[BUG]: CTAD is broken for thrust::pair

[jrhemstad]

Is this a duplicate?

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

Type of Bug

Compile-time Error

Component

Thrust

Describe the bug

The refactor to make thrust::pair an alias for cuda::std::pair ended up breaking CTAD because CTAD doesn't work for alias templates.

How to Reproduce

Try to compile:

#include <thrust/pair.h>
auto p = thrust::pair{42,42};

yields:

<source>(3): error: argument list for alias template "thrust::THRUST_200400_520_NS::pair" is missing
  auto p = thrust::pair{42,42};

Expected behavior

CTAD should work for thrust::pair.

Reproduction link

https://godbolt.org/z/1s6as73bb

Operating System

No response

nvidia-smi output

No response

NVCC version

No response

[bernhardmgruber]

I had this problem in the past, so here are some thoughts: I could not find an easy solution in C++17, but it works out of the box in C++20 (via P1021]. A deduction guide for the alias does not work. The best solution I found back than (for a more complex type) was to inherit instead of alias in C++17, and redefine the public API where necessary:

namespace thrust {
    template <typename T, typename U>
    struct pair : cuda::std::pair<T, U> { };
    // ... specialize tuple protocol for thrust::pair
}

Given the comparable simple implementation of cuda::std::pair, we may just have a distinct (duplicate) pair implementation in thrust until C++20.

[miscco]

Yeah that is definitely an option. I just hate the additional boilerplate that we need to use.

Mind that we then also need to specialize all of {cuda::}std::tuple_size and {cuda::}std::tuple_element