[libcxx-commits] [PATCH] D58019: Add is_nothrow_convertible (P0758R1)

[Louis Dionne via Phabricator via libcxx-commits]

ldionne added a comment.

Actually, I'm looking at this and `std::is_convertible` again, and I think this should be using `std::is_convertible` in the implementation. I know the paper suggests the implementation in this patch, however it's unclear to me whether that implementation properly handles all the cases that `std::is_convertible` handles (notably with references). Also, using `std::is_convertible` under the hood means that we will use the `__is_convertible_to` intrinsic when it's available instead.

I would do something like this instead (untested):

  template <typename _Fm, typename _To>
  struct __is_nothrow_convertible_impl {
  private:
      template <typename _Tp>
      static void __test_noexcept(_Tp) noexcept;
      
      template<typename __Fm, typename __To>
      static bool_constant<noexcept(__test_noexcept<__To>(declval<__Fm>()))>
      __test(int);
      
      template<typename, typename>
      static false_type
      __test(...);
      
  public:
      using type = decltype(__test<_Fm, _To>(0));
  };
  
  template <typename _Fm, typename _To>
  struct is_nothrow_convertible : __and_<is_convertible<_Fm, _To>, __is_nothrow_convertible_impl<_Fm, _To>>
  { };

This is very similar to what you have, but I also guard the whole thing with `std::is_convertible`. What do you think? Did I miss a subtlety that would make my draft implementation wrong (that's quite possible)?


Repository:
  rCXX libc++

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D58019/new/

https://reviews.llvm.org/D58019