3 files changed, 13 insertions, 9 deletions
diff --git a/gcell/src/include/spu/gc_jd_queue.h b/gcell/src/include/spu/gc_jd_queue.h
index f1ce1b3bd..7a6ac2e21 100644
--- a/gcell/src/include/spu/gc_jd_queue.h
+++ b/gcell/src/include/spu/gc_jd_queue.h
@@ -35,12 +35,14 @@ __GC_BEGIN_DECLS
  *
  * \param[in]  q is EA address of queue structure.
  * \param[out] item_ea is EA address of item at head of queue.
+ * \param[in]  jd_tag is the tag to use to get the LS copy of the item.
  * \param[out] item is local store copy of item at head of queue.
  * \returns false if the queue is empty, otherwise returns true
  *   and sets \p item_ea and DMA's job descriptor into \p item
  */
 bool
-gc_jd_queue_dequeue(gc_eaddr_t q, gc_eaddr_t *item_ea, gc_job_desc_t *item);
+gc_jd_queue_dequeue(gc_eaddr_t q, gc_eaddr_t *item_ea,
+		    int jd_tag, gc_job_desc_t *item);
 
 
 /*!
diff --git a/gcell/src/lib/runtime/spu/gc_main.c b/gcell/src/lib/runtime/spu/gc_main.c
index ef552f14a..867a21de8 100644
--- a/gcell/src/lib/runtime/spu/gc_main.c
+++ b/gcell/src/lib/runtime/spu/gc_main.c
@@ -542,9 +542,6 @@ process_job(gc_eaddr_t jd_ea, gc_job_desc_t *jd)
   int tag = ci_tags + ci_idx;			// use the current completion tag
   mfc_put(jd, jd_ea, sizeof(*jd), tag, 0, 0);
 
-  mfc_sync(tag);	// FIXME this makes it work, but is expensive
-
-  
   // Tell PPE we're done with the job.
   //
   // We queue these up until we run out of room, or until we can send
@@ -593,7 +590,7 @@ main_loop(void)
 	// by somebody doing something to the queue.  Go look and see
 	// if there's anything for us.
 	//
-	if (gc_jd_queue_dequeue(spu_args.queue, &jd_ea, &jd))
+	if (gc_jd_queue_dequeue(spu_args.queue, &jd_ea, ci_tags + ci_idx, &jd))
 	  process_job(jd_ea, &jd);
 
 	gc_jd_queue_getllar(spu_args.queue);	// get a new reservation
@@ -608,7 +605,7 @@ main_loop(void)
 #else
 
     // try to get a job from the job queue 
-    if (gc_jd_queue_dequeue(spu_args.queue, &jd_ea, &jd)){
+    if (gc_jd_queue_dequeue(spu_args.queue, &jd_ea, ci_tags + ci_idx, &jd)){
       total_jobs++;
       gc_log_write2(GCL_SS_SYS, 0x10, jd.sys.job_id, total_jobs);
 
diff --git a/gcell/src/lib/runtime/spu/gc_spu_jd_queue.c b/gcell/src/lib/runtime/spu/gc_spu_jd_queue.c
index ba4a1b9d2..22752fe68 100644
--- a/gcell/src/lib/runtime/spu/gc_spu_jd_queue.c
+++ b/gcell/src/lib/runtime/spu/gc_spu_jd_queue.c
@@ -26,7 +26,8 @@
 extern int gc_sys_tag;
 
 bool
-gc_jd_queue_dequeue(gc_eaddr_t q, gc_eaddr_t *item_ea, gc_job_desc_t *item)
+gc_jd_queue_dequeue(gc_eaddr_t q, gc_eaddr_t *item_ea,
+		    int jd_tag, gc_job_desc_t *item)
 {
   gc_jd_queue_t	local_q;
 
@@ -61,8 +62,12 @@ gc_jd_queue_dequeue(gc_eaddr_t q, gc_eaddr_t *item_ea, gc_job_desc_t *item)
 
   // copy in job descriptor at head of queue
   *item_ea = local_q.head;
-  mfc_get(item, local_q.head, sizeof(gc_job_desc_t), gc_sys_tag, 0, 0);
-  mfc_write_tag_mask(1 << gc_sys_tag);	// the tag we're interested in
+  
+  // We must use the fence with the jd_tag to ensure that any
+  // previously initiated put of a job desc is locally ordered before
+  // the get of the new one.
+  mfc_getf(item, local_q.head, sizeof(gc_job_desc_t), jd_tag, 0, 0);
+  mfc_write_tag_mask(1 << jd_tag);	// the tag we're interested in
   mfc_read_tag_status_all();		// wait for DMA to complete
 
   local_q.head = item->sys.next;