SPJ: Mayor refactoring of ha_ndbcluster_push.cc and the logic which build

the pushed query (or: NdbQueryDef).

Previously the NdbQueryDef was build incremenataly as we analyzed the query 
plan for pushable operations. This had the disadvantage that we could take 
descission about parent - child relations at this stage which later prevented
other child candidates to be appended to the pushed query.

This refactoring extends 'class ndb_pushed_builder_ctx' such that it
collects intermediate information about pushable operation, and possible
parent candidates for each child operation. Furthermore, it 
breaks the pushability logic into multiple passes:

1 ANALYZE:
    Analyze each child candidate and add it to 
    'ndb_pushed_builder_ctx::m_join_scope' as 
    'pushable' if it qualifies as such. In addition possible parent
    candidates, and required (forced) parent dependencies are collected
    for each table. All non pushable operations should be identified 
    during this pass such that only known pushable operations remains.

2 OPTIMIZE:
    Determine the parent to be used among the set of possible
    parents. This is decided based on simple heuristic where
    the goal is to employ conditional filters as soon as possible,  
    reduce the fanout of intermediate results, and utilize
    the parallelism of the SPJ block whenever considdered optimal.

3 BUILD:
    Build the NdbQueryDef.

The primary goal of this refactoring has been to provide a better platform 
for later improving the efficiency of the pushed queries being produced.
 
It has not been the goal of this refactoring to change the pushability
of any query in the existing MTR testsuite. However a few queries has improved
its pushability as the 'analyze' phase now are able to force an artificial 
sequential dependency between tables which enables more child operations to be 
appended to the pushed query.

Author: Ole John Aske

mysql-test/suite/ndb/r/ndb_gis.result

@@ -406,8 +406,8 @@ id	select_type	table	type	possible_keys	key	key_len	ref	rows	filtered	Extra
 1	SIMPLE	g1	ALL	NULL	NULL	NULL	NULL	2	100.00	Using temporary; Using filesort
 1	SIMPLE	g2	ALL	NULL	NULL	NULL	NULL	2	100.00	Using join buffer
 Warnings:
-Note	1644	Table 'g1' not pushable, select list can't contain BLOB columns
-Note	1644	Table 'g2' not pushable, select list can't contain BLOB columns
+Note	1644	Table 'g1' is not pushable: select list can't contain BLOB columns
+Note	1644	Table 'g2' is not pushable: select list can't contain BLOB columns
 Note	1003	select `test`.`g1`.`fid` AS `first`,`test`.`g2`.`fid` AS `second`,within(`test`.`g1`.`g`,`test`.`g2`.`g`) AS `w`,contains(`test`.`g1`.`g`,`test`.`g2`.`g`) AS `c`,overlaps(`test`.`g1`.`g`,`test`.`g2`.`g`) AS `o`,equals(`test`.`g1`.`g`,`test`.`g2`.`g`) AS `e`,disjoint(`test`.`g1`.`g`,`test`.`g2`.`g`) AS `d`,touches(`test`.`g1`.`g`,`test`.`g2`.`g`) AS `t`,intersects(`test`.`g1`.`g`,`test`.`g2`.`g`) AS `i`,crosses(`test`.`g1`.`g`,`test`.`g2`.`g`) AS `r` from `test`.`gis_geometrycollection` `g1` join `test`.`gis_geometrycollection` `g2` order by `test`.`g1`.`fid`,`test`.`g2`.`fid`
 DROP TABLE gis_point, gis_line, gis_polygon, gis_multi_point, gis_multi_line, gis_multi_polygon, gis_geometrycollection, gis_geometry;
 CREATE TABLE t1 (
@@ -960,8 +960,8 @@ id	select_type	table	type	possible_keys	key	key_len	ref	rows	filtered	Extra
 1	SIMPLE	g1	ALL	NULL	NULL	NULL	NULL	2	100.00	Using temporary; Using filesort
 1	SIMPLE	g2	ALL	NULL	NULL	NULL	NULL	2	100.00	Using join buffer
 Warnings:
-Note	1644	Table 'g1' not pushable, select list can't contain BLOB columns
-Note	1644	Table 'g2' not pushable, select list can't contain BLOB columns
+Note	1644	Table 'g1' is not pushable: select list can't contain BLOB columns
+Note	1644	Table 'g2' is not pushable: select list can't contain BLOB columns
 Note	1003	select `test`.`g1`.`fid` AS `first`,`test`.`g2`.`fid` AS `second`,within(`test`.`g1`.`g`,`test`.`g2`.`g`) AS `w`,contains(`test`.`g1`.`g`,`test`.`g2`.`g`) AS `c`,overlaps(`test`.`g1`.`g`,`test`.`g2`.`g`) AS `o`,equals(`test`.`g1`.`g`,`test`.`g2`.`g`) AS `e`,disjoint(`test`.`g1`.`g`,`test`.`g2`.`g`) AS `d`,touches(`test`.`g1`.`g`,`test`.`g2`.`g`) AS `t`,intersects(`test`.`g1`.`g`,`test`.`g2`.`g`) AS `i`,crosses(`test`.`g1`.`g`,`test`.`g2`.`g`) AS `r` from `test`.`gis_geometrycollection` `g1` join `test`.`gis_geometrycollection` `g2` order by `test`.`g1`.`fid`,`test`.`g2`.`fid`
 DROP TABLE gis_point, gis_line, gis_polygon, gis_multi_point, gis_multi_line, gis_multi_polygon, gis_geometrycollection, gis_geometry;
 CREATE TABLE t1 (

mysql-test/suite/ndb/r/ndb_join_pushdown.result

@@ -477,8 +477,7 @@ from t1
  join t1 as t2 on t2.a = t1.a and t2.b = t1.b
  join t1 as t3 on t3.a = t2.c and t3.b = t1.c;
 
-# t4 is NOT pushable as t2 & t3 does not have 
-# a parent / grandparent relationship
+# t4 is pushable iff we force an artificial parental dependency between t2 & t3.
 explain extended
 select straight_join *
 from t1
@@ -526,10 +525,10 @@ from t1
 # may be made pushable by equality set replacement.
 #
 # BEWARE: mysqld optimizer may do its own replacement
-#    before ha_ndbcluster analyze the AQP. We therefore 
+#    before ha_ndbcluster_push analyze the AQP. We therefore 
 #    provide multiple similar testcases and hope that
 #    some of them will trigger the replacement code in 
-#    ha_ndbcluster :-o
+#    ha_ndbcluster_push :-o
 explain extended
 select straight_join *
 from t1
@@ -608,6 +607,58 @@ from t1
  join t1 as t3 on t3.a = t1.c and t3.b = t1.d
  join t1 as t4 on t4.a = t3.c and t4.b = t1.b;
 
+# Added more multidependency tests;
+#
+
+explain extended
+select straight_join *
+from t1
+ join t1 as t2  on t2.a = t1.a and t2.b = t1.b
+ join t1 as t2x on t2x.a = t2.c and t2x.b = t2.d
+ join t1 as t3x on t3x.a = t1.c and t3x.b = t1.d
+ join t1 as t4  on t4.a = t3x.c and t4.b = t2x.c;
+
+explain extended
+select straight_join *
+from t1
+ join t1 as t2  on t2.a = t1.a and t2.b = t1.b
+ join t1 as t2x on t2x.a = t2.c and t2x.b = t2.d
+ join t1 as t3  on t3.a = t1.c and t3.b = t1.d
+ join t1 as t3x on t3x.a = t3.c and t3x.b = t3.d
+ join t1 as t4  on t4.a = t3x.c and t4.b = t2x.c; 
+
+explain extended
+select straight_join *
+from t1
+ join t1 as t2  on t2.a = t1.a and t2.b = t1.b
+ join t1 as t3  on t3.a = t1.c and t3.b = t1.d
+ join t1 as t2x on t2x.a = t2.c and t2x.b = t2.d
+ join t1 as t3x on t3x.a = t3.c and t3x.b = t3.d
+ join t1 as t4  on t4.a = t3x.c and t4.b = t2x.c; 
+
+# 't3' is not referred as ancestor and should not be
+# included in the forced dependencies
+# (Depends directly on 't1' and can be bushy wrt. to
+#  the other tables)
+explain extended
+select straight_join *
+from t1
+ join t1 as t2  on t2.a = t1.a and t2.b = t1.b
+ join t1 as t2x on t2x.a = t2.c and t2x.b = t2.d
+ join t1 as t3  on t3.a = t1.c and t3.b = t1.d
+ join t1 as t3x on t3x.a = t1.c and t3x.b = t1.d
+ join t1 as t4  on t4.a = t3x.c and t4.b = t2x.c;
+
+# 't3' is still independent - see comment above.
+explain extended
+select straight_join *
+from t1
+ join t1 as t2  on t2.a = t1.a and t2.b = t1.b
+ join t1 as t2x on t2x.a = t2.c and t2x.b = t2.d
+ join t1 as t3  on t3.a = t1.c and t3.b = t1.b
+ join t1 as t3x on t3x.a = t1.c and t3x.b = t1.d
+ join t1 as t4  on t4.a = t3x.c and t4.b = t2x.c;
+
 
 # Test a combination of pushed table scan (x, y)
 #  & pushed EQ-bound (indexScan) (z, t1)
@@ -1459,18 +1510,22 @@ select *
 from t1
 join t1 as t2 on (t2.b = t1.b or t2.b = t1.a)
 join t1 as t3 on t3.a = t2.a
-join t1 as t4 on t4.a = t3.b
-;
+join t1 as t4 on t4.a = t3.b;
 
 set ndb_join_pushdown=true;
 
+explain extended
+select *
+from t1
+join t1 as t2 on (t2.b = t1.b or t2.b = t1.a)
+join t1 as t3 on t3.a = t2.a
+join t1 as t4 on t4.a = t3.b;
 --sorted_result
 select *
 from t1
 join t1 as t2 on (t2.b = t1.b or t2.b = t1.a)
 join t1 as t3 on t3.a = t2.a
-join t1 as t4 on t4.a = t3.b
-;
+join t1 as t4 on t4.a = t3.b;
 
 ## Test subquery execution where 'Full scan on null key' strategy requires
 ## table scan execution in addition to the key lookup which was prepared