Skip to content

Latest commit

 

History

History
188 lines (137 loc) · 7.68 KB

0385-module-system-cleanup.md

File metadata and controls

188 lines (137 loc) · 7.68 KB

Module system cleanups

Summary

  • Lift the hard ordering restriction between extern crate, use and other items.
  • Allow pub extern crate as opposed to only private ones.
  • Allow extern crate in blocks/functions, and not just in modules.

Motivation

The main motivation is consistency and simplicity: None of the changes proposed here change the module system in any meaningful way, they just remove weird forbidden corner cases that are all already possible to express today with workarounds.

Thus, they make it easier to learn the system for beginners, and easier to for developers to evolve their module hierarchies

Lifting the ordering restriction between extern crate, use and other items.

Currently, certain items need to be written in a fixed order: First all extern crate, then all use and then all other items. This has historically reasons, due to the older, more complex resolution algorithm, which included that shadowing was allowed between those items in that order, and usability reasons, as it makes it easy to locate imports and library dependencies.

However, after RFC 50 got accepted there is only ever one item name in scope from any given source so the historical "hard" reasons loose validity: Any resolution algorithm that used to first process all extern crate, then all use and then all items can still do so, it just has to filter out the relevant items from the whole module body, rather then from sequential sections of it. And any usability reasons for keeping the order can be better addressed with conventions and lints, rather than hard parser rules.

(The exception here are the special cased prelude, and globs and macros, which are feature gated and out of scope for this proposal)

As it is, today the ordering rule is a unnecessary complication, as it routinely causes beginner to stumble over things like this:

mod foo;
use foo::bar; // ERROR: Imports have to precede items

In addition, it doesn't even prevent certain patterns, as it is possible to work around the order restriction by using a submodule:

struct Foo;
// One of many ways to expose the crate out of order:
mod bar { extern crate bar; pub use self::bar::x; pub use self::bar::y; ... }

Which with this RFC implemented would be identical to

struct Foo;
extern crate bar;

Another use case are item macros/attributes that want to automatically include their their crate dependencies. This is possible by having the macro expand to an item that links to the needed crate, eg like this:

#[my_attribute]
struct UserType;

Expands to:

struct UserType;
extern crate "MyCrate" as <gensymb>
impl <gensymb>::MyTrait for UserType { ... }

With the order restriction still in place, this requires the sub module workaround, which is unnecessary verbose.

As an example, gfx-rs currently employs this strategy.

Allow pub extern crate as opposed to only private ones.

extern crate semantically is somewhere between useing a module, and declaring one with mod, and is identical to both as far as as the module path to it is considered. As such, its surprising that its not possible to declare a extern crate as public, even though you can still make it so with an reexport:

mod foo {
    extern crate "bar" as bar_;
    pub use bar_ as bar;
}

While its generally not necessary to export a extern library directly, the need for it does arise occasionally during refactorings of huge crate collections, generally if a public module gets turned into its own crate.

As an example,the author recalls stumbling over it during a refactoring of gfx-rs.

Allow extern crate in blocks/functions, and not just in modules.

Similar to the point above, its currently possible to both import and declare a module in a block expression or function body, but not to link to an library:

fn foo() {
    let x = {
        extern crate qux; // ERROR: Extern crate not allowed here
        use bar::baz;     // OK
        mod bar { ... };  // OK
        qux::foo()
    };
}

This is again a unnecessary restriction considering that you can declare modules and imports there, and thus can make an extern library reachable at that point:

fn foo() {
    let x = {
        mod qux { extern crate "qux" as qux_; pub use self::qux_ as qux; }
        qux::foo()
    };
}

This again benefits macros and gives the developer the power to place external dependencies only needed for a single function lexically near it.

General benefits

In general, the simplification and freedom added by these changes would positively effect the docs of Rusts module system (which is already often regarded as too complex by outsiders), and possibly admit other simplifications or RFCs based on the now-equality of view items and items in the module system.

(As an example, the author is considering an RFC about merging the use and type features; by lifting the ordering restriction they become more similar and thus more redundant)

This also does not have to be a 1.0 feature, as it is entirely backwards compatible to implement, and strictly allows more programs to compile than before. However, as alluded to above it might be a good idea for 1.0 regardless

Detailed design

  • Remove the ordering restriction from resolve
  • If necessary, change resolve to look in the whole scope block for view items, not just in a prefix of it.
  • Make pub extern crate parse and teach privacy about it
  • Allow extern crate view items in blocks

Drawbacks

  • The source of names in scope might be harder to track down
  • Similarly, it might become confusing to see when a library dependency exist.

However, these issues already exist today in one form or another, and can be addressed by proper docs that make library dependencies clear, and by the fact that definitions are generally greppable in a file.

Alternatives

As this just cleans up a few aspects of the module system, there isn't really an alternative apart from not or only partially implementing it.

By not implementing this proposal, the module system remains more complex for the user than necessary.

Unresolved questions

  • Inner attributes occupy the same syntactic space as items and view items, and are currently also forced into a given order by needing to be written first. This is also potentially confusing or restrictive for the same reasons as for the view items mentioned above, especially in regard to the build-in crate attributes, and has one big issue: It is currently not possible to load a syntax extension that provides an crate-level attribute, as with the current macro system this would have to be written like this:

    #[phase(plugin)]
    extern crate mycrate;
    #![myattr]
    

    Which is impossible to write due to the ordering restriction. However, as attributes and the macro system are also not finalized, this has not been included in this RFC directly.

  • This RFC does also explicitly not talk about wildcard imports and macros in regard to resolution, as those are feature gated today and likely subject to change. In any case, it seems unlikely that they will conflict with the changes proposed here, as macros would likely follow the same module system rules where possible, and wildcard imports would either be removed, or allowed in a way that doesn't conflict with explicitly imported names to prevent compilation errors on upstream library changes (new public item may not conflict with downstream items).