The Standard should defend container constructors from bad CTAD

[StephanTLavavej]

From #4254:

@StephanTLavavej

Found by the upcoming libcxx update, where they test CTAD for (Iter, Iter, BadAlloc) and expect it to SFINAE away.

(The explicit deduction guides are already constrained to take only things that qualify as allocators. The issue is that every constructor generates an implicit guide, and those can lead to the container being instantiated with a BadAlloc unless we guard against that.)

It's actually unclear to me whether this is guaranteed by the Standard, but the defense is simple and we already do this in basic_string.

@CaseyCarter

It's not guaranteed but IMO it should be. The Standard tells us to do so explicitly for basic_string.

The Standard has wording for deduction guides but not for constructors. This is slightly confusing since pre-Standard language in WG21 used "implicit" and "explicit" deduction guides to distinguish between the function templates derived from constructors for use in CTAD and those derived from deduction guides.

We ought to file an LWG issue.

Note that this goes beyond the sequence container constructors with (Iter, Iter, BadAlloc); there are also scenarios with the associative containers found by libcxx that I haven't yet fixed, and presumably others.

[frederick-vs-ja]

This issue seems to supersede #4060 (or be its duplicate).

[CaseyCarter]

Yeah, let's close this as a duplicate of #4060 (which has more information).