[mcbain]

A while back I filed an issue on the MSVC std::vector implementation as the behaviour of a weird corner case differed from GCC/Clang.

The reply I got from STL himself was fantastic, pointing out that it was [paraphrasing here] a wart in the spec, and that neither were actually incorrect.

To my lasting regret, I didn't keep an ahem offsite backup of some of those corp emails. Hence I can't recall the exact details.

[StephanTLavavej]

This sounds like the vector<list<unique_ptr<T>>> fiasco, where a vector contains elements that are movable-only (you can't copy a list<unique_ptr<T>>, only move it), yet with throwing move constructors (as MSVC's list, set, and other node-based containers dynamically allocate sentinel nodes when default/move-constructed - this is a behavior difference from GCC's libstdc++ and Clang's libc++, where both dynamically-allocated and container-internal sentinel nodes are permitted by the Standard). The doom begins when STL containers like list (also vector itself) don't have "constrained copy constructors" - that is, list(const list&) always appears to be available, as far as type traits like is_copy_constructible are concerned, even if that copy constructor can't actually compile. The doom accelerates when vector<Elem> needs to ask the question "are you nothrow-movable? If so, I can move you efficiently. Otherwise, are you copyable? If so, I can fall back to copying you during reallocation, in order to preserve the strong exception guarantee. Otherwise, you're throwing movable-only, so I'll move you and you get the basic guarantee". The doom is complete when list<unique_ptr<T>> correctly reports that it isn't nothrow-move-constructible (it would throw due to allocating a sentinel), yet incorrectly claims to be is_copy_constructible. Therefore the vector tries to copy the list<unique_ptr<T>> and that fails to instantiate, resulting in an epic compiler error.

This situation is a fiasco because there is no good way out of it (constraining container copy constructors isn't possible because the Standard permits incomplete element types for many containers, and users use them for other containers even though they technically shouldn't, weakening EH guarantees is potentially problematic, changing the nature of sentinel nodes is ABI-breaking and affects iterator invalidation guarantees). It is also highly memorable, which is why I can guess what it was :-)

[mcbain]

I only just noticed this reply - many apologies.

Yes, that's the one! Our fix was explicitly deleting the copy constructor. In essence a one-liner for us, but you can't see the iceberg of an issue beneath the surface.

Thanks once again. :)