Add Arc::unwrap_or_drop for safely discarding Arcs without calling the destructor on the inner type.

library/alloc/src/sync.rs

@@ -469,6 +469,17 @@ impl<T> Arc<T> {
     ///
     /// This will succeed even if there are outstanding weak references.
     ///
+    /// It is strongly recommended to use [`Arc::unwrap_or_drop`] instead if you don't
+    /// want to keep the `Arc` in the [`Err`] case.
+    /// Immediately dropping the [`Err`] payload, like in the expression
+    /// `Arc::try_unwrap(this).ok()`, can still cause the strong count to
+    /// drop to zero and the inner value of the `Arc` to be dropped:
+    /// For instance if two threads execute this expression in parallel, then
+    /// there is a race condition. The threads could first both check whether they
+    /// have the last clone of their `Arc` via `try_unwrap`, and then
+    /// both drop their `Arc` in the call to [`ok`][`Result::ok`],
+    /// taking the strong count from two down to zero.
+    ///
     /// # Examples
     ///
     /// ```
@@ -500,6 +511,123 @@ impl<T> Arc<T> {
             Ok(elem)
         }
     }
+
+    /// Returns the inner value, if the `Arc` has exactly one strong reference.
+    ///
+    /// Otherwise, [`None`] is returned and the `Arc` is dropped.
+    ///
+    /// This will succeed even if there are outstanding weak references.
+    ///
+    /// If `unwrap_or_drop` is called on every clone of this `Arc`,
+    /// it is guaranteed that exactly one of the calls returns the inner value.
+    /// This means in particular that the inner value is not dropped.
+    ///
+    /// The similar expression `Arc::try_unwrap(this).ok()` does not
+    /// offer such a guarantee. See the last example below and the documentation
+    /// of [`try_unwrap`][`Arc::try_unwrap`].
+    ///
+    /// # Examples
+    ///
+    /// Minimal example demonstrating the guarantee that `unwrap_or_drop` gives.
+    /// ```
+    /// #![feature(unwrap_or_drop)]
+    ///
+    /// use std::sync::Arc;
+    ///
+    /// let x = Arc::new(3);
+    /// let y = Arc::clone(&x);
+    ///
+    /// // Two threads calling `unwrap_or_drop` on both clones of an `Arc`:
+    /// let x_unwrap_thread = std::thread::spawn(|| Arc::unwrap_or_drop(x));
+    /// let y_unwrap_thread = std::thread::spawn(|| Arc::unwrap_or_drop(y));
+    ///
+    /// let x_unwrapped_value = x_unwrap_thread.join().unwrap();
+    /// let y_unwrapped_value = y_unwrap_thread.join().unwrap();
+    ///
+    /// // One of the threads is guaranteed to receive the inner value:
+    /// assert!(matches!(
+    ///     (x_unwrapped_value, y_unwrapped_value),
+    ///     (None, Some(3)) | (Some(3), None)
+    /// ));
+    /// // The result could also be `(None, None)` if the threads called
+    /// // `Arc::try_unwrap(x).ok()` and `Arc::try_unwrap(y).ok()` instead.
+    /// ```
+    ///
+    /// A more practical example demonstrating the need for `unwrap_or_drop`:
+    /// ```
+    /// #![feature(unwrap_or_drop)]
+    ///
+    /// use std::sync::Arc;
+    ///
+    /// // Definition of a simple singly linked list using `Arc`:
+    /// #[derive(Clone)]
+    /// struct LinkedList<T>(Option<Arc<Node<T>>>);
+    /// struct Node<T>(T, Option<Arc<Node<T>>>);
+    ///
+    /// // Dropping a long `LinkedList<T>` relying on the destructor of `Arc`
+    /// // can cause a stack overflow. To prevent this, we can provide a
+    /// // manual `Drop` implementation that does the destruction in a loop:
+    /// impl<T> Drop for LinkedList<T> {
+    ///     fn drop(&mut self) {
+    ///         let mut x = self.0.take();
+    ///         while let Some(arc) = x.take() {
+    ///             Arc::unwrap_or_drop(arc).map(|node| x = node.1);
+    ///         }
+    ///     }
+    /// }
+    ///
+    /// // implementation of `new` and `push` omitted
+    /// impl<T> LinkedList<T> {
+    ///     /* ... */
+    /// #   fn new() -> Self {
+    /// #       LinkedList(None)
+    /// #   }
+    /// #   fn push(&mut self, x: T) {
+    /// #       self.0 = Some(Arc::new(Node(x, self.0.take())));
+    /// #   }
+    /// }
+    ///
+    /// // The following code could still cause a stack overflow
+    /// // despite the manual `Drop` impl if that `Drop` impl used
+    /// // `Arc::try_unwrap(arc).ok()` instead of `Arc::unwrap_or_drop(arc)`.
+    /// {
+    ///     // create a long list and clone it
+    ///     let mut x = LinkedList::new();
+    ///     for i in 0..100000 {
+    ///         x.push(i); // adds i to the front of x
+    ///     }
+    ///     let y = x.clone();
+    ///
+    ///     // drop the clones in parallel
+    ///     let t1 = std::thread::spawn(|| drop(x));
+    ///     let t2 = std::thread::spawn(|| drop(y));
+    ///     t1.join().unwrap();
+    ///     t2.join().unwrap();
+    /// }
+    /// ```
+    #[inline]
+    #[unstable(feature = "unwrap_or_drop", issue = "none")] // FIXME: add issue
+    // FIXME: should this copy all/some of the comments from drop and drop_slow?
+    pub fn unwrap_or_drop(this: Self) -> Option<T> {
+        // following the implementation of `drop` (and `drop_slow`)
+        let mut this = core::mem::ManuallyDrop::new(this);
+
+        if this.inner().strong.fetch_sub(1, Release) != 1 {
+            return None;
+        }
+
+        acquire!(this.inner().strong);
+
+        // FIXME: should the part below this be moved into a seperate #[inline(never)]
+        // function, like it's done with drop_slow in drop?
+
+        // using `ptr::read` where `drop_slow` was using `ptr::drop_in_place`
+        let inner = unsafe { ptr::read(Self::get_mut_unchecked(&mut this)) };
+
+        drop(Weak { ptr: this.ptr });
+
+        Some(inner)
+    }
 }
 
 impl<T> Arc<[T]> {

library/alloc/src/sync/tests.rs

@@ -101,6 +101,45 @@ fn try_unwrap() {
     assert_eq!(Arc::try_unwrap(x), Ok(5));
 }
 
+#[test]
+fn unwrap_or_drop() {
+    // FIXME: Is doing this kind of loop reasonable? I tested `Arc::try_unwrap(x).ok()`
+    // and it makes this kind of assertion fail in roughly every second run somewhere
+    // between 1000 and 5000 iterations; I feel like doing a single iteration is too
+    // unlikely to catch anything interesting but doing too many is way too slow
+    // for a test that wouldn't ever fail for any reasonable implementation
+
+    for _ in 0..100
+    // ^ increase chances of hitting uncommon race conditions
+    {
+        use std::sync::Arc;
+        let x = Arc::new(3);
+        let y = Arc::clone(&x);
+        let r_thread = std::thread::spawn(|| Arc::try_unwrap(x).ok());
+        let s_thread = std::thread::spawn(|| Arc::try_unwrap(y).ok());
+        let r = r_thread.join().expect("r_thread panicked");
+        let s = s_thread.join().expect("s_thread panicked");
+        assert!(
+            matches!((r, s), (None, Some(3)) | (Some(3), None)),
+            "assertion failed: unexpected result `{:?}`\
+            \n  expected `(None, Some(3))` or `(Some(3), None)`",
+            (r, s),
+        );
+    }
+
+    let x = Arc::new(3);
+    assert_eq!(Arc::unwrap_or_drop(x), Some(3));
+
+    let x = Arc::new(4);
+    let y = Arc::clone(&x);
+    assert_eq!(Arc::unwrap_or_drop(x), None);
+    assert_eq!(Arc::unwrap_or_drop(y), Some(4));
+
+    let x = Arc::new(5);
+    let _w = Arc::downgrade(&x);
+    assert_eq!(Arc::unwrap_or_drop(x), Some(5));
+}
+
 #[test]
 fn into_from_raw() {
     let x = Arc::new(box "hello");