Make WorkSync not reverse the batch prior to combining

This saves one iteration (via volatile reads) through the batch, and is allowed because the work units are commutative and associative.

community/primitive-collections/src/main/java/org/neo4j/concurrent/WorkSync.java

@@ -45,7 +45,7 @@
  *
  * @see Work
  */
-@SuppressWarnings( "unchecked" )
+@SuppressWarnings( {"unchecked", "NumberEquality"} )
 public class WorkSync<Material, W extends Work<Material,W>>
 {
     private final Material material;
@@ -96,7 +96,7 @@ public void apply( W work ) throws ExecutionException
     private WorkUnit<Material,W> enqueueWork( W work )
     {
         WorkUnit<Material,W> unit = new WorkUnit<>( work, Thread.currentThread() );
-        unit.next = stack.getAndSet( unit ); // benign race, see reverse()
+        unit.next = stack.getAndSet( unit ); // benign race, see the batch.next read-loop in combine()
         return unit;
     }
 
@@ -169,7 +169,7 @@ private void unlock()
 
     private WorkUnit<Material,W> grabBatch()
     {
-        return reverse( stack.getAndSet( (WorkUnit<Material,W>) stackEnd ) );
+        return stack.getAndSet( (WorkUnit<Material,W>) stackEnd );
     }
 
     private Throwable doSynchronizedWork( WorkUnit<Material,W> batch )
@@ -191,28 +191,6 @@ private Throwable doSynchronizedWork( WorkUnit<Material,W> batch )
         return failure;
     }
 
-    private WorkUnit<Material,W> reverse( WorkUnit<Material,W> batch )
-    {
-        WorkUnit<Material,W> result = (WorkUnit<Material,W>) stackEnd;
-        while ( batch != stackEnd )
-        {
-            WorkUnit<Material,W> tmp = batch.next;
-            while ( tmp == null )
-            {
-                // We may see 'null' via race, as work units are put on the
-                // stack before their 'next' pointers are updated. We just spin
-                // until we observe their volatile write to 'next'.
-                // todo Java9: Thread.onSpinWait() ?
-                Thread.yield();
-                tmp = batch.next;
-            }
-            batch.next = result;
-            result = batch;
-            batch = tmp;
-        }
-        return result;
-    }
-
     private W combine( WorkUnit<Material,W> batch )
     {
         W result = null;
@@ -227,7 +205,17 @@ private W combine( WorkUnit<Material,W> batch )
                 result = result.combine( batch.work );
             }
 
-            batch = batch.next;
+            WorkUnit<Material,W> tmp = batch.next;
+            while ( tmp == null )
+            {
+                // We may see 'null' via race, as work units are put on the
+                // stack before their 'next' pointers are updated. We just spin
+                // until we observe their volatile write to 'next'.
+                // todo Java9: Thread.onSpinWait() ?
+                Thread.yield();
+                tmp = batch.next;
+            }
+            batch = tmp;
         }
         return result;
     }