How to expose a static reference and an unmangled static pointer to the same memory location?

help
1

I have a struct called Encoding and I want to statically allocate a bunch of instances with different field values such that for a given set of field values there exists only instance with an address that's stable across crate boundaries and the FFI boundary.

Previously, I had in lib.rs stuff like:

/// The Big5 encoding.
pub const BIG5: &'static Encoding = &Encoding {
    name: "Big5",
    variant: VariantEncoding::Big5,
};

/// The EUC-JP encoding.
pub const EUC_JP: &'static Encoding = &Encoding {
    name: "EUC-JP",
    variant: VariantEncoding::EucJp,
};

And in re-exported ffi.rs stuff like:

/// Newtype for `*const Encoding` in order to be able to implement `Sync` for
/// it.
pub struct ConstEncoding(*const Encoding);

/// Required for `static` fields.
unsafe impl Sync for ConstEncoding {}

/// The Big5 encoding.
#[no_mangle]
pub static BIG5_ENCODING: ConstEncoding = ConstEncoding(BIG5);

/// The EUC-JP encoding.
#[no_mangle]
pub static EUC_JP_ENCODING: ConstEncoding = ConstEncoding(EUC_JP);

And in the C header file stuff like:

/// The Big5 encoding.
extern const Encoding* const BIG5_ENCODING;

/// The EUC-JP encoding.
extern const Encoding* const EUC_JP_ENCODING;

I expected

Encoding {
    name: "Big5",
    variant: VariantEncoding::Big5,
}

to be correspond to a unique range of statically allocated memory and BIG5 and BIG5_ENCODING both to refer/point to it.

As long as I had only the declaring Rust crate itself involved, this seemed to work within the crate and across FFI. However, when I started using the crate from another crate, use of BIG5 in the other crate resulted in different instance of

Encoding {
    name: "Big5",
    variant: VariantEncoding::Big5,
}

with references to it pointing to a different address.

On IRC, I was told that this was to be expected and that I should use static if I wanted a stable address across crates.

However, if I do s/const/static/ when declaring BIG5, the compiler no longer allows me to use BIG5 to initialize BIG5_ENCODING for FFI, because statics aren't allowed to refer to other statics.

What's the right way to have a single statically allocated instance of a struct and to export both a reference (for Rust consumers) and a pointer (for C consumers) to it?

Should I have

const BIG5_CONST: &'static Encoding = &Encoding {
    name: "Big5",
    variant: VariantEncoding::Big5,
};

and

/// The Big5 encoding.
pub static BIG5: &'static Encoding = BIG5_CONST;

and

#[no_mangle]
pub static BIG5_ENCODING: ConstEncoding = ConstEncoding(BIG5_CONST);

And rely on the undocumented(?) compiler behavior of these two instantiation of BIG5_CONST getting coalesced into one as long as they are in the same crate?

2

Well, statics are allowed to to refer to other statics. Statics are not allowed to refer to consts. Sometimes if you add too much &s you may get the error though.

First of all, I would stop declaring both the struct and the reference in one step, like you're doing (I'm not even sure what this syntax is a sugar for):

pub const BIG5: &'static Encoding = &Encoding { ... }; // don't do that

Instead, first declare the struct:

#[no_mangle]
pub static BIG5: Encoding = Encoding { ... };

This way, you're sure that &BIG5 is a stable address – a static reference of type &'static Encoding which can be used directly by Rust consumers. Now you can declare other statics if you want:

#[no_mangle]
pub static BIG5_ENCODING: ConstEncoding = ConstEncoding(&BIG5_CONST);

Now, in C you can either define:

extern Encoding const BIG5;

or

extern Encoding const * const BIG5_ENCODING;

(sorry for writing const on the right, but I find it more consistent)

1 Like
3

Thank you. Indeed, it seems to be that the key to the solution is not initializing the struct and taking a reference to it in one go.

Curiously, while taking a reference to a static and making another static out of the reference works, constants are prohibited from referring to statics to the point of a constant for the reference to a static struct being prohibited. Since constants seem to be mostly inlined expression I have a hard time seeing why this restriction needs to be in the language.