> False promotions

> ... When you do hit the end, you do a minor collection, walking the minor heap to figure out what's still live, and promoting that set of objects to the major heap.

I somewhat deal with that issue here:

http://www.kylheku.com/cgit/txr/tree/gc.c

When make_obj exhausts the available FRESHOBJ_VEC_SIZE in the freshobj array (nursery), it first tries to shake things out by doing a partial GC. The partial GC, during the sweep phase, will rebuild the freshobj array with just those gen 0 objects that are reachable. The sweep phase of the incremental GC marches through the freshobj array, and a second index is maintained where live objects are copied back into the array. At the end, that index becomes the new allocation pointer from that array.

Only if the array is still full after this do we then have to do a full gc, and that's when any "false promotion" will happen.

One problem with this approach is that this freshobj nursery becomes a miniature memory arena in itself; as it gets close to full, these partial garbage collections get more frequent. They are not that cheap.

Hmm, it occurs to me I can address that with a one-liner:

    diff --git a/gc.c b/gc.c
    index 6832905..7a0ee1c 100644
    --- a/gc.c
    +++ b/gc.c
    @@ -173,7 +173,7 @@ val make_obj(void)
           malloc_delta >= opt_gc_delta)
       {
         gc();
    -    if (freshobj_idx >= FRESHOBJ_VEC_SIZE)
    +    if (freshobj_idx >= FRESHOBJ_VEC_SIZE / 2)
           full_gc = 1;
         prev_malloc_bytes = malloc_bytes;
       }

I.e. after doing an ephemeral GC, if at least half of the nursery space isn't free, then set the flag for a full gc. (That will happen right away if the nursery has no more objects to satisfy the current request, or else the next time gc is triggered.)

This has been a useful HN submission since it got me thinking about something, leading to some findings in a different project. :)