[flang][cuda] Implicitly load cudadevice module in device/global subprogram

flang/include/flang/Semantics/semantics.h

@@ -214,9 +214,11 @@ class SemanticsContext {
   // Defines builtinsScope_ from the __Fortran_builtins module
   void UseFortranBuiltinsModule();
   const Scope *GetBuiltinsScope() const { return builtinsScope_; }
+
+  const Scope &GetCUDABuiltinsScope();
+  const Scope &GetCUDADeviceScope();
 
   void UsePPCBuiltinTypesModule();
-  const Scope &GetCUDABuiltinsScope();
   void UsePPCBuiltinsModule();
   Scope *GetPPCBuiltinTypesScope() { return ppcBuiltinTypesScope_; }
   const Scope *GetPPCBuiltinsScope() const { return ppcBuiltinsScope_; }
@@ -292,6 +294,7 @@ class SemanticsContext {
   const Scope *builtinsScope_{nullptr}; // module __Fortran_builtins
   Scope *ppcBuiltinTypesScope_{nullptr}; // module __Fortran_PPC_types
   std::optional<const Scope *> cudaBuiltinsScope_; // module __CUDA_builtins
+  std::optional<const Scope *> cudaDeviceScope_; // module cudadevice
   const Scope *ppcBuiltinsScope_{nullptr}; // module __ppc_intrinsics
   std::list<parser::Program> modFileParseTrees_;
   std::unique_ptr<CommonBlockMap> commonBlockMap_;

flang/lib/Semantics/check-cuda.cpp

@@ -82,6 +82,10 @@ struct DeviceExprChecker
           }
         }
       }
+      if (sym->owner().IsModule() &&
+          DEREF(sym->owner().symbol()).name() == "__cuda_device_builtins") {
+        return {};
+      }
     } else if (x.GetSpecificIntrinsic()) {
       // TODO(CUDA): Check for unsupported intrinsics here
       return {};

flang/lib/Semantics/resolve-names.cpp

@@ -3797,6 +3797,17 @@ bool SubprogramVisitor::Pre(const parser::PrefixSpec::Attributes &attrs) {
         subp->set_cudaSubprogramAttrs(attr);
       }
     }
+    if (auto attrs{subp->cudaSubprogramAttrs()}) {
+      if (*attrs == common::CUDASubprogramAttrs::Global ||
+          *attrs == common::CUDASubprogramAttrs::Device) {
+        // Implicitly USE the cudadevice module by copying its symbol in the
+        // current scope.
+        const Scope &scope{context().GetCUDADeviceScope()};
+        for (auto sym : scope.GetSymbols()) {
+          currScope().CopySymbol(sym);
+        }
+      }
+    }
   }
   return false;
 }

flang/lib/Semantics/semantics.cpp

@@ -543,6 +543,14 @@ const Scope &SemanticsContext::GetCUDABuiltinsScope() {
   return **cudaBuiltinsScope_;
 }
 
+const Scope &SemanticsContext::GetCUDADeviceScope() {
+  if (!cudaDeviceScope_) {
+    cudaDeviceScope_ = GetBuiltinModule("cudadevice");
+    CHECK(cudaDeviceScope_.value() != nullptr);
+  }
+  return **cudaDeviceScope_;
+}
+
 void SemanticsContext::UsePPCBuiltinsModule() {
   if (ppcBuiltinsScope_ == nullptr) {
     ppcBuiltinsScope_ = GetBuiltinModule("__ppc_intrinsics");

flang/module/__cuda_device_builtins.f90

@@ -0,0 +1,74 @@
+!===-- module/__cuda_device_builtins.f90 -----------------------------------===!
+!
+! Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+! See https://llvm.org/LICENSE.txt for license information.
+! SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+!
+!===------------------------------------------------------------------------===!
+
+! CUDA Fortran procedures available in device subprogram
+
+module __CUDA_device_builtins  
+
+  implicit none
+
+  ! Set PRIVATE by default to explicitly only export what is meant
+  ! to be exported by this MODULE.
+  private
+
+  ! Synchronization Functions
+
+  interface
+    subroutine __cuda_device_builtins_syncthreads()
+    end subroutine
+  end interface
+  public :: __cuda_device_builtins_syncthreads
+
+  interface
+    integer function __cuda_device_builtins_syncthreads_and(value)
+      integer :: value
+    end function
+  end interface
+  public :: __cuda_device_builtins_syncthreads_and
+
+  interface
+    integer function __cuda_device_builtins_syncthreads_count(value)
+      integer :: value
+    end function
+  end interface
+  public :: __cuda_device_builtins_syncthreads_count
+
+  interface
+    integer function __cuda_device_builtins_syncthreads_or(int_value)
+    end function
+  end interface
+  public :: __cuda_device_builtins_syncthreads_or
+
+  interface
+    subroutine __cuda_device_builtins_syncwarp(mask)
+      integer :: mask
+    end subroutine
+  end interface
+  public :: __cuda_device_builtins_syncwarp
+
+  ! Memory Fences
+
+  interface
+    subroutine __cuda_device_builtins_threadfence()
+    end subroutine
+  end interface
+  public :: __cuda_device_builtins_threadfence
+
+  interface
+    subroutine __cuda_device_builtins_threadfence_block()
+    end subroutine
+  end interface
+  public :: __cuda_device_builtins_threadfence_block
+
+  interface
+    subroutine __cuda_device_builtins_threadfence_system()
+    end subroutine
+  end interface
+  public :: __cuda_device_builtins_threadfence_system
+
+end module

flang/module/cudadevice.f90

@@ -0,0 +1,21 @@
+!===-- module/cudedevice.f90 -----------------------------------------------===!
+!
+! Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+! See https://llvm.org/LICENSE.txt for license information.
+! SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+!
+!===------------------------------------------------------------------------===!
+
+! CUDA Fortran procedures available in device subprogram
+
+module cudadevice
+  use __cuda_device_builtins, only: &
+    syncthreads => __cuda_device_builtins_syncthreads, &
+    syncthreads_and => __cuda_device_builtins_syncthreads_and, &
+    syncthreads_count => __cuda_device_builtins_syncthreads_count, &
+    syncthreads_or => __cuda_device_builtins_syncthreads_or, &
+    syncwarp => __cuda_device_builtins_syncwarp, &
+    threadfence => __cuda_device_builtins_threadfence, &
+    threadfence_block => __cuda_device_builtins_threadfence_block, &
+    threadfence_system => __cuda_device_builtins_threadfence_system
+end module

flang/test/Semantics/cuf-device-procedures.cuf

@@ -0,0 +1,35 @@
+! RUN: %flang_fc1 -fdebug-dump-symbols %s | FileCheck %s
+
+! Test CUDA Fortran intrinsic can pass semantic
+
+attributes(global) subroutine devsub()
+  implicit none
+  integer :: ret
+
+  ! 3.6.4. Synchronization Functions
+  call syncthreads()
+  call syncwarp(1)
+  call threadfence()
+  call threadfence_block()
+  call threadfence_system()
+  ret = syncthreads_and(1)
+  ret = syncthreads_count(1)
+  ret = syncthreads_or(1)
+end
+
+! CHECK-LABEL: Subprogram scope: devsub
+! CHECK: syncthreads, EXTERNAL, PUBLIC (Subroutine): Use from __cuda_device_builtins_syncthreads in __cuda_device_builtins
+! CHECK: syncthreads_and, EXTERNAL, PUBLIC (Function): Use from __cuda_device_builtins_syncthreads_and in __cuda_device_builtins
+! CHECK: syncthreads_count, EXTERNAL, PUBLIC (Function): Use from __cuda_device_builtins_syncthreads_count in __cuda_device_builtins
+! CHECK: syncthreads_or, EXTERNAL, PUBLIC (Function): Use from __cuda_device_builtins_syncthreads_or in __cuda_device_builtins
+! CHECK: syncwarp, EXTERNAL, PUBLIC (Subroutine): Use from __cuda_device_builtins_syncwarp in __cuda_device_builtins
+! CHECK: threadfence, EXTERNAL, PUBLIC (Subroutine): Use from __cuda_device_builtins_threadfence in __cuda_device_builtins
+! CHECK: threadfence_block, EXTERNAL, PUBLIC (Subroutine): Use from __cuda_device_builtins_threadfence_block in __cuda_device_builtins
+! CHECK: threadfence_system, EXTERNAL, PUBLIC (Subroutine): Use from __cuda_device_builtins_threadfence_system in __cuda_device_builtins
+
+subroutine host()
+  call syncthreads()
+end subroutine
+
+! CHECK-LABEL: Subprogram scope: host
+! CHECK: syncthreads, EXTERNAL: HostAssoc{{$}}

flang/tools/f18/CMakeLists.txt

@@ -12,6 +12,8 @@ set(MODULES
   "__ppc_intrinsics"
   "mma"
   "__cuda_builtins"
+  "__cuda_device_builtins"
+  "cudadevice"
   "ieee_arithmetic"
   "ieee_exceptions"
   "ieee_features"