Misc.

• Add region nodes after fork-joins to get better block placements.

hercules_rt/src/lib.rs

@@ -5,6 +5,7 @@ use std::future::Future;
 use std::marker::PhantomData;
 use std::ptr::{copy_nonoverlapping, write_bytes, NonNull};
 use std::slice::{from_raw_parts, from_raw_parts_mut};
+use std::sync::atomic::{AtomicUsize, Ordering};
 use std::sync::OnceLock;
 
 /*
@@ -928,3 +929,30 @@ unsafe impl GlobalAlloc for AlignedAlloc {
 
 #[global_allocator]
 static A: AlignedAlloc = AlignedAlloc;
+
+pub struct SpinBarrier {
+    num: usize,
+    waiting: AtomicUsize,
+    gen: AtomicUsize,
+}
+
+impl SpinBarrier {
+    pub const fn new(num: usize) -> Self {
+        SpinBarrier {
+            num,
+            waiting: AtomicUsize::new(0),
+            gen: AtomicUsize::new(0),
+        }
+    }
+
+    pub fn wait(&self) {
+        let old_gen = self.gen.load(Ordering::Acquire);
+        let old_waiting = self.waiting.fetch_add(1, Ordering::Relaxed);
+        if old_waiting + 1 == self.num {
+            self.waiting.store(0, Ordering::Relaxed);
+            self.gen.fetch_add(1, Ordering::Release);
+        } else {
+            while old_gen == self.gen.load(Ordering::Acquire) {}
+        }
+    }
+}