Task completion signatures ignore `error_types` from custom env

[wokron]

Hi,

When stdexec::task is given a custom environment with using error_types, the task's completion signatures still report set_error_t(std::exception_ptr) instead of the declared error type (e.g. set_error_t(std::error_code)).

Repro

// task_error_types_repro.cpp
#include <stdexec/execution.hpp>
#include <system_error>
#include <variant>

struct my_env
{
  using error_types = stdexec::completion_signatures<stdexec::set_error_t(std::error_code)>;
};

auto func() noexcept -> stdexec::task<int, my_env>
{
  co_yield stdexec::with_error{std::make_error_code(std::errc::invalid_argument)};
  co_return 1;
}

int main()
{
  auto s = func() | stdexec::upon_error([](auto err) noexcept { return err; })
         | stdexec::into_variant();
  auto [r] = stdexec::sync_wait(std::move(s)).value();
  bool fail = std::holds_alternative<std::tuple<std::error_code>>(r);
  return fail ? 1 : 0;
}

Build

• stdexec: e92245d
• clang: 23.0.0

clang++ task_error_types_repro.cpp -std=c++23 -I ./include/

This fails to compile with:

...
/usr/lib/gcc/x86_64-linux-gnu/14/../../../../include/c++/14/variant:1143:21: error: static assertion failed due to
      requirement '__detail::__variant::__exactly_once<std::tuple<std::error_code>, std::tuple<int>,
      std::tuple<std::__exception_ptr::exception_ptr>>': T must occur exactly once in alternatives
 1143 |       static_assert(__detail::__variant::__exactly_once<_Tp, _Types...>,
      |                     ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
task_error_types_repro.cpp:21:20: note: in instantiation of function template specialization
      'std::holds_alternative<std::tuple<std::error_code>, std::tuple<int>,
      std::tuple<std::__exception_ptr::exception_ptr>>' requested here
   21 |   bool fail = std::holds_alternative<std::tuple<std::error_code>>(r);
      |                    ^
...

[Cra3z]

It appears that the current implementation cannot retrieve the correct completion signatures for task: task is an awaitable sender that does not provide the member function template get_completion_signatures and connect, which causes the error channel in its completion signatures to always be set_error_t(exception_ptr), see

stdexec/include/stdexec/__detail/__connect.hpp

Lines 348 to 352 in f91f636

	else if constexpr (__connect::__with_co_await<__new_sndr_t, _Receiver>)
	{
	return __connect_awaitable(static_cast<__new_sndr_t&&>(__new_sndr),
	static_cast<_Receiver&&>(__rcvr));
	}

stdexec/include/stdexec/__detail/__get_completion_signatures.hpp

Lines 257 to 260 in f91f636

	else if constexpr (__with_co_await<_Sender, _Env>)
	{
	return __co_await_completions_t<_Sender, _Env>();
	}

stdexec/include/stdexec/__detail/__get_completion_signatures.hpp

Lines 106 to 111 in f91f636

	template <class _Sender, class... _Env>
	using __co_await_completions_t = //
	completion_signatures<
	__single_value_sig_t<__await_result_t<_Sender, __detail::__promise<_Env>...>>,
	set_error_t(std::exception_ptr),
	set_stopped_t()>;

[Cra3z]

I think this issue can be resolved by providing get_completion_signatures and connect members for task. The definition of task::get_completion_signatures is as follows:

template <__decays_to<task>, class _Env>
  requires __task::__has_compatible_environment_with<_Env, _TaskEnv>
static consteval auto get_completion_signatures() noexcept
{
  using signatures =
    __concat_completion_signatures_t<completion_signatures<__single_value_sig_t<_Ty>>,
                                     error_types,
                                     completion_signatures<set_stopped_t()>>;
  return signatures{};
}

However, I feel there isn't a good way to implement task::connect. If we follow the approach in the standard, it will most likely lead to stack overflow due to asymmetric transfer.