generalize the Option<NonNullablePtr> to libs

[thestinger]

An Option<Rc<T>> is now 1-word, like the built-in non-nullable pointers.

[alexcrichton]

This seems like a pretty cool feature, but it's also a little worrying to me. Right now you can declare unsafe_no_drop_flag and unsafe_non_zero_word and the compiler performs 0 checks to make sure you do anything sanely. I would expect unsafe_non_zero_word to at least verify that the struct's size is only 1 word.

It does make sense to flag them with unsafe, but it's not immediately clear why these attributes have unsafe in front of them. With unsafe blocks the compiler will tell you the specific reason that you want them, but with these attributes the compiler will just silently assume that you're obeying whatever contract is necessary.

I'm not really sure what the best answer to that is. Arguably the only way to actually learn of the attributes would be to read about them in some documentation, so the best solution could be to just document them well.

[thestinger]

I think you could also use this on a multiple-word struct as long as the first word was always zero. Although, it's not explicitly supported yet because I don't know a use case.

[huonw]

It seems like it would be relatively easy to pass the crate ctxt rather than the ty ctxt to give an error for size != 1 word.

[thestinger]

It would be equally unsafe with a check like that though.

[pcwalton]

I vote making sure it's 1 word for now to be conservative. BTW, @nikomatsakis and I would like to remove the drop flag entirely, rendering this moot.

[nikomatsakis]

Indeed. I am in fact actively working on this, though there is an annoyingly long set of blocker bugs in between here and there.

[nikomatsakis]

But does that truly render the issue moot? It would still affect how much space Option<RC<T>> takes.

[catamorphism]

I'm hearing that this patch will be subsumed by @nikomatsakis 's work-in-progress, so I'm closing the PR. Re-open if necessary.

[thestinger]

@catamorphism: I don't think it will - this is about making Option<Rc<int>> be 1 word instead of 2, like we do for &, @ and ~.

It does work but it could provide error messages on types that are not 1 word, instead of unspecified behaviour.

[catamorphism]

Closing again due to lack of activity. Please do reopen if you're able to make progress, @thestinger !

[thestinger]

I will eventually :).