Add a default-on std feature

[alexcrichton]

This adds a std Cargo feature which disables #![no_std] builds of libc, but
is enabled by default. The library will currently continue to link to the
standard library to maintain backwards compatibility with the 0.2 series and
older Rust compilers for now, but this default can possible be changed in the
future.

[alexcrichton]

r? @brson

[steveklabnik]

🙌🙌🙌🙌🙌🙌🙌🙌🙌

[bluss]

cargo features are additive, so this does not work under the cargo scheme. If any dependency activates "no-std", the whole project will have it (for libc).

cargo features can't turn things off. Could this instead use a positive feature "use-std" that is enabled by default? The unioning of cargo features will work correctly then.

[alexcrichton]

For this library specifically I think that this is an additive feature (because the API will not change based on it or not) as it's adding a feature, but for libraries in general I think that they're going to provide a more full suite of functionality with std than without.

Along those lines I'd be interested in starting to pioneer a convention for libraries with no_std support. An example of this is that the log crate now has a use_std feature. I wouldn't mind following suit there as well.

I also think that it makes more sense to "turn off the standard library" than "enable turning off the standard library", e.g. disable-features = false seems to jive better than features = ["no-std"].

cc @rust-lang/libs, this is at the very least a backwards compatibility hazard as once we've decided a name/interface for this it can't be changed. Thoughts?

[alexcrichton]

Ugh cross-org notifications, see my above message: ^^

cc @aturon, @brson, @sfackler, @huonw, @Kimundi, @BurntSushi, @Gankro

[bluss]

My bad! I did actually assume it would change its functionality depending on std / no std (this is the use case I've been thinking about). Very interested in that we have a common way to do this across crates.

[brson]

If I understand @bluss's concern, it 'no-std' is a convention, then it will be impossible to selectively turn off for specific crates; activating the cargo 'no-std' feature anywhere will turn it off for all crates. Is that right?

I also think that it makes more sense to "turn off the standard library" than "enable turning off the standard library", e.g. disable-features = false seems to jive better than features = ["no-std"].

I don't understand what this comment means. What does 'disable-features = false' do?

Could this instead use a positive feature "use-std" that is enabled by default? The unioning of cargo features will work correctly then.

Can you spell out why this is? Also, what do you do to disable a feature, either project-wide or per-crate?

[bluss]

You have a library D version "1" which supports core and/or std. D has the feature flag "use-std".

We have application A, with dependencies B and C.

• B depends on D version "1" with "use-std"
• C depends on D version "1" with no features

To build A, cargo has to build B, C, D. To build D, it unions all the features needed, so it builds D with "use-std". For this reason we can't use features to subtract, we must only add, since cargo treats them as additive (adding a feature flag must not break those that depend on the crate as it was without the feature flag).

Having default features is like having all deps specifying they depend on the default features automatically.

Specifying features is something you do in the dependencies section. For example.

[dependencies]
D = { version = "1", features = ["use-std"] }
E = { version = "1", default-features = false }

Edit: Note that since libc neither "adds" nor "subtracts", I don't think it matters for this package..

An example of adding would be to implement the trait "Read" when possible (when std is available).

[alexcrichton]

it will be impossible to selectively turn off for specific crates; activating the cargo 'no-std' feature anywhere will turn it off for all crates

Yeah this is correct

What does 'disable-features = false' do?

Oops, that actually does nothing, I meant default-features = false which is actually a Cargo thing!

---

I that I would want to concretely propose that std is the feature named used to activate linking to the standard library (although I'm also fine with bikeshedding the name). Although libc is not really additive or subtractive on this, it seems like a much more plausible convention for external crates to have this behavior (enabling std enables more features).

[Amanieu]

In this case the only reason we need a Cargo feature is to allow libc to work on older Rust versions that don't support no_std. If we didn't have to support older Rust version then we could just unconditionally make libc no_std.

[Stebalien]

Needs to be updated to yes-std.

[alexcrichton]

@Amanieu yes that's correct, it's currently expected that this crate works with versions like Rust 1.5 and below, so switching the defaults would be a bit of a breaking change for libc.

[tbu-]

Instead of having default-features = false, it would be better to selectively turn off default features, because otherwise, adding more default features would be a breaking change. Is that possible with Cargo or is there an issue for it?

[bluss]

That's a troubling thought. It seems cargo needs a new feature for this.

[alexcrichton]

@tbu- ah yes that's a good point, but it also only applies in the case where existing APIs are hidden behind a feature. If a default-on feature is used to gate just-newly-added APIs then it wouldn't cause breakage.

There is currently not a Cargo issue for disabling a feature (or just disabling a default feature), but it seems like it may be interesting to add.

[brson]

It's not clear to me we've arrived at the solution. This is a big convention to set. Is everybody comfortable with this: 'std' is the feature that enables linking to std, enabled by default, and you turn on no_std by disabling all default features?

[alexcrichton]

Yeah I agree, I would like to talk about this during libs triage.

[aturon]

I'm having trouble following the thrust of this thread. On the one hand, we seem to be talking about a hack in libc to get backwards-compatibility with older versions of rustc. On the other hand, we seem to be talking about a general convention. Is the latter also just about backcompat?

I feel like we continually run into issues like this and we should seriously consider a first-class way to straddle versions via cfg.

[alexcrichton]

Yes, there are two issues here, but the question for the libs team (in my mind) is about the feature naming/semantics itself.

To recap, the first issue is that it would be a breaking change right now to make libc #![no_std] by default because older compilers will not accept this.

The second issue is that there are many use cases throughout our crates of having a "libcore mode". For example the log crate recently picked this up. The idea is that you can provide what is a likely stripped down set of functionality without std, but you can still provide the core set of utilities. I'm thinking that we want to have features called std which are enabled by default to express this ability because:

• @bluss is right in that features are additive in Cargo, so having a feature to remove functionality could cause complications when compiling
• This provides a uniform interface of default-features = false to give the bare bones functionality of a crate.

As @tbu- mentioned, however, there is a backwards-compatibility hazard. Once the set of default features have been decided, then the existing API of the crate must always be available with default-features = false. Which is to say that over time we can add more default-on features for new APIs, but we cannot take existing functionality and move it behind a default feature.

---

@aturon I agree that straddling versions with some sort of cfg to just always support #![no_std] would be best, but... we don't have it :(

[Gankro]

@alexcrichton just have the build script detect the rustc version, and branch on that!

cargo/1.2 rustc/1.5 (llvm/3.8, like clang)

[alexcrichton]

The libs team discussed this during triage today and the conclusion was to take this strategy with the name "use_std" (to mirror the log crate). I've pushed a commit with these changes and I'll merge when CI is green!