1 | /* |
2 | Copyright (c) 2010, 2015, MariaDB |
3 | |
4 | This program is free software; you can redistribute it and/or modify |
5 | it under the terms of the GNU General Public License as published by |
6 | the Free Software Foundation; version 2 of the License. |
7 | |
8 | This program is distributed in the hope that it will be useful, |
9 | but WITHOUT ANY WARRANTY; without even the implied warranty of |
10 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
11 | GNU General Public License for more details. |
12 | |
13 | You should have received a copy of the GNU General Public License |
14 | along with this program; if not, write to the Free Software |
15 | Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02111-1301 USA */ |
16 | |
17 | /** |
18 | @file |
19 | |
20 | @brief |
21 | Semi-join subquery optimizations code |
22 | |
23 | */ |
24 | |
25 | #ifdef USE_PRAGMA_IMPLEMENTATION |
26 | #pragma implementation // gcc: Class implementation |
27 | #endif |
28 | |
29 | #include "mariadb.h" |
30 | #include "sql_base.h" |
31 | #include "sql_select.h" |
32 | #include "filesort.h" |
33 | #include "opt_subselect.h" |
34 | #include "sql_test.h" |
35 | #include <my_bit.h> |
36 | |
37 | /* |
38 | This file contains optimizations for semi-join subqueries. |
39 | |
40 | Contents |
41 | -------- |
42 | 1. What is a semi-join subquery |
43 | 2. General idea about semi-join execution |
44 | 2.1 Correlated vs uncorrelated semi-joins |
45 | 2.2 Mergeable vs non-mergeable semi-joins |
46 | 3. Code-level view of semi-join processing |
47 | 3.1 Conversion |
48 | 3.1.1 Merged semi-join TABLE_LIST object |
49 | 3.1.2 Non-merged semi-join data structure |
50 | 3.2 Semi-joins and query optimization |
51 | 3.2.1 Non-merged semi-joins and join optimization |
52 | 3.2.2 Merged semi-joins and join optimization |
53 | 3.3 Semi-joins and query execution |
54 | |
55 | 1. What is a semi-join subquery |
56 | ------------------------------- |
57 | We use this definition of semi-join: |
58 | |
59 | outer_tbl SEMI JOIN inner_tbl ON cond = {set of outer_tbl.row such that |
60 | exist inner_tbl.row, for which |
61 | cond(outer_tbl.row,inner_tbl.row) |
62 | is satisfied} |
63 | |
64 | That is, semi-join operation is similar to inner join operation, with |
65 | exception that we don't care how many matches a row from outer_tbl has in |
66 | inner_tbl. |
67 | |
68 | In SQL terms: a semi-join subquery is an IN subquery that is an AND-part of |
69 | the WHERE/ON clause. |
70 | |
71 | 2. General idea about semi-join execution |
72 | ----------------------------------------- |
73 | We can execute semi-join in a way similar to inner join, with exception that |
74 | we need to somehow ensure that we do not generate record combinations that |
75 | differ only in rows of inner tables. |
76 | There is a number of different ways to achieve this property, implemented by |
77 | a number of semi-join execution strategies. |
78 | Some strategies can handle any semi-joins, other can be applied only to |
79 | semi-joins that have certain properties that are described below: |
80 | |
81 | 2.1 Correlated vs uncorrelated semi-joins |
82 | ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ |
83 | Uncorrelated semi-joins are special in the respect that they allow to |
84 | - execute the subquery (possible as it's uncorrelated) |
85 | - somehow make sure that generated set does not have duplicates |
86 | - perform an inner join with outer tables. |
87 | |
88 | or, rephrasing in SQL form: |
89 | |
90 | SELECT ... FROM ot WHERE ot.col IN (SELECT it.col FROM it WHERE uncorr_cond) |
91 | -> |
92 | SELECT ... FROM ot JOIN (SELECT DISTINCT it.col FROM it WHERE uncorr_cond) |
93 | |
94 | 2.2 Mergeable vs non-mergeable semi-joins |
95 | ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ |
96 | Semi-join operation has some degree of commutability with inner join |
97 | operation: we can join subquery's tables with ouside table(s) and eliminate |
98 | duplicate record combination after that: |
99 | |
100 | ot1 JOIN ot2 SEMI_JOIN{it1,it2} (it1 JOIN it2) ON sjcond(ot2,it*) -> |
101 | | |
102 | +-------------------------------+ |
103 | v |
104 | ot1 SEMI_JOIN{it1,it2} (it1 JOIN it2 JOIN ot2) ON sjcond(ot2,it*) |
105 | |
106 | In order for this to work, subquery's top-level operation must be join, and |
107 | grouping or ordering with limit (grouping or ordering with limit are not |
108 | commutative with duplicate removal). In other words, the conversion is |
109 | possible when the subquery doesn't have GROUP BY clause, any aggregate |
110 | functions*, or ORDER BY ... LIMIT clause. |
111 | |
112 | Definitions: |
113 | - Subquery whose top-level operation is a join is called *mergeable semi-join* |
114 | - All other kinds of semi-join subqueries are considered non-mergeable. |
115 | |
116 | *- this requirement is actually too strong, but its exceptions are too |
117 | complicated to be considered here. |
118 | |
119 | 3. Code-level view of semi-join processing |
120 | ------------------------------------------ |
121 | |
122 | 3.1 Conversion and pre-optimization data structures |
123 | --------------------------------------------------- |
124 | * When doing JOIN::prepare for the subquery, we detect that it can be |
125 | converted into a semi-join and register it in parent_join->sj_subselects |
126 | |
127 | * At the start of parent_join->optimize(), the predicate is converted into |
128 | a semi-join node. A semi-join node is a TABLE_LIST object that is linked |
129 | somewhere in parent_join->join_list (either it is just present there, or |
130 | it is a descendant of some of its members). |
131 | |
132 | There are two kinds of semi-joins: |
133 | - Merged semi-joins |
134 | - Non-merged semi-joins |
135 | |
136 | 3.1.1 Merged semi-join TABLE_LIST object |
137 | ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ |
138 | Merged semi-join object is a TABLE_LIST that contains a sub-join of |
139 | subquery tables and the semi-join ON expression (in this respect it is |
140 | very similar to nested outer join representation) |
141 | Merged semi-join represents this SQL: |
142 | |
143 | ... SEMI JOIN (inner_tbl1 JOIN ... JOIN inner_tbl_n) ON sj_on_expr |
144 | |
145 | Semi-join objects of this kind have TABLE_LIST::sj_subq_pred set. |
146 | |
147 | 3.1.2 Non-merged semi-join data structure |
148 | ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ |
149 | Non-merged semi-join object is a leaf TABLE_LIST object that has a subquery |
150 | that produces rows. It is similar to a base table and represents this SQL: |
151 | |
152 | ... SEMI_JOIN (SELECT non_mergeable_select) ON sj_on_expr |
153 | |
154 | Subquery items that were converted into semi-joins are removed from the WHERE |
155 | clause. (They do remain in PS-saved WHERE clause, and they replace themselves |
156 | with Item_int(1) on subsequent re-executions). |
157 | |
158 | 3.2 Semi-joins and join optimization |
159 | ------------------------------------ |
160 | |
161 | 3.2.1 Non-merged semi-joins and join optimization |
162 | ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ |
163 | For join optimization purposes, non-merged semi-join nests are similar to |
164 | base tables. Each such nest is represented by one one JOIN_TAB, which has |
165 | two possible access strategies: |
166 | - full table scan (representing SJ-Materialization-Scan strategy) |
167 | - eq_ref-like table lookup (representing SJ-Materialization-Lookup) |
168 | |
169 | Unlike regular base tables, non-merged semi-joins have: |
170 | - non-zero JOIN_TAB::startup_cost, and |
171 | - join_tab->table->is_filled_at_execution()==TRUE, which means one |
172 | cannot do const table detection, range analysis or other dataset-dependent |
173 | optimizations. |
174 | Instead, get_delayed_table_estimates() will run optimization for the |
175 | subquery and produce an E(materialized table size). |
176 | |
177 | 3.2.2 Merged semi-joins and join optimization |
178 | ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ |
179 | - optimize_semijoin_nests() does pre-optimization |
180 | - during join optimization, the join has one JOIN_TAB (or is it POSITION?) |
181 | array, and suffix-based detection is used, see advance_sj_state() |
182 | - after join optimization is done, get_best_combination() switches |
183 | the data-structure to prefix-based, multiple JOIN_TAB ranges format. |
184 | |
185 | 3.3 Semi-joins and query execution |
186 | ---------------------------------- |
187 | * Join executor has hooks for all semi-join strategies. |
188 | TODO elaborate. |
189 | |
190 | */ |
191 | |
192 | /* |
193 | EqualityPropagationAndSjmNests |
194 | ****************************** |
195 | |
196 | Equalities are used for: |
197 | P1. Equality propagation |
198 | P2. Equality substitution [for a certain join order] |
199 | |
200 | The equality propagation is not affected by SJM nests. In fact, it is done |
201 | before we determine the execution plan, i.e. before we even know we will use |
202 | SJM-nests for execution. |
203 | |
204 | The equality substitution is affected. |
205 | |
206 | Substitution without SJMs |
207 | ========================= |
208 | When one doesn't have SJM nests, tables have a strict join order: |
209 | |
210 | ---------------------------------> |
211 | t1 -- t2 -- t3 -- t4 --- t5 |
212 | |
213 | |
214 | ? ^ |
215 | \ |
216 | --(part-of-WHERE) |
217 | |
218 | |
219 | parts WHERE/ON and ref. expressions are attached at some point along the axis. |
220 | Expression is allowed to refer to a table column if the table is to the left of |
221 | the attachment point. For any given expression, we have a goal: |
222 | |
223 | "Move leftmost allowed attachment point as much as possible to the left" |
224 | |
225 | Substitution with SJMs - task setting |
226 | ===================================== |
227 | |
228 | When SJM nests are present, there is no global strict table ordering anymore: |
229 | |
230 | |
231 | ---------------------------------> |
232 | |
233 | ot1 -- ot2 --- sjm -- ot4 --- ot5 |
234 | | |
235 | | Main execution |
236 | - - - - - - - - - - - - - - - - - - - - - - - - |
237 | | Materialization |
238 | it1 -- it2 --/ |
239 | |
240 | |
241 | Besides that, we must take into account that |
242 | - values for outer table columns, otN.col, are inaccessible at |
243 | materialization step (SJM-RULE) |
244 | - values for inner table columns, itN.col, are inaccessible at Main execution |
245 | step, except for SJ-Materialization-Scan and columns that are in the |
246 | subquery's select list. (SJM-RULE) |
247 | |
248 | Substitution with SJMs - solution |
249 | ================================= |
250 | |
251 | First, we introduce global strict table ordering like this: |
252 | |
253 | ot1 - ot2 --\ /--- ot3 -- ot5 |
254 | \--- it1 --- it2 --/ |
255 | |
256 | Now, let's see how to meet (SJM-RULE). |
257 | |
258 | SJ-Materialization is only applicable for uncorrelated subqueries. From this, it |
259 | follows that any multiple equality will either |
260 | 1. include only columns of outer tables, or |
261 | 2. include only columns of inner tables, or |
262 | 3. include columns of inner and outer tables, joined together through one |
263 | of IN-equalities. |
264 | |
265 | Cases #1 and #2 can be handled in the same way as with regular inner joins. |
266 | |
267 | Case #3 requires special handling, so that we don't construct violations of |
268 | (SJM-RULE). Let's consider possible ways to build violations. |
269 | |
270 | Equality propagation starts with the clause in this form |
271 | |
272 | top_query_where AND subquery_where AND in_equalities |
273 | |
274 | First, it builds multi-equalities. It can also build a mixed multi-equality |
275 | |
276 | multiple-equal(ot1.col, ot2.col, ... it1.col, itN.col) |
277 | |
278 | Multi-equalities are pushed down the OR-clauses in top_query_where and in |
279 | subquery_where, so it's possible that clauses like this one are built: |
280 | |
281 | subquery_cond OR (multiple-equal(it1.col, ot1.col,...) AND ...) |
282 | ^^^^^^^^^^^^^ \ |
283 | | this must be evaluated |
284 | \- can only be evaluated at the main phase. |
285 | at the materialization phase |
286 | |
287 | Finally, equality substitution is started. It does two operations: |
288 | |
289 | |
290 | 1. Field reference substitution |
291 | ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ |
292 | |
293 | (In the code, this is Item_field::replace_equal_field) |
294 | |
295 | This is a process of replacing each reference to "tblX.col" |
296 | with the first element of the multi-equality. (REF-SUBST-ORIG) |
297 | |
298 | This behaviour can cause problems with Semi-join nests. Suppose, we have a |
299 | condition: |
300 | |
301 | func(it1.col, it2.col) |
302 | |
303 | and a multi-equality(ot1.col, it1.col). Then, reference to "it1.col" will be |
304 | replaced with "ot1.col", constructing a condition |
305 | |
306 | func(ot1.col, it2.col) |
307 | |
308 | which will be a violation of (SJM-RULE). |
309 | |
310 | In order to avoid this, (REF-SUBST-ORIG) is amended as follows: |
311 | |
312 | - references to tables "itX.col" that are inner wrt some SJM nest, are |
313 | replaced with references to the first inner table from the same SJM nest. |
314 | |
315 | - references to top-level tables "otX.col" are replaced with references to |
316 | the first element of the multi-equality, no matter if that first element is |
317 | a column of a top-level table or of table from some SJM nest. |
318 | (REF-SUBST-SJM) |
319 | |
320 | The case where the first element is a table from an SJM nest $SJM is ok, |
321 | because it can be proven that $SJM uses SJ-Materialization-Scan, and |
322 | "unpacks" correct column values to the first element during the main |
323 | execution phase. |
324 | |
325 | 2. Item_equal elimination |
326 | ~~~~~~~~~~~~~~~~~~~~~~~~~ |
327 | (In the code: eliminate_item_equal) This is a process of taking |
328 | |
329 | multiple-equal(a,b,c,d,e) |
330 | |
331 | and replacing it with an equivalent expression which is an AND of pair-wise |
332 | equalities: |
333 | |
334 | a=b AND a=c AND ... |
335 | |
336 | The equalities are picked such that for any given join prefix (t1,t2...) the |
337 | subset of equalities that can be evaluated gives the most restrictive |
338 | filtering. |
339 | |
340 | Without SJM nests, it is sufficient to compare every multi-equality member |
341 | with the first one: |
342 | |
343 | elem1=elem2 AND elem1=elem3 AND elem1=elem4 ... |
344 | |
345 | When SJM nests are present, we should take care not to construct equalities |
346 | that violate the (SJM-RULE). This is achieved by generating separate sets of |
347 | equalites for top-level tables and for inner tables. That is, for the join |
348 | order |
349 | |
350 | ot1 - ot2 --\ /--- ot3 -- ot5 |
351 | \--- it1 --- it2 --/ |
352 | |
353 | we will generate |
354 | ot1.col=ot2.col |
355 | ot1.col=ot3.col |
356 | ot1.col=ot5.col |
357 | it2.col=it1.col |
358 | |
359 | |
360 | 2.1 The problem with Item_equals and ORs |
361 | ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ |
362 | As has been mentioned above, multiple equalities are pushed down into OR |
363 | clauses, possibly building clauses like this: |
364 | |
365 | func(it.col2) OR multiple-equal(it1.col1, it1.col2, ot1.col) (1) |
366 | |
367 | where the first part of the clause has references to inner tables, while the |
368 | second has references to the top-level tables, which is a violation of |
369 | (SJM-RULE). |
370 | |
371 | AND-clauses of this kind do not create problems, because make_cond_for_table() |
372 | will take them apart. OR-clauses will not be split. It is possible to |
373 | split-out the part that's dependent on the inner table: |
374 | |
375 | func(it.col2) OR it1.col1=it1.col2 |
376 | |
377 | but this is a less-restrictive condition than condition (1). Current execution |
378 | scheme will still try to generate the "remainder" condition: |
379 | |
380 | func(it.col2) OR it1.col1=ot1.col |
381 | |
382 | which is a violation of (SJM-RULE). |
383 | |
384 | QQ: "ot1.col=it1.col" is checked at the upper level. Why was it not removed |
385 | here? |
386 | AA: because has a proper subset of conditions that are found on this level. |
387 | consider a join order of ot, sjm(it) |
388 | and a condition |
389 | ot.col=it.col AND ( ot.col=it.col='foo' OR it.col2='bar') |
390 | |
391 | we will produce: |
392 | table ot: nothing |
393 | table it: ot.col=it.col AND (ot.col='foo' OR it.col2='bar') |
394 | ^^^^ ^^^^^^^^^^^^^^^^ |
395 | | \ the problem is that |
396 | | this part condition didnt |
397 | | receive a substitution |
398 | | |
399 | +--- it was correct to subst, 'ot' is |
400 | the left-most. |
401 | |
402 | |
403 | Does it make sense to push "inner=outer" down into ORs? |
404 | ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ |
405 | |
406 | Yes. Consider the query: |
407 | |
408 | select * from ot |
409 | where ot.col in (select it.col from it where (it.col='foo' OR it.col='bar')) |
410 | |
411 | here, it may be useful to infer that |
412 | |
413 | (ot.col='foo' OR ot.col='bar') (CASE-FOR-SUBST) |
414 | |
415 | and attach that condition to the table 'ot'. |
416 | |
417 | Possible solutions for Item_equals and ORs |
418 | ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ |
419 | |
420 | Solution #1 |
421 | ~~~~~~~~~~~ |
422 | Let make_cond_for_table() chop analyze the OR clauses it has produced and |
423 | discard them if they violate (SJM-RULE). This solution would allow to handle |
424 | cases like (CASE-FOR-SUBST) at the expense of making semantics of |
425 | make_cond_for_table() complicated. |
426 | |
427 | Solution #2 |
428 | ~~~~~~~~~~~ |
429 | Before the equality propagation phase, none of the OR clauses violate the |
430 | (SJM-RULE). This way, if we remember which tables the original equality |
431 | referred to, we can only generate equalities that refer to the outer (or inner) |
432 | tables. Note that this will disallow handling of cases like (CASE-FOR-SUBST). |
433 | |
434 | Currently, solution #2 is implemented. |
435 | */ |
436 | |
437 | LEX_CSTRING weedout_key= {STRING_WITH_LEN("weedout_key" )}; |
438 | |
439 | static |
440 | bool subquery_types_allow_materialization(Item_in_subselect *in_subs); |
441 | static bool replace_where_subcondition(JOIN *, Item **, Item *, Item *, bool); |
442 | static int subq_sj_candidate_cmp(Item_in_subselect* el1, Item_in_subselect* el2, |
443 | void *arg); |
444 | static bool convert_subq_to_sj(JOIN *parent_join, Item_in_subselect *subq_pred); |
445 | static bool convert_subq_to_jtbm(JOIN *parent_join, |
446 | Item_in_subselect *subq_pred, bool *remove); |
447 | static TABLE_LIST *alloc_join_nest(THD *thd); |
448 | static uint get_tmp_table_rec_length(Ref_ptr_array p_list, uint elements); |
449 | bool find_eq_ref_candidate(TABLE *table, table_map sj_inner_tables); |
450 | static SJ_MATERIALIZATION_INFO * |
451 | at_sjmat_pos(const JOIN *join, table_map remaining_tables, const JOIN_TAB *tab, |
452 | uint idx, bool *loose_scan); |
453 | void best_access_path(JOIN *join, JOIN_TAB *s, |
454 | table_map remaining_tables, uint idx, |
455 | bool disable_jbuf, double record_count, |
456 | POSITION *pos, POSITION *loose_scan_pos); |
457 | |
458 | static Item *create_subq_in_equalities(THD *thd, SJ_MATERIALIZATION_INFO *sjm, |
459 | Item_in_subselect *subq_pred); |
460 | static bool remove_sj_conds(THD *thd, Item **tree); |
461 | static bool is_cond_sj_in_equality(Item *item); |
462 | static bool sj_table_is_included(JOIN *join, JOIN_TAB *join_tab); |
463 | static Item *remove_additional_cond(Item* conds); |
464 | static void remove_subq_pushed_predicates(JOIN *join, Item **where); |
465 | |
466 | enum_nested_loop_state |
467 | end_sj_materialize(JOIN *join, JOIN_TAB *join_tab, bool end_of_records); |
468 | |
469 | |
470 | /* |
471 | Check if Materialization strategy is allowed for given subquery predicate. |
472 | |
473 | @param thd Thread handle |
474 | @param in_subs The subquery predicate |
475 | @param child_select The select inside predicate (the function will |
476 | check it is the only one) |
477 | |
478 | @return TRUE - Materialization is applicable |
479 | FALSE - Otherwise |
480 | */ |
481 | |
482 | bool is_materialization_applicable(THD *thd, Item_in_subselect *in_subs, |
483 | st_select_lex *child_select) |
484 | { |
485 | st_select_lex_unit* parent_unit= child_select->master_unit(); |
486 | /* |
487 | Check if the subquery predicate can be executed via materialization. |
488 | The required conditions are: |
489 | 0. The materialization optimizer switch was set. |
490 | 1. Subquery is a single SELECT (not a UNION). |
491 | TODO: this is a limitation that can be fixed |
492 | 2. Subquery is not a table-less query. In this case there is no |
493 | point in materializing. |
494 | 2A The upper query is not a table-less SELECT ... FROM DUAL. We |
495 | can't do materialization for SELECT .. FROM DUAL because it |
496 | does not call setup_subquery_materialization(). We could make |
497 | SELECT ... FROM DUAL call that function but that doesn't seem |
498 | to be the case that is worth handling. |
499 | 3. Either the subquery predicate is a top-level predicate, or at |
500 | least one partial match strategy is enabled. If no partial match |
501 | strategy is enabled, then materialization cannot be used for |
502 | non-top-level queries because it cannot handle NULLs correctly. |
503 | 4. Subquery is non-correlated |
504 | TODO: |
505 | This condition is too restrictive (limitation). It can be extended to: |
506 | (Subquery is non-correlated || |
507 | Subquery is correlated to any query outer to IN predicate || |
508 | (Subquery is correlated to the immediate outer query && |
509 | Subquery !contains {GROUP BY, ORDER BY [LIMIT], |
510 | aggregate functions}) && subquery predicate is not under "NOT IN")) |
511 | 5. Subquery does not contain recursive references |
512 | |
513 | A note about prepared statements: we want the if-branch to be taken on |
514 | PREPARE and each EXECUTE. The rewrites are only done once, but we need |
515 | select_lex->sj_subselects list to be populated for every EXECUTE. |
516 | |
517 | */ |
518 | if (optimizer_flag(thd, OPTIMIZER_SWITCH_MATERIALIZATION) && // 0 |
519 | !child_select->is_part_of_union() && // 1 |
520 | parent_unit->first_select()->leaf_tables.elements && // 2 |
521 | child_select->outer_select()->leaf_tables.elements && // 2A |
522 | subquery_types_allow_materialization(in_subs) && |
523 | (in_subs->is_top_level_item() || //3 |
524 | optimizer_flag(thd, |
525 | OPTIMIZER_SWITCH_PARTIAL_MATCH_ROWID_MERGE) || //3 |
526 | optimizer_flag(thd, |
527 | OPTIMIZER_SWITCH_PARTIAL_MATCH_TABLE_SCAN)) && //3 |
528 | !in_subs->is_correlated && //4 |
529 | !in_subs->with_recursive_reference) //5 |
530 | { |
531 | return TRUE; |
532 | } |
533 | return FALSE; |
534 | } |
535 | |
536 | |
537 | /* |
538 | Check if we need JOIN::prepare()-phase subquery rewrites and if yes, do them |
539 | |
540 | SYNOPSIS |
541 | check_and_do_in_subquery_rewrites() |
542 | join Subquery's join |
543 | |
544 | DESCRIPTION |
545 | Check if we need to do |
546 | - subquery -> mergeable semi-join rewrite |
547 | - if the subquery can be handled with materialization |
548 | - 'substitution' rewrite for table-less subqueries like "(select 1)" |
549 | - IN->EXISTS rewrite |
550 | and, depending on the rewrite, either do it, or record it to be done at a |
551 | later phase. |
552 | |
553 | RETURN |
554 | 0 - OK |
555 | Other - Some sort of query error |
556 | */ |
557 | |
558 | int check_and_do_in_subquery_rewrites(JOIN *join) |
559 | { |
560 | THD *thd=join->thd; |
561 | st_select_lex *select_lex= join->select_lex; |
562 | st_select_lex_unit* parent_unit= select_lex->master_unit(); |
563 | DBUG_ENTER("check_and_do_in_subquery_rewrites" ); |
564 | |
565 | /* |
566 | IN/ALL/ANY rewrites are not applicable for so called fake select |
567 | (this select exists only to filter results of union if it is needed). |
568 | */ |
569 | if (select_lex == select_lex->master_unit()->fake_select_lex) |
570 | DBUG_RETURN(0); |
571 | |
572 | /* |
573 | If |
574 | 1) this join is inside a subquery (of any type except FROM-clause |
575 | subquery) and |
576 | 2) we aren't just normalizing a VIEW |
577 | |
578 | Then perform early unconditional subquery transformations: |
579 | - Convert subquery predicate into semi-join, or |
580 | - Mark the subquery for execution using materialization, or |
581 | - Perform IN->EXISTS transformation, or |
582 | - Perform more/less ALL/ANY -> MIN/MAX rewrite |
583 | - Substitute trivial scalar-context subquery with its value |
584 | |
585 | TODO: for PS, make the whole block execute only on the first execution |
586 | */ |
587 | Item_subselect *subselect; |
588 | if (!thd->lex->is_view_context_analysis() && // (1) |
589 | (subselect= parent_unit->item)) // (2) |
590 | { |
591 | Item_in_subselect *in_subs= NULL; |
592 | Item_allany_subselect *allany_subs= NULL; |
593 | switch (subselect->substype()) { |
594 | case Item_subselect::IN_SUBS: |
595 | in_subs= (Item_in_subselect *)subselect; |
596 | break; |
597 | case Item_subselect::ALL_SUBS: |
598 | case Item_subselect::ANY_SUBS: |
599 | allany_subs= (Item_allany_subselect *)subselect; |
600 | break; |
601 | default: |
602 | break; |
603 | } |
604 | |
605 | |
606 | /* Resolve expressions and perform semantic analysis for IN query */ |
607 | if (in_subs != NULL) |
608 | /* |
609 | TODO: Add the condition below to this if statement when we have proper |
610 | support for is_correlated handling for materialized semijoins. |
611 | If we were to add this condition now, the fix_fields() call in |
612 | convert_subq_to_sj() would force the flag is_correlated to be set |
613 | erroneously for prepared queries. |
614 | |
615 | thd->stmt_arena->state != Query_arena::PREPARED) |
616 | */ |
617 | { |
618 | SELECT_LEX *current= thd->lex->current_select; |
619 | thd->lex->current_select= current->return_after_parsing(); |
620 | char const *save_where= thd->where; |
621 | thd->where= "IN/ALL/ANY subquery" ; |
622 | |
623 | bool failure= !in_subs->left_expr->fixed && |
624 | in_subs->left_expr->fix_fields(thd, &in_subs->left_expr); |
625 | thd->lex->current_select= current; |
626 | thd->where= save_where; |
627 | if (failure) |
628 | DBUG_RETURN(-1); /* purecov: deadcode */ |
629 | |
630 | /* |
631 | Check if the left and right expressions have the same # of |
632 | columns, i.e. we don't have a case like |
633 | (oe1, oe2) IN (SELECT ie1, ie2, ie3 ...) |
634 | |
635 | TODO why do we have this duplicated in IN->EXISTS transformers? |
636 | psergey-todo: fix these: grep for duplicated_subselect_card_check |
637 | */ |
638 | if (select_lex->item_list.elements != in_subs->left_expr->cols()) |
639 | { |
640 | my_error(ER_OPERAND_COLUMNS, MYF(0), in_subs->left_expr->cols()); |
641 | DBUG_RETURN(-1); |
642 | } |
643 | } |
644 | |
645 | DBUG_PRINT("info" , ("Checking if subq can be converted to semi-join" )); |
646 | /* |
647 | Check if we're in subquery that is a candidate for flattening into a |
648 | semi-join (which is done in flatten_subqueries()). The |
649 | requirements are: |
650 | 1. Subquery predicate is an IN/=ANY subq predicate |
651 | 2. Subquery is a single SELECT (not a UNION) |
652 | 3. Subquery does not have GROUP BY or ORDER BY |
653 | 4. Subquery does not use aggregate functions or HAVING |
654 | 5. Subquery predicate is at the AND-top-level of ON/WHERE clause |
655 | 6. We are not in a subquery of a single table UPDATE/DELETE that |
656 | doesn't have a JOIN (TODO: We should handle this at some |
657 | point by switching to multi-table UPDATE/DELETE) |
658 | 7. We're not in a table-less subquery like "SELECT 1" |
659 | 8. No execution method was already chosen (by a prepared statement) |
660 | 9. Parent select is not a table-less select |
661 | 10. Neither parent nor child select have STRAIGHT_JOIN option. |
662 | 11. It is first optimisation (the subquery could be moved from ON |
663 | clause during first optimisation and then be considered for SJ |
664 | on the second when it is too late) |
665 | */ |
666 | if (optimizer_flag(thd, OPTIMIZER_SWITCH_SEMIJOIN) && |
667 | in_subs && // 1 |
668 | !select_lex->is_part_of_union() && // 2 |
669 | !select_lex->group_list.elements && !join->order && // 3 |
670 | !join->having && !select_lex->with_sum_func && // 4 |
671 | in_subs->emb_on_expr_nest && // 5 |
672 | select_lex->outer_select()->join && // 6 |
673 | parent_unit->first_select()->leaf_tables.elements && // 7 |
674 | !in_subs->has_strategy() && // 8 |
675 | select_lex->outer_select()->leaf_tables.elements && // 9 |
676 | !((join->select_options | // 10 |
677 | select_lex->outer_select()->join->select_options) // 10 |
678 | & SELECT_STRAIGHT_JOIN) && // 10 |
679 | select_lex->first_cond_optimization) // 11 |
680 | { |
681 | DBUG_PRINT("info" , ("Subquery is semi-join conversion candidate" )); |
682 | |
683 | (void)subquery_types_allow_materialization(in_subs); |
684 | |
685 | in_subs->is_flattenable_semijoin= TRUE; |
686 | |
687 | /* Register the subquery for further processing in flatten_subqueries() */ |
688 | if (!in_subs->is_registered_semijoin) |
689 | { |
690 | Query_arena *arena, backup; |
691 | arena= thd->activate_stmt_arena_if_needed(&backup); |
692 | select_lex->outer_select()->sj_subselects.push_back(in_subs, |
693 | thd->mem_root); |
694 | if (arena) |
695 | thd->restore_active_arena(arena, &backup); |
696 | in_subs->is_registered_semijoin= TRUE; |
697 | } |
698 | } |
699 | else |
700 | { |
701 | DBUG_PRINT("info" , ("Subquery can't be converted to merged semi-join" )); |
702 | /* Test if the user has set a legal combination of optimizer switches. */ |
703 | if (!optimizer_flag(thd, OPTIMIZER_SWITCH_IN_TO_EXISTS) && |
704 | !optimizer_flag(thd, OPTIMIZER_SWITCH_MATERIALIZATION)) |
705 | my_error(ER_ILLEGAL_SUBQUERY_OPTIMIZER_SWITCHES, MYF(0)); |
706 | /* |
707 | Transform each subquery predicate according to its overloaded |
708 | transformer. |
709 | */ |
710 | if (subselect->select_transformer(join)) |
711 | DBUG_RETURN(-1); |
712 | |
713 | /* |
714 | If the subquery predicate is IN/=ANY, analyse and set all possible |
715 | subquery execution strategies based on optimizer switches and syntactic |
716 | properties. |
717 | */ |
718 | if (in_subs && !in_subs->has_strategy()) |
719 | { |
720 | if (is_materialization_applicable(thd, in_subs, select_lex)) |
721 | { |
722 | in_subs->add_strategy(SUBS_MATERIALIZATION); |
723 | |
724 | /* |
725 | If the subquery is an AND-part of WHERE register for being processed |
726 | with jtbm strategy |
727 | */ |
728 | if (in_subs->emb_on_expr_nest == NO_JOIN_NEST && |
729 | optimizer_flag(thd, OPTIMIZER_SWITCH_SEMIJOIN)) |
730 | { |
731 | in_subs->is_flattenable_semijoin= FALSE; |
732 | if (!in_subs->is_registered_semijoin) |
733 | { |
734 | Query_arena *arena, backup; |
735 | arena= thd->activate_stmt_arena_if_needed(&backup); |
736 | select_lex->outer_select()->sj_subselects.push_back(in_subs, |
737 | thd->mem_root); |
738 | if (arena) |
739 | thd->restore_active_arena(arena, &backup); |
740 | in_subs->is_registered_semijoin= TRUE; |
741 | } |
742 | } |
743 | } |
744 | |
745 | /* |
746 | IN-TO-EXISTS is the only universal strategy. Choose it if the user |
747 | allowed it via an optimizer switch, or if materialization is not |
748 | possible. |
749 | */ |
750 | if (optimizer_flag(thd, OPTIMIZER_SWITCH_IN_TO_EXISTS) || |
751 | !in_subs->has_strategy()) |
752 | in_subs->add_strategy(SUBS_IN_TO_EXISTS); |
753 | } |
754 | |
755 | /* Check if max/min optimization applicable */ |
756 | if (allany_subs && !allany_subs->is_set_strategy()) |
757 | { |
758 | uchar strategy= (allany_subs->is_maxmin_applicable(join) ? |
759 | (SUBS_MAXMIN_INJECTED | SUBS_MAXMIN_ENGINE) : |
760 | SUBS_IN_TO_EXISTS); |
761 | allany_subs->add_strategy(strategy); |
762 | } |
763 | |
764 | } |
765 | } |
766 | DBUG_RETURN(0); |
767 | } |
768 | |
769 | |
770 | /** |
771 | @brief Check if subquery's compared types allow materialization. |
772 | |
773 | @param in_subs Subquery predicate, updated as follows: |
774 | types_allow_materialization TRUE if subquery materialization is allowed. |
775 | sjm_scan_allowed If types_allow_materialization is TRUE, |
776 | indicates whether it is possible to use subquery |
777 | materialization and scan the materialized table. |
778 | |
779 | @retval TRUE If subquery types allow materialization. |
780 | @retval FALSE Otherwise. |
781 | |
782 | @details |
783 | This is a temporary fix for BUG#36752. |
784 | |
785 | There are two subquery materialization strategies: |
786 | |
787 | 1. Materialize and do index lookups in the materialized table. See |
788 | BUG#36752 for description of restrictions we need to put on the |
789 | compared expressions. |
790 | |
791 | 2. Materialize and then do a full scan of the materialized table. At the |
792 | moment, this strategy's applicability criteria are even stricter than |
793 | in #1. |
794 | |
795 | This is so because of the following: consider an uncorrelated subquery |
796 | |
797 | ...WHERE (ot1.col1, ot2.col2 ...) IN (SELECT ie1,ie2,... FROM it1 ...) |
798 | |
799 | and a join order that could be used to do sjm-materialization: |
800 | |
801 | SJM-Scan(it1, it1), ot1, ot2 |
802 | |
803 | IN-equalities will be parts of conditions attached to the outer tables: |
804 | |
805 | ot1: ot1.col1 = ie1 AND ... (C1) |
806 | ot2: ot1.col2 = ie2 AND ... (C2) |
807 | |
808 | besides those there may be additional references to ie1 and ie2 |
809 | generated by equality propagation. The problem with evaluating C1 and |
810 | C2 is that ie{1,2} refer to subquery tables' columns, while we only have |
811 | current value of materialization temptable. Our solution is to |
812 | * require that all ie{N} are table column references. This allows |
813 | to copy the values of materialization temptable columns to the |
814 | original table's columns (see setup_sj_materialization for more |
815 | details) |
816 | * require that compared columns have exactly the same type. This is |
817 | a temporary measure to avoid BUG#36752-type problems. |
818 | */ |
819 | |
820 | static |
821 | bool subquery_types_allow_materialization(Item_in_subselect *in_subs) |
822 | { |
823 | DBUG_ENTER("subquery_types_allow_materialization" ); |
824 | |
825 | DBUG_ASSERT(in_subs->left_expr->fixed); |
826 | |
827 | List_iterator<Item> it(in_subs->unit->first_select()->item_list); |
828 | uint elements= in_subs->unit->first_select()->item_list.elements; |
829 | |
830 | in_subs->types_allow_materialization= FALSE; // Assign default values |
831 | in_subs->sjm_scan_allowed= FALSE; |
832 | |
833 | bool all_are_fields= TRUE; |
834 | uint32 total_key_length = 0; |
835 | for (uint i= 0; i < elements; i++) |
836 | { |
837 | Item *outer= in_subs->left_expr->element_index(i); |
838 | Item *inner= it++; |
839 | all_are_fields &= (outer->real_item()->type() == Item::FIELD_ITEM && |
840 | inner->real_item()->type() == Item::FIELD_ITEM); |
841 | total_key_length += inner->max_length; |
842 | if (!inner->type_handler()->subquery_type_allows_materialization(inner, |
843 | outer)) |
844 | DBUG_RETURN(FALSE); |
845 | } |
846 | |
847 | /* |
848 | Make sure that create_tmp_table will not fail due to too long keys. |
849 | See MDEV-7122. This check is performed inside create_tmp_table also and |
850 | we must do it so that we know the table has keys created. |
851 | Make sure that the length of the key for the temp_table is atleast |
852 | greater than 0. |
853 | */ |
854 | if (!total_key_length || total_key_length > tmp_table_max_key_length() || |
855 | elements > tmp_table_max_key_parts()) |
856 | DBUG_RETURN(FALSE); |
857 | |
858 | in_subs->types_allow_materialization= TRUE; |
859 | in_subs->sjm_scan_allowed= all_are_fields; |
860 | DBUG_PRINT("info" ,("subquery_types_allow_materialization: ok, allowed" )); |
861 | DBUG_RETURN(TRUE); |
862 | } |
863 | |
864 | |
865 | /** |
866 | Apply max min optimization of all/any subselect |
867 | */ |
868 | |
869 | bool JOIN::transform_max_min_subquery() |
870 | { |
871 | DBUG_ENTER("JOIN::transform_max_min_subquery" ); |
872 | Item_subselect *subselect= unit->item; |
873 | if (!subselect || (subselect->substype() != Item_subselect::ALL_SUBS && |
874 | subselect->substype() != Item_subselect::ANY_SUBS)) |
875 | DBUG_RETURN(0); |
876 | DBUG_RETURN(((Item_allany_subselect *) subselect)-> |
877 | transform_into_max_min(this)); |
878 | } |
879 | |
880 | |
881 | /* |
882 | Finalize IN->EXISTS conversion in case we couldn't use materialization. |
883 | |
884 | DESCRIPTION Invoke the IN->EXISTS converter |
885 | Replace the Item_in_subselect with its wrapper Item_in_optimizer in WHERE. |
886 | |
887 | RETURN |
888 | FALSE - Ok |
889 | TRUE - Fatal error |
890 | */ |
891 | |
892 | bool make_in_exists_conversion(THD *thd, JOIN *join, Item_in_subselect *item) |
893 | { |
894 | DBUG_ENTER("make_in_exists_conversion" ); |
895 | JOIN *child_join= item->unit->first_select()->join; |
896 | bool res; |
897 | |
898 | /* |
899 | We're going to finalize IN->EXISTS conversion. |
900 | Normally, IN->EXISTS conversion takes place inside the |
901 | Item_subselect::fix_fields() call, where item_subselect->fixed==FALSE (as |
902 | fix_fields() haven't finished yet) and item_subselect->changed==FALSE (as |
903 | the conversion haven't been finalized) |
904 | |
905 | At the end of Item_subselect::fix_fields() we had to set fixed=TRUE, |
906 | changed=TRUE (the only other option would have been to return error). |
907 | |
908 | So, now we have to set these back for the duration of select_transformer() |
909 | call. |
910 | */ |
911 | item->changed= 0; |
912 | item->fixed= 0; |
913 | |
914 | SELECT_LEX *save_select_lex= thd->lex->current_select; |
915 | thd->lex->current_select= item->unit->first_select(); |
916 | |
917 | res= item->select_transformer(child_join); |
918 | |
919 | thd->lex->current_select= save_select_lex; |
920 | |
921 | if (res) |
922 | DBUG_RETURN(TRUE); |
923 | |
924 | item->changed= 1; |
925 | item->fixed= 1; |
926 | |
927 | Item *substitute= item->substitution; |
928 | bool do_fix_fields= !item->substitution->fixed; |
929 | /* |
930 | The Item_subselect has already been wrapped with Item_in_optimizer, so we |
931 | should search for item->optimizer, not 'item'. |
932 | */ |
933 | Item *replace_me= item->optimizer; |
934 | DBUG_ASSERT(replace_me==substitute); |
935 | |
936 | Item **tree= (item->emb_on_expr_nest == NO_JOIN_NEST)? |
937 | &join->conds : &(item->emb_on_expr_nest->on_expr); |
938 | if (replace_where_subcondition(join, tree, replace_me, substitute, |
939 | do_fix_fields)) |
940 | DBUG_RETURN(TRUE); |
941 | item->substitution= NULL; |
942 | |
943 | /* |
944 | If this is a prepared statement, repeat the above operation for |
945 | prep_where (or prep_on_expr). |
946 | */ |
947 | if (!thd->stmt_arena->is_conventional()) |
948 | { |
949 | tree= (item->emb_on_expr_nest == (TABLE_LIST*)NO_JOIN_NEST)? |
950 | &join->select_lex->prep_where : |
951 | &(item->emb_on_expr_nest->prep_on_expr); |
952 | |
953 | if (replace_where_subcondition(join, tree, replace_me, substitute, |
954 | FALSE)) |
955 | DBUG_RETURN(TRUE); |
956 | } |
957 | DBUG_RETURN(FALSE); |
958 | } |
959 | |
960 | |
961 | bool check_for_outer_joins(List<TABLE_LIST> *join_list) |
962 | { |
963 | TABLE_LIST *table; |
964 | NESTED_JOIN *nested_join; |
965 | List_iterator<TABLE_LIST> li(*join_list); |
966 | while ((table= li++)) |
967 | { |
968 | if ((nested_join= table->nested_join)) |
969 | { |
970 | if (check_for_outer_joins(&nested_join->join_list)) |
971 | return TRUE; |
972 | } |
973 | |
974 | if (table->outer_join) |
975 | return TRUE; |
976 | } |
977 | return FALSE; |
978 | } |
979 | |
980 | |
981 | void find_and_block_conversion_to_sj(Item *to_find, |
982 | List_iterator_fast<Item_in_subselect> &li) |
983 | { |
984 | if (to_find->type() == Item::FUNC_ITEM && |
985 | ((Item_func*)to_find)->functype() == Item_func::IN_OPTIMIZER_FUNC) |
986 | to_find= ((Item_in_optimizer*)to_find)->get_wrapped_in_subselect_item(); |
987 | |
988 | if (to_find->type() != Item::SUBSELECT_ITEM || |
989 | ((Item_subselect *) to_find)->substype() != Item_subselect::IN_SUBS) |
990 | return; |
991 | Item_in_subselect *in_subq; |
992 | li.rewind(); |
993 | while ((in_subq= li++)) |
994 | { |
995 | if (in_subq == to_find) |
996 | { |
997 | in_subq->block_conversion_to_sj(); |
998 | return; |
999 | } |
1000 | } |
1001 | } |
1002 | |
1003 | |
1004 | /* |
1005 | Convert semi-join subquery predicates into semi-join join nests |
1006 | |
1007 | SYNOPSIS |
1008 | convert_join_subqueries_to_semijoins() |
1009 | |
1010 | DESCRIPTION |
1011 | |
1012 | Convert candidate subquery predicates into semi-join join nests. This |
1013 | transformation is performed once in query lifetime and is irreversible. |
1014 | |
1015 | Conversion of one subquery predicate |
1016 | ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ |
1017 | We start with a join that has a semi-join subquery: |
1018 | |
1019 | SELECT ... |
1020 | FROM ot, ... |
1021 | WHERE oe IN (SELECT ie FROM it1 ... itN WHERE subq_where) AND outer_where |
1022 | |
1023 | and convert it into a semi-join nest: |
1024 | |
1025 | SELECT ... |
1026 | FROM ot SEMI JOIN (it1 ... itN), ... |
1027 | WHERE outer_where AND subq_where AND oe=ie |
1028 | |
1029 | that is, in order to do the conversion, we need to |
1030 | |
1031 | * Create the "SEMI JOIN (it1 .. itN)" part and add it into the parent |
1032 | query's FROM structure. |
1033 | * Add "AND subq_where AND oe=ie" into parent query's WHERE (or ON if |
1034 | the subquery predicate was in an ON expression) |
1035 | * Remove the subquery predicate from the parent query's WHERE |
1036 | |
1037 | Considerations when converting many predicates |
1038 | ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ |
1039 | A join may have at most MAX_TABLES tables. This may prevent us from |
1040 | flattening all subqueries when the total number of tables in parent and |
1041 | child selects exceeds MAX_TABLES. |
1042 | We deal with this problem by flattening children's subqueries first and |
1043 | then using a heuristic rule to determine each subquery predicate's |
1044 | "priority". |
1045 | |
1046 | RETURN |
1047 | FALSE OK |
1048 | TRUE Error |
1049 | */ |
1050 | |
1051 | bool convert_join_subqueries_to_semijoins(JOIN *join) |
1052 | { |
1053 | Query_arena *arena, backup; |
1054 | Item_in_subselect *in_subq; |
1055 | THD *thd= join->thd; |
1056 | DBUG_ENTER("convert_join_subqueries_to_semijoins" ); |
1057 | |
1058 | if (join->select_lex->sj_subselects.is_empty()) |
1059 | DBUG_RETURN(FALSE); |
1060 | |
1061 | List_iterator_fast<Item_in_subselect> li(join->select_lex->sj_subselects); |
1062 | |
1063 | while ((in_subq= li++)) |
1064 | { |
1065 | SELECT_LEX *subq_sel= in_subq->get_select_lex(); |
1066 | if (subq_sel->handle_derived(thd->lex, DT_MERGE)) |
1067 | DBUG_RETURN(TRUE); |
1068 | if (subq_sel->join->transform_in_predicates_into_in_subq(thd)) |
1069 | DBUG_RETURN(TRUE); |
1070 | subq_sel->update_used_tables(); |
1071 | } |
1072 | |
1073 | /* |
1074 | Check all candidates to semi-join conversion that occur |
1075 | in ON expressions of outer join. Set the flag blocking |
1076 | this conversion for them. |
1077 | */ |
1078 | TABLE_LIST *tbl; |
1079 | List_iterator<TABLE_LIST> ti(join->select_lex->leaf_tables); |
1080 | while ((tbl= ti++)) |
1081 | { |
1082 | TABLE_LIST *embedded; |
1083 | TABLE_LIST *embedding= tbl; |
1084 | do |
1085 | { |
1086 | embedded= embedding; |
1087 | bool block_conversion_to_sj= false; |
1088 | if (embedded->on_expr) |
1089 | { |
1090 | /* |
1091 | Conversion of an IN subquery predicate into semi-join |
1092 | is blocked now if the predicate occurs: |
1093 | - in the ON expression of an outer join |
1094 | - in the ON expression of an inner join embedded directly |
1095 | or indirectly in the inner nest of an outer join |
1096 | */ |
1097 | for (TABLE_LIST *tl= embedded; tl; tl= tl->embedding) |
1098 | { |
1099 | if (tl->outer_join) |
1100 | { |
1101 | block_conversion_to_sj= true; |
1102 | break; |
1103 | } |
1104 | } |
1105 | } |
1106 | if (block_conversion_to_sj) |
1107 | { |
1108 | Item *cond= embedded->on_expr; |
1109 | if (!cond) |
1110 | ; |
1111 | else if (cond->type() != Item::COND_ITEM) |
1112 | find_and_block_conversion_to_sj(cond, li); |
1113 | else if (((Item_cond*) cond)->functype() == |
1114 | Item_func::COND_AND_FUNC) |
1115 | { |
1116 | Item *item; |
1117 | List_iterator<Item> it(*(((Item_cond*) cond)->argument_list())); |
1118 | while ((item= it++)) |
1119 | { |
1120 | find_and_block_conversion_to_sj(item, li); |
1121 | } |
1122 | } |
1123 | } |
1124 | embedding= embedded->embedding; |
1125 | } |
1126 | while (embedding && |
1127 | embedding->nested_join->join_list.head() == embedded); |
1128 | } |
1129 | |
1130 | /* |
1131 | Block conversion to semi-joins for those candidates that |
1132 | are encountered in the WHERE condition of the multi-table view |
1133 | with CHECK OPTION if this view is used in UPDATE/DELETE. |
1134 | (This limitation can be, probably, easily lifted.) |
1135 | */ |
1136 | li.rewind(); |
1137 | while ((in_subq= li++)) |
1138 | { |
1139 | if (in_subq->emb_on_expr_nest != NO_JOIN_NEST && |
1140 | in_subq->emb_on_expr_nest->effective_with_check) |
1141 | { |
1142 | in_subq->block_conversion_to_sj(); |
1143 | } |
1144 | } |
1145 | |
1146 | if (join->select_options & SELECT_STRAIGHT_JOIN) |
1147 | { |
1148 | /* Block conversion to semijoins for all candidates */ |
1149 | li.rewind(); |
1150 | while ((in_subq= li++)) |
1151 | { |
1152 | in_subq->block_conversion_to_sj(); |
1153 | } |
1154 | } |
1155 | |
1156 | li.rewind(); |
1157 | /* First, convert child join's subqueries. We proceed bottom-up here */ |
1158 | while ((in_subq= li++)) |
1159 | { |
1160 | st_select_lex *child_select= in_subq->get_select_lex(); |
1161 | JOIN *child_join= child_select->join; |
1162 | child_join->outer_tables = child_join->table_count; |
1163 | |
1164 | /* |
1165 | child_select->where contains only the WHERE predicate of the |
1166 | subquery itself here. We may be selecting from a VIEW, which has its |
1167 | own predicate. The combined predicates are available in child_join->conds, |
1168 | which was built by setup_conds() doing prepare_where() for all views. |
1169 | */ |
1170 | child_select->where= child_join->conds; |
1171 | |
1172 | if (convert_join_subqueries_to_semijoins(child_join)) |
1173 | DBUG_RETURN(TRUE); |
1174 | |
1175 | |
1176 | in_subq->sj_convert_priority= |
1177 | MY_TEST(in_subq->do_not_convert_to_sj) * MAX_TABLES * 2 + |
1178 | in_subq->is_correlated * MAX_TABLES + child_join->outer_tables; |
1179 | } |
1180 | |
1181 | // Temporary measure: disable semi-joins when they are together with outer |
1182 | // joins. |
1183 | #if 0 |
1184 | if (check_for_outer_joins(join->join_list)) |
1185 | { |
1186 | in_subq= join->select_lex->sj_subselects.head(); |
1187 | arena= thd->activate_stmt_arena_if_needed(&backup); |
1188 | goto skip_conversion; |
1189 | } |
1190 | #endif |
1191 | //dump_TABLE_LIST_struct(select_lex, select_lex->leaf_tables); |
1192 | /* |
1193 | 2. Pick which subqueries to convert: |
1194 | sort the subquery array |
1195 | - prefer correlated subqueries over uncorrelated; |
1196 | - prefer subqueries that have greater number of outer tables; |
1197 | */ |
1198 | bubble_sort<Item_in_subselect>(&join->select_lex->sj_subselects, |
1199 | subq_sj_candidate_cmp, NULL); |
1200 | // #tables-in-parent-query + #tables-in-subquery < MAX_TABLES |
1201 | /* Replace all subqueries to be flattened with Item_int(1) */ |
1202 | arena= thd->activate_stmt_arena_if_needed(&backup); |
1203 | |
1204 | li.rewind(); |
1205 | while ((in_subq= li++)) |
1206 | { |
1207 | bool remove_item= TRUE; |
1208 | |
1209 | /* Stop processing if we've reached a subquery that's attached to the ON clause */ |
1210 | if (in_subq->do_not_convert_to_sj) |
1211 | break; |
1212 | |
1213 | if (in_subq->is_flattenable_semijoin) |
1214 | { |
1215 | if (join->table_count + |
1216 | in_subq->unit->first_select()->join->table_count >= MAX_TABLES) |
1217 | break; |
1218 | if (convert_subq_to_sj(join, in_subq)) |
1219 | goto restore_arena_and_fail; |
1220 | } |
1221 | else |
1222 | { |
1223 | if (join->table_count + 1 >= MAX_TABLES) |
1224 | break; |
1225 | if (convert_subq_to_jtbm(join, in_subq, &remove_item)) |
1226 | goto restore_arena_and_fail; |
1227 | } |
1228 | if (remove_item) |
1229 | { |
1230 | Item **tree= (in_subq->emb_on_expr_nest == NO_JOIN_NEST)? |
1231 | &join->conds : &(in_subq->emb_on_expr_nest->on_expr); |
1232 | Item *replace_me= in_subq->original_item(); |
1233 | if (replace_where_subcondition(join, tree, replace_me, |
1234 | new (thd->mem_root) Item_int(thd, 1), |
1235 | FALSE)) |
1236 | goto restore_arena_and_fail; |
1237 | } |
1238 | } |
1239 | //skip_conversion: |
1240 | /* |
1241 | 3. Finalize (perform IN->EXISTS rewrite) the subqueries that we didn't |
1242 | convert: |
1243 | */ |
1244 | while (in_subq) |
1245 | { |
1246 | JOIN *child_join= in_subq->unit->first_select()->join; |
1247 | in_subq->changed= 0; |
1248 | in_subq->fixed= 0; |
1249 | |
1250 | SELECT_LEX *save_select_lex= thd->lex->current_select; |
1251 | thd->lex->current_select= in_subq->unit->first_select(); |
1252 | |
1253 | bool res= in_subq->select_transformer(child_join); |
1254 | |
1255 | thd->lex->current_select= save_select_lex; |
1256 | |
1257 | if (res) |
1258 | DBUG_RETURN(TRUE); |
1259 | |
1260 | in_subq->changed= 1; |
1261 | in_subq->fixed= 1; |
1262 | |
1263 | Item *substitute= in_subq->substitution; |
1264 | bool do_fix_fields= !in_subq->substitution->fixed; |
1265 | Item **tree= (in_subq->emb_on_expr_nest == NO_JOIN_NEST)? |
1266 | &join->conds : &(in_subq->emb_on_expr_nest->on_expr); |
1267 | Item *replace_me= in_subq->original_item(); |
1268 | if (replace_where_subcondition(join, tree, replace_me, substitute, |
1269 | do_fix_fields)) |
1270 | DBUG_RETURN(TRUE); |
1271 | in_subq->substitution= NULL; |
1272 | /* |
1273 | If this is a prepared statement, repeat the above operation for |
1274 | prep_where (or prep_on_expr). Subquery-to-semijoin conversion is |
1275 | done once for prepared statement. |
1276 | */ |
1277 | if (!thd->stmt_arena->is_conventional()) |
1278 | { |
1279 | tree= (in_subq->emb_on_expr_nest == NO_JOIN_NEST)? |
1280 | &join->select_lex->prep_where : |
1281 | &(in_subq->emb_on_expr_nest->prep_on_expr); |
1282 | /* |
1283 | prep_on_expr/ prep_where may be NULL in some cases. |
1284 | If that is the case, do nothing - simplify_joins() will copy |
1285 | ON/WHERE expression into prep_on_expr/prep_where. |
1286 | */ |
1287 | if (*tree && replace_where_subcondition(join, tree, replace_me, substitute, |
1288 | FALSE)) |
1289 | DBUG_RETURN(TRUE); |
1290 | } |
1291 | /* |
1292 | Revert to the IN->EXISTS strategy in the rare case when the subquery could |
1293 | not be flattened. |
1294 | */ |
1295 | in_subq->reset_strategy(SUBS_IN_TO_EXISTS); |
1296 | if (is_materialization_applicable(thd, in_subq, |
1297 | in_subq->unit->first_select())) |
1298 | { |
1299 | in_subq->add_strategy(SUBS_MATERIALIZATION); |
1300 | } |
1301 | |
1302 | in_subq= li++; |
1303 | } |
1304 | |
1305 | if (arena) |
1306 | thd->restore_active_arena(arena, &backup); |
1307 | join->select_lex->sj_subselects.empty(); |
1308 | DBUG_RETURN(FALSE); |
1309 | |
1310 | restore_arena_and_fail: |
1311 | if (arena) |
1312 | thd->restore_active_arena(arena, &backup); |
1313 | DBUG_RETURN(TRUE); |
1314 | } |
1315 | |
1316 | |
1317 | /* |
1318 | Get #output_rows and scan_time estimates for a "delayed" table. |
1319 | |
1320 | SYNOPSIS |
1321 | get_delayed_table_estimates() |
1322 | table IN Table to get estimates for |
1323 | out_rows OUT E(#rows in the table) |
1324 | scan_time OUT E(scan_time). |
1325 | startup_cost OUT cost to populate the table. |
1326 | |
1327 | DESCRIPTION |
1328 | Get #output_rows and scan_time estimates for a "delayed" table. By |
1329 | "delayed" here we mean that the table is filled at the start of query |
1330 | execution. This means that the optimizer can't use table statistics to |
1331 | get #rows estimate for it, it has to call this function instead. |
1332 | |
1333 | This function is expected to make different actions depending on the nature |
1334 | of the table. At the moment there is only one kind of delayed tables, |
1335 | non-flattenable semi-joins. |
1336 | */ |
1337 | |
1338 | void get_delayed_table_estimates(TABLE *table, |
1339 | ha_rows *out_rows, |
1340 | double *scan_time, |
1341 | double *startup_cost) |
1342 | { |
1343 | Item_in_subselect *item= table->pos_in_table_list->jtbm_subselect; |
1344 | |
1345 | DBUG_ASSERT(item->engine->engine_type() == |
1346 | subselect_engine::HASH_SJ_ENGINE); |
1347 | |
1348 | subselect_hash_sj_engine *hash_sj_engine= |
1349 | ((subselect_hash_sj_engine*)item->engine); |
1350 | |
1351 | *out_rows= (ha_rows)item->jtbm_record_count; |
1352 | *startup_cost= item->jtbm_read_time; |
1353 | |
1354 | /* Calculate cost of scanning the temptable */ |
1355 | double data_size= item->jtbm_record_count * |
1356 | hash_sj_engine->tmp_table->s->reclength; |
1357 | /* Do like in handler::read_time */ |
1358 | *scan_time= data_size/IO_SIZE + 2; |
1359 | } |
1360 | |
1361 | |
1362 | /** |
1363 | @brief Replaces an expression destructively inside the expression tree of |
1364 | the WHERE clase. |
1365 | |
1366 | @note We substitute AND/OR structure because it was copied by |
1367 | copy_andor_structure and some changes could be done in the copy but |
1368 | should be left permanent, also there could be several layers of AND over |
1369 | AND and OR over OR because ::fix_field() possibly is not called. |
1370 | |
1371 | @param join The top-level query. |
1372 | @param old_cond The expression to be replaced. |
1373 | @param new_cond The expression to be substituted. |
1374 | @param do_fix_fields If true, Item::fix_fields(THD*, Item**) is called for |
1375 | the new expression. |
1376 | @return <code>true</code> if there was an error, <code>false</code> if |
1377 | successful. |
1378 | */ |
1379 | |
1380 | static bool replace_where_subcondition(JOIN *join, Item **expr, |
1381 | Item *old_cond, Item *new_cond, |
1382 | bool do_fix_fields) |
1383 | { |
1384 | if (*expr == old_cond) |
1385 | { |
1386 | *expr= new_cond; |
1387 | if (do_fix_fields) |
1388 | new_cond->fix_fields(join->thd, expr); |
1389 | return FALSE; |
1390 | } |
1391 | |
1392 | if ((*expr)->type() == Item::COND_ITEM) |
1393 | { |
1394 | List_iterator<Item> li(*((Item_cond*)(*expr))->argument_list()); |
1395 | Item *item; |
1396 | while ((item= li++)) |
1397 | { |
1398 | if (item == old_cond) |
1399 | { |
1400 | li.replace(new_cond); |
1401 | if (do_fix_fields) |
1402 | new_cond->fix_fields(join->thd, li.ref()); |
1403 | return FALSE; |
1404 | } |
1405 | else if (item->type() == Item::COND_ITEM) |
1406 | { |
1407 | replace_where_subcondition(join, li.ref(), |
1408 | old_cond, new_cond, |
1409 | do_fix_fields); |
1410 | } |
1411 | } |
1412 | } |
1413 | /* |
1414 | We can come to here when |
1415 | - we're doing replace operations on both on_expr and prep_on_expr |
1416 | - on_expr is the same as prep_on_expr, or they share a sub-tree |
1417 | (so, when we do replace in on_expr, we replace in prep_on_expr, too, |
1418 | and when we try doing a replace in prep_on_expr, the item we wanted |
1419 | to replace there has already been replaced) |
1420 | */ |
1421 | return FALSE; |
1422 | } |
1423 | |
1424 | static int subq_sj_candidate_cmp(Item_in_subselect* el1, Item_in_subselect* el2, |
1425 | void *arg) |
1426 | { |
1427 | return (el1->sj_convert_priority > el2->sj_convert_priority) ? -1 : |
1428 | ( (el1->sj_convert_priority == el2->sj_convert_priority)? 0 : 1); |
1429 | } |
1430 | |
1431 | |
1432 | /* |
1433 | Convert a subquery predicate into a TABLE_LIST semi-join nest |
1434 | |
1435 | SYNOPSIS |
1436 | convert_subq_to_sj() |
1437 | parent_join Parent join, the one that has subq_pred in its WHERE/ON |
1438 | clause |
1439 | subq_pred Subquery predicate to be converted |
1440 | |
1441 | DESCRIPTION |
1442 | Convert a subquery predicate into a TABLE_LIST semi-join nest. All the |
1443 | prerequisites are already checked, so the conversion is always successfull. |
1444 | |
1445 | Prepared Statements: the transformation is permanent: |
1446 | - Changes in TABLE_LIST structures are naturally permanent |
1447 | - Item tree changes are performed on statement MEM_ROOT: |
1448 | = we activate statement MEM_ROOT |
1449 | = this function is called before the first fix_prepare_information |
1450 | call. |
1451 | |
1452 | This is intended because the criteria for subquery-to-sj conversion remain |
1453 | constant for the lifetime of the Prepared Statement. |
1454 | |
1455 | RETURN |
1456 | FALSE OK |
1457 | TRUE Out of memory error |
1458 | */ |
1459 | |
1460 | static bool convert_subq_to_sj(JOIN *parent_join, Item_in_subselect *subq_pred) |
1461 | { |
1462 | SELECT_LEX *parent_lex= parent_join->select_lex; |
1463 | TABLE_LIST *emb_tbl_nest= NULL; |
1464 | List<TABLE_LIST> *emb_join_list= &parent_lex->top_join_list; |
1465 | THD *thd= parent_join->thd; |
1466 | DBUG_ENTER("convert_subq_to_sj" ); |
1467 | |
1468 | /* |
1469 | 1. Find out where to put the predicate into. |
1470 | Note: for "t1 LEFT JOIN t2" this will be t2, a leaf. |
1471 | */ |
1472 | if ((void*)subq_pred->emb_on_expr_nest != (void*)NO_JOIN_NEST) |
1473 | { |
1474 | if (subq_pred->emb_on_expr_nest->nested_join) |
1475 | { |
1476 | /* |
1477 | We're dealing with |
1478 | |
1479 | ... [LEFT] JOIN ( ... ) ON (subquery AND whatever) ... |
1480 | |
1481 | The sj-nest will be inserted into the brackets nest. |
1482 | */ |
1483 | emb_tbl_nest= subq_pred->emb_on_expr_nest; |
1484 | emb_join_list= &emb_tbl_nest->nested_join->join_list; |
1485 | } |
1486 | else if (!subq_pred->emb_on_expr_nest->outer_join) |
1487 | { |
1488 | /* |
1489 | We're dealing with |
1490 | |
1491 | ... INNER JOIN tblX ON (subquery AND whatever) ... |
1492 | |
1493 | The sj-nest will be tblX's "sibling", i.e. another child of its |
1494 | parent. This is ok because tblX is joined as an inner join. |
1495 | */ |
1496 | emb_tbl_nest= subq_pred->emb_on_expr_nest->embedding; |
1497 | if (emb_tbl_nest) |
1498 | emb_join_list= &emb_tbl_nest->nested_join->join_list; |
1499 | } |
1500 | else if (!subq_pred->emb_on_expr_nest->nested_join) |
1501 | { |
1502 | TABLE_LIST *outer_tbl= subq_pred->emb_on_expr_nest; |
1503 | TABLE_LIST *wrap_nest; |
1504 | LEX_CSTRING sj_wrap_name= { STRING_WITH_LEN("(sj-wrap)" ) }; |
1505 | /* |
1506 | We're dealing with |
1507 | |
1508 | ... LEFT JOIN tbl ON (on_expr AND subq_pred) ... |
1509 | |
1510 | we'll need to convert it into: |
1511 | |
1512 | ... LEFT JOIN ( tbl SJ (subq_tables) ) ON (on_expr AND subq_pred) ... |
1513 | | | |
1514 | |<----- wrap_nest ---->| |
1515 | |
1516 | Q: other subqueries may be pointing to this element. What to do? |
1517 | A1: simple solution: copy *subq_pred->expr_join_nest= *parent_nest. |
1518 | But we'll need to fix other pointers. |
1519 | A2: Another way: have TABLE_LIST::next_ptr so the following |
1520 | subqueries know the table has been nested. |
1521 | A3: changes in the TABLE_LIST::outer_join will make everything work |
1522 | automatically. |
1523 | */ |
1524 | if (!(wrap_nest= alloc_join_nest(thd))) |
1525 | { |
1526 | DBUG_RETURN(TRUE); |
1527 | } |
1528 | wrap_nest->embedding= outer_tbl->embedding; |
1529 | wrap_nest->join_list= outer_tbl->join_list; |
1530 | wrap_nest->alias= sj_wrap_name; |
1531 | |
1532 | wrap_nest->nested_join->join_list.empty(); |
1533 | wrap_nest->nested_join->join_list.push_back(outer_tbl, thd->mem_root); |
1534 | |
1535 | outer_tbl->embedding= wrap_nest; |
1536 | outer_tbl->join_list= &wrap_nest->nested_join->join_list; |
1537 | |
1538 | /* |
1539 | wrap_nest will take place of outer_tbl, so move the outer join flag |
1540 | and on_expr |
1541 | */ |
1542 | wrap_nest->outer_join= outer_tbl->outer_join; |
1543 | outer_tbl->outer_join= 0; |
1544 | |
1545 | wrap_nest->on_expr= outer_tbl->on_expr; |
1546 | outer_tbl->on_expr= NULL; |
1547 | |
1548 | List_iterator<TABLE_LIST> li(*wrap_nest->join_list); |
1549 | TABLE_LIST *tbl; |
1550 | while ((tbl= li++)) |
1551 | { |
1552 | if (tbl == outer_tbl) |
1553 | { |
1554 | li.replace(wrap_nest); |
1555 | break; |
1556 | } |
1557 | } |
1558 | /* |
1559 | Ok now wrap_nest 'contains' outer_tbl and we're ready to add the |
1560 | semi-join nest into it |
1561 | */ |
1562 | emb_join_list= &wrap_nest->nested_join->join_list; |
1563 | emb_tbl_nest= wrap_nest; |
1564 | } |
1565 | } |
1566 | |
1567 | TABLE_LIST *sj_nest; |
1568 | NESTED_JOIN *nested_join; |
1569 | LEX_CSTRING sj_nest_name= { STRING_WITH_LEN("(sj-nest)" ) }; |
1570 | if (!(sj_nest= alloc_join_nest(thd))) |
1571 | { |
1572 | DBUG_RETURN(TRUE); |
1573 | } |
1574 | nested_join= sj_nest->nested_join; |
1575 | |
1576 | sj_nest->join_list= emb_join_list; |
1577 | sj_nest->embedding= emb_tbl_nest; |
1578 | sj_nest->alias= sj_nest_name; |
1579 | sj_nest->sj_subq_pred= subq_pred; |
1580 | sj_nest->original_subq_pred_used_tables= subq_pred->used_tables() | |
1581 | subq_pred->left_expr->used_tables(); |
1582 | /* Nests do not participate in those 'chains', so: */ |
1583 | /* sj_nest->next_leaf= sj_nest->next_local= sj_nest->next_global == NULL*/ |
1584 | emb_join_list->push_back(sj_nest, thd->mem_root); |
1585 | |
1586 | /* |
1587 | nested_join->used_tables and nested_join->not_null_tables are |
1588 | initialized in simplify_joins(). |
1589 | */ |
1590 | |
1591 | /* |
1592 | 2. Walk through subquery's top list and set 'embedding' to point to the |
1593 | sj-nest. |
1594 | */ |
1595 | st_select_lex *subq_lex= subq_pred->unit->first_select(); |
1596 | DBUG_ASSERT(subq_lex->next_select() == NULL); |
1597 | nested_join->join_list.empty(); |
1598 | List_iterator_fast<TABLE_LIST> li(subq_lex->top_join_list); |
1599 | TABLE_LIST *tl; |
1600 | while ((tl= li++)) |
1601 | { |
1602 | tl->embedding= sj_nest; |
1603 | tl->join_list= &nested_join->join_list; |
1604 | nested_join->join_list.push_back(tl, thd->mem_root); |
1605 | } |
1606 | |
1607 | /* |
1608 | Reconnect the next_leaf chain. |
1609 | TODO: Do we have to put subquery's tables at the end of the chain? |
1610 | Inserting them at the beginning would be a bit faster. |
1611 | NOTE: We actually insert them at the front! That's because the order is |
1612 | reversed in this list. |
1613 | */ |
1614 | parent_lex->leaf_tables.append(&subq_lex->leaf_tables); |
1615 | |
1616 | if (subq_lex->options & OPTION_SCHEMA_TABLE) |
1617 | parent_lex->options |= OPTION_SCHEMA_TABLE; |
1618 | |
1619 | /* |
1620 | Same as above for next_local chain |
1621 | (a theory: a next_local chain always starts with ::leaf_tables |
1622 | because view's tables are inserted after the view) |
1623 | */ |
1624 | |
1625 | for (tl= (TABLE_LIST*)(parent_lex->table_list.first); tl->next_local; tl= tl->next_local) |
1626 | {} |
1627 | |
1628 | tl->next_local= subq_lex->join->tables_list; |
1629 | |
1630 | /* A theory: no need to re-connect the next_global chain */ |
1631 | |
1632 | /* 3. Remove the original subquery predicate from the WHERE/ON */ |
1633 | |
1634 | // The subqueries were replaced for Item_int(1) earlier |
1635 | subq_pred->reset_strategy(SUBS_SEMI_JOIN); // for subsequent executions |
1636 | /*TODO: also reset the 'm_with_subquery' there. */ |
1637 | |
1638 | /* n. Adjust the parent_join->table_count counter */ |
1639 | uint table_no= parent_join->table_count; |
1640 | /* n. Walk through child's tables and adjust table->map */ |
1641 | List_iterator_fast<TABLE_LIST> si(subq_lex->leaf_tables); |
1642 | while ((tl= si++)) |
1643 | { |
1644 | tl->set_tablenr(table_no); |
1645 | if (tl->is_jtbm()) |
1646 | { |
1647 | tl->jtbm_table_no= table_no; |
1648 | Item *dummy= tl->jtbm_subselect; |
1649 | tl->jtbm_subselect->fix_after_pullout(parent_lex, &dummy, true); |
1650 | DBUG_ASSERT(dummy == tl->jtbm_subselect); |
1651 | } |
1652 | SELECT_LEX *old_sl= tl->select_lex; |
1653 | tl->select_lex= parent_join->select_lex; |
1654 | for (TABLE_LIST *emb= tl->embedding; |
1655 | emb && emb->select_lex == old_sl; |
1656 | emb= emb->embedding) |
1657 | emb->select_lex= parent_join->select_lex; |
1658 | table_no++; |
1659 | } |
1660 | parent_join->table_count += subq_lex->join->table_count; |
1661 | //parent_join->table_count += subq_lex->leaf_tables.elements; |
1662 | |
1663 | /* |
1664 | Put the subquery's WHERE into semi-join's sj_on_expr |
1665 | Add the subquery-induced equalities too. |
1666 | */ |
1667 | SELECT_LEX *save_lex= thd->lex->current_select; |
1668 | thd->lex->current_select=subq_lex; |
1669 | if (!subq_pred->left_expr->fixed && |
1670 | subq_pred->left_expr->fix_fields(thd, &subq_pred->left_expr)) |
1671 | DBUG_RETURN(TRUE); |
1672 | thd->lex->current_select=save_lex; |
1673 | |
1674 | table_map subq_pred_used_tables= subq_pred->used_tables(); |
1675 | sj_nest->nested_join->sj_corr_tables= subq_pred_used_tables; |
1676 | sj_nest->nested_join->sj_depends_on= subq_pred_used_tables | |
1677 | subq_pred->left_expr->used_tables(); |
1678 | sj_nest->sj_on_expr= subq_lex->join->conds; |
1679 | |
1680 | /* |
1681 | Create the IN-equalities and inject them into semi-join's ON expression. |
1682 | Additionally, for LooseScan strategy |
1683 | - Record the number of IN-equalities. |
1684 | - Create list of pointers to (oe1, ..., ieN). We'll need the list to |
1685 | see which of the expressions are bound and which are not (for those |
1686 | we'll produce a distinct stream of (ie_i1,...ie_ik). |
1687 | |
1688 | (TODO: can we just create a list of pointers and hope the expressions |
1689 | will not substitute themselves on fix_fields()? or we need to wrap |
1690 | them into Item_direct_view_refs and store pointers to those. The |
1691 | pointers to Item_direct_view_refs are guaranteed to be stable as |
1692 | Item_direct_view_refs doesn't substitute itself with anything in |
1693 | Item_direct_view_ref::fix_fields. |
1694 | */ |
1695 | sj_nest->sj_in_exprs= subq_pred->left_expr->cols(); |
1696 | sj_nest->nested_join->sj_outer_expr_list.empty(); |
1697 | |
1698 | if (subq_pred->left_expr->cols() == 1) |
1699 | { |
1700 | /* add left = select_list_element */ |
1701 | nested_join->sj_outer_expr_list.push_back(&subq_pred->left_expr, |
1702 | thd->mem_root); |
1703 | /* |
1704 | Create Item_func_eq. Note that |
1705 | 1. this is done on the statement, not execution, arena |
1706 | 2. if it's a PS then this happens only once - on the first execution. |
1707 | On following re-executions, the item will be fix_field-ed normally. |
1708 | 3. Thus it should be created as if it was fix_field'ed, in particular |
1709 | all pointers to items in the execution arena should be protected |
1710 | with thd->change_item_tree |
1711 | */ |
1712 | Item_func_eq *item_eq= |
1713 | new (thd->mem_root) Item_func_eq(thd, subq_pred->left_expr_orig, |
1714 | subq_lex->ref_pointer_array[0]); |
1715 | if (!item_eq) |
1716 | DBUG_RETURN(TRUE); |
1717 | if (subq_pred->left_expr_orig != subq_pred->left_expr) |
1718 | thd->change_item_tree(item_eq->arguments(), subq_pred->left_expr); |
1719 | item_eq->in_equality_no= 0; |
1720 | sj_nest->sj_on_expr= and_items(thd, sj_nest->sj_on_expr, item_eq); |
1721 | } |
1722 | else if (subq_pred->left_expr->type() == Item::ROW_ITEM) |
1723 | { |
1724 | /* |
1725 | disassemple left expression and add |
1726 | left1 = select_list_element1 and left2 = select_list_element2 ... |
1727 | */ |
1728 | for (uint i= 0; i < subq_pred->left_expr->cols(); i++) |
1729 | { |
1730 | nested_join->sj_outer_expr_list.push_back(subq_pred->left_expr->addr(i), |
1731 | thd->mem_root); |
1732 | Item_func_eq *item_eq= |
1733 | new (thd->mem_root) |
1734 | Item_func_eq(thd, subq_pred->left_expr_orig->element_index(i), |
1735 | subq_lex->ref_pointer_array[i]); |
1736 | if (!item_eq) |
1737 | DBUG_RETURN(TRUE); |
1738 | DBUG_ASSERT(subq_pred->left_expr->element_index(i)->fixed); |
1739 | if (subq_pred->left_expr_orig->element_index(i) != |
1740 | subq_pred->left_expr->element_index(i)) |
1741 | thd->change_item_tree(item_eq->arguments(), |
1742 | subq_pred->left_expr->element_index(i)); |
1743 | item_eq->in_equality_no= i; |
1744 | sj_nest->sj_on_expr= and_items(thd, sj_nest->sj_on_expr, item_eq); |
1745 | } |
1746 | } |
1747 | else |
1748 | { |
1749 | /* |
1750 | add row operation |
1751 | left = (select_list_element1, select_list_element2, ...) |
1752 | */ |
1753 | Item_row *row= new (thd->mem_root) Item_row(thd, subq_lex->pre_fix); |
1754 | /* fix fields on subquery was call so they should be the same */ |
1755 | if (!row) |
1756 | DBUG_RETURN(TRUE); |
1757 | DBUG_ASSERT(subq_pred->left_expr->cols() == row->cols()); |
1758 | nested_join->sj_outer_expr_list.push_back(&subq_pred->left_expr); |
1759 | Item_func_eq *item_eq= |
1760 | new (thd->mem_root) Item_func_eq(thd, subq_pred->left_expr_orig, row); |
1761 | if (!item_eq) |
1762 | DBUG_RETURN(TRUE); |
1763 | for (uint i= 0; i < row->cols(); i++) |
1764 | { |
1765 | if (row->element_index(i) != subq_lex->ref_pointer_array[i]) |
1766 | thd->change_item_tree(row->addr(i), subq_lex->ref_pointer_array[i]); |
1767 | } |
1768 | item_eq->in_equality_no= 0; |
1769 | sj_nest->sj_on_expr= and_items(thd, sj_nest->sj_on_expr, item_eq); |
1770 | } |
1771 | /* |
1772 | Fix the created equality and AND |
1773 | |
1774 | Note that fix_fields() can actually fail in a meaningful way here. One |
1775 | example is when the IN-equality is not valid, because it compares columns |
1776 | with incompatible collations. (One can argue it would be more appropriate |
1777 | to check for this at name resolution stage, but as a legacy of IN->EXISTS |
1778 | we have in here). |
1779 | */ |
1780 | if (!sj_nest->sj_on_expr->fixed && |
1781 | sj_nest->sj_on_expr->fix_fields(thd, &sj_nest->sj_on_expr)) |
1782 | { |
1783 | DBUG_RETURN(TRUE); |
1784 | } |
1785 | |
1786 | /* |
1787 | Walk through sj nest's WHERE and ON expressions and call |
1788 | item->fix_table_changes() for all items. |
1789 | */ |
1790 | sj_nest->sj_on_expr->fix_after_pullout(parent_lex, &sj_nest->sj_on_expr, |
1791 | TRUE); |
1792 | fix_list_after_tbl_changes(parent_lex, &sj_nest->nested_join->join_list); |
1793 | |
1794 | |
1795 | /* Unlink the child select_lex so it doesn't show up in EXPLAIN: */ |
1796 | subq_lex->master_unit()->exclude_level(); |
1797 | |
1798 | DBUG_EXECUTE("where" , |
1799 | print_where(sj_nest->sj_on_expr,"SJ-EXPR" , QT_ORDINARY);); |
1800 | |
1801 | /* Inject sj_on_expr into the parent's WHERE or ON */ |
1802 | if (emb_tbl_nest) |
1803 | { |
1804 | emb_tbl_nest->on_expr= and_items(thd, emb_tbl_nest->on_expr, |
1805 | sj_nest->sj_on_expr); |
1806 | emb_tbl_nest->on_expr->top_level_item(); |
1807 | if (!emb_tbl_nest->on_expr->fixed && |
1808 | emb_tbl_nest->on_expr->fix_fields(thd, |
1809 | &emb_tbl_nest->on_expr)) |
1810 | { |
1811 | DBUG_RETURN(TRUE); |
1812 | } |
1813 | } |
1814 | else |
1815 | { |
1816 | /* Inject into the WHERE */ |
1817 | parent_join->conds= and_items(thd, parent_join->conds, sj_nest->sj_on_expr); |
1818 | parent_join->conds->top_level_item(); |
1819 | /* |
1820 | fix_fields must update the properties (e.g. st_select_lex::cond_count of |
1821 | the correct select_lex. |
1822 | */ |
1823 | save_lex= thd->lex->current_select; |
1824 | thd->lex->current_select=parent_join->select_lex; |
1825 | if (!parent_join->conds->fixed && |
1826 | parent_join->conds->fix_fields(thd, |
1827 | &parent_join->conds)) |
1828 | { |
1829 | DBUG_RETURN(1); |
1830 | } |
1831 | thd->lex->current_select=save_lex; |
1832 | parent_join->select_lex->where= parent_join->conds; |
1833 | } |
1834 | |
1835 | if (subq_lex->ftfunc_list->elements) |
1836 | { |
1837 | Item_func_match *ifm; |
1838 | List_iterator_fast<Item_func_match> li(*(subq_lex->ftfunc_list)); |
1839 | while ((ifm= li++)) |
1840 | parent_lex->ftfunc_list->push_front(ifm, thd->mem_root); |
1841 | } |
1842 | |
1843 | parent_lex->have_merged_subqueries= TRUE; |
1844 | /* Fatal error may have been set to by fix_after_pullout() */ |
1845 | DBUG_RETURN(thd->is_fatal_error); |
1846 | } |
1847 | |
1848 | |
1849 | const int SUBQERY_TEMPTABLE_NAME_MAX_LEN= 20; |
1850 | |
1851 | static void create_subquery_temptable_name(LEX_STRING *str, uint number) |
1852 | { |
1853 | char *to= str->str; |
1854 | DBUG_ASSERT(number < 10000); |
1855 | to= strmov(to, "<subquery" ); |
1856 | to= int10_to_str((int) number, to, 10); |
1857 | to[0]= '>'; |
1858 | to[1]= 0; |
1859 | str->length= (size_t) (to - str->str)+1; |
1860 | } |
1861 | |
1862 | |
1863 | /* |
1864 | Convert subquery predicate into non-mergeable semi-join nest. |
1865 | |
1866 | TODO: |
1867 | why does this do IN-EXISTS conversion? Can't we unify it with mergeable |
1868 | semi-joins? currently, convert_subq_to_sj() cannot fail to convert (unless |
1869 | fatal errors) |
1870 | |
1871 | |
1872 | RETURN |
1873 | FALSE - Ok |
1874 | TRUE - Fatal error |
1875 | */ |
1876 | |
1877 | static bool convert_subq_to_jtbm(JOIN *parent_join, |
1878 | Item_in_subselect *subq_pred, |
1879 | bool *remove_item) |
1880 | { |
1881 | SELECT_LEX *parent_lex= parent_join->select_lex; |
1882 | List<TABLE_LIST> *emb_join_list= &parent_lex->top_join_list; |
1883 | TABLE_LIST *emb_tbl_nest= NULL; // will change when we learn to handle outer joins |
1884 | TABLE_LIST *tl; |
1885 | bool optimization_delayed= TRUE; |
1886 | TABLE_LIST *jtbm; |
1887 | LEX_STRING tbl_alias; |
1888 | THD *thd= parent_join->thd; |
1889 | DBUG_ENTER("convert_subq_to_jtbm" ); |
1890 | |
1891 | subq_pred->set_strategy(SUBS_MATERIALIZATION); |
1892 | subq_pred->is_jtbm_merged= TRUE; |
1893 | |
1894 | *remove_item= TRUE; |
1895 | |
1896 | if (!(tbl_alias.str= (char*)thd->calloc(SUBQERY_TEMPTABLE_NAME_MAX_LEN)) || |
1897 | !(jtbm= alloc_join_nest(thd))) //todo: this is not a join nest! |
1898 | { |
1899 | DBUG_RETURN(TRUE); |
1900 | } |
1901 | |
1902 | jtbm->join_list= emb_join_list; |
1903 | jtbm->embedding= emb_tbl_nest; |
1904 | jtbm->jtbm_subselect= subq_pred; |
1905 | jtbm->nested_join= NULL; |
1906 | |
1907 | /* Nests do not participate in those 'chains', so: */ |
1908 | /* jtbm->next_leaf= jtbm->next_local= jtbm->next_global == NULL*/ |
1909 | emb_join_list->push_back(jtbm, thd->mem_root); |
1910 | |
1911 | /* |
1912 | Inject the jtbm table into TABLE_LIST::next_leaf list, so that |
1913 | make_join_statistics() and co. can find it. |
1914 | */ |
1915 | parent_lex->leaf_tables.push_back(jtbm, thd->mem_root); |
1916 | |
1917 | if (subq_pred->unit->first_select()->options & OPTION_SCHEMA_TABLE) |
1918 | parent_lex->options |= OPTION_SCHEMA_TABLE; |
1919 | |
1920 | /* |
1921 | Same as above for TABLE_LIST::next_local chain |
1922 | (a theory: a next_local chain always starts with ::leaf_tables |
1923 | because view's tables are inserted after the view) |
1924 | */ |
1925 | for (tl= (TABLE_LIST*)(parent_lex->table_list.first); tl->next_local; tl= tl->next_local) |
1926 | {} |
1927 | tl->next_local= jtbm; |
1928 | |
1929 | /* A theory: no need to re-connect the next_global chain */ |
1930 | if (optimization_delayed) |
1931 | { |
1932 | DBUG_ASSERT(parent_join->table_count < MAX_TABLES); |
1933 | |
1934 | jtbm->jtbm_table_no= parent_join->table_count; |
1935 | |
1936 | create_subquery_temptable_name(&tbl_alias, |
1937 | subq_pred->unit->first_select()->select_number); |
1938 | jtbm->alias.str= tbl_alias.str; |
1939 | jtbm->alias.length= tbl_alias.length; |
1940 | parent_join->table_count++; |
1941 | DBUG_RETURN(thd->is_fatal_error); |
1942 | } |
1943 | subselect_hash_sj_engine *hash_sj_engine= |
1944 | ((subselect_hash_sj_engine*)subq_pred->engine); |
1945 | jtbm->table= hash_sj_engine->tmp_table; |
1946 | |
1947 | jtbm->table->tablenr= parent_join->table_count; |
1948 | jtbm->table->map= table_map(1) << (parent_join->table_count); |
1949 | jtbm->jtbm_table_no= jtbm->table->tablenr; |
1950 | |
1951 | parent_join->table_count++; |
1952 | DBUG_ASSERT(parent_join->table_count < MAX_TABLES); |
1953 | |
1954 | Item *conds= hash_sj_engine->semi_join_conds; |
1955 | conds->fix_after_pullout(parent_lex, &conds, TRUE); |
1956 | |
1957 | DBUG_EXECUTE("where" , print_where(conds,"SJ-EXPR" , QT_ORDINARY);); |
1958 | |
1959 | create_subquery_temptable_name(&tbl_alias, hash_sj_engine->materialize_join-> |
1960 | select_lex->select_number); |
1961 | jtbm->alias.str= tbl_alias.str; |
1962 | jtbm->alias.length= tbl_alias.length; |
1963 | |
1964 | parent_lex->have_merged_subqueries= TRUE; |
1965 | |
1966 | /* Don't unlink the child subselect, as the subquery will be used. */ |
1967 | |
1968 | DBUG_RETURN(thd->is_fatal_error); |
1969 | } |
1970 | |
1971 | |
1972 | static TABLE_LIST *alloc_join_nest(THD *thd) |
1973 | { |
1974 | TABLE_LIST *tbl; |
1975 | if (!(tbl= (TABLE_LIST*) thd->calloc(ALIGN_SIZE(sizeof(TABLE_LIST))+ |
1976 | sizeof(NESTED_JOIN)))) |
1977 | return NULL; |
1978 | tbl->nested_join= (NESTED_JOIN*) ((uchar*)tbl + |
1979 | ALIGN_SIZE(sizeof(TABLE_LIST))); |
1980 | return tbl; |
1981 | } |
1982 | |
1983 | /* |
1984 | @Note thd->is_fatal_error can be set in case of OOM |
1985 | */ |
1986 | |
1987 | void fix_list_after_tbl_changes(SELECT_LEX *new_parent, List<TABLE_LIST> *tlist) |
1988 | { |
1989 | List_iterator<TABLE_LIST> it(*tlist); |
1990 | TABLE_LIST *table; |
1991 | while ((table= it++)) |
1992 | { |
1993 | if (table->on_expr) |
1994 | table->on_expr->fix_after_pullout(new_parent, &table->on_expr, TRUE); |
1995 | if (table->nested_join) |
1996 | fix_list_after_tbl_changes(new_parent, &table->nested_join->join_list); |
1997 | } |
1998 | } |
1999 | |
2000 | |
2001 | static void set_emb_join_nest(List<TABLE_LIST> *tables, TABLE_LIST *emb_sj_nest) |
2002 | { |
2003 | List_iterator<TABLE_LIST> it(*tables); |
2004 | TABLE_LIST *tbl; |
2005 | while ((tbl= it++)) |
2006 | { |
2007 | /* |
2008 | Note: check for nested_join first. |
2009 | derived-merged tables have tbl->table!=NULL && |
2010 | tbl->table->reginfo==NULL. |
2011 | */ |
2012 | if (tbl->nested_join) |
2013 | set_emb_join_nest(&tbl->nested_join->join_list, emb_sj_nest); |
2014 | else if (tbl->table) |
2015 | tbl->table->reginfo.join_tab->emb_sj_nest= emb_sj_nest; |
2016 | |
2017 | } |
2018 | } |
2019 | |
2020 | /* |
2021 | Pull tables out of semi-join nests, if possible |
2022 | |
2023 | SYNOPSIS |
2024 | pull_out_semijoin_tables() |
2025 | join The join where to do the semi-join flattening |
2026 | |
2027 | DESCRIPTION |
2028 | Try to pull tables out of semi-join nests. |
2029 | |
2030 | PRECONDITIONS |
2031 | When this function is called, the join may have several semi-join nests |
2032 | but it is guaranteed that one semi-join nest does not contain another. |
2033 | |
2034 | ACTION |
2035 | A table can be pulled out of the semi-join nest if |
2036 | - It is a constant table, or |
2037 | - It is accessed via eq_ref(outer_tables) |
2038 | |
2039 | POSTCONDITIONS |
2040 | * Tables that were pulled out have JOIN_TAB::emb_sj_nest == NULL |
2041 | * Tables that were not pulled out have JOIN_TAB::emb_sj_nest pointing |
2042 | to semi-join nest they are in. |
2043 | * Semi-join nests' TABLE_LIST::sj_inner_tables is updated accordingly |
2044 | |
2045 | This operation is (and should be) performed at each PS execution since |
2046 | tables may become/cease to be constant across PS reexecutions. |
2047 | |
2048 | NOTE |
2049 | Table pullout may make uncorrelated subquery correlated. Consider this |
2050 | example: |
2051 | |
2052 | ... WHERE oe IN (SELECT it1.primary_key WHERE p(it1, it2) ... ) |
2053 | |
2054 | here table it1 can be pulled out (we have it1.primary_key=oe which gives |
2055 | us functional dependency). Once it1 is pulled out, all references to it1 |
2056 | from p(it1, it2) become references to outside of the subquery and thus |
2057 | make the subquery (i.e. its semi-join nest) correlated. |
2058 | Making the subquery (i.e. its semi-join nest) correlated prevents us from |
2059 | using Materialization or LooseScan to execute it. |
2060 | |
2061 | RETURN |
2062 | 0 - OK |
2063 | 1 - Out of memory error |
2064 | */ |
2065 | |
2066 | int pull_out_semijoin_tables(JOIN *join) |
2067 | { |
2068 | TABLE_LIST *sj_nest; |
2069 | DBUG_ENTER("pull_out_semijoin_tables" ); |
2070 | List_iterator<TABLE_LIST> sj_list_it(join->select_lex->sj_nests); |
2071 | |
2072 | /* Try pulling out of the each of the semi-joins */ |
2073 | while ((sj_nest= sj_list_it++)) |
2074 | { |
2075 | List_iterator<TABLE_LIST> child_li(sj_nest->nested_join->join_list); |
2076 | TABLE_LIST *tbl; |
2077 | |
2078 | /* |
2079 | Don't do table pull-out for nested joins (if we get nested joins here, it |
2080 | means these are outer joins. It is theoretically possible to do pull-out |
2081 | for some of the outer tables but we dont support this currently. |
2082 | */ |
2083 | bool have_join_nest_children= FALSE; |
2084 | |
2085 | set_emb_join_nest(&sj_nest->nested_join->join_list, sj_nest); |
2086 | |
2087 | while ((tbl= child_li++)) |
2088 | { |
2089 | if (tbl->nested_join) |
2090 | { |
2091 | have_join_nest_children= TRUE; |
2092 | break; |
2093 | } |
2094 | } |
2095 | |
2096 | table_map pulled_tables= 0; |
2097 | table_map dep_tables= 0; |
2098 | if (have_join_nest_children) |
2099 | goto skip; |
2100 | |
2101 | /* |
2102 | Calculate set of tables within this semi-join nest that have |
2103 | other dependent tables |
2104 | */ |
2105 | child_li.rewind(); |
2106 | while ((tbl= child_li++)) |
2107 | { |
2108 | TABLE *const table= tbl->table; |
2109 | if (table && |
2110 | (table->reginfo.join_tab->dependent & |
2111 | sj_nest->nested_join->used_tables)) |
2112 | dep_tables|= table->reginfo.join_tab->dependent; |
2113 | } |
2114 | |
2115 | /* Action #1: Mark the constant tables to be pulled out */ |
2116 | child_li.rewind(); |
2117 | while ((tbl= child_li++)) |
2118 | { |
2119 | if (tbl->table) |
2120 | { |
2121 | tbl->table->reginfo.join_tab->emb_sj_nest= sj_nest; |
2122 | #if 0 |
2123 | /* |
2124 | Do not pull out tables because they are constant. This operation has |
2125 | a problem: |
2126 | - Some constant tables may become/cease to be constant across PS |
2127 | re-executions |
2128 | - Contrary to our initial assumption, it turned out that table pullout |
2129 | operation is not easily undoable. |
2130 | |
2131 | The solution is to leave constant tables where they are. This will |
2132 | affect only constant tables that are 1-row or empty, tables that are |
2133 | constant because they are accessed via eq_ref(const) access will |
2134 | still be pulled out as functionally-dependent. |
2135 | |
2136 | This will cause us to miss the chance to flatten some of the |
2137 | subqueries, but since const tables do not generate many duplicates, |
2138 | it really doesn't matter that much whether they were pulled out or |
2139 | not. |
2140 | |
2141 | All of this was done as fix for BUG#43768. |
2142 | */ |
2143 | if (tbl->table->map & join->const_table_map) |
2144 | { |
2145 | pulled_tables |= tbl->table->map; |
2146 | DBUG_PRINT("info" , ("Table %s pulled out (reason: constant)" , |
2147 | tbl->table->alias)); |
2148 | } |
2149 | #endif |
2150 | } |
2151 | } |
2152 | |
2153 | /* |
2154 | Action #2: Find which tables we can pull out based on |
2155 | update_ref_and_keys() data. Note that pulling one table out can allow |
2156 | us to pull out some other tables too. |
2157 | */ |
2158 | bool pulled_a_table; |
2159 | do |
2160 | { |
2161 | pulled_a_table= FALSE; |
2162 | child_li.rewind(); |
2163 | while ((tbl= child_li++)) |
2164 | { |
2165 | if (tbl->table && !(pulled_tables & tbl->table->map) && |
2166 | !(dep_tables & tbl->table->map)) |
2167 | { |
2168 | if (find_eq_ref_candidate(tbl->table, |
2169 | sj_nest->nested_join->used_tables & |
2170 | ~pulled_tables)) |
2171 | { |
2172 | pulled_a_table= TRUE; |
2173 | pulled_tables |= tbl->table->map; |
2174 | DBUG_PRINT("info" , ("Table %s pulled out (reason: func dep)" , |
2175 | tbl->table->alias.c_ptr())); |
2176 | /* |
2177 | Pulling a table out of uncorrelated subquery in general makes |
2178 | makes it correlated. See the NOTE to this funtion. |
2179 | */ |
2180 | sj_nest->sj_subq_pred->is_correlated= TRUE; |
2181 | sj_nest->nested_join->sj_corr_tables|= tbl->table->map; |
2182 | sj_nest->nested_join->sj_depends_on|= tbl->table->map; |
2183 | } |
2184 | } |
2185 | } |
2186 | } while (pulled_a_table); |
2187 | |
2188 | child_li.rewind(); |
2189 | skip: |
2190 | /* |
2191 | Action #3: Move the pulled out TABLE_LIST elements to the parents. |
2192 | */ |
2193 | table_map inner_tables= sj_nest->nested_join->used_tables & |
2194 | ~pulled_tables; |
2195 | /* Record the bitmap of inner tables */ |
2196 | sj_nest->sj_inner_tables= inner_tables; |
2197 | if (pulled_tables) |
2198 | { |
2199 | List<TABLE_LIST> *upper_join_list= (sj_nest->embedding != NULL)? |
2200 | (&sj_nest->embedding->nested_join->join_list): |
2201 | (&join->select_lex->top_join_list); |
2202 | Query_arena *arena, backup; |
2203 | arena= join->thd->activate_stmt_arena_if_needed(&backup); |
2204 | while ((tbl= child_li++)) |
2205 | { |
2206 | if (tbl->table) |
2207 | { |
2208 | if (inner_tables & tbl->table->map) |
2209 | { |
2210 | /* This table is not pulled out */ |
2211 | tbl->table->reginfo.join_tab->emb_sj_nest= sj_nest; |
2212 | } |
2213 | else |
2214 | { |
2215 | /* This table has been pulled out of the semi-join nest */ |
2216 | tbl->table->reginfo.join_tab->emb_sj_nest= NULL; |
2217 | /* |
2218 | Pull the table up in the same way as simplify_joins() does: |
2219 | update join_list and embedding pointers but keep next[_local] |
2220 | pointers. |
2221 | */ |
2222 | child_li.remove(); |
2223 | sj_nest->nested_join->used_tables &= ~tbl->table->map; |
2224 | upper_join_list->push_back(tbl, join->thd->mem_root); |
2225 | tbl->join_list= upper_join_list; |
2226 | tbl->embedding= sj_nest->embedding; |
2227 | } |
2228 | } |
2229 | } |
2230 | |
2231 | /* Remove the sj-nest itself if we've removed everything from it */ |
2232 | if (!inner_tables) |
2233 | { |
2234 | List_iterator<TABLE_LIST> li(*upper_join_list); |
2235 | /* Find the sj_nest in the list. */ |
2236 | while (sj_nest != li++) ; |
2237 | li.remove(); |
2238 | /* Also remove it from the list of SJ-nests: */ |
2239 | sj_list_it.remove(); |
2240 | } |
2241 | |
2242 | if (arena) |
2243 | join->thd->restore_active_arena(arena, &backup); |
2244 | } |
2245 | } |
2246 | DBUG_RETURN(0); |
2247 | } |
2248 | |
2249 | |
2250 | /* |
2251 | Optimize semi-join nests that could be run with sj-materialization |
2252 | |
2253 | SYNOPSIS |
2254 | optimize_semijoin_nests() |
2255 | join The join to optimize semi-join nests for |
2256 | all_table_map Bitmap of all tables in the join |
2257 | |
2258 | DESCRIPTION |
2259 | Optimize each of the semi-join nests that can be run with |
2260 | materialization. For each of the nests, we |
2261 | - Generate the best join order for this "sub-join" and remember it; |
2262 | - Remember the sub-join execution cost (it's part of materialization |
2263 | cost); |
2264 | - Calculate other costs that will be incurred if we decide |
2265 | to use materialization strategy for this semi-join nest. |
2266 | |
2267 | All obtained information is saved and will be used by the main join |
2268 | optimization pass. |
2269 | |
2270 | NOTES |
2271 | Because of Join::reoptimize(), this function may be called multiple times. |
2272 | |
2273 | RETURN |
2274 | FALSE Ok |
2275 | TRUE Out of memory error |
2276 | */ |
2277 | |
2278 | bool optimize_semijoin_nests(JOIN *join, table_map all_table_map) |
2279 | { |
2280 | DBUG_ENTER("optimize_semijoin_nests" ); |
2281 | List_iterator<TABLE_LIST> sj_list_it(join->select_lex->sj_nests); |
2282 | TABLE_LIST *sj_nest; |
2283 | while ((sj_nest= sj_list_it++)) |
2284 | { |
2285 | /* semi-join nests with only constant tables are not valid */ |
2286 | /// DBUG_ASSERT(sj_nest->sj_inner_tables & ~join->const_table_map); |
2287 | |
2288 | sj_nest->sj_mat_info= NULL; |
2289 | /* |
2290 | The statement may have been executed with 'semijoin=on' earlier. |
2291 | We need to verify that 'semijoin=on' still holds. |
2292 | */ |
2293 | if (optimizer_flag(join->thd, OPTIMIZER_SWITCH_SEMIJOIN) && |
2294 | optimizer_flag(join->thd, OPTIMIZER_SWITCH_MATERIALIZATION)) |
2295 | { |
2296 | if ((sj_nest->sj_inner_tables & ~join->const_table_map) && /* not everything was pulled out */ |
2297 | !sj_nest->sj_subq_pred->is_correlated && |
2298 | sj_nest->sj_subq_pred->types_allow_materialization) |
2299 | { |
2300 | join->emb_sjm_nest= sj_nest; |
2301 | if (choose_plan(join, all_table_map &~join->const_table_map)) |
2302 | DBUG_RETURN(TRUE); /* purecov: inspected */ |
2303 | /* |
2304 | The best plan to run the subquery is now in join->best_positions, |
2305 | save it. |
2306 | */ |
2307 | uint n_tables= my_count_bits(sj_nest->sj_inner_tables & ~join->const_table_map); |
2308 | SJ_MATERIALIZATION_INFO* sjm; |
2309 | if (!(sjm= new SJ_MATERIALIZATION_INFO) || |
2310 | !(sjm->positions= (POSITION*)join->thd->alloc(sizeof(POSITION)* |
2311 | n_tables))) |
2312 | DBUG_RETURN(TRUE); /* purecov: inspected */ |
2313 | sjm->tables= n_tables; |
2314 | sjm->is_used= FALSE; |
2315 | double subjoin_out_rows, subjoin_read_time; |
2316 | |
2317 | /* |
2318 | join->get_partial_cost_and_fanout(n_tables + join->const_tables, |
2319 | table_map(-1), |
2320 | &subjoin_read_time, |
2321 | &subjoin_out_rows); |
2322 | */ |
2323 | join->get_prefix_cost_and_fanout(n_tables, |
2324 | &subjoin_read_time, |
2325 | &subjoin_out_rows); |
2326 | |
2327 | sjm->materialization_cost.convert_from_cost(subjoin_read_time); |
2328 | sjm->rows= subjoin_out_rows; |
2329 | |
2330 | // Don't use the following list because it has "stale" items. use |
2331 | // ref_pointer_array instead: |
2332 | // |
2333 | //List<Item> &right_expr_list= |
2334 | // sj_nest->sj_subq_pred->unit->first_select()->item_list; |
2335 | /* |
2336 | Adjust output cardinality estimates. If the subquery has form |
2337 | |
2338 | ... oe IN (SELECT t1.colX, t2.colY, func(X,Y,Z) ) |
2339 | |
2340 | then the number of distinct output record combinations has an |
2341 | upper bound of product of number of records matching the tables |
2342 | that are used by the SELECT clause. |
2343 | TODO: |
2344 | We can get a more precise estimate if we |
2345 | - use rec_per_key cardinality estimates. For simple cases like |
2346 | "oe IN (SELECT t.key ...)" it is trivial. |
2347 | - Functional dependencies between the tables in the semi-join |
2348 | nest (the payoff is probably less here?) |
2349 | |
2350 | See also get_post_group_estimate(). |
2351 | */ |
2352 | SELECT_LEX *subq_select= sj_nest->sj_subq_pred->unit->first_select(); |
2353 | { |
2354 | for (uint i=0 ; i < join->const_tables + sjm->tables ; i++) |
2355 | { |
2356 | JOIN_TAB *tab= join->best_positions[i].table; |
2357 | join->map2table[tab->table->tablenr]= tab; |
2358 | } |
2359 | table_map map= 0; |
2360 | for (uint i=0; i < subq_select->item_list.elements; i++) |
2361 | map|= subq_select->ref_pointer_array[i]->used_tables(); |
2362 | map= map & ~PSEUDO_TABLE_BITS; |
2363 | Table_map_iterator tm_it(map); |
2364 | int tableno; |
2365 | double rows= 1.0; |
2366 | while ((tableno = tm_it.next_bit()) != Table_map_iterator::BITMAP_END) |
2367 | rows *= join->map2table[tableno]->table->quick_condition_rows; |
2368 | sjm->rows= MY_MIN(sjm->rows, rows); |
2369 | } |
2370 | memcpy((uchar*) sjm->positions, |
2371 | (uchar*) (join->best_positions + join->const_tables), |
2372 | sizeof(POSITION) * n_tables); |
2373 | |
2374 | /* |
2375 | Calculate temporary table parameters and usage costs |
2376 | */ |
2377 | uint rowlen= get_tmp_table_rec_length(subq_select->ref_pointer_array, |
2378 | subq_select->item_list.elements); |
2379 | double lookup_cost= get_tmp_table_lookup_cost(join->thd, |
2380 | subjoin_out_rows, rowlen); |
2381 | double write_cost= get_tmp_table_write_cost(join->thd, |
2382 | subjoin_out_rows, rowlen); |
2383 | |
2384 | /* |
2385 | Let materialization cost include the cost to write the data into the |
2386 | temporary table: |
2387 | */ |
2388 | sjm->materialization_cost.add_io(subjoin_out_rows, write_cost); |
2389 | |
2390 | /* |
2391 | Set the cost to do a full scan of the temptable (will need this to |
2392 | consider doing sjm-scan): |
2393 | */ |
2394 | sjm->scan_cost.reset(); |
2395 | sjm->scan_cost.add_io(sjm->rows, lookup_cost); |
2396 | |
2397 | sjm->lookup_cost.convert_from_cost(lookup_cost); |
2398 | sj_nest->sj_mat_info= sjm; |
2399 | DBUG_EXECUTE("opt" , print_sjm(sjm);); |
2400 | } |
2401 | } |
2402 | } |
2403 | join->emb_sjm_nest= NULL; |
2404 | DBUG_RETURN(FALSE); |
2405 | } |
2406 | |
2407 | |
2408 | /* |
2409 | Get estimated record length for semi-join materialization temptable |
2410 | |
2411 | SYNOPSIS |
2412 | get_tmp_table_rec_length() |
2413 | items IN subquery's select list. |
2414 | |
2415 | DESCRIPTION |
2416 | Calculate estimated record length for semi-join materialization |
2417 | temptable. It's an estimate because we don't follow every bit of |
2418 | create_tmp_table()'s logic. This isn't necessary as the return value of |
2419 | this function is used only for cost calculations. |
2420 | |
2421 | RETURN |
2422 | Length of the temptable record, in bytes |
2423 | */ |
2424 | |
2425 | static uint get_tmp_table_rec_length(Ref_ptr_array p_items, uint elements) |
2426 | { |
2427 | uint len= 0; |
2428 | Item *item; |
2429 | //List_iterator<Item> it(items); |
2430 | for (uint i= 0; i < elements ; i++) |
2431 | { |
2432 | item = p_items[i]; |
2433 | switch (item->result_type()) { |
2434 | case REAL_RESULT: |
2435 | len += sizeof(double); |
2436 | break; |
2437 | case INT_RESULT: |
2438 | if (item->max_length >= (MY_INT32_NUM_DECIMAL_DIGITS - 1)) |
2439 | len += 8; |
2440 | else |
2441 | len += 4; |
2442 | break; |
2443 | case STRING_RESULT: |
2444 | enum enum_field_types type; |
2445 | /* DATE/TIME and GEOMETRY fields have STRING_RESULT result type. */ |
2446 | if ((type= item->field_type()) == MYSQL_TYPE_DATETIME || |
2447 | type == MYSQL_TYPE_TIME || type == MYSQL_TYPE_DATE || |
2448 | type == MYSQL_TYPE_TIMESTAMP || type == MYSQL_TYPE_GEOMETRY) |
2449 | len += 8; |
2450 | else |
2451 | len += item->max_length; |
2452 | break; |
2453 | case DECIMAL_RESULT: |
2454 | len += 10; |
2455 | break; |
2456 | case ROW_RESULT: |
2457 | default: |
2458 | DBUG_ASSERT(0); /* purecov: deadcode */ |
2459 | break; |
2460 | } |
2461 | } |
2462 | return len; |
2463 | } |
2464 | |
2465 | |
2466 | /** |
2467 | The cost of a lookup into a unique hash/btree index on a temporary table |
2468 | with 'row_count' rows each of size 'row_size'. |
2469 | |
2470 | @param thd current query context |
2471 | @param row_count number of rows in the temp table |
2472 | @param row_size average size in bytes of the rows |
2473 | |
2474 | @return the cost of one lookup |
2475 | */ |
2476 | |
2477 | double |
2478 | get_tmp_table_lookup_cost(THD *thd, double row_count, uint row_size) |
2479 | { |
2480 | if (row_count * row_size > thd->variables.max_heap_table_size) |
2481 | return (double) DISK_TEMPTABLE_LOOKUP_COST; |
2482 | else |
2483 | return (double) HEAP_TEMPTABLE_LOOKUP_COST; |
2484 | } |
2485 | |
2486 | /** |
2487 | The cost of writing a row into a temporary table with 'row_count' unique |
2488 | rows each of size 'row_size'. |
2489 | |
2490 | @param thd current query context |
2491 | @param row_count number of rows in the temp table |
2492 | @param row_size average size in bytes of the rows |
2493 | |
2494 | @return the cost of writing one row |
2495 | */ |
2496 | |
2497 | double |
2498 | get_tmp_table_write_cost(THD *thd, double row_count, uint row_size) |
2499 | { |
2500 | double lookup_cost= get_tmp_table_lookup_cost(thd, row_count, row_size); |
2501 | /* |
2502 | TODO: |
2503 | This is an optimistic estimate. Add additional costs resulting from |
2504 | actually writing the row to memory/disk and possible index reorganization. |
2505 | */ |
2506 | return lookup_cost; |
2507 | } |
2508 | |
2509 | |
2510 | /* |
2511 | Check if table's KEYUSE elements have an eq_ref(outer_tables) candidate |
2512 | |
2513 | SYNOPSIS |
2514 | find_eq_ref_candidate() |
2515 | table Table to be checked |
2516 | sj_inner_tables Bitmap of inner tables. eq_ref(inner_table) doesn't |
2517 | count. |
2518 | |
2519 | DESCRIPTION |
2520 | Check if table's KEYUSE elements have an eq_ref(outer_tables) candidate |
2521 | |
2522 | TODO |
2523 | Check again if it is feasible to factor common parts with constant table |
2524 | search |
2525 | |
2526 | Also check if it's feasible to factor common parts with table elimination |
2527 | |
2528 | RETURN |
2529 | TRUE - There exists an eq_ref(outer-tables) candidate |
2530 | FALSE - Otherwise |
2531 | */ |
2532 | |
2533 | bool find_eq_ref_candidate(TABLE *table, table_map sj_inner_tables) |
2534 | { |
2535 | KEYUSE *keyuse= table->reginfo.join_tab->keyuse; |
2536 | |
2537 | if (keyuse) |
2538 | { |
2539 | do |
2540 | { |
2541 | uint key= keyuse->key; |
2542 | KEY *keyinfo; |
2543 | key_part_map bound_parts= 0; |
2544 | bool is_excluded_key= keyuse->is_for_hash_join(); |
2545 | if (!is_excluded_key) |
2546 | { |
2547 | keyinfo= table->key_info + key; |
2548 | is_excluded_key= !MY_TEST(keyinfo->flags & HA_NOSAME); |
2549 | } |
2550 | if (!is_excluded_key) |
2551 | { |
2552 | do /* For all equalities on all key parts */ |
2553 | { |
2554 | /* Check if this is "t.keypart = expr(outer_tables) */ |
2555 | if (!(keyuse->used_tables & sj_inner_tables) && |
2556 | !(keyuse->optimize & KEY_OPTIMIZE_REF_OR_NULL)) |
2557 | { |
2558 | bound_parts |= 1 << keyuse->keypart; |
2559 | } |
2560 | keyuse++; |
2561 | } while (keyuse->key == key && keyuse->table == table); |
2562 | |
2563 | if (bound_parts == PREV_BITS(uint, keyinfo->user_defined_key_parts)) |
2564 | return TRUE; |
2565 | } |
2566 | else |
2567 | { |
2568 | do |
2569 | { |
2570 | keyuse++; |
2571 | } while (keyuse->key == key && keyuse->table == table); |
2572 | } |
2573 | } while (keyuse->table == table); |
2574 | } |
2575 | return FALSE; |
2576 | } |
2577 | |
2578 | |
2579 | /* |
2580 | Do semi-join optimization step after we've added a new tab to join prefix |
2581 | |
2582 | SYNOPSIS |
2583 | advance_sj_state() |
2584 | join The join we're optimizing |
2585 | remaining_tables Tables not in the join prefix |
2586 | new_join_tab Join tab we've just added to the join prefix |
2587 | idx Index of this join tab (i.e. number of tables |
2588 | in the prefix minus one) |
2589 | current_record_count INOUT Estimate of #records in join prefix's output |
2590 | current_read_time INOUT Cost to execute the join prefix |
2591 | loose_scan_pos IN A POSITION with LooseScan plan to access |
2592 | table new_join_tab |
2593 | (produced by the last best_access_path call) |
2594 | |
2595 | DESCRIPTION |
2596 | Update semi-join optimization state after we've added another tab (table |
2597 | and access method) to the join prefix. |
2598 | |
2599 | The state is maintained in join->positions[#prefix_size]. Each of the |
2600 | available strategies has its own state variables. |
2601 | |
2602 | for each semi-join strategy |
2603 | { |
2604 | update strategy's state variables; |
2605 | |
2606 | if (join prefix has all the tables that are needed to consider |
2607 | using this strategy for the semi-join(s)) |
2608 | { |
2609 | calculate cost of using the strategy |
2610 | if ((this is the first strategy to handle the semi-join nest(s) || |
2611 | the cost is less than other strategies)) |
2612 | { |
2613 | // Pick this strategy |
2614 | pos->sj_strategy= .. |
2615 | .. |
2616 | } |
2617 | } |
2618 | |
2619 | Most of the new state is saved join->positions[idx] (and hence no undo |
2620 | is necessary). Several members of class JOIN are updated also, these |
2621 | changes can be rolled back with restore_prev_sj_state(). |
2622 | |
2623 | See setup_semijoin_dups_elimination() for a description of what kinds of |
2624 | join prefixes each strategy can handle. |
2625 | */ |
2626 | |
2627 | bool is_multiple_semi_joins(JOIN *join, POSITION *prefix, uint idx, table_map inner_tables) |
2628 | { |
2629 | for (int i= (int)idx; i >= 0; i--) |
2630 | { |
2631 | TABLE_LIST *emb_sj_nest; |
2632 | if ((emb_sj_nest= prefix[i].table->emb_sj_nest)) |
2633 | { |
2634 | if (inner_tables & emb_sj_nest->sj_inner_tables) |
2635 | return !MY_TEST(inner_tables == (emb_sj_nest->sj_inner_tables & |
2636 | ~join->const_table_map)); |
2637 | } |
2638 | } |
2639 | return FALSE; |
2640 | } |
2641 | |
2642 | |
2643 | void advance_sj_state(JOIN *join, table_map remaining_tables, uint idx, |
2644 | double *current_record_count, double *current_read_time, |
2645 | POSITION *loose_scan_pos) |
2646 | { |
2647 | POSITION *pos= join->positions + idx; |
2648 | const JOIN_TAB *new_join_tab= pos->table; |
2649 | Semi_join_strategy_picker *pickers[]= |
2650 | { |
2651 | &pos->firstmatch_picker, |
2652 | &pos->loosescan_picker, |
2653 | &pos->sjmat_picker, |
2654 | &pos->dups_weedout_picker, |
2655 | NULL, |
2656 | }; |
2657 | |
2658 | if (join->emb_sjm_nest) |
2659 | { |
2660 | /* |
2661 | We're performing optimization inside SJ-Materialization nest: |
2662 | - there are no other semi-joins inside semi-join nests |
2663 | - attempts to build semi-join strategies here will confuse |
2664 | the optimizer, so bail out. |
2665 | */ |
2666 | pos->sj_strategy= SJ_OPT_NONE; |
2667 | return; |
2668 | } |
2669 | |
2670 | /* |
2671 | Update join->cur_sj_inner_tables (Used by FirstMatch in this function and |
2672 | LooseScan detector in best_access_path) |
2673 | */ |
2674 | remaining_tables &= ~new_join_tab->table->map; |
2675 | table_map dups_producing_tables, UNINIT_VAR(prev_dups_producing_tables), |
2676 | UNINIT_VAR(prev_sjm_lookup_tables); |
2677 | |
2678 | if (idx == join->const_tables) |
2679 | dups_producing_tables= 0; |
2680 | else |
2681 | dups_producing_tables= pos[-1].dups_producing_tables; |
2682 | |
2683 | TABLE_LIST *emb_sj_nest; |
2684 | if ((emb_sj_nest= new_join_tab->emb_sj_nest)) |
2685 | dups_producing_tables |= emb_sj_nest->sj_inner_tables; |
2686 | |
2687 | Semi_join_strategy_picker **strategy, **prev_strategy= 0; |
2688 | if (idx == join->const_tables) |
2689 | { |
2690 | /* First table, initialize pickers */ |
2691 | for (strategy= pickers; *strategy != NULL; strategy++) |
2692 | (*strategy)->set_empty(); |
2693 | pos->inner_tables_handled_with_other_sjs= 0; |
2694 | } |
2695 | else |
2696 | { |
2697 | for (strategy= pickers; *strategy != NULL; strategy++) |
2698 | { |
2699 | (*strategy)->set_from_prev(pos - 1); |
2700 | } |
2701 | pos->inner_tables_handled_with_other_sjs= |
2702 | pos[-1].inner_tables_handled_with_other_sjs; |
2703 | } |
2704 | |
2705 | pos->prefix_cost.convert_from_cost(*current_read_time); |
2706 | pos->prefix_record_count= *current_record_count; |
2707 | |
2708 | { |
2709 | pos->sj_strategy= SJ_OPT_NONE; |
2710 | |
2711 | for (strategy= pickers; *strategy != NULL; strategy++) |
2712 | { |
2713 | table_map handled_fanout; |
2714 | sj_strategy_enum sj_strategy; |
2715 | double rec_count= *current_record_count; |
2716 | double read_time= *current_read_time; |
2717 | if ((*strategy)->check_qep(join, idx, remaining_tables, |
2718 | new_join_tab, |
2719 | &rec_count, |
2720 | &read_time, |
2721 | &handled_fanout, |
2722 | &sj_strategy, |
2723 | loose_scan_pos)) |
2724 | { |
2725 | /* |
2726 | It's possible to use the strategy. Use it, if |
2727 | - it removes semi-join fanout that was not removed before |
2728 | - using it is cheaper than using something else, |
2729 | and {if some other strategy has removed fanout |
2730 | that this strategy is trying to remove, then it |
2731 | did remove the fanout only for one semi-join} |
2732 | This is to avoid a situation when |
2733 | 1. strategy X removes fanout for semijoin X,Y |
2734 | 2. using strategy Z is cheaper, but it only removes |
2735 | fanout from semijoin X. |
2736 | 3. We have no clue what to do about fanount of semi-join Y. |
2737 | */ |
2738 | if ((dups_producing_tables & handled_fanout) || |
2739 | (read_time < *current_read_time && |
2740 | !(handled_fanout & pos->inner_tables_handled_with_other_sjs))) |
2741 | { |
2742 | DBUG_ASSERT(pos->sj_strategy != sj_strategy); |
2743 | /* |
2744 | If the strategy choosen first time or |
2745 | the strategy replace strategy which was used to exectly the same |
2746 | tables |
2747 | */ |
2748 | if (pos->sj_strategy == SJ_OPT_NONE || |
2749 | handled_fanout == |
2750 | (prev_dups_producing_tables ^ dups_producing_tables)) |
2751 | { |
2752 | prev_strategy= strategy; |
2753 | if (pos->sj_strategy == SJ_OPT_NONE) |
2754 | { |
2755 | prev_dups_producing_tables= dups_producing_tables; |
2756 | prev_sjm_lookup_tables= join->sjm_lookup_tables; |
2757 | } |
2758 | /* Mark strategy as used */ |
2759 | (*strategy)->mark_used(); |
2760 | pos->sj_strategy= sj_strategy; |
2761 | if (sj_strategy == SJ_OPT_MATERIALIZE) |
2762 | join->sjm_lookup_tables |= handled_fanout; |
2763 | else |
2764 | join->sjm_lookup_tables &= ~handled_fanout; |
2765 | *current_read_time= read_time; |
2766 | *current_record_count= rec_count; |
2767 | dups_producing_tables &= ~handled_fanout; |
2768 | //TODO: update bitmap of semi-joins that were handled together with |
2769 | // others. |
2770 | if (is_multiple_semi_joins(join, join->positions, idx, |
2771 | handled_fanout)) |
2772 | pos->inner_tables_handled_with_other_sjs |= handled_fanout; |
2773 | } |
2774 | else |
2775 | { |
2776 | /* Conflict fall to most general variant */ |
2777 | (*prev_strategy)->set_empty(); |
2778 | dups_producing_tables= prev_dups_producing_tables; |
2779 | join->sjm_lookup_tables= prev_sjm_lookup_tables; |
2780 | // mark it 'none' to avpoid loops |
2781 | pos->sj_strategy= SJ_OPT_NONE; |
2782 | // next skip to last; |
2783 | strategy= pickers + |
2784 | (sizeof(pickers)/sizeof(Semi_join_strategy_picker*) - 3); |
2785 | continue; |
2786 | } |
2787 | } |
2788 | else |
2789 | { |
2790 | /* We decided not to apply the strategy. */ |
2791 | (*strategy)->set_empty(); |
2792 | } |
2793 | } |
2794 | } |
2795 | } |
2796 | |
2797 | if ((emb_sj_nest= new_join_tab->emb_sj_nest)) |
2798 | { |
2799 | join->cur_sj_inner_tables |= emb_sj_nest->sj_inner_tables; |
2800 | |
2801 | /* Remove the sj_nest if all of its SJ-inner tables are in cur_table_map */ |
2802 | if (!(remaining_tables & |
2803 | emb_sj_nest->sj_inner_tables & ~new_join_tab->table->map)) |
2804 | join->cur_sj_inner_tables &= ~emb_sj_nest->sj_inner_tables; |
2805 | } |
2806 | |
2807 | pos->prefix_cost.convert_from_cost(*current_read_time); |
2808 | pos->prefix_record_count= *current_record_count; |
2809 | pos->dups_producing_tables= dups_producing_tables; |
2810 | } |
2811 | |
2812 | |
2813 | void Sj_materialization_picker::set_from_prev(struct st_position *prev) |
2814 | { |
2815 | if (prev->sjmat_picker.is_used) |
2816 | set_empty(); |
2817 | else |
2818 | { |
2819 | sjm_scan_need_tables= prev->sjmat_picker.sjm_scan_need_tables; |
2820 | sjm_scan_last_inner= prev->sjmat_picker.sjm_scan_last_inner; |
2821 | } |
2822 | is_used= FALSE; |
2823 | } |
2824 | |
2825 | |
2826 | bool Sj_materialization_picker::check_qep(JOIN *join, |
2827 | uint idx, |
2828 | table_map remaining_tables, |
2829 | const JOIN_TAB *new_join_tab, |
2830 | double *record_count, |
2831 | double *read_time, |
2832 | table_map *handled_fanout, |
2833 | sj_strategy_enum *strategy, |
2834 | POSITION *loose_scan_pos) |
2835 | { |
2836 | bool sjm_scan; |
2837 | SJ_MATERIALIZATION_INFO *mat_info; |
2838 | if ((mat_info= at_sjmat_pos(join, remaining_tables, |
2839 | new_join_tab, idx, &sjm_scan))) |
2840 | { |
2841 | if (sjm_scan) |
2842 | { |
2843 | /* |
2844 | We can't yet evaluate this option yet. This is because we can't |
2845 | accout for fanout of sj-inner tables yet: |
2846 | |
2847 | ntX SJM-SCAN(it1 ... itN) | ot1 ... otN | |
2848 | ^(1) ^(2) |
2849 | |
2850 | we're now at position (1). SJM temptable in general has multiple |
2851 | records, so at point (1) we'll get the fanout from sj-inner tables (ie |
2852 | there will be multiple record combinations). |
2853 | |
2854 | The final join result will not contain any semi-join produced |
2855 | fanout, i.e. tables within SJM-SCAN(...) will not contribute to |
2856 | the cardinality of the join output. Extra fanout produced by |
2857 | SJM-SCAN(...) will be 'absorbed' into fanout produced by ot1 ... otN. |
2858 | |
2859 | The simple way to model this is to remove SJM-SCAN(...) fanout once |
2860 | we reach the point #2. |
2861 | */ |
2862 | sjm_scan_need_tables= |
2863 | new_join_tab->emb_sj_nest->sj_inner_tables | |
2864 | new_join_tab->emb_sj_nest->nested_join->sj_depends_on | |
2865 | new_join_tab->emb_sj_nest->nested_join->sj_corr_tables; |
2866 | sjm_scan_last_inner= idx; |
2867 | } |
2868 | else |
2869 | { |
2870 | /* This is SJ-Materialization with lookups */ |
2871 | Cost_estimate prefix_cost; |
2872 | signed int first_tab= (int)idx - mat_info->tables; |
2873 | double prefix_rec_count; |
2874 | if (first_tab < (int)join->const_tables) |
2875 | { |
2876 | prefix_cost.reset(); |
2877 | prefix_rec_count= 1.0; |
2878 | } |
2879 | else |
2880 | { |
2881 | prefix_cost= join->positions[first_tab].prefix_cost; |
2882 | prefix_rec_count= join->positions[first_tab].prefix_record_count; |
2883 | } |
2884 | |
2885 | double mat_read_time= prefix_cost.total_cost(); |
2886 | mat_read_time += mat_info->materialization_cost.total_cost() + |
2887 | prefix_rec_count * mat_info->lookup_cost.total_cost(); |
2888 | |
2889 | /* |
2890 | NOTE: When we pick to use SJM[-Scan] we don't memcpy its POSITION |
2891 | elements to join->positions as that makes it hard to return things |
2892 | back when making one step back in join optimization. That's done |
2893 | after the QEP has been chosen. |
2894 | */ |
2895 | *read_time= mat_read_time; |
2896 | *record_count= prefix_rec_count; |
2897 | *handled_fanout= new_join_tab->emb_sj_nest->sj_inner_tables; |
2898 | *strategy= SJ_OPT_MATERIALIZE; |
2899 | return TRUE; |
2900 | } |
2901 | } |
2902 | |
2903 | /* 4.A SJM-Scan second phase check */ |
2904 | if (sjm_scan_need_tables && /* Have SJM-Scan prefix */ |
2905 | !(sjm_scan_need_tables & remaining_tables)) |
2906 | { |
2907 | TABLE_LIST *mat_nest= |
2908 | join->positions[sjm_scan_last_inner].table->emb_sj_nest; |
2909 | SJ_MATERIALIZATION_INFO *mat_info= mat_nest->sj_mat_info; |
2910 | |
2911 | double prefix_cost; |
2912 | double prefix_rec_count; |
2913 | int first_tab= sjm_scan_last_inner + 1 - mat_info->tables; |
2914 | /* Get the prefix cost */ |
2915 | if (first_tab == (int)join->const_tables) |
2916 | { |
2917 | prefix_rec_count= 1.0; |
2918 | prefix_cost= 0.0; |
2919 | } |
2920 | else |
2921 | { |
2922 | prefix_cost= join->positions[first_tab - 1].prefix_cost.total_cost(); |
2923 | prefix_rec_count= join->positions[first_tab - 1].prefix_record_count; |
2924 | } |
2925 | |
2926 | /* Add materialization cost */ |
2927 | prefix_cost += mat_info->materialization_cost.total_cost() + |
2928 | prefix_rec_count * mat_info->scan_cost.total_cost(); |
2929 | prefix_rec_count *= mat_info->rows; |
2930 | |
2931 | uint i; |
2932 | table_map rem_tables= remaining_tables; |
2933 | for (i= idx; i != (first_tab + mat_info->tables - 1); i--) |
2934 | rem_tables |= join->positions[i].table->table->map; |
2935 | |
2936 | POSITION curpos, dummy; |
2937 | /* Need to re-run best-access-path as we prefix_rec_count has changed */ |
2938 | bool disable_jbuf= (join->thd->variables.join_cache_level == 0); |
2939 | for (i= first_tab + mat_info->tables; i <= idx; i++) |
2940 | { |
2941 | best_access_path(join, join->positions[i].table, rem_tables, i, |
2942 | disable_jbuf, prefix_rec_count, &curpos, &dummy); |
2943 | prefix_rec_count *= curpos.records_read; |
2944 | prefix_cost += curpos.read_time; |
2945 | } |
2946 | |
2947 | *strategy= SJ_OPT_MATERIALIZE_SCAN; |
2948 | *read_time= prefix_cost; |
2949 | *record_count= prefix_rec_count; |
2950 | *handled_fanout= mat_nest->sj_inner_tables; |
2951 | return TRUE; |
2952 | } |
2953 | return FALSE; |
2954 | } |
2955 | |
2956 | |
2957 | void LooseScan_picker::set_from_prev(struct st_position *prev) |
2958 | { |
2959 | if (prev->loosescan_picker.is_used) |
2960 | set_empty(); |
2961 | else |
2962 | { |
2963 | first_loosescan_table= prev->loosescan_picker.first_loosescan_table; |
2964 | loosescan_need_tables= prev->loosescan_picker.loosescan_need_tables; |
2965 | } |
2966 | is_used= FALSE; |
2967 | } |
2968 | |
2969 | |
2970 | bool LooseScan_picker::check_qep(JOIN *join, |
2971 | uint idx, |
2972 | table_map remaining_tables, |
2973 | const JOIN_TAB *new_join_tab, |
2974 | double *record_count, |
2975 | double *read_time, |
2976 | table_map *handled_fanout, |
2977 | sj_strategy_enum *strategy, |
2978 | struct st_position *loose_scan_pos) |
2979 | { |
2980 | POSITION *first= join->positions + first_loosescan_table; |
2981 | /* |
2982 | LooseScan strategy can't handle interleaving between tables from the |
2983 | semi-join that LooseScan is handling and any other tables. |
2984 | |
2985 | If we were considering LooseScan for the join prefix (1) |
2986 | and the table we're adding creates an interleaving (2) |
2987 | then |
2988 | stop considering loose scan |
2989 | */ |
2990 | if ((first_loosescan_table != MAX_TABLES) && // (1) |
2991 | (first->table->emb_sj_nest->sj_inner_tables & remaining_tables) && //(2) |
2992 | new_join_tab->emb_sj_nest != first->table->emb_sj_nest) //(2) |
2993 | { |
2994 | first_loosescan_table= MAX_TABLES; |
2995 | } |
2996 | |
2997 | /* |
2998 | If we got an option to use LooseScan for the current table, start |
2999 | considering using LooseScan strategy |
3000 | */ |
3001 | if (loose_scan_pos->read_time != DBL_MAX && !join->outer_join) |
3002 | { |
3003 | first_loosescan_table= idx; |
3004 | loosescan_need_tables= |
3005 | new_join_tab->emb_sj_nest->sj_inner_tables | |
3006 | new_join_tab->emb_sj_nest->nested_join->sj_depends_on | |
3007 | new_join_tab->emb_sj_nest->nested_join->sj_corr_tables; |
3008 | } |
3009 | |
3010 | if ((first_loosescan_table != MAX_TABLES) && |
3011 | !(remaining_tables & loosescan_need_tables) && |
3012 | (new_join_tab->table->map & loosescan_need_tables)) |
3013 | { |
3014 | /* |
3015 | Ok we have LooseScan plan and also have all LooseScan sj-nest's |
3016 | inner tables and outer correlated tables into the prefix. |
3017 | */ |
3018 | |
3019 | first= join->positions + first_loosescan_table; |
3020 | uint n_tables= my_count_bits(first->table->emb_sj_nest->sj_inner_tables); |
3021 | /* Got a complete LooseScan range. Calculate its cost */ |
3022 | /* |
3023 | The same problem as with FirstMatch - we need to save POSITIONs |
3024 | somewhere but reserving space for all cases would require too |
3025 | much space. We will re-calculate POSITION structures later on. |
3026 | */ |
3027 | bool disable_jbuf= (join->thd->variables.join_cache_level == 0); |
3028 | optimize_wo_join_buffering(join, first_loosescan_table, idx, |
3029 | remaining_tables, |
3030 | TRUE, //first_alt |
3031 | disable_jbuf ? join->table_count : |
3032 | first_loosescan_table + n_tables, |
3033 | record_count, |
3034 | read_time); |
3035 | /* |
3036 | We don't yet have any other strategies that could handle this |
3037 | semi-join nest (the other options are Duplicate Elimination or |
3038 | Materialization, which need at least the same set of tables in |
3039 | the join prefix to be considered) so unconditionally pick the |
3040 | LooseScan. |
3041 | */ |
3042 | *strategy= SJ_OPT_LOOSE_SCAN; |
3043 | *handled_fanout= first->table->emb_sj_nest->sj_inner_tables; |
3044 | return TRUE; |
3045 | } |
3046 | return FALSE; |
3047 | } |
3048 | |
3049 | void Firstmatch_picker::set_from_prev(struct st_position *prev) |
3050 | { |
3051 | if (prev->firstmatch_picker.is_used) |
3052 | invalidate_firstmatch_prefix(); |
3053 | else |
3054 | { |
3055 | first_firstmatch_table= prev->firstmatch_picker.first_firstmatch_table; |
3056 | first_firstmatch_rtbl= prev->firstmatch_picker.first_firstmatch_rtbl; |
3057 | firstmatch_need_tables= prev->firstmatch_picker.firstmatch_need_tables; |
3058 | } |
3059 | is_used= FALSE; |
3060 | } |
3061 | |
3062 | bool Firstmatch_picker::check_qep(JOIN *join, |
3063 | uint idx, |
3064 | table_map remaining_tables, |
3065 | const JOIN_TAB *new_join_tab, |
3066 | double *record_count, |
3067 | double *read_time, |
3068 | table_map *handled_fanout, |
3069 | sj_strategy_enum *strategy, |
3070 | POSITION *loose_scan_pos) |
3071 | { |
3072 | if (new_join_tab->emb_sj_nest && |
3073 | optimizer_flag(join->thd, OPTIMIZER_SWITCH_FIRSTMATCH) && |
3074 | !join->outer_join) |
3075 | { |
3076 | const table_map outer_corr_tables= |
3077 | new_join_tab->emb_sj_nest->nested_join->sj_corr_tables | |
3078 | new_join_tab->emb_sj_nest->nested_join->sj_depends_on; |
3079 | const table_map sj_inner_tables= |
3080 | new_join_tab->emb_sj_nest->sj_inner_tables & ~join->const_table_map; |
3081 | |
3082 | /* |
3083 | Enter condition: |
3084 | 1. The next join tab belongs to semi-join nest |
3085 | (verified for the encompassing code block above). |
3086 | 2. We're not in a duplicate producer range yet |
3087 | 3. All outer tables that |
3088 | - the subquery is correlated with, or |
3089 | - referred to from the outer_expr |
3090 | are in the join prefix |
3091 | 4. All inner tables are still part of remaining_tables. |
3092 | */ |
3093 | if (!join->cur_sj_inner_tables && // (2) |
3094 | !(remaining_tables & outer_corr_tables) && // (3) |
3095 | (sj_inner_tables == // (4) |
3096 | ((remaining_tables | new_join_tab->table->map) & sj_inner_tables))) |
3097 | { |
3098 | /* Start tracking potential FirstMatch range */ |
3099 | first_firstmatch_table= idx; |
3100 | firstmatch_need_tables= sj_inner_tables; |
3101 | first_firstmatch_rtbl= remaining_tables; |
3102 | } |
3103 | |
3104 | if (in_firstmatch_prefix()) |
3105 | { |
3106 | if (outer_corr_tables & first_firstmatch_rtbl) |
3107 | { |
3108 | /* |
3109 | Trying to add an sj-inner table whose sj-nest has an outer correlated |
3110 | table that was not in the prefix. This means FirstMatch can't be used. |
3111 | */ |
3112 | invalidate_firstmatch_prefix(); |
3113 | } |
3114 | else |
3115 | { |
3116 | /* Record that we need all of this semi-join's inner tables, too */ |
3117 | firstmatch_need_tables|= sj_inner_tables; |
3118 | } |
3119 | |
3120 | if (in_firstmatch_prefix() && |
3121 | !(firstmatch_need_tables & remaining_tables)) |
3122 | { |
3123 | /* |
3124 | Got a complete FirstMatch range. Calculate correct costs and fanout |
3125 | */ |
3126 | |
3127 | if (idx == first_firstmatch_table && |
3128 | optimizer_flag(join->thd, OPTIMIZER_SWITCH_SEMIJOIN_WITH_CACHE)) |
3129 | { |
3130 | /* |
3131 | An important special case: only one inner table, and @@optimizer_switch |
3132 | allows join buffering. |
3133 | - read_time is the same (i.e. FirstMatch doesn't add any cost |
3134 | - remove fanout added by the last table |
3135 | */ |
3136 | if (*record_count) |
3137 | *record_count /= join->positions[idx].records_read; |
3138 | } |
3139 | else |
3140 | { |
3141 | optimize_wo_join_buffering(join, first_firstmatch_table, idx, |
3142 | remaining_tables, FALSE, idx, |
3143 | record_count, |
3144 | read_time); |
3145 | } |
3146 | /* |
3147 | We ought to save the alternate POSITIONs produced by |
3148 | optimize_wo_join_buffering but the problem is that providing save |
3149 | space uses too much space. Instead, we will re-calculate the |
3150 | alternate POSITIONs after we've picked the best QEP. |
3151 | */ |
3152 | *handled_fanout= firstmatch_need_tables; |
3153 | /* *record_count and *read_time were set by the above call */ |
3154 | *strategy= SJ_OPT_FIRST_MATCH; |
3155 | return TRUE; |
3156 | } |
3157 | } |
3158 | } |
3159 | else |
3160 | invalidate_firstmatch_prefix(); |
3161 | return FALSE; |
3162 | } |
3163 | |
3164 | |
3165 | void Duplicate_weedout_picker::set_from_prev(POSITION *prev) |
3166 | { |
3167 | if (prev->dups_weedout_picker.is_used) |
3168 | set_empty(); |
3169 | else |
3170 | { |
3171 | dupsweedout_tables= prev->dups_weedout_picker.dupsweedout_tables; |
3172 | first_dupsweedout_table= prev->dups_weedout_picker.first_dupsweedout_table; |
3173 | } |
3174 | is_used= FALSE; |
3175 | } |
3176 | |
3177 | |
3178 | bool Duplicate_weedout_picker::check_qep(JOIN *join, |
3179 | uint idx, |
3180 | table_map remaining_tables, |
3181 | const JOIN_TAB *new_join_tab, |
3182 | double *record_count, |
3183 | double *read_time, |
3184 | table_map *handled_fanout, |
3185 | sj_strategy_enum *strategy, |
3186 | POSITION *loose_scan_pos |
3187 | ) |
3188 | { |
3189 | TABLE_LIST *nest; |
3190 | if ((nest= new_join_tab->emb_sj_nest)) |
3191 | { |
3192 | if (!dupsweedout_tables) |
3193 | first_dupsweedout_table= idx; |
3194 | |
3195 | dupsweedout_tables |= nest->sj_inner_tables | |
3196 | nest->nested_join->sj_depends_on | |
3197 | nest->nested_join->sj_corr_tables; |
3198 | } |
3199 | |
3200 | if (dupsweedout_tables) |
3201 | { |
3202 | /* we're in the process of constructing a DuplicateWeedout range */ |
3203 | TABLE_LIST *emb= new_join_tab->table->pos_in_table_list->embedding; |
3204 | /* and we've entered an inner side of an outer join*/ |
3205 | if (emb && emb->on_expr) |
3206 | dupsweedout_tables |= emb->nested_join->used_tables; |
3207 | } |
3208 | |
3209 | /* If this is the last table that we need for DuplicateWeedout range */ |
3210 | if (dupsweedout_tables && !(remaining_tables & ~new_join_tab->table->map & |
3211 | dupsweedout_tables)) |
3212 | { |
3213 | /* |
3214 | Ok, reached a state where we could put a dups weedout point. |
3215 | Walk back and calculate |
3216 | - the join cost (this is needed as the accumulated cost may assume |
3217 | some other duplicate elimination method) |
3218 | - extra fanout that will be removed by duplicate elimination |
3219 | - duplicate elimination cost |
3220 | There are two cases: |
3221 | 1. We have other strategy/ies to remove all of the duplicates. |
3222 | 2. We don't. |
3223 | |
3224 | We need to calculate the cost in case #2 also because we need to make |
3225 | choice between this join order and others. |
3226 | */ |
3227 | uint first_tab= first_dupsweedout_table; |
3228 | double dups_cost; |
3229 | double prefix_rec_count; |
3230 | double sj_inner_fanout= 1.0; |
3231 | double sj_outer_fanout= 1.0; |
3232 | uint temptable_rec_size; |
3233 | if (first_tab == join->const_tables) |
3234 | { |
3235 | prefix_rec_count= 1.0; |
3236 | temptable_rec_size= 0; |
3237 | dups_cost= 0.0; |
3238 | } |
3239 | else |
3240 | { |
3241 | dups_cost= join->positions[first_tab - 1].prefix_cost.total_cost(); |
3242 | prefix_rec_count= join->positions[first_tab - 1].prefix_record_count; |
3243 | temptable_rec_size= 8; /* This is not true but we'll make it so */ |
3244 | } |
3245 | |
3246 | table_map dups_removed_fanout= 0; |
3247 | double current_fanout= prefix_rec_count; |
3248 | for (uint j= first_dupsweedout_table; j <= idx; j++) |
3249 | { |
3250 | POSITION *p= join->positions + j; |
3251 | current_fanout *= p->records_read; |
3252 | dups_cost += p->read_time + current_fanout / TIME_FOR_COMPARE; |
3253 | if (p->table->emb_sj_nest) |
3254 | { |
3255 | sj_inner_fanout *= p->records_read; |
3256 | dups_removed_fanout |= p->table->table->map; |
3257 | } |
3258 | else |
3259 | { |
3260 | sj_outer_fanout *= p->records_read; |
3261 | temptable_rec_size += p->table->table->file->ref_length; |
3262 | } |
3263 | } |
3264 | |
3265 | /* |
3266 | Add the cost of temptable use. The table will have sj_outer_fanout |
3267 | records, and we will make |
3268 | - sj_outer_fanout table writes |
3269 | - sj_inner_fanout*sj_outer_fanout lookups. |
3270 | |
3271 | */ |
3272 | double one_lookup_cost= get_tmp_table_lookup_cost(join->thd, |
3273 | sj_outer_fanout, |
3274 | temptable_rec_size); |
3275 | double one_write_cost= get_tmp_table_write_cost(join->thd, |
3276 | sj_outer_fanout, |
3277 | temptable_rec_size); |
3278 | |
3279 | double write_cost= join->positions[first_tab].prefix_record_count* |
3280 | sj_outer_fanout * one_write_cost; |
3281 | double full_lookup_cost= join->positions[first_tab].prefix_record_count* |
3282 | sj_outer_fanout* sj_inner_fanout * |
3283 | one_lookup_cost; |
3284 | dups_cost += write_cost + full_lookup_cost; |
3285 | |
3286 | *read_time= dups_cost; |
3287 | *record_count= prefix_rec_count * sj_outer_fanout; |
3288 | *handled_fanout= dups_removed_fanout; |
3289 | *strategy= SJ_OPT_DUPS_WEEDOUT; |
3290 | return TRUE; |
3291 | } |
3292 | return FALSE; |
3293 | } |
3294 | |
3295 | |
3296 | /* |
3297 | Remove the last join tab from from join->cur_sj_inner_tables bitmap |
3298 | we assume remaining_tables doesnt contain @tab. |
3299 | */ |
3300 | |
3301 | void restore_prev_sj_state(const table_map remaining_tables, |
3302 | const JOIN_TAB *tab, uint idx) |
3303 | { |
3304 | TABLE_LIST *emb_sj_nest; |
3305 | |
3306 | if (tab->emb_sj_nest) |
3307 | { |
3308 | table_map subq_tables= tab->emb_sj_nest->sj_inner_tables; |
3309 | tab->join->sjm_lookup_tables &= ~subq_tables; |
3310 | } |
3311 | |
3312 | if ((emb_sj_nest= tab->emb_sj_nest)) |
3313 | { |
3314 | /* If we're removing the last SJ-inner table, remove the sj-nest */ |
3315 | if ((remaining_tables & emb_sj_nest->sj_inner_tables) == |
3316 | (emb_sj_nest->sj_inner_tables & ~tab->table->map)) |
3317 | { |
3318 | tab->join->cur_sj_inner_tables &= ~emb_sj_nest->sj_inner_tables; |
3319 | } |
3320 | } |
3321 | } |
3322 | |
3323 | |
3324 | /* |
3325 | Given a semi-join nest, find out which of the IN-equalities are bound |
3326 | |
3327 | SYNOPSIS |
3328 | get_bound_sj_equalities() |
3329 | sj_nest Semi-join nest |
3330 | remaining_tables Tables that are not yet bound |
3331 | |
3332 | DESCRIPTION |
3333 | Given a semi-join nest, find out which of the IN-equalities have their |
3334 | left part expression bound (i.e. the said expression doesn't refer to |
3335 | any of remaining_tables and can be evaluated). |
3336 | |
3337 | RETURN |
3338 | Bitmap of bound IN-equalities. |
3339 | */ |
3340 | |
3341 | ulonglong get_bound_sj_equalities(TABLE_LIST *sj_nest, |
3342 | table_map remaining_tables) |
3343 | { |
3344 | List_iterator<Item_ptr> li(sj_nest->nested_join->sj_outer_expr_list); |
3345 | Item **item; |
3346 | uint i= 0; |
3347 | ulonglong res= 0; |
3348 | while ((item= li++)) |
3349 | { |
3350 | /* |
3351 | Q: should this take into account equality propagation and how? |
3352 | A: If e->outer_side is an Item_field, walk over the equality |
3353 | class and see if there is an element that is bound? |
3354 | (this is an optional feature) |
3355 | */ |
3356 | if (!(item[0]->used_tables() & remaining_tables)) |
3357 | { |
3358 | res |= 1ULL << i; |
3359 | } |
3360 | i++; |
3361 | } |
3362 | return res; |
3363 | } |
3364 | |
3365 | |
3366 | /* |
3367 | Check if the last tables of the partial join order allow to use |
3368 | sj-materialization strategy for them |
3369 | |
3370 | SYNOPSIS |
3371 | at_sjmat_pos() |
3372 | join |
3373 | remaining_tables |
3374 | tab the last table's join tab |
3375 | idx last table's index |
3376 | loose_scan OUT TRUE <=> use LooseScan |
3377 | |
3378 | RETURN |
3379 | TRUE Yes, can apply sj-materialization |
3380 | FALSE No, some of the requirements are not met |
3381 | */ |
3382 | |
3383 | static SJ_MATERIALIZATION_INFO * |
3384 | at_sjmat_pos(const JOIN *join, table_map remaining_tables, const JOIN_TAB *tab, |
3385 | uint idx, bool *loose_scan) |
3386 | { |
3387 | /* |
3388 | Check if |
3389 | 1. We're in a semi-join nest that can be run with SJ-materialization |
3390 | 2. All the tables correlated through the IN subquery are in the prefix |
3391 | */ |
3392 | TABLE_LIST *emb_sj_nest= tab->emb_sj_nest; |
3393 | table_map suffix= remaining_tables & ~tab->table->map; |
3394 | if (emb_sj_nest && emb_sj_nest->sj_mat_info && |
3395 | !(suffix & emb_sj_nest->sj_inner_tables)) |
3396 | { |
3397 | /* |
3398 | Walk back and check if all immediately preceding tables are from |
3399 | this semi-join. |
3400 | */ |
3401 | uint n_tables= my_count_bits(tab->emb_sj_nest->sj_inner_tables); |
3402 | for (uint i= 1; i < n_tables ; i++) |
3403 | { |
3404 | if (join->positions[idx - i].table->emb_sj_nest != tab->emb_sj_nest) |
3405 | return NULL; |
3406 | } |
3407 | *loose_scan= MY_TEST(remaining_tables & ~tab->table->map & |
3408 | (emb_sj_nest->sj_inner_tables | |
3409 | emb_sj_nest->nested_join->sj_depends_on)); |
3410 | if (*loose_scan && !emb_sj_nest->sj_subq_pred->sjm_scan_allowed) |
3411 | return NULL; |
3412 | else |
3413 | return emb_sj_nest->sj_mat_info; |
3414 | } |
3415 | return NULL; |
3416 | } |
3417 | |
3418 | |
3419 | /* |
3420 | Re-calculate values of join->best_positions[start..end].prefix_record_count |
3421 | */ |
3422 | |
3423 | static void recalculate_prefix_record_count(JOIN *join, uint start, uint end) |
3424 | { |
3425 | for (uint j= start; j < end ;j++) |
3426 | { |
3427 | double prefix_count; |
3428 | if (j == join->const_tables) |
3429 | prefix_count= 1.0; |
3430 | else |
3431 | prefix_count= join->best_positions[j-1].prefix_record_count * |
3432 | join->best_positions[j-1].records_read; |
3433 | |
3434 | join->best_positions[j].prefix_record_count= prefix_count; |
3435 | } |
3436 | } |
3437 | |
3438 | |
3439 | /* |
3440 | Fix semi-join strategies for the picked join order |
3441 | |
3442 | SYNOPSIS |
3443 | fix_semijoin_strategies_for_picked_join_order() |
3444 | join The join with the picked join order |
3445 | |
3446 | DESCRIPTION |
3447 | Fix semi-join strategies for the picked join order. This is a step that |
3448 | needs to be done right after we have fixed the join order. What we do |
3449 | here is switch join's semi-join strategy description from backward-based |
3450 | to forwards based. |
3451 | |
3452 | When join optimization is in progress, we re-consider semi-join |
3453 | strategies after we've added another table. Here's an illustration. |
3454 | Suppose the join optimization is underway: |
3455 | |
3456 | 1) ot1 it1 it2 |
3457 | sjX -- looking at (ot1, it1, it2) join prefix, we decide |
3458 | to use semi-join strategy sjX. |
3459 | |
3460 | 2) ot1 it1 it2 ot2 |
3461 | sjX sjY -- Having added table ot2, we now may consider |
3462 | another semi-join strategy and decide to use a |
3463 | different strategy sjY. Note that the record |
3464 | of sjX has remained under it2. That is |
3465 | necessary because we need to be able to get |
3466 | back to (ot1, it1, it2) join prefix. |
3467 | what makes things even worse is that there are cases where the choice |
3468 | of sjY changes the way we should access it2. |
3469 | |
3470 | 3) [ot1 it1 it2 ot2 ot3] |
3471 | sjX sjY -- This means that after join optimization is |
3472 | finished, semi-join info should be read |
3473 | right-to-left (while nearly all plan refinement |
3474 | functions, EXPLAIN, etc proceed from left to |
3475 | right) |
3476 | |
3477 | This function does the needed reversal, making it possible to read the |
3478 | join and semi-join order from left to right. |
3479 | */ |
3480 | |
3481 | void fix_semijoin_strategies_for_picked_join_order(JOIN *join) |
3482 | { |
3483 | uint table_count=join->table_count; |
3484 | uint tablenr; |
3485 | table_map remaining_tables= 0; |
3486 | table_map handled_tabs= 0; |
3487 | join->sjm_lookup_tables= 0; |
3488 | join->sjm_scan_tables= 0; |
3489 | for (tablenr= table_count - 1 ; tablenr != join->const_tables - 1; tablenr--) |
3490 | { |
3491 | POSITION *pos= join->best_positions + tablenr; |
3492 | JOIN_TAB *s= pos->table; |
3493 | uint UNINIT_VAR(first); // Set by every branch except SJ_OPT_NONE which doesn't use it |
3494 | |
3495 | if ((handled_tabs & s->table->map) || pos->sj_strategy == SJ_OPT_NONE) |
3496 | { |
3497 | remaining_tables |= s->table->map; |
3498 | continue; |
3499 | } |
3500 | |
3501 | if (pos->sj_strategy == SJ_OPT_MATERIALIZE) |
3502 | { |
3503 | SJ_MATERIALIZATION_INFO *sjm= s->emb_sj_nest->sj_mat_info; |
3504 | sjm->is_used= TRUE; |
3505 | sjm->is_sj_scan= FALSE; |
3506 | memcpy((uchar*) (pos - sjm->tables + 1), (uchar*) sjm->positions, |
3507 | sizeof(POSITION) * sjm->tables); |
3508 | recalculate_prefix_record_count(join, tablenr - sjm->tables + 1, |
3509 | tablenr); |
3510 | first= tablenr - sjm->tables + 1; |
3511 | join->best_positions[first].n_sj_tables= sjm->tables; |
3512 | join->best_positions[first].sj_strategy= SJ_OPT_MATERIALIZE; |
3513 | join->sjm_lookup_tables|= s->table->map; |
3514 | } |
3515 | else if (pos->sj_strategy == SJ_OPT_MATERIALIZE_SCAN) |
3516 | { |
3517 | POSITION *first_inner= join->best_positions + pos->sjmat_picker.sjm_scan_last_inner; |
3518 | SJ_MATERIALIZATION_INFO *sjm= first_inner->table->emb_sj_nest->sj_mat_info; |
3519 | sjm->is_used= TRUE; |
3520 | sjm->is_sj_scan= TRUE; |
3521 | first= pos->sjmat_picker.sjm_scan_last_inner - sjm->tables + 1; |
3522 | memcpy((uchar*) (join->best_positions + first), |
3523 | (uchar*) sjm->positions, sizeof(POSITION) * sjm->tables); |
3524 | recalculate_prefix_record_count(join, first, first + sjm->tables); |
3525 | join->best_positions[first].sj_strategy= SJ_OPT_MATERIALIZE_SCAN; |
3526 | join->best_positions[first].n_sj_tables= sjm->tables; |
3527 | /* |
3528 | Do what advance_sj_state did: re-run best_access_path for every table |
3529 | in the [last_inner_table + 1; pos..) range |
3530 | */ |
3531 | double prefix_rec_count; |
3532 | /* Get the prefix record count */ |
3533 | if (first == join->const_tables) |
3534 | prefix_rec_count= 1.0; |
3535 | else |
3536 | prefix_rec_count= join->best_positions[first-1].prefix_record_count; |
3537 | |
3538 | /* Add materialization record count*/ |
3539 | prefix_rec_count *= sjm->rows; |
3540 | |
3541 | uint i; |
3542 | table_map rem_tables= remaining_tables; |
3543 | for (i= tablenr; i != (first + sjm->tables - 1); i--) |
3544 | rem_tables |= join->best_positions[i].table->table->map; |
3545 | |
3546 | for (i= first; i < first+ sjm->tables; i++) |
3547 | join->sjm_scan_tables |= join->best_positions[i].table->table->map; |
3548 | |
3549 | POSITION dummy; |
3550 | join->cur_sj_inner_tables= 0; |
3551 | for (i= first + sjm->tables; i <= tablenr; i++) |
3552 | { |
3553 | best_access_path(join, join->best_positions[i].table, rem_tables, i, |
3554 | FALSE, prefix_rec_count, |
3555 | join->best_positions + i, &dummy); |
3556 | prefix_rec_count *= join->best_positions[i].records_read; |
3557 | rem_tables &= ~join->best_positions[i].table->table->map; |
3558 | } |
3559 | } |
3560 | |
3561 | if (pos->sj_strategy == SJ_OPT_FIRST_MATCH) |
3562 | { |
3563 | first= pos->firstmatch_picker.first_firstmatch_table; |
3564 | join->best_positions[first].sj_strategy= SJ_OPT_FIRST_MATCH; |
3565 | join->best_positions[first].n_sj_tables= tablenr - first + 1; |
3566 | POSITION dummy; // For loose scan paths |
3567 | double record_count= (first== join->const_tables)? 1.0: |
3568 | join->best_positions[tablenr - 1].prefix_record_count; |
3569 | |
3570 | table_map rem_tables= remaining_tables; |
3571 | uint idx; |
3572 | for (idx= first; idx <= tablenr; idx++) |
3573 | { |
3574 | rem_tables |= join->best_positions[idx].table->table->map; |
3575 | } |
3576 | /* |
3577 | Re-run best_access_path to produce best access methods that do not use |
3578 | join buffering |
3579 | */ |
3580 | join->cur_sj_inner_tables= 0; |
3581 | for (idx= first; idx <= tablenr; idx++) |
3582 | { |
3583 | if (join->best_positions[idx].use_join_buffer) |
3584 | { |
3585 | best_access_path(join, join->best_positions[idx].table, |
3586 | rem_tables, idx, TRUE /* no jbuf */, |
3587 | record_count, join->best_positions + idx, &dummy); |
3588 | } |
3589 | record_count *= join->best_positions[idx].records_read; |
3590 | rem_tables &= ~join->best_positions[idx].table->table->map; |
3591 | } |
3592 | } |
3593 | |
3594 | if (pos->sj_strategy == SJ_OPT_LOOSE_SCAN) |
3595 | { |
3596 | first= pos->loosescan_picker.first_loosescan_table; |
3597 | POSITION *first_pos= join->best_positions + first; |
3598 | POSITION loose_scan_pos; // For loose scan paths |
3599 | double record_count= (first== join->const_tables)? 1.0: |
3600 | join->best_positions[tablenr - 1].prefix_record_count; |
3601 | |
3602 | table_map rem_tables= remaining_tables; |
3603 | uint idx; |
3604 | for (idx= first; idx <= tablenr; idx++) |
3605 | rem_tables |= join->best_positions[idx].table->table->map; |
3606 | /* |
3607 | Re-run best_access_path to produce best access methods that do not use |
3608 | join buffering |
3609 | */ |
3610 | join->cur_sj_inner_tables= 0; |
3611 | for (idx= first; idx <= tablenr; idx++) |
3612 | { |
3613 | if (join->best_positions[idx].use_join_buffer || (idx == first)) |
3614 | { |
3615 | best_access_path(join, join->best_positions[idx].table, |
3616 | rem_tables, idx, TRUE /* no jbuf */, |
3617 | record_count, join->best_positions + idx, |
3618 | &loose_scan_pos); |
3619 | if (idx==first) |
3620 | { |
3621 | join->best_positions[idx]= loose_scan_pos; |
3622 | /* |
3623 | If LooseScan is based on ref access (including the "degenerate" |
3624 | one with 0 key parts), we should use full index scan. |
3625 | |
3626 | Unfortunately, lots of code assumes that if tab->type==JT_ALL && |
3627 | tab->quick!=NULL, then quick select should be used. The only |
3628 | simple way to fix this is to remove the quick select: |
3629 | */ |
3630 | if (join->best_positions[idx].key) |
3631 | { |
3632 | delete join->best_positions[idx].table->quick; |
3633 | join->best_positions[idx].table->quick= NULL; |
3634 | } |
3635 | } |
3636 | } |
3637 | rem_tables &= ~join->best_positions[idx].table->table->map; |
3638 | record_count *= join->best_positions[idx].records_read; |
3639 | } |
3640 | first_pos->sj_strategy= SJ_OPT_LOOSE_SCAN; |
3641 | first_pos->n_sj_tables= my_count_bits(first_pos->table->emb_sj_nest->sj_inner_tables); |
3642 | } |
3643 | |
3644 | if (pos->sj_strategy == SJ_OPT_DUPS_WEEDOUT) |
3645 | { |
3646 | /* |
3647 | Duplicate Weedout starting at pos->first_dupsweedout_table, ending at |
3648 | this table. |
3649 | */ |
3650 | first= pos->dups_weedout_picker.first_dupsweedout_table; |
3651 | join->best_positions[first].sj_strategy= SJ_OPT_DUPS_WEEDOUT; |
3652 | join->best_positions[first].n_sj_tables= tablenr - first + 1; |
3653 | } |
3654 | |
3655 | uint i_end= first + join->best_positions[first].n_sj_tables; |
3656 | for (uint i= first; i < i_end; i++) |
3657 | { |
3658 | if (i != first) |
3659 | join->best_positions[i].sj_strategy= SJ_OPT_NONE; |
3660 | handled_tabs |= join->best_positions[i].table->table->map; |
3661 | } |
3662 | |
3663 | if (tablenr != first) |
3664 | pos->sj_strategy= SJ_OPT_NONE; |
3665 | remaining_tables |= s->table->map; |
3666 | join->join_tab[first].sj_strategy= join->best_positions[first].sj_strategy; |
3667 | join->join_tab[first].n_sj_tables= join->best_positions[first].n_sj_tables; |
3668 | } |
3669 | } |
3670 | |
3671 | |
3672 | /* |
3673 | Setup semi-join materialization strategy for one semi-join nest |
3674 | |
3675 | SYNOPSIS |
3676 | |
3677 | setup_sj_materialization() |
3678 | tab The first tab in the semi-join |
3679 | |
3680 | DESCRIPTION |
3681 | Setup execution structures for one semi-join materialization nest: |
3682 | - Create the materialization temporary table |
3683 | - If we're going to do index lookups |
3684 | create TABLE_REF structure to make the lookus |
3685 | - else (if we're going to do a full scan of the temptable) |
3686 | create Copy_field structures to do copying. |
3687 | |
3688 | RETURN |
3689 | FALSE Ok |
3690 | TRUE Error |
3691 | */ |
3692 | |
3693 | bool setup_sj_materialization_part1(JOIN_TAB *sjm_tab) |
3694 | { |
3695 | JOIN_TAB *tab= sjm_tab->bush_children->start; |
3696 | TABLE_LIST *emb_sj_nest= tab->table->pos_in_table_list->embedding; |
3697 | SJ_MATERIALIZATION_INFO *sjm; |
3698 | THD *thd; |
3699 | |
3700 | DBUG_ENTER("setup_sj_materialization" ); |
3701 | |
3702 | /* Walk out of outer join nests until we reach the semi-join nest we're in */ |
3703 | while (!emb_sj_nest->sj_mat_info) |
3704 | emb_sj_nest= emb_sj_nest->embedding; |
3705 | |
3706 | sjm= emb_sj_nest->sj_mat_info; |
3707 | thd= tab->join->thd; |
3708 | /* First the calls come to the materialization function */ |
3709 | |
3710 | DBUG_ASSERT(sjm->is_used); |
3711 | /* |
3712 | Set up the table to write to, do as select_union::create_result_table does |
3713 | */ |
3714 | sjm->sjm_table_param.init(); |
3715 | sjm->sjm_table_param.bit_fields_as_long= TRUE; |
3716 | SELECT_LEX *subq_select= emb_sj_nest->sj_subq_pred->unit->first_select(); |
3717 | const LEX_CSTRING sj_materialize_name= { STRING_WITH_LEN("sj-materialize" ) }; |
3718 | List_iterator<Item> it(subq_select->item_list); |
3719 | Item *item; |
3720 | while((item= it++)) |
3721 | { |
3722 | /* |
3723 | This semi-join replaced the subquery (subq_select) and so on |
3724 | re-executing it will not be prepared. To use the Items from its |
3725 | select list we have to prepare (fix_fields) them |
3726 | */ |
3727 | if (!item->fixed && item->fix_fields(thd, it.ref())) |
3728 | DBUG_RETURN(TRUE); |
3729 | item= *(it.ref()); // it can be changed by fix_fields |
3730 | DBUG_ASSERT(!item->name.length || item->name.length == strlen(item->name.str)); |
3731 | sjm->sjm_table_cols.push_back(item, thd->mem_root); |
3732 | } |
3733 | |
3734 | sjm->sjm_table_param.field_count= subq_select->item_list.elements; |
3735 | sjm->sjm_table_param.force_not_null_cols= TRUE; |
3736 | |
3737 | if (!(sjm->table= create_tmp_table(thd, &sjm->sjm_table_param, |
3738 | sjm->sjm_table_cols, (ORDER*) 0, |
3739 | TRUE /* distinct */, |
3740 | 1, /*save_sum_fields*/ |
3741 | thd->variables.option_bits | TMP_TABLE_ALL_COLUMNS, |
3742 | HA_POS_ERROR /*rows_limit */, |
3743 | &sj_materialize_name))) |
3744 | DBUG_RETURN(TRUE); /* purecov: inspected */ |
3745 | sjm->table->map= emb_sj_nest->nested_join->used_tables; |
3746 | sjm->table->file->extra(HA_EXTRA_WRITE_CACHE); |
3747 | sjm->table->file->extra(HA_EXTRA_IGNORE_DUP_KEY); |
3748 | |
3749 | tab->join->sj_tmp_tables.push_back(sjm->table, thd->mem_root); |
3750 | tab->join->sjm_info_list.push_back(sjm, thd->mem_root); |
3751 | |
3752 | sjm->materialized= FALSE; |
3753 | sjm_tab->table= sjm->table; |
3754 | sjm->table->pos_in_table_list= emb_sj_nest; |
3755 | |
3756 | DBUG_RETURN(FALSE); |
3757 | } |
3758 | |
3759 | /** |
3760 | @retval |
3761 | FALSE ok |
3762 | TRUE error |
3763 | */ |
3764 | |
3765 | bool setup_sj_materialization_part2(JOIN_TAB *sjm_tab) |
3766 | { |
3767 | DBUG_ENTER("setup_sj_materialization_part2" ); |
3768 | JOIN_TAB *tab= sjm_tab->bush_children->start; |
3769 | TABLE_LIST *emb_sj_nest= tab->table->pos_in_table_list->embedding; |
3770 | /* Walk out of outer join nests until we reach the semi-join nest we're in */ |
3771 | while (!emb_sj_nest->sj_mat_info) |
3772 | emb_sj_nest= emb_sj_nest->embedding; |
3773 | SJ_MATERIALIZATION_INFO *sjm= emb_sj_nest->sj_mat_info; |
3774 | THD *thd= tab->join->thd; |
3775 | uint i; |
3776 | |
3777 | if (!sjm->is_sj_scan) |
3778 | { |
3779 | KEY *tmp_key; /* The only index on the temporary table. */ |
3780 | uint tmp_key_parts; /* Number of keyparts in tmp_key. */ |
3781 | tmp_key= sjm->table->key_info; |
3782 | tmp_key_parts= tmp_key->user_defined_key_parts; |
3783 | |
3784 | /* |
3785 | Create/initialize everything we will need to index lookups into the |
3786 | temptable. |
3787 | */ |
3788 | TABLE_REF *tab_ref; |
3789 | tab_ref= &sjm_tab->ref; |
3790 | tab_ref->key= 0; /* The only temp table index. */ |
3791 | tab_ref->key_length= tmp_key->key_length; |
3792 | if (!(tab_ref->key_buff= |
3793 | (uchar*) thd->calloc(ALIGN_SIZE(tmp_key->key_length) * 2)) || |
3794 | !(tab_ref->key_copy= |
3795 | (store_key**) thd->alloc((sizeof(store_key*) * |
3796 | (tmp_key_parts + 1)))) || |
3797 | !(tab_ref->items= |
3798 | (Item**) thd->alloc(sizeof(Item*) * tmp_key_parts))) |
3799 | DBUG_RETURN(TRUE); /* purecov: inspected */ |
3800 | |
3801 | tab_ref->key_buff2=tab_ref->key_buff+ALIGN_SIZE(tmp_key->key_length); |
3802 | tab_ref->key_err=1; |
3803 | tab_ref->null_rejecting= 1; |
3804 | tab_ref->disable_cache= FALSE; |
3805 | |
3806 | KEY_PART_INFO *cur_key_part= tmp_key->key_part; |
3807 | store_key **ref_key= tab_ref->key_copy; |
3808 | uchar *cur_ref_buff= tab_ref->key_buff; |
3809 | |
3810 | for (i= 0; i < tmp_key_parts; i++, cur_key_part++, ref_key++) |
3811 | { |
3812 | tab_ref->items[i]= emb_sj_nest->sj_subq_pred->left_expr->element_index(i); |
3813 | int null_count= MY_TEST(cur_key_part->field->real_maybe_null()); |
3814 | *ref_key= new store_key_item(thd, cur_key_part->field, |
3815 | /* TODO: |
3816 | the NULL byte is taken into account in |
3817 | cur_key_part->store_length, so instead of |
3818 | cur_ref_buff + MY_TEST(maybe_null), we could |
3819 | use that information instead. |
3820 | */ |
3821 | cur_ref_buff + null_count, |
3822 | null_count ? cur_ref_buff : 0, |
3823 | cur_key_part->length, tab_ref->items[i], |
3824 | FALSE); |
3825 | if (!*ref_key) |
3826 | DBUG_RETURN(TRUE); |
3827 | cur_ref_buff+= cur_key_part->store_length; |
3828 | } |
3829 | *ref_key= NULL; /* End marker. */ |
3830 | |
3831 | /* |
3832 | We don't ever have guarded conditions for SJM tables, but code at SQL |
3833 | layer depends on cond_guards array being alloced. |
3834 | */ |
3835 | if (!(tab_ref->cond_guards= (bool**) thd->calloc(sizeof(uint*)*tmp_key_parts))) |
3836 | { |
3837 | DBUG_RETURN(TRUE); |
3838 | } |
3839 | |
3840 | tab_ref->key_err= 1; |
3841 | tab_ref->key_parts= tmp_key_parts; |
3842 | sjm->tab_ref= tab_ref; |
3843 | |
3844 | /* |
3845 | Remove the injected semi-join IN-equalities from join_tab conds. This |
3846 | needs to be done because the IN-equalities refer to columns of |
3847 | sj-inner tables which are not available after the materialization |
3848 | has been finished. |
3849 | */ |
3850 | for (i= 0; i < sjm->tables; i++) |
3851 | { |
3852 | if (remove_sj_conds(thd, &tab[i].select_cond) || |
3853 | (tab[i].select && remove_sj_conds(thd, &tab[i].select->cond))) |
3854 | DBUG_RETURN(TRUE); |
3855 | } |
3856 | if (!(sjm->in_equality= create_subq_in_equalities(thd, sjm, |
3857 | emb_sj_nest->sj_subq_pred))) |
3858 | DBUG_RETURN(TRUE); /* purecov: inspected */ |
3859 | sjm_tab->type= JT_EQ_REF; |
3860 | sjm_tab->select_cond= sjm->in_equality; |
3861 | } |
3862 | else |
3863 | { |
3864 | /* |
3865 | We'll be doing full scan of the temptable. |
3866 | Setup copying of temptable columns back to the record buffers |
3867 | for their source tables. We need this because IN-equalities |
3868 | refer to the original tables. |
3869 | |
3870 | EXAMPLE |
3871 | |
3872 | Consider the query: |
3873 | SELECT * FROM ot WHERE ot.col1 IN (SELECT it.col2 FROM it) |
3874 | |
3875 | Suppose it's executed with SJ-Materialization-scan. We choose to do scan |
3876 | if we can't do the lookup, i.e. the join order is (it, ot). The plan |
3877 | would look as follows: |
3878 | |
3879 | table access method condition |
3880 | it materialize+scan - |
3881 | ot (whatever) ot1.col1=it.col2 (C2) |
3882 | |
3883 | The condition C2 refers to current row of table it. The problem is |
3884 | that by the time we evaluate C2, we would have finished with scanning |
3885 | it itself and will be scanning the temptable. |
3886 | |
3887 | At the moment, our solution is to copy back: when we get the next |
3888 | temptable record, we copy its columns to their corresponding columns |
3889 | in the record buffers for the source tables. |
3890 | */ |
3891 | if (!(sjm->copy_field= new Copy_field[sjm->sjm_table_cols.elements])) |
3892 | DBUG_RETURN(TRUE); |
3893 | |
3894 | //it.rewind(); |
3895 | Ref_ptr_array p_items= emb_sj_nest->sj_subq_pred->unit->first_select()->ref_pointer_array; |
3896 | for (uint i=0; i < sjm->sjm_table_cols.elements; i++) |
3897 | { |
3898 | bool dummy; |
3899 | Item_equal *item_eq; |
3900 | //Item *item= (it++)->real_item(); |
3901 | Item *item= p_items[i]->real_item(); |
3902 | DBUG_ASSERT(item->type() == Item::FIELD_ITEM); |
3903 | Field *copy_to= ((Item_field*)item)->field; |
3904 | /* |
3905 | Tricks with Item_equal are due to the following: suppose we have a |
3906 | query: |
3907 | |
3908 | ... WHERE cond(ot.col) AND ot.col IN (SELECT it2.col FROM it1,it2 |
3909 | WHERE it1.col= it2.col) |
3910 | then equality propagation will create an |
3911 | |
3912 | Item_equal(it1.col, it2.col, ot.col) |
3913 | |
3914 | then substitute_for_best_equal_field() will change the conditions |
3915 | according to the join order: |
3916 | |
3917 | table | attached condition |
3918 | ------+-------------------- |
3919 | it1 | |
3920 | it2 | it1.col=it2.col |
3921 | ot | cond(it1.col) |
3922 | |
3923 | although we've originally had "SELECT it2.col", conditions attached |
3924 | to subsequent outer tables will refer to it1.col, so SJM-Scan will |
3925 | need to unpack data to there. |
3926 | That is, if an element from subquery's select list participates in |
3927 | equality propagation, then we need to unpack it to the first |
3928 | element equality propagation member that refers to table that is |
3929 | within the subquery. |
3930 | */ |
3931 | item_eq= find_item_equal(tab->join->cond_equal, copy_to, &dummy); |
3932 | |
3933 | if (item_eq) |
3934 | { |
3935 | List_iterator<Item> it(item_eq->equal_items); |
3936 | /* We're interested in field items only */ |
3937 | if (item_eq->get_const()) |
3938 | it++; |
3939 | Item *item; |
3940 | while ((item= it++)) |
3941 | { |
3942 | if (!(item->used_tables() & ~emb_sj_nest->sj_inner_tables)) |
3943 | { |
3944 | DBUG_ASSERT(item->real_item()->type() == Item::FIELD_ITEM); |
3945 | copy_to= ((Item_field *) (item->real_item()))->field; |
3946 | break; |
3947 | } |
3948 | } |
3949 | } |
3950 | sjm->copy_field[i].set(copy_to, sjm->table->field[i], FALSE); |
3951 | /* The write_set for source tables must be set up to allow the copying */ |
3952 | bitmap_set_bit(copy_to->table->write_set, copy_to->field_index); |
3953 | } |
3954 | sjm_tab->type= JT_ALL; |
3955 | |
3956 | /* Initialize full scan */ |
3957 | sjm_tab->read_first_record= join_read_record_no_init; |
3958 | sjm_tab->read_record.copy_field= sjm->copy_field; |
3959 | sjm_tab->read_record.copy_field_end= sjm->copy_field + |
3960 | sjm->sjm_table_cols.elements; |
3961 | sjm_tab->read_record.read_record_func= rr_sequential_and_unpack; |
3962 | } |
3963 | |
3964 | sjm_tab->bush_children->end[-1].next_select= end_sj_materialize; |
3965 | |
3966 | DBUG_RETURN(FALSE); |
3967 | } |
3968 | |
3969 | |
3970 | |
3971 | /* |
3972 | Create subquery IN-equalities assuming use of materialization strategy |
3973 | |
3974 | SYNOPSIS |
3975 | create_subq_in_equalities() |
3976 | thd Thread handle |
3977 | sjm Semi-join materialization structure |
3978 | subq_pred The subquery predicate |
3979 | |
3980 | DESCRIPTION |
3981 | Create subquery IN-equality predicates. That is, for a subquery |
3982 | |
3983 | (oe1, oe2, ...) IN (SELECT ie1, ie2, ... FROM ...) |
3984 | |
3985 | create "oe1=ie1 AND ie1=ie2 AND ..." expression, such that ie1, ie2, .. |
3986 | refer to the columns of the table that's used to materialize the |
3987 | subquery. |
3988 | |
3989 | RETURN |
3990 | Created condition |
3991 | */ |
3992 | |
3993 | static Item *create_subq_in_equalities(THD *thd, SJ_MATERIALIZATION_INFO *sjm, |
3994 | Item_in_subselect *subq_pred) |
3995 | { |
3996 | Item *res= NULL; |
3997 | if (subq_pred->left_expr->cols() == 1) |
3998 | { |
3999 | if (!(res= new (thd->mem_root) Item_func_eq(thd, subq_pred->left_expr, |
4000 | new (thd->mem_root) Item_field(thd, sjm->table->field[0])))) |
4001 | return NULL; /* purecov: inspected */ |
4002 | } |
4003 | else |
4004 | { |
4005 | Item *conj; |
4006 | for (uint i= 0; i < subq_pred->left_expr->cols(); i++) |
4007 | { |
4008 | if (!(conj= new (thd->mem_root) Item_func_eq(thd, subq_pred->left_expr->element_index(i), |
4009 | new (thd->mem_root) Item_field(thd, sjm->table->field[i]))) || |
4010 | !(res= and_items(thd, res, conj))) |
4011 | return NULL; /* purecov: inspected */ |
4012 | } |
4013 | } |
4014 | if (res->fix_fields(thd, &res)) |
4015 | return NULL; /* purecov: inspected */ |
4016 | return res; |
4017 | } |
4018 | |
4019 | |
4020 | /** |
4021 | @retval |
4022 | 0 ok |
4023 | 1 error |
4024 | */ |
4025 | |
4026 | static bool remove_sj_conds(THD *thd, Item **tree) |
4027 | { |
4028 | if (*tree) |
4029 | { |
4030 | if (is_cond_sj_in_equality(*tree)) |
4031 | { |
4032 | *tree= NULL; |
4033 | return 0; |
4034 | } |
4035 | else if ((*tree)->type() == Item::COND_ITEM) |
4036 | { |
4037 | Item *item; |
4038 | List_iterator<Item> li(*(((Item_cond*)*tree)->argument_list())); |
4039 | while ((item= li++)) |
4040 | { |
4041 | if (is_cond_sj_in_equality(item)) |
4042 | { |
4043 | Item_int *tmp= new (thd->mem_root) Item_int(thd, 1); |
4044 | if (!tmp) |
4045 | return 1; |
4046 | li.replace(tmp); |
4047 | } |
4048 | } |
4049 | } |
4050 | } |
4051 | return 0; |
4052 | } |
4053 | |
4054 | |
4055 | /* Check if given Item was injected by semi-join equality */ |
4056 | static bool is_cond_sj_in_equality(Item *item) |
4057 | { |
4058 | if (item->type() == Item::FUNC_ITEM && |
4059 | ((Item_func*)item)->functype()== Item_func::EQ_FUNC) |
4060 | { |
4061 | Item_func_eq *item_eq= (Item_func_eq*)item; |
4062 | return MY_TEST(item_eq->in_equality_no != UINT_MAX); |
4063 | } |
4064 | return FALSE; |
4065 | } |
4066 | |
4067 | |
4068 | /* |
4069 | Create a temporary table to weed out duplicate rowid combinations |
4070 | |
4071 | SYNOPSIS |
4072 | |
4073 | create_sj_weedout_tmp_table() |
4074 | thd Thread handle |
4075 | |
4076 | DESCRIPTION |
4077 | Create a temporary table to weed out duplicate rowid combinations. The |
4078 | table has a single column that is a concatenation of all rowids in the |
4079 | combination. |
4080 | |
4081 | Depending on the needed length, there are two cases: |
4082 | |
4083 | 1. When the length of the column < max_key_length: |
4084 | |
4085 | CREATE TABLE tmp (col VARBINARY(n) NOT NULL, UNIQUE KEY(col)); |
4086 | |
4087 | 2. Otherwise (not a valid SQL syntax but internally supported): |
4088 | |
4089 | CREATE TABLE tmp (col VARBINARY NOT NULL, UNIQUE CONSTRAINT(col)); |
4090 | |
4091 | The code in this function was produced by extraction of relevant parts |
4092 | from create_tmp_table(). |
4093 | |
4094 | RETURN |
4095 | created table |
4096 | NULL on error |
4097 | */ |
4098 | |
4099 | bool |
4100 | SJ_TMP_TABLE::create_sj_weedout_tmp_table(THD *thd) |
4101 | { |
4102 | MEM_ROOT *mem_root_save, own_root; |
4103 | TABLE *table; |
4104 | TABLE_SHARE *share; |
4105 | uint temp_pool_slot=MY_BIT_NONE; |
4106 | char *tmpname,path[FN_REFLEN]; |
4107 | Field **reg_field; |
4108 | KEY_PART_INFO *key_part_info; |
4109 | KEY *keyinfo; |
4110 | uchar *group_buff; |
4111 | uchar *bitmaps; |
4112 | uint *blob_field; |
4113 | bool using_unique_constraint=FALSE; |
4114 | bool use_packed_rows= FALSE; |
4115 | Field *field, *key_field; |
4116 | uint null_pack_length, null_count; |
4117 | uchar *null_flags; |
4118 | uchar *pos; |
4119 | DBUG_ENTER("create_sj_weedout_tmp_table" ); |
4120 | DBUG_ASSERT(!is_degenerate); |
4121 | |
4122 | tmp_table= NULL; |
4123 | uint uniq_tuple_length_arg= rowid_len + null_bytes; |
4124 | /* |
4125 | STEP 1: Get temporary table name |
4126 | */ |
4127 | if (use_temp_pool && !(test_flags & TEST_KEEP_TMP_TABLES)) |
4128 | temp_pool_slot = bitmap_lock_set_next(&temp_pool); |
4129 | |
4130 | if (temp_pool_slot != MY_BIT_NONE) // we got a slot |
4131 | sprintf(path, "%s_%lx_%i" , tmp_file_prefix, |
4132 | current_pid, temp_pool_slot); |
4133 | else |
4134 | { |
4135 | /* if we run out of slots or we are not using tempool */ |
4136 | sprintf(path,"%s%lx_%lx_%x" , tmp_file_prefix,current_pid, |
4137 | (ulong) thd->thread_id, thd->tmp_table++); |
4138 | } |
4139 | fn_format(path, path, mysql_tmpdir, "" , MY_REPLACE_EXT|MY_UNPACK_FILENAME); |
4140 | |
4141 | /* STEP 2: Figure if we'll be using a key or blob+constraint */ |
4142 | /* it always has my_charset_bin, so mbmaxlen==1 */ |
4143 | if (uniq_tuple_length_arg >= CONVERT_IF_BIGGER_TO_BLOB) |
4144 | using_unique_constraint= TRUE; |
4145 | |
4146 | /* STEP 3: Allocate memory for temptable description */ |
4147 | init_sql_alloc(&own_root, "SJ_TMP_TABLE" , |
4148 | TABLE_ALLOC_BLOCK_SIZE, 0, MYF(MY_THREAD_SPECIFIC)); |
4149 | if (!multi_alloc_root(&own_root, |
4150 | &table, sizeof(*table), |
4151 | &share, sizeof(*share), |
4152 | ®_field, sizeof(Field*) * (1+1), |
4153 | &blob_field, sizeof(uint)*2, |
4154 | &keyinfo, sizeof(*keyinfo), |
4155 | &key_part_info, sizeof(*key_part_info) * 2, |
4156 | &start_recinfo, |
4157 | sizeof(*recinfo)*(1*2+4), |
4158 | &tmpname, (uint) strlen(path)+1, |
4159 | &group_buff, (!using_unique_constraint ? |
4160 | uniq_tuple_length_arg : 0), |
4161 | &bitmaps, bitmap_buffer_size(1)*6, |
4162 | NullS)) |
4163 | { |
4164 | if (temp_pool_slot != MY_BIT_NONE) |
4165 | bitmap_lock_clear_bit(&temp_pool, temp_pool_slot); |
4166 | DBUG_RETURN(TRUE); |
4167 | } |
4168 | strmov(tmpname,path); |
4169 | |
4170 | |
4171 | /* STEP 4: Create TABLE description */ |
4172 | bzero((char*) table,sizeof(*table)); |
4173 | bzero((char*) reg_field,sizeof(Field*)*2); |
4174 | |
4175 | table->mem_root= own_root; |
4176 | mem_root_save= thd->mem_root; |
4177 | thd->mem_root= &table->mem_root; |
4178 | |
4179 | table->field=reg_field; |
4180 | table->alias.set("weedout-tmp" , sizeof("weedout-tmp" )-1, |
4181 | table_alias_charset); |
4182 | table->reginfo.lock_type=TL_WRITE; /* Will be updated */ |
4183 | table->db_stat=HA_OPEN_KEYFILE; |
4184 | table->map=1; |
4185 | table->temp_pool_slot = temp_pool_slot; |
4186 | table->copy_blobs= 1; |
4187 | table->in_use= thd; |
4188 | table->quick_keys.init(); |
4189 | table->covering_keys.init(); |
4190 | table->keys_in_use_for_query.init(); |
4191 | |
4192 | table->s= share; |
4193 | init_tmp_table_share(thd, share, "" , 0, tmpname, tmpname); |
4194 | share->blob_field= blob_field; |
4195 | share->table_charset= NULL; |
4196 | share->primary_key= MAX_KEY; // Indicate no primary key |
4197 | share->keys_for_keyread.init(); |
4198 | share->keys_in_use.init(); |
4199 | |
4200 | /* Create the field */ |
4201 | { |
4202 | LEX_CSTRING field_name= {STRING_WITH_LEN("rowids" ) }; |
4203 | /* |
4204 | For the sake of uniformity, always use Field_varstring (altough we could |
4205 | use Field_string for shorter keys) |
4206 | */ |
4207 | field= new Field_varstring(uniq_tuple_length_arg, FALSE, &field_name, |
4208 | share, &my_charset_bin); |
4209 | if (!field) |
4210 | DBUG_RETURN(0); |
4211 | field->table= table; |
4212 | field->key_start.init(0); |
4213 | field->part_of_key.init(0); |
4214 | field->part_of_sortkey.init(0); |
4215 | field->unireg_check= Field::NONE; |
4216 | field->flags= (NOT_NULL_FLAG | BINARY_FLAG | NO_DEFAULT_VALUE_FLAG); |
4217 | field->reset_fields(); |
4218 | field->init(table); |
4219 | field->orig_table= NULL; |
4220 | |
4221 | field->field_index= 0; |
4222 | |
4223 | *(reg_field++)= field; |
4224 | *blob_field= 0; |
4225 | *reg_field= 0; |
4226 | |
4227 | share->fields= 1; |
4228 | share->blob_fields= 0; |
4229 | } |
4230 | |
4231 | uint reclength= field->pack_length(); |
4232 | if (using_unique_constraint) |
4233 | { |
4234 | share->db_plugin= ha_lock_engine(0, TMP_ENGINE_HTON); |
4235 | table->file= get_new_handler(share, &table->mem_root, |
4236 | share->db_type()); |
4237 | } |
4238 | else |
4239 | { |
4240 | share->db_plugin= ha_lock_engine(0, heap_hton); |
4241 | table->file= get_new_handler(share, &table->mem_root, |
4242 | share->db_type()); |
4243 | DBUG_ASSERT(!table->file || uniq_tuple_length_arg <= table->file->max_key_length()); |
4244 | } |
4245 | if (!table->file) |
4246 | goto err; |
4247 | |
4248 | if (table->file->set_ha_share_ref(&share->ha_share)) |
4249 | { |
4250 | delete table->file; |
4251 | goto err; |
4252 | } |
4253 | |
4254 | null_count=1; |
4255 | |
4256 | null_pack_length= 1; |
4257 | reclength += null_pack_length; |
4258 | |
4259 | share->reclength= reclength; |
4260 | { |
4261 | uint alloc_length=ALIGN_SIZE(share->reclength + MI_UNIQUE_HASH_LENGTH+1); |
4262 | share->rec_buff_length= alloc_length; |
4263 | if (!(table->record[0]= (uchar*) |
4264 | alloc_root(&table->mem_root, alloc_length*3))) |
4265 | goto err; |
4266 | table->record[1]= table->record[0]+alloc_length; |
4267 | share->default_values= table->record[1]+alloc_length; |
4268 | } |
4269 | setup_tmp_table_column_bitmaps(table, bitmaps); |
4270 | |
4271 | recinfo= start_recinfo; |
4272 | null_flags=(uchar*) table->record[0]; |
4273 | pos=table->record[0]+ null_pack_length; |
4274 | if (null_pack_length) |
4275 | { |
4276 | bzero((uchar*) recinfo,sizeof(*recinfo)); |
4277 | recinfo->type=FIELD_NORMAL; |
4278 | recinfo->length=null_pack_length; |
4279 | recinfo++; |
4280 | bfill(null_flags,null_pack_length,255); // Set null fields |
4281 | |
4282 | table->null_flags= (uchar*) table->record[0]; |
4283 | share->null_fields= null_count; |
4284 | share->null_bytes= null_pack_length; |
4285 | } |
4286 | null_count=1; |
4287 | |
4288 | { |
4289 | //Field *field= *reg_field; |
4290 | uint length; |
4291 | bzero((uchar*) recinfo,sizeof(*recinfo)); |
4292 | field->move_field(pos,(uchar*) 0,0); |
4293 | |
4294 | field->reset(); |
4295 | /* |
4296 | Test if there is a default field value. The test for ->ptr is to skip |
4297 | 'offset' fields generated by initalize_tables |
4298 | */ |
4299 | // Initialize the table field: |
4300 | bzero(field->ptr, field->pack_length()); |
4301 | |
4302 | length=field->pack_length(); |
4303 | pos+= length; |
4304 | |
4305 | /* Make entry for create table */ |
4306 | recinfo->length=length; |
4307 | if (field->flags & BLOB_FLAG) |
4308 | recinfo->type= FIELD_BLOB; |
4309 | else if (use_packed_rows && |
4310 | field->real_type() == MYSQL_TYPE_STRING && |
4311 | length >= MIN_STRING_LENGTH_TO_PACK_ROWS) |
4312 | recinfo->type=FIELD_SKIP_ENDSPACE; |
4313 | else |
4314 | recinfo->type=FIELD_NORMAL; |
4315 | |
4316 | field->set_table_name(&table->alias); |
4317 | } |
4318 | |
4319 | if (thd->variables.tmp_memory_table_size == ~ (ulonglong) 0) // No limit |
4320 | share->max_rows= ~(ha_rows) 0; |
4321 | else |
4322 | share->max_rows= (ha_rows) (((share->db_type() == heap_hton) ? |
4323 | MY_MIN(thd->variables.tmp_memory_table_size, |
4324 | thd->variables.max_heap_table_size) : |
4325 | thd->variables.tmp_memory_table_size) / |
4326 | share->reclength); |
4327 | set_if_bigger(share->max_rows,1); // For dummy start options |
4328 | |
4329 | |
4330 | //// keyinfo= param->keyinfo; |
4331 | if (TRUE) |
4332 | { |
4333 | DBUG_PRINT("info" ,("Creating group key in temporary table" )); |
4334 | share->keys=1; |
4335 | share->uniques= MY_TEST(using_unique_constraint); |
4336 | table->key_info=keyinfo; |
4337 | keyinfo->key_part=key_part_info; |
4338 | keyinfo->flags=HA_NOSAME; |
4339 | keyinfo->usable_key_parts= keyinfo->user_defined_key_parts= 1; |
4340 | keyinfo->key_length=0; |
4341 | keyinfo->rec_per_key=0; |
4342 | keyinfo->algorithm= HA_KEY_ALG_UNDEF; |
4343 | keyinfo->name= weedout_key; |
4344 | { |
4345 | key_part_info->null_bit=0; |
4346 | key_part_info->field= field; |
4347 | key_part_info->offset= field->offset(table->record[0]); |
4348 | key_part_info->length= (uint16) field->key_length(); |
4349 | key_part_info->type= (uint8) field->key_type(); |
4350 | key_part_info->key_type = FIELDFLAG_BINARY; |
4351 | if (!using_unique_constraint) |
4352 | { |
4353 | if (!(key_field= field->new_key_field(thd->mem_root, table, |
4354 | group_buff, |
4355 | key_part_info->length, |
4356 | field->null_ptr, |
4357 | field->null_bit))) |
4358 | goto err; |
4359 | } |
4360 | keyinfo->key_length+= key_part_info->length; |
4361 | } |
4362 | } |
4363 | |
4364 | if (unlikely(thd->is_fatal_error)) // If end of memory |
4365 | goto err; |
4366 | share->db_record_offset= 1; |
4367 | table->no_rows= 1; // We don't need the data |
4368 | |
4369 | // recinfo must point after last field |
4370 | recinfo++; |
4371 | if (share->db_type() == TMP_ENGINE_HTON) |
4372 | { |
4373 | if (unlikely(create_internal_tmp_table(table, keyinfo, start_recinfo, |
4374 | &recinfo, 0))) |
4375 | goto err; |
4376 | } |
4377 | if (unlikely(open_tmp_table(table))) |
4378 | goto err; |
4379 | |
4380 | thd->mem_root= mem_root_save; |
4381 | tmp_table= table; |
4382 | DBUG_RETURN(FALSE); |
4383 | |
4384 | err: |
4385 | thd->mem_root= mem_root_save; |
4386 | free_tmp_table(thd,table); /* purecov: inspected */ |
4387 | if (temp_pool_slot != MY_BIT_NONE) |
4388 | bitmap_lock_clear_bit(&temp_pool, temp_pool_slot); |
4389 | DBUG_RETURN(TRUE); /* purecov: inspected */ |
4390 | } |
4391 | |
4392 | |
4393 | /* |
4394 | SemiJoinDuplicateElimination: Reset the temporary table |
4395 | */ |
4396 | |
4397 | int SJ_TMP_TABLE::sj_weedout_delete_rows() |
4398 | { |
4399 | DBUG_ENTER("SJ_TMP_TABLE::sj_weedout_delete_rows" ); |
4400 | if (tmp_table) |
4401 | { |
4402 | int rc= tmp_table->file->ha_delete_all_rows(); |
4403 | DBUG_RETURN(rc); |
4404 | } |
4405 | have_degenerate_row= FALSE; |
4406 | DBUG_RETURN(0); |
4407 | } |
4408 | |
4409 | |
4410 | /* |
4411 | SemiJoinDuplicateElimination: Weed out duplicate row combinations |
4412 | |
4413 | SYNPOSIS |
4414 | sj_weedout_check_row() |
4415 | thd Thread handle |
4416 | |
4417 | DESCRIPTION |
4418 | Try storing current record combination of outer tables (i.e. their |
4419 | rowids) in the temporary table. This records the fact that we've seen |
4420 | this record combination and also tells us if we've seen it before. |
4421 | |
4422 | RETURN |
4423 | -1 Error |
4424 | 1 The row combination is a duplicate (discard it) |
4425 | 0 The row combination is not a duplicate (continue) |
4426 | */ |
4427 | |
4428 | int SJ_TMP_TABLE::sj_weedout_check_row(THD *thd) |
4429 | { |
4430 | int error; |
4431 | SJ_TMP_TABLE::TAB *tab= tabs; |
4432 | SJ_TMP_TABLE::TAB *tab_end= tabs_end; |
4433 | uchar *ptr; |
4434 | uchar *nulls_ptr; |
4435 | |
4436 | DBUG_ENTER("SJ_TMP_TABLE::sj_weedout_check_row" ); |
4437 | |
4438 | if (is_degenerate) |
4439 | { |
4440 | if (have_degenerate_row) |
4441 | DBUG_RETURN(1); |
4442 | |
4443 | have_degenerate_row= TRUE; |
4444 | DBUG_RETURN(0); |
4445 | } |
4446 | |
4447 | ptr= tmp_table->record[0] + 1; |
4448 | |
4449 | /* Put the the rowids tuple into table->record[0]: */ |
4450 | |
4451 | // 1. Store the length |
4452 | if (((Field_varstring*)(tmp_table->field[0]))->length_bytes == 1) |
4453 | { |
4454 | *ptr= (uchar)(rowid_len + null_bytes); |
4455 | ptr++; |
4456 | } |
4457 | else |
4458 | { |
4459 | int2store(ptr, rowid_len + null_bytes); |
4460 | ptr += 2; |
4461 | } |
4462 | |
4463 | nulls_ptr= ptr; |
4464 | // 2. Zero the null bytes |
4465 | if (null_bytes) |
4466 | { |
4467 | bzero(ptr, null_bytes); |
4468 | ptr += null_bytes; |
4469 | } |
4470 | |
4471 | // 3. Put the rowids |
4472 | for (uint i=0; tab != tab_end; tab++, i++) |
4473 | { |
4474 | handler *h= tab->join_tab->table->file; |
4475 | if (tab->join_tab->table->maybe_null && tab->join_tab->table->null_row) |
4476 | { |
4477 | /* It's a NULL-complemented row */ |
4478 | *(nulls_ptr + tab->null_byte) |= tab->null_bit; |
4479 | bzero(ptr + tab->rowid_offset, h->ref_length); |
4480 | } |
4481 | else |
4482 | { |
4483 | /* Copy the rowid value */ |
4484 | memcpy(ptr + tab->rowid_offset, h->ref, h->ref_length); |
4485 | } |
4486 | } |
4487 | |
4488 | error= tmp_table->file->ha_write_tmp_row(tmp_table->record[0]); |
4489 | if (unlikely(error)) |
4490 | { |
4491 | /* create_internal_tmp_table_from_heap will generate error if needed */ |
4492 | if (!tmp_table->file->is_fatal_error(error, HA_CHECK_DUP)) |
4493 | DBUG_RETURN(1); /* Duplicate */ |
4494 | |
4495 | bool is_duplicate; |
4496 | if (create_internal_tmp_table_from_heap(thd, tmp_table, start_recinfo, |
4497 | &recinfo, error, 1, &is_duplicate)) |
4498 | DBUG_RETURN(-1); |
4499 | if (is_duplicate) |
4500 | DBUG_RETURN(1); |
4501 | } |
4502 | DBUG_RETURN(0); |
4503 | } |
4504 | |
4505 | |
4506 | int init_dups_weedout(JOIN *join, uint first_table, int first_fanout_table, uint n_tables) |
4507 | { |
4508 | THD *thd= join->thd; |
4509 | DBUG_ENTER("init_dups_weedout" ); |
4510 | SJ_TMP_TABLE::TAB sjtabs[MAX_TABLES]; |
4511 | SJ_TMP_TABLE::TAB *last_tab= sjtabs; |
4512 | uint jt_rowid_offset= 0; // # tuple bytes are already occupied (w/o NULL bytes) |
4513 | uint jt_null_bits= 0; // # null bits in tuple bytes |
4514 | /* |
4515 | Walk through the range and remember |
4516 | - tables that need their rowids to be put into temptable |
4517 | - the last outer table |
4518 | */ |
4519 | for (JOIN_TAB *j=join->join_tab + first_table; |
4520 | j < join->join_tab + first_table + n_tables; j++) |
4521 | { |
4522 | if (sj_table_is_included(join, j)) |
4523 | { |
4524 | last_tab->join_tab= j; |
4525 | last_tab->rowid_offset= jt_rowid_offset; |
4526 | jt_rowid_offset += j->table->file->ref_length; |
4527 | if (j->table->maybe_null) |
4528 | { |
4529 | last_tab->null_byte= jt_null_bits / 8; |
4530 | last_tab->null_bit= jt_null_bits++; |
4531 | } |
4532 | last_tab++; |
4533 | j->table->prepare_for_position(); |
4534 | j->keep_current_rowid= TRUE; |
4535 | } |
4536 | } |
4537 | |
4538 | SJ_TMP_TABLE *sjtbl; |
4539 | if (jt_rowid_offset) /* Temptable has at least one rowid */ |
4540 | { |
4541 | size_t tabs_size= (last_tab - sjtabs) * sizeof(SJ_TMP_TABLE::TAB); |
4542 | if (!(sjtbl= (SJ_TMP_TABLE*)thd->alloc(sizeof(SJ_TMP_TABLE))) || |
4543 | !(sjtbl->tabs= (SJ_TMP_TABLE::TAB*) thd->alloc(tabs_size))) |
4544 | DBUG_RETURN(TRUE); /* purecov: inspected */ |
4545 | memcpy(sjtbl->tabs, sjtabs, tabs_size); |
4546 | sjtbl->is_degenerate= FALSE; |
4547 | sjtbl->tabs_end= sjtbl->tabs + (last_tab - sjtabs); |
4548 | sjtbl->rowid_len= jt_rowid_offset; |
4549 | sjtbl->null_bits= jt_null_bits; |
4550 | sjtbl->null_bytes= (jt_null_bits + 7)/8; |
4551 | if (sjtbl->create_sj_weedout_tmp_table(thd)) |
4552 | DBUG_RETURN(TRUE); |
4553 | join->sj_tmp_tables.push_back(sjtbl->tmp_table, thd->mem_root); |
4554 | } |
4555 | else |
4556 | { |
4557 | /* |
4558 | This is a special case where the entire subquery predicate does |
4559 | not depend on anything at all, ie this is |
4560 | WHERE const IN (uncorrelated select) |
4561 | */ |
4562 | if (!(sjtbl= (SJ_TMP_TABLE*)thd->alloc(sizeof(SJ_TMP_TABLE)))) |
4563 | DBUG_RETURN(TRUE); /* purecov: inspected */ |
4564 | sjtbl->tmp_table= NULL; |
4565 | sjtbl->is_degenerate= TRUE; |
4566 | sjtbl->have_degenerate_row= FALSE; |
4567 | } |
4568 | |
4569 | sjtbl->next_flush_table= join->join_tab[first_table].flush_weedout_table; |
4570 | join->join_tab[first_table].flush_weedout_table= sjtbl; |
4571 | join->join_tab[first_fanout_table].first_weedout_table= sjtbl; |
4572 | join->join_tab[first_table + n_tables - 1].check_weed_out_table= sjtbl; |
4573 | DBUG_RETURN(0); |
4574 | } |
4575 | |
4576 | |
4577 | /* |
4578 | @brief |
4579 | Set up semi-join Loose Scan strategy for execution |
4580 | |
4581 | @detail |
4582 | Other strategies are done in setup_semijoin_dups_elimination(), |
4583 | however, we need to set up Loose Scan earlier, before make_join_select is |
4584 | called. This is to prevent make_join_select() from switching full index |
4585 | scans into quick selects (which will break Loose Scan access). |
4586 | |
4587 | @return |
4588 | 0 OK |
4589 | 1 Error |
4590 | */ |
4591 | |
4592 | int setup_semijoin_loosescan(JOIN *join) |
4593 | { |
4594 | uint i; |
4595 | DBUG_ENTER("setup_semijoin_loosescan" ); |
4596 | |
4597 | POSITION *pos= join->best_positions + join->const_tables; |
4598 | for (i= join->const_tables ; i < join->top_join_tab_count; ) |
4599 | { |
4600 | JOIN_TAB *tab=join->join_tab + i; |
4601 | switch (pos->sj_strategy) { |
4602 | case SJ_OPT_MATERIALIZE: |
4603 | case SJ_OPT_MATERIALIZE_SCAN: |
4604 | i+= 1; /* join tabs are embedded in the nest */ |
4605 | pos += pos->n_sj_tables; |
4606 | break; |
4607 | case SJ_OPT_LOOSE_SCAN: |
4608 | { |
4609 | /* We jump from the last table to the first one */ |
4610 | tab->loosescan_match_tab= tab + pos->n_sj_tables - 1; |
4611 | |
4612 | /* LooseScan requires records to be produced in order */ |
4613 | if (tab->select && tab->select->quick) |
4614 | tab->select->quick->need_sorted_output(); |
4615 | |
4616 | for (uint j= i; j < i + pos->n_sj_tables; j++) |
4617 | join->join_tab[j].inside_loosescan_range= TRUE; |
4618 | |
4619 | /* Calculate key length */ |
4620 | uint keylen= 0; |
4621 | uint keyno= pos->loosescan_picker.loosescan_key; |
4622 | for (uint kp=0; kp < pos->loosescan_picker.loosescan_parts; kp++) |
4623 | keylen += tab->table->key_info[keyno].key_part[kp].store_length; |
4624 | |
4625 | tab->loosescan_key= keyno; |
4626 | tab->loosescan_key_len= keylen; |
4627 | if (pos->n_sj_tables > 1) |
4628 | tab[pos->n_sj_tables - 1].do_firstmatch= tab; |
4629 | i+= pos->n_sj_tables; |
4630 | pos+= pos->n_sj_tables; |
4631 | break; |
4632 | } |
4633 | default: |
4634 | { |
4635 | i++; |
4636 | pos++; |
4637 | break; |
4638 | } |
4639 | } |
4640 | } |
4641 | DBUG_RETURN(FALSE); |
4642 | } |
4643 | |
4644 | |
4645 | /* |
4646 | Setup the strategies to eliminate semi-join duplicates. |
4647 | |
4648 | SYNOPSIS |
4649 | setup_semijoin_dups_elimination() |
4650 | join Join to process |
4651 | options Join options (needed to see if join buffering will be |
4652 | used or not) |
4653 | no_jbuf_after Another bit of information re where join buffering will |
4654 | be used. |
4655 | |
4656 | DESCRIPTION |
4657 | Setup the strategies to eliminate semi-join duplicates. ATM there are 4 |
4658 | strategies: |
4659 | |
4660 | 1. DuplicateWeedout (use of temptable to remove duplicates based on rowids |
4661 | of row combinations) |
4662 | 2. FirstMatch (pick only the 1st matching row combination of inner tables) |
4663 | 3. LooseScan (scanning the sj-inner table in a way that groups duplicates |
4664 | together and picking the 1st one) |
4665 | 4. SJ-Materialization. |
4666 | |
4667 | The join order has "duplicate-generating ranges", and every range is |
4668 | served by one strategy or a combination of FirstMatch with with some |
4669 | other strategy. |
4670 | |
4671 | "Duplicate-generating range" is defined as a range within the join order |
4672 | that contains all of the inner tables of a semi-join. All ranges must be |
4673 | disjoint, if tables of several semi-joins are interleaved, then the ranges |
4674 | are joined together, which is equivalent to converting |
4675 | SELECT ... WHERE oe1 IN (SELECT ie1 ...) AND oe2 IN (SELECT ie2 ) |
4676 | to |
4677 | SELECT ... WHERE (oe1, oe2) IN (SELECT ie1, ie2 ... ...) |
4678 | . |
4679 | |
4680 | Applicability conditions are as follows: |
4681 | |
4682 | DuplicateWeedout strategy |
4683 | ~~~~~~~~~~~~~~~~~~~~~~~~~ |
4684 | |
4685 | (ot|nt)* [ it ((it|ot|nt)* (it|ot))] (nt)* |
4686 | +------+ +=========================+ +---+ |
4687 | (1) (2) (3) |
4688 | |
4689 | (1) - Prefix of OuterTables (those that participate in |
4690 | IN-equality and/or are correlated with subquery) and outer |
4691 | Non-correlated tables. |
4692 | (2) - The handled range. The range starts with the first sj-inner |
4693 | table, and covers all sj-inner and outer tables |
4694 | Within the range, Inner, Outer, outer non-correlated tables |
4695 | may follow in any order. |
4696 | (3) - The suffix of outer non-correlated tables. |
4697 | |
4698 | FirstMatch strategy |
4699 | ~~~~~~~~~~~~~~~~~~~ |
4700 | |
4701 | (ot|nt)* [ it ((it|nt)* it) ] (nt)* |
4702 | +------+ +==================+ +---+ |
4703 | (1) (2) (3) |
4704 | |
4705 | (1) - Prefix of outer and non-correlated tables |
4706 | (2) - The handled range, which may contain only inner and |
4707 | non-correlated tables. |
4708 | (3) - The suffix of outer non-correlated tables. |
4709 | |
4710 | LooseScan strategy |
4711 | ~~~~~~~~~~~~~~~~~~ |
4712 | |
4713 | (ot|ct|nt) [ loosescan_tbl (ot|nt|it)* it ] (ot|nt)* |
4714 | +--------+ +===========+ +=============+ +------+ |
4715 | (1) (2) (3) (4) |
4716 | |
4717 | (1) - Prefix that may contain any outer tables. The prefix must contain |
4718 | all the non-trivially correlated outer tables. (non-trivially means |
4719 | that the correlation is not just through the IN-equality). |
4720 | |
4721 | (2) - Inner table for which the LooseScan scan is performed. |
4722 | |
4723 | (3) - The remainder of the duplicate-generating range. It is served by |
4724 | application of FirstMatch strategy, with the exception that |
4725 | outer IN-correlated tables are considered to be non-correlated. |
4726 | |
4727 | (4) - THe suffix of outer and outer non-correlated tables. |
4728 | |
4729 | |
4730 | The choice between the strategies is made by the join optimizer (see |
4731 | advance_sj_state() and fix_semijoin_strategies_for_picked_join_order()). |
4732 | This function sets up all fields/structures/etc needed for execution except |
4733 | for setup/initialization of semi-join materialization which is done in |
4734 | setup_sj_materialization() (todo: can't we move that to here also?) |
4735 | |
4736 | RETURN |
4737 | FALSE OK |
4738 | TRUE Out of memory error |
4739 | */ |
4740 | |
4741 | int setup_semijoin_dups_elimination(JOIN *join, ulonglong options, |
4742 | uint no_jbuf_after) |
4743 | { |
4744 | uint i; |
4745 | DBUG_ENTER("setup_semijoin_dups_elimination" ); |
4746 | |
4747 | join->complex_firstmatch_tables= table_map(0); |
4748 | |
4749 | POSITION *pos= join->best_positions + join->const_tables; |
4750 | for (i= join->const_tables ; i < join->top_join_tab_count; ) |
4751 | { |
4752 | JOIN_TAB *tab=join->join_tab + i; |
4753 | switch (pos->sj_strategy) { |
4754 | case SJ_OPT_MATERIALIZE: |
4755 | case SJ_OPT_MATERIALIZE_SCAN: |
4756 | /* Do nothing */ |
4757 | i+= 1;// It used to be pos->n_sj_tables, but now they are embedded in a nest |
4758 | pos += pos->n_sj_tables; |
4759 | break; |
4760 | case SJ_OPT_LOOSE_SCAN: |
4761 | { |
4762 | /* Setup already handled by setup_semijoin_loosescan */ |
4763 | i+= pos->n_sj_tables; |
4764 | pos+= pos->n_sj_tables; |
4765 | break; |
4766 | } |
4767 | case SJ_OPT_DUPS_WEEDOUT: |
4768 | { |
4769 | /* |
4770 | Check for join buffering. If there is one, move the first table |
4771 | forwards, but do not destroy other duplicate elimination methods. |
4772 | */ |
4773 | uint first_table= i; |
4774 | |
4775 | uint join_cache_level= join->thd->variables.join_cache_level; |
4776 | for (uint j= i; j < i + pos->n_sj_tables; j++) |
4777 | { |
4778 | /* |
4779 | When we'll properly take join buffering into account during |
4780 | join optimization, the below check should be changed to |
4781 | "if (join->best_positions[j].use_join_buffer && |
4782 | j <= no_jbuf_after)". |
4783 | For now, use a rough criteria: |
4784 | */ |
4785 | JOIN_TAB *js_tab=join->join_tab + j; |
4786 | if (j != join->const_tables && js_tab->use_quick != 2 && |
4787 | j <= no_jbuf_after && |
4788 | ((js_tab->type == JT_ALL && join_cache_level != 0) || |
4789 | (join_cache_level > 2 && (js_tab->type == JT_REF || |
4790 | js_tab->type == JT_EQ_REF)))) |
4791 | { |
4792 | /* Looks like we'll be using join buffer */ |
4793 | first_table= join->const_tables; |
4794 | /* |
4795 | Make sure that possible sorting of rows from the head table |
4796 | is not to be employed. |
4797 | */ |
4798 | if (join->get_sort_by_join_tab()) |
4799 | { |
4800 | join->simple_order= 0; |
4801 | join->simple_group= 0; |
4802 | join->need_tmp= join->test_if_need_tmp_table(); |
4803 | } |
4804 | break; |
4805 | } |
4806 | } |
4807 | |
4808 | init_dups_weedout(join, first_table, i, i + pos->n_sj_tables - first_table); |
4809 | i+= pos->n_sj_tables; |
4810 | pos+= pos->n_sj_tables; |
4811 | break; |
4812 | } |
4813 | case SJ_OPT_FIRST_MATCH: |
4814 | { |
4815 | JOIN_TAB *j; |
4816 | JOIN_TAB *jump_to= tab-1; |
4817 | |
4818 | bool complex_range= FALSE; |
4819 | table_map tables_in_range= table_map(0); |
4820 | |
4821 | for (j= tab; j != tab + pos->n_sj_tables; j++) |
4822 | { |
4823 | tables_in_range |= j->table->map; |
4824 | if (!j->emb_sj_nest) |
4825 | { |
4826 | /* |
4827 | Got a table that's not within any semi-join nest. This is a case |
4828 | like this: |
4829 | |
4830 | SELECT * FROM ot1, nt1 WHERE ot1.col IN (SELECT expr FROM it1, it2) |
4831 | |
4832 | with a join order of |
4833 | |
4834 | +----- FirstMatch range ----+ |
4835 | | | |
4836 | ot1 it1 nt1 nt2 it2 it3 ... |
4837 | | ^ |
4838 | | +-------- 'j' points here |
4839 | +------------- SJ_OPT_FIRST_MATCH was set for this table as |
4840 | it's the first one that produces duplicates |
4841 | |
4842 | */ |
4843 | DBUG_ASSERT(j != tab); /* table ntX must have an itX before it */ |
4844 | |
4845 | /* |
4846 | If the table right before us is an inner table (like it1 in the |
4847 | picture), it should be set to jump back to previous outer-table |
4848 | */ |
4849 | if (j[-1].emb_sj_nest) |
4850 | j[-1].do_firstmatch= jump_to; |
4851 | |
4852 | jump_to= j; /* Jump back to us */ |
4853 | complex_range= TRUE; |
4854 | } |
4855 | else |
4856 | { |
4857 | j->first_sj_inner_tab= tab; |
4858 | j->last_sj_inner_tab= tab + pos->n_sj_tables - 1; |
4859 | } |
4860 | } |
4861 | j[-1].do_firstmatch= jump_to; |
4862 | i+= pos->n_sj_tables; |
4863 | pos+= pos->n_sj_tables; |
4864 | |
4865 | if (complex_range) |
4866 | join->complex_firstmatch_tables|= tables_in_range; |
4867 | break; |
4868 | } |
4869 | case SJ_OPT_NONE: |
4870 | i++; |
4871 | pos++; |
4872 | break; |
4873 | } |
4874 | } |
4875 | DBUG_RETURN(FALSE); |
4876 | } |
4877 | |
4878 | |
4879 | /* |
4880 | Destroy all temporary tables created by NL-semijoin runtime |
4881 | */ |
4882 | |
4883 | void destroy_sj_tmp_tables(JOIN *join) |
4884 | { |
4885 | List_iterator<TABLE> it(join->sj_tmp_tables); |
4886 | TABLE *table; |
4887 | while ((table= it++)) |
4888 | { |
4889 | /* |
4890 | SJ-Materialization tables are initialized for either sequential reading |
4891 | or index lookup, DuplicateWeedout tables are not initialized for read |
4892 | (we only write to them), so need to call ha_index_or_rnd_end. |
4893 | */ |
4894 | table->file->ha_index_or_rnd_end(); |
4895 | free_tmp_table(join->thd, table); |
4896 | } |
4897 | join->sj_tmp_tables.empty(); |
4898 | join->sjm_info_list.empty(); |
4899 | } |
4900 | |
4901 | |
4902 | /* |
4903 | Remove all records from all temp tables used by NL-semijoin runtime |
4904 | |
4905 | SYNOPSIS |
4906 | clear_sj_tmp_tables() |
4907 | join The join to remove tables for |
4908 | |
4909 | DESCRIPTION |
4910 | Remove all records from all temp tables used by NL-semijoin runtime. This |
4911 | must be done before every join re-execution. |
4912 | */ |
4913 | |
4914 | int clear_sj_tmp_tables(JOIN *join) |
4915 | { |
4916 | int res; |
4917 | List_iterator<TABLE> it(join->sj_tmp_tables); |
4918 | TABLE *table; |
4919 | while ((table= it++)) |
4920 | { |
4921 | if ((res= table->file->ha_delete_all_rows())) |
4922 | return res; /* purecov: inspected */ |
4923 | } |
4924 | |
4925 | SJ_MATERIALIZATION_INFO *sjm; |
4926 | List_iterator<SJ_MATERIALIZATION_INFO> it2(join->sjm_info_list); |
4927 | while ((sjm= it2++)) |
4928 | { |
4929 | sjm->materialized= FALSE; |
4930 | } |
4931 | return 0; |
4932 | } |
4933 | |
4934 | |
4935 | /* |
4936 | Check if the table's rowid is included in the temptable |
4937 | |
4938 | SYNOPSIS |
4939 | sj_table_is_included() |
4940 | join The join |
4941 | join_tab The table to be checked |
4942 | |
4943 | DESCRIPTION |
4944 | SemiJoinDuplicateElimination: check the table's rowid should be included |
4945 | in the temptable. This is so if |
4946 | |
4947 | 1. The table is not embedded within some semi-join nest |
4948 | 2. The has been pulled out of a semi-join nest, or |
4949 | |
4950 | 3. The table is functionally dependent on some previous table |
4951 | |
4952 | [4. This is also true for constant tables that can't be |
4953 | NULL-complemented but this function is not called for such tables] |
4954 | |
4955 | RETURN |
4956 | TRUE - Include table's rowid |
4957 | FALSE - Don't |
4958 | */ |
4959 | |
4960 | static bool sj_table_is_included(JOIN *join, JOIN_TAB *join_tab) |
4961 | { |
4962 | if (join_tab->emb_sj_nest) |
4963 | return FALSE; |
4964 | |
4965 | /* Check if this table is functionally dependent on the tables that |
4966 | are within the same outer join nest |
4967 | */ |
4968 | TABLE_LIST *embedding= join_tab->table->pos_in_table_list->embedding; |
4969 | if (join_tab->type == JT_EQ_REF) |
4970 | { |
4971 | table_map depends_on= 0; |
4972 | uint idx; |
4973 | |
4974 | for (uint kp= 0; kp < join_tab->ref.key_parts; kp++) |
4975 | depends_on |= join_tab->ref.items[kp]->used_tables(); |
4976 | |
4977 | Table_map_iterator it(depends_on & ~PSEUDO_TABLE_BITS); |
4978 | while ((idx= it.next_bit())!=Table_map_iterator::BITMAP_END) |
4979 | { |
4980 | JOIN_TAB *ref_tab= join->map2table[idx]; |
4981 | if (embedding != ref_tab->table->pos_in_table_list->embedding) |
4982 | return TRUE; |
4983 | } |
4984 | /* Ok, functionally dependent */ |
4985 | return FALSE; |
4986 | } |
4987 | /* Not functionally dependent => need to include*/ |
4988 | return TRUE; |
4989 | } |
4990 | |
4991 | |
4992 | /* |
4993 | Index lookup-based subquery: save some flags for EXPLAIN output |
4994 | |
4995 | SYNOPSIS |
4996 | save_index_subquery_explain_info() |
4997 | join_tab Subquery's join tab (there is only one as index lookup is |
4998 | only used for subqueries that are single-table SELECTs) |
4999 | where Subquery's WHERE clause |
5000 | |
5001 | DESCRIPTION |
5002 | For index lookup-based subquery (i.e. one executed with |
5003 | subselect_uniquesubquery_engine or subselect_indexsubquery_engine), |
5004 | check its EXPLAIN output row should contain |
5005 | "Using index" (TAB_INFO_FULL_SCAN_ON_NULL) |
5006 | "Using Where" (TAB_INFO_USING_WHERE) |
5007 | "Full scan on NULL key" (TAB_INFO_FULL_SCAN_ON_NULL) |
5008 | and set appropriate flags in join_tab->packed_info. |
5009 | */ |
5010 | |
5011 | static void save_index_subquery_explain_info(JOIN_TAB *join_tab, Item* where) |
5012 | { |
5013 | join_tab->packed_info= TAB_INFO_HAVE_VALUE; |
5014 | if (join_tab->table->covering_keys.is_set(join_tab->ref.key)) |
5015 | join_tab->packed_info |= TAB_INFO_USING_INDEX; |
5016 | if (where) |
5017 | join_tab->packed_info |= TAB_INFO_USING_WHERE; |
5018 | for (uint i = 0; i < join_tab->ref.key_parts; i++) |
5019 | { |
5020 | if (join_tab->ref.cond_guards[i]) |
5021 | { |
5022 | join_tab->packed_info |= TAB_INFO_FULL_SCAN_ON_NULL; |
5023 | break; |
5024 | } |
5025 | } |
5026 | } |
5027 | |
5028 | |
5029 | /* |
5030 | Check if the join can be rewritten to [unique_]indexsubquery_engine |
5031 | |
5032 | DESCRIPTION |
5033 | Check if the join can be changed into [unique_]indexsubquery_engine. |
5034 | |
5035 | The check is done after join optimization, the idea is that if the join |
5036 | has only one table and uses a [eq_]ref access generated from subselect's |
5037 | IN-equality then we replace it with a subselect_indexsubquery_engine or a |
5038 | subselect_uniquesubquery_engine. |
5039 | |
5040 | RETURN |
5041 | 0 - Ok, rewrite done (stop join optimization and return) |
5042 | 1 - Fatal error (stop join optimization and return) |
5043 | -1 - No rewrite performed, continue with join optimization |
5044 | */ |
5045 | |
5046 | int rewrite_to_index_subquery_engine(JOIN *join) |
5047 | { |
5048 | THD *thd= join->thd; |
5049 | JOIN_TAB* join_tab=join->join_tab; |
5050 | SELECT_LEX_UNIT *unit= join->unit; |
5051 | DBUG_ENTER("rewrite_to_index_subquery_engine" ); |
5052 | |
5053 | /* |
5054 | is this simple IN subquery? |
5055 | */ |
5056 | /* TODO: In order to use these more efficient subquery engines in more cases, |
5057 | the following problems need to be solved: |
5058 | - the code that removes GROUP BY (group_list), also adds an ORDER BY |
5059 | (order), thus GROUP BY queries (almost?) never pass through this branch. |
5060 | Solution: remove the test below '!join->order', because we remove the |
5061 | ORDER clase for subqueries anyway. |
5062 | - in order to set a more efficient engine, the optimizer needs to both |
5063 | decide to remove GROUP BY, *and* select one of the JT_[EQ_]REF[_OR_NULL] |
5064 | access methods, *and* loose scan should be more expensive or |
5065 | inapliccable. When is that possible? |
5066 | - Consider expanding the applicability of this rewrite for loose scan |
5067 | for group by queries. |
5068 | */ |
5069 | if (!join->group_list && !join->order && |
5070 | join->unit->item && |
5071 | join->unit->item->substype() == Item_subselect::IN_SUBS && |
5072 | join->table_count == 1 && join->conds && |
5073 | !join->unit->is_unit_op()) |
5074 | { |
5075 | if (!join->having) |
5076 | { |
5077 | Item *where= join->conds; |
5078 | if (join_tab[0].type == JT_EQ_REF && |
5079 | join_tab[0].ref.items[0]->name.str == in_left_expr_name.str) |
5080 | { |
5081 | remove_subq_pushed_predicates(join, &where); |
5082 | save_index_subquery_explain_info(join_tab, where); |
5083 | join_tab[0].type= JT_UNIQUE_SUBQUERY; |
5084 | join->error= 0; |
5085 | DBUG_RETURN(unit->item-> |
5086 | change_engine(new |
5087 | subselect_uniquesubquery_engine(thd, |
5088 | join_tab, |
5089 | unit->item, |
5090 | where))); |
5091 | } |
5092 | else if (join_tab[0].type == JT_REF && |
5093 | join_tab[0].ref.items[0]->name.str == in_left_expr_name.str) |
5094 | { |
5095 | remove_subq_pushed_predicates(join, &where); |
5096 | save_index_subquery_explain_info(join_tab, where); |
5097 | join_tab[0].type= JT_INDEX_SUBQUERY; |
5098 | join->error= 0; |
5099 | DBUG_RETURN(unit->item-> |
5100 | change_engine(new |
5101 | subselect_indexsubquery_engine(thd, |
5102 | join_tab, |
5103 | unit->item, |
5104 | where, |
5105 | NULL, |
5106 | 0))); |
5107 | } |
5108 | } else if (join_tab[0].type == JT_REF_OR_NULL && |
5109 | join_tab[0].ref.items[0]->name.str == in_left_expr_name.str && |
5110 | join->having->name.str == in_having_cond.str) |
5111 | { |
5112 | join_tab[0].type= JT_INDEX_SUBQUERY; |
5113 | join->error= 0; |
5114 | join->conds= remove_additional_cond(join->conds); |
5115 | save_index_subquery_explain_info(join_tab, join->conds); |
5116 | DBUG_RETURN(unit->item-> |
5117 | change_engine(new subselect_indexsubquery_engine(thd, |
5118 | join_tab, |
5119 | unit->item, |
5120 | join->conds, |
5121 | join->having, |
5122 | 1))); |
5123 | } |
5124 | } |
5125 | |
5126 | DBUG_RETURN(-1); /* Haven't done the rewrite */ |
5127 | } |
5128 | |
5129 | |
5130 | /** |
5131 | Remove additional condition inserted by IN/ALL/ANY transformation. |
5132 | |
5133 | @param conds condition for processing |
5134 | |
5135 | @return |
5136 | new conditions |
5137 | */ |
5138 | |
5139 | static Item *remove_additional_cond(Item* conds) |
5140 | { |
5141 | if (conds->name.str == in_additional_cond.str) |
5142 | return 0; |
5143 | if (conds->type() == Item::COND_ITEM) |
5144 | { |
5145 | Item_cond *cnd= (Item_cond*) conds; |
5146 | List_iterator<Item> li(*(cnd->argument_list())); |
5147 | Item *item; |
5148 | while ((item= li++)) |
5149 | { |
5150 | if (item->name.str == in_additional_cond.str) |
5151 | { |
5152 | li.remove(); |
5153 | if (cnd->argument_list()->elements == 1) |
5154 | return cnd->argument_list()->head(); |
5155 | return conds; |
5156 | } |
5157 | } |
5158 | } |
5159 | return conds; |
5160 | } |
5161 | |
5162 | |
5163 | /* |
5164 | Remove the predicates pushed down into the subquery |
5165 | |
5166 | SYNOPSIS |
5167 | remove_subq_pushed_predicates() |
5168 | where IN Must be NULL |
5169 | OUT The remaining WHERE condition, or NULL |
5170 | |
5171 | DESCRIPTION |
5172 | Given that this join will be executed using (unique|index)_subquery, |
5173 | without "checking NULL", remove the predicates that were pushed down |
5174 | into the subquery. |
5175 | |
5176 | If the subquery compares scalar values, we can remove the condition that |
5177 | was wrapped into trig_cond (it will be checked when needed by the subquery |
5178 | engine) |
5179 | |
5180 | If the subquery compares row values, we need to keep the wrapped |
5181 | equalities in the WHERE clause: when the left (outer) tuple has both NULL |
5182 | and non-NULL values, we'll do a full table scan and will rely on the |
5183 | equalities corresponding to non-NULL parts of left tuple to filter out |
5184 | non-matching records. |
5185 | |
5186 | TODO: We can remove the equalities that will be guaranteed to be true by the |
5187 | fact that subquery engine will be using index lookup. This must be done only |
5188 | for cases where there are no conversion errors of significance, e.g. 257 |
5189 | that is searched in a byte. But this requires homogenization of the return |
5190 | codes of all Field*::store() methods. |
5191 | */ |
5192 | |
5193 | static void remove_subq_pushed_predicates(JOIN *join, Item **where) |
5194 | { |
5195 | if (join->conds->type() == Item::FUNC_ITEM && |
5196 | ((Item_func *)join->conds)->functype() == Item_func::EQ_FUNC && |
5197 | ((Item_func *)join->conds)->arguments()[0]->type() == Item::REF_ITEM && |
5198 | ((Item_func *)join->conds)->arguments()[1]->type() == Item::FIELD_ITEM && |
5199 | test_if_ref (join->conds, |
5200 | (Item_field *)((Item_func *)join->conds)->arguments()[1], |
5201 | ((Item_func *)join->conds)->arguments()[0])) |
5202 | { |
5203 | *where= 0; |
5204 | return; |
5205 | } |
5206 | } |
5207 | |
5208 | |
5209 | |
5210 | |
5211 | /** |
5212 | Optimize all subqueries of a query that were not flattened into a semijoin. |
5213 | |
5214 | @details |
5215 | Optimize all immediate children subqueries of a query. |
5216 | |
5217 | This phase must be called after substitute_for_best_equal_field() because |
5218 | that function may replace items with other items from a multiple equality, |
5219 | and we need to reference the correct items in the index access method of the |
5220 | IN predicate. |
5221 | |
5222 | @return Operation status |
5223 | @retval FALSE success. |
5224 | @retval TRUE error occurred. |
5225 | */ |
5226 | |
5227 | bool JOIN::optimize_unflattened_subqueries() |
5228 | { |
5229 | return select_lex->optimize_unflattened_subqueries(false); |
5230 | } |
5231 | |
5232 | /** |
5233 | Optimize all constant subqueries of a query that were not flattened into |
5234 | a semijoin. |
5235 | |
5236 | @details |
5237 | Similar to other constant conditions, constant subqueries can be used in |
5238 | various constant optimizations. Having optimized constant subqueries before |
5239 | these constant optimizations, makes it possible to estimate if a subquery |
5240 | is "cheap" enough to be executed during the optimization phase. |
5241 | |
5242 | Constant subqueries can be optimized and evaluated independent of the outer |
5243 | query, therefore if const_only = true, this method can be called early in |
5244 | the optimization phase of the outer query. |
5245 | |
5246 | @return Operation status |
5247 | @retval FALSE success. |
5248 | @retval TRUE error occurred. |
5249 | */ |
5250 | |
5251 | bool JOIN::optimize_constant_subqueries() |
5252 | { |
5253 | ulonglong save_options= select_lex->options; |
5254 | bool res; |
5255 | /* |
5256 | Constant subqueries may be executed during the optimization phase. |
5257 | In EXPLAIN mode the optimizer doesn't initialize many of the data structures |
5258 | needed for execution. In order to make it possible to execute subqueries |
5259 | during optimization, constant subqueries must be optimized for execution, |
5260 | not for EXPLAIN. |
5261 | */ |
5262 | select_lex->options&= ~SELECT_DESCRIBE; |
5263 | res= select_lex->optimize_unflattened_subqueries(true); |
5264 | select_lex->options= save_options; |
5265 | return res; |
5266 | } |
5267 | |
5268 | |
5269 | /* |
5270 | Join tab execution startup function. |
5271 | |
5272 | SYNOPSIS |
5273 | join_tab_execution_startup() |
5274 | tab Join tab to perform startup actions for |
5275 | |
5276 | DESCRIPTION |
5277 | Join tab execution startup function. This is different from |
5278 | tab->read_first_record in the regard that this has actions that are to be |
5279 | done once per join execution. |
5280 | |
5281 | Currently there are only two possible startup functions, so we have them |
5282 | both here inside if (...) branches. In future we could switch to function |
5283 | pointers. |
5284 | |
5285 | TODO: consider moving this together with JOIN_TAB::preread_init |
5286 | |
5287 | RETURN |
5288 | NESTED_LOOP_OK - OK |
5289 | NESTED_LOOP_ERROR| NESTED_LOOP_KILLED - Error, abort the join execution |
5290 | */ |
5291 | |
5292 | enum_nested_loop_state join_tab_execution_startup(JOIN_TAB *tab) |
5293 | { |
5294 | Item_in_subselect *in_subs; |
5295 | DBUG_ENTER("join_tab_execution_startup" ); |
5296 | |
5297 | if (tab->table->pos_in_table_list && |
5298 | (in_subs= tab->table->pos_in_table_list->jtbm_subselect)) |
5299 | { |
5300 | /* It's a non-merged SJM nest */ |
5301 | DBUG_ASSERT(in_subs->engine->engine_type() == |
5302 | subselect_engine::HASH_SJ_ENGINE); |
5303 | subselect_hash_sj_engine *hash_sj_engine= |
5304 | ((subselect_hash_sj_engine*)in_subs->engine); |
5305 | if (!hash_sj_engine->is_materialized) |
5306 | { |
5307 | hash_sj_engine->materialize_join->exec(); |
5308 | hash_sj_engine->is_materialized= TRUE; |
5309 | |
5310 | if (unlikely(hash_sj_engine->materialize_join->error) || |
5311 | unlikely(tab->join->thd->is_fatal_error)) |
5312 | DBUG_RETURN(NESTED_LOOP_ERROR); |
5313 | } |
5314 | } |
5315 | else if (tab->bush_children) |
5316 | { |
5317 | /* It's a merged SJM nest */ |
5318 | enum_nested_loop_state rc; |
5319 | SJ_MATERIALIZATION_INFO *sjm= tab->bush_children->start->emb_sj_nest->sj_mat_info; |
5320 | |
5321 | if (!sjm->materialized) |
5322 | { |
5323 | JOIN *join= tab->join; |
5324 | JOIN_TAB *join_tab= tab->bush_children->start; |
5325 | JOIN_TAB *save_return_tab= join->return_tab; |
5326 | /* |
5327 | Now run the join for the inner tables. The first call is to run the |
5328 | join, the second one is to signal EOF (this is essential for some |
5329 | join strategies, e.g. it will make join buffering flush the records) |
5330 | */ |
5331 | if ((rc= sub_select(join, join_tab, FALSE/* no EOF */)) < 0 || |
5332 | (rc= sub_select(join, join_tab, TRUE/* now EOF */)) < 0) |
5333 | { |
5334 | join->return_tab= save_return_tab; |
5335 | DBUG_RETURN(rc); /* it's NESTED_LOOP_(ERROR|KILLED)*/ |
5336 | } |
5337 | join->return_tab= save_return_tab; |
5338 | sjm->materialized= TRUE; |
5339 | } |
5340 | } |
5341 | |
5342 | DBUG_RETURN(NESTED_LOOP_OK); |
5343 | } |
5344 | |
5345 | |
5346 | /* |
5347 | Create a dummy temporary table, useful only for the sake of having a |
5348 | TABLE* object with map,tablenr and maybe_null properties. |
5349 | |
5350 | This is used by non-mergeable semi-join materilization code to handle |
5351 | degenerate cases where materialized subquery produced "Impossible WHERE" |
5352 | and thus wasn't materialized. |
5353 | */ |
5354 | |
5355 | TABLE *create_dummy_tmp_table(THD *thd) |
5356 | { |
5357 | DBUG_ENTER("create_dummy_tmp_table" ); |
5358 | TABLE *table; |
5359 | TMP_TABLE_PARAM sjm_table_param; |
5360 | sjm_table_param.init(); |
5361 | sjm_table_param.field_count= 1; |
5362 | List<Item> sjm_table_cols; |
5363 | const LEX_CSTRING dummy_name= { STRING_WITH_LEN("dummy" ) }; |
5364 | Item *column_item= new (thd->mem_root) Item_int(thd, 1); |
5365 | if (!column_item) |
5366 | DBUG_RETURN(NULL); |
5367 | |
5368 | sjm_table_cols.push_back(column_item, thd->mem_root); |
5369 | if (!(table= create_tmp_table(thd, &sjm_table_param, |
5370 | sjm_table_cols, (ORDER*) 0, |
5371 | TRUE /* distinct */, |
5372 | 1, /*save_sum_fields*/ |
5373 | thd->variables.option_bits | |
5374 | TMP_TABLE_ALL_COLUMNS, |
5375 | HA_POS_ERROR /*rows_limit */, |
5376 | &dummy_name, TRUE /* Do not open */))) |
5377 | { |
5378 | DBUG_RETURN(NULL); |
5379 | } |
5380 | DBUG_RETURN(table); |
5381 | } |
5382 | |
5383 | |
5384 | /* |
5385 | A class that is used to catch one single tuple that is sent to the join |
5386 | output, and save it in Item_cache element(s). |
5387 | |
5388 | It is very similar to select_singlerow_subselect but doesn't require a |
5389 | Item_singlerow_subselect item. |
5390 | */ |
5391 | |
5392 | class select_value_catcher :public select_subselect |
5393 | { |
5394 | public: |
5395 | select_value_catcher(THD *thd_arg, Item_subselect *item_arg): |
5396 | select_subselect(thd_arg, item_arg) |
5397 | {} |
5398 | int send_data(List<Item> &items); |
5399 | int setup(List<Item> *items); |
5400 | bool assigned; /* TRUE <=> we've caught a value */ |
5401 | uint n_elements; /* How many elements we get */ |
5402 | Item_cache **row; /* Array of cache elements */ |
5403 | }; |
5404 | |
5405 | |
5406 | int select_value_catcher::setup(List<Item> *items) |
5407 | { |
5408 | assigned= FALSE; |
5409 | n_elements= items->elements; |
5410 | |
5411 | if (!(row= (Item_cache**) thd->alloc(sizeof(Item_cache*) * n_elements))) |
5412 | return TRUE; |
5413 | |
5414 | Item *sel_item; |
5415 | List_iterator<Item> li(*items); |
5416 | for (uint i= 0; (sel_item= li++); i++) |
5417 | { |
5418 | if (!(row[i]= sel_item->get_cache(thd))) |
5419 | return TRUE; |
5420 | row[i]->setup(thd, sel_item); |
5421 | } |
5422 | return FALSE; |
5423 | } |
5424 | |
5425 | |
5426 | int select_value_catcher::send_data(List<Item> &items) |
5427 | { |
5428 | DBUG_ENTER("select_value_catcher::send_data" ); |
5429 | DBUG_ASSERT(!assigned); |
5430 | DBUG_ASSERT(items.elements == n_elements); |
5431 | |
5432 | if (unit->offset_limit_cnt) |
5433 | { // Using limit offset,count |
5434 | unit->offset_limit_cnt--; |
5435 | DBUG_RETURN(0); |
5436 | } |
5437 | |
5438 | Item *val_item; |
5439 | List_iterator_fast<Item> li(items); |
5440 | for (uint i= 0; (val_item= li++); i++) |
5441 | { |
5442 | row[i]->store(val_item); |
5443 | row[i]->cache_value(); |
5444 | } |
5445 | assigned= TRUE; |
5446 | DBUG_RETURN(0); |
5447 | } |
5448 | |
5449 | |
5450 | /* |
5451 | Setup JTBM join tabs for execution |
5452 | */ |
5453 | |
5454 | bool setup_jtbm_semi_joins(JOIN *join, List<TABLE_LIST> *join_list, |
5455 | Item **join_where) |
5456 | { |
5457 | TABLE_LIST *table; |
5458 | NESTED_JOIN *nested_join; |
5459 | List_iterator<TABLE_LIST> li(*join_list); |
5460 | THD *thd= join->thd; |
5461 | DBUG_ENTER("setup_jtbm_semi_joins" ); |
5462 | |
5463 | while ((table= li++)) |
5464 | { |
5465 | Item_in_subselect *item; |
5466 | |
5467 | if ((item= table->jtbm_subselect)) |
5468 | { |
5469 | Item_in_subselect *subq_pred= item; |
5470 | double rows; |
5471 | double read_time; |
5472 | |
5473 | /* |
5474 | Perform optimization of the subquery, so that we know estmated |
5475 | - cost of materialization process |
5476 | - how many records will be in the materialized temp.table |
5477 | */ |
5478 | if (subq_pred->optimize(&rows, &read_time)) |
5479 | DBUG_RETURN(TRUE); |
5480 | |
5481 | subq_pred->jtbm_read_time= read_time; |
5482 | subq_pred->jtbm_record_count=rows; |
5483 | JOIN *subq_join= subq_pred->unit->first_select()->join; |
5484 | |
5485 | if (!subq_join->tables_list || !subq_join->table_count) |
5486 | { |
5487 | /* |
5488 | A special case; subquery's join is degenerate, and it either produces |
5489 | 0 or 1 record. Examples of both cases: |
5490 | |
5491 | select * from ot where col in (select ... from it where 2>3) |
5492 | select * from ot where col in (select MY_MIN(it.key) from it) |
5493 | |
5494 | in this case, the subquery predicate has not been setup for |
5495 | materialization. In particular, there is no materialized temp.table. |
5496 | We'll now need to |
5497 | 1. Check whether 1 or 0 records are produced, setup this as a |
5498 | constant join tab. |
5499 | 2. Create a dummy temporary table, because all of the join |
5500 | optimization code relies on TABLE object being present (here we |
5501 | follow a bad tradition started by derived tables) |
5502 | */ |
5503 | DBUG_ASSERT(subq_pred->engine->engine_type() == |
5504 | subselect_engine::SINGLE_SELECT_ENGINE); |
5505 | subselect_single_select_engine *engine= |
5506 | (subselect_single_select_engine*)subq_pred->engine; |
5507 | select_value_catcher *new_sink; |
5508 | if (!(new_sink= |
5509 | new (thd->mem_root) select_value_catcher(thd, subq_pred))) |
5510 | DBUG_RETURN(TRUE); |
5511 | if (new_sink->setup(&engine->select_lex->join->fields_list) || |
5512 | engine->select_lex->join->change_result(new_sink, NULL) || |
5513 | engine->exec()) |
5514 | { |
5515 | DBUG_RETURN(TRUE); |
5516 | } |
5517 | subq_pred->is_jtbm_const_tab= TRUE; |
5518 | |
5519 | if (new_sink->assigned) |
5520 | { |
5521 | subq_pred->jtbm_const_row_found= TRUE; |
5522 | /* |
5523 | Subselect produced one row, which is saved in new_sink->row. |
5524 | Inject "left_expr[i] == row[i] equalities into parent's WHERE. |
5525 | */ |
5526 | Item *eq_cond; |
5527 | for (uint i= 0; i < subq_pred->left_expr->cols(); i++) |
5528 | { |
5529 | eq_cond= new (thd->mem_root) |
5530 | Item_func_eq(thd, subq_pred->left_expr->element_index(i), |
5531 | new_sink->row[i]); |
5532 | if (!eq_cond) |
5533 | DBUG_RETURN(1); |
5534 | |
5535 | if (!((*join_where)= and_items(thd, *join_where, eq_cond)) || |
5536 | (*join_where)->fix_fields(thd, join_where)) |
5537 | DBUG_RETURN(1); |
5538 | } |
5539 | } |
5540 | else |
5541 | { |
5542 | /* Subselect produced no rows. Just set the flag, */ |
5543 | subq_pred->jtbm_const_row_found= FALSE; |
5544 | } |
5545 | |
5546 | /* Set up a dummy TABLE*, optimizer code needs JOIN_TABs to have TABLE */ |
5547 | TABLE *dummy_table; |
5548 | if (!(dummy_table= create_dummy_tmp_table(thd))) |
5549 | DBUG_RETURN(1); |
5550 | table->table= dummy_table; |
5551 | table->table->pos_in_table_list= table; |
5552 | /* |
5553 | Note: the table created above may be freed by: |
5554 | 1. JOIN_TAB::cleanup(), when the parent join is a regular join. |
5555 | 2. cleanup_empty_jtbm_semi_joins(), when the parent join is a |
5556 | degenerate join (e.g. one with "Impossible where"). |
5557 | */ |
5558 | setup_table_map(table->table, table, table->jtbm_table_no); |
5559 | } |
5560 | else |
5561 | { |
5562 | DBUG_ASSERT(subq_pred->test_set_strategy(SUBS_MATERIALIZATION)); |
5563 | subq_pred->is_jtbm_const_tab= FALSE; |
5564 | subselect_hash_sj_engine *hash_sj_engine= |
5565 | ((subselect_hash_sj_engine*)item->engine); |
5566 | |
5567 | table->table= hash_sj_engine->tmp_table; |
5568 | table->table->pos_in_table_list= table; |
5569 | |
5570 | setup_table_map(table->table, table, table->jtbm_table_no); |
5571 | |
5572 | Item *sj_conds= hash_sj_engine->semi_join_conds; |
5573 | |
5574 | (*join_where)= and_items(thd, *join_where, sj_conds); |
5575 | if (!(*join_where)->fixed) |
5576 | (*join_where)->fix_fields(thd, join_where); |
5577 | } |
5578 | table->table->maybe_null= MY_TEST(join->mixed_implicit_grouping); |
5579 | } |
5580 | |
5581 | if ((nested_join= table->nested_join)) |
5582 | { |
5583 | if (setup_jtbm_semi_joins(join, &nested_join->join_list, join_where)) |
5584 | DBUG_RETURN(TRUE); |
5585 | } |
5586 | } |
5587 | DBUG_RETURN(FALSE); |
5588 | } |
5589 | |
5590 | |
5591 | /* |
5592 | Cleanup non-merged semi-joins (JBMs) that have empty. |
5593 | |
5594 | This function is to cleanups for a special case: |
5595 | Consider a query like |
5596 | |
5597 | select * from t1 where 1=2 AND t1.col IN (select max(..) ... having 1=2) |
5598 | |
5599 | For this query, optimization of subquery will short-circuit, and |
5600 | setup_jtbm_semi_joins() will call create_dummy_tmp_table() so that we have |
5601 | empty, constant temp.table to stand in as materialized temp. table. |
5602 | |
5603 | Now, suppose that the upper join is also found to be degenerate. In that |
5604 | case, no JOIN_TAB array will be produced, and hence, JOIN::cleanup() will |
5605 | have a problem with cleaning up empty JTBMs (non-empty ones are cleaned up |
5606 | through Item::cleanup() calls). |
5607 | */ |
5608 | |
5609 | void cleanup_empty_jtbm_semi_joins(JOIN *join, List<TABLE_LIST> *join_list) |
5610 | { |
5611 | List_iterator<TABLE_LIST> li(*join_list); |
5612 | TABLE_LIST *table; |
5613 | while ((table= li++)) |
5614 | { |
5615 | if ((table->jtbm_subselect && table->jtbm_subselect->is_jtbm_const_tab)) |
5616 | { |
5617 | if (table->table) |
5618 | { |
5619 | free_tmp_table(join->thd, table->table); |
5620 | table->table= NULL; |
5621 | } |
5622 | } |
5623 | else if (table->nested_join && table->sj_subq_pred) |
5624 | { |
5625 | cleanup_empty_jtbm_semi_joins(join, &table->nested_join->join_list); |
5626 | } |
5627 | } |
5628 | } |
5629 | |
5630 | |
5631 | /** |
5632 | Choose an optimal strategy to execute an IN/ALL/ANY subquery predicate |
5633 | based on cost. |
5634 | |
5635 | @param join_tables the set of tables joined in the subquery |
5636 | |
5637 | @notes |
5638 | The method chooses between the materialization and IN=>EXISTS rewrite |
5639 | strategies for the execution of a non-flattened subquery IN predicate. |
5640 | The cost-based decision is made as follows: |
5641 | |
5642 | 1. compute materialize_strategy_cost based on the unmodified subquery |
5643 | 2. reoptimize the subquery taking into account the IN-EXISTS predicates |
5644 | 3. compute in_exists_strategy_cost based on the reoptimized plan |
5645 | 4. compare and set the cheaper strategy |
5646 | if (materialize_strategy_cost >= in_exists_strategy_cost) |
5647 | in_strategy = MATERIALIZATION |
5648 | else |
5649 | in_strategy = IN_TO_EXISTS |
5650 | 5. if in_strategy = MATERIALIZATION and it is not possible to initialize it |
5651 | revert to IN_TO_EXISTS |
5652 | 6. if (in_strategy == MATERIALIZATION) |
5653 | revert the subquery plan to the original one before reoptimizing |
5654 | else |
5655 | inject the IN=>EXISTS predicates into the new EXISTS subquery plan |
5656 | |
5657 | The implementation itself is a bit more complicated because it takes into |
5658 | account two more factors: |
5659 | - whether the user allowed both strategies through an optimizer_switch, and |
5660 | - if materialization was the cheaper strategy, whether it can be executed |
5661 | or not. |
5662 | |
5663 | @retval FALSE success. |
5664 | @retval TRUE error occurred. |
5665 | */ |
5666 | |
5667 | bool JOIN::choose_subquery_plan(table_map join_tables) |
5668 | { |
5669 | enum_reopt_result reopt_result= REOPT_NONE; |
5670 | Item_in_subselect *in_subs; |
5671 | |
5672 | /* |
5673 | IN/ALL/ANY optimizations are not applicable for so called fake select |
5674 | (this select exists only to filter results of union if it is needed). |
5675 | */ |
5676 | if (select_lex == select_lex->master_unit()->fake_select_lex) |
5677 | return 0; |
5678 | |
5679 | if (is_in_subquery()) |
5680 | { |
5681 | in_subs= (Item_in_subselect*) unit->item; |
5682 | if (in_subs->create_in_to_exists_cond(this)) |
5683 | return true; |
5684 | } |
5685 | else |
5686 | return false; |
5687 | |
5688 | /* A strategy must be chosen earlier. */ |
5689 | DBUG_ASSERT(in_subs->has_strategy()); |
5690 | DBUG_ASSERT(in_to_exists_where || in_to_exists_having); |
5691 | DBUG_ASSERT(!in_to_exists_where || in_to_exists_where->fixed); |
5692 | DBUG_ASSERT(!in_to_exists_having || in_to_exists_having->fixed); |
5693 | |
5694 | /* The original QEP of the subquery. */ |
5695 | Join_plan_state save_qep(table_count); |
5696 | |
5697 | /* |
5698 | Compute and compare the costs of materialization and in-exists if both |
5699 | strategies are possible and allowed by the user (checked during the prepare |
5700 | phase. |
5701 | */ |
5702 | if (in_subs->test_strategy(SUBS_MATERIALIZATION) && |
5703 | in_subs->test_strategy(SUBS_IN_TO_EXISTS)) |
5704 | { |
5705 | JOIN *outer_join; |
5706 | JOIN *inner_join= this; |
5707 | /* Number of unique value combinations filtered by the IN predicate. */ |
5708 | double outer_lookup_keys; |
5709 | /* Cost and row count of the unmodified subquery. */ |
5710 | double inner_read_time_1, inner_record_count_1; |
5711 | /* Cost of the subquery with injected IN-EXISTS predicates. */ |
5712 | double inner_read_time_2; |
5713 | /* The cost to compute IN via materialization. */ |
5714 | double materialize_strategy_cost; |
5715 | /* The cost of the IN->EXISTS strategy. */ |
5716 | double in_exists_strategy_cost; |
5717 | double dummy; |
5718 | |
5719 | /* |
5720 | A. Estimate the number of rows of the outer table that will be filtered |
5721 | by the IN predicate. |
5722 | */ |
5723 | outer_join= unit->outer_select() ? unit->outer_select()->join : NULL; |
5724 | /* |
5725 | Get the cost of the outer join if: |
5726 | (1) It has at least one table, and |
5727 | (2) It has been already optimized (if there is no join_tab, then the |
5728 | outer join has not been optimized yet). |
5729 | */ |
5730 | if (outer_join && outer_join->table_count > 0 && // (1) |
5731 | outer_join->join_tab && // (2) |
5732 | !in_subs->const_item()) |
5733 | { |
5734 | /* |
5735 | TODO: |
5736 | Currently outer_lookup_keys is computed as the number of rows in |
5737 | the partial join including the JOIN_TAB where the IN predicate is |
5738 | pushed to. In the general case this is a gross overestimate because |
5739 | due to caching we are interested only in the number of unique keys. |
5740 | The search key may be formed by columns from much fewer than all |
5741 | tables in the partial join. Example: |
5742 | select * from t1, t2 where t1.c1 = t2.key AND t2.c2 IN (select ...); |
5743 | If the join order: t1, t2, the number of unique lookup keys is ~ to |
5744 | the number of unique values t2.c2 in the partial join t1 join t2. |
5745 | */ |
5746 | outer_join->get_partial_cost_and_fanout(in_subs->get_join_tab_idx(), |
5747 | table_map(-1), |
5748 | &dummy, |
5749 | &outer_lookup_keys); |
5750 | } |
5751 | else |
5752 | { |
5753 | /* |
5754 | TODO: outer_join can be NULL for DELETE statements. |
5755 | How to compute its cost? |
5756 | */ |
5757 | outer_lookup_keys= 1; |
5758 | } |
5759 | |
5760 | /* |
5761 | B. Estimate the cost and number of records of the subquery both |
5762 | unmodified, and with injected IN->EXISTS predicates. |
5763 | */ |
5764 | inner_read_time_1= inner_join->best_read; |
5765 | inner_record_count_1= inner_join->join_record_count; |
5766 | |
5767 | if (in_to_exists_where && const_tables != table_count) |
5768 | { |
5769 | /* |
5770 | Re-optimize and cost the subquery taking into account the IN-EXISTS |
5771 | conditions. |
5772 | */ |
5773 | reopt_result= reoptimize(in_to_exists_where, join_tables, &save_qep); |
5774 | if (reopt_result == REOPT_ERROR) |
5775 | return TRUE; |
5776 | |
5777 | /* Get the cost of the modified IN-EXISTS plan. */ |
5778 | inner_read_time_2= inner_join->best_read; |
5779 | |
5780 | } |
5781 | else |
5782 | { |
5783 | /* Reoptimization would not produce any better plan. */ |
5784 | inner_read_time_2= inner_read_time_1; |
5785 | } |
5786 | |
5787 | /* |
5788 | C. Compute execution costs. |
5789 | */ |
5790 | /* C.1 Compute the cost of the materialization strategy. */ |
5791 | //uint rowlen= get_tmp_table_rec_length(unit->first_select()->item_list); |
5792 | uint rowlen= get_tmp_table_rec_length(ref_ptrs, |
5793 | select_lex->item_list.elements); |
5794 | /* The cost of writing one row into the temporary table. */ |
5795 | double write_cost= get_tmp_table_write_cost(thd, inner_record_count_1, |
5796 | rowlen); |
5797 | /* The cost of a lookup into the unique index of the materialized table. */ |
5798 | double lookup_cost= get_tmp_table_lookup_cost(thd, inner_record_count_1, |
5799 | rowlen); |
5800 | /* |
5801 | The cost of executing the subquery and storing its result in an indexed |
5802 | temporary table. |
5803 | */ |
5804 | double materialization_cost= inner_read_time_1 + |
5805 | write_cost * inner_record_count_1; |
5806 | |
5807 | materialize_strategy_cost= materialization_cost + |
5808 | outer_lookup_keys * lookup_cost; |
5809 | |
5810 | /* C.2 Compute the cost of the IN=>EXISTS strategy. */ |
5811 | in_exists_strategy_cost= outer_lookup_keys * inner_read_time_2; |
5812 | |
5813 | /* C.3 Compare the costs and choose the cheaper strategy. */ |
5814 | if (materialize_strategy_cost >= in_exists_strategy_cost) |
5815 | in_subs->set_strategy(SUBS_IN_TO_EXISTS); |
5816 | else |
5817 | in_subs->set_strategy(SUBS_MATERIALIZATION); |
5818 | |
5819 | DBUG_PRINT("info" , |
5820 | ("mat_strategy_cost: %.2f, mat_cost: %.2f, write_cost: %.2f, lookup_cost: %.2f" , |
5821 | materialize_strategy_cost, materialization_cost, write_cost, lookup_cost)); |
5822 | DBUG_PRINT("info" , |
5823 | ("inx_strategy_cost: %.2f, inner_read_time_2: %.2f" , |
5824 | in_exists_strategy_cost, inner_read_time_2)); |
5825 | DBUG_PRINT("info" ,("outer_lookup_keys: %.2f" , outer_lookup_keys)); |
5826 | } |
5827 | |
5828 | /* |
5829 | If (1) materialization is a possible strategy based on semantic analysis |
5830 | during the prepare phase, then if |
5831 | (2) it is more expensive than the IN->EXISTS transformation, and |
5832 | (3) it is not possible to create usable indexes for the materialization |
5833 | strategy, |
5834 | fall back to IN->EXISTS. |
5835 | otherwise |
5836 | use materialization. |
5837 | */ |
5838 | if (in_subs->test_strategy(SUBS_MATERIALIZATION) && |
5839 | in_subs->setup_mat_engine()) |
5840 | { |
5841 | /* |
5842 | If materialization was the cheaper or the only user-selected strategy, |
5843 | but it is not possible to execute it due to limitations in the |
5844 | implementation, fall back to IN-TO-EXISTS. |
5845 | */ |
5846 | in_subs->set_strategy(SUBS_IN_TO_EXISTS); |
5847 | } |
5848 | |
5849 | if (in_subs->test_strategy(SUBS_MATERIALIZATION)) |
5850 | { |
5851 | /* Restore the original query plan used for materialization. */ |
5852 | if (reopt_result == REOPT_NEW_PLAN) |
5853 | restore_query_plan(&save_qep); |
5854 | |
5855 | in_subs->unit->uncacheable&= ~UNCACHEABLE_DEPENDENT_INJECTED; |
5856 | select_lex->uncacheable&= ~UNCACHEABLE_DEPENDENT_INJECTED; |
5857 | |
5858 | /* |
5859 | Reset the "LIMIT 1" set in Item_exists_subselect::fix_length_and_dec. |
5860 | TODO: |
5861 | Currently we set the subquery LIMIT to infinity, and this is correct |
5862 | because we forbid at parse time LIMIT inside IN subqueries (see |
5863 | Item_in_subselect::test_limit). However, once we allow this, here |
5864 | we should set the correct limit if given in the query. |
5865 | */ |
5866 | in_subs->unit->global_parameters()->select_limit= NULL; |
5867 | in_subs->unit->set_limit(unit->global_parameters()); |
5868 | /* |
5869 | Set the limit of this JOIN object as well, because normally its being |
5870 | set in the beginning of JOIN::optimize, which was already done. |
5871 | */ |
5872 | select_limit= in_subs->unit->select_limit_cnt; |
5873 | } |
5874 | else if (in_subs->test_strategy(SUBS_IN_TO_EXISTS)) |
5875 | { |
5876 | if (reopt_result == REOPT_NONE && in_to_exists_where && |
5877 | const_tables != table_count) |
5878 | { |
5879 | /* |
5880 | The subquery was not reoptimized with the newly injected IN-EXISTS |
5881 | conditions either because the user allowed only the IN-EXISTS strategy, |
5882 | or because materialization was not possible based on semantic analysis. |
5883 | */ |
5884 | reopt_result= reoptimize(in_to_exists_where, join_tables, NULL); |
5885 | if (reopt_result == REOPT_ERROR) |
5886 | return TRUE; |
5887 | } |
5888 | |
5889 | if (in_subs->inject_in_to_exists_cond(this)) |
5890 | return TRUE; |
5891 | /* |
5892 | If the injected predicate is correlated the IN->EXISTS transformation |
5893 | make the subquery dependent. |
5894 | */ |
5895 | if ((in_to_exists_where && |
5896 | in_to_exists_where->used_tables() & OUTER_REF_TABLE_BIT) || |
5897 | (in_to_exists_having && |
5898 | in_to_exists_having->used_tables() & OUTER_REF_TABLE_BIT)) |
5899 | { |
5900 | in_subs->unit->uncacheable|= UNCACHEABLE_DEPENDENT_INJECTED; |
5901 | select_lex->uncacheable|= UNCACHEABLE_DEPENDENT_INJECTED; |
5902 | } |
5903 | select_limit= 1; |
5904 | } |
5905 | else |
5906 | DBUG_ASSERT(FALSE); |
5907 | |
5908 | return FALSE; |
5909 | } |
5910 | |
5911 | |
5912 | /** |
5913 | Choose a query plan for a table-less subquery. |
5914 | |
5915 | @notes |
5916 | |
5917 | @retval FALSE success. |
5918 | @retval TRUE error occurred. |
5919 | */ |
5920 | |
5921 | bool JOIN::choose_tableless_subquery_plan() |
5922 | { |
5923 | DBUG_ASSERT(!tables_list || !table_count); |
5924 | if (unit->item) |
5925 | { |
5926 | DBUG_ASSERT(unit->item->type() == Item::SUBSELECT_ITEM); |
5927 | Item_subselect *subs_predicate= unit->item; |
5928 | |
5929 | /* |
5930 | If the optimizer determined that his query has an empty result, |
5931 | in most cases the subquery predicate is a known constant value - |
5932 | either of TRUE, FALSE or NULL. The implementation of |
5933 | Item_subselect::no_rows_in_result() determines which one. |
5934 | */ |
5935 | if (zero_result_cause) |
5936 | { |
5937 | if (!implicit_grouping) |
5938 | { |
5939 | /* |
5940 | Both group by queries and non-group by queries without aggregate |
5941 | functions produce empty subquery result. There is no need to further |
5942 | rewrite the subquery because it will not be executed at all. |
5943 | */ |
5944 | return FALSE; |
5945 | } |
5946 | |
5947 | /* @todo |
5948 | A further optimization is possible when a non-group query with |
5949 | MIN/MAX/COUNT is optimized by opt_sum_query. Then, if there are |
5950 | only MIN/MAX functions over an empty result set, the subquery |
5951 | result is a NULL value/row, thus the value of subs_predicate is |
5952 | NULL. |
5953 | */ |
5954 | } |
5955 | |
5956 | /* |
5957 | For IN subqueries, use IN->EXISTS transfomation, unless the subquery |
5958 | has been converted to a JTBM semi-join. In that case, just leave |
5959 | everything as-is, setup_jtbm_semi_joins() has special handling for cases |
5960 | like this. |
5961 | */ |
5962 | if (subs_predicate->is_in_predicate() && |
5963 | !(subs_predicate->substype() == Item_subselect::IN_SUBS && |
5964 | ((Item_in_subselect*)subs_predicate)->is_jtbm_merged)) |
5965 | { |
5966 | Item_in_subselect *in_subs; |
5967 | in_subs= (Item_in_subselect*) subs_predicate; |
5968 | in_subs->set_strategy(SUBS_IN_TO_EXISTS); |
5969 | if (in_subs->create_in_to_exists_cond(this) || |
5970 | in_subs->inject_in_to_exists_cond(this)) |
5971 | return TRUE; |
5972 | tmp_having= having; |
5973 | } |
5974 | } |
5975 | exec_const_cond= conds; |
5976 | return FALSE; |
5977 | } |
5978 | |