Limitations

[nvzqz]

Currently the following macros are only available from within the context of a function:

• assert_eq_size (commit f3da7e4)
  • Relies on forget, transmute, and uninitialized as provided by core::mem. transmute and uninitialized are marked as unsafe, which makes it unlikely for them to become const.
  • Could provide an assertion built on const_assert if mem::size_of were a const fn. To resolve the limitation, const_assert would need to be resolved first.
• const_assert (commit b7ca44e)
• const_assert_eq (commit b7ca44e)
  • Dependent on const_assert, so resolving that macro would resolve this one.

---

Commits b7ca44e and f3da7e4 have somewhat resolved this.

The following macros now require labels (identifiers) unique to their namespace to be used in a non-function context:

• assert_eq_size
• assert_obj_safe
• assert_impl
• const_assert
• const_assert_eq

---

This, along with having more assertions available (#2), would allow for a 1.0.0 release.

[nvzqz]

See rust-lang/rust#54912 for the fix to this issue. With this, we can declare macros like const_assert! as such (run in playground):

#![feature(underscore_const_names)]

macro_rules! const_assert {
    ($($xs:expr),+ $(,)*) => {
        const _: [(); 0 - !($($xs)&&+) as usize] = [];
    };
}

[asomers]

Another limitation is that the size of a generic type cannot be determined. For example, the following does not work:

#[test]
fn test_generic_types() {
    fn foo<T, U>(_t: T, _u: U) {
        assert_eq_size!(T, U);
    }
}

error[E0512]: transmute called with types of different sizes
   --> tests/eq_size.rs:113:9
    |
113 |         assert_eq_size!(T, U);
    |         ^^^^^^^^^^^^^^^^^^^^^^
    |
    = note: source type: T (this type's size can vary)
    = note: target type: U (this type's size can vary)
    = note: this error originates in a macro outside of the current crate (in Nightly builds, run with -Z external-macro-backtrace for more info)

I don't know if it's possible to remove this limitation, but it would be extremely useful to do so.

[nvzqz]

@asomers Unfortunately, if we try taking the const_assert! approach to fix this, other errors show up. 😕

Say you have the following (test in playground):

use std::mem::size_of;

fn test<T, U>() {
    // Essentially const_assert!()'s internals
    const ASSERT: [
        ();
        0 - !(size_of::<T>() == size_of::<U>()) as usize
    ] = [];
}

This generates the following errors:

error[E0401]: can't use type parameters from outer function
 --> src/main.rs:7:25
  |
3 | fn test<T, U>() {
  |    ---- - type variable from outer function
  |    |
  |    try adding a local type parameter in this method instead
...
7 |         0 - !(size_of::<T>() == size_of::<U>()) as usize
  |                         ^ use of type variable from outer function

error[E0401]: can't use type parameters from outer function
 --> src/main.rs:7:43
  |
3 | fn test<T, U>() {
  |    ----    - type variable from outer function
  |    |
  |    try adding a local type parameter in this method instead
...
7 |         0 - !(size_of::<T>() == size_of::<U>()) as usize
  |                                           ^ use of type variable from outer function

I'm not sure if anyone's working to fix this issue. I agree that it would be quite nice to be able to have these assurances in monomorphized functions.

[asomers]

Interesting. The reference to an "outer function" is what you would expect if [(); _] were a macro, but it isn't. I'm of the opinion that this is a compiler bug. It's an incorrect error message, if nothing else.

What you're doing looks basically like creating a variable-sized array, which of course isn't supported in Rust. I don't have any better ideas about how to do it, though. I suppose you could wrap the generic function in a macro, but that's ugly.

[nvzqz]

Not just in an array context; doing this won't work either:

const X: usize = size_of::<T>();

[regiontog]

I found that I can create unique identifiers for macros on stable using procedual macros. Maybe this could help remove some of your limitations?

Take a look at this crate for the implementation.

[nvzqz]

Is a separate crate still necessary for procedural macros or can they now be written within the same crate in which they're used?

Also, I consider bumping the minimum rustc version from 1.24.0 to 1.30.0 a breaking change, which means static_assertions going from 0.3.x to 0.4. I'm fine with that, but I'd like to see if I come up with some features to add or introduce other breaking changes before I go down that path.

Eventually I want to go down the const _ route since that feels cleaner to me. This means bumping rustc to whatever future version underscore_const_names gets stabilized in, which again is a breaking change. So I think at that point, I'd be comfortable with making this crate 1.0. However, that would make the only difference between 0.4 and 1.0 just be a rustc version bump, which is underwhelming. I can probably come up with ideas for other things to include in 1.0 but that's not a guarantee, and I don't want to create very basic features that don't add much value.

[regiontog]

AFAIK you still need a seperate crate, at least for crates with both #[macro_export] and #[proc_macro]. You could have a multi-crate setup I think e.g. like serde.

After testing multi-crates it seems like it works, however users of the #[macro_export] library have to explicitly depend on both crates. If not then you get a use of undeclared type or module "gensym" from the macro expansion.

[Centril]

I'm of the opinion that this is a compiler bug.

It's not. A const, type, ... item may not refer to the parameters of the outer scope unless the immediate scope is an impl. This is by design. Rust does not support lexically scoped type variables.

[carbotaniuman]

Why is it still not possible to use this outside of a function even though underscore_const_names has stabilized?