[FEA]: libcu++ iterator_traits should pick up std iterator_traits specializations

[bernhardmgruber]

Is this a duplicate?

• I confirmed there appear to be no duplicate issues for this request and that I agree to the Code of Conduct

Area

libcu++

Is your feature request related to a problem? Please describe.

In cub/util_type.cuh we have something like:

#ifdef _CCCL_COMPILER_NVRTC
#  include <cuda/std/iterator>
#else // !defined(_CCCL_COMPILER_NVRTC)
#  include <iterator>
#endif // defined(_CCCL_COMPILER_NVRTC)
...
template <typename Iterator>
using value_t =
#  ifdef _CCCL_COMPILER_NVRTC
  typename ::cuda::std::iterator_traits<Iterator>::value_type;
#  else // !defined(_CCCL_COMPILER_NVRTC)
  typename std::iterator_traits<Iterator>::value_type;
#  endif // defined(_CCCL_COMPILER_NVRTC)

IIRC, @gevtushenko pointed out that the special treatment for NVRTC is necessary, because there may be no standard library around to supply std::iterator_traits. However, this seems like a problem that should be solved by libcu++ instead. Furthermore, we use plain std::iterator_traits in a lot of places without this special handling.

We cannot just use ::cuda::std::iterator_traits directly, because a user is free to specialize std::iterator_traits for their iterator type, which currently is not detected by ::cuda::std::iterator_traits.

Describe the solution you'd like

I want to use only ::cuda::std::iterator_traits<T> and it should pick up a specialization of std::iterator_traits<T> when it exists.

Describe alternatives you've considered

No response

Additional context

No response

[bernhardmgruber]

What already works today is that ::cuda::std::xxx_iterator_tag refers to the std:: ones when a host compiler is available.