[arm64] Add sgen support.

[mono.git] / mono / metadata / sgen-minor-copy-object.h

diff --git a/mono/metadata/sgen-minor-copy-object.h b/mono/metadata/sgen-minor-copy-object.h
index 343e74c51dc7cca01d2b673e2559e5b8b63f8460..782bbdf8de2dfe8b9994ccd92034b261a06cd250 100644 (file)
--- a/mono/metadata/sgen-minor-copy-object.h
+++ b/mono/metadata/sgen-minor-copy-object.h
@@ -160,6 +160,43 @@ SERIAL_COPY_OBJECT_FROM_OBJ (void **obj_slot, SgenGrayQueue *queue)
                SGEN_ASSERT (9, ((MonoVTable*)SGEN_LOAD_VTABLE(obj))->gc_descr, "to space object %p has no gc descriptor", obj);
                SGEN_LOG (9, " (tospace, no change)");
                HEAVY_STAT (++stat_nursery_copy_object_failed_to_space);                
+
+               /*
+                * FIXME:
+                *
+                * The card table scanning code sometimes clears cards
+                * that have just been set for a global remset.  In
+                * the split nursery the following situation can
+                * occur:
+                *
+                * Let's say object A starts in card C but continues
+                * into C+1.  Within A, at offset O there's a
+                * reference to a new nursery object X.  A+O is in
+                * card C+1.  Now card C is scanned, and as part of
+                * it, object A.  The reference at A+O is processed by
+                * copying X into nursery to-space at Y.  Since it's
+                * still in the nursery, a global remset must be added
+                * for A+O, so card C+1 is marked.  Now, however, card
+                * C+1 is scanned, which means that it's cleared
+                * first.  This wouldn't be terribly bad if reference
+                * A+O were re-scanned and the global remset re-added,
+                * but since the reference points to to-space, that
+                * doesn't happen, and C+1 remains cleared: the remset
+                * is lost.
+                *
+                * There's at least two ways to fix this.  The easy
+                * one is to re-add the remset on the re-scan.  This
+                * is that - the following two lines of code.
+                *
+                * The proper solution appears to be to first make a
+                * copy of the cards before scanning a block, then to
+                * clear all the cards and scan from the copy, so no
+                * remsets will be overwritten.  Scanning objects at
+                * most once would be the icing on the cake.
+                */
+               if (!sgen_ptr_in_nursery (obj_slot))
+                       sgen_add_to_global_remset (obj_slot, obj);
+
                return;
        }
 #endif