remove static_assert

[steveklabnik]

rendered

[lilyball]

I don't see how CTFE obviates the desire for static_assert. If anything, I think CTFE makes static_assert much more powerful. CTFE allows us to calculate things at compile-time, and static_assert then lets you abort compilation if a given CTFE expression doesn't have the expected value.

That said, with CTFE, I'd expect static_assert to take the form of a macro static_assert!(condition[, message]).

[steveklabnik]

I was under the assumption that in a macro, you could just panic, or use something like https://github.com/huonw/compile_msg 's compile_error, which would outright replace the current static_assert.

[lilyball]

I don't know what the current thinking is on CTFE, but I'm a little skeptical that panicking is something that would be compatible with CTFE. Similarly, I don't know whether CTFE would be able to use syntax extensions like compile_error (as opposed to a syntax extension evaluating a CTFE, which is what my proposed static_assert!() would be).

[lilyball]

I suppose with CTFE, this static_assert!() macro could be provided by a crate, although it seems like a reasonable candidate to bake into the language (e.g. because the standard libraries might want to use it).

[pythonesque]

Yeah, I don't see how CTFE obviates it.

[P1start]

Ideally, assuming a CTFE where begin_unwind[_fmt] (and whatever other functions are needed for panic!) are const fns (throwing compile errors at runtime) then we could just use the regular assert! macro:

static _assert: () = assert!(FOO);
// Or, in a future Rust where statements are allowed at module-level:
assert!(FOO);

[lilyball]

The current form of static_assert is pretty weird. If we really aren't actually using it for anything, then 👍 on removing it. But with full-blown CTFE I'd want to have it reintroduced as a macro static_assert(cond[, message]) as I mentioned in my comment.

[Ericson2314]

+1

[pythonesque]

As someone who uses static_assert, I'd rather not see it removed without a more compelling reason (unless full CTFE can actually fully reintroduce it, but it's not clear to me that it would be able to; panic! diverges, after all).

[msiemens]

Typo: Should be [text](url), not [text][url]

[steveklabnik]

Thanks, I'll fix this if it's about to get merged.

[ben0x539]

alternative: provide a macro static_assert!(condition); that expands to #[static_assert] static new_identifier: bool = condition; and stabilize the macro but not the implementation :P

[nagisa]

The “want” to remove static_assert is intense here. It infects compiler all the way to trans. If we’re keeping it, I’d rather see this implemented within bounds of CTFE somehow.

[steveklabnik]

@nrc given that I'm the author, I probably shouldn't shepard?

[cmr]

@pythonesque what's the use case? When I implemented this I planned on usnig it to assert that sizes of things were the same but constant expressions aren't powerful enough.

As the original designer and implementor, 👍. This is a weird feature that is extremely limited.

[pythonesque]

@cmr I use it to assert that u8::MAX >= usize::MAX_BITS, which is required for correctness for my union find algorithm. With more powerful constant expressions (which I think have been accepted now) I would be able to assert more interesting things (particularly around things like correct alignments / gaps in data structures, where I rely on two representations matching up for memory safety).

[huonw]

I'm in favour of removing this and just using #[test]s until we get a nicer solution.

[ssokolow]

#[test]s are less reliable. In fact, at the moment, I've got an experimental project where, even if I create tests exactly as shown in the book, zero tests are found by cargo test. I haven't had time to figure out why yet.

[huonw]

No tests running at all seems like an orthogonal problem that should be solved anyway. :)

In any case, replacing the #[static_assert] const NAME: bool = expression; with #[test] fn name() { assert!(expression); } in a library should "just work" (unless one has explicitly opted for test = false in the [lib] section for that library).

[ssokolow]

Wait... the [lib] section? The project in question is currently a single-file binary project. Did the book neglect to mention #[test] only working in library projects?

[steveklabnik]

I distinctly remember writing a patch that explains this, but I cannot find it in the testing guide now :( Yes, rustdoc is only able to test libraries, not binaries.

[steveklabnik]

rust-lang/rust#24508 <- hmmmmmm

[steveklabnik]

oh! it's in http://doc.rust-lang.org/nightly/book/documentation.html#running-documentation-tests but not in the testing guide. that's an oversight. I'll fix that now.

[steveklabnik]

wooo reading github first thing in the morning. I should drink more coffee. It's true that _rustdoc_ cannot test binary crates, but #test doesn't work either? I'm not aware of that.

[huonw]

@ssokolow hm, a src/main.rs like the following gets compiled and tested when I run cargo test.

fn main() {
    println!("Hello, world!");
}

#[test]
fn foo() {}

(If you're still having problems I'm happy to help, but we should take it to stackoverflow or users rather than here.)

[oli-obk]

Note: I have created a replacement lint-plugin: https://crates.io/crates/static_assert

So now there's zero reason for this to stay in the compiler.

Usage example:

#![feature(plugin, custom_attribute)]
#![plugin(static_assert_)]

fn main() {
    #[static_assert_]
    const TEST: bool = false;
}

I haven't figured out yet how to silence the "unused" warnings, but I'm working on it. Also the next step is to create a macro that automatically creates a temporary constant + adds the attribute (maybe even the #[allow(dead_code)] attribute) so it can be used just like the C++ static_assert on any constant expression

Note2: once static_assert isn't a builtin-attribute anymore, i can remove the trailing underscore

Edit: I just noticed the original was on static fields not on constants... so it's not a drop-in replacement. sorry about that.

[SimonSapin]

FWIW, I’ve replaced one usage of #[static_assert] by #[cfg(target_endian = "little")] so that the function would be missing on big-endian (and compilation would fail). The other usage in that library I’ve made a debug_assert!:

servo/string-cache#85

[nikomatsakis]

Hear ye, hear ye. This RFC is now in final comment period until June 2nd.

[steveklabnik]

🎊

[ssokolow]

@SimonSapin Unless I missed an announcement that compiler plugins will work in stable channel, I'd hardly call it zero reason.

[huonw]

@ssokolow NB. #[static_assert] also doesn't work in the stable channel, so there's not much loss.

[ssokolow]

Ahh. That makes sense. It can get a bit difficult to keep track of which features I'm not currently using are stable.

[SimonSapin]

@ssokolow I don’t see how my comment is related to plugins. Or did you mean @oli-obk ?

[nikomatsakis]

I think I agree with @huonw that, in practice, adding a #[test] (or a plugin, if you're so inclined) is better than keeping this feature around. I'm nervous about keeping random features around in unstable fashion for an indefinite period of time without strong motivation -- I don't want to allow de facto dependencies to form that prove troublesome to change. cough macro rules cough

[pnkfelix]

I'm fine with removing #[static_assert]; when we do, we should probably open a corresponding RFC issue at the same time stating that we'd like to see this functionality provided in some manner in the future (in the abstract; its just an issue, not an RFC, after all). At least, that's my take-away from the comment thread.

[nikomatsakis]

Based on the comment thread, we've decided to accept this RFC to remove static_assert. The major arguments in favor of removing it are that it complicates many parts of the compiler tool-chain and we know we would never stabilize the current design. Existing uses can be converted into tests or via a plugin. Per @pnkfelix's comment, we've created issue #1146.

--- Language Subteam

[steveklabnik]

I will rebase my implementation PR later today or tomorrow.