`extended_key_value_attributes` cannot be enabled conditionally

[jyn514]

I tried this code:

#![cfg_attr(
    any(),
    feature(extended_key_value_attributes),
    doc = include_str!("../README.md")
)]

I expected to see this happen: The compiler gives no error, because doc = include_str! should only be compiled if the predicate is true.

Instead, this happened: The compiler gives an error that extended_key_value_attributes is unstable.

I guess this happens because the compiler has to parse the attribute even if the cfg is false? @petrochenkov is giving an error here intentional?

Meta

rustc --version: 1.52.0-nightly (2021-03-01 4f20caa)

cc #78835
Originally posted by @daxpedda in #82539 (comment)

[petrochenkov]

fn process_cfg_attr parses all "attribute items" inside cfg_attr first, and only then checks the predicate and possibly throws the attribute items away. (extended_key_value_attributes feature gate is reported during such parsing.)

It's possible to check the predicate first and parse the rest of cfg_attr arguments only if the predicate is true, but in that case cfg_attr will also accept arbitrary garbage like #[cfg_attr(FALSE, a b < 7 . / asjd &)].

[petrochenkov]

The workaround is to add one more layer of cfg_attr:

#![cfg_attr(FALSE, cfg_attr(TRUE, doc = include_str!("../README.md")))] // OK

[jyn514]

The workaround is to add one more layer of cfg_attr:

#![cfg_attr(FALSE, cfg_attr(TRUE, doc = include_str!("../README.md")))] // OK

Thanks! Would it be possible to add that as a structured suggestion in the parser? I can take a look at this weekend if you think it's a good idea :)

[petrochenkov]

I don't think it's a reasonable investment of time, especially given that we'll be able to stabilize extended_key_value_attributes pretty soon.

[pksunkara]

@daxpedda In what scenario are you checking for the feature existence?

In clap, I do:

#![cfg_attr(feature = "doc", feature(external_doc))]
#![cfg_attr(feature = "doc", doc(include = "../README.md"))]

[jyn514]

@pksunkara doc(include) does not use this feature; see the linked thread: #82539 (comment). doc(include) is only relevant in that it should be replaced with this feature.

[pksunkara]

My bad, I meant to paste the updated one as a probable workaround for people who run into the above bug when they don't need to check the feature existence

#![cfg_attr(feature = "doc", feature(extended_key_value_attributes))]
#![cfg_attr(feature = "doc", doc = include_str!("../README.md"))]

[jyn514]

@pksunkara that has the same issue, when the doc feature isn't passed it still gives an error that the feature is unstable: https://play.rust-lang.org/?version=stable&mode=debug&edition=2018&gist=d6f0b2a9d584e889d14fd1688827d510

[Nemo157]

So, my understanding of the workaround is that you would write it like this:

#![cfg_attr(feature = "doc", feature(extended_key_value_attributes))]
#![cfg_attr(feature = "doc", cfg_attr(feature = "doc", doc = include_str!("../README.md")))]

This compiles all the way back to rustc 1.18.0 (03fc9d622 2017-06-06).

(edit: Given that this is related to the behaviour of existing compilers, I don't see that there is any change that could happen in new compilers to mitigate it, other than changing the syntax, which would sort of defeat the point of just reusing the existing syntax).