Disable automatic inlining of resource dtors and other things

[nikomatsakis]

Right now, resource destructors seem to always be inlined. This recently led to #2170 because the reachability map was incorrect. I fixed this the wrong way: by marking all resource bodies as reachable. The right fix seems to me to be not inlining resource bodies by default, but that change seemed a bit more involved. In particular, I wanted to check with @marijnh whether resources are being inlined intentionally? If so, is there a deep reason for that?

In general I'd like to be able to say that "no user-specified code is inlined unless either generic or marked #[inline]".

[marijnh]

There was no deep reason for this. Embarrassingly, I think it was an accident. I agree it should be changed.

[nikomatsakis]

One other, related, thing: right now, we always inline generic functions. I think the original plan was to having generic functions only be inlined if they were tagged #[inline]. In the event that you made a cross-crate call to non-inlined generic function, the call would only be permitted the generic types were bound to box types (or something like that). Then, for all exported generic functions, we would make sure to export a version that operates on box types.

[marijnh]

I don't think that adds much, beyond the annoying need to specify all exported generics as #[inline]. Yes, you'd have some check on implicit code duplication, but I don't believe implicit code duplication is a big enough problem to warrant such a heavy measure.

[nikomatsakis]

I'm not sure. I think the goal was to make the model of what code is
inlined simple to describe: "those things marked with #[inline]". But
maybe it's better to have the model be "generic things or things that are
marked #[inline]". You can always create monomorphic wrappers if you'd
prefer.