I was trying to write a gist index extension, and, after some debugging,
it looks like I found a bug somewhere in the gist.c code ...
I can't be quite sure, because I am not familiar with the postgres
code... but, here is what I see happenning (this is 7.1, but I compared
the sources to 7.2, and did not see this fixed - although, I did not
inspect it too carefully)...
First of all, gistPageAddItem () calls gistdentryinit() with a pointer
to what's stored in the tuple, so, 'by-value' types do not work (because
gistcentryinit () would be passed the value itself, when called from
gistinsert(), and then, in gistPageAddItem (), it is passed a pointer,
coming from gistdentryinit () - so, it just doesn't know really how to
treat the argument)...
Secondly, gist_tuple_replacekey() seems to have incorrect logic figuring
out if there is enough space in the tuple (it checks for '<', instead of
'<=') - this causes a new tuple to get always created (this one, seems
to be fixed in 7.2)
Thirdly, gist_tuple_replace_key () sends a pointer to entry.pred (which
is already a pointer to the actual value) to index_formtuple (), that
looks at the tuple, sees that the type is 'pass-by-value', and puts that
pointer directly into the tuple, so that, the resulting tuple now
contains a pointer to a pointer to the actual value...
Now, if more then one split is required, this sequence is repeated again
and again and again, so that, by the time the tuple gets actually
written, it contains something like a pointer to a pointer to a pointer
to a pointer to the actual data :-(
Once again, I've seen some comments in the 7.2 branch about gists and
pass-by-value types, but brief looking at the differences in the source
did not make me conveinced that it was indeed fixed...
Anyone knows otherwise?
Thanks a lot!
Dima