(NON_PORTABLE) insure good cache alignment of NRE_callbackmig_opt2_tmp

2 files changed, 19 insertions, 5 deletions

diff --git a/TODO_DONE b/TODO_DONE
index 96a275b..6e59acf 100644
--- a/TODO_DONE
+++ b/TODO_DONE
@@ -15,10 +15,6 @@
 * mig-alloc-reform: what can be brought in? At least forget -DPURIFY, make
   it usable from normal builds.
 
-* mod the callback struct: make it have only 3 data fields, 4 in
-  total. This will reduce the size, possible improvements due to cache and 
-  memory alignment issues.
-
 *************************************************************************
 **** DONE ***************************************************************
 *************************************************************************
@@ -71,3 +67,10 @@
   This is done by defining HAVE_FAST_TSD in the defaut build; can remove it
   later on for the default - handy while developing. FIXME!!
 
+* mod the callback struct: make it have only 3 data fields, 4 in
+  total. This will reduce the size, possible improvements due to cache and 
+  memory alignment issues.
+
+* INCOMPATIBILITY ... or at least, no attention paid to windows. We are
+  using __atribute__ to insure good cache alignment of NRE_callback, this
+  is certainly not portable.
diff --git a/generic/tclNRE.h b/generic/tclNRE.h
index a00e75f..7cf2a55 100644
--- a/generic/tclNRE.h
+++ b/generic/tclNRE.h
@@ -18,10 +18,21 @@ typedef struct NRE_callback {
     ClientData data[3];
 } NRE_callback;
 
+/*
+ * - NOTES ON ALIGNMENT -
+ *
+ * . Cache lines are 64B long; callbacks are 32B on x86_64 and 16B on x32. In 
+ *   order to insure that no callback ever strides cache line boundaries, we
+ *   require the first callback to be aligned to 32b (on x32, 16B would be
+ *   enough).
+ * . we use a gcc'ism to insure alignment, portability not being the priority
+ *   right now.
+ */
+
 typedef struct NRE_stack {
     struct NRE_callback items[NRE_STACK_SIZE];
     struct NRE_stack *next;
-} NRE_stack;
+} NRE_stack  __attribute__ ((aligned (32)));
 
 #define NRE_newExtra(ptr)			\
     TclSmallAlloc(5*sizeof(ClientData), ptr)