Restore the data block size to 62.

The former block size traded away good fit within cache lines in
order to gain faster division in deque_item().  However, compilers
are getting smarter and can now replace the slow division operation
with a fast integer multiply and right shift.  Accordingly, it makes
sense to go back to a size that lets blocks neatly fill entire
cache-lines.

GCC-4.8 and CLANG 4.0 both compute "x // 62" with something
roughly equivalent to "x * 9520900167075897609 >> 69".

diff --git a/Lib/test/test_deque.py b/Lib/test/test_deque.py
index ae1de9ab46da42e8ec658436af130a1b6be7a28f..7bff1d2798b4ebb85f6e65b25c0a1b8941f5cb71 100644 (file)
--- a/Lib/test/test_deque.py
+++ b/Lib/test/test_deque.py
@@ -536,7 +536,7 @@ class TestBasic(unittest.TestCase):
 
     @support.cpython_only
     def test_sizeof(self):
-        BLOCKLEN = 64
+        BLOCKLEN = 62
         basesize = support.calcobjsize('2P4nlP')
         blocksize = struct.calcsize('2P%dP' % BLOCKLEN)
         self.assertEqual(object.__sizeof__(deque()), basesize)

diff --git a/Modules/_collectionsmodule.c b/Modules/_collectionsmodule.c
index 1f583b8c78df13d4e6920ef4c13284a7401faafd..e5dfdb424215653713c092ce034d21c359b41471 100644 (file)
--- a/Modules/_collectionsmodule.c
+++ b/Modules/_collectionsmodule.c
@@ -10,11 +10,14 @@
 /* The block length may be set to any number over 1.  Larger numbers
  * reduce the number of calls to the memory allocator, give faster
  * indexing and rotation, and reduce the link::data overhead ratio.
- * If the block length is a power-of-two, we also get faster
- * division/modulo computations during indexing.
+ *
+ * Ideally, the block length will be set to two less than some
+ * multiple of the cache-line length (so that the full block
+ * including the leftlink and rightlink will fit neatly into
+ * cache lines).
  */
 
-#define BLOCKLEN 64
+#define BLOCKLEN 62
 #define CENTER ((BLOCKLEN - 1) / 2)
 
 /* A `dequeobject` is composed of a doubly-linked list of `block` nodes.