Rollup merge of rust-lang#109970 - danielhenrymantilla:add-poll-fn-pi…

…n-clarifications, r=thomcc

[doc] `poll_fn`: explain how to `pin` captured state safely

Usage of `Pin::new_unchecked(&mut …)` is dangerous with `poll_fn`, even though the `!Unpin`-infectiousness has made things smoother. Nonetheless, there are easy ways to avoid the need for any `unsafe` altogether, be it through `Box::pin`ning, or the `pin!` macro. Since the latter only works within an `async` context, showing an example artificially introducing one ought to help people navigate this subtlety with safety and confidence.

## Preview

https://user-images.githubusercontent.com/9920355/230092494-da22fdcb-0b8f-4ff4-a2ac-aa7d9ead077a.mov

`@rustbot` label +A-docs

library/core/src/future/poll_fn.rs

@@ -24,6 +24,93 @@ use crate::task::{Context, Poll};
 /// assert_eq!(read_future.await, "Hello, World!".to_owned());
 /// # }
 /// ```
+///
+/// ## Capturing a pinned state
+///
+/// Example of a closure wrapping inner futures:
+///
+/// ```
+/// # async fn run() {
+/// use core::future::{self, Future};
+/// use core::task::Poll;
+///
+/// /// Resolves to the first future that completes. In the event of a tie, `a` wins.
+/// fn naive_select<T>(
+///     a: impl Future<Output = T>,
+///     b: impl Future<Output = T>,
+/// ) -> impl Future<Output = T>
+/// {
+///     let (mut a, mut b) = (Box::pin(a), Box::pin(b));
+///     future::poll_fn(move |cx| {
+///         if let Poll::Ready(r) = a.as_mut().poll(cx) {
+///             Poll::Ready(r)
+///         } else if let Poll::Ready(r) = b.as_mut().poll(cx) {
+///             Poll::Ready(r)
+///         } else {
+///             Poll::Pending
+///         }
+///     })
+/// }
+///
+/// let a = async { 42 };
+/// let b = future::pending();
+/// let v = naive_select(a, b).await;
+/// assert_eq!(v, 42);
+///
+/// let a = future::pending();
+/// let b = async { 27 };
+/// let v = naive_select(a, b).await;
+/// assert_eq!(v, 27);
+///
+/// let a = async { 42 };
+/// let b = async { 27 };
+/// let v = naive_select(a, b).await;
+/// assert_eq!(v, 42); // biased towards `a` in case of tie!
+/// # }
+/// ```
+///
+/// This time without [`Box::pin`]ning:
+///
+/// [`Box::pin`]: ../../std/boxed/struct.Box.html#method.pin
+///
+/// ```
+/// # async fn run() {
+/// use core::future::{self, Future};
+/// use core::pin::pin;
+/// use core::task::Poll;
+///
+/// /// Resolves to the first future that completes. In the event of a tie, `a` wins.
+/// fn naive_select<T>(
+///     a: impl Future<Output = T>,
+///     b: impl Future<Output = T>,
+/// ) -> impl Future<Output = T>
+/// {
+///     async {
+///         let (mut a, mut b) = (pin!(a), pin!(b));
+///         future::poll_fn(move |cx| {
+///             if let Poll::Ready(r) = a.as_mut().poll(cx) {
+///                 Poll::Ready(r)
+///             } else if let Poll::Ready(r) = b.as_mut().poll(cx) {
+///                 Poll::Ready(r)
+///             } else {
+///                 Poll::Pending
+///             }
+///         }).await
+///     }
+/// }
+///
+/// let a = async { 42 };
+/// let b = future::pending();
+/// let v = naive_select(a, b).await;
+/// assert_eq!(v, 42);
+/// # }
+/// ```
+///
+///   - Notice how, by virtue of being in an `async` context, we have been able to make the [`pin!`]
+///     macro work, thereby avoiding any need for the `unsafe`
+///     <code>[Pin::new_unchecked](&mut fut)</code> constructor.
+///
+/// [`pin!`]: crate::pin::pin!
 #[stable(feature = "future_poll_fn", since = "1.64.0")]
 pub fn poll_fn<T, F>(f: F) -> PollFn<F>
 where