change the box_free lang item to accept pointers to unsized types

[oli-obk]

in miri we use the box_free lang item as the destructor for Box objects, since the function's api matches that of an fn drop(&mut self) in a hypothetical impl<T: ?Sized> Drop for Box<T> exactly.

This works fine except if we insert a check in the size_of intrinsic to ensure that it is only called with sized types, since the box_free lang item calls that intrinsic.

cc @eddyb

no clue who to r? here, probably lang team?

[rust-highfive]

r? @sfackler

(rust_highfive has picked a reviewer for you, use r? to override)

[sfackler]

r? @eddyb

[eddyb]

Do you think you could also change trans to take advantage of this? IIRC it does its own DST logic.
I prefer this form because it's closer to an eventual Drop impl for Box.
r=me with an issue open about changing trans otherwise.

[oli-obk]

The only places rustc generates calls to box_free are in librustc_mir and librustc_borrowck, and these don't check for sizedness, but instead are at locations where the Box is guaranteed to point to a Sized type. There's some interaction with exchange_free that I don't grok. It would probably be better to leave this until Box gets a real Drop impl. I'll open an issue as advised.

[oli-obk]

travis likes it + issue reported.

[eddyb]

@bors r+

[bors]

📌 Commit de371e5 has been approved by eddyb

[bluss]

Doesn't have the [breaking-change] text in the commit log, which we use to signal breaking changes in unstable APIs. (I'll continue nagging about this until we officially abandon the practice...)

[oli-obk]

added [breaking-change] to commit message

[eddyb]

@bors r+

[bors]

📌 Commit 323c20c has been approved by eddyb