Closures accepting structs with lifetimes

[Michael-F-Bryan]

I ran into an annoying issue with BoxDynFnMut1 and friends when you try to pass around a struct with a non-'static lifetime due to the lack of higher-kinded types.

Basically, I want to let users set their own logging function. In normal Rust you'd write something like this:

pub fn set_logger(log_func: Box<dyn Fn(&log::Record<'_>)>) { ... }

The equivalent using safer_ffi would be (in theory) something like this:

#[ffi_export]
pub fn set_logger(log_func: for<'a> BoxDynFnMut1<(), LogRecord<'a>>) { ... }

#[derive_ReprC]
#[repr(C)]
pub struct LogRecord<'a> {
    level: LogLevel,
    target: str_ref<'a>,
}

Obviously for<'a> SomeStruct isn't valid syntax and you get an "Error: expected trait, found struct BoxDynFnMut1" error, however without the for<'a> you get "Error: cannot infer an appropriate lifetime due to conflicting requirements" when you try to use it in practice.

This is a concrete example of the "improve support for closures" mentioned in #47.

[danielhenrymantilla]

Indeed! FWIW, there is already support for higher-ranked function pointers (stateless), which can test with the ditto internal branch:

#[derive_ReprC] // Both these attrs will likely become `#[safer_ffi::repr(transparent)]`
#[repr(transparent)]
struct MyLoggerFnPointer(
    extern "C" fn(&log::Record<'_>),
);

One API for higher-ranked closures, right now, would be an API very similar to https://docs.rs/stackbox/0.1.2/stackbox/macro.custom_dyn.html (or at least that would be the general solution to #47):

#[safer_ffi::repr(C)] // Defines `BoxDynMyLogger<'_>`, `ArcDynMyLogger<'_>` types.
trait MyLogger : Send + Sync {
    fn call (self: &'_ Self, _: &log::Record<'_>);
}

• Btw @Michael-F-Bryan, thanks to your blogpost, there may even be a #[safer_ffi::repr(C, thin)] option 🙃

If you have suggestions ideas for these things, I'm all ears 🙂

For the specific case of closures, I could also imagine using the same #[repr(transparent)] syntax showcased above, but directly supporting the repr_c::closure::… types:

#[derive_ReprC]
#[repr(transparent)]
struct MyLoggerClosure (
    repr_c::BoxDynClosure<fn(&log::Record<'_>)>, // <- optional future syntax for the closure types
);

[danielhenrymantilla]

I know this is an old issue, but I wanted to gauge the situation at the moment, @Michael-F-Bryan:

• safer-ffi nowadays supports FFI-safe dyn traits which can be Boxed, and which hopefully do support this generic lifetime param in input;
• thanks to https://docs.rs/higher-kinded-types, it would be possible to express some ForFn<ForLt!(&log::Record<'_>)> kind of type, at least in the covariant case
  • (or ForFn<For!(<'a, 'b> = &'a log::Record<'b>)> in the invariant case)