I can't wrap my brain around why we need extern "C++" for partial specializations but not explicit specializations of extern "C++" templates. (I suspect you are similarly confused, hence annotating explicit specializations despite that they don't seem to need it.) It makes sense to me that e.g. program-defined specializations of STL templates (e.g., iterator_traits) would need to attach to a different module than the library-defined primary template, but I can't see why the behavior for explicit and partial specializations would differ. Any idea/intuition here?

(To be clear: this isn't a concern with the content of the PR - I'm happy to apply this on the basis that it "works" - so much as trying to get my head straight so I can spot similar problems in the future.)

I suspect that extern "C++" works the same way for partial specializations and explicit specializations (i.e. they affect whether the specialization is attached to the global module via name mangling), and that the difference is in how the STL has separately compiled its facets (which I have never properly understood). It appears that both time_put and ctype are dllexported, but perhaps there is a behavior difference that I haven't found yet.

Aha, I think I have a better guess! time_put's _OutIt defaults to ostreambuf_iterator, but the test uses wstring::iterator. Thus, we need to instantiate a specialization for wstring::iterator (I think those are the right words), there's no way we can consume it from the DLL. My guess is that when the primary template is marked extern "C++" but not the partial specialization, something goes wrong (not sure exactly what) and we don't get an implicit instantiation. Therefore we end up with an unresolved external.

Again, not totally sure what's really happening here, but this does seem like the relevant difference that would explain why ctype seemed to work as-is.

I'm mirroring this to the MSVC-internal repo - please notify me if any further changes are pushed.

Thanks for taking the time to put this right!