It could be duplicate of MDEV-5356 (I am cheking it).
NO, it is not duplicate.
IF() has the same problem, GREATEST() has more interesting result:
execute my_stmt;
v1_field1
NULL
NULL
NULL
NULL
NULL
NULL
NULL
NULL
execute my_stmt;
v1_field1
10
10
10
10
2
2
2
2
deallocate prepare my_stmt;
Adding fields to SELECT list of the query fixes the problem, so probably read flags set incorrectly during second execution.
according to EXPLAIN EXTENDED ON expression lost on second execution:
execute my_stmt;
id select_type table type possible_keys key key_len ref rows filtered Extra
1 SIMPLE t4 ALL NULL NULL NULL NULL 2 100.00
1 SIMPLE t2 ALL NULL NULL NULL NULL 2 100.00 Using join buffer (flat, BNL join)
1 SIMPLE t1 eq_ref PRIMARY PRIMARY 4 test.t2.i2 1 100.00 Using where
1 SIMPLE t3 ALL NULL NULL NULL NULL 2 100.00 Using join buffer (flat, BNL join)
Warnings:
Note 1003 select coalesce(`test`.`t1`.`j1`,`test`.`t3`.`i3`) AS `v1_field1` from `test`.`t4` join `test`.`t2` join `test`.`t3` left join `test`.`t1` on(((`test`.`t1`.`i1` = `test`.`t2`.`i2`) and (`test`.`t2`.`i2` is not null))) where 1
execute my_stmt;
id select_type table type possible_keys key key_len ref rows filtered Extra
1 SIMPLE t4 ALL NULL NULL NULL NULL 2 100.00
1 SIMPLE t2 ALL NULL NULL NULL NULL 2 100.00 Using join buffer (flat, BNL join)
1 SIMPLE t3 ALL NULL NULL NULL NULL 2 100.00 Using join buffer (incremental, BNL join)
Warnings:
Note 1003 select coalesce(`test`.`t1`.`j1`,`test`.`t3`.`i3`) AS `v1_field1` from `test`.`t4` join `test`.`t2` join `test`.`t3` where 1
eliminate_tables() got wrong used tables map from SELECT item list (6 instead of 12) on second execution
it looks like lack of update_used_tables() after handling derived tables 
(worked for usual queries because there was other call only for firest execution)
Patch sent for review.
Debugging the example, I see:
(gdb) p join->fields_list.head()
$39 = (Item_direct_view_ref *) 0x7fffcd4eefb8
(gdb) p *join->fields_list.head()->ref
$44 = (Item_direct_view_ref *) 0x7fffcd4eed78
(gdb) p *(*join->fields_list.head()>ref)>ref
$49 = (Item_func_coalesce *) 0x7fffcd4683d0
This is probably not related to this bug, but why is one Item_direct_view_ref
wrapped within the other Item_direct_view_ref.
http://bazaar.launchpad.net/~maria-captains/maria/5.5/revision/2502.567.182