exec(fork_join): coalesce empty and unary calls

[romintomasetti]

• When exec::fork_join(sndr, clsr) is given a single closure, do no attempt to wrap it in the (thereby useless) fork_join machinery. Return sndr | clsr.
• When exec::fork_join(sndr) is given no closure, it behaves like sndr.

As a drive-by, I've added a few noexcept to then.

Similar to:

• when_all: Conditional Stop Source #2124.

See P4269R1.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[romintomasetti]

@RobertLeahy @ericniebler Could you please have a look? Thanks!

[RobertLeahy]

I don't think this noexcept is correct. Just because Closure can be decay-copied doesn't mean binding it together with something of type Sndr can't.

[romintomasetti]

Indeed! How would you do it, apart from

noexcept(noexcept(static_cast<Sndr&&>(sndr) | static_cast<Closure&&>(clsr)))

?

[maartenarnst]

Use inspiration from include/stdexec/__detail/__sender_adaptor_closure.hpp?

[romintomasetti]

Done. What do you think?

[maartenarnst]

We could also omit this line. Because it seems the reason is that the function has no noexcept clause, rather than that one operations may throw.

[romintomasetti]

Yeah, but at least it's tested.

[maartenarnst]

OK.

[RobertLeahy]

I would expect constructing such an exec::fork_join sender not to throw, which noexcept is missing that prevents that?

[romintomasetti]

There are 2 functions in the fork_join_t struct I haven't touched yet.

[maartenarnst]

My two cents would be to add the noexcept clause and flip the static_assert here.

We could write something like (perhaps there's a way to make it less verbose with other concepts from __concepts.hpp):

noexcept(STDEXEC::__nothrow_decay_copyable<Closures..., Sndr>
         && (std::is_nothrow_move_constructible_v<STDEXEC::__decay_t<Closures>> && ...)
         && std::is_nothrow_move_constructible_v<STDEXEC::__decay_t<Sndr>>)

The last two checks would be needed for the moves here:

stdexec/include/stdexec/__detail/__basic_sender.hpp

Lines 460 to 470 in f4c123f

	template <class _Tag>
	struct __make_sexpr_t
	{
	template <class _Data = __, class... _Child>
	constexpr auto operator()(_Data __data = {}, _Child... __child) const
	{
	return __sexpr_t<_Tag, _Data, _Child...>{
	{_Tag(), static_cast<_Data&&>(__data), static_cast<_Child&&>(__child)...}
	};
	}
	};

[romintomasetti]

@ericniebler What do you think?

[ericniebler]

we should probably fix __make_sexpr_t to perfectly forward the arguments and give it a noexcept clause. i can make a PR for that.

[ericniebler]

That change has now landed and i have merged it into this PR, so the noexcept clause here can be simply noexcept(STDEXEC::__nothrow_decay_copyable<Closures..., Sndr>).

[maartenarnst]

Ok! Thank you very much for the pr and merging it here. To write the noexcept clause, one remaining question seems to be whether we should take into account the move of the __tuple that fork_join constructs and then passes to __make_sexpr?