fix deadlock

Kudos to @Ekleog for finding it!

src/imp_pl.rs

@@ -120,6 +120,10 @@ struct Guard<'a> {
 impl<'a> Drop for Guard<'a> {
     fn drop(&mut self) {
         self.state.store(self.new_state, Ordering::Release);
+        unsafe {
+            let key = self.state as *const _ as usize;
+            parking_lot_core::unpark_all(key, parking_lot_core::DEFAULT_UNPARK_TOKEN);
+        }
     }
 }
 
@@ -136,9 +140,6 @@ fn initialize_inner(state: &AtomicU8, init: &mut dyn FnMut() -> bool) {
                 if init() {
                     guard.new_state = COMPLETE;
                 }
-                unsafe {
-                    parking_lot_core::unpark_all(key, parking_lot_core::DEFAULT_UNPARK_TOKEN);
-                }
                 return;
             }
             Err(COMPLETE) => return,

tests/it.rs

@@ -319,6 +319,28 @@ mod sync {
         assert_eq!(cell.get(), Some(&"hello".to_string()));
     }
 
+    #[test]
+    #[cfg_attr(miri, ignore)] // miri doesn't support Barrier
+    fn get_or_init_stress() {
+        use std::sync::Barrier;
+        let n_threads = 1_000;
+        let n_cells = 1_000;
+        let cells: Vec<_> = std::iter::repeat_with(|| (Barrier::new(n_threads), OnceCell::new()))
+            .take(n_cells)
+            .collect();
+        scope(|s| {
+            for _ in 0..n_threads {
+                s.spawn(|_| {
+                    for (i, (b, s)) in cells.iter().enumerate() {
+                        b.wait();
+                        let j = s.get_or_init(|| i);
+                        assert_eq!(*j, i);
+                    }
+                });
+            }
+        }).unwrap();
+    }
+
     #[test]
     fn from_impl() {
         assert_eq!(OnceCell::from("value").get(), Some(&"value"));