10 releases

0.0.10 Mar 31, 2020
0.0.9 Feb 19, 2020
0.0.6 Nov 14, 2019

#11 in #next-gen

34 downloads per month

Custom license

8KB
179 lines

::next_gen

Safe generators on stable Rust.

Repository Latest version Documentation MSRV License CI

Examples

Reimplementing a range iterator

use ::next_gen::prelude::*;

#[generator(yield(u8))]
fn range (start: u8, end: u8)
{
    let mut current = start;
    while current < end {
        yield_!(current);
        current += 1;
    }
}

mk_gen!(let generator = range(3, 10));
assert_eq!(
    generator.collect::<Vec<_>>(),
    (3.. 10).collect::<Vec<_>>(),
);

Implementing an iterator over prime numbers using the sieve of Eratosthenes

use ::next_gen::prelude::*;

enum NeverSome {}

/// Generator over all the primes less or equal to `up_to`.
#[generator(yield(usize))]
fn primes_up_to (up_to: usize)
  -> Option<NeverSome>
{
    if up_to < 2 { return None; }
    let mut sieve = vec![true; up_to.checked_add(1).expect("Overflow")];
    let mut p: usize = 1;
    loop {
        p += 1 + sieve
                    .get(p + 1..)?
                    .iter()
                    .position(|&is_prime| is_prime)?
        ;
        yield_!(p);
        let p2 = if let Some(p2) = p.checked_mul(p) { p2 } else {
            continue
        };
        if p2 >= up_to { continue; }
        sieve[p2..]
            .iter_mut()
            .step_by(p)
            .for_each(|is_prime| *is_prime = false)
        ;
    }
}

mk_gen!(let primes = primes_up_to(10_000));
for prime in primes {
    assert!(
        (2_usize..)
            .take_while(|&n| n.saturating_mul(n) <= prime)
            .all(|n| prime % n != 0)
    );
}

Defining an iterator with self-borrowed state

This is surely the most useful feature of a generator.

Consider, for instance, the following problem:

fn iter_locked (elems: &'_ Mutex<Set<i32>>)
  -> impl '_ + Iterator<Item = i32>

Miserable attempts without generators

No matter how hard you try, without using unsafe, or some other unsafe-using self-referential library/tool, you won't be able to feature such a signature!

  • The following fails:

    fn iter_locked (mutexed_elems: &'_ Mutex<Set<i32>>)
      -> impl '_ + Iterator<Item = i32>
    {
        ::std::iter::from_fn({
            let locked_elems = mutexed_elems.lock().unwrap();
            let mut elems = locked_elems.iter().copied();
            move || {
                // let _ = locked_elems;
                elems.next()
            } // Error, borrowed `locked_elems` is not captured and is thus dropped!
        })
    }
    
    error[E0515]: cannot return value referencing local variable `locked_elems`
      --> src/lib.rs:122:5
       |
    11 | /     ::std::iter::from_fn({
    12 | |         let locked_elems = mutexed_elems.lock().unwrap();
    13 | |         let mut elems = locked_elems.iter().copied();
       | |                         ------------------- `locked_elems` is borrowed here
    14 | |         move || {
    ...  |
    17 | |         } // Error, borrowed `locked_elems` is not captured and is thus dropped!
    18 | |     })
       | |______^ returns a value referencing data owned by the current function
       |
       = help: use `.collect()` to allocate the iterator
    
  • as well as this:

    fn iter_locked (mutexed_elems: &'_ Mutex<Set<i32>>)
      -> impl '_ + Iterator<Item = i32>
    {
        ::std::iter::from_fn({
            let locked_elems = mutexed_elems.lock().unwrap();
            let mut elems = locked_elems.iter().copied();
            move || {
                let _ = &locked_elems; // ensure `locked_elems` is captured (and thus moved)
                elems.next() // Error, can't use borrow of moved value!
            }
        })
    }
    
    error[E0515]: cannot return value referencing local variable `locked_elems`
      --> src/lib.rs:144:5
       |
    11 | /     ::std::iter::from_fn({
    12 | |         let locked_elems = mutexed_elems.lock().unwrap();
    13 | |         let mut elems = locked_elems.iter().copied();
       | |                         ------------------- `locked_elems` is borrowed here
    14 | |         move || {
    ...  |
    17 | |         }
    18 | |     })
       | |______^ returns a value referencing data owned by the current function
       |
       = help: use `.collect()` to allocate the iterator
    
    error[E0505]: cannot move out of `locked_elems` because it is borrowed
      --> src/lib.rs:147:9
       |
    8  |   fn iter_locked (mutexed_elems: &'_ Mutex<Set<i32>>)
       |                                  - let's call the lifetime of this reference `'1`
    ...
    11 | /     ::std::iter::from_fn({
    12 | |         let locked_elems = mutexed_elems.lock().unwrap();
    13 | |         let mut elems = locked_elems.iter().copied();
       | |                         ------------------- borrow of `locked_elems` occurs here
    14 | |         move || {
       | |         ^^^^^^^ move out of `locked_elems` occurs here
    15 | |             let _ = &locked_elems; // ensure `locked_elems` is captured (and thus moved)
       | |                      ------------ move occurs due to use in closure
    16 | |             elems.next() // Error, can't use borrow of moved value!
    17 | |         }
    18 | |     })
       | |______- returning this value requires that `locked_elems` is borrowed for `'1`
    
    error: aborting due to 2 previous errors
    
  • In other cases sub-efficient workarounds may be available

    Such as when that Set would be a Vec instead. In that case, we can use indices as a poorman's self-reference, with no "official" lifetimes and thus Rust not complaining:

    fn iter_locked (mutexed_vec: &'_ Mutex<Vec<i32>>)
      -> impl '_ + Iterator<Item = i32>
    {
        ::std::iter::from_fn({
            let locked_vec = mutexed_vec.lock().unwrap();
            let mut indices = 0.. locked_vec.len();
            move /* locked_vec, indices */ || {
                let i = indices.next()?;
                Some(locked_vec[i]) // copies, so OK.
            }
        })
    }
    let mutexed_elems = Mutex::new(vec![27, 42]);
    let mut iter = iter_locked(&mutexed_elems);
    assert_eq!(iter.next(), Some(27));
    assert_eq!(iter.next(), Some(42));
    assert_eq!(iter.next(), None);
    

But with generators this is easy:

use ::next_gen::prelude::*;

#[generator(yield(i32))]
fn gen_iter_locked (mutexed_elems: &'_ Mutex<Set<i32>>)
{
    let locked_elems = mutexed_elems.lock().unwrap();
    for elem in locked_elems.iter().copied() {
        yield_!(elem);
    }
}

and voilà!

That #[generator] fn is the key constructor for our safe self-referential iterator!

Now, instantiating an iterator off a self-referential generator has a subtle aspect, muck alike that of polling a self-referential Future (that's what a missing Unpin bound means): we need to get it pinned before it can be polled!

About pinning "before use", and the two forms of pinning
  1. Getting a Future:

    let future = async { ... };
    // or
    let future = some_async_fn(...);
    
  • Pinning an instantiated Future in the heap (Boxed):

    // Pinning it in the heap (boxed):
    let mut pinned_future = Box::pin(future)
    // or, through an extension trait (`::futures::future::FutureExt`):
    let mut pinned_future = future.boxed() // this also incidentally `dyn`-erases the future.
    
    • Now we can return it, or poll it:

      if true {
          pinned_future.as_mut().poll(...);
      }
      // and/or return it:
      return pinned_future;
      
  • Pinning an instantiated Future in the stack (pinned to the local scope):

    use ::some_lib::some_pinning_macro as stack_pinned;
    // Pinning it in the "stack"
    stack_pinned!(mut future);
    /* the above shadows `future`, thus acting as:
    let mut future = magic::Stack::pin(future); // */
    
    // Let's rename it for clarity:
    let mut pinned_future = future;
    
    • Now we can poll it / use it within the current stack frame, but we cannot return it.

      pinned_future.as_mut().poll(...)
      

Well, it turns out that for generators it's similar:

  1. Once you have a #[generator] fn "generator constructor"

    use ::next_gen::prelude::*;
    
    #[generator(yield(u8))]
    fn foo ()
    {
        yield_!(42);
    }
    
  2. Instantiation requires pinning, and thus:

    • Stack-pinning: cheap, no_std compatible, usable within the same scope. But it cannot be returned.

      mk_gen!(let mut generator = foo());
      
      // can be used within the same scope
      assert_eq!(generator.next(), Some(42));
      assert_eq!(generator.next(), None);
      
      // but it can't be returned
      // return generator; /* Error, can't return borrow to local value */
      
    • Heap-pinning: a bit more expensive, requires an ::allocator or not being no_std, but the so-pinned generator can be returned.

      mk_gen!(let mut generator = box foo());
      
      // can be used within the same scope
      if some_condition {
          assert_eq!(generator.next(), Some(42));
          assert_eq!(generator.next(), None);
      }
      
      // and/or it can be returned
      return generator; // OK
      

So, back to our example, this is what we need to do:


use ::next_gen::prelude::*;

/// We already have:
#[generator(yield(i32))]
fn gen_iter_locked (mutexed_elems: &'_ Mutex<Set<i32>>)
...

/// Now let's wrap-it so that it yields a nice iterator:
fn iter_locked (mutexed_elems: &'_ Mutex<Set<i32>>)
  -> impl '_ + Iterator<Item = i32>
{
    if true {
        // One possible syntax to instantiate the generator
        mk_gen!(let generator = box gen_iter_locked(mutexed_elems));
        generator
    } else {
        // or, since we are `box`-ing, we can directly do:
        gen_iter_locked.call_boxed((mutexed_elems, ))
    }
    // : Pin<Box<impl '_ + Generator<Yield = i32>>>
    // : impl '_ + Iterator<Item = i32>
}

let mutexed_elems = Mutex::new([27, 42].iter().copied().collect::<Set<_>>());
let mut iter = iter_locked(&mutexed_elems);
assert_eq!(iter.next(), Some(27));
assert_eq!(iter.next(), Some(42));
assert_eq!(iter.next(), None);
  • If the iter_locked() function you are trying to implement is part of a trait definition and thus need to name the type, at which point the impl '_ + Iterator existential syntax can be problematic, you can then use dyn instead of impl, at the cost of having to mention the Pin<Box<>> layer:

    // instead of
      -> impl '_ + Iterator<Item = i32>
    // write:
      -> Pin<Box<dyn '_ + Generator<Yield = i32, Return = ()>>>
    
    An example
    use ::next_gen::prelude::*;
    
    struct Once<T>(T);
    impl<T : 'static> IntoIterator for Once<T> {
        type Item = T;
        type IntoIter = Pin<Box<dyn Generator<(), Yield = T, Return = ()> + 'static>>;
    
        fn into_iter (self: Once<T>)
          -> Self::IntoIter
        {
            #[generator(yield(T))]
            fn once_generator<T> (value: T)
            {
                yield_!(value);
            }
    
            once_generator.call_boxed((self.0, ))
        }
    }
    assert_eq!(Once(42).into_iter().next(), Some(42));
    

Resume arguments

This crate has been updated to support resume arguments: the Generator trait is now generic over a ResumeArg parameter (which defaults to ()), and its .resume() method now takes a parameter of that type:

let _: GeneratorState<Yield, Return> = generator.as_mut().resume(resume_arg);

this makes it so the yield_!() expressions inside the generator evaluate to ResumeArg rather than ():

let _: ResumeArg = yield_!(value);

Macro syntax

In order to express this using the #[generator] attribute, add a resume(Type) parameter to it:

use ::next_gen::prelude::*;

type ShouldContinue = bool;

#[generator(yield(i32), resume(ShouldContinue))]
fn g ()
{
    for i in 0.. {
        let should_continue = yield_!(i);
        if should_continue.not() {
            break;
        }
    }
}

mk_gen!(let mut generator = g());
assert!(matches!(
    generator.as_mut().resume(bool::default()), // <- this resume arg is being ignored
    GeneratorState::Yielded(0),
));
assert!(matches!(
    generator.as_mut().resume(true),
    GeneratorState::Yielded(1),
));
assert!(matches!(
    generator.as_mut().resume(true),
    GeneratorState::Yielded(2),
));
assert!(matches!(
    generator.as_mut().resume(true),
    GeneratorState::Yielded(3),
));

assert!(matches!(
    generator.as_mut().resume(false),
    GeneratorState::Complete,
));

If you don't want to ignore/disregard the first resume argument (the "start argument" we could call it), then you can append a as <binding> after the resume(ResumeArgTy) annotation:

use ::next_gen::prelude::*;

type ShouldContinue = bool;

#[generator(
    yield(i32),
    resume(ShouldContinue) as mut should_continue,
)]
fn g ()
{
    for i in 0.. {
        if should_continue.not() {
            break;
        }
        should_continue = yield_!(i);
    }
}
  • A mind-bending example of recursion with an "automagically segmented stack"
    use ::next_gen::prelude::*;
    
    /// A silly recursive function, computing the sum of integers up to `n`.
    ///
    /// If you know your math, you know this equals `n * (n + 1) / 2`.
    ///
    /// This result is quite "obvious" from the geometric representation:
    ///
    /// ```text
    /// # . . . . .   <- Amount of #: 1
    /// # # . . . .   <- Amount of #: 2
    /// # # # . . .   <- Amount of #: 3
    /// # # # # . .   <- Amount of #: 4
    /// ⋮   …   ⋱ ⋮
    /// # # # # # #   <- Amount of #: N
    /// Height = N + 1         Total: 1 + 2 + … + N
    /// Width  = N
    /// Half the area of the "square": (N + 1) * N / 2
    /// ```
    ///
    /// As you can see, computing the sum `1 + 2 + 3 + … + N` is the same as
    /// counting the number of `#` in that diagram. And those `#` fill half a
    /// "square". But it's actually not exactly a `N x N` square since we have
    /// from `1` to `N` rows, that is, `N + 1` rows, and a width of `N`.
    ///
    /// This results in `(n + 1) * n` for the area of the "square", followed by
    /// the `/ 2` halving operation.
    fn triangular (n: u64)
      -> u64
    {
        #[generator(yield(u64), resume(u64))]
        fn triangular (n: u64)
          -> u64
        {
            use yield_ as recurse;
            if n == 0 {
                0
            } else {
                n + recurse!(n - 1)
            }
        }
    
        drive_recursion(n, |n| triangular.call_boxed((n, )))
    }
    
    const N: u64 = 10_000;
    assert_eq!(
        triangular(N),
        N * (N + 1) / 2
    );
    
    // where the `drive_recursion` "runtime" is defined as:
    
    /// A recursive computation can be seen as a "suspensible coroutine",
    /// whereby, when needing to "compute-recurse" into new inputs,
    /// that current computation just suspends and yields the new input
    /// for which it requests a computation.
    ///
    /// The driver / "executor", thus starts with the initial input, and
    /// polls the suspensible coroutine until reaching a suspension point.
    ///
    /// Such suspension point gives the driver a new computation it needs to
    /// perform (updates `input`), and a new "customer" waiting for that new
    /// result: that suspended computation. These stack onto each other as
    /// we recurse, and when the innermost computation _completes_ / _returns_
    /// rather than yield-enqueuing a new one, we can then feed that result to
    /// the top-most suspended computation, _resuming_ it.
    fn drive_recursion<Input, SuspendedComputation, Result> (
        input: Input,
        mut start_computing: impl FnMut(Input) -> Pin<Box<SuspendedComputation>>,
    ) -> Result
    where
        SuspendedComputation
            : Generator<
                /* ResumedWith = */ Result, // recursive result
                Yield = Input, // recursive "query"
                Return = Result,
            >
        ,
        Result : Default, // to feed the initial dummies.
    {
        // This is the "recursive state stack", when you think about this,
        // and with this approach we automagically get it heap-allocated
        // (the `Pin<Box<GeneratorFn…>>` state machines are the main things
        // heap-allocating the "recursively captured local state".
        // This vec is just storing these `Pin<Box<…>>` things, to avoid
        // stack-allocating those (which naively recursing within this very body
        // would achieve).
        let mut suspended_computations = Vec::new();
    
        let mut last_suspended_computation = start_computing(input);
        let mut computation_result = Result::default(); // start with a dummy one
    
        loop {
            match last_suspended_computation.resume_unpin(computation_result) {
                // We reached `return`: completion of the current computation.
                | GeneratorState::Returned(computation_result_) => {
                    match suspended_computations.pop() {
                        // If it was the outer-most computation, we've finished.
                        | None => return computation_result_,
                        // Otherwise, feed the current result to the outer
                        // computation that had previously yield-requested the
                        // current computation.
                        | Some(suspended_computation) => {
                            last_suspended_computation = suspended_computation;
                            computation_result = computation_result_;
                        },
                    }
                },
                // We need to "compute-recurse" ourselves with this new `arg`
                | GeneratorState::Yielded(arg) => {
                    suspended_computations.push(last_suspended_computation);
                    last_suspended_computation = start_computing(arg);
                    computation_result = Result::default();
                },
            }
        }
    }
    

    The "stacks" (storage for the local variables captured within non-terminal / non-tail recursive calls) are thus, in practice, state that crosses the yield_!() points, resulting in state captured by the Generator. And since the Generator instance is Boxed, it means such stack ends up in the heap, behind a pointer. This happens for each and every recursion step.

    This means that the stack has successfully been segmented (within each Generator instance) into the heap; which is otherwise a cumbersome manual process that is nonetheless needed for non-trivial recursive functions.

Features

Performance

The crate enables no-allocation generators, thanks the usage of stack pinning. When used in that fashion, it should thus be close to zero-cost.

Ergonomics / sugar

A lot of effort has been put into macros and an attribute macro providing the most ergonomic experience when dealing with these generators, despite the complex / subtle internals involved, such as stack pinning.

Safe

Almost no unsafe is used, the exception being:

  • Stack pinning, where it uses the official ::pin_utils::pin_mut implementation;

  • Using the pinning guarantee to extend a lifetime;

no_std support

This crates supports #![no_std]. For it, just disable the default "std" feature:

[dependencies]
next-gen.version = "..."
next-gen.default-features = false  # <- ADD THIS to disable `std`&`alloc` for `no_std` compat
next-gen.features = [
    "alloc",  # If your no_std platform has access to allocators.
  # "std",  # `default-features` bundles this.
]

Idea

Since generators and coroutines rely on the same internals, one can derive a safe implementation of generators using the async / await machinery, which is only already in stable Rust.

A similar idea has also been implemented in https://docs.rs/genawaiter.

Dependencies

~1.5MB
~37K SLoC