The constructor

Code domains are 100% constexpr to construct and destruct, as are status codes. This enables the compiler to 100% instantiate both only in its mind, and to emit zero code and thus zero overhead.

Unfortunately it also means that it must be possible for each domain to be instantiated an infinite number of times, and being 100% in constexpr, any instances never have a unique address in memory either. Thus we cannot compare domains for equivalence using their address in memory, as std::error_category does.

We solve this by using a very random 64 bit number taken from a hard random number source. The website https://www.random.org/cgi-bin/randbyte?nbytes=8&format=h is strongly suggested as the source for this number.

(In case you are wondering about the chance of collision for a 64 bit integer, SG14 estimated that approximately 190,000 separate domains would need to exist in a single process for there to be a 0.00000001% probability of collision if the random number source is very random)

  // unique id must be from a hard random number source
  // Use https://www.random.org/cgi-bin/randbyte?nbytes=8&format=h to get a hard random 64 bit id.
  // Do NOT make up your own value. Do NOT use zero.
  constexpr explicit _file_io_error_domain(typename _base::unique_id_type id = 0x230f170194fcc6c7) noexcept : _base(id) {}
  static inline constexpr const _file_io_error_domain &get();
View this code on Github

A nice side effect of this approach is that custom error domains in header-only libraries are safe, unlike custom <system_error> error categories. Boost.System’s error categories can now opt into this same safe mechanism in order to also be safe in header only library use cases.