wrong field access on unsized types about miri HOT 11 CLOSED

Generative types from existentials would allow perfectly typed safe vtables from even user code ;).

from miri.

@Aatch tells me that they had this problem in the rustc compiler, too, and solved it by computing these trait object field offsets dynamically by reading the alignment stored in the vtable.

from miri.

Specifically, we need to do it dynamically for only the DST field, only when it's not the first field, and only for non-packed structs.

from miri.

We can make use of our size_and_align_of_dst. Here is the relevant code in rustc trans with comments.

from miri.

While I agree that this is the solution, shouldn't rustc's layout computation fail on dsts? I mean the layout is just plain wrong.

from miri.

@oli-obk No, the layout computation for DSTs returns information you can't possibly get otherwise.

from miri.

How about protecting the offsets field (access only through an accessor) in case sized: false?

from miri.

@oli-obk You still need the offset of the last field. It's correct in terms of being the base offset but each backend has to generate its own dynamic alignment code. I don't think it's possible to ergonomically hide it.

from miri.

@eddyb We could ergonomically hide it if MIR desugared to a form with explicit vtable accesses and expanded DST field offsets to MIR statements doing this offset calculation in a backend-independent way. =)

from miri.

Unfortunately MIR is typed, so you would need.... vtable types? A person can dream, though.

from miri.

I mean, other than using the generative typing trick through lifetime HRTB.

from miri.