Change the scavenging collector so that promoted objects do not reside...

Change the scavenging collector so that promoted objects do not reside
in the old generation while its remembered set is being swept for
pointers into the young generation.  This is done by delaying the
copying of promoted objects from when they are first encountered and
promoted to only when their body is visited.


Committed: http://code.google.com/p/v8/source/detail?r=1971

[Kevin Millikin (Chromium), 2009-05-07 14:36:15]

Because you really like reading GC code.

[bak, 2009-05-12 12:17:51]

LGTM,
  Lars

http://codereview.chromium.org/115086/diff/1/2
File src/heap.cc (right):

http://codereview.chromium.org/115086/diff/1/2#newcode629
Line 629: promoted_rear =
reinterpret_cast<HeapObject**>(new_space_.ToSpaceHigh());
How about:
   promoted_rear = promoted_front;

http://codereview.chromium.org/115086/diff/1/2#newcode659
Line 659: // to-be-promoted objects.  Process them until the queue is empty.
It would be much easier to read if you used an abstraction for the queue. This
also reduces the complexity of this function.

http://codereview.chromium.org/115086/diff/1/2#newcode662
Line 662: Map* map = Map::cast(*(--promoted_front));
while (!queue.is_empty()) {
  HeapObject* source;
  Map* map;
  queue.remove(&source, &map);
  ...
}

http://codereview.chromium.org/115086/diff/1/2#newcode893
Line 893: *(--promoted_rear) = first_word.ToMap();
queue.insert(object, first_word.ToMap());

http://codereview.chromium.org/115086/diff/1/2#newcode902
Line 902: // Update *p.
This comment does not add anything.

http://codereview.chromium.org/115086/diff/1/2#newcode905
Line 905: // Objects promoted to the data space can be copied immediately
NOTE: this is only an optimization.

[Kevin Millikin (Chromium), 2009-05-13 08:17:21]

http://codereview.chromium.org/115086/diff/1/2
File src/heap.cc (right):

http://codereview.chromium.org/115086/diff/1/2#newcode659
Line 659: // to-be-promoted objects.  Process them until the queue is empty.
On 2009/05/12 12:17:51, bak wrote:
> It would be much easier to read if you used an abstraction for the queue. This
> also reduces the complexity of this function.

I thought about that.  I got hung up on trying to come up with something simple
that worked for both the promoted pointers and non-promoted objects.

I'll adopt your suggestion of just doing it for the promoted ones, which is the
messy part anyway.

[bak, 2009-05-15 13:38:06]

LGTM,
  Lars