[flang] CUDA Fortran - part 2/5: symbols & scopes

[klausler]

Summary

Add representations of CUDA Fortran data and subprogram attributes
to the symbol table and scopes of semantics. Set them in name
resolution, and emit them to module files.

Depends on https://reviews.llvm.org/D150159.

[klausler]

Remove redundantly added tests.

[clementval]

LGTM

[wangzpgi]

flang/lib/Semantics/resolve-names.cpp

3527

Can launch_bounds take third argument of max cluster rank for Hopper?

3552

Can cluster_dim take variable argument such as "cluster_dim(x,y,z)", where x, y, and z are integer variables?

[wangzpgi]

flang/include/flang/Semantics/symbol.h

152

misspelling: LANCH_BOUNDS -> LAUNCH_BOUNDS

[klausler]

flang/lib/Semantics/resolve-names.cpp

3527

Not according to https://docs.nvidia.com/hpc-sdk/compilers/cuda-fortran-prog-guide/index.html#cfref-new-subs-func-attr which has "The form used in CUDA Fortran is:

attributes(global) launch_bounds(maxTPB, minBPM) subroutine sub(...)

where maxTPB is the maxThreadsPerBlock, the maximum number of threads per block with which the application will ever launch, and minBPM is the desired minimum number of resident blocks per multiprocessor. Both values must be numeric constants." in 3.1.2.

3552

No, not according to https://docs.nvidia.com/hpc-sdk/compilers/cuda-fortran-prog-guide/index.html#cfref-new-subs-func-attr , which has "Beginning with the 23.3 release, support for thread block clusters is enabled for Hopper (cc90) and later targets. To specify the dimensions of the cluster, use the cluster_dims syntax and specify each x, y, and z dimension. Values must be numeric constants." in 3.1.3.

[klausler]

Fix misspelling in comment.

[klausler]

Generate diff against part 1 rather than main (oops).

[wangzpgi]

flang/lib/Semantics/resolve-names.cpp

3552

It mentioned at the end of 3.1.2: "Also, as part of the Hopper support, the launch_bounds syntax has been extended to accept a third argument, an upper bound on the cluster size."

[klausler]

flang/lib/Semantics/resolve-names.cpp

3552

That's for LAUNCH_BOUNDS(), not CLUSTER_DIMS().

[klausler]

Allow third parameter to LAUNCH_BOUNDS().

[klausler]

Upload correct diff (sorry).

[wangzpgi]

flang/lib/Semantics/resolve-names.cpp

3552

Yes, allowing third argument was for launch_bounds, not cluster_dims, I commented at the wrong place. Not quite sure how this Phabricator review works.

For cluster_dims as well as launch_bounds, Brent has requested to take maybe const expressions. He would like to use module parameters, for global routines that are in that module. Neither Michael nor I implemented that in nvfortran.

[wangzpgi]

Is there checks for CUDA Version that some features are allowed for certain CUDA Version?

[klausler]

In D150161#4331054, @wangzpgi wrote:

Is there checks for CUDA Version that some features are allowed for certain CUDA Version?

Not yet.

flang/lib/Semantics/resolve-names.cpp

3552

That's what I implemented -- any expression that can be folded to a constant value at compilation time will work. They don't have to integer literals.

[wangzpgi]

flang/lib/Semantics/resolve-names.cpp

3552

Great. Thanks.

[klausler]

Rebase to current head.

[klausler]

Rebasing to current head in preparation for merging.

flang/include/flang/Semantics/scope.h

 
   // Create the Global scope -- the root of the scope tree
   explicit Scope(SemanticsContext &context)
       : Scope{*this, Kind::Global, nullptr, context} {}
   Scope(Scope &parent, Kind kind, Symbol *symbol, SemanticsContext &context)
-      : parent_{parent}, kind_{kind}, symbol_{symbol}, context_{context} {
+      : parent_{&parent}, kind_{kind}, symbol_{symbol}, context_{context} {
     if (symbol) {
       symbol->set_scope(this);
     }
   }
   Scope(const Scope &) = delete;
 
   bool operator==(const Scope &that) const { return this == &that; }
   bool operator!=(const Scope &that) const { return this != &that; }
 
   Scope &parent() {
-    CHECK(&parent_ != this);
-    return parent_;
+    CHECK(parent_ != this);
+    return *parent_;
   }
   const Scope &parent() const {
-    CHECK(&parent_ != this);
-    return parent_;
+    CHECK(parent_ != this);
+    return *parent_;
   }
   Kind kind() const { return kind_; }
   bool IsGlobal() const { return kind_ == Kind::Global; }
   bool IsIntrinsicModules() const { return kind_ == Kind::IntrinsicModules; }
   bool IsTopLevel() const {

   inline std::optional<SourceName> GetName() const;
   // Returns true if this scope contains, or is, another scope.
   bool Contains(const Scope &) const;
   /// Make a scope nested in this one
   Scope &MakeScope(Kind kind, Symbol *symbol = nullptr);
+
   SemanticsContext &GetMutableSemanticsContext() const {
     return const_cast<SemanticsContext &>(context());
   }
 
   using size_type = mapType::size_type;

   void set_runtimeDerivedTypeDescription(const Symbol &symbol) {
     runtimeDerivedTypeDescription_ = &symbol;
   }
 
 private:
-  Scope &parent_; // this is enclosing scope, not extended derived type base
+  Scope *parent_{
+      nullptr}; // this is enclosing scope, not extended derived type base
   const Kind kind_;
   std::size_t size_{0}; // size in bytes
   std::optional<std::size_t> alignment_; // required alignment in bytes
   parser::CharBlock sourceRange_;
   const parser::CookedSource *cookedSource_{nullptr};

flang/include/flang/Semantics/semantics.h

   Scope *GetBuiltinModule(const char *name);
 
   // Defines builtinsScope_ from the __Fortran_builtins module
   void UseFortranBuiltinsModule();
   const Scope *GetBuiltinsScope() const { return builtinsScope_; }
-
   void UsePPCFortranBuiltinTypesModule();
+  const Scope *GetCUDABuiltinsScope();
   void UsePPCFortranBuiltinsModule();
   Scope *GetPPCBuiltinTypesScope() { return ppcBuiltinTypesScope_; }
   const Scope *GetPPCBuiltinsScope() const { return ppcBuiltinsScope_; }
 
   // Saves a module file's parse tree so that it remains available

   void CheckIndexVarRedefine(
       const parser::CharBlock &, const Symbol &, parser::MessageFixedText &&);
   void CheckError(const Symbol &);
 
   const common::IntrinsicTypeDefaultKinds &defaultKinds_;
-  const common::LanguageFeatureControl languageFeatures_;
+  const common::LanguageFeatureControl &languageFeatures_;
   parser::AllCookedSources &allCookedSources_;
   std::optional<parser::CharBlock> location_;
   std::vector<std::string> searchDirectories_;
   std::vector<std::string> intrinsicModuleDirectories_;
   std::string moduleDirectory_{"."s};

       activeIndexVars_;
   UnorderedSymbolSet errorSymbols_;
   std::set<std::string> tempNames_;
   const Scope *builtinsScope_{nullptr}; // module __Fortran_builtins
   Scope *ppcBuiltinTypesScope_{nullptr}; // module __Fortran_PPC_types
+  std::optional<const Scope *> CUDABuiltinsScope_; // module __CUDA_builtins
   const Scope *ppcBuiltinsScope_{nullptr}; // module __Fortran_PPC_intrinsics
   std::list<parser::Program> modFileParseTrees_;
   std::unique_ptr<CommonBlockMap> commonBlockMap_;
   bool anyDefinedIntrinsicOperator_{false};
 };

flang/include/flang/Semantics/symbol.h

     CHECK(result_ != nullptr);
     result_ = &result;
   }
   bool defaultIgnoreTKR() const { return defaultIgnoreTKR_; }
   void set_defaultIgnoreTKR(bool yes) { defaultIgnoreTKR_ = yes; }
+  std::optional<common::CUDASubprogramAttrs> cudaSubprogramAttrs() const {
+    return cudaSubprogramAttrs_;
+  }
+  void set_cudaSubprogramAttrs(common::CUDASubprogramAttrs csas) {
+    cudaSubprogramAttrs_ = csas;
+  }
+  std::vector<std::int64_t> &cudaLaunchBounds() { return cudaLaunchBounds_; }
+  const std::vector<std::int64_t> &cudaLaunchBounds() const {
+    return cudaLaunchBounds_;
+  }
+  void set_cudaLaunchBounds(std::vector<std::int64_t> &&x) {
+    cudaLaunchBounds_ = std::move(x);
+  }
+  std::vector<std::int64_t> &cudaClusterDims() { return cudaClusterDims_; }
+  const std::vector<std::int64_t> &cudaClusterDims() const {
+    return cudaClusterDims_;
+  }
+  void set_cudaClusterDims(std::vector<std::int64_t> &&x) {
+    cudaClusterDims_ = std::move(x);
+  }
 
 private:
   bool isInterface_{false}; // true if this represents an interface-body
   bool isDummy_{false}; // true when interface of dummy procedure
   std::vector<Symbol *> dummyArgs_; // nullptr -> alternate return indicator

   // For MODULE FUNCTION or SUBROUTINE, this is the symbol of its declared
   // interface.  For MODULE PROCEDURE, this is the declared interface if it
   // appeared in an ancestor (sub)module.
   Symbol *moduleInterface_{nullptr};
   bool defaultIgnoreTKR_{false};
+  // CUDA ATTRIBUTES(...) from subroutine/function prefix
+  std::optional<common::CUDASubprogramAttrs> cudaSubprogramAttrs_;
+  // CUDA LAUNCH_BOUNDS(...) & CLUSTER_DIMS(...) from prefix
+  std::vector<std::int64_t> cudaLaunchBounds_, cudaClusterDims_;
 
   friend llvm::raw_ostream &operator<<(
       llvm::raw_ostream &, const SubprogramDetails &);
 };
 

     return isDummy() && shape_.CanBeAssumedShape();
   }
   bool CanBeDeferredShape() const { return shape_.CanBeDeferredShape(); }
   bool IsAssumedSize() const { return isDummy() && shape_.CanBeAssumedSize(); }
   bool IsAssumedRank() const { return isDummy() && shape_.IsAssumedRank(); }
+  std::optional<common::CUDADataAttr> cudaDataAttr() const {
+    return cudaDataAttr_;
+  }
+  void set_cudaDataAttr(std::optional<common::CUDADataAttr> attr) {
+    cudaDataAttr_ = attr;
+  }
 
 private:
   MaybeExpr init_;
   const parser::Expr *unanalyzedPDTComponentInit_{nullptr};
   ArraySpec shape_;
   ArraySpec coshape_;
   common::IgnoreTKRSet ignoreTKR_;
   const Symbol *commonBlock_{nullptr}; // common block this object is in
+  std::optional<common::CUDADataAttr> cudaDataAttr_;
   friend llvm::raw_ostream &operator<<(
       llvm::raw_ostream &, const ObjectEntityDetails &);
 };
 
 // Mixin for details with passed-object dummy argument.

   // Be advised: !init().has_value() => uninitialized pointer,
   // while *init() == nullptr => explicit NULL() initialization.
   std::optional<const Symbol *> init() const { return init_; }
   void set_init(const Symbol &symbol) { init_ = &symbol; }
   void set_init(std::nullptr_t) { init_ = nullptr; }
+  bool isCUDAKernel() const { return isCUDAKernel_; }
+  void set_isCUDAKernel(bool yes = true) { isCUDAKernel_ = yes; }
 
 private:
   const Symbol *procInterface_{nullptr};
   std::optional<const Symbol *> init_;
+  bool isCUDAKernel_{false};
   friend llvm::raw_ostream &operator<<(
       llvm::raw_ostream &, const ProcEntityDetails &);
 };
 
 // These derived type details represent the characteristics of a derived

flang/include/flang/Semantics/tools.h

 bool IsAssumedType(const Symbol &);
 bool IsPolymorphic(const Symbol &);
 bool IsUnlimitedPolymorphic(const Symbol &);
 bool IsPolymorphicAllocatable(const Symbol &);
 
+inline bool IsCUDADeviceContext(const Scope *scope) {
+  if (scope) {
+    if (const Symbol * symbol{scope->symbol()}) {
+      if (const auto *subp{symbol->detailsIf<SubprogramDetails>()}) {
+        if (auto attrs{subp->cudaSubprogramAttrs()}) {
+          return *attrs != common::CUDASubprogramAttrs::Host;
+        }
+      }
+    }
+  }
+  return false;
+}
+
+const Scope *FindCUDADeviceContext(const Scope *);
+std::optional<common::CUDADataAttr> GetCUDADataAttr(const Symbol *);
+
 // Return an error if a symbol is not accessible from a scope
 std::optional<parser::MessageFormattedText> CheckAccessibleSymbol(
     const semantics::Scope &, const Symbol &);
 
 // Analysis of image control statements

 //   - Potential subobject components of a derived type are the closure of
 //     its non-POINTER components and the potential subobject components of
 //     its non-POINTER derived type components.  (The lifetime of each
 //     potential subobject component is that of the entire instance.)
 //   - PotentialAndPointer subobject components of a derived type are the
-//   closure of
-//     its components (including POINTERs) and the PotentialAndPointer subobject
-//     components of its non-POINTER derived type components.
+//     closure of its components (including POINTERs) and the
+//     PotentialAndPointer subobject components of its non-POINTER derived type
+//     components.
 // Parent and procedure components are considered against these definitions.
 // For this kind of iterator, the component tree is recursively visited in the
 // following order:
 //  - the parent component first (if relevant)
 //  - then, the components of the parent type (if relevant)

flang/lib/Semantics/mod-file.cpp

   llvm::raw_ostream &os{isInterface ? decls_ : contains_};
   if (isInterface) {
     os << (isAbstract ? "abstract " : "") << "interface\n";
   }
   PutAttrs(os, prefixAttrs, nullptr, false, ""s, " "s);
+  if (auto attrs{details.cudaSubprogramAttrs()}) {
+    if (*attrs == common::CUDASubprogramAttrs::HostDevice) {
+      os << "attributes(host,device) ";
+    } else {
+      PutLower(os << "attributes(", common::EnumToString(*attrs)) << ") ";
+    }
+    if (!details.cudaLaunchBounds().empty()) {
+      os << "launch_bounds";
+      char sep{'('};
+      for (auto x : details.cudaLaunchBounds()) {
+        os << sep << x;
+        sep = ',';
+      }
+      os << ") ";
+    }
+    if (!details.cudaClusterDims().empty()) {
+      os << "cluster_dims";
+      char sep{'('};
+      for (auto x : details.cudaClusterDims()) {
+        os << sep << x;
+        sep = ',';
+      }
+      os << ") ";
+    }
+  }
   os << (details.isFunction() ? "function " : "subroutine ");
   os << symbol.name() << '(';
   int n = 0;
   for (const auto &dummy : details.dummyArgs()) {
     if (n++ > 0) {

         break;
       }
     });
     os << ") " << symbol.name() << '\n';
   }
+  if (auto attr{details.cudaDataAttr()}) {
+    PutLower(os << "attributes(", common::EnumToString(*attr))
+        << ") " << symbol.name() << '\n';
+  }
 }
 
 void ModFileWriter::PutProcEntity(llvm::raw_ostream &os, const Symbol &symbol) {
   if (symbol.attrs().test(Attr::INTRINSIC)) {
     os << "intrinsic::" << symbol.name() << '\n';

   parser::Parsing parsing{context_.allCookedSources()};
   parser::Options options;
   options.isModuleFile = true;
   options.features.Enable(common::LanguageFeature::BackslashEscapes);
   options.features.Enable(common::LanguageFeature::OpenMP);
+  options.features.Enable(common::LanguageFeature::CUDA);
   if (!isIntrinsic.value_or(false) && !notAModule) {
     // The search for this module file will scan non-intrinsic module
     // directories.  If a directory is in both the intrinsic and non-intrinsic
     // directory lists, the intrinsic module directory takes precedence.
     options.searchDirectories = context_.searchDirectories();

flang/lib/Semantics/resolve-names.cpp

 // Provide Post methods to collect attributes into a member variable.
 class AttrsVisitor : public virtual BaseVisitor {
 public:
   bool BeginAttrs(); // always returns true
   Attrs GetAttrs();
+  std::optional<common::CUDADataAttr> cudaDataAttr() { return cudaDataAttr_; }
   Attrs EndAttrs();
   bool SetPassNameOn(Symbol &);
   void SetBindNameOn(Symbol &);
   void Post(const parser::LanguageBindingSpec &);
   bool Pre(const parser::IntentSpec &);

   HANDLE_ATTR_CLASS(Save, SAVE)
   HANDLE_ATTR_CLASS(Target, TARGET)
   HANDLE_ATTR_CLASS(Value, VALUE)
   HANDLE_ATTR_CLASS(Volatile, VOLATILE)
 #undef HANDLE_ATTR_CLASS
+  bool Pre(const common::CUDADataAttr);
 
 protected:
   std::optional<Attrs> attrs_;
+  std::optional<common::CUDADataAttr> cudaDataAttr_;
 
   Attr AccessSpecToAttr(const parser::AccessSpec &x) {
     switch (x.v) {
     case parser::AccessSpec::Kind::Public:
       return Attr::PUBLIC;

 
   bool HandleImplicitNone(const std::list<ImplicitNoneNameSpec> &nameSpecs);
 };
 
 // Track array specifications. They can occur in AttrSpec, EntityDecl,
-// ObjectDecl, DimensionStmt, CommonBlockObject, or BasedPointerStmt.
+// ObjectDecl, DimensionStmt, CommonBlockObject, BasedPointerStmt, and
+// ComponentDecl.
 // 1. INTEGER, DIMENSION(10) :: x
 // 2. INTEGER :: x(10)
 // 3. ALLOCATABLE :: x(:)
 // 4. DIMENSION :: x(10)
 // 5. COMMON x(10)

   }
   void SetImplicitAttr(Symbol &symbol, Attr attr) const {
     symbol.attrs().set(attr);
     symbol.implicitAttrs().set(attr);
   }
+  void SetCUDADataAttr(
+      SourceName, Symbol &, std::optional<common::CUDADataAttr>);
 
 protected:
   FuncResultStack &funcResultStack() { return funcResultStack_; }
 
   // Apply the implicit type rules to this symbol.

   void Post(const parser::InterfaceBody::Subroutine &);
   bool Pre(const parser::InterfaceBody::Function &);
   void Post(const parser::InterfaceBody::Function &);
   bool Pre(const parser::Suffix &);
   bool Pre(const parser::PrefixSpec &);
+  bool Pre(const parser::PrefixSpec::Attributes &);
+  void Post(const parser::PrefixSpec::Launch_Bounds &);
+  void Post(const parser::PrefixSpec::Cluster_Dims &);
 
   bool BeginSubprogram(const parser::Name &, Symbol::Flag,
       bool hasModulePrefix = false,
       const parser::LanguageBindingSpec * = nullptr,
       const ProgramTree::EntryStmtList * = nullptr);

   void Post(const parser::AllocatableStmt &) { objectDeclAttr_ = std::nullopt; }
   bool Pre(const parser::TargetStmt &) {
     objectDeclAttr_ = Attr::TARGET;
     return true;
   }
+  bool Pre(const parser::CUDAAttributesStmt &);
   void Post(const parser::TargetStmt &) { objectDeclAttr_ = std::nullopt; }
   void Post(const parser::DimensionStmt::Declaration &);
   void Post(const parser::CodimensionDecl &);
   bool Pre(const parser::TypeDeclarationStmt &) { return BeginDecl(); }
   void Post(const parser::TypeDeclarationStmt &);

   }
   bool Pre(const parser::BlockData &) {
     llvm_unreachable("This node is handled in ProgramUnit");
   }
 
-  void NoteExecutablePartCall(Symbol::Flag, const parser::Call &);
+  void NoteExecutablePartCall(
+      Symbol::Flag, const parser::Call &, bool hasCUDAChevrons);
 
   friend void ResolveSpecificationParts(SemanticsContext &, const Symbol &);
 
 private:
   // Kind of procedure we are expecting to see in a ProcedureDesignator

   bool BeginScopeForNode(const ProgramTree &);
   void EndScopeForNode(const ProgramTree &);
   void FinishSpecificationParts(const ProgramTree &);
   void FinishDerivedTypeInstantiation(Scope &);
   void ResolveExecutionParts(const ProgramTree &);
+  void UseCUDABuiltinNames();
 };
 
 // ImplicitRules implementation
 
 bool ImplicitRules::isImplicitNoneType() const {

 }
 
 // AttrsVisitor implementation
 
 bool AttrsVisitor::BeginAttrs() {
-  CHECK(!attrs_);
-  attrs_ = std::make_optional<Attrs>();
+  CHECK(!attrs_ && !cudaDataAttr_);
+  attrs_ = Attrs{};
   return true;
 }
 Attrs AttrsVisitor::GetAttrs() {
   CHECK(attrs_);
   return *attrs_;
 }
 Attrs AttrsVisitor::EndAttrs() {
   Attrs result{GetAttrs()};
   attrs_.reset();
+  cudaDataAttr_.reset();
   passName_ = std::nullopt;
   bindName_.reset();
   return result;
 }
 

     return false;
   }
   attrs_->set(attrName);
   return true;
 }
+bool AttrsVisitor::Pre(const common::CUDADataAttr x) {
+  if (cudaDataAttr_.value_or(x) != x) {
+    Say(currStmtSource().value(),
+        "CUDA data attributes '%s' and '%s' may not both be specified"_err_en_US,
+        common::EnumToString(*cudaDataAttr_), common::EnumToString(x));
+  }
+  cudaDataAttr_ = x;
+  return false;
+}
 
 // DeclTypeSpecVisitor implementation
 
 const DeclTypeSpec *DeclTypeSpecVisitor::GetDeclTypeSpec() {
   return state_.declTypeSpec;

   attrs.IterateOverMembers(
       [&](Attr x) { ok &= CheckDuplicatedAttr(name, symbol, x); });
   return ok;
 }
 
+void ScopeHandler::SetCUDADataAttr(SourceName source, Symbol &symbol,
+    std::optional<common::CUDADataAttr> attr) {
+  if (attr) {
+    ConvertToObjectEntity(symbol);
+    if (auto *object{symbol.detailsIf<ObjectEntityDetails>()}) {
+      if (*attr != object->cudaDataAttr().value_or(*attr)) {
+        Say(source,
+            "'%s' already has another CUDA data attribute ('%s')"_err_en_US,
+            symbol.name(),
+            common::EnumToString(*object->cudaDataAttr()).substr());
+      } else {
+        object->set_cudaDataAttr(attr);
+      }
+    } else {
+      Say(source,
+          "'%s' is not an object and may not have a CUDA data attribute"_err_en_US,
+          symbol.name());
+    }
+  }
+}
+
 // ModuleVisitor implementation
 
 bool ModuleVisitor::Pre(const parser::Only &x) {
   common::visit(common::visitors{
                     [&](const Indirection<parser::GenericSpec> &generic) {

   if (const auto *parsedType{std::get_if<parser::DeclarationTypeSpec>(&x.u)}) {
     FuncResultStack::FuncInfo &info{DEREF(funcResultStack().Top())};
     if (info.parsedType) { // C1543
       Say(currStmtSource().value(),
           "FUNCTION prefix cannot specify the type more than once"_err_en_US);
-      return false;
     } else {
       info.parsedType = parsedType;
       info.source = currStmtSource();
-      return false;
     }
+    return false;
   } else {
     return true;
   }
 }
 
+bool SubprogramVisitor::Pre(const parser::PrefixSpec::Attributes &attrs) {
+  if (auto *subp{currScope().symbol()
+              ? currScope().symbol()->detailsIf<SubprogramDetails>()
+              : nullptr}) {
+    for (auto attr : attrs.v) {
+      if (auto current{subp->cudaSubprogramAttrs()}) {
+        if (attr == *current ||
+            (*current == common::CUDASubprogramAttrs::HostDevice &&
+                (attr == common::CUDASubprogramAttrs::Host ||
+                    attr == common::CUDASubprogramAttrs::Device))) {
+          Say(currStmtSource().value(),
+              "ATTRIBUTES(%s) appears more than once"_warn_en_US,
+              common::EnumToString(attr));
+        } else if ((attr == common::CUDASubprogramAttrs::Host ||
+                       attr == common::CUDASubprogramAttrs::Device) &&
+            (*current == common::CUDASubprogramAttrs::Host ||
+                *current == common::CUDASubprogramAttrs::Device ||
+                *current == common::CUDASubprogramAttrs::HostDevice)) {
+          // HOST,DEVICE or DEVICE,HOST -> HostDevice
+          subp->set_cudaSubprogramAttrs(
+              common::CUDASubprogramAttrs::HostDevice);
+        } else {
+          Say(currStmtSource().value(),
+              "ATTRIBUTES(%s) conflicts with earlier ATTRIBUTES(%s)"_err_en_US,
+              common::EnumToString(attr), common::EnumToString(*current));
+        }
+      } else {
+        subp->set_cudaSubprogramAttrs(attr);
+      }
+    }
+  }
+  return false;
+}
+
+void SubprogramVisitor::Post(const parser::PrefixSpec::Launch_Bounds &x) {
+  std::vector<std::int64_t> bounds;
+  bool ok{true};
+  for (const auto &sicx : x.v) {
+    if (auto value{evaluate::ToInt64(EvaluateExpr(sicx))}) {
+      bounds.push_back(*value);
+    } else {
+      ok = false;
+    }
+  }
+  if (!ok || bounds.size() < 2 || bounds.size() > 3) {
+    Say(currStmtSource().value(),
+        "Operands of LAUNCH_BOUNDS() must be 2 or 3 integer constants"_err_en_US);
+  } else if (auto *subp{currScope().symbol()
+                     ? currScope().symbol()->detailsIf<SubprogramDetails>()
+                     : nullptr}) {
+    if (subp->cudaLaunchBounds().empty()) {
+      subp->set_cudaLaunchBounds(std::move(bounds));
+    } else {
+      Say(currStmtSource().value(),
+          "LAUNCH_BOUNDS() may only appear once"_err_en_US);
+    }
+  }
+}
+
+void SubprogramVisitor::Post(const parser::PrefixSpec::Cluster_Dims &x) {
+  std::vector<std::int64_t> dims;
+  bool ok{true};
+  for (const auto &sicx : x.v) {
+    if (auto value{evaluate::ToInt64(EvaluateExpr(sicx))}) {
+      dims.push_back(*value);
+    } else {
+      ok = false;
+    }
+  }
+  if (!ok || dims.size() != 3) {
+    Say(currStmtSource().value(),
+        "Operands of CLUSTER_DIMS() must be three integer constants"_err_en_US);
+  } else if (auto *subp{currScope().symbol()
+                     ? currScope().symbol()->detailsIf<SubprogramDetails>()
+                     : nullptr}) {
+    if (subp->cudaClusterDims().empty()) {
+      subp->set_cudaClusterDims(std::move(dims));
+    } else {
+      Say(currStmtSource().value(),
+          "CLUSTER_DIMS() may only appear once"_err_en_US);
+    }
+  }
+}
+
 bool SubprogramVisitor::Pre(const parser::InterfaceBody::Subroutine &x) {
   const auto &name{std::get<parser::Name>(
       std::get<parser::Statement<parser::SubroutineStmt>>(x.t).statement.t)};
   return BeginSubprogram(name, Symbol::Flag::Subroutine);
 }

     SayAlreadyDeclared(name, entrySymbol);
     return;
   }
   SubprogramDetails &entryDetails{entrySymbol.get<SubprogramDetails>()};
   CHECK(entryDetails.entryScope() == &inclusiveScope);
+  SetCUDADataAttr(name.source, entrySymbol, cudaDataAttr());
   entrySymbol.attrs() |= GetAttrs();
   SetBindNameOn(entrySymbol);
   for (const auto &dummyArg : std::get<std::list<parser::DummyArg>>(stmt.t)) {
     if (const auto *dummyName{std::get_if<parser::Name>(&dummyArg.u)}) {
       if (Symbol * dummy{FindInScope(*dummyName)}) {

 void DeclarationVisitor::Post(const parser::EntityDecl &x) {
   const auto &name{std::get<parser::ObjectName>(x.t)};
   Attrs attrs{attrs_ ? HandleSaveName(name.source, *attrs_) : Attrs{}};
   Symbol &symbol{DeclareUnknownEntity(name, attrs)};
   symbol.ReplaceName(name.source);
+  SetCUDADataAttr(name.source, symbol, cudaDataAttr());
   if (const auto &init{std::get<std::optional<parser::Initialization>>(x.t)}) {
     ConvertToObjectEntity(symbol) || ConvertToProcEntity(symbol);
     symbol.set(
         Symbol::Flag::EntryDummyArgument, false); // forestall excessive errors
     Initialization(name, *init, false);

       HandleAttributeStmt(Attr::VALUE, x.v);
 }
 bool DeclarationVisitor::Pre(const parser::VolatileStmt &x) {
   return HandleAttributeStmt(Attr::VOLATILE, x.v);
 }
+bool DeclarationVisitor::Pre(const parser::CUDAAttributesStmt &x) {
+  auto attr{std::get<common::CUDADataAttr>(x.t)};
+  for (const auto &name : std::get<std::list<parser::Name>>(x.t)) {
+    auto *symbol{FindInScope(name)};
+    if (symbol && symbol->has<UseDetails>()) {
+      Say(currStmtSource().value(),
+          "Cannot apply CUDA data attribute to use-associated '%s'"_err_en_US,
+          name.source);
+    } else {
+      if (!symbol) {
+        symbol = &MakeSymbol(name, ObjectEntityDetails{});
+      }
+      SetCUDADataAttr(name.source, *symbol, attr);
+    }
+  }
+  return false;
+}
 // Handle a statement that sets an attribute on a list of names.
 bool DeclarationVisitor::HandleAttributeStmt(
     Attr attr, const std::list<parser::Name> &names) {
   for (const auto &name : names) {
     HandleAttributeStmt(attr, name);

       }
     }
   }
   if (OkToAddComponent(name)) {
     auto &symbol{DeclareObjectEntity(name, attrs)};
+    SetCUDADataAttr(name.source, symbol, cudaDataAttr());
     if (symbol.has<ObjectEntityDetails>()) {
       if (auto &init{std::get<std::optional<parser::Initialization>>(x.t)}) {
         Initialization(name, *init, true);
       }
     }

   }
   if (!dtDetails) {
     attrs.set(Attr::EXTERNAL);
   }
   Symbol &symbol{DeclareProcEntity(name, attrs, procInterface)};
+  SetCUDADataAttr(name.source, symbol, cudaDataAttr()); // for error
   symbol.ReplaceName(name.source);
   if (dtDetails) {
     dtDetails->add_component(symbol);
   }
 }

         !attrs.HasAny({Attr::PUBLIC, Attr::PRIVATE}) &&
         std::holds_alternative<ProcBindingDetails>(details)) {
       attrs.set(Attr::PRIVATE);
     }
     Symbol &result{MakeSymbol(name, attrs, std::move(details))};
+    SetCUDADataAttr(name, result, cudaDataAttr());
     if (result.has<TypeParamDetails>()) {
       derivedType.symbol()->get<DerivedTypeDetails>().add_paramDecl(result);
     }
     return &result;
   }

   HandleCall(Symbol::Flag::Function, x.v);
   return false;
 }
 bool ResolveNamesVisitor::Pre(const parser::CallStmt &x) {
   HandleCall(Symbol::Flag::Subroutine, x.call);
+  Walk(x.chevrons);
   return false;
 }
 
 bool ResolveNamesVisitor::Pre(const parser::ImportStmt &x) {
   auto &scope{currScope()};

 // Variant of HandleProcedureName() for use while skimming the executable
 // part of a subprogram to catch calls to dummy procedures that are part
 // of the subprogram's interface, and to mark as procedures any symbols
 // that might otherwise have been miscategorized as objects.
 void ResolveNamesVisitor::NoteExecutablePartCall(
-    Symbol::Flag flag, const parser::Call &call) {
+    Symbol::Flag flag, const parser::Call &call, bool hasCUDAChevrons) {
   auto &designator{std::get<parser::ProcedureDesignator>(call.t)};
   if (const auto *name{std::get_if<parser::Name>(&designator.u)}) {
     // Subtlety: The symbol pointers in the parse tree are not set, because
     // they might end up resolving elsewhere (e.g., construct entities in
     // SELECT TYPE).

       Symbol::Flag other{flag == Symbol::Flag::Subroutine
               ? Symbol::Flag::Function
               : Symbol::Flag::Subroutine};
       if (!symbol->test(other)) {
         ConvertToProcEntity(*symbol);
-        if (symbol->has<ProcEntityDetails>()) {
+        if (auto *details{symbol->detailsIf<ProcEntityDetails>()}) {
           symbol->set(flag);
           if (IsDummy(*symbol)) {
             SetImplicitAttr(*symbol, Attr::EXTERNAL);
           }
           ApplyImplicitRules(*symbol);
+          if (hasCUDAChevrons) {
+            details->set_isCUDAKernel();
+          }
         }
       }
     }
   }
 }

       common::ScopedSet(specPartState_, SpecificationPartState{})};
   Walk(accDecls);
   Walk(ompDecls);
   Walk(compilerDirectives);
   Walk(useStmts);
+  UseCUDABuiltinNames();
   ClearUseRenames();
   ClearUseOnly();
   ClearModuleUses();
   Walk(importStmts);
   Walk(implicitPart);

   Walk(decls);
   FinishSpecificationPart(decls);
   return false;
 }
 
+void ResolveNamesVisitor::UseCUDABuiltinNames() {
+  if (FindCUDADeviceContext(&currScope())) {
+    if (const Scope * CUDABuiltins{context().GetCUDABuiltinsScope()}) {
+      for (const auto &[name, symbol] : *CUDABuiltins) {
+        if (!FindInScope(name)) {
+          auto &localSymbol{MakeSymbol(name)};
+          localSymbol.set_details(UseDetails{name, *symbol});
+          localSymbol.flags() = symbol->flags();
+        }
+      }
+    }
+  }
+}
+
 // Initial processing on specification constructs, before visiting them.
 void ResolveNamesVisitor::PreSpecificationConstruct(
     const parser::SpecificationConstruct &spec) {
   common::visit(
       common::visitors{

   }
 
   template <typename A> bool Pre(const A &) { return true; }
   template <typename A> void Post(const A &) {}
   void Post(const parser::FunctionReference &fr) {
-    resolver_.NoteExecutablePartCall(Symbol::Flag::Function, fr.v);
+    resolver_.NoteExecutablePartCall(Symbol::Flag::Function, fr.v, false);
   }
   void Post(const parser::CallStmt &cs) {
-    resolver_.NoteExecutablePartCall(Symbol::Flag::Subroutine, cs.call);
+    resolver_.NoteExecutablePartCall(
+        Symbol::Flag::Subroutine, cs.call, cs.chevrons.has_value());
   }
 
 private:
   ResolveNamesVisitor &resolver_;
 };

 }
 
 void ResolveNamesVisitor::Post(const parser::Program &) {
   // ensure that all temps were deallocated
   CHECK(!attrs_);
+  CHECK(!cudaDataAttr_);
   CHECK(!GetDeclTypeSpec());
 }
 
 // A singleton instance of the scope -> IMPLICIT rules mapping is
 // shared by all instances of ResolveNamesVisitor and accessed by this

flang/lib/Semantics/scope.cpp

     return &*it->second;
   } else if (IsSubmodule()) {
     const Scope *parent{symbol_->get<ModuleDetails>().parent()};
     return parent ? parent->FindSymbol(name) : nullptr;
   } else if (CanImport(name)) {
-    return parent_.FindSymbol(name);
+    return parent_->FindSymbol(name);
   } else {
     return nullptr;
   }
 }
 

   importNames_.insert(name);
 }
 
 // true if name can be imported or host-associated from parent scope.
 bool Scope::CanImport(const SourceName &name) const {
-  if (IsTopLevel() || parent_.IsTopLevel()) {
+  if (IsTopLevel() || parent_->IsTopLevel()) {
     return false;
   }
   switch (GetImportKind()) {
     SWITCH_COVERS_ALL_CASES
   case ImportKind::None:

flang/lib/Semantics/semantics.cpp

   if (ppcBuiltinTypesScope_ == nullptr) {
     ppcBuiltinTypesScope_ = GetBuiltinModule("__fortran_ppc_types");
   }
 }
 
+const Scope *SemanticsContext::GetCUDABuiltinsScope() {
+  if (!CUDABuiltinsScope_) {
+    CUDABuiltinsScope_ = GetBuiltinModule("__cuda_builtins");
+  }
+  return *CUDABuiltinsScope_;
+}
+
 void SemanticsContext::UsePPCFortranBuiltinsModule() {
   if (ppcBuiltinsScope_ == nullptr) {
     ppcBuiltinsScope_ = GetBuiltinModule("__fortran_ppc_intrinsics");
   }
 }

flang/lib/Semantics/symbol.cpp

     os << " moduleInterface: " << *x.moduleInterface_;
   }
   if (x.defaultIgnoreTKR_) {
     os << " defaultIgnoreTKR";
   }
+  if (x.cudaSubprogramAttrs_) {
+    os << " cudaSubprogramAttrs: "
+       << common::EnumToString(*x.cudaSubprogramAttrs_);
+  }
+  if (!x.cudaLaunchBounds_.empty()) {
+    os << " cudaLaunchBounds:";
+    for (auto x : x.cudaLaunchBounds_) {
+      os << ' ' << x;
+    }
+  }
+  if (!x.cudaClusterDims_.empty()) {
+    os << " cudaClusterDims:";
+    for (auto x : x.cudaClusterDims_) {
+      os << ' ' << x;
+    }
+  }
   return os;
 }
 
 void EntityDetails::set_type(const DeclTypeSpec &type) {
   CHECK(!type_);

     os << " (has unanalyzedPDTComponentInit)";
   }
   if (!x.ignoreTKR_.empty()) {
     x.ignoreTKR_.Dump(os << ' ', common::EnumToString);
   }
+  if (x.cudaDataAttr()) {
+    os << " cudaDataAttr: " << common::EnumToString(*x.cudaDataAttr());
+  }
   return os;
 }
 
 llvm::raw_ostream &operator<<(
     llvm::raw_ostream &os, const AssocEntityDetails &x) {

       os << " => " << target->name();
     } else {
       os << " => NULL()";
     }
   }
+  if (x.isCUDAKernel()) {
+    os << " isCUDAKernel";
+  }
   return os;
 }
 
 llvm::raw_ostream &operator<<(
     llvm::raw_ostream &os, const DerivedTypeDetails &x) {

flang/lib/Semantics/tools.cpp

 
 bool IsPolymorphicAllocatable(const Symbol &symbol) {
   return IsAllocatable(symbol) && IsPolymorphic(symbol);
 }
 
+const Scope *FindCUDADeviceContext(const Scope *scope) {
+  return !scope ? nullptr : FindScopeContaining(*scope, [](const Scope &s) {
+    return IsCUDADeviceContext(&s);
+  });
+}
+
+std::optional<common::CUDADataAttr> GetCUDADataAttr(const Symbol *symbol) {
+  const auto *object{
+      symbol ? symbol->detailsIf<ObjectEntityDetails>() : nullptr};
+  return object ? object->cudaDataAttr() : std::nullopt;
+}
+
 std::optional<parser::MessageFormattedText> CheckAccessibleSymbol(
     const Scope &scope, const Symbol &symbol) {
   if (symbol.attrs().test(Attr::PRIVATE)) {
     if (FindModuleFileContaining(scope)) {
       // Don't enforce component accessibility checks in module files;

flang/module/__fortran_builtins.f90

   end type
 
   integer, parameter :: __builtin_atomic_int_kind = selected_int_kind(18)
   integer, parameter :: __builtin_atomic_logical_kind = __builtin_atomic_int_kind
 
+  procedure(type(__builtin_c_ptr)) :: __builtin_c_loc
+
+  type :: __builtin_dim3
+    integer :: x=1, y=1, z=1
+  end type
+  type(__builtin_dim3) :: &
+    __builtin_threadIdx, __builtin_blockDim, __builtin_blockIdx, __builtin_gridDim
+  integer, parameter :: __builtin_warpsize = 32
+
   intrinsic :: __builtin_ieee_is_nan, __builtin_ieee_is_negative, &
     __builtin_ieee_is_normal
   intrinsic :: __builtin_ieee_next_after, __builtin_ieee_next_down, &
     __builtin_ieee_next_up
   intrinsic :: scale ! for ieee_scalb

flang/test/Parser/cuf-sanity-common

   attributes(grid_global) subroutine gridglobalsub; end
   attributes(host) subroutine hostsub; end
   attributes(global) launch_bounds(1, 2) subroutine lbsub; end
   attributes(global) cluster_dims(1, 2, 3) subroutine cdsub; end
   attributes(device) subroutine attrs
-! enable with name resolution:    attributes(device) :: devx1
+    attributes(device) :: devx1
     real, device :: devx2
   end subroutine
   subroutine test
     logical isPinned
     !$cuf kernel do(1) <<<*, *, stream = 1>>>

flang/test/Semantics/cuf01.cuf

+! RUN: %python %S/test_errors.py %s %flang_fc1
+! Test conflicting CUDA subprogram attributes
+module m1
+ contains
+  !WARNING: ATTRIBUTES(Host) appears more than once
+  attributes(host,host) subroutine ok1; end
+  !WARNING: ATTRIBUTES(Host) appears more than once
+  attributes(host) attributes(host) subroutine ok2; end
+  attributes(host,device) subroutine ok3; end
+  attributes(device,host) subroutine ok4; end
+  !WARNING: ATTRIBUTES(Host) appears more than once
+  attributes(host,device,host) subroutine ok5; end
+  !WARNING: ATTRIBUTES(Device) appears more than once
+  attributes(device,host,device) subroutine ok6; end
+  !ERROR: ATTRIBUTES(Global) conflicts with earlier ATTRIBUTES(Host)
+  attributes(host,global) subroutine conflict1; end
+  !ERROR: ATTRIBUTES(Host) conflicts with earlier ATTRIBUTES(Global)
+  attributes(global,host) subroutine conflict2; end
+  !ERROR: ATTRIBUTES(Grid_Global) conflicts with earlier ATTRIBUTES(Host)
+  attributes(host,grid_global) subroutine conflict3; end
+  !TODO: more with launch_bounds & cluster_dims
+end module

flang/test/Semantics/modfile55.cuf

+! RUN: %python %S/test_modfile.py %s %flang_fc1
+! Sanity check for CUDA Fortran attributes in module files
+module m
+  attributes(device) dd
+  real, managed, allocatable :: md
+  real, pinned, allocatable :: mp
+  attributes(constant) cd
+ contains
+  attributes(global) subroutine globsub(x,y,z)
+    real, value :: x
+    real, device :: y
+    real, managed :: z
+  end subroutine
+  attributes(host,device) real function foo(x)
+    foo = x + 1.
+  end function
+end
+
+!Expect: m.mod
+!module m
+!real(4)::dd
+!attributes(device) dd
+!real(4),allocatable::md
+!attributes(managed) md
+!real(4),allocatable::mp
+!attributes(pinned) mp
+!real(4)::cd
+!attributes(constant) cd
+!contains
+!attributes(global) subroutine globsub(x,y,z)
+!real(4),value::x
+!real(4)::y
+!attributes(device) y
+!real(4)::z
+!attributes(managed) z
+!end
+!attributes(host,device) function foo(x)
+!real(4)::x
+!real(4)::foo
+!end
+!end