From 5bf04a4afcc866f52e86d70dc5688e7a1799c57b Mon Sep 17 00:00:00 2001
From: Matthias Geier <Matthias.Geier@gmail.com>
Date: Mon, 6 Dec 2021 20:54:38 +0100
Subject: [PATCH] Use "Relaxed" for reading tail in producer and head in
 consumer

---
 src/chunks.rs | 15 ++++++++++-----
 src/lib.rs    | 12 ++++++++----
 2 files changed, 18 insertions(+), 9 deletions(-)

diff --git a/src/chunks.rs b/src/chunks.rs
index ee99352..e6d9d3b 100644
--- a/src/chunks.rs
+++ b/src/chunks.rs
@@ -237,7 +237,8 @@ impl<T> Producer<T> {
     /// For a safe alternative that provides mutable slices of [`Default`]-initialized slots,
     /// see [`Producer::write_chunk()`].
     pub fn write_chunk_uninit(&mut self, n: usize) -> Result<WriteChunkUninit<'_, T>, ChunkError> {
-        let tail = self.buffer.tail.load(Ordering::Acquire);
+        // "tail" is only ever written by the producer thread, "Relaxed" is enough
+        let tail = self.buffer.tail.load(Ordering::Relaxed);
 
         // Check if the queue has *possibly* not enough slots.
         if self.buffer.capacity - self.buffer.distance(self.cached_head.get(), tail) < n {
@@ -286,7 +287,8 @@ impl<T> Consumer<T> {
     ///
     /// See the documentation of the [`chunks`](crate::chunks#examples) module.
     pub fn read_chunk(&mut self, n: usize) -> Result<ReadChunk<'_, T>, ChunkError> {
-        let head = self.buffer.head.load(Ordering::Acquire);
+        // "head" is only ever written by the consumer thread, "Relaxed" is enough
+        let head = self.buffer.head.load(Ordering::Relaxed);
 
         // Check if the queue has *possibly* not enough slots.
         if self.buffer.distance(head, self.cached_tail.get()) < n {
@@ -481,7 +483,8 @@ impl<T> WriteChunkUninit<'_, T> {
 
     unsafe fn commit_unchecked(self, n: usize) -> usize {
         let p = self.producer;
-        let tail = p.buffer.tail.load(Ordering::Acquire);
+        // "tail" is only ever written by the producer thread, "Relaxed" is enough
+        let tail = p.buffer.tail.load(Ordering::Relaxed);
         let tail = p.buffer.increment(tail, n);
         p.buffer.tail.store(tail, Ordering::Release);
         n
@@ -681,7 +684,8 @@ impl<T> ReadChunk<'_, T> {
             self.second_ptr.add(i).drop_in_place();
         }
         let c = self.consumer;
-        let head = c.buffer.head.load(Ordering::Acquire);
+        // "head" is only ever written by the consumer thread, "Relaxed" is enough
+        let head = c.buffer.head.load(Ordering::Relaxed);
         let head = c.buffer.increment(head, n);
         c.buffer.head.store(head, Ordering::Release);
         n
@@ -735,7 +739,8 @@ impl<'a, T> Drop for ReadChunkIntoIter<'a, T> {
     /// Non-iterated items remain in the ring buffer and are *not* dropped.
     fn drop(&mut self) {
         let c = &self.chunk.consumer;
-        let head = c.buffer.head.load(Ordering::Acquire);
+        // "head" is only ever written by the consumer thread, "Relaxed" is enough
+        let head = c.buffer.head.load(Ordering::Relaxed);
         let head = c.buffer.increment(head, self.iterated);
         c.buffer.head.store(head, Ordering::Release);
     }
diff --git a/src/lib.rs b/src/lib.rs
index 9f1e55e..441274c 100644
--- a/src/lib.rs
+++ b/src/lib.rs
@@ -338,7 +338,8 @@ impl<T> Producer<T> {
     pub fn slots(&self) -> usize {
         let head = self.buffer.head.load(Ordering::Acquire);
         self.cached_head.set(head);
-        let tail = self.buffer.tail.load(Ordering::Acquire);
+        // "tail" is only ever written by the producer thread, "Relaxed" is enough
+        let tail = self.buffer.tail.load(Ordering::Relaxed);
         self.buffer.capacity - self.buffer.distance(head, tail)
     }
 
@@ -434,7 +435,8 @@ impl<T> Producer<T> {
     /// This is a strict subset of the functionality implemented in `write_chunk_uninit()`.
     /// For performance, this special case is immplemented separately.
     fn next_tail(&self) -> Option<usize> {
-        let tail = self.buffer.tail.load(Ordering::Acquire);
+        // "tail" is only ever written by the producer thread, "Relaxed" is enough
+        let tail = self.buffer.tail.load(Ordering::Relaxed);
 
         // Check if the queue is *possibly* full.
         if self.buffer.distance(self.cached_head.get(), tail) == self.buffer.capacity {
@@ -570,9 +572,10 @@ impl<T> Consumer<T> {
     /// assert_eq!(c.slots(), 0);
     /// ```
     pub fn slots(&self) -> usize {
-        let head = self.buffer.head.load(Ordering::Acquire);
         let tail = self.buffer.tail.load(Ordering::Acquire);
         self.cached_tail.set(tail);
+        // "head" is only ever written by the consumer thread, "Relaxed" is enough
+        let head = self.buffer.head.load(Ordering::Relaxed);
         self.buffer.distance(head, tail)
     }
 
@@ -667,7 +670,8 @@ impl<T> Consumer<T> {
     /// This is a strict subset of the functionality implemented in `read_chunk()`.
     /// For performance, this special case is immplemented separately.
     fn next_head(&self) -> Option<usize> {
-        let head = self.buffer.head.load(Ordering::Acquire);
+        // "head" is only ever written by the consumer thread, "Relaxed" is enough
+        let head = self.buffer.head.load(Ordering::Relaxed);
 
         // Check if the queue is *possibly* empty.
         if head == self.cached_tail.get() {