On Thu, Nov 28, 2013 at 10:17 PM, Carsten Haitzler ras...@rasterman.com wrote:
On Thu, 28 Nov 2013 20:22:21 -0200 Gustavo Sverzut Barbieri
barbi...@gmail.com said:
Hi,
I was debugging the crash with elm_genlist_clear() in 1.8 and I'm
wondering the rationale of rel. Currently it is badly broken given
that you use anything other than append/prepend without parent.
The current code keeps a list of items in the smart data (sd-items).
Although these are ordered as they would be displayed, the code will
replicate that information in the rel (it-items-rel), that means
the item the current one is relative to. Then it keeps a boolean
before flag and a reverse relative list (who is relative to it).
i remember putting in rel ONLY for the purpose of the queue handling. ie the
item is queued but since it is relative TO something, rel is filled in with
before true/false. after the item is no longer in the queue - rel should be
unused and ignored. we could happily set it to NULL if we wanted.
Good to know, reseting these flags would make lots of sense, avoid
bugs (as the one I fixed with _item_del_pre_hook, or at least make it
harder to show as it would still happen if the item wasn't realized
yet) and save memory (you'd not keep the list in memory).
While at that, I'd say it's better to remove the rel and before
from items and make it a separate structure that is allocated once the
item is relative to something (you'd only have a single pointer wasted
in the main structure for the item, not a pointer + bitflag, that if
padding is wrong may spawn over one pointer in size!). Similarly we
could have rel_revs to not be a list, but a structure with two
pointers... if we fix this bug keeping the rel, then we'll always
have at most one before and another after, so 2 ptrs.
Bad things happens if you insert or move before, after or sorted as
the siblings are not updated. Take _item_move_before(it, before):
sd-items =
eina_inlist_remove(sd-items, EINA_INLIST_GET(it));
if (it-item-block) _item_block_del(it);
sd-items = eina_inlist_prepend_relative
(sd-items, EINA_INLIST_GET(it), EINA_INLIST_GET(before));
So far, so good. remove from old position and place it where it should
be. Delete the block so it would be recreated when needed.
if (it-item-rel)
it-item-rel-item-rel_revs =
eina_list_remove(it-item-rel-item-rel_revs, it);
It was relative to another item, then remove the current item from the
reverse relative list of that other item as it's not relative to it
anymore. Good.
it-item-rel = before;
before-item-rel_revs = eina_list_append(before-item-rel_revs, it);
it-item-before = EINA_TRUE;
Now make the item relative to the given parameter before, also
appending itself to its reverse relative list. Flag as before. Just
that?
Did you notice the problem? If not take the given list as input:
- a (rel = , before=False, rel_revs=[c])
- c (rel = a, before=False, rel_revs=[d])
- d (rel = c, before=False, rel_revs=[b])
- b (rel = d, before=False, rel_revs=[])
Now move b before c. The result will be:
- a (rel = , before=False, rel_revs=[c])
- b (rel = c, before=True, rel_revs=[])
- c (rel = a, before=False, rel_revs=[d, b])
- d (rel = c, before=False, rel_revs=[]).
Wait, c is still relative to a even if should be after b... To
fix this is quite cumbersome as other elements would be relative given
they can be before=True or False. Would need to walk the reverse
relative's list of c and see if b got in the way of them, changing
their relative members.
Looking at the code it seems the user of that is _item_block_add(),
does it need all of that complexity to do its job? I didn't review
such function, but if we could remove all those pointer+boolean+list
we would save memory AND complexity.
but then you'd be forced to evaluate all queued items immediately. either that
or re-architect how the queue works. ie insert items immediately into the
actual list AND keep a queue entry, but this will have all sorts of nasty
knock-on effects when a user scrolls as they will encounter entire block or
sets of items not yet sized and identified.
well, the current approach is also wrong, so we need to find a solution.
Can't we insert in the sd-items list and somehow flag before/after
item blocks as gone, then we just go and regenerate them? (that may
not make sense, I never looked into the block handling of genlist)
--
Gustavo Sverzut Barbieri
--
Mobile: +55 (19) 9225-2202
Contact: http://www.gustavobarbieri.com.br/contact
--
Rapidly troubleshoot problems before they affect your business. Most IT
organizations don't have a clear picture of how application performance
affects their revenue. With AppDynamics, you get 100% visibility into your
Java,.NET, PHP application. Start your 15-day FREE
