Rust: is `#[cfg]` applicable to *some* non-stmt expressions?

[pnkfelix]

Consider the following:

http://is.gd/28Q5H5

#![feature(stmt_expr_attributes)]

fn foo(_x: &[u32]) { }

fn main() {
    foo(&[1, #[cfg(not(now))] 2, 3])
}

Currently this is accepted (see https://github.com/rust-lang/rust/pull/29850) but RFC 16 as written does not clearly say whether this is acceptable or not (see https://github.com/rust-lang/rfcs/blob/master/text/0016-more-attributes.md#cfg )

[pnkfelix]

cc @Kimundi

[pnkfelix]

Lang team discussed this.

Opinions were mixed about whether this should be allowed or not.

But the more fundamental conclusion is that the decision here should go through the RFC process (merely as an amendment to RFC 16), rather than just having the lang team make a unilateral decision; we want to get input from the community about what is best here.

Note: It would be good to keep in mind that some use cases, such as function calls of the form

bar(1, #[cfg(no)] 2, 3)

are at the very least difficult to try to use, because one cannot write

fn bar(a: u32, #[cfg(no)] b: u32, c: u32) { ... }

[tbu-]

I'm against such fine-grained cfg attributes. I think it'll produce unreadable code, similar to many #ifdefs inside the same term.

if(
#ifdef A
a
#endif
#if defined(A) && defined(B)
&&
#endif
#if defined(B)
b
#endif
#if defined(A) || defined(B)
&&
#endif
c)
{}

Compare that to

#[cfg(A)]      fn a() -> bool { a }
#[cfg(not(A))] fn a() -> bool { true }

#[cfg(B)]      fn b() -> bool { b }
#[cfg(not(B))] fn b() -> bool { true }

if a() && b() && c {
}

[pnkfelix]

cc #15701 ... this is effectively an unresolved question that needs to be added to that tracking issue before we stabilize this feature.

[Kimundi]

So, obviously unreadable messes like what @tbu- showed should be considered unidiomatic in any case, but I can see the use cases where just plain item-level cfgs won't cut it.

One of the motivating examples that lead to me implementing RFC #16 was the use case of a huge struct with many #[cfg]-guarded fields, and the desire to be able to write a constructor function for it without having combinatorial explosion of the possible variations.

While what I implemented did not fully reach that state due to things like struct constructors like Foo { #[cfg] a: 5 } or as you said function argument lists not accepting attributes, it does allow for this:

#![feature(stmt_expr_attributes)]

struct Foo(
    u8,
    #[cfg(foo)] u16
);

fn main() {
    let x = Foo(1, #[cfg(foo)] 2);
}

As well as compile-time-cfg switching code like

let x = (#[cfg(a)] foo(), #[cfg(not(a))] bar()).0;

(Though I guess the latter might be replicable with blocks and assignment statements)

In my opinion, the fact that we already support attributes on struct fields and match arms might indicate that there is a desire to have them at more specific locations than just the "top level" items, types and statements, so maybe it would make sense to investigate whether it can be extend to places like patterns or argument lists so that things like #[cfg] can become more usable there.

---

Also, the way I see it all more complicated/unreadable constructions you can build with this would end up behind a macro anyway. For example , have a cfg-match macro:

macro_rules! cfg_match {
    (_ => $b:block) => {
        $b
    };
    ($cfg:meta => $b:block $($t:tt)*) => {
        (
            #[cfg($cfg)] $b,
            #[cfg(not($cfg))] (cfg_match!($($t)*)),
        ).0
    };
    () => {
        {
            panic!("Conditional code not supported under this configuration");
        }
    }
}

let x = cfg_match! {
    unix => { foo() }
    windows => { bar() }
};

[tbu-]

That should be possible today, if I understand it correctly:

let x;
#[cfg(windows)]
{ x = 5; }
#[cfg(unix)]
{ x = 3; }

or even

#[cfg(windows)]
let x = 3;
#[cfg(unix)]
let x = 5;

[bluss]

@tbu- That needs the same feature (stmt_expr_attributes) to compile

[plietar]

There's a usecase for #[cfg] on expressions I find useful :

let backends = [
    #[cfg(feature = "foo-backend")] foo_backend,
    #[cfg(feature = "bar-backend")] bar_backend,
];

allowing multiple backends to be enabled at compile time using cargo features, and choosing one at runtime.
The alternative is to list all 2^N combinations, or use some macro magic to generate them (I've had to resort to this in order to stay on stable).

I guess you could consider this as a #[cfg] on a list element rather than an expression.

I'm more skeptical about other uses of it, especially on function calls.
if cfg!(...) { 3 } else { 5 } can be used for simpler cases (I assume this gets optimized out at compile time)

[brson]

This looks like the only thing blocking stmt_expr_attributes from stabilizing.

[brson]

Perhaps we can turn this behavior off and move forward with stmt_expr_attributes stabilization @rust-lang/lang? Readress this issue later.

[Kimundi]

Turn off in the sense of removal, or in the sense of adding a feature for it specifically?

[nrc]

I'm in favour of stabilising statement attributes and leaving expression ones under the existing feature flag (probably the easiest solution for users and the compiler, I'd also be happy with a new feature flag or removing expression attributes entirely, but either option seems less friendly).

[Mark-Simulacrum]

So I'm not really clear what this issue is addressing. I think it's about attributes working on some expression and not others... but I don't see representative examples. Could someone summarize the current situation (@nrc?)?

[durka]

Right now [1, #[cfg(unix)] 2, 3] is accepted on stable with no feature flag, so I'm not sure there is anything left to talk about here.

[nikomatsakis]

It's so clearly useful... I think I'd be against trying to remove it. But I'd like to know the current status too.

[pnkfelix]

yes it appears that #36995 may have (accidentally?) stabilized it for list elements, which was the main case this issue was discussing... at this point I'm not sure we can afford to remove it...

[pnkfelix]

So, as noted above, it appears that this (play):

fn foo(_x: &[u32]) { }

fn main() {
    foo(&[1, #[cfg(not(now))] 2, 3])
}

is now accepted on stable.

So regardless of whatever resoluton I was hoping to reach via discussion, it appears the matter has been resolved independently of the discussion here.

(Note to self: Maybe in the future when I discover oddities like this, I should immediately add compile-fail tests that are lacking the feature flag to ensure they don't get accidentally stabilized. At least, my interpretation of @nrc's stated position leads me to think that they did not intend to stabilize this aspect of attributes)

• At least, this helps give me perspective on why some projects eagerly add test cases that are expected to fail (and even check them into the repository, tagged as expected failures) before a feature is implemented