Bin: Move stats closer to the mutex.

This is a slight cache locality optimization.

include/jemalloc/internal/bin.h

@@ -16,6 +16,12 @@ struct bin_s {
 	/* All operations on bin_t fields require lock ownership. */
 	malloc_mutex_t		lock;
 
+	/*
+	 * Bin statistics.  These get touched every time the lock is acquired,
+	 * so put them close by in the hopes of getting some cache locality.
+	 */
+	bin_stats_t	stats;
+
 	/*
 	 * Current slab being used to service allocations of this bin's size
 	 * class.  slabcur is independent of slabs_{nonfull,full}; whenever
@@ -33,9 +39,6 @@ struct bin_s {
 
 	/* List used to track full slabs. */
 	edata_list_active_t	slabs_full;
-
-	/* Bin statistics. */
-	bin_stats_t	stats;
 };
 
 /* A set of sharded bins of the same size class. */

src/tcache.c

@@ -347,6 +347,13 @@ tcache_bin_flush_impl(tsd_t *tsd, tcache_t *tcache, cache_bin_t *cache_bin,
 			cur_bin = arena_get_bin(cur_arena, binind,
 			    cur_binshard);
 			assert(cur_binshard < bin_infos[binind].n_shards);
+			/*
+			 * If you're looking at profiles, you might think this
+			 * is a good place to prefetch the bin stats, which are
+			 * often a cache miss.  This turns out not to be
+			 * helpful on the workloads we've looked at, with moving
+			 * the bin stats next to the lock seeming to do better.
+			 */
 		}
 
 		if (small) {