RFC: Stabilize std::prelude

[alexcrichton]

Stabilize the std::prelude module by removing some of the less commonly used
functionality of it.

Rendered

[alexcrichton]

cc #497, an RFC specifically about Ordering, but reaching the same conclusion.

[SimonSapin]

If user-defined preludes are allowed in some form, it is also unclear about how this would impact the inclusion of reexports in the standard library's prelude in some form.

Assuming that user-defined preludes need to be opted into, the syntax for that could be use my_prelude::*; and require no new language feature. Or am I missing something?

[aturon]

@SimonSapin

That doesn't quite give you the effect of the prelude, because it applies only to the current scope. There are some proposals in the works for allowing you to specify a custom prelude in your crate root that will be wholly imported into all scopes within the crate.

A feature like this would make it relatively easy to make "changes" to the std prelude without breakage by introducing separate versions, e.g. std::prelude::v1_2 (where the standard prelude would be std::prelude::v1_0).

[steveklabnik]

I am 👍 for this.

[frewsxcv]

mod

Is this a typo?

[sfackler]

Nope, it imports the Ordering enum, like cmp::{mod, Ordering} would import cmp and cmp::Ordering.

[frewsxcv]

woah awesome, TIL

[kmcallister]

// The prelude will be responsible for providing unicode-respecting methods on
// primitives while requiring that ascii-specific manipulation is imported
// manually.

I'm skeptical of providing Unicode-respecting methods in the prelude, because the operations are so complicated, with locale-dependence and many choices to make. It might be better to require the user to specify, via imports, which flavor of operations they're using.

There are a lot of formats (like HTTP and HTML) that use machine-interpreted text with ASCII-case-insensitive matching. This is a subtly different data type from "text written by humans for human consumption", but it's not worth forking &str and String. For this reason, you're just as likely to want ASCII ops as the (slow and complicated and locale-dependent) Unicode case folding. So I don't think either one should be in the prelude-imported extension traits.

[kmcallister]

// While currently present in the prelude, these traits do not need to be in
// scope to use the language syntax associated with each trait. These traits are
// also only rarely used in bounds on generics and are consequently
// predominately used for `impl` blocks. Due to this lack of need to be included
// into all modules in Rust, these traits are all removed from the prelude.
...
pub use ops::{Deref, DerefMut};

Because of rust-lang/rust#15609, it is sometimes necessary to introduce an explicit deref_mut() call. Moving the trait out of the prelude will increase the effort to work around this bug. That said it's not a situation I run into a whole lot, and it would be better to just fix the bug :)

[SimonSapin]

@kmcallister

For this reason, you're just as likely to want ASCII ops as the (slow and complicated and locale-dependent) Unicode case folding.

I disagree with "just as likely". You and I use ASCII-only case-insensitive matching routinely, but it’s fairly domain-specific: it only makes senses in a namespace like HTML element names where every "valid" name is ASCII, it’s just wrong on full Unicode text. It definitely should not be in the prelude.

As to Unicode case folding, some variations of it are locale-independent. Sure, they’re also wrong for the a few cases like the dotted/dotless iİıI in Turkish, but they’re much less wrong than ASCII-only. (SpecialCasing.txt doesn’t have that many language-sensitive mappings. CLDR might have more, but I haven’t figured out how that works yet.)

I think that some people will want to "just do XYZ" and not care about Unicode subtleties, much less being forced to specify a content language (which might just end up hard-coded to en-US). So we should give them a default, and it should be the "less wrong" one.

[alexcrichton]

@kmcallister right now we are quite strict about str being UTF-8 at all times, and I think it may be odd if we don't offer basically any functionality on it by default due to worries about unicode. The functionality I'm thinking of are things like uppercase, lowercase, alphanuemrics, and whitespace detection to get functions like words(), is_alphanumeric, to_uppercase, etc. I definitely agree that unicode can be super complicated about this, but as @SimonSapin mentioned, almost everyone expects something of this form to work.

I expect we will very clearly document precisely what is_whitespace means (it's just doing a table lookup) and more flavorful functionality (like locale-specific mappings) can evolve in an external crate.

[sinistersnare]

nota => not a

[alexcrichton]

The discussion here seems to be slowing down, so I may try to look to merge this soon.

ping @kmcallister about unicode in the prelude
ping @aturon do you feel confident enough to propose std::prelude::v1 here?
cc @brson, @wycats (curious as to more thoughts)

[aturon]

@alexcrichton

@aturon do you feel confident enough to propose std::prelude::v1 here?

I think we should do it, yes (probably v1_0). It won't be visible unless you go looking for it, and it leaves space to add new prelude versions without them being part of the normal prelude glob import.

Broader context for others reading: we hope to eventually allow you to specify a custom prelude at the top of your crate (which would be glob imported everywhere, just as std::prelude normally is). This is useful both for non-std libraries providing preludes that work as nicely std::prelude, and to allow for breaking evolution of std::prelude in the future -- you could opt in to a new version.

If we don't introduce internal namespacing like this, there would not be a clear place for new prelude versions to go.

On the other hand, if we never end up evolving the prelude in this way, having v1_0 nested doesn't seem too harmful, given that you'd never write it in actual code.

[sfackler]

I'm sitting on a proposal that should hopefully eliminate much of the need for third party preludes and the addition of anything to the libstd prelude, but sticking the current prelude in a v1_0 module seems pretty reasonable.

[aturon]

@sfackler Interesting! FWIW, the plan I had in mind was roughly to add a #[prelude] attribute that can be used at crate root on any item defining a module or bringing one into scope, e.g. #[prelude] use std::prelude::v1_2. If the attribute is used anywhere, the normal prelude is not injected; instead, the contents of all marked modules are glob imported into every namespace.

Curious to hear what you have in mind!

[sfackler]

The tl;dr is to provide a way of marking a trait impl as always being in scope (e.g. inherent impl Foo for Bar). A significant portion of items in the current prelude are traits, and most of the utility of those reexports is not having the trait name in scope, but the methods defined on the trait. Many of those are basically never used as type bounds, like StrPrelude for example. Other traits like Iterator are used as trait bounds, but if they weren't in the prelude, they'd be imported for their methods far more often than they'd be imported for the name itself, if that makes sense.

In any case, it's a pretty clear post-1.0 improvement, so I'm not planning on writing it up formally until after the initial release.

[aturon]

@sfackler Ah ok, we've discussed that in the past. I don't think that completely covers what preludes offer, but it would certainly help!

[alexcrichton]

I've pushed an update which recommends std::prelude::v1 containing the reexports instead of std::prelude (invisible change to all current users essentially).

[Gankra]

Collections Reform v1 adds repeat to the prelude, which I think was never impl'd? Does this RFC officially override that decision? Note that v2's ergonomics partially assumed this would be something very easy to use (in deprecating certain Vec methods).

[aturon]

@gankro I think for the alpha, we're going to leave repeat and friends out; if this turns out to be a problem, we can add it back in for the beta.

[aturon]

While there has been surprisingly little discussion on this RFC thread, it's been brought up in the meeting and seen fairly widely in the community, and we need to move forward before the alpha. The core team has reviewed the RFC and is happy with the direction it's taking. We will also have an opportunity to tweak the exact contents before beta/1.0.

Tracking issue