Re: Sporadic crash / valgrind warning in register_field_in_read_map() with semijoin

unsimplified test case
LPexportBug860580_bug860580.test

Re: Sporadic crash / valgrind warning in register_field_in_read_map() with semijoin
A simplified testcase:

SET SESSION optimizer_switch='semijoin=ON,materialization=ON,firstmatch=OFF,loosescan=OFF';

SELECT
alias1.col_int_nokey AS field1 ,
(SELECT
MAX( SQ2_alias1.col_varchar_key) AS SQ2_field1
FROM
t2 AS SQ2_alias1
WHERE
(SQ2_alias1.pk , SQ2_alias1.pk) IN
(
SELECT
C_SQ1_alias2.col_int_nokey AS C_SQ1_field1,
C_SQ1_alias1.col_int_key AS C_SQ1_field2
FROM
(
t3 AS C_SQ1_alias1,
t3 AS C_SQ1_alias2,
t1 AS C_SQ1_alias3
)
WHERE C_SQ1_alias3.col_int_key > 98 AND
C_SQ1_alias3.pk = C_SQ1_alias2.pk AND
C_SQ1_alias3.col_varchar_key = C_SQ1_alias2.col_varchar_key
GROUP BY C_SQ1_field1 , C_SQ1_field2
)
) AS field3
FROM
t3 AS alias1,
t4 AS alias2,
t1 AS alias3

GROUP BY
field1, field3
HAVING
field3 < 'e' AND field1 < 242
ORDER BY
alias1.col_datetime_key ASC , field3 ASC, field1, field3
LIMIT 100 OFFSET 9 ;

drop table t1,t2,t3,t4;

Re: Sporadic crash / valgrind warning in register_field_in_read_map() with semijoin
EXPLAIN:
------------------------------------------------------------------------------------------------------------------------------------------------------------+

------------------------------------------------------------------------------------------------------------------------------------------------------------+

------------------------------------------------------------------------------------------------------------------------------------------------------------+
8 rows in set (0.02 sec)

Re: Sporadic crash / valgrind warning in register_field_in_read_map() with semijoin
An outline of the crashing scenario:

== Query ==
The query has this structure:

select
single_row_subselect (select with non-mergeable
semi-join subselect) as field3
from
...
group by
... field3
having
field3 < 'e' and ...
order by
..., field3, ...
limit ..

Things to note:

• single_row_subselect is uncorrelated
• it is used in select list, group list, order by list.

== Execution ==
Execution proceeds as follows:

• single_row_subselect is put into top-level select's
  exec_const_order_group_cond list.
• JOIN::exec calls cur_const_item->val_str, which evaluates the
  single_row_subselect:

= single_row_subselect has a non-mergeable semi-join which is executed
as materialized join tab (a temporary table is created, and the
IN-equality is injected into subquery's WHERE condition)

= After the subquery has finished executing, the code for materialized
join_tab makes this call:

table->pos_in_table_list->jtbm_subselect->cleanup();

the stack trace at this point looks as follows:

#1 in Item_in_subselect::cleanup
#2 in st_join_table::cleanup
#3 in JOIN::cleanup
#4 in JOIN::join_free
#5 in do_select
#6 in JOIN::exec
#7 in subselect_single_select_engine::exec
#8 in Item_subselect::exec
#9 in Item_singlerow_subselect::val_str
...

= single_row_subselect execution is completed. The code for top-level select
makes multiple attempts to evaluate the single_row_subselect but since it
is uncorrelated, any subsequent attempts only fetch the cached value.

• evenually, execution of upper select reaches filesort() which calls
  find_all_keys() which runs this line:

select->cond->walk(&Item::register_field_in_read_map, 1,
(uchar*) sort_form);

where
select->cond = ((`field3` < 'e') and (`field1` < 242))

where field3 is the single_row_subselect.

field3->Item_subselect::walk is called which will at some point invoke

Item_field::register_field_in_read_map

for the field object that refers to already-freed temporary table.

Launchpad bug id: 860580