`#[repr(packed(N))]` (tracking issue for RFC 1399)

[nikomatsakis]

Tracking issue for rust-lang/rfcs#1399: #[repr(packed(N))]

Unresolved issues

• Should borrows to packed fields be safe if the field's natural alignment is greater than or equal to the packing value, or should all fields of a packed struct always be unsafe to borrow?
• Currently #[repr(packed(_))] structs cannot transitively contain #[repr(align(_))] structs due to differing behavior between msvc and gcc. Do we want to keep this a hard error, pick one behavior over the other, or provide some way to choose which behavior is desired?

[retep998]

#27060 is closely related to this.

[bitshifter]

This should probably use attr_literals now like repr(align) does, e.g. repr(pack(2)).

[ahicks92]

Is anyone working on this? If not I think I mostly know how to do it, modulo someone telling me how the attribute parsing and verification needs to be extended higher up.

[ahicks92]

Unless someone comes in and says they're doing this before then, I'm going to take a crack at it Friday or thereabouts. My intent is to use the attribute literals approach (or support both if that's what people want), plus potentially submitting a PR against the RFC repo to update the RFC to attribute literals.

@eddyb or whoever else might know: is inserting a bunch of [0 x u8] in packed structs going to degrade performance? The easiest approach here seems to be to multiply everything in the type index to memory index vector by 2, then insert such dummy fields in packed structs. I think that everything else after that may just work except maybe constructing constants for them, but iirc constant construction is in one place and relatively easy for me to fix in this way.

[eddyb]

@camlorn everything currently has those usually-empty arrays, for #[repr(align(N))].
I also have a work-in-progress branch where I've simplified a lot of trans (including making constant ADTs boring), but there's more to be done - maybe I should try to polish it and open a PR soon.

[ahicks92]

@eddyb
Should I wait on you?

[eddyb]

I'd say so, yes, at least for the trans part. In fact, I'm not sure there is anything left to do once I'm done, other than use the value from the attribute in the align field.