[flang][cuda] Avoid to issue data transfer in device context

flang/lib/Lower/Bridge.cpp

@@ -3807,16 +3807,34 @@ class FirConverter : public Fortran::lower::AbstractConverter {
     return temps;
   }
 
+  // Check if the insertion point is currently in a device context. HostDevice
+  // subprogram are not considered fully device context so it will return false
+  // for it.
+  static bool isDeviceContext(fir::FirOpBuilder &builder) {
+    if (builder.getRegion().getParentOfType<fir::CUDAKernelOp>())
+      return true;
+    if (auto funcOp =
+            builder.getRegion().getParentOfType<mlir::func::FuncOp>()) {
+      if (auto cudaProcAttr =
+              funcOp.getOperation()->getAttrOfType<fir::CUDAProcAttributeAttr>(
+                  fir::getCUDAAttrName())) {
+        return cudaProcAttr.getValue() != fir::CUDAProcAttribute::Host &&
+               cudaProcAttr.getValue() != fir::CUDAProcAttribute::HostDevice;
+      }
+    }
+    return false;
+  }
+
   void genDataAssignment(
       const Fortran::evaluate::Assignment &assign,
       const Fortran::evaluate::ProcedureRef *userDefinedAssignment) {
     mlir::Location loc = getCurrentLocation();
     fir::FirOpBuilder &builder = getFirOpBuilder();
 
-    bool isInDeviceContext =
-        builder.getRegion().getParentOfType<fir::CUDAKernelOp>();
-    bool isCUDATransfer = Fortran::evaluate::HasCUDAAttrs(assign.lhs) ||
-                          Fortran::evaluate::HasCUDAAttrs(assign.rhs);
+    bool isInDeviceContext = isDeviceContext(builder);
+    bool isCUDATransfer = (Fortran::evaluate::HasCUDAAttrs(assign.lhs) ||
+                           Fortran::evaluate::HasCUDAAttrs(assign.rhs)) &&
+                          !isInDeviceContext;
     bool hasCUDAImplicitTransfer =
         Fortran::evaluate::HasCUDAImplicitTransfer(assign.rhs);
     llvm::SmallVector<mlir::Value> implicitTemps;
@@ -3879,7 +3897,7 @@ class FirConverter : public Fortran::lower::AbstractConverter {
       Fortran::lower::StatementContext localStmtCtx;
       hlfir::Entity rhs = evaluateRhs(localStmtCtx);
       hlfir::Entity lhs = evaluateLhs(localStmtCtx);
-      if (isCUDATransfer && !hasCUDAImplicitTransfer && !isInDeviceContext)
+      if (isCUDATransfer && !hasCUDAImplicitTransfer)
         genCUDADataTransfer(builder, loc, assign, lhs, rhs);
       else
         builder.create<hlfir::AssignOp>(loc, rhs, lhs,

flang/test/Lower/CUDA/cuda-data-transfer.cuf

@@ -141,3 +141,21 @@ end subroutine
 ! CHECK: fir.cuda_kernel<<<*, *>>>
 ! CHECK-NOT: fir.cuda_data_transfer
 ! CHECK: hlfir.assign
+
+attributes(global) subroutine sub5(a)
+  integer, device :: a
+  integer :: i
+  a = i
+end subroutine
+
+! CHECK-LABEL: func.func @_QPsub5
+! CHECK-NOT: fir.cuda_data_transfer
+
+attributes(host,device) subroutine sub6(a)
+  integer, device :: a
+  integer :: i
+  a = i
+end subroutine
+
+! CHECK-LABEL: func.func @_QPsub6
+! CHECK: fir.cuda_data_transfer