| 1 | //===--- CGDeclCXX.cpp - Emit LLVM Code for C++ 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 dealing with code generation of C++ declarations |
| 10 | // |
| 11 | //===----------------------------------------------------------------------===// |
| 12 | |
| 13 | #include "CGCXXABI.h" |
| 14 | #include "CGDebugInfo.h" |
| 15 | #include "CGHLSLRuntime.h" |
| 16 | #include "CGObjCRuntime.h" |
| 17 | #include "CGOpenMPRuntime.h" |
| 18 | #include "CodeGenFunction.h" |
| 19 | #include "TargetInfo.h" |
| 20 | #include "clang/AST/Attr.h" |
| 21 | #include "clang/Basic/LangOptions.h" |
| 22 | #include "llvm/ADT/StringExtras.h" |
| 23 | #include "llvm/IR/Intrinsics.h" |
| 24 | #include "llvm/IR/MDBuilder.h" |
| 25 | #include "llvm/Support/Path.h" |
| 26 | |
| 27 | using namespace clang; |
| 28 | using namespace CodeGen; |
| 29 | |
| 30 | static void EmitDeclInit(CodeGenFunction &CGF, const VarDecl &D, |
| 31 | ConstantAddress DeclPtr) { |
| 32 | assert( |
| 33 | (D.hasGlobalStorage() || |
| 34 | (D.hasLocalStorage() && CGF.getContext().getLangOpts().OpenCLCPlusPlus)) && |
| 35 | "VarDecl must have global or local (in the case of OpenCL) storage!" ); |
| 36 | assert(!D.getType()->isReferenceType() && |
| 37 | "Should not call EmitDeclInit on a reference!" ); |
| 38 | |
| 39 | QualType type = D.getType(); |
| 40 | LValue lv = CGF.MakeAddrLValue(Addr: DeclPtr, T: type); |
| 41 | |
| 42 | const Expr *Init = D.getInit(); |
| 43 | switch (CGF.getEvaluationKind(T: type)) { |
| 44 | case TEK_Scalar: { |
| 45 | CodeGenModule &CGM = CGF.CGM; |
| 46 | if (lv.isObjCStrong()) |
| 47 | CGM.getObjCRuntime().EmitObjCGlobalAssign(CGF, src: CGF.EmitScalarExpr(E: Init), |
| 48 | dest: DeclPtr, threadlocal: D.getTLSKind()); |
| 49 | else if (lv.isObjCWeak()) |
| 50 | CGM.getObjCRuntime().EmitObjCWeakAssign(CGF, src: CGF.EmitScalarExpr(E: Init), |
| 51 | dest: DeclPtr); |
| 52 | else |
| 53 | CGF.EmitScalarInit(init: Init, D: &D, lvalue: lv, capturedByInit: false); |
| 54 | return; |
| 55 | } |
| 56 | case TEK_Complex: |
| 57 | CGF.EmitComplexExprIntoLValue(E: Init, dest: lv, /*isInit*/ true); |
| 58 | return; |
| 59 | case TEK_Aggregate: |
| 60 | CGF.EmitAggExpr(E: Init, |
| 61 | AS: AggValueSlot::forLValue(LV: lv, isDestructed: AggValueSlot::IsDestructed, |
| 62 | needsGC: AggValueSlot::DoesNotNeedGCBarriers, |
| 63 | isAliased: AggValueSlot::IsNotAliased, |
| 64 | mayOverlap: AggValueSlot::DoesNotOverlap)); |
| 65 | return; |
| 66 | } |
| 67 | llvm_unreachable("bad evaluation kind" ); |
| 68 | } |
| 69 | |
| 70 | /// Emit code to cause the destruction of the given variable with |
| 71 | /// static storage duration. |
| 72 | static void EmitDeclDestroy(CodeGenFunction &CGF, const VarDecl &D, |
| 73 | ConstantAddress Addr) { |
| 74 | // Honor __attribute__((no_destroy)) and bail instead of attempting |
| 75 | // to emit a reference to a possibly nonexistent destructor, which |
| 76 | // in turn can cause a crash. This will result in a global constructor |
| 77 | // that isn't balanced out by a destructor call as intended by the |
| 78 | // attribute. This also checks for -fno-c++-static-destructors and |
| 79 | // bails even if the attribute is not present. |
| 80 | QualType::DestructionKind DtorKind = D.needsDestruction(Ctx: CGF.getContext()); |
| 81 | |
| 82 | // FIXME: __attribute__((cleanup)) ? |
| 83 | |
| 84 | switch (DtorKind) { |
| 85 | case QualType::DK_none: |
| 86 | return; |
| 87 | |
| 88 | case QualType::DK_cxx_destructor: |
| 89 | break; |
| 90 | |
| 91 | case QualType::DK_objc_strong_lifetime: |
| 92 | case QualType::DK_objc_weak_lifetime: |
| 93 | case QualType::DK_nontrivial_c_struct: |
| 94 | // We don't care about releasing objects during process teardown. |
| 95 | assert(!D.getTLSKind() && "should have rejected this" ); |
| 96 | return; |
| 97 | } |
| 98 | |
| 99 | llvm::FunctionCallee Func; |
| 100 | llvm::Constant *Argument; |
| 101 | |
| 102 | CodeGenModule &CGM = CGF.CGM; |
| 103 | QualType Type = D.getType(); |
| 104 | |
| 105 | // Special-case non-array C++ destructors, if they have the right signature. |
| 106 | // Under some ABIs, destructors return this instead of void, and cannot be |
| 107 | // passed directly to __cxa_atexit if the target does not allow this |
| 108 | // mismatch. |
| 109 | const CXXRecordDecl *Record = Type->getAsCXXRecordDecl(); |
| 110 | bool CanRegisterDestructor = |
| 111 | Record && (!CGM.getCXXABI().HasThisReturn( |
| 112 | GD: GlobalDecl(Record->getDestructor(), Dtor_Complete)) || |
| 113 | CGM.getCXXABI().canCallMismatchedFunctionType()); |
| 114 | // If __cxa_atexit is disabled via a flag, a different helper function is |
| 115 | // generated elsewhere which uses atexit instead, and it takes the destructor |
| 116 | // directly. |
| 117 | bool UsingExternalHelper = !CGM.getCodeGenOpts().CXAAtExit; |
| 118 | if (Record && (CanRegisterDestructor || UsingExternalHelper)) { |
| 119 | assert(!Record->hasTrivialDestructor()); |
| 120 | CXXDestructorDecl *Dtor = Record->getDestructor(); |
| 121 | |
| 122 | Func = CGM.getAddrAndTypeOfCXXStructor(GD: GlobalDecl(Dtor, Dtor_Complete)); |
| 123 | if (CGF.getContext().getLangOpts().OpenCL) { |
| 124 | auto DestAS = |
| 125 | CGM.getTargetCodeGenInfo().getAddrSpaceOfCxaAtexitPtrParam(); |
| 126 | auto DestTy = llvm::PointerType::get( |
| 127 | C&: CGM.getLLVMContext(), AddressSpace: CGM.getContext().getTargetAddressSpace(AS: DestAS)); |
| 128 | auto SrcAS = D.getType().getQualifiers().getAddressSpace(); |
| 129 | if (DestAS == SrcAS) |
| 130 | Argument = Addr.getPointer(); |
| 131 | else |
| 132 | // FIXME: On addr space mismatch we are passing NULL. The generation |
| 133 | // of the global destructor function should be adjusted accordingly. |
| 134 | Argument = llvm::ConstantPointerNull::get(T: DestTy); |
| 135 | } else { |
| 136 | Argument = Addr.getPointer(); |
| 137 | } |
| 138 | // Otherwise, the standard logic requires a helper function. |
| 139 | } else { |
| 140 | Addr = Addr.withElementType(ElemTy: CGF.ConvertTypeForMem(T: Type)); |
| 141 | Func = CodeGenFunction(CGM) |
| 142 | .generateDestroyHelper(addr: Addr, type: Type, destroyer: CGF.getDestroyer(destructionKind: DtorKind), |
| 143 | useEHCleanupForArray: CGF.needsEHCleanup(kind: DtorKind), VD: &D); |
| 144 | Argument = llvm::Constant::getNullValue(Ty: CGF.Int8PtrTy); |
| 145 | } |
| 146 | |
| 147 | CGM.getCXXABI().registerGlobalDtor(CGF, D, Dtor: Func, Addr: Argument); |
| 148 | } |
| 149 | |
| 150 | /// Emit code to cause the variable at the given address to be considered as |
| 151 | /// constant from this point onwards. |
| 152 | static void EmitDeclInvariant(CodeGenFunction &CGF, const VarDecl &D, |
| 153 | llvm::Constant *Addr) { |
| 154 | return CGF.EmitInvariantStart( |
| 155 | Addr, Size: CGF.getContext().getTypeSizeInChars(T: D.getType())); |
| 156 | } |
| 157 | |
| 158 | void CodeGenFunction::EmitInvariantStart(llvm::Constant *Addr, CharUnits Size) { |
| 159 | // Do not emit the intrinsic if we're not optimizing. |
| 160 | if (!CGM.getCodeGenOpts().OptimizationLevel) |
| 161 | return; |
| 162 | |
| 163 | // Grab the llvm.invariant.start intrinsic. |
| 164 | llvm::Intrinsic::ID InvStartID = llvm::Intrinsic::invariant_start; |
| 165 | // Overloaded address space type. |
| 166 | assert(Addr->getType()->isPointerTy() && "Address must be a pointer" ); |
| 167 | llvm::Type *ObjectPtr[1] = {Addr->getType()}; |
| 168 | llvm::Function *InvariantStart = CGM.getIntrinsic(IID: InvStartID, Tys: ObjectPtr); |
| 169 | |
| 170 | // Emit a call with the size in bytes of the object. |
| 171 | uint64_t Width = Size.getQuantity(); |
| 172 | llvm::Value *Args[2] = {llvm::ConstantInt::getSigned(Ty: Int64Ty, V: Width), Addr}; |
| 173 | Builder.CreateCall(Callee: InvariantStart, Args); |
| 174 | } |
| 175 | |
| 176 | void CodeGenFunction::EmitCXXGlobalVarDeclInit(const VarDecl &D, |
| 177 | llvm::GlobalVariable *GV, |
| 178 | bool PerformInit) { |
| 179 | |
| 180 | const Expr *Init = D.getInit(); |
| 181 | QualType T = D.getType(); |
| 182 | |
| 183 | // The address space of a static local variable (DeclPtr) may be different |
| 184 | // from the address space of the "this" argument of the constructor. In that |
| 185 | // case, we need an addrspacecast before calling the constructor. |
| 186 | // |
| 187 | // struct StructWithCtor { |
| 188 | // __device__ StructWithCtor() {...} |
| 189 | // }; |
| 190 | // __device__ void foo() { |
| 191 | // __shared__ StructWithCtor s; |
| 192 | // ... |
| 193 | // } |
| 194 | // |
| 195 | // For example, in the above CUDA code, the static local variable s has a |
| 196 | // "shared" address space qualifier, but the constructor of StructWithCtor |
| 197 | // expects "this" in the "generic" address space. |
| 198 | unsigned ExpectedAddrSpace = getTypes().getTargetAddressSpace(T); |
| 199 | unsigned ActualAddrSpace = GV->getAddressSpace(); |
| 200 | llvm::Constant *DeclPtr = GV; |
| 201 | if (ActualAddrSpace != ExpectedAddrSpace) { |
| 202 | llvm::PointerType *PTy = |
| 203 | llvm::PointerType::get(C&: getLLVMContext(), AddressSpace: ExpectedAddrSpace); |
| 204 | DeclPtr = llvm::ConstantExpr::getAddrSpaceCast(C: DeclPtr, Ty: PTy); |
| 205 | } |
| 206 | |
| 207 | ConstantAddress DeclAddr( |
| 208 | DeclPtr, GV->getValueType(), getContext().getDeclAlign(D: &D)); |
| 209 | |
| 210 | if (!T->isReferenceType()) { |
| 211 | if (getLangOpts().OpenMP && !getLangOpts().OpenMPSimd && |
| 212 | D.hasAttr<OMPThreadPrivateDeclAttr>()) { |
| 213 | (void)CGM.getOpenMPRuntime().emitThreadPrivateVarDefinition( |
| 214 | VD: &D, VDAddr: DeclAddr, Loc: D.getAttr<OMPThreadPrivateDeclAttr>()->getLocation(), |
| 215 | PerformInit, CGF: this); |
| 216 | } |
| 217 | bool NeedsDtor = |
| 218 | D.needsDestruction(Ctx: getContext()) == QualType::DK_cxx_destructor; |
| 219 | if (PerformInit) |
| 220 | EmitDeclInit(CGF&: *this, D, DeclPtr: DeclAddr); |
| 221 | if (D.getType().isConstantStorage(Ctx: getContext(), ExcludeCtor: true, ExcludeDtor: !NeedsDtor)) |
| 222 | EmitDeclInvariant(CGF&: *this, D, Addr: DeclPtr); |
| 223 | else |
| 224 | EmitDeclDestroy(CGF&: *this, D, Addr: DeclAddr); |
| 225 | return; |
| 226 | } |
| 227 | |
| 228 | assert(PerformInit && "cannot have constant initializer which needs " |
| 229 | "destruction for reference" ); |
| 230 | RValue RV = EmitReferenceBindingToExpr(E: Init); |
| 231 | EmitStoreOfScalar(Value: RV.getScalarVal(), Addr: DeclAddr, Volatile: false, Ty: T); |
| 232 | } |
| 233 | |
| 234 | /// Create a stub function, suitable for being passed to atexit, |
| 235 | /// which passes the given address to the given destructor function. |
| 236 | llvm::Constant *CodeGenFunction::createAtExitStub(const VarDecl &VD, |
| 237 | llvm::FunctionCallee dtor, |
| 238 | llvm::Constant *addr) { |
| 239 | // Get the destructor function type, void(*)(void). |
| 240 | llvm::FunctionType *ty = llvm::FunctionType::get(Result: CGM.VoidTy, isVarArg: false); |
| 241 | SmallString<256> FnName; |
| 242 | { |
| 243 | llvm::raw_svector_ostream Out(FnName); |
| 244 | CGM.getCXXABI().getMangleContext().mangleDynamicAtExitDestructor(D: &VD, Out); |
| 245 | } |
| 246 | |
| 247 | const CGFunctionInfo &FI = CGM.getTypes().arrangeNullaryFunction(); |
| 248 | llvm::Function *fn = CGM.CreateGlobalInitOrCleanUpFunction( |
| 249 | ty, name: FnName.str(), FI, Loc: VD.getLocation()); |
| 250 | |
| 251 | CodeGenFunction CGF(CGM); |
| 252 | |
| 253 | CGF.StartFunction(GD: GlobalDecl(&VD, DynamicInitKind::AtExit), |
| 254 | RetTy: CGM.getContext().VoidTy, Fn: fn, FnInfo: FI, Args: FunctionArgList(), |
| 255 | Loc: VD.getLocation(), StartLoc: VD.getInit()->getExprLoc()); |
| 256 | // Emit an artificial location for this function. |
| 257 | auto AL = ApplyDebugLocation::CreateArtificial(CGF); |
| 258 | |
| 259 | llvm::CallInst *call = CGF.Builder.CreateCall(Callee: dtor, Args: addr); |
| 260 | |
| 261 | // Make sure the call and the callee agree on calling convention. |
| 262 | if (auto *dtorFn = dyn_cast<llvm::Function>( |
| 263 | Val: dtor.getCallee()->stripPointerCastsAndAliases())) |
| 264 | call->setCallingConv(dtorFn->getCallingConv()); |
| 265 | |
| 266 | CGF.FinishFunction(); |
| 267 | |
| 268 | // Get a proper function pointer. |
| 269 | FunctionProtoType::ExtProtoInfo EPI(getContext().getDefaultCallingConvention( |
| 270 | /*IsVariadic=*/false, /*IsCXXMethod=*/false)); |
| 271 | QualType fnType = getContext().getFunctionType(ResultTy: getContext().VoidTy, |
| 272 | Args: {getContext().VoidPtrTy}, EPI); |
| 273 | return CGM.getFunctionPointer(Pointer: fn, FunctionType: fnType); |
| 274 | } |
| 275 | |
| 276 | /// Create a stub function, suitable for being passed to __pt_atexit_np, |
| 277 | /// which passes the given address to the given destructor function. |
| 278 | llvm::Function *CodeGenFunction::createTLSAtExitStub( |
| 279 | const VarDecl &D, llvm::FunctionCallee Dtor, llvm::Constant *Addr, |
| 280 | llvm::FunctionCallee &AtExit) { |
| 281 | SmallString<256> FnName; |
| 282 | { |
| 283 | llvm::raw_svector_ostream Out(FnName); |
| 284 | CGM.getCXXABI().getMangleContext().mangleDynamicAtExitDestructor(D: &D, Out); |
| 285 | } |
| 286 | |
| 287 | const CGFunctionInfo &FI = CGM.getTypes().arrangeLLVMFunctionInfo( |
| 288 | returnType: getContext().IntTy, opts: FnInfoOpts::None, argTypes: {getContext().IntTy}, |
| 289 | info: FunctionType::ExtInfo(), paramInfos: {}, args: RequiredArgs::All); |
| 290 | |
| 291 | // Get the stub function type, int(*)(int,...). |
| 292 | llvm::FunctionType *StubTy = |
| 293 | llvm::FunctionType::get(Result: CGM.IntTy, Params: {CGM.IntTy}, isVarArg: true); |
| 294 | |
| 295 | llvm::Function *DtorStub = CGM.CreateGlobalInitOrCleanUpFunction( |
| 296 | ty: StubTy, name: FnName.str(), FI, Loc: D.getLocation()); |
| 297 | |
| 298 | CodeGenFunction CGF(CGM); |
| 299 | |
| 300 | FunctionArgList Args; |
| 301 | ImplicitParamDecl IPD(CGM.getContext(), CGM.getContext().IntTy, |
| 302 | ImplicitParamKind::Other); |
| 303 | Args.push_back(Elt: &IPD); |
| 304 | QualType ResTy = CGM.getContext().IntTy; |
| 305 | |
| 306 | CGF.StartFunction(GD: GlobalDecl(&D, DynamicInitKind::AtExit), RetTy: ResTy, Fn: DtorStub, |
| 307 | FnInfo: FI, Args, Loc: D.getLocation(), StartLoc: D.getInit()->getExprLoc()); |
| 308 | |
| 309 | // Emit an artificial location for this function. |
| 310 | auto AL = ApplyDebugLocation::CreateArtificial(CGF); |
| 311 | |
| 312 | llvm::CallInst *call = CGF.Builder.CreateCall(Callee: Dtor, Args: Addr); |
| 313 | |
| 314 | // Make sure the call and the callee agree on calling convention. |
| 315 | if (auto *DtorFn = dyn_cast<llvm::Function>( |
| 316 | Val: Dtor.getCallee()->stripPointerCastsAndAliases())) |
| 317 | call->setCallingConv(DtorFn->getCallingConv()); |
| 318 | |
| 319 | // Return 0 from function |
| 320 | CGF.Builder.CreateStore(Val: llvm::Constant::getNullValue(Ty: CGM.IntTy), |
| 321 | Addr: CGF.ReturnValue); |
| 322 | |
| 323 | CGF.FinishFunction(); |
| 324 | |
| 325 | return DtorStub; |
| 326 | } |
| 327 | |
| 328 | /// Register a global destructor using the C atexit runtime function. |
| 329 | void CodeGenFunction::registerGlobalDtorWithAtExit(const VarDecl &VD, |
| 330 | llvm::FunctionCallee dtor, |
| 331 | llvm::Constant *addr) { |
| 332 | // Create a function which calls the destructor. |
| 333 | llvm::Constant *dtorStub = createAtExitStub(VD, dtor, addr); |
| 334 | registerGlobalDtorWithAtExit(dtorStub); |
| 335 | } |
| 336 | |
| 337 | /// Register a global destructor using the LLVM 'llvm.global_dtors' global. |
| 338 | void CodeGenFunction::registerGlobalDtorWithLLVM(const VarDecl &VD, |
| 339 | llvm::FunctionCallee Dtor, |
| 340 | llvm::Constant *Addr) { |
| 341 | // Create a function which calls the destructor. |
| 342 | llvm::Function *dtorStub = |
| 343 | cast<llvm::Function>(Val: createAtExitStub(VD, dtor: Dtor, addr: Addr)); |
| 344 | CGM.AddGlobalDtor(Dtor: dtorStub); |
| 345 | } |
| 346 | |
| 347 | void CodeGenFunction::registerGlobalDtorWithAtExit(llvm::Constant *dtorStub) { |
| 348 | // extern "C" int atexit(void (*f)(void)); |
| 349 | assert(dtorStub->getType()->isPointerTy() && |
| 350 | "Argument to atexit has a wrong type." ); |
| 351 | |
| 352 | llvm::FunctionType *atexitTy = |
| 353 | llvm::FunctionType::get(Result: IntTy, Params: dtorStub->getType(), isVarArg: false); |
| 354 | |
| 355 | llvm::FunctionCallee atexit = |
| 356 | CGM.CreateRuntimeFunction(Ty: atexitTy, Name: "atexit" , ExtraAttrs: llvm::AttributeList(), |
| 357 | /*Local=*/true); |
| 358 | if (llvm::Function *atexitFn = dyn_cast<llvm::Function>(Val: atexit.getCallee())) |
| 359 | atexitFn->setDoesNotThrow(); |
| 360 | |
| 361 | EmitNounwindRuntimeCall(callee: atexit, args: dtorStub); |
| 362 | } |
| 363 | |
| 364 | llvm::Value * |
| 365 | CodeGenFunction::unregisterGlobalDtorWithUnAtExit(llvm::Constant *dtorStub) { |
| 366 | // The unatexit subroutine unregisters __dtor functions that were previously |
| 367 | // registered by the atexit subroutine. If the referenced function is found, |
| 368 | // it is removed from the list of functions that are called at normal program |
| 369 | // termination and the unatexit returns a value of 0, otherwise a non-zero |
| 370 | // value is returned. |
| 371 | // |
| 372 | // extern "C" int unatexit(void (*f)(void)); |
| 373 | assert(dtorStub->getType()->isPointerTy() && |
| 374 | "Argument to unatexit has a wrong type." ); |
| 375 | |
| 376 | llvm::FunctionType *unatexitTy = |
| 377 | llvm::FunctionType::get(Result: IntTy, Params: {dtorStub->getType()}, /*isVarArg=*/false); |
| 378 | |
| 379 | llvm::FunctionCallee unatexit = |
| 380 | CGM.CreateRuntimeFunction(Ty: unatexitTy, Name: "unatexit" , ExtraAttrs: llvm::AttributeList()); |
| 381 | |
| 382 | cast<llvm::Function>(Val: unatexit.getCallee())->setDoesNotThrow(); |
| 383 | |
| 384 | return EmitNounwindRuntimeCall(callee: unatexit, args: dtorStub); |
| 385 | } |
| 386 | |
| 387 | void CodeGenFunction::EmitCXXGuardedInit(const VarDecl &D, |
| 388 | llvm::GlobalVariable *DeclPtr, |
| 389 | bool PerformInit) { |
| 390 | // If we've been asked to forbid guard variables, emit an error now. |
| 391 | // This diagnostic is hard-coded for Darwin's use case; we can find |
| 392 | // better phrasing if someone else needs it. |
| 393 | if (CGM.getCodeGenOpts().ForbidGuardVariables) |
| 394 | CGM.Error(loc: D.getLocation(), |
| 395 | error: "this initialization requires a guard variable, which " |
| 396 | "the kernel does not support" ); |
| 397 | |
| 398 | CGM.getCXXABI().EmitGuardedInit(CGF&: *this, D, DeclPtr, PerformInit); |
| 399 | } |
| 400 | |
| 401 | void CodeGenFunction::EmitCXXGuardedInitBranch(llvm::Value *NeedsInit, |
| 402 | llvm::BasicBlock *InitBlock, |
| 403 | llvm::BasicBlock *NoInitBlock, |
| 404 | GuardKind Kind, |
| 405 | const VarDecl *D) { |
| 406 | assert((Kind == GuardKind::TlsGuard || D) && "no guarded variable" ); |
| 407 | |
| 408 | // A guess at how many times we will enter the initialization of a |
| 409 | // variable, depending on the kind of variable. |
| 410 | static const uint64_t InitsPerTLSVar = 1024; |
| 411 | static const uint64_t InitsPerLocalVar = 1024 * 1024; |
| 412 | |
| 413 | llvm::MDNode *Weights; |
| 414 | if (Kind == GuardKind::VariableGuard && !D->isLocalVarDecl()) { |
| 415 | // For non-local variables, don't apply any weighting for now. Due to our |
| 416 | // use of COMDATs, we expect there to be at most one initialization of the |
| 417 | // variable per DSO, but we have no way to know how many DSOs will try to |
| 418 | // initialize the variable. |
| 419 | Weights = nullptr; |
| 420 | } else { |
| 421 | uint64_t NumInits; |
| 422 | // FIXME: For the TLS case, collect and use profiling information to |
| 423 | // determine a more accurate brach weight. |
| 424 | if (Kind == GuardKind::TlsGuard || D->getTLSKind()) |
| 425 | NumInits = InitsPerTLSVar; |
| 426 | else |
| 427 | NumInits = InitsPerLocalVar; |
| 428 | |
| 429 | // The probability of us entering the initializer is |
| 430 | // 1 / (total number of times we attempt to initialize the variable). |
| 431 | llvm::MDBuilder MDHelper(CGM.getLLVMContext()); |
| 432 | Weights = MDHelper.createBranchWeights(TrueWeight: 1, FalseWeight: NumInits - 1); |
| 433 | } |
| 434 | |
| 435 | Builder.CreateCondBr(Cond: NeedsInit, True: InitBlock, False: NoInitBlock, BranchWeights: Weights); |
| 436 | } |
| 437 | |
| 438 | llvm::Function *CodeGenModule::CreateGlobalInitOrCleanUpFunction( |
| 439 | llvm::FunctionType *FTy, const Twine &Name, const CGFunctionInfo &FI, |
| 440 | SourceLocation Loc, bool TLS, llvm::GlobalVariable::LinkageTypes Linkage) { |
| 441 | llvm::Function *Fn = llvm::Function::Create(Ty: FTy, Linkage, N: Name, M: &getModule()); |
| 442 | |
| 443 | if (!getLangOpts().AppleKext && !TLS) { |
| 444 | // Set the section if needed. |
| 445 | if (const char *Section = getTarget().getStaticInitSectionSpecifier()) |
| 446 | Fn->setSection(Section); |
| 447 | } |
| 448 | |
| 449 | if (Linkage == llvm::GlobalVariable::InternalLinkage) |
| 450 | SetInternalFunctionAttributes(GD: GlobalDecl(), F: Fn, FI); |
| 451 | else { |
| 452 | SetLLVMFunctionAttributes(GD: GlobalDecl(), Info: FI, F: Fn, IsThunk: false); |
| 453 | SetLLVMFunctionAttributesForDefinition(D: nullptr, F: Fn); |
| 454 | getTargetCodeGenInfo().setTargetAttributes(D: nullptr, GV: Fn, M&: *this); |
| 455 | } |
| 456 | |
| 457 | Fn->setCallingConv(getRuntimeCC()); |
| 458 | |
| 459 | if (!getLangOpts().Exceptions) |
| 460 | Fn->setDoesNotThrow(); |
| 461 | |
| 462 | if (getLangOpts().Sanitize.has(K: SanitizerKind::Address) && |
| 463 | !isInNoSanitizeList(Kind: SanitizerKind::Address, Fn, Loc)) |
| 464 | Fn->addFnAttr(Kind: llvm::Attribute::SanitizeAddress); |
| 465 | |
| 466 | if (getLangOpts().Sanitize.has(K: SanitizerKind::KernelAddress) && |
| 467 | !isInNoSanitizeList(Kind: SanitizerKind::KernelAddress, Fn, Loc)) |
| 468 | Fn->addFnAttr(Kind: llvm::Attribute::SanitizeAddress); |
| 469 | |
| 470 | if (getLangOpts().Sanitize.has(K: SanitizerKind::HWAddress) && |
| 471 | !isInNoSanitizeList(Kind: SanitizerKind::HWAddress, Fn, Loc)) |
| 472 | Fn->addFnAttr(Kind: llvm::Attribute::SanitizeHWAddress); |
| 473 | |
| 474 | if (getLangOpts().Sanitize.has(K: SanitizerKind::KernelHWAddress) && |
| 475 | !isInNoSanitizeList(Kind: SanitizerKind::KernelHWAddress, Fn, Loc)) |
| 476 | Fn->addFnAttr(Kind: llvm::Attribute::SanitizeHWAddress); |
| 477 | |
| 478 | if (getLangOpts().Sanitize.has(K: SanitizerKind::MemtagStack) && |
| 479 | !isInNoSanitizeList(Kind: SanitizerKind::MemtagStack, Fn, Loc)) |
| 480 | Fn->addFnAttr(Kind: llvm::Attribute::SanitizeMemTag); |
| 481 | |
| 482 | if (getLangOpts().Sanitize.has(K: SanitizerKind::Type) && |
| 483 | !isInNoSanitizeList(Kind: SanitizerKind::Type, Fn, Loc)) |
| 484 | Fn->addFnAttr(Kind: llvm::Attribute::SanitizeType); |
| 485 | |
| 486 | if (getLangOpts().Sanitize.has(K: SanitizerKind::Thread) && |
| 487 | !isInNoSanitizeList(Kind: SanitizerKind::Thread, Fn, Loc)) |
| 488 | Fn->addFnAttr(Kind: llvm::Attribute::SanitizeThread); |
| 489 | |
| 490 | if (getLangOpts().Sanitize.has(K: SanitizerKind::NumericalStability) && |
| 491 | !isInNoSanitizeList(Kind: SanitizerKind::NumericalStability, Fn, Loc)) |
| 492 | Fn->addFnAttr(Kind: llvm::Attribute::SanitizeNumericalStability); |
| 493 | |
| 494 | if (getLangOpts().Sanitize.has(K: SanitizerKind::Memory) && |
| 495 | !isInNoSanitizeList(Kind: SanitizerKind::Memory, Fn, Loc)) |
| 496 | Fn->addFnAttr(Kind: llvm::Attribute::SanitizeMemory); |
| 497 | |
| 498 | if (getLangOpts().Sanitize.has(K: SanitizerKind::KernelMemory) && |
| 499 | !isInNoSanitizeList(Kind: SanitizerKind::KernelMemory, Fn, Loc)) |
| 500 | Fn->addFnAttr(Kind: llvm::Attribute::SanitizeMemory); |
| 501 | |
| 502 | if (getLangOpts().Sanitize.has(K: SanitizerKind::SafeStack) && |
| 503 | !isInNoSanitizeList(Kind: SanitizerKind::SafeStack, Fn, Loc)) |
| 504 | Fn->addFnAttr(Kind: llvm::Attribute::SafeStack); |
| 505 | |
| 506 | if (getLangOpts().Sanitize.has(K: SanitizerKind::ShadowCallStack) && |
| 507 | !isInNoSanitizeList(Kind: SanitizerKind::ShadowCallStack, Fn, Loc)) |
| 508 | Fn->addFnAttr(Kind: llvm::Attribute::ShadowCallStack); |
| 509 | |
| 510 | return Fn; |
| 511 | } |
| 512 | |
| 513 | /// Create a global pointer to a function that will initialize a global |
| 514 | /// variable. The user has requested that this pointer be emitted in a specific |
| 515 | /// section. |
| 516 | void CodeGenModule::EmitPointerToInitFunc(const VarDecl *D, |
| 517 | llvm::GlobalVariable *GV, |
| 518 | llvm::Function *InitFunc, |
| 519 | InitSegAttr *ISA) { |
| 520 | llvm::GlobalVariable *PtrArray = new llvm::GlobalVariable( |
| 521 | TheModule, InitFunc->getType(), /*isConstant=*/true, |
| 522 | llvm::GlobalValue::PrivateLinkage, InitFunc, "__cxx_init_fn_ptr" ); |
| 523 | PtrArray->setSection(ISA->getSection()); |
| 524 | addUsedGlobal(GV: PtrArray); |
| 525 | |
| 526 | // If the GV is already in a comdat group, then we have to join it. |
| 527 | if (llvm::Comdat *C = GV->getComdat()) |
| 528 | PtrArray->setComdat(C); |
| 529 | } |
| 530 | |
| 531 | void |
| 532 | CodeGenModule::EmitCXXGlobalVarDeclInitFunc(const VarDecl *D, |
| 533 | llvm::GlobalVariable *Addr, |
| 534 | bool PerformInit) { |
| 535 | |
| 536 | // According to E.2.3.1 in CUDA-7.5 Programming guide: __device__, |
| 537 | // __constant__ and __shared__ variables defined in namespace scope, |
| 538 | // that are of class type, cannot have a non-empty constructor. All |
| 539 | // the checks have been done in Sema by now. Whatever initializers |
| 540 | // are allowed are empty and we just need to ignore them here. |
| 541 | if (getLangOpts().CUDAIsDevice && !getLangOpts().GPUAllowDeviceInit && |
| 542 | (D->hasAttr<CUDADeviceAttr>() || D->hasAttr<CUDAConstantAttr>() || |
| 543 | D->hasAttr<CUDASharedAttr>())) |
| 544 | return; |
| 545 | |
| 546 | // Check if we've already initialized this decl. |
| 547 | auto I = DelayedCXXInitPosition.find(Val: D); |
| 548 | if (I != DelayedCXXInitPosition.end() && I->second == ~0U) |
| 549 | return; |
| 550 | |
| 551 | llvm::FunctionType *FTy = llvm::FunctionType::get(Result: VoidTy, isVarArg: false); |
| 552 | SmallString<256> FnName; |
| 553 | { |
| 554 | llvm::raw_svector_ostream Out(FnName); |
| 555 | getCXXABI().getMangleContext().mangleDynamicInitializer(D, Out); |
| 556 | } |
| 557 | |
| 558 | // Create a variable initialization function. |
| 559 | llvm::Function *Fn = CreateGlobalInitOrCleanUpFunction( |
| 560 | FTy, Name: FnName.str(), FI: getTypes().arrangeNullaryFunction(), Loc: D->getLocation()); |
| 561 | |
| 562 | auto *ISA = D->getAttr<InitSegAttr>(); |
| 563 | CodeGenFunction(*this).GenerateCXXGlobalVarDeclInitFunc(Fn, D, Addr, |
| 564 | PerformInit); |
| 565 | |
| 566 | llvm::GlobalVariable *COMDATKey = |
| 567 | supportsCOMDAT() && D->isExternallyVisible() ? Addr : nullptr; |
| 568 | |
| 569 | if (D->getTLSKind()) { |
| 570 | // FIXME: Should we support init_priority for thread_local? |
| 571 | // FIXME: We only need to register one __cxa_thread_atexit function for the |
| 572 | // entire TU. |
| 573 | CXXThreadLocalInits.push_back(x: Fn); |
| 574 | CXXThreadLocalInitVars.push_back(x: D); |
| 575 | } else if (PerformInit && ISA) { |
| 576 | // Contract with backend that "init_seg(compiler)" corresponds to priority |
| 577 | // 200 and "init_seg(lib)" corresponds to priority 400. |
| 578 | int Priority = -1; |
| 579 | if (ISA->getSection() == ".CRT$XCC" ) |
| 580 | Priority = 200; |
| 581 | else if (ISA->getSection() == ".CRT$XCL" ) |
| 582 | Priority = 400; |
| 583 | |
| 584 | if (Priority != -1) |
| 585 | AddGlobalCtor(Ctor: Fn, Priority, LexOrder: ~0U, AssociatedData: COMDATKey); |
| 586 | else |
| 587 | EmitPointerToInitFunc(D, GV: Addr, InitFunc: Fn, ISA); |
| 588 | } else if (auto *IPA = D->getAttr<InitPriorityAttr>()) { |
| 589 | OrderGlobalInitsOrStermFinalizers Key(IPA->getPriority(), |
| 590 | PrioritizedCXXGlobalInits.size()); |
| 591 | PrioritizedCXXGlobalInits.push_back(Elt: std::make_pair(x&: Key, y&: Fn)); |
| 592 | } else if (isTemplateInstantiation(Kind: D->getTemplateSpecializationKind()) || |
| 593 | !isUniqueGVALinkage(L: getContext().GetGVALinkageForVariable(VD: D)) || |
| 594 | D->hasAttr<SelectAnyAttr>()) { |
| 595 | // For vague linkage globals, put the initializer into its own global_ctors |
| 596 | // entry with the global as a comdat key. This ensures at most one |
| 597 | // initializer per DSO runs during DSO dynamic initialization. |
| 598 | // |
| 599 | // For ELF platforms, this is an important code size and startup time |
| 600 | // optimization. For dynamic, non-hidden symbols, the weak guard variable |
| 601 | // remains to ensure that other DSOs do not re-initialize the global. |
| 602 | // |
| 603 | // For PE-COFF platforms, there is no guard variable, and COMDAT |
| 604 | // associativity is the only way to ensure vauge linkage globals are |
| 605 | // initialized exactly once. |
| 606 | // |
| 607 | // MachO is the only remaining platform with no comdats that doesn't |
| 608 | // benefit from this optimization. The rest are mainly modeled on ELF |
| 609 | // behavior. |
| 610 | // |
| 611 | // C++ requires that inline global variables are initialized in source |
| 612 | // order, but this requirement does not exist for templated entities. |
| 613 | // llvm.global_ctors does not guarantee initialization order, so in |
| 614 | // general, Clang does not fully conform to the ordering requirement. |
| 615 | // However, in practice, LLVM emits global_ctors in the provided order, and |
| 616 | // users typically don't rely on ordering between inline globals in |
| 617 | // different headers which are then transitively included in varying order. |
| 618 | // Clang's current behavior is a practical tradeoff, since dropping the |
| 619 | // comdat would lead to unacceptable impact on code size and startup time. |
| 620 | // |
| 621 | // FIXME: Find a solution to guarantee source-order initialization of |
| 622 | // inline variables. |
| 623 | // |
| 624 | // C++ [basic.start.init]p2: |
| 625 | // Definitions of explicitly specialized class template static data |
| 626 | // members have ordered initialization. Other class template static data |
| 627 | // members (i.e., implicitly or explicitly instantiated specializations) |
| 628 | // have unordered initialization. |
| 629 | // |
| 630 | // CXXGlobalInits.size() is the lex order number for the next deferred |
| 631 | // VarDecl. Use it when the current VarDecl is non-deferred. Although this |
| 632 | // lex order number is shared between current VarDecl and some following |
| 633 | // VarDecls, their order of insertion into `llvm.global_ctors` is the same |
| 634 | // as the lexing order and the following stable sort would preserve such |
| 635 | // order. |
| 636 | I = DelayedCXXInitPosition.find(Val: D); |
| 637 | unsigned LexOrder = |
| 638 | I == DelayedCXXInitPosition.end() ? CXXGlobalInits.size() : I->second; |
| 639 | AddGlobalCtor(Ctor: Fn, Priority: 65535, LexOrder, AssociatedData: COMDATKey); |
| 640 | if (COMDATKey && (getTriple().isOSBinFormatELF() || |
| 641 | getTarget().getCXXABI().isMicrosoft())) { |
| 642 | // When COMDAT is used on ELF or in the MS C++ ABI, the key must be in |
| 643 | // llvm.used to prevent linker GC. |
| 644 | addUsedGlobal(GV: COMDATKey); |
| 645 | } |
| 646 | |
| 647 | // If we used a COMDAT key for the global ctor, the init function can be |
| 648 | // discarded if the global ctor entry is discarded. |
| 649 | // FIXME: Do we need to restrict this to ELF and Wasm? |
| 650 | llvm::Comdat *C = Addr->getComdat(); |
| 651 | if (COMDATKey && C && |
| 652 | (getTarget().getTriple().isOSBinFormatELF() || |
| 653 | getTarget().getTriple().isOSBinFormatWasm())) { |
| 654 | Fn->setComdat(C); |
| 655 | } |
| 656 | } else { |
| 657 | I = DelayedCXXInitPosition.find(Val: D); // Re-do lookup in case of re-hash. |
| 658 | if (I == DelayedCXXInitPosition.end()) { |
| 659 | CXXGlobalInits.push_back(x: Fn); |
| 660 | } else if (I->second != ~0U) { |
| 661 | assert(I->second < CXXGlobalInits.size() && |
| 662 | CXXGlobalInits[I->second] == nullptr); |
| 663 | CXXGlobalInits[I->second] = Fn; |
| 664 | } |
| 665 | } |
| 666 | |
| 667 | // Remember that we already emitted the initializer for this global. |
| 668 | DelayedCXXInitPosition[D] = ~0U; |
| 669 | } |
| 670 | |
| 671 | void CodeGenModule::EmitCXXThreadLocalInitFunc() { |
| 672 | getCXXABI().EmitThreadLocalInitFuncs( |
| 673 | CGM&: *this, CXXThreadLocals, CXXThreadLocalInits, CXXThreadLocalInitVars); |
| 674 | |
| 675 | CXXThreadLocalInits.clear(); |
| 676 | CXXThreadLocalInitVars.clear(); |
| 677 | CXXThreadLocals.clear(); |
| 678 | } |
| 679 | |
| 680 | /* Build the initializer for a C++20 module: |
| 681 | This is arranged to be run only once regardless of how many times the module |
| 682 | might be included transitively. This arranged by using a guard variable. |
| 683 | |
| 684 | If there are no initializers at all (and also no imported modules) we reduce |
| 685 | this to an empty function (since the Itanium ABI requires that this function |
| 686 | be available to a caller, which might be produced by a different |
| 687 | implementation). |
| 688 | |
| 689 | First we call any initializers for imported modules. |
| 690 | We then call initializers for the Global Module Fragment (if present) |
| 691 | We then call initializers for the current module. |
| 692 | We then call initializers for the Private Module Fragment (if present) |
| 693 | */ |
| 694 | |
| 695 | void CodeGenModule::EmitCXXModuleInitFunc(Module *Primary) { |
| 696 | assert(Primary->isInterfaceOrPartition() && |
| 697 | "The function should only be called for C++20 named module interface" |
| 698 | " or partition." ); |
| 699 | |
| 700 | while (!CXXGlobalInits.empty() && !CXXGlobalInits.back()) |
| 701 | CXXGlobalInits.pop_back(); |
| 702 | |
| 703 | // As noted above, we create the function, even if it is empty. |
| 704 | // Module initializers for imported modules are emitted first. |
| 705 | |
| 706 | // Collect all the modules that we import |
| 707 | llvm::SmallSetVector<Module *, 8> AllImports; |
| 708 | // Ones that we export |
| 709 | for (auto I : Primary->Exports) |
| 710 | AllImports.insert(X: I.getPointer()); |
| 711 | // Ones that we only import. |
| 712 | AllImports.insert_range(R&: Primary->Imports); |
| 713 | // Ones that we import in the global module fragment or the private module |
| 714 | // fragment. |
| 715 | for (Module *SubM : Primary->submodules()) { |
| 716 | assert((SubM->isGlobalModule() || SubM->isPrivateModule()) && |
| 717 | "The sub modules of C++20 module unit should only be global module " |
| 718 | "fragments or private module framents." ); |
| 719 | assert(SubM->Exports.empty() && |
| 720 | "The global mdoule fragments and the private module fragments are " |
| 721 | "not allowed to export import modules." ); |
| 722 | AllImports.insert_range(R&: SubM->Imports); |
| 723 | } |
| 724 | |
| 725 | SmallVector<llvm::Function *, 8> ModuleInits; |
| 726 | for (Module *M : AllImports) { |
| 727 | // No Itanium initializer in header like modules. |
| 728 | if (M->isHeaderLikeModule()) |
| 729 | continue; // TODO: warn of mixed use of module map modules and C++20? |
| 730 | // We're allowed to skip the initialization if we are sure it doesn't |
| 731 | // do any thing. |
| 732 | if (!M->isNamedModuleInterfaceHasInit()) |
| 733 | continue; |
| 734 | llvm::FunctionType *FTy = llvm::FunctionType::get(Result: VoidTy, isVarArg: false); |
| 735 | SmallString<256> FnName; |
| 736 | { |
| 737 | llvm::raw_svector_ostream Out(FnName); |
| 738 | cast<ItaniumMangleContext>(Val&: getCXXABI().getMangleContext()) |
| 739 | .mangleModuleInitializer(Module: M, Out); |
| 740 | } |
| 741 | assert(!GetGlobalValue(FnName.str()) && |
| 742 | "We should only have one use of the initializer call" ); |
| 743 | llvm::Function *Fn = llvm::Function::Create( |
| 744 | Ty: FTy, Linkage: llvm::Function::ExternalLinkage, N: FnName.str(), M: &getModule()); |
| 745 | ModuleInits.push_back(Elt: Fn); |
| 746 | } |
| 747 | |
| 748 | // Add any initializers with specified priority; this uses the same approach |
| 749 | // as EmitCXXGlobalInitFunc(). |
| 750 | if (!PrioritizedCXXGlobalInits.empty()) { |
| 751 | llvm::array_pod_sort(Start: PrioritizedCXXGlobalInits.begin(), |
| 752 | End: PrioritizedCXXGlobalInits.end()); |
| 753 | for (SmallVectorImpl<GlobalInitData>::iterator |
| 754 | I = PrioritizedCXXGlobalInits.begin(), |
| 755 | E = PrioritizedCXXGlobalInits.end(); |
| 756 | I != E;) { |
| 757 | SmallVectorImpl<GlobalInitData>::iterator PrioE = |
| 758 | std::upper_bound(first: I + 1, last: E, val: *I, comp: GlobalInitPriorityCmp()); |
| 759 | |
| 760 | for (; I < PrioE; ++I) |
| 761 | ModuleInits.push_back(Elt: I->second); |
| 762 | } |
| 763 | } |
| 764 | |
| 765 | // Now append the ones without specified priority. |
| 766 | for (auto *F : CXXGlobalInits) |
| 767 | ModuleInits.push_back(Elt: F); |
| 768 | |
| 769 | llvm::FunctionType *FTy = llvm::FunctionType::get(Result: VoidTy, isVarArg: false); |
| 770 | const CGFunctionInfo &FI = getTypes().arrangeNullaryFunction(); |
| 771 | |
| 772 | // We now build the initializer for this module, which has a mangled name |
| 773 | // as per the Itanium ABI . The action of the initializer is guarded so that |
| 774 | // each init is run just once (even though a module might be imported |
| 775 | // multiple times via nested use). |
| 776 | llvm::Function *Fn; |
| 777 | { |
| 778 | SmallString<256> InitFnName; |
| 779 | llvm::raw_svector_ostream Out(InitFnName); |
| 780 | cast<ItaniumMangleContext>(Val&: getCXXABI().getMangleContext()) |
| 781 | .mangleModuleInitializer(Module: Primary, Out); |
| 782 | Fn = CreateGlobalInitOrCleanUpFunction( |
| 783 | FTy, Name: llvm::Twine(InitFnName), FI, Loc: SourceLocation(), TLS: false, |
| 784 | Linkage: llvm::GlobalVariable::ExternalLinkage); |
| 785 | |
| 786 | // If we have a completely empty initializer then we do not want to create |
| 787 | // the guard variable. |
| 788 | ConstantAddress GuardAddr = ConstantAddress::invalid(); |
| 789 | if (!ModuleInits.empty()) { |
| 790 | // Create the guard var. |
| 791 | llvm::GlobalVariable *Guard = new llvm::GlobalVariable( |
| 792 | getModule(), Int8Ty, /*isConstant=*/false, |
| 793 | llvm::GlobalVariable::InternalLinkage, |
| 794 | llvm::ConstantInt::get(Ty: Int8Ty, V: 0), InitFnName.str() + "__in_chrg" ); |
| 795 | CharUnits GuardAlign = CharUnits::One(); |
| 796 | Guard->setAlignment(GuardAlign.getAsAlign()); |
| 797 | GuardAddr = ConstantAddress(Guard, Int8Ty, GuardAlign); |
| 798 | } |
| 799 | CodeGenFunction(*this).GenerateCXXGlobalInitFunc(Fn, CXXThreadLocals: ModuleInits, |
| 800 | Guard: GuardAddr); |
| 801 | } |
| 802 | |
| 803 | // We allow for the case that a module object is added to a linked binary |
| 804 | // without a specific call to the the initializer. This also ensures that |
| 805 | // implementation partition initializers are called when the partition |
| 806 | // is not imported as an interface. |
| 807 | AddGlobalCtor(Ctor: Fn); |
| 808 | |
| 809 | // See the comment in EmitCXXGlobalInitFunc about OpenCL global init |
| 810 | // functions. |
| 811 | if (getLangOpts().OpenCL) { |
| 812 | GenKernelArgMetadata(FN: Fn); |
| 813 | Fn->setCallingConv(llvm::CallingConv::SPIR_KERNEL); |
| 814 | } |
| 815 | |
| 816 | assert(!getLangOpts().CUDA || !getLangOpts().CUDAIsDevice || |
| 817 | getLangOpts().GPUAllowDeviceInit); |
| 818 | if (getLangOpts().HIP && getLangOpts().CUDAIsDevice) { |
| 819 | if (getTriple().isSPIRV()) |
| 820 | Fn->setCallingConv(llvm::CallingConv::SPIR_KERNEL); |
| 821 | else |
| 822 | Fn->setCallingConv(llvm::CallingConv::AMDGPU_KERNEL); |
| 823 | Fn->addFnAttr(Kind: "device-init" ); |
| 824 | } |
| 825 | |
| 826 | // We are done with the inits. |
| 827 | AllImports.clear(); |
| 828 | PrioritizedCXXGlobalInits.clear(); |
| 829 | CXXGlobalInits.clear(); |
| 830 | ModuleInits.clear(); |
| 831 | } |
| 832 | |
| 833 | static SmallString<128> getTransformedFileName(llvm::Module &M) { |
| 834 | SmallString<128> FileName = llvm::sys::path::filename(path: M.getName()); |
| 835 | |
| 836 | if (FileName.empty()) |
| 837 | FileName = "<null>" ; |
| 838 | |
| 839 | for (size_t i = 0; i < FileName.size(); ++i) { |
| 840 | // Replace everything that's not [a-zA-Z0-9._] with a _. This set happens |
| 841 | // to be the set of C preprocessing numbers. |
| 842 | if (!isPreprocessingNumberBody(c: FileName[i])) |
| 843 | FileName[i] = '_'; |
| 844 | } |
| 845 | |
| 846 | return FileName; |
| 847 | } |
| 848 | |
| 849 | static std::string getPrioritySuffix(unsigned int Priority) { |
| 850 | assert(Priority <= 65535 && "Priority should always be <= 65535." ); |
| 851 | |
| 852 | // Compute the function suffix from priority. Prepend with zeroes to make |
| 853 | // sure the function names are also ordered as priorities. |
| 854 | std::string PrioritySuffix = llvm::utostr(X: Priority); |
| 855 | PrioritySuffix = std::string(6 - PrioritySuffix.size(), '0') + PrioritySuffix; |
| 856 | |
| 857 | return PrioritySuffix; |
| 858 | } |
| 859 | |
| 860 | void |
| 861 | CodeGenModule::EmitCXXGlobalInitFunc() { |
| 862 | while (!CXXGlobalInits.empty() && !CXXGlobalInits.back()) |
| 863 | CXXGlobalInits.pop_back(); |
| 864 | |
| 865 | // When we import C++20 modules, we must run their initializers first. |
| 866 | SmallVector<llvm::Function *, 8> ModuleInits; |
| 867 | if (CXX20ModuleInits) |
| 868 | for (Module *M : ImportedModules) { |
| 869 | // No Itanium initializer in header like modules. |
| 870 | if (M->isHeaderLikeModule()) |
| 871 | continue; |
| 872 | // We're allowed to skip the initialization if we are sure it doesn't |
| 873 | // do any thing. |
| 874 | if (!M->isNamedModuleInterfaceHasInit()) |
| 875 | continue; |
| 876 | llvm::FunctionType *FTy = llvm::FunctionType::get(Result: VoidTy, isVarArg: false); |
| 877 | SmallString<256> FnName; |
| 878 | { |
| 879 | llvm::raw_svector_ostream Out(FnName); |
| 880 | cast<ItaniumMangleContext>(Val&: getCXXABI().getMangleContext()) |
| 881 | .mangleModuleInitializer(Module: M, Out); |
| 882 | } |
| 883 | assert(!GetGlobalValue(FnName.str()) && |
| 884 | "We should only have one use of the initializer call" ); |
| 885 | llvm::Function *Fn = llvm::Function::Create( |
| 886 | Ty: FTy, Linkage: llvm::Function::ExternalLinkage, N: FnName.str(), M: &getModule()); |
| 887 | ModuleInits.push_back(Elt: Fn); |
| 888 | } |
| 889 | |
| 890 | if (ModuleInits.empty() && CXXGlobalInits.empty() && |
| 891 | PrioritizedCXXGlobalInits.empty()) |
| 892 | return; |
| 893 | |
| 894 | llvm::FunctionType *FTy = llvm::FunctionType::get(Result: VoidTy, isVarArg: false); |
| 895 | const CGFunctionInfo &FI = getTypes().arrangeNullaryFunction(); |
| 896 | |
| 897 | // Create our global prioritized initialization function. |
| 898 | if (!PrioritizedCXXGlobalInits.empty()) { |
| 899 | SmallVector<llvm::Function *, 8> LocalCXXGlobalInits; |
| 900 | llvm::array_pod_sort(Start: PrioritizedCXXGlobalInits.begin(), |
| 901 | End: PrioritizedCXXGlobalInits.end()); |
| 902 | // Iterate over "chunks" of ctors with same priority and emit each chunk |
| 903 | // into separate function. Note - everything is sorted first by priority, |
| 904 | // second - by lex order, so we emit ctor functions in proper order. |
| 905 | for (SmallVectorImpl<GlobalInitData >::iterator |
| 906 | I = PrioritizedCXXGlobalInits.begin(), |
| 907 | E = PrioritizedCXXGlobalInits.end(); I != E; ) { |
| 908 | SmallVectorImpl<GlobalInitData >::iterator |
| 909 | PrioE = std::upper_bound(first: I + 1, last: E, val: *I, comp: GlobalInitPriorityCmp()); |
| 910 | |
| 911 | LocalCXXGlobalInits.clear(); |
| 912 | |
| 913 | unsigned int Priority = I->first.priority; |
| 914 | llvm::Function *Fn = CreateGlobalInitOrCleanUpFunction( |
| 915 | FTy, Name: "_GLOBAL__I_" + getPrioritySuffix(Priority), FI); |
| 916 | |
| 917 | // Prepend the module inits to the highest priority set. |
| 918 | if (!ModuleInits.empty()) { |
| 919 | for (auto *F : ModuleInits) |
| 920 | LocalCXXGlobalInits.push_back(Elt: F); |
| 921 | ModuleInits.clear(); |
| 922 | } |
| 923 | |
| 924 | for (; I < PrioE; ++I) |
| 925 | LocalCXXGlobalInits.push_back(Elt: I->second); |
| 926 | |
| 927 | CodeGenFunction(*this).GenerateCXXGlobalInitFunc(Fn, CXXThreadLocals: LocalCXXGlobalInits); |
| 928 | AddGlobalCtor(Ctor: Fn, Priority); |
| 929 | } |
| 930 | PrioritizedCXXGlobalInits.clear(); |
| 931 | } |
| 932 | |
| 933 | if (getCXXABI().useSinitAndSterm() && ModuleInits.empty() && |
| 934 | CXXGlobalInits.empty()) |
| 935 | return; |
| 936 | |
| 937 | for (auto *F : CXXGlobalInits) |
| 938 | ModuleInits.push_back(Elt: F); |
| 939 | CXXGlobalInits.clear(); |
| 940 | |
| 941 | // Include the filename in the symbol name. Including "sub_" matches gcc |
| 942 | // and makes sure these symbols appear lexicographically behind the symbols |
| 943 | // with priority emitted above. Module implementation units behave the same |
| 944 | // way as a non-modular TU with imports. |
| 945 | llvm::Function *Fn; |
| 946 | if (CXX20ModuleInits && getContext().getCurrentNamedModule() && |
| 947 | !getContext().getCurrentNamedModule()->isModuleImplementation()) { |
| 948 | SmallString<256> InitFnName; |
| 949 | llvm::raw_svector_ostream Out(InitFnName); |
| 950 | cast<ItaniumMangleContext>(Val&: getCXXABI().getMangleContext()) |
| 951 | .mangleModuleInitializer(Module: getContext().getCurrentNamedModule(), Out); |
| 952 | Fn = CreateGlobalInitOrCleanUpFunction( |
| 953 | FTy, Name: llvm::Twine(InitFnName), FI, Loc: SourceLocation(), TLS: false, |
| 954 | Linkage: llvm::GlobalVariable::ExternalLinkage); |
| 955 | } else |
| 956 | Fn = CreateGlobalInitOrCleanUpFunction( |
| 957 | FTy, |
| 958 | Name: llvm::Twine("_GLOBAL__sub_I_" , getTransformedFileName(M&: getModule())), |
| 959 | FI); |
| 960 | |
| 961 | CodeGenFunction(*this).GenerateCXXGlobalInitFunc(Fn, CXXThreadLocals: ModuleInits); |
| 962 | AddGlobalCtor(Ctor: Fn); |
| 963 | |
| 964 | // In OpenCL global init functions must be converted to kernels in order to |
| 965 | // be able to launch them from the host. |
| 966 | // FIXME: Some more work might be needed to handle destructors correctly. |
| 967 | // Current initialization function makes use of function pointers callbacks. |
| 968 | // We can't support function pointers especially between host and device. |
| 969 | // However it seems global destruction has little meaning without any |
| 970 | // dynamic resource allocation on the device and program scope variables are |
| 971 | // destroyed by the runtime when program is released. |
| 972 | if (getLangOpts().OpenCL) { |
| 973 | GenKernelArgMetadata(FN: Fn); |
| 974 | Fn->setCallingConv(llvm::CallingConv::SPIR_KERNEL); |
| 975 | } |
| 976 | |
| 977 | assert(!getLangOpts().CUDA || !getLangOpts().CUDAIsDevice || |
| 978 | getLangOpts().GPUAllowDeviceInit); |
| 979 | if (getLangOpts().HIP && getLangOpts().CUDAIsDevice) { |
| 980 | if (getTriple().isSPIRV()) |
| 981 | Fn->setCallingConv(llvm::CallingConv::SPIR_KERNEL); |
| 982 | else |
| 983 | Fn->setCallingConv(llvm::CallingConv::AMDGPU_KERNEL); |
| 984 | Fn->addFnAttr(Kind: "device-init" ); |
| 985 | } |
| 986 | |
| 987 | ModuleInits.clear(); |
| 988 | } |
| 989 | |
| 990 | void CodeGenModule::EmitCXXGlobalCleanUpFunc() { |
| 991 | if (CXXGlobalDtorsOrStermFinalizers.empty() && |
| 992 | PrioritizedCXXStermFinalizers.empty()) |
| 993 | return; |
| 994 | |
| 995 | llvm::FunctionType *FTy = llvm::FunctionType::get(Result: VoidTy, isVarArg: false); |
| 996 | const CGFunctionInfo &FI = getTypes().arrangeNullaryFunction(); |
| 997 | |
| 998 | // Create our global prioritized cleanup function. |
| 999 | if (!PrioritizedCXXStermFinalizers.empty()) { |
| 1000 | SmallVector<CXXGlobalDtorsOrStermFinalizer_t, 8> LocalCXXStermFinalizers; |
| 1001 | llvm::array_pod_sort(Start: PrioritizedCXXStermFinalizers.begin(), |
| 1002 | End: PrioritizedCXXStermFinalizers.end()); |
| 1003 | // Iterate over "chunks" of dtors with same priority and emit each chunk |
| 1004 | // into separate function. Note - everything is sorted first by priority, |
| 1005 | // second - by lex order, so we emit dtor functions in proper order. |
| 1006 | for (SmallVectorImpl<StermFinalizerData>::iterator |
| 1007 | I = PrioritizedCXXStermFinalizers.begin(), |
| 1008 | E = PrioritizedCXXStermFinalizers.end(); |
| 1009 | I != E;) { |
| 1010 | SmallVectorImpl<StermFinalizerData>::iterator PrioE = |
| 1011 | std::upper_bound(first: I + 1, last: E, val: *I, comp: StermFinalizerPriorityCmp()); |
| 1012 | |
| 1013 | LocalCXXStermFinalizers.clear(); |
| 1014 | |
| 1015 | unsigned int Priority = I->first.priority; |
| 1016 | llvm::Function *Fn = CreateGlobalInitOrCleanUpFunction( |
| 1017 | FTy, Name: "_GLOBAL__a_" + getPrioritySuffix(Priority), FI); |
| 1018 | |
| 1019 | for (; I < PrioE; ++I) { |
| 1020 | llvm::FunctionCallee DtorFn = I->second; |
| 1021 | LocalCXXStermFinalizers.emplace_back(Args: DtorFn.getFunctionType(), |
| 1022 | Args: DtorFn.getCallee(), Args: nullptr); |
| 1023 | } |
| 1024 | |
| 1025 | CodeGenFunction(*this).GenerateCXXGlobalCleanUpFunc( |
| 1026 | Fn, DtorsOrStermFinalizers: LocalCXXStermFinalizers); |
| 1027 | AddGlobalDtor(Dtor: Fn, Priority); |
| 1028 | } |
| 1029 | PrioritizedCXXStermFinalizers.clear(); |
| 1030 | } |
| 1031 | |
| 1032 | if (CXXGlobalDtorsOrStermFinalizers.empty()) |
| 1033 | return; |
| 1034 | |
| 1035 | // Create our global cleanup function. |
| 1036 | llvm::Function *Fn = |
| 1037 | CreateGlobalInitOrCleanUpFunction(FTy, Name: "_GLOBAL__D_a" , FI); |
| 1038 | |
| 1039 | CodeGenFunction(*this).GenerateCXXGlobalCleanUpFunc( |
| 1040 | Fn, DtorsOrStermFinalizers: CXXGlobalDtorsOrStermFinalizers); |
| 1041 | AddGlobalDtor(Dtor: Fn); |
| 1042 | CXXGlobalDtorsOrStermFinalizers.clear(); |
| 1043 | } |
| 1044 | |
| 1045 | /// Emit the code necessary to initialize the given global variable. |
| 1046 | void CodeGenFunction::GenerateCXXGlobalVarDeclInitFunc(llvm::Function *Fn, |
| 1047 | const VarDecl *D, |
| 1048 | llvm::GlobalVariable *Addr, |
| 1049 | bool PerformInit) { |
| 1050 | // Check if we need to emit debug info for variable initializer. |
| 1051 | if (D->hasAttr<NoDebugAttr>()) |
| 1052 | DebugInfo = nullptr; // disable debug info indefinitely for this function |
| 1053 | |
| 1054 | CurEHLocation = D->getBeginLoc(); |
| 1055 | |
| 1056 | StartFunction(GD: GlobalDecl(D, DynamicInitKind::Initializer), |
| 1057 | RetTy: getContext().VoidTy, Fn, FnInfo: getTypes().arrangeNullaryFunction(), |
| 1058 | Args: FunctionArgList()); |
| 1059 | // Emit an artificial location for this function. |
| 1060 | auto AL = ApplyDebugLocation::CreateArtificial(CGF&: *this); |
| 1061 | |
| 1062 | // Use guarded initialization if the global variable is weak. This |
| 1063 | // occurs for, e.g., instantiated static data members and |
| 1064 | // definitions explicitly marked weak. |
| 1065 | // |
| 1066 | // Also use guarded initialization for a variable with dynamic TLS and |
| 1067 | // unordered initialization. (If the initialization is ordered, the ABI |
| 1068 | // layer will guard the whole-TU initialization for us.) |
| 1069 | if (Addr->hasWeakLinkage() || Addr->hasLinkOnceLinkage() || |
| 1070 | (D->getTLSKind() == VarDecl::TLS_Dynamic && |
| 1071 | isTemplateInstantiation(Kind: D->getTemplateSpecializationKind()))) { |
| 1072 | EmitCXXGuardedInit(D: *D, DeclPtr: Addr, PerformInit); |
| 1073 | } else { |
| 1074 | EmitCXXGlobalVarDeclInit(D: *D, GV: Addr, PerformInit); |
| 1075 | } |
| 1076 | |
| 1077 | FinishFunction(); |
| 1078 | } |
| 1079 | |
| 1080 | void |
| 1081 | CodeGenFunction::GenerateCXXGlobalInitFunc(llvm::Function *Fn, |
| 1082 | ArrayRef<llvm::Function *> Decls, |
| 1083 | ConstantAddress Guard) { |
| 1084 | { |
| 1085 | auto NL = ApplyDebugLocation::CreateEmpty(CGF&: *this); |
| 1086 | StartFunction(GD: GlobalDecl(), RetTy: getContext().VoidTy, Fn, |
| 1087 | FnInfo: getTypes().arrangeNullaryFunction(), Args: FunctionArgList()); |
| 1088 | // Emit an artificial location for this function. |
| 1089 | auto AL = ApplyDebugLocation::CreateArtificial(CGF&: *this); |
| 1090 | |
| 1091 | llvm::BasicBlock *ExitBlock = nullptr; |
| 1092 | if (Guard.isValid()) { |
| 1093 | // If we have a guard variable, check whether we've already performed |
| 1094 | // these initializations. This happens for TLS initialization functions. |
| 1095 | llvm::Value *GuardVal = Builder.CreateLoad(Addr: Guard); |
| 1096 | llvm::Value *Uninit = Builder.CreateIsNull(Arg: GuardVal, |
| 1097 | Name: "guard.uninitialized" ); |
| 1098 | llvm::BasicBlock *InitBlock = createBasicBlock(name: "init" ); |
| 1099 | ExitBlock = createBasicBlock(name: "exit" ); |
| 1100 | EmitCXXGuardedInitBranch(NeedsInit: Uninit, InitBlock, NoInitBlock: ExitBlock, |
| 1101 | Kind: GuardKind::TlsGuard, D: nullptr); |
| 1102 | EmitBlock(BB: InitBlock); |
| 1103 | // Mark as initialized before initializing anything else. If the |
| 1104 | // initializers use previously-initialized thread_local vars, that's |
| 1105 | // probably supposed to be OK, but the standard doesn't say. |
| 1106 | Builder.CreateStore(Val: llvm::ConstantInt::get(Ty: GuardVal->getType(),V: 1), Addr: Guard); |
| 1107 | |
| 1108 | // The guard variable can't ever change again. |
| 1109 | EmitInvariantStart( |
| 1110 | Addr: Guard.getPointer(), |
| 1111 | Size: CharUnits::fromQuantity( |
| 1112 | Quantity: CGM.getDataLayout().getTypeAllocSize(Ty: GuardVal->getType()))); |
| 1113 | } |
| 1114 | |
| 1115 | RunCleanupsScope Scope(*this); |
| 1116 | |
| 1117 | // When building in Objective-C++ ARC mode, create an autorelease pool |
| 1118 | // around the global initializers. |
| 1119 | if (getLangOpts().ObjCAutoRefCount && getLangOpts().CPlusPlus) { |
| 1120 | llvm::Value *token = EmitObjCAutoreleasePoolPush(); |
| 1121 | EmitObjCAutoreleasePoolCleanup(Ptr: token); |
| 1122 | } |
| 1123 | |
| 1124 | for (llvm::Function *Decl : Decls) |
| 1125 | if (Decl) |
| 1126 | EmitRuntimeCall(callee: Decl); |
| 1127 | |
| 1128 | Scope.ForceCleanup(); |
| 1129 | |
| 1130 | if (ExitBlock) { |
| 1131 | Builder.CreateBr(Dest: ExitBlock); |
| 1132 | EmitBlock(BB: ExitBlock); |
| 1133 | } |
| 1134 | } |
| 1135 | |
| 1136 | FinishFunction(); |
| 1137 | } |
| 1138 | |
| 1139 | void CodeGenFunction::GenerateCXXGlobalCleanUpFunc( |
| 1140 | llvm::Function *Fn, |
| 1141 | ArrayRef<std::tuple<llvm::FunctionType *, llvm::WeakTrackingVH, |
| 1142 | llvm::Constant *>> |
| 1143 | DtorsOrStermFinalizers) { |
| 1144 | { |
| 1145 | auto NL = ApplyDebugLocation::CreateEmpty(CGF&: *this); |
| 1146 | StartFunction(GD: GlobalDecl(), RetTy: getContext().VoidTy, Fn, |
| 1147 | FnInfo: getTypes().arrangeNullaryFunction(), Args: FunctionArgList()); |
| 1148 | // Emit an artificial location for this function. |
| 1149 | auto AL = ApplyDebugLocation::CreateArtificial(CGF&: *this); |
| 1150 | |
| 1151 | // Emit the cleanups, in reverse order from construction. |
| 1152 | for (unsigned i = 0, e = DtorsOrStermFinalizers.size(); i != e; ++i) { |
| 1153 | llvm::FunctionType *CalleeTy; |
| 1154 | llvm::Value *Callee; |
| 1155 | llvm::Constant *Arg; |
| 1156 | std::tie(args&: CalleeTy, args&: Callee, args&: Arg) = DtorsOrStermFinalizers[e - i - 1]; |
| 1157 | |
| 1158 | llvm::CallBase *CI = nullptr; |
| 1159 | if (Arg == nullptr) { |
| 1160 | assert( |
| 1161 | CGM.getCXXABI().useSinitAndSterm() && |
| 1162 | "Arg could not be nullptr unless using sinit and sterm functions." ); |
| 1163 | CI = Builder.CreateCall(FTy: CalleeTy, Callee); |
| 1164 | } else { |
| 1165 | // If the object lives in a different address space, the `this` pointer |
| 1166 | // address space won't match the dtor `this` param. An addrspacecast is |
| 1167 | // required. |
| 1168 | assert(Arg->getType()->isPointerTy()); |
| 1169 | assert(CalleeTy->getParamType(0)->isPointerTy()); |
| 1170 | unsigned ActualAddrSpace = Arg->getType()->getPointerAddressSpace(); |
| 1171 | unsigned ExpectedAddrSpace = |
| 1172 | CalleeTy->getParamType(i: 0)->getPointerAddressSpace(); |
| 1173 | if (ActualAddrSpace != ExpectedAddrSpace) { |
| 1174 | llvm::PointerType *PTy = |
| 1175 | llvm::PointerType::get(C&: getLLVMContext(), AddressSpace: ExpectedAddrSpace); |
| 1176 | Arg = llvm::ConstantExpr::getAddrSpaceCast(C: Arg, Ty: PTy); |
| 1177 | } |
| 1178 | CI = Builder.CreateCall(FTy: CalleeTy, Callee, Args: Arg); |
| 1179 | } |
| 1180 | |
| 1181 | // Make sure the call and the callee agree on calling convention. |
| 1182 | if (llvm::Function *F = dyn_cast<llvm::Function>(Val: Callee)) |
| 1183 | CI->setCallingConv(F->getCallingConv()); |
| 1184 | |
| 1185 | if (CGM.shouldEmitConvergenceTokens() && CI->isConvergent()) |
| 1186 | CI = addConvergenceControlToken(Input: CI); |
| 1187 | } |
| 1188 | } |
| 1189 | |
| 1190 | FinishFunction(); |
| 1191 | } |
| 1192 | |
| 1193 | /// generateDestroyHelper - Generates a helper function which, when |
| 1194 | /// invoked, destroys the given object. The address of the object |
| 1195 | /// should be in global memory. |
| 1196 | llvm::Function *CodeGenFunction::generateDestroyHelper( |
| 1197 | Address addr, QualType type, Destroyer *destroyer, |
| 1198 | bool useEHCleanupForArray, const VarDecl *VD) { |
| 1199 | FunctionArgList args; |
| 1200 | ImplicitParamDecl Dst(getContext(), getContext().VoidPtrTy, |
| 1201 | ImplicitParamKind::Other); |
| 1202 | args.push_back(Elt: &Dst); |
| 1203 | |
| 1204 | const CGFunctionInfo &FI = |
| 1205 | CGM.getTypes().arrangeBuiltinFunctionDeclaration(resultType: getContext().VoidTy, args); |
| 1206 | llvm::FunctionType *FTy = CGM.getTypes().GetFunctionType(Info: FI); |
| 1207 | llvm::Function *fn = CGM.CreateGlobalInitOrCleanUpFunction( |
| 1208 | FTy, Name: "__cxx_global_array_dtor" , FI, Loc: VD->getLocation()); |
| 1209 | |
| 1210 | CurEHLocation = VD->getBeginLoc(); |
| 1211 | |
| 1212 | StartFunction(GD: GlobalDecl(VD, DynamicInitKind::GlobalArrayDestructor), |
| 1213 | RetTy: getContext().VoidTy, Fn: fn, FnInfo: FI, Args: args); |
| 1214 | // Emit an artificial location for this function. |
| 1215 | auto AL = ApplyDebugLocation::CreateArtificial(CGF&: *this); |
| 1216 | |
| 1217 | emitDestroy(addr, type, destroyer, useEHCleanupForArray); |
| 1218 | |
| 1219 | FinishFunction(); |
| 1220 | |
| 1221 | return fn; |
| 1222 | } |
| 1223 | |