Nvcc don't recognize a type on my machine but compile on another

Accelerated Computing CUDA CUDA NVCC Compiler
1

Hi,
I have the following code in a single main.cu file:

namespace my
{
  using Arr = float[5];

  void __global__ foo(Arr const){};
}

int main(int argc, char *[])
{
  my::foo<<<1, 1>>>(nullptr);

  return 0;
}

I compile it with the command nvcc main.cu. Also, I tried to build it with the CMake and set different properties, like architecture, optimization, standard version, etc. On my machine (Ubuntu 18.04) I have the following error:

nvcc main.cu 
In file included from tmpxft_00004f21_00000000-6_main.cudafe1.stub.c:1:0:
main.cu:5:21: error: ‘_ZN2my3ArrE’ does not name a type
   void __global__ foo(Arr const){};
                     ^~~~~~~~~~~
main.cu:5:41: error: ‘void my::foo(int)’ should have been declared inside ‘my’
   void __global__ foo(Arr const){};
                                         ^
In file included from tmpxft_00004f21_00000000-6_main.cudafe1.stub.c:1:0:
main.cu: In function ‘void my::foo(int)’:
main.cu:5:40: error: invalid conversion from ‘int’ to ‘const float*’ [-fpermissive]
   void __global__ foo(Arr const){};
                                        ^
In file included from tmpxft_00004f21_00000000-6_main.cudafe1.stub.c:1:0:
/tmp/tmpxft_00004f21_00000000-6_main.cudafe1.stub.c:13:6: note:   initializing argument 1 of ‘void __device_stub__ZN2my3fooEPKf(const float*)’
 void __device_stub__ZN2my3fooEPKf(const float *__par0){__cudaLaunchPrologue(1);__cudaSetupArgSimple(__par0, 0UL);__cudaLaunch(((char *)((void ( *)(const float *))my::foo)));}

It complains about the Arr const argument of the kernel foo. If I remove the const keyword, the code compiles. Also, if I move the using directive to the global namespace, it compiles even with the const. I tried NVCC 11.4 and 10.2. Also, I tried GCC 7.5 and 7.3 as the host compiler.

The same code compiles fine on godbolt and this confuses me a lot. Is it the problem in NVCC? Or OS? Or the host compiler?