Use const instead of macros

[sosthene-nitrokey]

This is pretty much a complete rewrite. It gives a much nicer API to
understand from the documentation. It also means that it doesn't rely on
#[doc(hidden)] methods (except for Interchange::const_new so that it
can be removed once the required compiler features are stabilized)

Because everything doesn't need to be 'static anymore, it can also be
tested efficiently with Loom, ensuring that there are no possible race
conditions.

It is a significantly breaking change

There may be some possible improvement. For example the claims mechanism could be implemented in the Channel struct and use RAII to be made safe and remove the need for reset-claims. This could make the API 100% safe which would be nice, but it would not be zero cost, given that it's unlikely a Requester or a Responder would ever be dropped in most use cases.

[sosthene-nitrokey]

I've updated the PR with this trick which makes it possible to have a const constructor and completely remove all macros.

I also added an RAII mechanism instead of the reset_claims which should not be that big of an overhead. It also makes Channel::split safe which means that no public API is unsafe, and that Channel can be used by itself without using the Interchange structure.

[sosthene-nitrokey]

I've made a branch that also removes the Clone trait bound on the request and response structs. sosthene-nitrokey#1

[robin-nitrokey]

It would be helpful to also implement Default for Interchange and Channel.

[sosthene-nitrokey]

Isn't it already the case?

• https://github.com/trussed-dev/interchange/pull/5/files#diff-b1a35a68f14e696205874893c07fd24fdb88882b47c23cc0e0c80a30c7d53759R405
• https://github.com/trussed-dev/interchange/pull/5/files#diff-b1a35a68f14e696205874893c07fd24fdb88882b47c23cc0e0c80a30c7d53759R913

[robin-nitrokey]

Ah, sorry, I probably looked at the wrong branch.

[nickray]

Not worried about the overhead of the atomic bools to implement the RAII, just to state the obvious though, it's not guaranteed destructors (drop) are run. Seems fine for the intended use cases.

Overall, I think these are good changes, particularly adding the testing with Loom. Do we have any chance of upstreaming constness as needed to Loom, or is this a property of the implementations there?

Also.. did Loom catch any bugs already?

[nickray]

I presume Q = query/question, and A = answer? Maybe use longer names in the struct definition (and switch to the short ones in the implementation)? In any case, please add a comment on what the names mean, so it's easier to understand.

[sosthene-nitrokey]

You're right, I'm not very satisfied with Q,A since everything else mentions Request/Response. I'll replace it with Rq and Rp

[sosthene-nitrokey]

This is the implementation of new for loom's atomic mocks.

It registers itself to loom's runtime so I doubt it will ever be made const.
It may be possible to make it const by making the registration happen lazily on the first operation, but I don't know if it's worth it or if it can work.

[sosthene-nitrokey]

I also added an InterchangeRef type that removes the need to have a const parameter everywhere.

static INTERCHANGE_INNER: Interchange<Request, Response, 1> = Interchange::new();

// The size of the interchange is absent from the type
static INTERCHANGE: InterchangeRef<'static, Request, Response> = INTERCHANGE_INNER.as_interchange_ref();

let (mut rq, mut rp) = INTERCHANGE.claim().unwrap();