ParamValue Arg, |
2519 | unsigned ArgNo) { |
2520 | bool NoDebugInfo = false; |
2521 | // FIXME: Why isn't ImplicitParamDecl a ParmVarDecl? |
2522 | assert((isa<ParmVarDecl>(D) || isa<ImplicitParamDecl>(D)) && |
2523 | "Invalid argument to EmitParmDecl" ); |
2524 | |
2525 | // Set the name of the parameter's initial value to make IR easier to |
2526 | // read. Don't modify the names of globals. |
2527 | if (!isa<llvm::GlobalValue>(Val: Arg.getAnyValue())) |
2528 | Arg.getAnyValue()->setName(D.getName()); |
2529 | |
2530 | QualType Ty = D.getType(); |
2531 | |
2532 | // Use better IR generation for certain implicit parameters. |
2533 | if (auto IPD = dyn_cast<ImplicitParamDecl>(Val: &D)) { |
2534 | // The only implicit argument a block has is its literal. |
2535 | // This may be passed as an inalloca'ed value on Windows x86. |
2536 | if (BlockInfo) { |
2537 | llvm::Value *V = Arg.isIndirect() |
2538 | ? Builder.CreateLoad(Addr: Arg.getIndirectAddress()) |
2539 | : Arg.getDirectValue(); |
2540 | setBlockContextParameter(D: IPD, argNum: ArgNo, ptr: V); |
2541 | return; |
2542 | } |
2543 | // Suppressing debug info for ThreadPrivateVar parameters, else it hides |
2544 | // debug info of TLS variables. |
2545 | NoDebugInfo = |
2546 | (IPD->getParameterKind() == ImplicitParamKind::ThreadPrivateVar); |
2547 | } |
2548 | |
2549 | Address DeclPtr = Address::invalid(); |
2550 | RawAddress AllocaPtr = Address::invalid(); |
2551 | bool DoStore = false; |
2552 | bool IsScalar = hasScalarEvaluationKind(T: Ty); |
2553 | bool UseIndirectDebugAddress = false; |
2554 | |
2555 | // If we already have a pointer to the argument, reuse the input pointer. |
2556 | if (Arg.isIndirect()) { |
2557 | DeclPtr = Arg.getIndirectAddress(); |
2558 | DeclPtr = DeclPtr.withElementType(ElemTy: ConvertTypeForMem(T: Ty)); |
2559 | // Indirect argument is in alloca address space, which may be different |
2560 | // from the default address space. |
2561 | auto AllocaAS = CGM.getASTAllocaAddressSpace(); |
2562 | auto *V = DeclPtr.emitRawPointer(CGF&: *this); |
2563 | AllocaPtr = RawAddress(V, DeclPtr.getElementType(), DeclPtr.getAlignment()); |
2564 | |
2565 | // For truly ABI indirect arguments -- those that are not `byval` -- store |
2566 | // the address of the argument on the stack to preserve debug information. |
2567 | ABIArgInfo ArgInfo = CurFnInfo->arguments()[ArgNo - 1].info; |
2568 | if (ArgInfo.isIndirect()) |
2569 | UseIndirectDebugAddress = !ArgInfo.getIndirectByVal(); |
2570 | if (UseIndirectDebugAddress) { |
2571 | auto PtrTy = getContext().getPointerType(T: Ty); |
2572 | AllocaPtr = CreateMemTemp(PtrTy, getContext().getTypeAlignInChars(PtrTy), |
2573 | D.getName() + ".indirect_addr" ); |
2574 | EmitStoreOfScalar(V, AllocaPtr, /* Volatile */ false, PtrTy); |
2575 | } |
2576 | |
2577 | auto SrcLangAS = getLangOpts().OpenCL ? LangAS::opencl_private : AllocaAS; |
2578 | auto DestLangAS = |
2579 | getLangOpts().OpenCL ? LangAS::opencl_private : LangAS::Default; |
2580 | if (SrcLangAS != DestLangAS) { |
2581 | assert(getContext().getTargetAddressSpace(SrcLangAS) == |
2582 | CGM.getDataLayout().getAllocaAddrSpace()); |
2583 | auto DestAS = getContext().getTargetAddressSpace(AS: DestLangAS); |
2584 | auto *T = llvm::PointerType::get(C&: getLLVMContext(), AddressSpace: DestAS); |
2585 | DeclPtr = |
2586 | DeclPtr.withPointer(NewPointer: getTargetHooks().performAddrSpaceCast( |
2587 | *this, V, SrcLangAS, DestLangAS, T, true), |
2588 | IsKnownNonNull: DeclPtr.isKnownNonNull()); |
2589 | } |
2590 | |
2591 | // Push a destructor cleanup for this parameter if the ABI requires it. |
2592 | // Don't push a cleanup in a thunk for a method that will also emit a |
2593 | // cleanup. |
2594 | if (Ty->isRecordType() && !CurFuncIsThunk && |
2595 | Ty->castAs<RecordType>()->getDecl()->isParamDestroyedInCallee()) { |
2596 | if (QualType::DestructionKind DtorKind = |
2597 | D.needsDestruction(Ctx: getContext())) { |
2598 | assert((DtorKind == QualType::DK_cxx_destructor || |
2599 | DtorKind == QualType::DK_nontrivial_c_struct) && |
2600 | "unexpected destructor type" ); |
2601 | pushDestroy(dtorKind: DtorKind, addr: DeclPtr, type: Ty); |
2602 | CalleeDestructedParamCleanups[cast<ParmVarDecl>(Val: &D)] = |
2603 | EHStack.stable_begin(); |
2604 | } |
2605 | } |
2606 | } else { |
2607 | // Check if the parameter address is controlled by OpenMP runtime. |
2608 | Address OpenMPLocalAddr = |
2609 | getLangOpts().OpenMP |
2610 | ? CGM.getOpenMPRuntime().getAddressOfLocalVariable(CGF&: *this, VD: &D) |
2611 | : Address::invalid(); |
2612 | if (getLangOpts().OpenMP && OpenMPLocalAddr.isValid()) { |
2613 | DeclPtr = OpenMPLocalAddr; |
2614 | AllocaPtr = DeclPtr; |
2615 | } else { |
2616 | // Otherwise, create a temporary to hold the value. |
2617 | DeclPtr = CreateMemTemp(Ty, getContext().getDeclAlign(&D), |
2618 | D.getName() + ".addr" , &AllocaPtr); |
2619 | } |
2620 | DoStore = true; |
2621 | } |
2622 | |
2623 | llvm::Value *ArgVal = (DoStore ? Arg.getDirectValue() : nullptr); |
2624 | |
2625 | LValue lv = MakeAddrLValue(Addr: DeclPtr, T: Ty); |
2626 | if (IsScalar) { |
2627 | Qualifiers qs = Ty.getQualifiers(); |
2628 | if (Qualifiers::ObjCLifetime lt = qs.getObjCLifetime()) { |
2629 | // We honor __attribute__((ns_consumed)) for types with lifetime. |
2630 | // For __strong, it's handled by just skipping the initial retain; |
2631 | // otherwise we have to balance out the initial +1 with an extra |
2632 | // cleanup to do the release at the end of the function. |
2633 | bool isConsumed = D.hasAttr<NSConsumedAttr>(); |
2634 | |
2635 | // If a parameter is pseudo-strong then we can omit the implicit retain. |
2636 | if (D.isARCPseudoStrong()) { |
2637 | assert(lt == Qualifiers::OCL_Strong && |
2638 | "pseudo-strong variable isn't strong?" ); |
2639 | assert(qs.hasConst() && "pseudo-strong variable should be const!" ); |
2640 | lt = Qualifiers::OCL_ExplicitNone; |
2641 | } |
2642 | |
2643 | // Load objects passed indirectly. |
2644 | if (Arg.isIndirect() && !ArgVal) |
2645 | ArgVal = Builder.CreateLoad(Addr: DeclPtr); |
2646 | |
2647 | if (lt == Qualifiers::OCL_Strong) { |
2648 | if (!isConsumed) { |
2649 | if (CGM.getCodeGenOpts().OptimizationLevel == 0) { |
2650 | // use objc_storeStrong(&dest, value) for retaining the |
2651 | // object. But first, store a null into 'dest' because |
2652 | // objc_storeStrong attempts to release its old value. |
2653 | llvm::Value *Null = CGM.EmitNullConstant(T: D.getType()); |
2654 | EmitStoreOfScalar(value: Null, lvalue: lv, /* isInitialization */ isInit: true); |
2655 | EmitARCStoreStrongCall(addr: lv.getAddress(CGF&: *this), value: ArgVal, resultIgnored: true); |
2656 | DoStore = false; |
2657 | } |
2658 | else |
2659 | // Don't use objc_retainBlock for block pointers, because we |
2660 | // don't want to Block_copy something just because we got it |
2661 | // as a parameter. |
2662 | ArgVal = EmitARCRetainNonBlock(value: ArgVal); |
2663 | } |
2664 | } else { |
2665 | // Push the cleanup for a consumed parameter. |
2666 | if (isConsumed) { |
2667 | ARCPreciseLifetime_t precise = (D.hasAttr<ObjCPreciseLifetimeAttr>() |
2668 | ? ARCPreciseLifetime : ARCImpreciseLifetime); |
2669 | EHStack.pushCleanup<ConsumeARCParameter>(Kind: getARCCleanupKind(), A: ArgVal, |
2670 | A: precise); |
2671 | } |
2672 | |
2673 | if (lt == Qualifiers::OCL_Weak) { |
2674 | EmitARCInitWeak(addr: DeclPtr, value: ArgVal); |
2675 | DoStore = false; // The weak init is a store, no need to do two. |
2676 | } |
2677 | } |
2678 | |
2679 | // Enter the cleanup scope. |
2680 | EmitAutoVarWithLifetime(CGF&: *this, var: D, addr: DeclPtr, lifetime: lt); |
2681 | } |
2682 | } |
2683 | |
2684 | // Store the initial value into the alloca. |
2685 | if (DoStore) |
2686 | EmitStoreOfScalar(value: ArgVal, lvalue: lv, /* isInitialization */ isInit: true); |
2687 | |
2688 | setAddrOfLocalVar(VD: &D, Addr: DeclPtr); |
2689 | |
2690 | // Emit debug info for param declarations in non-thunk functions. |
2691 | if (CGDebugInfo *DI = getDebugInfo()) { |
2692 | if (CGM.getCodeGenOpts().hasReducedDebugInfo() && !CurFuncIsThunk && |
2693 | !NoDebugInfo) { |
2694 | llvm::DILocalVariable *DILocalVar = DI->EmitDeclareOfArgVariable( |
2695 | Decl: &D, AI: AllocaPtr.getPointer(), ArgNo, Builder, UsePointerValue: UseIndirectDebugAddress); |
2696 | if (const auto *Var = dyn_cast_or_null<ParmVarDecl>(Val: &D)) |
2697 | DI->getParamDbgMappings().insert(KV: {Var, DILocalVar}); |
2698 | } |
2699 | } |
2700 | |
2701 | if (D.hasAttr<AnnotateAttr>()) |
2702 | EmitVarAnnotations(D: &D, V: DeclPtr.emitRawPointer(CGF&: *this)); |
2703 | |
2704 | // We can only check return value nullability if all arguments to the |
2705 | // function satisfy their nullability preconditions. This makes it necessary |
2706 | // to emit null checks for args in the function body itself. |
2707 | if (requiresReturnValueNullabilityCheck()) { |
2708 | auto Nullability = Ty->getNullability(); |
2709 | if (Nullability && *Nullability == NullabilityKind::NonNull) { |
2710 | SanitizerScope SanScope(this); |
2711 | RetValNullabilityPrecondition = |
2712 | Builder.CreateAnd(LHS: RetValNullabilityPrecondition, |
2713 | RHS: Builder.CreateIsNotNull(Arg: Arg.getAnyValue())); |
2714 | } |
2715 | } |
2716 | } |
2717 | |
2718 | void CodeGenModule::EmitOMPDeclareReduction(const OMPDeclareReductionDecl *D, |
2719 | CodeGenFunction *CGF) { |
2720 | if (!LangOpts.OpenMP || (!LangOpts.EmitAllDecls && !D->isUsed())) |
2721 | return; |
2722 | getOpenMPRuntime().emitUserDefinedReduction(CGF, D); |
2723 | } |
2724 | |
2725 | void CodeGenModule::EmitOMPDeclareMapper(const OMPDeclareMapperDecl *D, |
2726 | CodeGenFunction *CGF) { |
2727 | if (!LangOpts.OpenMP || LangOpts.OpenMPSimd || |
2728 | (!LangOpts.EmitAllDecls && !D->isUsed())) |
2729 | return; |
2730 | getOpenMPRuntime().emitUserDefinedMapper(D, CGF); |
2731 | } |
2732 | |
2733 | void CodeGenModule::EmitOMPRequiresDecl(const OMPRequiresDecl *D) { |
2734 | getOpenMPRuntime().processRequiresDirective(D); |
2735 | } |
2736 | |
2737 | void CodeGenModule::EmitOMPAllocateDecl(const OMPAllocateDecl *D) { |
2738 | for (const Expr *E : D->varlists()) { |
2739 | const auto *DE = cast<DeclRefExpr>(Val: E); |
2740 | const auto *VD = cast<VarDecl>(Val: DE->getDecl()); |
2741 | |
2742 | // Skip all but globals. |
2743 | if (!VD->hasGlobalStorage()) |
2744 | continue; |
2745 | |
2746 | // Check if the global has been materialized yet or not. If not, we are done |
2747 | // as any later generation will utilize the OMPAllocateDeclAttr. However, if |
2748 | // we already emitted the global we might have done so before the |
2749 | // OMPAllocateDeclAttr was attached, leading to the wrong address space |
2750 | // (potentially). While not pretty, common practise is to remove the old IR |
2751 | // global and generate a new one, so we do that here too. Uses are replaced |
2752 | // properly. |
2753 | StringRef MangledName = getMangledName(GD: VD); |
2754 | llvm::GlobalValue *Entry = GetGlobalValue(Ref: MangledName); |
2755 | if (!Entry) |
2756 | continue; |
2757 | |
2758 | // We can also keep the existing global if the address space is what we |
2759 | // expect it to be, if not, it is replaced. |
2760 | QualType ASTTy = VD->getType(); |
2761 | clang::LangAS GVAS = GetGlobalVarAddressSpace(D: VD); |
2762 | auto TargetAS = getContext().getTargetAddressSpace(AS: GVAS); |
2763 | if (Entry->getType()->getAddressSpace() == TargetAS) |
2764 | continue; |
2765 | |
2766 | // Make a new global with the correct type / address space. |
2767 | llvm::Type *Ty = getTypes().ConvertTypeForMem(T: ASTTy); |
2768 | llvm::PointerType *PTy = llvm::PointerType::get(ElementType: Ty, AddressSpace: TargetAS); |
2769 | |
2770 | // Replace all uses of the old global with a cast. Since we mutate the type |
2771 | // in place we neeed an intermediate that takes the spot of the old entry |
2772 | // until we can create the cast. |
2773 | llvm::GlobalVariable *DummyGV = new llvm::GlobalVariable( |
2774 | getModule(), Entry->getValueType(), false, |
2775 | llvm::GlobalValue::CommonLinkage, nullptr, "dummy" , nullptr, |
2776 | llvm::GlobalVariable::NotThreadLocal, Entry->getAddressSpace()); |
2777 | Entry->replaceAllUsesWith(V: DummyGV); |
2778 | |
2779 | Entry->mutateType(Ty: PTy); |
2780 | llvm::Constant *NewPtrForOldDecl = |
2781 | llvm::ConstantExpr::getPointerBitCastOrAddrSpaceCast( |
2782 | C: Entry, Ty: DummyGV->getType()); |
2783 | |
2784 | // Now we have a casted version of the changed global, the dummy can be |
2785 | // replaced and deleted. |
2786 | DummyGV->replaceAllUsesWith(V: NewPtrForOldDecl); |
2787 | DummyGV->eraseFromParent(); |
2788 | } |
2789 | } |
2790 | |
2791 | std::optional<CharUnits> |
2792 | CodeGenModule::getOMPAllocateAlignment(const VarDecl *VD) { |
2793 | if (const auto *AA = VD->getAttr<OMPAllocateDeclAttr>()) { |
2794 | if (Expr *Alignment = AA->getAlignment()) { |
2795 | unsigned UserAlign = |
2796 | Alignment->EvaluateKnownConstInt(Ctx: getContext()).getExtValue(); |
2797 | CharUnits NaturalAlign = |
2798 | getNaturalTypeAlignment(T: VD->getType().getNonReferenceType()); |
2799 | |
2800 | // OpenMP5.1 pg 185 lines 7-10 |
2801 | // Each item in the align modifier list must be aligned to the maximum |
2802 | // of the specified alignment and the type's natural alignment. |
2803 | return CharUnits::fromQuantity( |
2804 | Quantity: std::max<unsigned>(a: UserAlign, b: NaturalAlign.getQuantity())); |
2805 | } |
2806 | } |
2807 | return std::nullopt; |
2808 | } |
2809 | |