Don't use repr(C)

[Jules-Bertholet]

Fixes #41.

[dtolnay]

Thanks!

[afeistel]

this seems to set off some clippy warnings (#[warn(clippy::repr_packed_without_abi)]); maybe #[repr(Rust, packed)] instead?

[RalfJung]

What is the reason why this uses a 0-length array to begin with, rather than core::marker::PhantomData?

I'm considering to ban arrays from being ignored in repr(transparent) since that is an ABI promise that I am not sure we can hold (rust-lang/rust#156114). With that change, ghost's phantom types would no longer be ignored by repr(transparent). No regression showed up on crater so that does not seem to be a pattern that's very widely used, though it'd be nice to keep it working.

[RalfJung]

Never mind, @Jules-Bertholet found a wording in rust-lang/rust#157973 that will keep this crate working.

However I am still curious why you are using a 0-lement array instead of PhantomData.

[dtolnay]

It is a demonstration that a PhantomData-like can be built from primitives without using an existing PhantomData.

[Jules-Bertholet]

Note that rust-lang/rust#157973 doesn't really answer all the issues here. That PR describes what types have trivial ABI on any specific target, but for deciding what to allow in repr(transparent) we will (IMO) want to ensure target independence.

[RalfJung]

Yeah, that's why rust-lang/rust#155984 is so complicated -- and the code explains basically none of that reasoning (especially not the high-level goals that the logic is trying to achieve) which is why I was so confused when trying to review it.^^

It is a demonstration that a PhantomData-like can be built from primitives without using an existing PhantomData.

That is a cute trick, but IMO not a great argument for actually putting this in production. This will cause a whole bunch of extra work in rust-lang/rust#155984...