This is a simple reproducer for NVIDIA bug 4216808 (https://developer.nvidia.com/bugs/4216808). The problem has been fixed in CUDA 12.4.0.

Description

When nvcc compiles a C++ file, in some corner cases it produces a host binary with the wrong symbol type, b (a BSS local symbol) instead of u (a unique global symbol).

For example, compiling this file with gcc

#include <type_traits>

// user defined type used as template argument
class Type {};

// the SFINAE condition is a possible cause of the problem
template <typename T1,
          typename T2,
          typename = std::enable_if_t<std::is_class_v<T1> and std::is_class_v<T2>>>
class Resource {
public:
  explicit Resource(int value) : value_{value} {}

  int value_;
};

// initialise the function-static variable on the first call
template <typename T>
inline Resource<Type, T>& getResource() {
  static Resource<Type, T> resource(42);
  return resource;
}

// explicit call to instantiate the templates
void call() {
  getResource<Type>();
}

generates unique global symbols for the resource static variable:

$ g++-12 -O2 -std=c++17 -c test.cc -o test.gcc.o
$ nm -C test.gcc.o | grep '::resource$'
0000000000000000 u guard variable for getResource<Type>()::resource
0000000000000000 u getResource<Type>()::resource

Compiling the same file with nvcc generates a BSS local symbol:

$ /usr/local/cuda-12.2/bin/nvcc -ccbin g++-12 -O2 -std=c++17 -x cu -c test.cc -o test.nvcc.o
$ nm -C test.nvcc.o | grep '::resource$'
0000000000000000 b guard variable for getResource<Type>()::resource
0000000000000008 b getResource<Type>()::resource

Minor changes to the test (for example, removing the SFINAE condition on the the Resource class template) do not trigger the problem any more, resulting in unique global symbols also when compiling with nvcc.

Details

Tested on Ubuntu 22.04 and RHEL 8.7, with CUDA 11.5, 11.8, 12.1, 12.2 and 12.3, with GCC 11.x and 12.x (where supported).