optimize memory layout for much better cache hits

Moave element status out of batch_obj_t because we
get a *much* better cache hit ratio this way.
Note that this is really a HUGE saving, even if it
doesn't look so (both profiler data as well as
practical tests indicate that!).

1 files changed, 11 insertions, 4 deletions

diff --git a/runtime/batch.h b/runtime/batch.h
index 097e104a..03122bd0 100644
--- a/runtime/batch.h
+++ b/runtime/batch.h
@@ -46,7 +46,6 @@ typedef unsigned char batch_state_t;
  */
 struct batch_obj_s {
 	msg_t *pMsg;
-	batch_state_t state;	/* associated state */
 	/* work variables for action processing; these are reused for each action (or block of
 	 * actions)
 	 */
@@ -84,6 +83,13 @@ struct batch_s {
 	sbool *active;		/* which messages are active for processing, NULL=all */
 	sbool bSingleRuleset;	/* do all msgs of this batch use a single ruleset? */
 	batch_obj_t *pElem;	/* batch elements */
+	batch_state_t *eltState;/* state (array!) for individual objects.
+	   			   NOTE: we have moved this out of batch_obj_t because we
+				         get a *much* better cache hit ratio this way. So do not
+					 move it back into this structure! Note that this is really
+					 a HUGE saving, even if it doesn't look so (both profiler
+					 data as well as practical tests indicate that!).
+				*/
 };
 
 
@@ -118,8 +124,8 @@ batchNumMsgs(batch_t *pBatch) {
  */
 static inline void
 batchSetElemState(batch_t *pBatch, int i, batch_state_t newState) {
-	if(pBatch->pElem[i].state != BATCH_STATE_DISC)
-		pBatch->pElem[i].state = newState;
+	if(pBatch->eltState[i] != BATCH_STATE_DISC)
+		pBatch->eltState[i] = newState;
 }
 
 
@@ -128,7 +134,7 @@ batchSetElemState(batch_t *pBatch, int i, batch_state_t newState) {
  */
 static inline int
 batchIsValidElem(batch_t *pBatch, int i) {
-	return(   (pBatch->pElem[i].state != BATCH_STATE_DISC)
+	return(   (pBatch->eltState[i] != BATCH_STATE_DISC)
 	       && (pBatch->active == NULL || pBatch->active[i]));
 }
 
@@ -163,6 +169,7 @@ batchInit(batch_t *pBatch, int maxElem) {
 	pBatch->iDoneUpTo = 0;
 	pBatch->maxElem = maxElem;
 	CHKmalloc(pBatch->pElem = calloc((size_t)maxElem, sizeof(batch_obj_t)));
+	CHKmalloc(pBatch->eltState = calloc((size_t)maxElem, sizeof(batch_state_t)));
 	// TODO: replace calloc by inidividual writes?
 finalize_it:
 	RETiRet;