Jay Pipes wrote:
> Øystein Grøvlen wrote:
>> From my point of view, one should try to pull out t3 if t2 is
>> pulled-out, but not otherwise. Does anyone disagree?
>
> No, that makes sense to me. :) It is also what the WL#3740 seems to
> indicate is the correct behaviour:
<snip>
> So, perhaps the current code has a bug in it somewhere?
I think there might be more that one bug preventing the table from
being pulled out. The relevant code from make_join_statistics is as
follows:
if (s->dependent) // If dependent on some table
{
// All dep. must be constants
if (s->dependent & ~(found_const_table_map))
continue;
if (table->file->stats.records <= 1L &&
(table->file->ha_table_flags() & HA_STATS_RECORDS_IS_EXACT) &&
!table->pos_in_table_list->embedding)
{ // system table
<Register table as a constant table>
}
Note that s->dependent will refer to the table of the left side of
STRAIGHT_JOIN. There are two issues with this code:
1. As far as I can tell, found_const_table_map will only contain tables
with one row. To also include tables with zero rows, one should test
against join->const_table_map.
2. At least in my particular example,
table->pos_in_table_list->embedding is non-zero. I have not quite
figured out what this means, but I observe that if I remove this
criteria some existing tests start failing.
--
Øystein
