1 //===--- SemaCUDA.cpp - Semantic Analysis for CUDA constructs -------------===//
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
7 //===----------------------------------------------------------------------===//
11 //===----------------------------------------------------------------------===//
36   if (auto *A = D->getAttr<AttrT>())  in hasExplicitAttr()
37     return !A->isImplicit();  in hasExplicitAttr()
50   ForceHostDeviceDepth--;  in PopForceHostDevice()
61   QualType ConfigQTy = ConfigDecl->getType();  in ActOnExecConfigExpr()
105     return CUDAFunctionTarget::Device;  in IdentifyTarget()
112   return D->hasAttrs() && llvm::any_of(D->getAttrs(), [&](Attr *Attribute) {  in hasAttr()
114                   !(IgnoreImplicitAttr && Attribute->isImplicit());  in hasAttr()
124   if (VD && VD->hasGlobalStorage() && !VD->isStaticLocal()) {  in CUDATargetContextRAII()
130       Target = CUDAFunctionTarget::Device;  in CUDATargetContextRAII()
135 /// IdentifyTarget - Determine the CUDA compilation target for this function
142   if (D->hasAttr<CUDAInvalidTargetAttr>())  in IdentifyTarget()
145   if (D->hasAttr<CUDAGlobalAttr>())  in IdentifyTarget()
151     return CUDAFunctionTarget::Device;  in IdentifyTarget()
154   } else if ((D->isImplicit() || !D->isUserProvided()) &&  in IdentifyTarget()
164 /// IdentifyTarget - Determine the CUDA compilation target for this variable.
166   if (Var->hasAttr<HIPManagedAttr>())  in IdentifyTarget()
169   // are emitted on both sides. Such variables are promoted to device side  in IdentifyTarget()
170   // only if they have static constant intializers on device side.  in IdentifyTarget()
171   if ((Var->isConstexpr() || Var->getType().isConstQualified()) &&  in IdentifyTarget()
172       Var->hasAttr<CUDAConstantAttr>() &&  in IdentifyTarget()
175   if (Var->hasAttr<CUDADeviceAttr>() || Var->hasAttr<CUDAConstantAttr>() ||  in IdentifyTarget()
176       Var->hasAttr<CUDASharedAttr>() ||  in IdentifyTarget()
177       Var->getType()->isCUDADeviceBuiltinSurfaceType() ||  in IdentifyTarget()
178       Var->getType()->isCUDADeviceBuiltinTextureType())  in IdentifyTarget()
180   // Function-scope static variable without explicit device or constant  in IdentifyTarget()
182   //  - on both sides in host device functions  in IdentifyTarget()
183   //  - on device side in device or global functions  in IdentifyTarget()
184   if (auto *FD = dyn_cast<FunctionDecl>(Var->getDeclContext())) {  in IdentifyTarget()
188     case CUDAFunctionTarget::Device:  in IdentifyTarget()
200 // F - from,
201 // T - to
202 // Ph - preference in host mode
203 // Pd - preference in device mode
204 // H  - handled in (x)
205 // Preferences: N:native, SS:same side, HD:host-device, WS:wrong side, --:never.
208 // |----+----+-----+-----+-----+
210 // | d  | g  | --  | --  | (a) |
211 // | d  | h  | --  | --  | (e) |
214 // | g  | g  | --  | --  | (a) |
215 // | g  | h  | --  | --  | (e) |
217 // | h  | d  | --  | --  | (e) |
222 // | hd | g  | SS  | --  |(d/a)|
231   // Treat ctor/dtor as host device function in device var initializer to allow  in IdentifyPreference()
232   // trivial ctor/dtor without device attr to be used. Non-trivial ctor/dtor  in IdentifyPreference()
235       CurCUDATargetCtx.Target == CUDAFunctionTarget::Device &&  in IdentifyPreference()
252        CallerTarget == CUDAFunctionTarget::Device))  in IdentifyPreference()
264        CalleeTarget == CUDAFunctionTarget::Device))  in IdentifyPreference()
267   // HipStdPar mode is special, in that assessing whether a device side call to  in IdentifyPreference()
272        CallerTarget == CUDAFunctionTarget::Device ||  in IdentifyPreference()
279     // It's OK to call a compilation-mode matching function from an HD one.  in IdentifyPreference()
281          CalleeTarget == CUDAFunctionTarget::Device) ||  in IdentifyPreference()
287     // Calls from HD to non-mode-matching functions (i.e., to host functions  in IdentifyPreference()
288     // when compiling in device mode or to device functions when compiling in  in IdentifyPreference()
294   // (e) Calling across device/host boundary is not something you should do.  in IdentifyPreference()
296        CalleeTarget == CUDAFunctionTarget::Device) ||  in IdentifyPreference()
297       (CallerTarget == CUDAFunctionTarget::Device &&  in IdentifyPreference()
309   if (auto *A = D->getAttr<AttrT>())  in hasImplicitAttr()
310     return A->isImplicit();  in hasImplicitAttr()
311   return D->isImplicit();  in hasImplicitAttr()
343 /// When an implicitly-declared special member has to invoke more than one
378   // owning class, or the special member already has explicit device or host  in inferTargetForImplicitSpecialMember()
380   bool InClass = MemberDecl->getLexicalParent() == MemberDecl->getParent();  in inferTargetForImplicitSpecialMember()
381   bool HasH = MemberDecl->hasAttr<CUDAHostAttr>();  in inferTargetForImplicitSpecialMember()
382   bool HasD = MemberDecl->hasAttr<CUDADeviceAttr>();  in inferTargetForImplicitSpecialMember()
384       (HasD && !MemberDecl->getAttr<CUDADeviceAttr>()->isImplicit()) ||  in inferTargetForImplicitSpecialMember()
385       (HasH && !MemberDecl->getAttr<CUDAHostAttr>()->isImplicit());  in inferTargetForImplicitSpecialMember()
399   for (const auto &B : ClassDecl->bases()) {  in inferTargetForImplicitSpecialMember()
405   if (!ClassDecl->isAbstract()) {  in inferTargetForImplicitSpecialMember()
406     llvm::append_range(Bases, llvm::make_pointer_range(ClassDecl->vbases()));  in inferTargetForImplicitSpecialMember()
410     const RecordType *BaseType = B->getType()->getAs<RecordType>();  in inferTargetForImplicitSpecialMember()
415     CXXRecordDecl *BaseClassDecl = cast<CXXRecordDecl>(BaseType->getDecl());  in inferTargetForImplicitSpecialMember()
435           Diag(ClassDecl->getLocation(),  in inferTargetForImplicitSpecialMember()
440         MemberDecl->addAttr(  in inferTargetForImplicitSpecialMember()
448   for (const auto *F : ClassDecl->fields()) {  in inferTargetForImplicitSpecialMember()
449     if (F->isInvalidDecl()) {  in inferTargetForImplicitSpecialMember()
454         getASTContext().getBaseElementType(F->getType())->getAs<RecordType>();  in inferTargetForImplicitSpecialMember()
459     CXXRecordDecl *FieldRecDecl = cast<CXXRecordDecl>(FieldType->getDecl());  in inferTargetForImplicitSpecialMember()
462                                     /* ConstArg */ ConstRHS && !F->isMutable(),  in inferTargetForImplicitSpecialMember()
479           Diag(ClassDecl->getLocation(),  in inferTargetForImplicitSpecialMember()
484         MemberDecl->addAttr(  in inferTargetForImplicitSpecialMember()
496     if (*InferredTarget == CUDAFunctionTarget::Device)  in inferTargetForImplicitSpecialMember()
505     MemberDecl->addAttr(CUDADeviceAttr::CreateImplicit(getASTContext()));  in inferTargetForImplicitSpecialMember()
507     MemberDecl->addAttr(CUDAHostAttr::CreateImplicit(getASTContext()));  in inferTargetForImplicitSpecialMember()
513   if (!CD->isDefined() && CD->isTemplateInstantiation())  in isEmptyConstructor()
514     SemaRef.InstantiateFunctionDefinition(Loc, CD->getFirstDecl());  in isEmptyConstructor()
519   if (CD->isTrivial())  in isEmptyConstructor()
525   // and the function body is an empty compound statement.  in isEmptyConstructor()
526   if (!(CD->hasTrivialBody() && CD->getNumParams() == 0))  in isEmptyConstructor()
530   if (CD->getParent()->isDynamicClass())  in isEmptyConstructor()
534   if (CD->getParent()->isUnion())  in isEmptyConstructor()
539   if (!llvm::all_of(CD->inits(), [&](const CXXCtorInitializer *CI) {  in isEmptyConstructor()
541                 dyn_cast<CXXConstructExpr>(CI->getInit()))  in isEmptyConstructor()
542           return isEmptyConstructor(Loc, CE->getConstructor());  in isEmptyConstructor()
551   // No destructor -> no problem.  in isEmptyDestructor()
555   if (!DD->isDefined() && DD->isTemplateInstantiation())  in isEmptyDestructor()
556     SemaRef.InstantiateFunctionDefinition(Loc, DD->getFirstDecl());  in isEmptyDestructor()
561   if (DD->isTrivial())  in isEmptyDestructor()
566   // and the function body is an empty compound statement.  in isEmptyDestructor()
567   if (!DD->hasTrivialBody())  in isEmptyDestructor()
570   const CXXRecordDecl *ClassDecl = DD->getParent();  in isEmptyDestructor()
573   if (ClassDecl->isDynamicClass())  in isEmptyDestructor()
578   if (DD->getParent()->isUnion())  in isEmptyDestructor()
583   if (!llvm::all_of(ClassDecl->bases(), [&](const CXXBaseSpecifier &BS) {  in isEmptyDestructor()
584         if (CXXRecordDecl *RD = BS.getType()->getAsCXXRecordDecl())  in isEmptyDestructor()
585           return isEmptyDestructor(Loc, RD->getDestructor());  in isEmptyDestructor()
591   if (!llvm::all_of(ClassDecl->fields(), [&](const FieldDecl *Field) {  in isEmptyDestructor()
592         if (CXXRecordDecl *RD = Field->getType()  in isEmptyDestructor()
593                                     ->getBaseElementTypeUnsafe()  in isEmptyDestructor()
594                                     ->getAsCXXRecordDecl())  in isEmptyDestructor()
595           return isEmptyDestructor(Loc, RD->getDestructor());  in isEmptyDestructor()
605   CICK_DeviceOrConstant, // Check initializer for device/constant variable
610   if (VD->getType()->isDependentType())  in IsDependentVar()
612   if (const auto *Init = VD->getInit())  in IsDependentVar()
613     return Init->isValueDependent();  in IsDependentVar()
617 // Check whether a variable has an allowed initializer for a CUDA device side
619 // potential promotion to device side variable.
628   assert(!VD->isInvalidDecl() && VD->hasGlobalStorage());  in HasAllowedCUDADeviceStaticInitializer()
630   const Expr *Init = VD->getInit();  in HasAllowedCUDADeviceStaticInitializer()
635       return S.isEmptyConstructor(VD->getLocation(), CE->getConstructor());  in HasAllowedCUDADeviceStaticInitializer()
643     return Init->isConstantInitializer(S.getASTContext(),  in HasAllowedCUDADeviceStaticInitializer()
644                                        VD->getType()->isReferenceType());  in HasAllowedCUDADeviceStaticInitializer()
647     if (const auto *RD = VD->getType()->getAsCXXRecordDecl())  in HasAllowedCUDADeviceStaticInitializer()
648       return S.isEmptyDestructor(VD->getLocation(), RD->getDestructor());  in HasAllowedCUDADeviceStaticInitializer()
659   // Return early if VD is inside a non-instantiated template function since  in checkAllowedInitializer()
662           dyn_cast_or_null<FunctionDecl>(VD->getDeclContext()))  in checkAllowedInitializer()
663     if (FD->isDependentContext())  in checkAllowedInitializer()
668   if (VD->isInvalidDecl() || !VD->hasInit() || !VD->hasGlobalStorage() ||  in checkAllowedInitializer()
671   const Expr *Init = VD->getInit();  in checkAllowedInitializer()
672   bool IsSharedVar = VD->hasAttr<CUDASharedAttr>();  in checkAllowedInitializer()
675       (VD->hasAttr<CUDADeviceAttr>() || VD->hasAttr<CUDAConstantAttr>());  in checkAllowedInitializer()
680     Diag(VD->getLocation(),  in checkAllowedInitializer()
682         << Init->getSourceRange();  in checkAllowedInitializer()
683     VD->setInvalidDecl();  in checkAllowedInitializer()
685     // This is a host-side global variable.  Check that the initializer is  in checkAllowedInitializer()
689       InitFn = CE->getConstructor();  in checkAllowedInitializer()
691       InitFn = CE->getDirectCallee();  in checkAllowedInitializer()
697         Diag(VD->getLocation(), diag::err_ref_bad_target_global_initializer)  in checkAllowedInitializer()
699         Diag(InitFn->getLocation(), diag::note_previous_decl) << InitFn;  in checkAllowedInitializer()
700         VD->setInvalidDecl();  in checkAllowedInitializer()
717   // Record whether an implicit host device function is used on device side.  in RecordImplicitHostDeviceFuncUsedByDevice()
718   if (CallerTarget != CUDAFunctionTarget::Device &&  in RecordImplicitHostDeviceFuncUsedByDevice()
728 // With -fcuda-host-device-constexpr, an unattributed constexpr function is
730 //  * it is a variadic function (device-side variadic functions are not
745     if (!NewD->hasAttr<CUDAHostAttr>())  in maybeAddHostDeviceAttrs()
746       NewD->addAttr(CUDAHostAttr::CreateImplicit(getASTContext()));  in maybeAddHostDeviceAttrs()
747     if (!NewD->hasAttr<CUDADeviceAttr>())  in maybeAddHostDeviceAttrs()
748       NewD->addAttr(CUDADeviceAttr::CreateImplicit(getASTContext()));  in maybeAddHostDeviceAttrs()
752   // If a template function has no host/device/global attributes,  in maybeAddHostDeviceAttrs()
753   // make it implicitly host device function.  in maybeAddHostDeviceAttrs()
755       !NewD->hasAttr<CUDAHostAttr>() && !NewD->hasAttr<CUDADeviceAttr>() &&  in maybeAddHostDeviceAttrs()
756       !NewD->hasAttr<CUDAGlobalAttr>() &&  in maybeAddHostDeviceAttrs()
757       (NewD->getDescribedFunctionTemplate() ||  in maybeAddHostDeviceAttrs()
758        NewD->isFunctionTemplateSpecialization())) {  in maybeAddHostDeviceAttrs()
759     NewD->addAttr(CUDAHostAttr::CreateImplicit(getASTContext()));  in maybeAddHostDeviceAttrs()
760     NewD->addAttr(CUDADeviceAttr::CreateImplicit(getASTContext()));  in maybeAddHostDeviceAttrs()
764   if (!getLangOpts().CUDAHostDeviceConstexpr || !NewD->isConstexpr() ||  in maybeAddHostDeviceAttrs()
765       NewD->isVariadic() || NewD->hasAttr<CUDAHostAttr>() ||  in maybeAddHostDeviceAttrs()
766       NewD->hasAttr<CUDADeviceAttr>() || NewD->hasAttr<CUDAGlobalAttr>())  in maybeAddHostDeviceAttrs()
773       D = Using->getTargetDecl();  in maybeAddHostDeviceAttrs()
774     FunctionDecl *OldD = D->getAsFunction();  in maybeAddHostDeviceAttrs()
775     return OldD && OldD->hasAttr<CUDADeviceAttr>() &&  in maybeAddHostDeviceAttrs()
776            !OldD->hasAttr<CUDAHostAttr>() &&  in maybeAddHostDeviceAttrs()
786     // host+device.  in maybeAddHostDeviceAttrs()
788     if (!SemaRef.getSourceManager().isInSystemHeader(Match->getLocation())) {  in maybeAddHostDeviceAttrs()
789       Diag(NewD->getLocation(),  in maybeAddHostDeviceAttrs()
792       Diag(Match->getLocation(),  in maybeAddHostDeviceAttrs()
798   NewD->addAttr(CUDAHostAttr::CreateImplicit(getASTContext()));  in maybeAddHostDeviceAttrs()
799   NewD->addAttr(CUDADeviceAttr::CreateImplicit(getASTContext()));  in maybeAddHostDeviceAttrs()
808   if (getLangOpts().CUDAIsDevice && !VD->hasAttr<CUDAConstantAttr>() &&  in MaybeAddConstantAttr()
809       !VD->hasAttr<CUDASharedAttr>() &&  in MaybeAddConstantAttr()
810       (VD->isFileVarDecl() || VD->isStaticDataMember()) &&  in MaybeAddConstantAttr()
812       ((VD->isConstexpr() || VD->getType().isConstQualified()) &&  in MaybeAddConstantAttr()
815     VD->addAttr(CUDAConstantAttr::CreateImplicit(getASTContext()));  in MaybeAddConstantAttr()
829     case CUDAFunctionTarget::Device:  in DiagIfDeviceCode()
833       // device code if we're compiling for device.  Defer any errors in device  in DiagIfDeviceCode()
834       // mode until the function is known-emitted.  in DiagIfDeviceCode()
838           getDiagnostics().getDiagnosticIDs()->isBuiltinNote(DiagID))  in DiagIfDeviceCode()
864       // device code if we're compiling for device.  Defer any errors in device  in DiagIfHostCode()
865       // mode until the function is known-emitted.  in DiagIfHostCode()
869           getDiagnostics().getDiagnosticIDs()->isBuiltinNote(DiagID))  in DiagIfHostCode()
896   // If the caller is known-emitted, mark the callee as known-emitted.  in CheckCall()
905       assert(Caller && "Never/wrongSide calls require a non-null caller");  in CheckCall()
906       // If we know the caller will be emitted, we know this wrong-side call  in CheckCall()
918     // For -fgpu-rdc, keep track of external kernels used by host functions.  in CheckCall()
920         Callee->hasAttr<CUDAGlobalAttr>() && !Callee->isDefined() &&  in CheckCall()
921         (!Caller || (!Caller->getDescribedFunctionTemplate() &&  in CheckCall()
939   if (!Callee->getBuiltinID())  in CheckCall()
940     SemaDiagnosticBuilder(DiagKind, Callee->getLocation(),  in CheckCall()
947 // Check the wrong-sided reference capture of lambda for CUDA/HIP.
956   // if the lambda structure is populated by a device function or kernel then  in CheckLambdaCapture()
958   // since a device function or kernel can only call a device function, also a  in CheckLambdaCapture()
965   // File-scope lambda can only do init captures for global variables, which  in CheckLambdaCapture()
971   // In device compilation, we only need to check lambda functions which are  in CheckLambdaCapture()
972   // emitted on device side. For such lambdas, a reference capture is invalid  in CheckLambdaCapture()
974   // to and called in a device function or kernel.  in CheckLambdaCapture()
975   bool CalleeIsDevice = Callee->hasAttr<CUDADeviceAttr>();  in CheckLambdaCapture()
977       !Caller->hasAttr<CUDAGlobalAttr>() && !Caller->hasAttr<CUDADeviceAttr>();  in CheckLambdaCapture()
988     // managed memory which is accessible on both device and host sides. It only  in CheckLambdaCapture()
990     // accessible on device side.  in CheckLambdaCapture()
999   if (Method->hasAttr<CUDAHostAttr>() || Method->hasAttr<CUDADeviceAttr>())  in SetLambdaAttrs()
1001   Method->addAttr(CUDADeviceAttr::CreateImplicit(getASTContext()));  in SetLambdaAttrs()
1002   Method->addAttr(CUDAHostAttr::CreateImplicit(getASTContext()));  in SetLambdaAttrs()
1010     FunctionDecl *OldFD = OldND->getAsFunction();  in checkTargetOverload()
1017     // functions can have different implementations on the host and device, but  in checkTargetOverload()
1018     // HD/global functions "exist" in some sense on both the host and device, so  in checkTargetOverload()
1026              OldTarget == CUDAFunctionTarget::Device)) ||  in checkTargetOverload()
1030              NewTarget == CUDAFunctionTarget::Device)) ||  in checkTargetOverload()
1033         Diag(NewFD->getLocation(), diag::err_cuda_ovl_target)  in checkTargetOverload()
1034             << llvm::to_underlying(NewTarget) << NewFD->getDeclName()  in checkTargetOverload()
1036         Diag(OldFD->getLocation(), diag::note_previous_declaration);  in checkTargetOverload()
1037         NewFD->setInvalidDecl();  in checkTargetOverload()
1041            OldTarget == CUDAFunctionTarget::Device) ||  in checkTargetOverload()
1042           (NewTarget == CUDAFunctionTarget::Device &&  in checkTargetOverload()
1044         Diag(NewFD->getLocation(), diag::warn_offload_incompatible_redeclare)  in checkTargetOverload()
1046         Diag(OldFD->getLocation(), diag::note_previous_declaration);  in checkTargetOverload()
1056     AttrTy *Clone = Attribute->clone(S.Context);  in copyAttrIfPresent()
1057     Clone->setInherited(true);  in copyAttrIfPresent()
1058     FD->addAttr(Clone);  in copyAttrIfPresent()