X-Git-Url: http://git.megacz.com/?a=blobdiff_plain;f=rts%2Fsm%2FStorage.c;h=2e43bffdfc5ab6b39325664d99ca8e8acf1a3718;hb=cebecc110be6e53c310326940a311cab7fad84ca;hp=3b9775ede13d3004ec2df2be23461f350c0f89de;hpb=1373cd3085b8dec456e6118c58e940718cb9d559;p=ghc-hetmet.git

diff --git a/rts/sm/Storage.c b/rts/sm/Storage.c
index 3b9775e..2e43bff 100644
--- a/rts/sm/Storage.c
+++ b/rts/sm/Storage.c
@@ -290,7 +290,9 @@ newCAF(StgRegTable *reg, StgClosure* caf)
   {
     // Put this CAF on the mutable list for the old generation.
     ((StgIndStatic *)caf)->saved_info = NULL;
-    recordMutableCap(caf, regTableToCapability(reg), oldest_gen->no);
+    if (oldest_gen->no != 0) {
+        recordMutableCap(caf, regTableToCapability(reg), oldest_gen->no);
+    }
   }
 }
 
@@ -330,31 +332,45 @@ static bdescr *
 allocNursery (bdescr *tail, nat blocks)
 {
     bdescr *bd;
-    nat i;
+    nat i, n;
 
-    // Allocate a nursery: we allocate fresh blocks one at a time and
-    // cons them on to the front of the list, not forgetting to update
-    // the back pointer on the tail of the list to point to the new block.
-    for (i=0; i < blocks; i++) {
-	// @LDV profiling
-	/*
-	  processNursery() in LdvProfile.c assumes that every block group in
-	  the nursery contains only a single block. So, if a block group is
-	  given multiple blocks, change processNursery() accordingly.
-	*/
-	bd = allocBlock();
-	bd->link = tail;
-	// double-link the nursery: we might need to insert blocks
-	if (tail != NULL) {
-	    tail->u.back = bd;
-	}
-        initBdescr(bd, g0, g0);
-	bd->flags = 0;
-	bd->free = bd->start;
-	tail = bd;
+    // We allocate the nursery as a single contiguous block and then
+    // divide it into single blocks manually.  This way we guarantee
+    // that the nursery blocks are adjacent, so that the processor's
+    // automatic prefetching works across nursery blocks.  This is a
+    // tiny optimisation (~0.5%), but it's free.
+
+    while (blocks > 0) {
+        n = stg_min(blocks, BLOCKS_PER_MBLOCK);
+        blocks -= n;
+
+        bd = allocGroup(n);
+        for (i = 0; i < n; i++) {
+            initBdescr(&bd[i], g0, g0);
+
+            bd[i].blocks = 1;
+            bd[i].flags = 0;
+
+            if (i > 0) {
+                bd[i].u.back = &bd[i-1];
+            }
+
+            if (i+1 < n) {
+                bd[i].link = &bd[i+1];
+            } else {
+                bd[i].link = tail;
+                if (tail != NULL) {
+                    tail->u.back = &bd[i];
+                }
+            }
+
+            bd[i].free = bd[i].start;
+        }
+
+        tail = &bd[0];
     }
-    tail->u.back = NULL;
-    return tail;
+
+    return &bd[0];
 }
 
 static void
@@ -721,6 +737,16 @@ setTSOLink (Capability *cap, StgTSO *tso, StgTSO *target)
 }
 
 void
+setTSOPrev (Capability *cap, StgTSO *tso, StgTSO *target)
+{
+    if (tso->dirty == 0 && (tso->flags & TSO_LINK_DIRTY) == 0) {
+        tso->flags |= TSO_LINK_DIRTY;
+        recordClosureMutated(cap,(StgClosure*)tso);
+    }
+    tso->block_info.prev = target;
+}
+
+void
 dirty_TSO (Capability *cap, StgTSO *tso)
 {
     if (tso->dirty == 0 && (tso->flags & TSO_LINK_DIRTY) == 0) {