#closure #signature #higher-order #lifetime #hrtb

higher-order-closure

Allow function lifetime elision and explicit for<'a> annotations on closures

5 releases

0.0.5 Jan 30, 2022
0.0.4 Jan 30, 2022
0.0.3 Jan 30, 2022
0.0.2 Jan 30, 2022
0.0.1 Jan 30, 2022

#5 in #closure

Download history 39/week @ 2022-03-09 345/week @ 2022-03-16 227/week @ 2022-03-23 160/week @ 2022-03-30 110/week @ 2022-04-06 353/week @ 2022-04-13 71/week @ 2022-04-20 170/week @ 2022-04-27 91/week @ 2022-05-04 128/week @ 2022-05-11 63/week @ 2022-05-18 116/week @ 2022-05-25 178/week @ 2022-06-01 163/week @ 2022-06-08 75/week @ 2022-06-15 45/week @ 2022-06-22

514 downloads per month
Used in linotype

Zlib OR MIT OR Apache-2.0

14KB
94 lines

::higher-order-closure

Allow function lifetime elision and explicit for<'a> annotations on closures.

Repository Latest version Documentation MSRV unsafe forbidden License CI

Motivation / Rationale

See the RFC #3216: this crate is a Proof-of-Concept of the ideas laid out there[^1].

[^1]: with the exception of allowing elided lifetimes to have a meaning (chosen higher-order), which the RFC cannot do in order to be future-proof.

Click to expand

The following example fails to compile:

let f = |x| {
    let _: &i32 = x;
    x
};
{
    let scoped = 42;
    f(&scoped);
} // <- scoped dropped here.
f(&42);
error[E0597]: `scoped` does not live long enough
  --> src/lib.rs:10:7
   |
10 |     f(&scoped);
   |       ^^^^^^^ borrowed value does not live long enough
11 | } // <- scoped dropped here.
   | - `scoped` dropped here while still borrowed
12 | f(&42);
   | - borrow later used here

For more information about this error, try `rustc --explain E0597`.

Indeed, the signature of f in that example is that of:

impl Fn(&'inferred i32) -> &'inferred i32

wherein 'inferred represents some not yet known (to be inferred) but fixed lifetime.

Then,

{
    let scoped = 42;
    f(&scoped); // `'inferred` must "fit" into this borrow…
} // <- and thus can't span beyond this point.
f(&42) // And yet `'inferred` is used here as well => Error!

The solution, then, is to explicitly annotate the types involved in the closure signature, and more importantly, the lifetime "holes" / placeholders / parameters involved in that signature:

           // Rust sees this "hole" early enough in its compiler pass
           //                       to figure out that the closure signature
           // vv                    needs to be higher-order, **input-wise**
let f = |_x: &'_ i32| {
};
{
    let scoped = 42;
    f(&scoped);
}
f(&42);

This makes it so the input-side of the closure signature effectively gets to be higher-order. Instead of:

impl Fn(&'inferred_and_thus_fixed i32)...

for some outer inferred (and thus, fixed) lifetime 'inferred_and_thus_fixed, we now have:

impl for<'any> Fn(&'any i32)...

This works, but quid of returning borrows? (all while remaining higher-order)

Indeed, the following fails to compile:

let f = |x: &'_ i32| -> &'_ i32 {
    x // <- Error, does not live long enough.
};
error: lifetime may not live long enough
 --> src/lib.rs:5:5
  |
4 | let f = |x: &'_ i32| -> &'_ i32 {
  |             -           - let's call the lifetime of this reference `'2`
  |             |
  |             let's call the lifetime of this reference `'1`
5 |     x // <- Error, does not live long enough.
  |     ^ returning this value requires that `'1` must outlive `'2`

The reason for this is that "explicit lifetime 'holes' / placeholders become higher-order lifetime parameters in the closure signature" mechanism only works for the input-side of the signature.

The return side keeps using an inferred lifetime:

let f = /* for<'any> */ |x: &'any i32| -> &'inferred i32 {
    x // <- Error, does not live long enough (when `'any < 'inferred`)
};

we'd like for f there to have the fn(&'any i32) -> &'any i32 signature that functions get from the lifetime elision rules for function signatures.

Hence the reason for using this crate.

Examples

This crate provides a higher_order_closure! macro, with which one can properly annotate the closure signatures so that they become higher-order, featuring universally quantified ("forall" quantification, for<> in Rust) lifetimes:

#[macro_use]
extern crate higher_order_closure;

fn main ()
{
    let f = higher_order_closure! {
        for<'any> |x: &'any i32| -> &'any i32 {
            x
        }
    };
    {
        let local = 42;
        f(&local);
    }
    f(&42);
}

The lifetime elision rules of function signatures apply, which means the previous signature can even be simplified down to:

#[macro_use]
extern crate higher_order_closure;

fn main ()
{
    let f =
        higher_order_closure! {
            for<> |x: &'_ i32| -> &'_ i32 {
                x
            }
        }
    ;
}

or even:

#[macro_use]
extern crate higher_order_closure;

fn main ()
{
    let f =
        higher_order_closure! {
            |x: &'_ i32| -> &'_ i32 {
                x
            }
        }
    ;
}
  • Because of these lifetime elision in function signatures semantics, it is highly advisable that the elided_lifetimes_in_paths be, at the very least, on warn when using this macro:

    //! At the root of the `src/{lib,main}.rs`.
    #![warn(elided_lifetimes_in_paths)]
    

Extra features

Macro shorthand

The main macro is re-exported as a hrtb! shorthand, to allow inlining closures with less rightward drift:

#[macro_use]
extern crate higher_order_closure;

fn main ()
{
    let f = {
        let y = 42;
        hrtb!(for<'any> move |x: &'any i32| -> &'any i32 {
            println!("{y}");
            x
        })
    };
}

Outer generic parameters

Given how the macro internally works[^2], generic parameters "in scope" won't, by default, be available in the closure signature (similar to consts and nested function or type definitions).

In order to make them available, higher_order_signature! accepts an initial optional #![with<simple generics…>] parameter (or even #![with<simple generics…> where clauses] if the "simple shape" restrictions for the generic parameters are too restrictive).

"simple shaped" generics macro restrictions

The generics parameters inside #![with<…>] have to be of the form:

<
    'a, 'b : 'a, ...
    T, U : ?Sized + 'a + ::core::fmt::Debug, V, ...
>

Mainly:

  • at most one super-lifetime bound on each lifetime,

  • the super-bounds on the types must be exactly of the form:

    1. optional ?Sized,
    2. followed by an optional lifetime bound,
    3. followed by an optional trait bound.
    4. And nothing more. If you need more versatility, use the where clauses.

In practice, however, the bounds are seldom needed, since such generics are only used for the signature of the closure, not its body / implementation.

#[macro_use]
extern crate higher_order_closure;

use ::core::fmt::Display;

fn example<T : Display> (iter: impl IntoIterator<Item = (bool, T)>)
{
    let mb_display = higher_order_closure! {
        #![with<T>]
        |hide: bool, elem: &'_ T| -> &'_ dyn Display {
            if hide {
                &"<hidden>"
            } else {
                elem
            }
        }
    };
    for (hide, elem) in iter {
        println!("{}", mb_display(hide, &elem));
    }
}

fn main ()
{
    example([(false, 42), (true, 27), (false, 0)]);
}

Mixing higher-order lifetime parameters with inferred ones

Since inside higher_order_closure!, '_ has the semantics of lifetime elision for function signatures, it means that, by default, all the lifetime parameters appearing in such closure signatures will necessarily be higher-order.

In some more contrived or rare scenarios, this may be undesirable.

In that case, a nice way to work around that is by artificially introducing generic lifetime parameters as seen above, and using these newly introduced named lifetime parameters where the inferred lifetime would be desired.

#[macro_use]
extern crate higher_order_closure;

fn main ()
{
    let y = 42;
    let f = higher_order_closure! {
        #![with<'inferred>]
        for<'any> |x: &'any i32| -> (&'any i32, &'inferred i32) {
            (x, &y)
        }
    };
    let (&a, &b) = f(&27);
    assert_eq!(a + b, 42 + 27);
}

[^2]: it generates a "closure identity" helper function, with the desired higher-order signatures embedded as Fn bounds on its parameters, thus making it act as a "funnel" that only lets closure with the right signature pass through).

No runtime deps

Features

  • better-docs
  • ui-tests