1039.2.2
by Jay Pipes
Phase 2 of JOIN refactoring. |
1 |
/* - mode: c; c-basic-offset: 2; indent-tabs-mode: nil; -*-
|
2 |
* vim:expandtab:shiftwidth=2:tabstop=2:smarttab:
|
|
3 |
*
|
|
4 |
* Copyright (C) 2008-2009 Sun Microsystems
|
|
5 |
*
|
|
6 |
* This program is free software; you can redistribute it and/or modify
|
|
7 |
* it under the terms of the GNU General Public License as published by
|
|
8 |
* the Free Software Foundation; either version 2 of the License, or
|
|
9 |
* (at your option) any later version.
|
|
10 |
*
|
|
11 |
* This program is distributed in the hope that it will be useful,
|
|
12 |
* but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
13 |
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
|
14 |
* GNU General Public License for more details.
|
|
15 |
*
|
|
16 |
* You should have received a copy of the GNU General Public License
|
|
17 |
* along with this program; if not, write to the Free Software
|
|
18 |
* Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
|
|
19 |
*/
|
|
20 |
||
21 |
/**
|
|
22 |
* @file
|
|
23 |
*
|
|
24 |
* Implementation of the JOIN class
|
|
25 |
*
|
|
26 |
* @defgroup Query_Optimizer Query Optimizer
|
|
27 |
* @{
|
|
28 |
*/
|
|
29 |
||
30 |
#include "drizzled/server_includes.h" |
|
31 |
#include "drizzled/sj_tmp_table.h" |
|
32 |
#include "drizzled/table_map_iterator.h" |
|
33 |
#include "drizzled/item/cache.h" |
|
34 |
#include "drizzled/item/cmpfunc.h" |
|
35 |
#include "drizzled/item/copy_string.h" |
|
36 |
#include "drizzled/item/uint.h" |
|
37 |
#include "drizzled/cached_item.h" |
|
38 |
#include "drizzled/sql_base.h" |
|
39 |
#include "drizzled/sql_select.h" /* include join.h */ |
|
40 |
#include "drizzled/lock.h" |
|
41 |
#include "drizzled/nested_join.h" |
|
42 |
#include "drizzled/join.h" |
|
1039.2.7
by Jay Pipes
Yet more style and indentation cleanups. |
43 |
#include "drizzled/join_cache.h" |
1039.2.2
by Jay Pipes
Phase 2 of JOIN refactoring. |
44 |
#include "drizzled/show.h" |
45 |
#include "drizzled/field/blob.h" |
|
46 |
#include "mysys/my_bit.h" |
|
47 |
||
1067.4.4
by Nathan Williams
The rest of the files in the drizzled directory were purged of the cmin macro and replace with std::min (except for the definition in globals.h and 1 usage in stacktrace.cc). |
48 |
#include <algorithm> |
49 |
||
50 |
using namespace std; |
|
51 |
||
1039.2.2
by Jay Pipes
Phase 2 of JOIN refactoring. |
52 |
/** Declarations of static functions used in this source file. */
|
53 |
static bool make_group_fields(JOIN *main_join, JOIN *curr_join); |
|
54 |
static void calc_group_buffer(JOIN *join,order_st *group); |
|
55 |
static bool alloc_group_fields(JOIN *join,order_st *group); |
|
56 |
/*
|
|
57 |
TODO: 'find_best' is here only temporarily until 'greedy_search' is
|
|
58 |
tested and approved.
|
|
59 |
*/
|
|
60 |
static bool find_best(JOIN *join,table_map rest_tables,uint32_t index, double record_count,double read_time); |
|
61 |
static uint32_t cache_record_length(JOIN *join, uint32_t index); |
|
62 |
static double prev_record_reads(JOIN *join, uint32_t idx, table_map found_ref); |
|
63 |
static bool get_best_combination(JOIN *join); |
|
64 |
static void set_position(JOIN *join,uint32_t index,JOIN_TAB *table,KEYUSE *key); |
|
65 |
static bool choose_plan(JOIN *join,table_map join_tables); |
|
66 |
static void best_access_path(JOIN *join, JOIN_TAB *s, |
|
67 |
Session *session, |
|
68 |
table_map remaining_tables, |
|
69 |
uint32_t idx, |
|
70 |
double record_count, |
|
71 |
double read_time); |
|
72 |
static void optimize_straight_join(JOIN *join, table_map join_tables); |
|
73 |
static bool greedy_search(JOIN *join, table_map remaining_tables, uint32_t depth, uint32_t prune_level); |
|
74 |
static bool best_extension_by_limited_search(JOIN *join, |
|
75 |
table_map remaining_tables, |
|
76 |
uint32_t idx, |
|
77 |
double record_count, |
|
78 |
double read_time, |
|
79 |
uint32_t depth, |
|
80 |
uint32_t prune_level); |
|
81 |
static uint32_t determine_search_depth(JOIN* join); |
|
82 |
static bool make_simple_join(JOIN *join,Table *tmp_table); |
|
83 |
static void make_outerjoin_info(JOIN *join); |
|
84 |
static bool make_join_select(JOIN *join,SQL_SELECT *select,COND *item); |
|
85 |
static bool make_join_readinfo(JOIN *join, uint64_t options, uint32_t no_jbuf_after); |
|
86 |
static void update_depend_map(JOIN *join); |
|
87 |
static void update_depend_map(JOIN *join, order_st *order); |
|
88 |
static order_st *remove_constants(JOIN *join,order_st *first_order,COND *cond, bool change_list, bool *simple_order); |
|
89 |
static int return_zero_rows(JOIN *join, |
|
90 |
select_result *res, |
|
91 |
TableList *tables, |
|
92 |
List<Item> &fields, |
|
93 |
bool send_row, |
|
94 |
uint64_t select_options, |
|
95 |
const char *info, |
|
96 |
Item *having); |
|
97 |
static COND *simplify_joins(JOIN *join, List<TableList> *join_list, COND *conds, bool top, bool in_sj); |
|
98 |
static int remove_duplicates(JOIN *join,Table *entry,List<Item> &fields, Item *having); |
|
99 |
static int setup_without_group(Session *session, |
|
100 |
Item **ref_pointer_array, |
|
101 |
TableList *tables, |
|
102 |
TableList *, |
|
103 |
List<Item> &fields, |
|
104 |
List<Item> &all_fields, |
|
105 |
COND **conds, |
|
106 |
order_st *order, |
|
107 |
order_st *group, |
|
108 |
bool *hidden_group_fields); |
|
109 |
static bool make_join_statistics(JOIN *join, TableList *leaves, COND *conds, DYNAMIC_ARRAY *keyuse); |
|
110 |
static uint32_t build_bitmap_for_nested_joins(List<TableList> *join_list, uint32_t first_unused); |
|
111 |
static Table *get_sort_by_table(order_st *a,order_st *b,TableList *tables); |
|
112 |
static void reset_nj_counters(List<TableList> *join_list); |
|
113 |
static bool test_if_subpart(order_st *a,order_st *b); |
|
114 |
static void restore_prev_nj_state(JOIN_TAB *last); |
|
115 |
static uint32_t make_join_orderinfo(JOIN *join); |
|
116 |
static int setup_semijoin_dups_elimination(JOIN *join, uint64_t options, uint32_t no_jbuf_after); |
|
117 |
static void cleanup_sj_tmp_tables(JOIN *join); |
|
118 |
static bool add_ref_to_table_cond(Session *session, JOIN_TAB *join_tab); |
|
119 |
static bool replace_where_subcondition(JOIN *join, Item *old_cond, Item *new_cond, bool fix_fields); |
|
120 |
static int pull_out_semijoin_tables(JOIN *join); |
|
121 |
static int do_sj_dups_weedout(Session *session, SJ_TMP_TABLE *sjtbl); |
|
122 |
static void free_blobs(Field **ptr); /* Rename this method...conflicts with another in global namespace... */ |
|
123 |
static bool bitmap_covers(const table_map x, const table_map y); |
|
124 |
static bool sj_table_is_included(JOIN *join, JOIN_TAB *join_tab); |
|
125 |
||
126 |
/**
|
|
127 |
Prepare of whole select (including sub queries in future).
|
|
128 |
||
129 |
@todo
|
|
130 |
Add check of calculation of GROUP functions and fields:
|
|
131 |
SELECT COUNT(*)+table.col1 from table1;
|
|
132 |
||
133 |
@retval
|
|
134 |
-1 on error
|
|
135 |
@retval
|
|
136 |
0 on success
|
|
137 |
*/
|
|
138 |
int JOIN::prepare(Item ***rref_pointer_array, |
|
139 |
TableList *tables_init, |
|
140 |
uint32_t wild_num, |
|
141 |
COND *conds_init, |
|
142 |
uint32_t og_num, |
|
143 |
order_st *order_init, |
|
144 |
order_st *group_init, |
|
145 |
Item *having_init, |
|
146 |
Select_Lex *select_lex_arg, |
|
147 |
Select_Lex_Unit *unit_arg) |
|
148 |
{
|
|
149 |
// to prevent double initialization on EXPLAIN
|
|
150 |
if (optimized) |
|
151 |
return 0; |
|
152 |
||
153 |
conds= conds_init; |
|
154 |
order= order_init; |
|
155 |
group_list= group_init; |
|
156 |
having= having_init; |
|
157 |
tables_list= tables_init; |
|
158 |
select_lex= select_lex_arg; |
|
159 |
select_lex->join= this; |
|
160 |
join_list= &select_lex->top_join_list; |
|
161 |
union_part= unit_arg->is_union(); |
|
162 |
||
163 |
session->lex->current_select->is_item_list_lookup= 1; |
|
164 |
/*
|
|
165 |
If we have already executed SELECT, then it have not sense to prevent
|
|
166 |
its table from update (see unique_table())
|
|
167 |
*/
|
|
168 |
if (session->derived_tables_processing) |
|
169 |
select_lex->exclude_from_table_unique_test= true; |
|
170 |
||
171 |
/* Check that all tables, fields, conds and order are ok */
|
|
172 |
||
173 |
if (!(select_options & OPTION_SETUP_TABLES_DONE) && |
|
174 |
setup_tables_and_check_access(session, &select_lex->context, join_list, |
|
175 |
tables_list, &select_lex->leaf_tables, |
|
176 |
false)) |
|
177 |
return(-1); |
|
178 |
||
179 |
TableList *table_ptr; |
|
180 |
for (table_ptr= select_lex->leaf_tables; |
|
181 |
table_ptr; |
|
182 |
table_ptr= table_ptr->next_leaf) |
|
183 |
tables++; |
|
184 |
||
185 |
if (setup_wild(session, fields_list, &all_fields, wild_num) || |
|
186 |
select_lex->setup_ref_array(session, og_num) || |
|
187 |
setup_fields(session, (*rref_pointer_array), fields_list, MARK_COLUMNS_READ, |
|
188 |
&all_fields, 1) || |
|
189 |
setup_without_group(session, (*rref_pointer_array), tables_list, |
|
190 |
select_lex->leaf_tables, fields_list, |
|
191 |
all_fields, &conds, order, group_list, |
|
192 |
&hidden_group_fields)) |
|
193 |
return(-1); /* purecov: inspected */ |
|
194 |
||
195 |
ref_pointer_array= *rref_pointer_array; |
|
196 |
||
197 |
if (having) |
|
198 |
{
|
|
199 |
nesting_map save_allow_sum_func= session->lex->allow_sum_func; |
|
200 |
session->where="having clause"; |
|
201 |
session->lex->allow_sum_func|= 1 << select_lex_arg->nest_level; |
|
202 |
select_lex->having_fix_field= 1; |
|
203 |
bool having_fix_rc= (!having->fixed && |
|
204 |
(having->fix_fields(session, &having) || |
|
205 |
having->check_cols(1))); |
|
206 |
select_lex->having_fix_field= 0; |
|
207 |
if (having_fix_rc || session->is_error()) |
|
208 |
return(-1); /* purecov: inspected */ |
|
209 |
session->lex->allow_sum_func= save_allow_sum_func; |
|
210 |
}
|
|
211 |
||
212 |
{
|
|
213 |
Item_subselect *subselect; |
|
214 |
Item_in_subselect *in_subs= NULL; |
|
215 |
/*
|
|
216 |
Are we in a subquery predicate?
|
|
217 |
TODO: the block below will be executed for every PS execution without need.
|
|
218 |
*/
|
|
219 |
if ((subselect= select_lex->master_unit()->item)) |
|
220 |
{
|
|
221 |
bool do_semijoin= !test(session->variables.optimizer_switch & |
|
222 |
OPTIMIZER_SWITCH_NO_SEMIJOIN); |
|
223 |
if (subselect->substype() == Item_subselect::IN_SUBS) |
|
224 |
in_subs= (Item_in_subselect*)subselect; |
|
225 |
||
226 |
/*
|
|
227 |
Check if we're in subquery that is a candidate for flattening into a
|
|
228 |
semi-join (which is done done in flatten_subqueries()). The
|
|
229 |
requirements are:
|
|
230 |
1. Subquery predicate is an IN/=ANY subq predicate
|
|
231 |
2. Subquery is a single SELECT (not a UNION)
|
|
232 |
3. Subquery does not have GROUP BY or order_st BY
|
|
233 |
4. Subquery does not use aggregate functions or HAVING
|
|
234 |
5. Subquery predicate is at the AND-top-level of ON/WHERE clause
|
|
235 |
6. No execution method was already chosen (by a prepared statement).
|
|
236 |
||
237 |
(*). We are not in a subquery of a single table UPDATE/DELETE that
|
|
238 |
doesn't have a JOIN (TODO: We should handle this at some
|
|
239 |
point by switching to multi-table UPDATE/DELETE)
|
|
240 |
||
241 |
(**). We're not in a confluent table-less subquery, like
|
|
242 |
"SELECT 1".
|
|
243 |
*/
|
|
244 |
if (in_subs && // 1 |
|
245 |
!select_lex->master_unit()->first_select()->next_select() && // 2 |
|
246 |
!select_lex->group_list.elements && !order && // 3 |
|
247 |
!having && !select_lex->with_sum_func && // 4 |
|
248 |
session->session_marker && // 5 |
|
249 |
select_lex->outer_select()->join && // (*) |
|
250 |
select_lex->master_unit()->first_select()->leaf_tables && // (**) |
|
251 |
do_semijoin && |
|
252 |
in_subs->exec_method == Item_in_subselect::NOT_TRANSFORMED) // 6 |
|
253 |
{
|
|
254 |
{
|
|
255 |
if (!in_subs->left_expr->fixed && |
|
256 |
in_subs->left_expr->fix_fields(session, &in_subs->left_expr)) |
|
257 |
{
|
|
258 |
return(-1); |
|
259 |
}
|
|
260 |
/*
|
|
261 |
Check that the right part of the subselect contains no more than one
|
|
262 |
column. E.g. in SELECT 1 IN (SELECT * ..) the right part is (SELECT * ...)
|
|
263 |
*/
|
|
264 |
if (subselect->substype() == Item_subselect::IN_SUBS && |
|
265 |
(select_lex->item_list.elements != |
|
266 |
((Item_in_subselect*)subselect)->left_expr->cols())) |
|
267 |
{
|
|
268 |
my_error(ER_OPERAND_COLUMNS, MYF(0), ((Item_in_subselect*)subselect)->left_expr->cols()); |
|
269 |
return(-1); |
|
270 |
}
|
|
271 |
}
|
|
272 |
||
273 |
/* Register the subquery for further processing */
|
|
274 |
select_lex->outer_select()->join->sj_subselects.append(session->mem_root, in_subs); |
|
275 |
in_subs->expr_join_nest= (TableList*)session->session_marker; |
|
276 |
}
|
|
277 |
else
|
|
278 |
{
|
|
279 |
bool do_materialize= !test(session->variables.optimizer_switch & |
|
280 |
OPTIMIZER_SWITCH_NO_MATERIALIZATION); |
|
281 |
/*
|
|
282 |
Check if the subquery predicate can be executed via materialization.
|
|
283 |
The required conditions are:
|
|
284 |
1. Subquery predicate is an IN/=ANY subq predicate
|
|
285 |
2. Subquery is a single SELECT (not a UNION)
|
|
286 |
3. Subquery is not a table-less query. In this case there is no
|
|
287 |
point in materializing.
|
|
288 |
4. Subquery predicate is a top-level predicate
|
|
289 |
(this implies it is not negated)
|
|
290 |
TODO: this is a limitation that should be lifeted once we
|
|
291 |
implement correct NULL semantics (WL#3830)
|
|
292 |
5. Subquery is non-correlated
|
|
293 |
TODO:
|
|
294 |
This is an overly restrictive condition. It can be extended to:
|
|
295 |
(Subquery is non-correlated ||
|
|
296 |
Subquery is correlated to any query outer to IN predicate ||
|
|
297 |
(Subquery is correlated to the immediate outer query &&
|
|
298 |
Subquery !contains {GROUP BY, order_st BY [LIMIT],
|
|
299 |
aggregate functions) && subquery predicate is not under "NOT IN"))
|
|
300 |
6. No execution method was already chosen (by a prepared statement).
|
|
301 |
||
302 |
(*) The subquery must be part of a SELECT statement. The current
|
|
303 |
condition also excludes multi-table update statements.
|
|
304 |
||
305 |
We have to determine whether we will perform subquery materialization
|
|
306 |
before calling the IN=>EXISTS transformation, so that we know whether to
|
|
307 |
perform the whole transformation or only that part of it which wraps
|
|
308 |
Item_in_subselect in an Item_in_optimizer.
|
|
309 |
*/
|
|
310 |
if (do_materialize && |
|
311 |
in_subs && // 1 |
|
312 |
!select_lex->master_unit()->first_select()->next_select() && // 2 |
|
313 |
select_lex->master_unit()->first_select()->leaf_tables && // 3 |
|
314 |
session->lex->sql_command == SQLCOM_SELECT) // * |
|
315 |
{
|
|
316 |
if (in_subs->is_top_level_item() && // 4 |
|
317 |
!in_subs->is_correlated && // 5 |
|
318 |
in_subs->exec_method == Item_in_subselect::NOT_TRANSFORMED) // 6 |
|
319 |
in_subs->exec_method= Item_in_subselect::MATERIALIZATION; |
|
320 |
}
|
|
321 |
||
322 |
Item_subselect::trans_res trans_res; |
|
323 |
if ((trans_res= subselect->select_transformer(this)) != |
|
324 |
Item_subselect::RES_OK) |
|
325 |
{
|
|
326 |
return((trans_res == Item_subselect::RES_ERROR)); |
|
327 |
}
|
|
328 |
}
|
|
329 |
}
|
|
330 |
}
|
|
331 |
||
332 |
if (order) |
|
333 |
{
|
|
334 |
order_st *ord; |
|
335 |
for (ord= order; ord; ord= ord->next) |
|
336 |
{
|
|
337 |
Item *item= *ord->item; |
|
338 |
if (item->with_sum_func && item->type() != Item::SUM_FUNC_ITEM) |
|
339 |
item->split_sum_func(session, ref_pointer_array, all_fields); |
|
340 |
}
|
|
341 |
}
|
|
342 |
||
343 |
if (having && having->with_sum_func) |
|
344 |
having->split_sum_func(session, ref_pointer_array, all_fields, |
|
345 |
&having, true); |
|
346 |
if (select_lex->inner_sum_func_list) |
|
347 |
{
|
|
348 |
Item_sum *end=select_lex->inner_sum_func_list; |
|
349 |
Item_sum *item_sum= end; |
|
350 |
do
|
|
351 |
{
|
|
352 |
item_sum= item_sum->next; |
|
353 |
item_sum->split_sum_func(session, ref_pointer_array, |
|
354 |
all_fields, item_sum->ref_by, false); |
|
355 |
} while (item_sum != end); |
|
356 |
}
|
|
357 |
||
358 |
if (select_lex->inner_refs_list.elements && |
|
359 |
fix_inner_refs(session, all_fields, select_lex, ref_pointer_array)) |
|
360 |
return(-1); |
|
361 |
||
362 |
/*
|
|
363 |
Check if there are references to un-aggregated columns when computing
|
|
364 |
aggregate functions with implicit grouping (there is no GROUP BY).
|
|
365 |
||
366 |
MODE_ONLY_FULL_GROUP_BY is enabled here by default
|
|
367 |
*/
|
|
368 |
if (!group_list && select_lex->full_group_by_flag == (NON_AGG_FIELD_USED | SUM_FUNC_USED)) |
|
369 |
{
|
|
370 |
my_message(ER_MIX_OF_GROUP_FUNC_AND_FIELDS, |
|
371 |
ER(ER_MIX_OF_GROUP_FUNC_AND_FIELDS), MYF(0)); |
|
372 |
return(-1); |
|
373 |
}
|
|
374 |
{
|
|
375 |
/* Caclulate the number of groups */
|
|
376 |
send_group_parts= 0; |
|
377 |
for (order_st *group_tmp= group_list ; group_tmp ; group_tmp= group_tmp->next) |
|
378 |
send_group_parts++; |
|
379 |
}
|
|
380 |
||
381 |
if (error) |
|
382 |
goto err; /* purecov: inspected */ |
|
383 |
||
384 |
if (result && result->prepare(fields_list, unit_arg)) |
|
385 |
goto err; /* purecov: inspected */ |
|
386 |
||
387 |
/* Init join struct */
|
|
388 |
count_field_types(select_lex, &tmp_table_param, all_fields, 0); |
|
389 |
ref_pointer_array_size= all_fields.elements*sizeof(Item*); |
|
390 |
this->group= group_list != 0; |
|
391 |
unit= unit_arg; |
|
392 |
||
393 |
#ifdef RESTRICTED_GROUP
|
|
394 |
if (sum_func_count && !group_list && (func_count || field_count)) |
|
395 |
{
|
|
396 |
my_message(ER_WRONG_SUM_SELECT,ER(ER_WRONG_SUM_SELECT),MYF(0)); |
|
397 |
goto err; |
|
398 |
}
|
|
399 |
#endif
|
|
400 |
if (select_lex->olap == ROLLUP_TYPE && rollup_init()) |
|
401 |
goto err; |
|
402 |
if (alloc_func_list()) |
|
403 |
goto err; |
|
404 |
||
405 |
return(0); // All OK |
|
406 |
||
407 |
err: |
|
408 |
return(-1); /* purecov: inspected */ |
|
409 |
}
|
|
410 |
||
411 |
/*
|
|
412 |
Remove the predicates pushed down into the subquery
|
|
413 |
||
414 |
SYNOPSIS
|
|
415 |
JOIN::remove_subq_pushed_predicates()
|
|
416 |
where IN Must be NULL
|
|
417 |
OUT The remaining WHERE condition, or NULL
|
|
418 |
||
419 |
DESCRIPTION
|
|
420 |
Given that this join will be executed using (unique|index)_subquery,
|
|
421 |
without "checking NULL", remove the predicates that were pushed down
|
|
422 |
into the subquery.
|
|
423 |
||
424 |
If the subquery compares scalar values, we can remove the condition that
|
|
425 |
was wrapped into trig_cond (it will be checked when needed by the subquery
|
|
426 |
engine)
|
|
427 |
||
428 |
If the subquery compares row values, we need to keep the wrapped
|
|
429 |
equalities in the WHERE clause: when the left (outer) tuple has both NULL
|
|
430 |
and non-NULL values, we'll do a full table scan and will rely on the
|
|
431 |
equalities corresponding to non-NULL parts of left tuple to filter out
|
|
432 |
non-matching records.
|
|
433 |
||
434 |
TODO: We can remove the equalities that will be guaranteed to be true by the
|
|
435 |
fact that subquery engine will be using index lookup. This must be done only
|
|
436 |
for cases where there are no conversion errors of significance, e.g. 257
|
|
437 |
that is searched in a byte. But this requires homogenization of the return
|
|
438 |
codes of all Field*::store() methods.
|
|
439 |
*/
|
|
440 |
void JOIN::remove_subq_pushed_predicates(Item **where) |
|
441 |
{
|
|
442 |
if (conds->type() == Item::FUNC_ITEM && |
|
443 |
((Item_func *)this->conds)->functype() == Item_func::EQ_FUNC && |
|
444 |
((Item_func *)conds)->arguments()[0]->type() == Item::REF_ITEM && |
|
445 |
((Item_func *)conds)->arguments()[1]->type() == Item::FIELD_ITEM && |
|
446 |
test_if_ref ((Item_field *)((Item_func *)conds)->arguments()[1], |
|
447 |
((Item_func *)conds)->arguments()[0])) |
|
448 |
{
|
|
449 |
*where= 0; |
|
450 |
return; |
|
451 |
}
|
|
452 |
}
|
|
453 |
||
454 |
/**
|
|
455 |
global select optimisation.
|
|
456 |
||
457 |
@note
|
|
458 |
error code saved in field 'error'
|
|
459 |
||
460 |
@retval
|
|
461 |
0 success
|
|
462 |
@retval
|
|
463 |
1 error
|
|
464 |
*/
|
|
1039.2.7
by Jay Pipes
Yet more style and indentation cleanups. |
465 |
int JOIN::optimize() |
1039.2.2
by Jay Pipes
Phase 2 of JOIN refactoring. |
466 |
{
|
467 |
// to prevent double initialization on EXPLAIN
|
|
468 |
if (optimized) |
|
469 |
return(0); |
|
470 |
optimized= 1; |
|
471 |
||
472 |
session->set_proc_info("optimizing"); |
|
473 |
row_limit= ((select_distinct || order || group_list) ? HA_POS_ERROR : |
|
474 |
unit->select_limit_cnt); |
|
475 |
/* select_limit is used to decide if we are likely to scan the whole table */
|
|
476 |
select_limit= unit->select_limit_cnt; |
|
477 |
if (having || (select_options & OPTION_FOUND_ROWS)) |
|
478 |
select_limit= HA_POS_ERROR; |
|
479 |
do_send_rows = (unit->select_limit_cnt) ? 1 : 0; |
|
480 |
// Ignore errors of execution if option IGNORE present
|
|
481 |
if (session->lex->ignore) |
|
482 |
session->lex->current_select->no_error= 1; |
|
483 |
||
484 |
#ifdef HAVE_REF_TO_FIELDS // Not done yet |
|
485 |
/* Add HAVING to WHERE if possible */
|
|
486 |
if (having && !group_list && !sum_func_count) |
|
487 |
{
|
|
488 |
if (!conds) |
|
489 |
{
|
|
490 |
conds= having; |
|
491 |
having= 0; |
|
492 |
}
|
|
493 |
else if ((conds=new Item_cond_and(conds,having))) |
|
494 |
{
|
|
495 |
/*
|
|
496 |
Item_cond_and can't be fixed after creation, so we do not check
|
|
497 |
conds->fixed
|
|
498 |
*/
|
|
499 |
conds->fix_fields(session, &conds); |
|
500 |
conds->change_ref_to_fields(session, tables_list); |
|
501 |
conds->top_level_item(); |
|
502 |
having= 0; |
|
503 |
}
|
|
504 |
}
|
|
505 |
#endif
|
|
506 |
||
507 |
/* Convert all outer joins to inner joins if possible */
|
|
508 |
conds= simplify_joins(this, join_list, conds, true, false); |
|
509 |
build_bitmap_for_nested_joins(join_list, 0); |
|
510 |
||
511 |
conds= optimize_cond(this, conds, join_list, &cond_value); |
|
512 |
if (session->is_error()) |
|
513 |
{
|
|
514 |
error= 1; |
|
515 |
return(1); |
|
516 |
}
|
|
517 |
||
518 |
{
|
|
519 |
having= optimize_cond(this, having, join_list, &having_value); |
|
520 |
if (session->is_error()) |
|
521 |
{
|
|
522 |
error= 1; |
|
523 |
return(1); |
|
524 |
}
|
|
525 |
if (select_lex->where) |
|
526 |
select_lex->cond_value= cond_value; |
|
527 |
if (select_lex->having) |
|
528 |
select_lex->having_value= having_value; |
|
529 |
||
530 |
if (cond_value == Item::COND_FALSE || having_value == Item::COND_FALSE || |
|
531 |
(!unit->select_limit_cnt && !(select_options & OPTION_FOUND_ROWS))) |
|
532 |
{ /* Impossible cond */ |
|
533 |
zero_result_cause= having_value == Item::COND_FALSE ? |
|
534 |
"Impossible HAVING" : "Impossible WHERE"; |
|
535 |
error= 0; |
|
536 |
return(0); |
|
537 |
}
|
|
538 |
}
|
|
539 |
||
540 |
/* Optimize count(*), cmin() and cmax() */
|
|
541 |
if (tables_list && tmp_table_param.sum_func_count && ! group_list) |
|
542 |
{
|
|
543 |
int res; |
|
544 |
/*
|
|
545 |
opt_sum_query() returns HA_ERR_KEY_NOT_FOUND if no rows match
|
|
546 |
to the WHERE conditions,
|
|
547 |
or 1 if all items were resolved,
|
|
548 |
or 0, or an error number HA_ERR_...
|
|
549 |
*/
|
|
550 |
if ((res=opt_sum_query(select_lex->leaf_tables, all_fields, conds))) |
|
551 |
{
|
|
552 |
if (res == HA_ERR_KEY_NOT_FOUND) |
|
553 |
{
|
|
1039.2.7
by Jay Pipes
Yet more style and indentation cleanups. |
554 |
zero_result_cause= "No matching min/max row"; |
555 |
error=0; |
|
556 |
return(0); |
|
1039.2.2
by Jay Pipes
Phase 2 of JOIN refactoring. |
557 |
}
|
558 |
if (res > 1) |
|
559 |
{
|
|
560 |
error= res; |
|
561 |
return(1); |
|
562 |
}
|
|
563 |
if (res < 0) |
|
564 |
{
|
|
565 |
zero_result_cause= "No matching min/max row"; |
|
566 |
error=0; |
|
567 |
return(0); |
|
568 |
}
|
|
569 |
zero_result_cause= "Select tables optimized away"; |
|
570 |
tables_list= 0; // All tables resolved |
|
571 |
/*
|
|
572 |
Extract all table-independent conditions and replace the WHERE
|
|
573 |
clause with them. All other conditions were computed by opt_sum_query
|
|
574 |
and the MIN/MAX/COUNT function(s) have been replaced by constants,
|
|
575 |
so there is no need to compute the whole WHERE clause again.
|
|
576 |
Notice that make_cond_for_table() will always succeed to remove all
|
|
577 |
computed conditions, because opt_sum_query() is applicable only to
|
|
578 |
conjunctions.
|
|
579 |
Preserve conditions for EXPLAIN.
|
|
580 |
*/
|
|
581 |
if (conds && !(session->lex->describe & DESCRIBE_EXTENDED)) |
|
582 |
{
|
|
583 |
COND *table_independent_conds= make_cond_for_table(conds, PSEUDO_TABLE_BITS, 0, 0); |
|
584 |
conds= table_independent_conds; |
|
585 |
}
|
|
586 |
}
|
|
587 |
}
|
|
588 |
if (!tables_list) |
|
589 |
{
|
|
590 |
error= 0; |
|
591 |
return(0); |
|
592 |
}
|
|
593 |
error= -1; // Error is sent to client |
|
594 |
sort_by_table= get_sort_by_table(order, group_list, select_lex->leaf_tables); |
|
595 |
||
596 |
/* Calculate how to do the join */
|
|
597 |
session->set_proc_info("statistics"); |
|
598 |
if (make_join_statistics(this, select_lex->leaf_tables, conds, &keyuse) || |
|
599 |
session->is_fatal_error) |
|
600 |
{
|
|
601 |
return(1); |
|
602 |
}
|
|
603 |
||
604 |
/* Remove distinct if only const tables */
|
|
605 |
select_distinct= select_distinct && (const_tables != tables); |
|
606 |
session->set_proc_info("preparing"); |
|
607 |
if (result->initialize_tables(this)) |
|
608 |
{
|
|
609 |
return(1); // error == -1 |
|
610 |
}
|
|
611 |
if (const_table_map != found_const_table_map && |
|
612 |
!(select_options & SELECT_DESCRIBE) && |
|
613 |
(!conds || |
|
614 |
!(conds->used_tables() & RAND_TABLE_BIT) || |
|
615 |
select_lex->master_unit() == &session->lex->unit)) // upper level SELECT |
|
616 |
{
|
|
617 |
zero_result_cause= "no matching row in const table"; |
|
618 |
error= 0; |
|
619 |
return(0); |
|
620 |
}
|
|
621 |
if (!(session->options & OPTION_BIG_SELECTS) && |
|
622 |
best_read > (double) session->variables.max_join_size && |
|
623 |
!(select_options & SELECT_DESCRIBE)) |
|
624 |
{ /* purecov: inspected */ |
|
625 |
my_message(ER_TOO_BIG_SELECT, ER(ER_TOO_BIG_SELECT), MYF(0)); |
|
626 |
error= -1; |
|
627 |
return(1); |
|
628 |
}
|
|
1054.1.8
by Brian Aker
Remove lock_tables list from session. |
629 |
if (const_tables && !(select_options & SELECT_NO_UNLOCK)) |
1039.2.2
by Jay Pipes
Phase 2 of JOIN refactoring. |
630 |
mysql_unlock_some_tables(session, table, const_tables); |
631 |
if (!conds && outer_join) |
|
632 |
{
|
|
633 |
/* Handle the case where we have an OUTER JOIN without a WHERE */
|
|
634 |
conds=new Item_int((int64_t) 1,1); // Always true |
|
635 |
}
|
|
636 |
select= make_select(*table, const_table_map, |
|
637 |
const_table_map, conds, 1, &error); |
|
638 |
if (error) |
|
639 |
{ /* purecov: inspected */ |
|
640 |
error= -1; /* purecov: inspected */ |
|
641 |
return(1); |
|
642 |
}
|
|
643 |
||
644 |
reset_nj_counters(join_list); |
|
645 |
make_outerjoin_info(this); |
|
646 |
||
647 |
/*
|
|
648 |
Among the equal fields belonging to the same multiple equality
|
|
649 |
choose the one that is to be retrieved first and substitute
|
|
650 |
all references to these in where condition for a reference for
|
|
651 |
the selected field.
|
|
652 |
*/
|
|
653 |
if (conds) |
|
654 |
{
|
|
655 |
conds= substitute_for_best_equal_field(conds, cond_equal, map2table); |
|
656 |
conds->update_used_tables(); |
|
657 |
}
|
|
658 |
||
659 |
/*
|
|
660 |
Permorm the the optimization on fields evaluation mentioned above
|
|
661 |
for all on expressions.
|
|
662 |
*/
|
|
663 |
for (JOIN_TAB *tab= join_tab + const_tables; tab < join_tab + tables ; tab++) |
|
664 |
{
|
|
665 |
if (*tab->on_expr_ref) |
|
666 |
{
|
|
667 |
*tab->on_expr_ref= substitute_for_best_equal_field(*tab->on_expr_ref, |
|
668 |
tab->cond_equal, |
|
669 |
map2table); |
|
670 |
(*tab->on_expr_ref)->update_used_tables(); |
|
671 |
}
|
|
672 |
}
|
|
673 |
||
674 |
if (conds &&!outer_join && const_table_map != found_const_table_map && |
|
675 |
(select_options & SELECT_DESCRIBE) && |
|
676 |
select_lex->master_unit() == &session->lex->unit) // upper level SELECT |
|
677 |
{
|
|
678 |
conds=new Item_int((int64_t) 0,1); // Always false |
|
679 |
}
|
|
680 |
if (make_join_select(this, select, conds)) |
|
681 |
{
|
|
682 |
zero_result_cause= |
|
683 |
"Impossible WHERE noticed after reading const tables"; |
|
684 |
return(0); // error == 0 |
|
685 |
}
|
|
686 |
||
687 |
error= -1; /* if goto err */ |
|
688 |
||
689 |
/* Optimize distinct away if possible */
|
|
690 |
{
|
|
691 |
order_st *org_order= order; |
|
1039.2.7
by Jay Pipes
Yet more style and indentation cleanups. |
692 |
order= remove_constants(this, order,conds,1, &simple_order); |
1039.2.2
by Jay Pipes
Phase 2 of JOIN refactoring. |
693 |
if (session->is_error()) |
694 |
{
|
|
695 |
error= 1; |
|
696 |
return(1); |
|
697 |
}
|
|
698 |
||
699 |
/*
|
|
700 |
If we are using order_st BY NULL or order_st BY const_expression,
|
|
701 |
return result in any order (even if we are using a GROUP BY)
|
|
702 |
*/
|
|
703 |
if (!order && org_order) |
|
704 |
skip_sort_order= 1; |
|
705 |
}
|
|
706 |
/*
|
|
707 |
Check if we can optimize away GROUP BY/DISTINCT.
|
|
708 |
We can do that if there are no aggregate functions, the
|
|
709 |
fields in DISTINCT clause (if present) and/or columns in GROUP BY
|
|
710 |
(if present) contain direct references to all key parts of
|
|
711 |
an unique index (in whatever order) and if the key parts of the
|
|
712 |
unique index cannot contain NULLs.
|
|
713 |
Note that the unique keys for DISTINCT and GROUP BY should not
|
|
714 |
be the same (as long as they are unique).
|
|
715 |
||
716 |
The FROM clause must contain a single non-constant table.
|
|
717 |
*/
|
|
718 |
if (tables - const_tables == 1 && (group_list || select_distinct) && |
|
719 |
!tmp_table_param.sum_func_count && |
|
720 |
(!join_tab[const_tables].select || |
|
721 |
!join_tab[const_tables].select->quick || |
|
722 |
join_tab[const_tables].select->quick->get_type() != |
|
723 |
QUICK_SELECT_I::QS_TYPE_GROUP_MIN_MAX)) |
|
724 |
{
|
|
725 |
if (group_list && list_contains_unique_index(join_tab[const_tables].table, find_field_in_order_list, (void *) group_list)) |
|
726 |
{
|
|
727 |
/*
|
|
728 |
We have found that grouping can be removed since groups correspond to
|
|
729 |
only one row anyway, but we still have to guarantee correct result
|
|
730 |
order. The line below effectively rewrites the query from GROUP BY
|
|
731 |
<fields> to order_st BY <fields>. There are two exceptions:
|
|
732 |
- if skip_sort_order is set (see above), then we can simply skip
|
|
733 |
GROUP BY;
|
|
734 |
- we can only rewrite order_st BY if the order_st BY fields are 'compatible'
|
|
735 |
with the GROUP BY ones, i.e. either one is a prefix of another.
|
|
736 |
We only check if the order_st BY is a prefix of GROUP BY. In this case
|
|
737 |
test_if_subpart() copies the ASC/DESC attributes from the original
|
|
738 |
order_st BY fields.
|
|
739 |
If GROUP BY is a prefix of order_st BY, then it is safe to leave
|
|
740 |
'order' as is.
|
|
741 |
*/
|
|
742 |
if (!order || test_if_subpart(group_list, order)) |
|
743 |
order= skip_sort_order ? 0 : group_list; |
|
744 |
/*
|
|
745 |
If we have an IGNORE INDEX FOR GROUP BY(fields) clause, this must be
|
|
746 |
rewritten to IGNORE INDEX FOR order_st BY(fields).
|
|
747 |
*/
|
|
748 |
join_tab->table->keys_in_use_for_order_by= |
|
749 |
join_tab->table->keys_in_use_for_group_by; |
|
750 |
group_list= 0; |
|
751 |
group= 0; |
|
752 |
}
|
|
753 |
if (select_distinct && |
|
754 |
list_contains_unique_index(join_tab[const_tables].table, |
|
755 |
find_field_in_item_list, |
|
756 |
(void *) &fields_list)) |
|
757 |
{
|
|
758 |
select_distinct= 0; |
|
759 |
}
|
|
760 |
}
|
|
761 |
if (group_list || tmp_table_param.sum_func_count) |
|
762 |
{
|
|
763 |
if (! hidden_group_fields && rollup.state == ROLLUP::STATE_NONE) |
|
764 |
select_distinct=0; |
|
765 |
}
|
|
766 |
else if (select_distinct && tables - const_tables == 1) |
|
767 |
{
|
|
768 |
/*
|
|
769 |
We are only using one table. In this case we change DISTINCT to a
|
|
770 |
GROUP BY query if:
|
|
771 |
- The GROUP BY can be done through indexes (no sort) and the order_st
|
|
772 |
BY only uses selected fields.
|
|
1039.2.7
by Jay Pipes
Yet more style and indentation cleanups. |
773 |
(In this case we can later optimize away GROUP BY and order_st BY)
|
1039.2.2
by Jay Pipes
Phase 2 of JOIN refactoring. |
774 |
- We are scanning the whole table without LIMIT
|
775 |
This can happen if:
|
|
776 |
- We are using CALC_FOUND_ROWS
|
|
777 |
- We are using an order_st BY that can't be optimized away.
|
|
778 |
||
779 |
We don't want to use this optimization when we are using LIMIT
|
|
780 |
because in this case we can just create a temporary table that
|
|
781 |
holds LIMIT rows and stop when this table is full.
|
|
782 |
*/
|
|
783 |
JOIN_TAB *tab= &join_tab[const_tables]; |
|
784 |
bool all_order_fields_used; |
|
785 |
if (order) |
|
786 |
skip_sort_order= test_if_skip_sort_order(tab, order, select_limit, 1, |
|
787 |
&tab->table->keys_in_use_for_order_by); |
|
788 |
if ((group_list=create_distinct_group(session, select_lex->ref_pointer_array, |
|
789 |
order, fields_list, all_fields, |
|
790 |
&all_order_fields_used))) |
|
791 |
{
|
|
792 |
bool skip_group= (skip_sort_order && |
|
793 |
test_if_skip_sort_order(tab, group_list, select_limit, 1, |
|
794 |
&tab->table->keys_in_use_for_group_by) != 0); |
|
795 |
count_field_types(select_lex, &tmp_table_param, all_fields, 0); |
|
796 |
if ((skip_group && all_order_fields_used) || |
|
1039.2.7
by Jay Pipes
Yet more style and indentation cleanups. |
797 |
select_limit == HA_POS_ERROR || |
798 |
(order && !skip_sort_order)) |
|
1039.2.2
by Jay Pipes
Phase 2 of JOIN refactoring. |
799 |
{
|
1039.2.7
by Jay Pipes
Yet more style and indentation cleanups. |
800 |
/* Change DISTINCT to GROUP BY */
|
801 |
select_distinct= 0; |
|
802 |
no_order= !order; |
|
803 |
if (all_order_fields_used) |
|
804 |
{
|
|
805 |
if (order && skip_sort_order) |
|
806 |
{
|
|
807 |
/*
|
|
808 |
Force MySQL to read the table in sorted order to get result in
|
|
809 |
order_st BY order.
|
|
810 |
*/
|
|
811 |
tmp_table_param.quick_group=0; |
|
812 |
}
|
|
813 |
order=0; |
|
1039.2.2
by Jay Pipes
Phase 2 of JOIN refactoring. |
814 |
}
|
1039.2.7
by Jay Pipes
Yet more style and indentation cleanups. |
815 |
group=1; // For end_write_group |
1039.2.2
by Jay Pipes
Phase 2 of JOIN refactoring. |
816 |
}
|
817 |
else
|
|
1039.2.7
by Jay Pipes
Yet more style and indentation cleanups. |
818 |
group_list= 0; |
1039.2.2
by Jay Pipes
Phase 2 of JOIN refactoring. |
819 |
}
|
820 |
else if (session->is_fatal_error) // End of memory |
|
821 |
return(1); |
|
822 |
}
|
|
823 |
simple_group= 0; |
|
824 |
{
|
|
825 |
order_st *old_group_list; |
|
826 |
group_list= remove_constants(this, (old_group_list= group_list), conds, |
|
827 |
rollup.state == ROLLUP::STATE_NONE, |
|
828 |
&simple_group); |
|
829 |
if (session->is_error()) |
|
830 |
{
|
|
831 |
error= 1; |
|
832 |
return(1); |
|
833 |
}
|
|
834 |
if (old_group_list && !group_list) |
|
835 |
select_distinct= 0; |
|
836 |
}
|
|
837 |
if (!group_list && group) |
|
838 |
{
|
|
839 |
order=0; // The output has only one row |
|
840 |
simple_order=1; |
|
841 |
select_distinct= 0; // No need in distinct for 1 row |
|
842 |
group_optimized_away= 1; |
|
843 |
}
|
|
844 |
||
845 |
calc_group_buffer(this, group_list); |
|
846 |
send_group_parts= tmp_table_param.group_parts; /* Save org parts */ |
|
847 |
||
848 |
if (test_if_subpart(group_list, order) || |
|
849 |
(!group_list && tmp_table_param.sum_func_count)) |
|
850 |
order=0; |
|
851 |
||
852 |
// Can't use sort on head table if using row cache
|
|
853 |
if (full_join) |
|
854 |
{
|
|
855 |
if (group_list) |
|
856 |
simple_group=0; |
|
857 |
if (order) |
|
858 |
simple_order=0; |
|
859 |
}
|
|
860 |
||
861 |
/*
|
|
862 |
Check if we need to create a temporary table.
|
|
863 |
This has to be done if all tables are not already read (const tables)
|
|
864 |
and one of the following conditions holds:
|
|
865 |
- We are using DISTINCT (simple distinct's are already optimized away)
|
|
866 |
- We are using an order_st BY or GROUP BY on fields not in the first table
|
|
867 |
- We are using different order_st BY and GROUP BY orders
|
|
868 |
- The user wants us to buffer the result.
|
|
869 |
*/
|
|
870 |
need_tmp= (const_tables != tables && |
|
871 |
((select_distinct || !simple_order || !simple_group) || |
|
872 |
(group_list && order) || |
|
873 |
test(select_options & OPTION_BUFFER_RESULT))); |
|
874 |
||
875 |
uint32_t no_jbuf_after= make_join_orderinfo(this); |
|
876 |
uint64_t select_opts_for_readinfo= |
|
877 |
(select_options & (SELECT_DESCRIBE | SELECT_NO_JOIN_CACHE)) | (0); |
|
878 |
||
879 |
sj_tmp_tables= NULL; |
|
880 |
if (!select_lex->sj_nests.is_empty()) |
|
1039.2.7
by Jay Pipes
Yet more style and indentation cleanups. |
881 |
setup_semijoin_dups_elimination(this, select_opts_for_readinfo, no_jbuf_after); |
1039.2.2
by Jay Pipes
Phase 2 of JOIN refactoring. |
882 |
|
883 |
// No cache for MATCH == 'Don't use join buffering when we use MATCH'.
|
|
884 |
if (make_join_readinfo(this, select_opts_for_readinfo, no_jbuf_after)) |
|
885 |
return(1); |
|
886 |
||
887 |
/* Create all structures needed for materialized subquery execution. */
|
|
888 |
if (setup_subquery_materialization()) |
|
889 |
return(1); |
|
890 |
||
891 |
/*
|
|
892 |
is this simple IN subquery?
|
|
893 |
*/
|
|
894 |
if (!group_list && !order && |
|
895 |
unit->item && unit->item->substype() == Item_subselect::IN_SUBS && |
|
896 |
tables == 1 && conds && |
|
897 |
!unit->is_union()) |
|
898 |
{
|
|
899 |
if (!having) |
|
900 |
{
|
|
901 |
Item *where= conds; |
|
1039.2.7
by Jay Pipes
Yet more style and indentation cleanups. |
902 |
if (join_tab[0].type == JT_EQ_REF && join_tab[0].ref.items[0]->name == in_left_expr_name) |
1039.2.2
by Jay Pipes
Phase 2 of JOIN refactoring. |
903 |
{
|
904 |
remove_subq_pushed_predicates(&where); |
|
905 |
save_index_subquery_explain_info(join_tab, where); |
|
906 |
join_tab[0].type= JT_UNIQUE_SUBQUERY; |
|
907 |
error= 0; |
|
908 |
return(unit->item-> |
|
909 |
change_engine(new |
|
910 |
subselect_uniquesubquery_engine(session, |
|
911 |
join_tab, |
|
912 |
unit->item, |
|
913 |
where))); |
|
914 |
}
|
|
915 |
else if (join_tab[0].type == JT_REF && |
|
916 |
join_tab[0].ref.items[0]->name == in_left_expr_name) |
|
917 |
{
|
|
1039.2.7
by Jay Pipes
Yet more style and indentation cleanups. |
918 |
remove_subq_pushed_predicates(&where); |
1039.2.2
by Jay Pipes
Phase 2 of JOIN refactoring. |
919 |
save_index_subquery_explain_info(join_tab, where); |
920 |
join_tab[0].type= JT_INDEX_SUBQUERY; |
|
921 |
error= 0; |
|
922 |
return(unit->item-> |
|
923 |
change_engine(new |
|
924 |
subselect_indexsubquery_engine(session, |
|
925 |
join_tab, |
|
926 |
unit->item, |
|
927 |
where, |
|
928 |
NULL, |
|
929 |
0))); |
|
930 |
}
|
|
1039.2.7
by Jay Pipes
Yet more style and indentation cleanups. |
931 |
}
|
932 |
else if (join_tab[0].type == JT_REF_OR_NULL && |
|
1039.2.2
by Jay Pipes
Phase 2 of JOIN refactoring. |
933 |
join_tab[0].ref.items[0]->name == in_left_expr_name && |
934 |
having->name == in_having_cond) |
|
935 |
{
|
|
936 |
join_tab[0].type= JT_INDEX_SUBQUERY; |
|
937 |
error= 0; |
|
938 |
conds= remove_additional_cond(conds); |
|
939 |
save_index_subquery_explain_info(join_tab, conds); |
|
940 |
return(unit->item-> |
|
941 |
change_engine(new subselect_indexsubquery_engine(session, |
|
942 |
join_tab, |
|
943 |
unit->item, |
|
944 |
conds, |
|
945 |
having, |
|
946 |
1))); |
|
947 |
}
|
|
948 |
||
949 |
}
|
|
950 |
/*
|
|
951 |
Need to tell handlers that to play it safe, it should fetch all
|
|
952 |
columns of the primary key of the tables: this is because MySQL may
|
|
953 |
build row pointers for the rows, and for all columns of the primary key
|
|
954 |
the read set has not necessarily been set by the server code.
|
|
955 |
*/
|
|
956 |
if (need_tmp || select_distinct || group_list || order) |
|
957 |
{
|
|
958 |
for (uint32_t i = const_tables; i < tables; i++) |
|
959 |
join_tab[i].table->prepare_for_position(); |
|
960 |
}
|
|
961 |
||
962 |
if (const_tables != tables) |
|
963 |
{
|
|
964 |
/*
|
|
965 |
Because filesort always does a full table scan or a quick range scan
|
|
966 |
we must add the removed reference to the select for the table.
|
|
967 |
We only need to do this when we have a simple_order or simple_group
|
|
968 |
as in other cases the join is done before the sort.
|
|
969 |
*/
|
|
970 |
if ((order || group_list) && |
|
971 |
(join_tab[const_tables].type != JT_ALL) && |
|
972 |
(join_tab[const_tables].type != JT_REF_OR_NULL) && |
|
973 |
((order && simple_order) || (group_list && simple_group))) |
|
974 |
{
|
|
975 |
if (add_ref_to_table_cond(session,&join_tab[const_tables])) { |
|
976 |
return(1); |
|
977 |
}
|
|
978 |
}
|
|
979 |
||
980 |
if (!(select_options & SELECT_BIG_RESULT) && |
|
981 |
((group_list && |
|
982 |
(!simple_group || |
|
983 |
!test_if_skip_sort_order(&join_tab[const_tables], group_list, |
|
984 |
unit->select_limit_cnt, 0, |
|
985 |
&join_tab[const_tables].table-> |
|
986 |
keys_in_use_for_group_by))) || |
|
987 |
select_distinct) && |
|
988 |
tmp_table_param.quick_group) |
|
989 |
{
|
|
990 |
need_tmp=1; simple_order=simple_group=0; // Force tmp table without sort |
|
991 |
}
|
|
992 |
if (order) |
|
993 |
{
|
|
994 |
/*
|
|
995 |
Force using of tmp table if sorting by a SP or UDF function due to
|
|
996 |
their expensive and probably non-deterministic nature.
|
|
997 |
*/
|
|
998 |
for (order_st *tmp_order= order; tmp_order ; tmp_order=tmp_order->next) |
|
999 |
{
|
|
1000 |
Item *item= *tmp_order->item; |
|
1001 |
if (item->is_expensive()) |
|
1002 |
{
|
|
1003 |
/* Force tmp table without sort */
|
|
1004 |
need_tmp=1; simple_order=simple_group=0; |
|
1005 |
break; |
|
1006 |
}
|
|
1007 |
}
|
|
1008 |
}
|
|
1009 |
}
|
|
1010 |
||
1011 |
tmp_having= having; |
|
1012 |
if (select_options & SELECT_DESCRIBE) |
|
1013 |
{
|
|
1014 |
error= 0; |
|
1015 |
return(0); |
|
1016 |
}
|
|
1017 |
having= 0; |
|
1018 |
||
1019 |
/*
|
|
1020 |
The loose index scan access method guarantees that all grouping or
|
|
1021 |
duplicate row elimination (for distinct) is already performed
|
|
1022 |
during data retrieval, and that all MIN/MAX functions are already
|
|
1023 |
computed for each group. Thus all MIN/MAX functions should be
|
|
1024 |
treated as regular functions, and there is no need to perform
|
|
1025 |
grouping in the main execution loop.
|
|
1026 |
Notice that currently loose index scan is applicable only for
|
|
1027 |
single table queries, thus it is sufficient to test only the first
|
|
1028 |
join_tab element of the plan for its access method.
|
|
1029 |
*/
|
|
1030 |
if (join_tab->is_using_loose_index_scan()) |
|
1031 |
tmp_table_param.precomputed_group_by= true; |
|
1032 |
||
1033 |
/* Create a tmp table if distinct or if the sort is too complicated */
|
|
1034 |
if (need_tmp) |
|
1035 |
{
|
|
1036 |
session->set_proc_info("Creating tmp table"); |
|
1037 |
||
1038 |
init_items_ref_array(); |
|
1039 |
||
1040 |
tmp_table_param.hidden_field_count= (all_fields.elements - |
|
1041 |
fields_list.elements); |
|
1042 |
order_st *tmp_group= ((!simple_group && !(test_flags & TEST_NO_KEY_GROUP)) ? group_list : |
|
1043 |
(order_st*) 0); |
|
1044 |
/*
|
|
1045 |
Pushing LIMIT to the temporary table creation is not applicable
|
|
1046 |
when there is order_st BY or GROUP BY or there is no GROUP BY, but
|
|
1047 |
there are aggregate functions, because in all these cases we need
|
|
1048 |
all result rows.
|
|
1049 |
*/
|
|
1050 |
ha_rows tmp_rows_limit= ((order == 0 || skip_sort_order) && |
|
1051 |
!tmp_group && |
|
1052 |
!session->lex->current_select->with_sum_func) ? |
|
1053 |
select_limit : HA_POS_ERROR; |
|
1054 |
||
1055 |
if (!(exec_tmp_table1= |
|
1056 |
create_tmp_table(session, &tmp_table_param, all_fields, |
|
1057 |
tmp_group, |
|
1058 |
group_list ? 0 : select_distinct, |
|
1059 |
group_list && simple_group, |
|
1060 |
select_options, |
|
1061 |
tmp_rows_limit, |
|
1062 |
(char *) ""))) |
|
1063 |
{
|
|
1064 |
return(1); |
|
1065 |
}
|
|
1066 |
||
1067 |
/*
|
|
1068 |
We don't have to store rows in temp table that doesn't match HAVING if:
|
|
1069 |
- we are sorting the table and writing complete group rows to the
|
|
1070 |
temp table.
|
|
1071 |
- We are using DISTINCT without resolving the distinct as a GROUP BY
|
|
1072 |
on all columns.
|
|
1073 |
||
1074 |
If having is not handled here, it will be checked before the row
|
|
1075 |
is sent to the client.
|
|
1076 |
*/
|
|
1039.2.7
by Jay Pipes
Yet more style and indentation cleanups. |
1077 |
if (tmp_having && (sort_and_group || (exec_tmp_table1->distinct && !group_list))) |
1039.2.2
by Jay Pipes
Phase 2 of JOIN refactoring. |
1078 |
having= tmp_having; |
1079 |
||
1080 |
/* if group or order on first table, sort first */
|
|
1081 |
if (group_list && simple_group) |
|
1082 |
{
|
|
1083 |
session->set_proc_info("Sorting for group"); |
|
1084 |
if (create_sort_index(session, this, group_list, |
|
1085 |
HA_POS_ERROR, HA_POS_ERROR, false) || |
|
1039.2.7
by Jay Pipes
Yet more style and indentation cleanups. |
1086 |
alloc_group_fields(this, group_list) || |
1039.2.2
by Jay Pipes
Phase 2 of JOIN refactoring. |
1087 |
make_sum_func_list(all_fields, fields_list, 1) || |
1088 |
setup_sum_funcs(session, sum_funcs)) |
|
1089 |
{
|
|
1090 |
return(1); |
|
1091 |
}
|
|
1092 |
group_list=0; |
|
1093 |
}
|
|
1094 |
else
|
|
1095 |
{
|
|
1096 |
if (make_sum_func_list(all_fields, fields_list, 0) || |
|
1097 |
setup_sum_funcs(session, sum_funcs)) |
|
1098 |
{
|
|
1099 |
return(1); |
|
1100 |
}
|
|
1101 |
||
1102 |
if (!group_list && ! exec_tmp_table1->distinct && order && simple_order) |
|
1103 |
{
|
|
1104 |
session->set_proc_info("Sorting for order"); |
|
1105 |
if (create_sort_index(session, this, order, |
|
1106 |
HA_POS_ERROR, HA_POS_ERROR, true)) |
|
1107 |
{
|
|
1108 |
return(1); |
|
1109 |
}
|
|
1110 |
order=0; |
|
1111 |
}
|
|
1112 |
}
|
|
1113 |
||
1114 |
/*
|
|
1115 |
Optimize distinct when used on some of the tables
|
|
1116 |
SELECT DISTINCT t1.a FROM t1,t2 WHERE t1.b=t2.b
|
|
1117 |
In this case we can stop scanning t2 when we have found one t1.a
|
|
1118 |
*/
|
|
1119 |
||
1120 |
if (exec_tmp_table1->distinct) |
|
1121 |
{
|
|
1122 |
table_map used_tables= session->used_tables; |
|
1123 |
JOIN_TAB *last_join_tab= join_tab+tables-1; |
|
1124 |
do
|
|
1125 |
{
|
|
1039.2.7
by Jay Pipes
Yet more style and indentation cleanups. |
1126 |
if (used_tables & last_join_tab->table->map) |
1127 |
break; |
|
1128 |
last_join_tab->not_used_in_distinct=1; |
|
1039.2.2
by Jay Pipes
Phase 2 of JOIN refactoring. |
1129 |
} while (last_join_tab-- != join_tab); |
1130 |
/* Optimize "select distinct b from t1 order by key_part_1 limit #" */
|
|
1131 |
if (order && skip_sort_order) |
|
1132 |
{
|
|
1039.2.7
by Jay Pipes
Yet more style and indentation cleanups. |
1133 |
/* Should always succeed */
|
1134 |
if (test_if_skip_sort_order(&join_tab[const_tables], |
|
1135 |
order, unit->select_limit_cnt, 0, |
|
1136 |
&join_tab[const_tables].table-> |
|
1137 |
keys_in_use_for_order_by)) |
|
1138 |
order= 0; |
|
1039.2.2
by Jay Pipes
Phase 2 of JOIN refactoring. |
1139 |
}
|
1140 |
}
|
|
1141 |
||
1142 |
/*
|
|
1143 |
If this join belongs to an uncacheable subquery save
|
|
1144 |
the original join
|
|
1145 |
*/
|
|
1146 |
if (select_lex->uncacheable && !is_top_level_join() && |
|
1147 |
init_save_join_tab()) |
|
1148 |
return(-1); /* purecov: inspected */ |
|
1149 |
}
|
|
1150 |
||
1151 |
error= 0; |
|
1152 |
return(0); |
|
1153 |
}
|
|
1154 |
||
1155 |
/**
|
|
1156 |
Restore values in temporary join.
|
|
1157 |
*/
|
|
1158 |
void JOIN::restore_tmp() |
|
1159 |
{
|
|
1160 |
memcpy(tmp_join, this, (size_t) sizeof(JOIN)); |
|
1161 |
}
|
|
1162 |
||
1039.2.7
by Jay Pipes
Yet more style and indentation cleanups. |
1163 |
int JOIN::reinit() |
1039.2.2
by Jay Pipes
Phase 2 of JOIN refactoring. |
1164 |
{
|
1165 |
unit->offset_limit_cnt= (ha_rows)(select_lex->offset_limit ? |
|
1166 |
select_lex->offset_limit->val_uint() : |
|
1167 |
0UL); |
|
1168 |
||
1169 |
first_record= 0; |
|
1170 |
||
1171 |
if (exec_tmp_table1) |
|
1172 |
{
|
|
1173 |
exec_tmp_table1->file->extra(HA_EXTRA_RESET_STATE); |
|
1174 |
exec_tmp_table1->file->ha_delete_all_rows(); |
|
1175 |
free_io_cache(exec_tmp_table1); |
|
1176 |
filesort_free_buffers(exec_tmp_table1,0); |
|
1177 |
}
|
|
1178 |
if (exec_tmp_table2) |
|
1179 |
{
|
|
1180 |
exec_tmp_table2->file->extra(HA_EXTRA_RESET_STATE); |
|
1181 |
exec_tmp_table2->file->ha_delete_all_rows(); |
|
1182 |
free_io_cache(exec_tmp_table2); |
|
1183 |
filesort_free_buffers(exec_tmp_table2,0); |
|
1184 |
}
|
|
1185 |
if (items0) |
|
1186 |
set_items_ref_array(items0); |
|
1187 |
||
1188 |
if (join_tab_save) |
|
1189 |
memcpy(join_tab, join_tab_save, sizeof(JOIN_TAB) * tables); |
|
1190 |
||
1191 |
if (tmp_join) |
|
1192 |
restore_tmp(); |
|
1193 |
||
1194 |
/* Reset of sum functions */
|
|
1195 |
if (sum_funcs) |
|
1196 |
{
|
|
1197 |
Item_sum *func, **func_ptr= sum_funcs; |
|
1198 |
while ((func= *(func_ptr++))) |
|
1199 |
func->clear(); |
|
1200 |
}
|
|
1201 |
||
1202 |
return(0); |
|
1203 |
}
|
|
1204 |
||
1205 |
/**
|
|
1206 |
@brief Save the original join layout
|
|
1207 |
||
1208 |
@details Saves the original join layout so it can be reused in
|
|
1209 |
re-execution and for EXPLAIN.
|
|
1210 |
||
1211 |
@return Operation status
|
|
1212 |
@retval 0 success.
|
|
1213 |
@retval 1 error occurred.
|
|
1214 |
*/
|
|
1039.2.7
by Jay Pipes
Yet more style and indentation cleanups. |
1215 |
bool JOIN::init_save_join_tab() |
1039.2.2
by Jay Pipes
Phase 2 of JOIN refactoring. |
1216 |
{
|
1217 |
if (!(tmp_join= (JOIN*)session->alloc(sizeof(JOIN)))) |
|
1218 |
return 1; /* purecov: inspected */ |
|
1219 |
error= 0; // Ensure that tmp_join.error= 0 |
|
1220 |
restore_tmp(); |
|
1221 |
return 0; |
|
1222 |
}
|
|
1223 |
||
1039.2.7
by Jay Pipes
Yet more style and indentation cleanups. |
1224 |
bool JOIN::save_join_tab() |
1039.2.2
by Jay Pipes
Phase 2 of JOIN refactoring. |
1225 |
{
|
1226 |
if (!join_tab_save && select_lex->master_unit()->uncacheable) |
|
1227 |
{
|
|
1228 |
if (!(join_tab_save= (JOIN_TAB*)session->memdup((unsigned char*) join_tab, |
|
1229 |
sizeof(JOIN_TAB) * tables))) |
|
1230 |
return 1; |
|
1231 |
}
|
|
1232 |
return 0; |
|
1233 |
}
|
|
1234 |
||
1235 |
/**
|
|
1236 |
Exec select.
|
|
1237 |
||
1238 |
@todo
|
|
1239 |
Note, that create_sort_index calls test_if_skip_sort_order and may
|
|
1240 |
finally replace sorting with index scan if there is a LIMIT clause in
|
|
1241 |
the query. It's never shown in EXPLAIN!
|
|
1242 |
||
1243 |
@todo
|
|
1244 |
When can we have here session->net.report_error not zero?
|
|
1245 |
*/
|
|
1246 |
void JOIN::exec() |
|
1247 |
{
|
|
1248 |
List<Item> *columns_list= &fields_list; |
|
1249 |
int tmp_error; |
|
1250 |
||
1251 |
session->set_proc_info("executing"); |
|
1252 |
error= 0; |
|
1253 |
||
1254 |
if (!tables_list && (tables || !select_lex->with_sum_func)) |
|
1255 |
{
|
|
1256 |
/* Only test of functions */
|
|
1257 |
if (select_options & SELECT_DESCRIBE) |
|
1258 |
select_describe(this, false, false, false, (zero_result_cause?zero_result_cause:"No tables used")); |
|
1259 |
else
|
|
1260 |
{
|
|
1261 |
result->send_fields(*columns_list, Protocol::SEND_NUM_ROWS | Protocol::SEND_EOF); |
|
1262 |
/*
|
|
1263 |
We have to test for 'conds' here as the WHERE may not be constant
|
|
1264 |
even if we don't have any tables for prepared statements or if
|
|
1265 |
conds uses something like 'rand()'.
|
|
1266 |
*/
|
|
1267 |
if (cond_value != Item::COND_FALSE && |
|
1268 |
(!conds || conds->val_int()) && |
|
1269 |
(!having || having->val_int())) |
|
1270 |
{
|
|
1271 |
if (do_send_rows && result->send_data(fields_list)) |
|
1272 |
error= 1; |
|
1273 |
else
|
|
1274 |
{
|
|
1275 |
error= (int) result->send_eof(); |
|
1276 |
send_records= ((select_options & OPTION_FOUND_ROWS) ? 1 : session->sent_row_count); |
|
1277 |
}
|
|
1278 |
}
|
|
1279 |
else
|
|
1280 |
{
|
|
1281 |
error= (int) result->send_eof(); |
|
1282 |
send_records= 0; |
|
1283 |
}
|
|
1284 |
}
|
|
1285 |
/* Single select (without union) always returns 0 or 1 row */
|
|
1286 |
session->limit_found_rows= send_records; |
|
1287 |
session->examined_row_count= 0; |
|
1288 |
return; |
|
1289 |
}
|
|
1290 |
/*
|
|
1291 |
Don't reset the found rows count if there're no tables as
|
|
1292 |
FOUND_ROWS() may be called. Never reset the examined row count here.
|
|
1293 |
It must be accumulated from all join iterations of all join parts.
|
|
1294 |
*/
|
|
1295 |
if (tables) |
|
1296 |
session->limit_found_rows= 0; |
|
1297 |
||
1298 |
if (zero_result_cause) |
|
1299 |
{
|
|
1300 |
(void) return_zero_rows(this, result, select_lex->leaf_tables, |
|
1301 |
*columns_list, |
|
1302 |
send_row_on_empty_set(), |
|
1303 |
select_options, |
|
1304 |
zero_result_cause, |
|
1305 |
having); |
|
1306 |
return; |
|
1307 |
}
|
|
1308 |
||
1309 |
if ((this->select_lex->options & OPTION_SCHEMA_TABLE) && get_schema_tables_result(this, PROCESSED_BY_JOIN_EXEC)) |
|
1310 |
return; |
|
1311 |
||
1312 |
if (select_options & SELECT_DESCRIBE) |
|
1313 |
{
|
|
1314 |
/*
|
|
1315 |
Check if we managed to optimize order_st BY away and don't use temporary
|
|
1316 |
table to resolve order_st BY: in that case, we only may need to do
|
|
1317 |
filesort for GROUP BY.
|
|
1318 |
*/
|
|
1319 |
if (!order && !no_order && (!skip_sort_order || !need_tmp)) |
|
1320 |
{
|
|
1321 |
/* Reset 'order' to 'group_list' and reinit variables describing 'order' */
|
|
1322 |
order= group_list; |
|
1323 |
simple_order= simple_group; |
|
1324 |
skip_sort_order= 0; |
|
1325 |
}
|
|
1326 |
if (order && (order != group_list || !(select_options & SELECT_BIG_RESULT))) |
|
1327 |
{
|
|
1328 |
if (const_tables == tables |
|
1329 |
|| ((simple_order || skip_sort_order) |
|
1330 |
&& test_if_skip_sort_order(&join_tab[const_tables], order, select_limit, 0, &join_tab[const_tables].table->keys_in_use_for_query))) |
|
1331 |
order= 0; |
|
1332 |
}
|
|
1333 |
having= tmp_having; |
|
1334 |
select_describe(this, need_tmp, order != 0 && !skip_sort_order, select_distinct, !tables ? "No tables used" : NULL); |
|
1335 |
return; |
|
1336 |
}
|
|
1337 |
||
1338 |
JOIN *curr_join= this; |
|
1339 |
List<Item> *curr_all_fields= &all_fields; |
|
1340 |
List<Item> *curr_fields_list= &fields_list; |
|
1341 |
Table *curr_tmp_table= 0; |
|
1342 |
/*
|
|
1343 |
Initialize examined rows here because the values from all join parts
|
|
1344 |
must be accumulated in examined_row_count. Hence every join
|
|
1345 |
iteration must count from zero.
|
|
1346 |
*/
|
|
1347 |
curr_join->examined_rows= 0; |
|
1348 |
||
1349 |
/* Create a tmp table if distinct or if the sort is too complicated */
|
|
1350 |
if (need_tmp) |
|
1351 |
{
|
|
1352 |
if (tmp_join) |
|
1353 |
{
|
|
1354 |
/*
|
|
1355 |
We are in a non cacheable sub query. Get the saved join structure
|
|
1356 |
after optimization.
|
|
1357 |
(curr_join may have been modified during last exection and we need
|
|
1358 |
to reset it)
|
|
1359 |
*/
|
|
1360 |
curr_join= tmp_join; |
|
1361 |
}
|
|
1362 |
curr_tmp_table= exec_tmp_table1; |
|
1363 |
||
1364 |
/* Copy data to the temporary table */
|
|
1365 |
session->set_proc_info("Copying to tmp table"); |
|
1366 |
if (! curr_join->sort_and_group && curr_join->const_tables != curr_join->tables) |
|
1367 |
curr_join->join_tab[curr_join->const_tables].sorted= 0; |
|
1368 |
if ((tmp_error= do_select(curr_join, (List<Item> *) 0, curr_tmp_table))) |
|
1369 |
{
|
|
1370 |
error= tmp_error; |
|
1371 |
return; |
|
1372 |
}
|
|
1373 |
curr_tmp_table->file->info(HA_STATUS_VARIABLE); |
|
1374 |
||
1375 |
if (curr_join->having) |
|
1376 |
curr_join->having= curr_join->tmp_having= 0; // Allready done |
|
1377 |
||
1378 |
/* Change sum_fields reference to calculated fields in tmp_table */
|
|
1379 |
curr_join->all_fields= *curr_all_fields; |
|
1380 |
if (!items1) |
|
1381 |
{
|
|
1382 |
items1= items0 + all_fields.elements; |
|
1383 |
if (sort_and_group || curr_tmp_table->group) |
|
1384 |
{
|
|
1385 |
if (change_to_use_tmp_fields(session, items1, |
|
1386 |
tmp_fields_list1, tmp_all_fields1, |
|
1387 |
fields_list.elements, all_fields)) |
|
1388 |
return; |
|
1389 |
}
|
|
1390 |
else
|
|
1391 |
{
|
|
1392 |
if (change_refs_to_tmp_fields(session, items1, |
|
1393 |
tmp_fields_list1, tmp_all_fields1, |
|
1394 |
fields_list.elements, all_fields)) |
|
1395 |
return; |
|
1396 |
}
|
|
1397 |
curr_join->tmp_all_fields1= tmp_all_fields1; |
|
1398 |
curr_join->tmp_fields_list1= tmp_fields_list1; |
|
1399 |
curr_join->items1= items1; |
|
1400 |
}
|
|
1401 |
curr_all_fields= &tmp_all_fields1; |
|
1402 |
curr_fields_list= &tmp_fields_list1; |
|
1403 |
curr_join->set_items_ref_array(items1); |
|
1404 |
||
1405 |
if (sort_and_group || curr_tmp_table->group) |
|
1406 |
{
|
|
1407 |
curr_join->tmp_table_param.field_count+= curr_join->tmp_table_param.sum_func_count |
|
1408 |
+ curr_join->tmp_table_param.func_count; |
|
1409 |
curr_join->tmp_table_param.sum_func_count= 0; |
|
1410 |
curr_join->tmp_table_param.func_count= 0; |
|
1411 |
}
|
|
1412 |
else
|
|
1413 |
{
|
|
1414 |
curr_join->tmp_table_param.field_count+= curr_join->tmp_table_param.func_count; |
|
1415 |
curr_join->tmp_table_param.func_count= 0; |
|
1416 |
}
|
|
1417 |
||
1418 |
if (curr_tmp_table->group) |
|
1419 |
{ // Already grouped |
|
1420 |
if (!curr_join->order && !curr_join->no_order && !skip_sort_order) |
|
1421 |
curr_join->order= curr_join->group_list; /* order by group */ |
|
1422 |
curr_join->group_list= 0; |
|
1423 |
}
|
|
1424 |
||
1425 |
/*
|
|
1426 |
If we have different sort & group then we must sort the data by group
|
|
1427 |
and copy it to another tmp table
|
|
1428 |
This code is also used if we are using distinct something
|
|
1429 |
we haven't been able to store in the temporary table yet
|
|
1430 |
like SEC_TO_TIME(SUM(...)).
|
|
1431 |
*/
|
|
1432 |
||
1433 |
if ((curr_join->group_list && (!test_if_subpart(curr_join->group_list, curr_join->order) || curr_join->select_distinct)) |
|
1434 |
|| (curr_join->select_distinct && curr_join->tmp_table_param.using_indirect_summary_function)) |
|
1435 |
{ /* Must copy to another table */ |
|
1436 |
/* Free first data from old join */
|
|
1437 |
curr_join->join_free(); |
|
1438 |
if (make_simple_join(curr_join, curr_tmp_table)) |
|
1439 |
return; |
|
1440 |
calc_group_buffer(curr_join, group_list); |
|
1441 |
count_field_types(select_lex, &curr_join->tmp_table_param, |
|
1442 |
curr_join->tmp_all_fields1, |
|
1443 |
curr_join->select_distinct && !curr_join->group_list); |
|
1444 |
curr_join->tmp_table_param.hidden_field_count= curr_join->tmp_all_fields1.elements |
|
1445 |
- curr_join->tmp_fields_list1.elements; |
|
1446 |
||
1447 |
if (exec_tmp_table2) |
|
1448 |
curr_tmp_table= exec_tmp_table2; |
|
1449 |
else
|
|
1450 |
{
|
|
1451 |
/* group data to new table */
|
|
1452 |
||
1453 |
/*
|
|
1454 |
If the access method is loose index scan then all MIN/MAX
|
|
1455 |
functions are precomputed, and should be treated as regular
|
|
1456 |
functions. See extended comment in JOIN::exec.
|
|
1457 |
*/
|
|
1458 |
if (curr_join->join_tab->is_using_loose_index_scan()) |
|
1459 |
curr_join->tmp_table_param.precomputed_group_by= true; |
|
1460 |
||
1461 |
if (!(curr_tmp_table= |
|
1462 |
exec_tmp_table2= create_tmp_table(session, |
|
1463 |
&curr_join->tmp_table_param, |
|
1464 |
*curr_all_fields, |
|
1465 |
(order_st*) 0, |
|
1466 |
curr_join->select_distinct && |
|
1467 |
!curr_join->group_list, |
|
1468 |
1, curr_join->select_options, |
|
1469 |
HA_POS_ERROR, |
|
1470 |
(char *) ""))) |
|
1471 |
return; |
|
1472 |
curr_join->exec_tmp_table2= exec_tmp_table2; |
|
1473 |
}
|
|
1474 |
if (curr_join->group_list) |
|
1475 |
{
|
|
1476 |
session->set_proc_info("Creating sort index"); |
|
1477 |
if (curr_join->join_tab == join_tab && save_join_tab()) |
|
1478 |
{
|
|
1479 |
return; |
|
1480 |
}
|
|
1481 |
if (create_sort_index(session, curr_join, curr_join->group_list, |
|
1482 |
HA_POS_ERROR, HA_POS_ERROR, false) || |
|
1483 |
make_group_fields(this, curr_join)) |
|
1484 |
{
|
|
1485 |
return; |
|
1486 |
}
|
|
1487 |
sortorder= curr_join->sortorder; |
|
1488 |
}
|
|
1489 |
||
1490 |
session->set_proc_info("Copying to group table"); |
|
1491 |
tmp_error= -1; |
|
1492 |
if (curr_join != this) |
|
1493 |
{
|
|
1494 |
if (sum_funcs2) |
|
1495 |
{
|
|
1496 |
curr_join->sum_funcs= sum_funcs2; |
|
1497 |
curr_join->sum_funcs_end= sum_funcs_end2; |
|
1498 |
}
|
|
1499 |
else
|
|
1500 |
{
|
|
1501 |
curr_join->alloc_func_list(); |
|
1502 |
sum_funcs2= curr_join->sum_funcs; |
|
1503 |
sum_funcs_end2= curr_join->sum_funcs_end; |
|
1504 |
}
|
|
1505 |
}
|
|
1506 |
if (curr_join->make_sum_func_list(*curr_all_fields, *curr_fields_list, 1, true)) |
|
1507 |
return; |
|
1508 |
curr_join->group_list= 0; |
|
1509 |
||
1510 |
if (!curr_join->sort_and_group && (curr_join->const_tables != curr_join->tables)) |
|
1511 |
curr_join->join_tab[curr_join->const_tables].sorted= 0; |
|
1512 |
||
1513 |
if (setup_sum_funcs(curr_join->session, curr_join->sum_funcs) |
|
1514 |
|| (tmp_error= do_select(curr_join, (List<Item> *) 0, curr_tmp_table))) |
|
1515 |
{
|
|
1516 |
error= tmp_error; |
|
1517 |
return; |
|
1518 |
}
|
|
1519 |
end_read_record(&curr_join->join_tab->read_record); |
|
1520 |
curr_join->const_tables= curr_join->tables; // Mark free for cleanup() |
|
1521 |
curr_join->join_tab[0].table= 0; // Table is freed |
|
1522 |
||
1523 |
// No sum funcs anymore
|
|
1524 |
if (!items2) |
|
1525 |
{
|
|
1526 |
items2= items1 + all_fields.elements; |
|
1527 |
if (change_to_use_tmp_fields(session, items2, |
|
1528 |
tmp_fields_list2, tmp_all_fields2, |
|
1529 |
fields_list.elements, tmp_all_fields1)) |
|
1530 |
return; |
|
1531 |
curr_join->tmp_fields_list2= tmp_fields_list2; |
|
1532 |
curr_join->tmp_all_fields2= tmp_all_fields2; |
|
1533 |
}
|
|
1534 |
curr_fields_list= &curr_join->tmp_fields_list2; |
|
1535 |
curr_all_fields= &curr_join->tmp_all_fields2; |
|
1536 |
curr_join->set_items_ref_array(items2); |
|
1537 |
curr_join->tmp_table_param.field_count+= curr_join->tmp_table_param.sum_func_count; |
|
1538 |
curr_join->tmp_table_param.sum_func_count= 0; |
|
1539 |
}
|
|
1540 |
if (curr_tmp_table->distinct) |
|
1541 |
curr_join->select_distinct=0; /* Each row is unique */ |
|
1542 |
||
1543 |
curr_join->join_free(); /* Free quick selects */ |
|
1544 |
if (curr_join->select_distinct && ! curr_join->group_list) |
|
1545 |
{
|
|
1546 |
session->set_proc_info("Removing duplicates"); |
|
1547 |
if (curr_join->tmp_having) |
|
1548 |
curr_join->tmp_having->update_used_tables(); |
|
1549 |
||
1550 |
if (remove_duplicates(curr_join, curr_tmp_table, |
|
1551 |
*curr_fields_list, curr_join->tmp_having)) |
|
1552 |
return; |
|
1553 |
||
1554 |
curr_join->tmp_having=0; |
|
1555 |
curr_join->select_distinct=0; |
|
1556 |
}
|
|
1557 |
curr_tmp_table->reginfo.lock_type= TL_UNLOCK; |
|
1558 |
if (make_simple_join(curr_join, curr_tmp_table)) |
|
1559 |
return; |
|
1560 |
calc_group_buffer(curr_join, curr_join->group_list); |
|
1561 |
count_field_types(select_lex, &curr_join->tmp_table_param, *curr_all_fields, 0); |
|
1562 |
||
1563 |
}
|
|
1564 |
||
1565 |
if (curr_join->group || curr_join->tmp_table_param.sum_func_count) |
|
1566 |
{
|
|
1567 |
if (make_group_fields(this, curr_join)) |
|
1568 |
return; |
|
1569 |
||
1570 |
if (! items3) |
|
1571 |
{
|
|
1572 |
if (! items0) |
|
1573 |
init_items_ref_array(); |
|
1574 |
items3= ref_pointer_array + (all_fields.elements*4); |
|
1575 |
setup_copy_fields(session, &curr_join->tmp_table_param, |
|
1576 |
items3, tmp_fields_list3, tmp_all_fields3, |
|
1577 |
curr_fields_list->elements, *curr_all_fields); |
|
1578 |
tmp_table_param.save_copy_funcs= curr_join->tmp_table_param.copy_funcs; |
|
1579 |
tmp_table_param.save_copy_field= curr_join->tmp_table_param.copy_field; |
|
1580 |
tmp_table_param.save_copy_field_end= curr_join->tmp_table_param.copy_field_end; |
|
1581 |
curr_join->tmp_all_fields3= tmp_all_fields3; |
|
1582 |
curr_join->tmp_fields_list3= tmp_fields_list3; |
|
1583 |
}
|
|
1584 |
else
|
|
1585 |
{
|
|
1586 |
curr_join->tmp_table_param.copy_funcs= tmp_table_param.save_copy_funcs; |
|
1587 |
curr_join->tmp_table_param.copy_field= tmp_table_param.save_copy_field; |
|
1588 |
curr_join->tmp_table_param.copy_field_end= tmp_table_param.save_copy_field_end; |
|
1589 |
}
|
|
1590 |
curr_fields_list= &tmp_fields_list3; |
|
1591 |
curr_all_fields= &tmp_all_fields3; |
|
1592 |
curr_join->set_items_ref_array(items3); |
|
1593 |
||
1594 |
if (curr_join->make_sum_func_list(*curr_all_fields, *curr_fields_list, |
|
1595 |
1, true) || |
|
1596 |
setup_sum_funcs(curr_join->session, curr_join->sum_funcs) || |
|
1597 |
session->is_fatal_error) |
|
1598 |
return; |
|
1599 |
}
|
|
1600 |
if (curr_join->group_list || curr_join->order) |
|
1601 |
{
|
|
1602 |
session->set_proc_info("Sorting result"); |
|
1603 |
/* If we have already done the group, add HAVING to sorted table */
|
|
1604 |
if (curr_join->tmp_having && ! curr_join->group_list && ! curr_join->sort_and_group) |
|
1605 |
{
|
|
1606 |
// Some tables may have been const
|
|
1607 |
curr_join->tmp_having->update_used_tables(); |
|
1608 |
JOIN_TAB *curr_table= &curr_join->join_tab[curr_join->const_tables]; |
|
1609 |
table_map used_tables= (curr_join->const_table_map | |
|
1610 |
curr_table->table->map); |
|
1611 |
||
1612 |
Item* sort_table_cond= make_cond_for_table(curr_join->tmp_having, used_tables, used_tables, 0); |
|
1613 |
if (sort_table_cond) |
|
1614 |
{
|
|
1615 |
if (!curr_table->select) |
|
1616 |
if (!(curr_table->select= new SQL_SELECT)) |
|
1617 |
return; |
|
1618 |
if (!curr_table->select->cond) |
|
1619 |
curr_table->select->cond= sort_table_cond; |
|
1620 |
else // This should never happen |
|
1621 |
{
|
|
1622 |
if (!(curr_table->select->cond= |
|
1623 |
new Item_cond_and(curr_table->select->cond, |
|
1624 |
sort_table_cond))) |
|
1625 |
return; |
|
1626 |
/*
|
|
1627 |
Item_cond_and do not need fix_fields for execution, its parameters
|
|
1628 |
are fixed or do not need fix_fields, too
|
|
1629 |
*/
|
|
1630 |
curr_table->select->cond->quick_fix_field(); |
|
1631 |
}
|
|
1632 |
curr_table->select_cond= curr_table->select->cond; |
|
1633 |
curr_table->select_cond->top_level_item(); |
|
1634 |
curr_join->tmp_having= make_cond_for_table(curr_join->tmp_having, |
|
1635 |
~ (table_map) 0, |
|
1636 |
~used_tables, 0); |
|
1637 |
}
|
|
1638 |
}
|
|
1639 |
{
|
|
1640 |
if (group) |
|
1641 |
curr_join->select_limit= HA_POS_ERROR; |
|
1642 |
else
|
|
1643 |
{
|
|
1644 |
/*
|
|
1645 |
We can abort sorting after session->select_limit rows if we there is no
|
|
1646 |
WHERE clause for any tables after the sorted one.
|
|
1647 |
*/
|
|
1648 |
JOIN_TAB *curr_table= &curr_join->join_tab[curr_join->const_tables+1]; |
|
1649 |
JOIN_TAB *end_table= &curr_join->join_tab[curr_join->tables]; |
|
1650 |
for (; curr_table < end_table ; curr_table++) |
|
1651 |
{
|
|
1652 |
/*
|
|
1653 |
table->keyuse is set in the case there was an original WHERE clause
|
|
1654 |
on the table that was optimized away.
|
|
1655 |
*/
|
|
1656 |
if (curr_table->select_cond || |
|
1657 |
(curr_table->keyuse && !curr_table->first_inner)) |
|
1658 |
{
|
|
1659 |
/* We have to sort all rows */
|
|
1660 |
curr_join->select_limit= HA_POS_ERROR; |
|
1661 |
break; |
|
1662 |
}
|
|
1663 |
}
|
|
1664 |
}
|
|
1665 |
if (curr_join->join_tab == join_tab && save_join_tab()) |
|
1666 |
return; |
|
1667 |
/*
|
|
1668 |
Here we sort rows for order_st BY/GROUP BY clause, if the optimiser
|
|
1669 |
chose FILESORT to be faster than INDEX SCAN or there is no
|
|
1670 |
suitable index present.
|
|
1671 |
Note, that create_sort_index calls test_if_skip_sort_order and may
|
|
1672 |
finally replace sorting with index scan if there is a LIMIT clause in
|
|
1673 |
the query. XXX: it's never shown in EXPLAIN!
|
|
1674 |
OPTION_FOUND_ROWS supersedes LIMIT and is taken into account.
|
|
1675 |
*/
|
|
1676 |
if (create_sort_index(session, curr_join, |
|
1677 |
curr_join->group_list ? |
|
1678 |
curr_join->group_list : curr_join->order, |
|
1679 |
curr_join->select_limit, |
|
1680 |
(select_options & OPTION_FOUND_ROWS ? |
|
1681 |
HA_POS_ERROR : unit->select_limit_cnt), |
|
1682 |
curr_join->group_list ? true : false)) |
|
1683 |
return; |
|
1684 |
||
1685 |
sortorder= curr_join->sortorder; |
|
1686 |
if (curr_join->const_tables != curr_join->tables && |
|
1687 |
!curr_join->join_tab[curr_join->const_tables].table->sort.io_cache) |
|
1688 |
{
|
|
1689 |
/*
|
|
1690 |
If no IO cache exists for the first table then we are using an
|
|
1691 |
INDEX SCAN and no filesort. Thus we should not remove the sorted
|
|
1692 |
attribute on the INDEX SCAN.
|
|
1693 |
*/
|
|
1694 |
skip_sort_order= 1; |
|
1695 |
}
|
|
1696 |
}
|
|
1697 |
}
|
|
1698 |
/* XXX: When can we have here session->is_error() not zero? */
|
|
1699 |
if (session->is_error()) |
|
1700 |
{
|
|
1701 |
error= session->is_error(); |
|
1702 |
return; |
|
1703 |
}
|
|
1704 |
curr_join->having= curr_join->tmp_having; |
|
1705 |
curr_join->fields= curr_fields_list; |
|
1706 |
||
1707 |
session->set_proc_info("Sending data"); |
|
1708 |
result->send_fields(*curr_fields_list, |
|
1709 |
Protocol::SEND_NUM_ROWS | Protocol::SEND_EOF); |
|
1710 |
error= do_select(curr_join, curr_fields_list, NULL); |
|
1711 |
session->limit_found_rows= curr_join->send_records; |
|
1712 |
||
1713 |
/* Accumulate the counts from all join iterations of all join parts. */
|
|
1714 |
session->examined_row_count+= curr_join->examined_rows; |
|
1715 |
||
1716 |
/*
|
|
1717 |
With EXPLAIN EXTENDED we have to restore original ref_array
|
|
1718 |
for a derived table which is always materialized.
|
|
1719 |
Otherwise we would not be able to print the query correctly.
|
|
1720 |
*/
|
|
1721 |
if (items0 && (session->lex->describe & DESCRIBE_EXTENDED) && select_lex->linkage == DERIVED_TABLE_TYPE) |
|
1722 |
set_items_ref_array(items0); |
|
1723 |
||
1724 |
return; |
|
1725 |
}
|
|
1726 |
||
1727 |
/**
|
|
1728 |
Clean up join.
|
|
1729 |
||
1730 |
@return
|
|
1731 |
Return error that hold JOIN.
|
|
1732 |
*/
|
|
1733 |
int JOIN::destroy() |
|
1734 |
{
|
|
1735 |
select_lex->join= 0; |
|
1736 |
||
1737 |
if (tmp_join) |
|
1738 |
{
|
|
1739 |
if (join_tab != tmp_join->join_tab) |
|
1740 |
{
|
|
1741 |
JOIN_TAB *tab, *end; |
|
1742 |
for (tab= join_tab, end= tab+tables ; tab != end ; tab++) |
|
1743 |
tab->cleanup(); |
|
1744 |
}
|
|
1745 |
tmp_join->tmp_join= 0; |
|
1746 |
tmp_table_param.copy_field=0; |
|
1747 |
return(tmp_join->destroy()); |
|
1748 |
}
|
|
1749 |
cond_equal= 0; |
|
1750 |
||
1751 |
cleanup(1); |
|
1752 |
if (exec_tmp_table1) |
|
1753 |
exec_tmp_table1->free_tmp_table(session); |
|
1754 |
if (exec_tmp_table2) |
|
1755 |
exec_tmp_table2->free_tmp_table(session); |
|
1756 |
delete select; |
|
1757 |
delete_dynamic(&keyuse); |
|
1758 |
return(error); |
|
1759 |
}
|
|
1760 |
||
1761 |
/*
|
|
1762 |
Convert candidate subquery predicates to semi-joins
|
|
1763 |
||
1764 |
SYNOPSIS
|
|
1765 |
JOIN::flatten_subqueries()
|
|
1766 |
||
1767 |
DESCRIPTION
|
|
1768 |
Convert candidate subquery predicates to semi-joins.
|
|
1769 |
||
1770 |
RETURN
|
|
1771 |
false OK
|
|
1772 |
true Error
|
|
1773 |
*/
|
|
1774 |
bool JOIN::flatten_subqueries() |
|
1775 |
{
|
|
1776 |
Item_in_subselect **in_subq; |
|
1777 |
Item_in_subselect **in_subq_end; |
|
1778 |
||
1779 |
if (sj_subselects.elements() == 0) |
|
1780 |
return(false); |
|
1781 |
||
1782 |
/* 1. Fix children subqueries */
|
|
1783 |
for (in_subq= sj_subselects.front(), in_subq_end= sj_subselects.back(); |
|
1784 |
in_subq != in_subq_end; in_subq++) |
|
1785 |
{
|
|
1786 |
JOIN *child_join= (*in_subq)->unit->first_select()->join; |
|
1787 |
child_join->outer_tables = child_join->tables; |
|
1788 |
if (child_join->flatten_subqueries()) |
|
1789 |
return(true); |
|
1790 |
(*in_subq)->sj_convert_priority= |
|
1791 |
(*in_subq)->is_correlated * MAX_TABLES + child_join->outer_tables; |
|
1792 |
}
|
|
1793 |
||
1794 |
bool outer_join_disable_semi_join= false; |
|
1795 |
/*
|
|
1796 |
* Temporary measure: disable semi-joins when they are together with outer
|
|
1797 |
* joins.
|
|
1798 |
*
|
|
1799 |
* @see LP Bug #314911
|
|
1800 |
*/
|
|
1801 |
for (TableList *tbl= select_lex->leaf_tables; tbl; tbl=tbl->next_leaf) |
|
1802 |
{
|
|
1803 |
TableList *embedding= tbl->embedding; |
|
1804 |
if (tbl->on_expr || (tbl->embedding && !(embedding->sj_on_expr && |
|
1805 |
!embedding->embedding))) |
|
1806 |
{
|
|
1807 |
in_subq= sj_subselects.front(); |
|
1808 |
outer_join_disable_semi_join= true; |
|
1809 |
}
|
|
1810 |
}
|
|
1811 |
||
1812 |
if (! outer_join_disable_semi_join) |
|
1813 |
{
|
|
1814 |
/*
|
|
1815 |
2. Pick which subqueries to convert:
|
|
1816 |
sort the subquery array
|
|
1817 |
- prefer correlated subqueries over uncorrelated;
|
|
1818 |
- prefer subqueries that have greater number of outer tables;
|
|
1819 |
*/
|
|
1820 |
sj_subselects.sort(subq_sj_candidate_cmp); |
|
1821 |
// #tables-in-parent-query + #tables-in-subquery < MAX_TABLES
|
|
1822 |
/* Replace all subqueries to be flattened with Item_int(1) */
|
|
1823 |
for (in_subq= sj_subselects.front(); |
|
1824 |
in_subq != in_subq_end && |
|
1825 |
tables + ((*in_subq)->sj_convert_priority % MAX_TABLES) < MAX_TABLES; |
|
1826 |
in_subq++) |
|
1827 |
{
|
|
1828 |
if (replace_where_subcondition(this, *in_subq, new Item_int(1), false)) |
|
1829 |
return(true); |
|
1830 |
}
|
|
1831 |
||
1832 |
for (in_subq= sj_subselects.front(); |
|
1833 |
in_subq != in_subq_end && |
|
1834 |
tables + ((*in_subq)->sj_convert_priority % MAX_TABLES) < MAX_TABLES; |
|
1835 |
in_subq++) |
|
1836 |
{
|
|
1837 |
if (convert_subq_to_sj(this, *in_subq)) |
|
1838 |
return(true); |
|
1839 |
}
|
|
1840 |
}
|
|
1841 |
||
1842 |
/* 3. Finalize those we didn't convert */
|
|
1843 |
for (; in_subq!= in_subq_end; in_subq++) |
|
1844 |
{
|
|
1845 |
JOIN *child_join= (*in_subq)->unit->first_select()->join; |
|
1846 |
Item_subselect::trans_res res; |
|
1847 |
(*in_subq)->changed= 0; |
|
1848 |
(*in_subq)->fixed= 0; |
|
1849 |
res= (*in_subq)->select_transformer(child_join); |
|
1850 |
if (res == Item_subselect::RES_ERROR) |
|
1851 |
return(true); |
|
1852 |
||
1853 |
(*in_subq)->changed= 1; |
|
1854 |
(*in_subq)->fixed= 1; |
|
1855 |
||
1856 |
Item *substitute= (*in_subq)->substitution; |
|
1857 |
bool do_fix_fields= !(*in_subq)->substitution->fixed; |
|
1858 |
if (replace_where_subcondition(this, *in_subq, substitute, do_fix_fields)) |
|
1859 |
return(true); |
|
1860 |
||
1861 |
//if ((*in_subq)->fix_fields(session, (*in_subq)->ref_ptr))
|
|
1862 |
// return(true);
|
|
1863 |
}
|
|
1864 |
sj_subselects.clear(); |
|
1865 |
return(false); |
|
1866 |
}
|
|
1867 |
||
1868 |
/**
|
|
1869 |
Setup for execution all subqueries of a query, for which the optimizer
|
|
1870 |
chose hash semi-join.
|
|
1871 |
||
1872 |
@details Iterate over all subqueries of the query, and if they are under an
|
|
1873 |
IN predicate, and the optimizer chose to compute it via hash semi-join:
|
|
1874 |
- try to initialize all data structures needed for the materialized execution
|
|
1875 |
of the IN predicate,
|
|
1876 |
- if this fails, then perform the IN=>EXISTS transformation which was
|
|
1877 |
previously blocked during JOIN::prepare.
|
|
1878 |
||
1879 |
This method is part of the "code generation" query processing phase.
|
|
1880 |
||
1881 |
This phase must be called after substitute_for_best_equal_field() because
|
|
1882 |
that function may replace items with other items from a multiple equality,
|
|
1883 |
and we need to reference the correct items in the index access method of the
|
|
1884 |
IN predicate.
|
|
1885 |
||
1886 |
@return Operation status
|
|
1887 |
@retval false success.
|
|
1888 |
@retval true error occurred.
|
|
1889 |
*/
|
|
1890 |
bool JOIN::setup_subquery_materialization() |
|
1891 |
{
|
|
1892 |
for (Select_Lex_Unit *un= select_lex->first_inner_unit(); un; |
|
1893 |
un= un->next_unit()) |
|
1894 |
{
|
|
1895 |
for (Select_Lex *sl= un->first_select(); sl; sl= sl->next_select()) |
|
1896 |
{
|
|
1897 |
Item_subselect *subquery_predicate= sl->master_unit()->item; |
|
1898 |
if (subquery_predicate && |
|
1899 |
subquery_predicate->substype() == Item_subselect::IN_SUBS) |
|
1900 |
{
|
|
1901 |
Item_in_subselect *in_subs= (Item_in_subselect*) subquery_predicate; |
|
1902 |
if (in_subs->exec_method == Item_in_subselect::MATERIALIZATION && |
|
1903 |
in_subs->setup_engine()) |
|
1904 |
return true; |
|
1905 |
}
|
|
1906 |
}
|
|
1907 |
}
|
|
1908 |
return false; |
|
1909 |
}
|
|
1039.2.7
by Jay Pipes
Yet more style and indentation cleanups. |
1910 |
|
1039.2.2
by Jay Pipes
Phase 2 of JOIN refactoring. |
1911 |
/**
|
1912 |
Partially cleanup JOIN after it has executed: close index or rnd read
|
|
1913 |
(table cursors), free quick selects.
|
|
1914 |
||
1915 |
This function is called in the end of execution of a JOIN, before the used
|
|
1916 |
tables are unlocked and closed.
|
|
1917 |
||
1918 |
For a join that is resolved using a temporary table, the first sweep is
|
|
1919 |
performed against actual tables and an intermediate result is inserted
|
|
1920 |
into the temprorary table.
|
|
1921 |
The last sweep is performed against the temporary table. Therefore,
|
|
1922 |
the base tables and associated buffers used to fill the temporary table
|
|
1923 |
are no longer needed, and this function is called to free them.
|
|
1924 |
||
1925 |
For a join that is performed without a temporary table, this function
|
|
1926 |
is called after all rows are sent, but before EOF packet is sent.
|
|
1927 |
||
1928 |
For a simple SELECT with no subqueries this function performs a full
|
|
1929 |
cleanup of the JOIN and calls mysql_unlock_read_tables to free used base
|
|
1930 |
tables.
|
|
1931 |
||
1932 |
If a JOIN is executed for a subquery or if it has a subquery, we can't
|
|
1933 |
do the full cleanup and need to do a partial cleanup only.
|
|
1934 |
- If a JOIN is not the top level join, we must not unlock the tables
|
|
1935 |
because the outer select may not have been evaluated yet, and we
|
|
1936 |
can't unlock only selected tables of a query.
|
|
1937 |
- Additionally, if this JOIN corresponds to a correlated subquery, we
|
|
1938 |
should not free quick selects and join buffers because they will be
|
|
1939 |
needed for the next execution of the correlated subquery.
|
|
1940 |
- However, if this is a JOIN for a [sub]select, which is not
|
|
1941 |
a correlated subquery itself, but has subqueries, we can free it
|
|
1942 |
fully and also free JOINs of all its subqueries. The exception
|
|
1943 |
is a subquery in SELECT list, e.g: @n
|
|
1944 |
SELECT a, (select cmax(b) from t1) group by c @n
|
|
1945 |
This subquery will not be evaluated at first sweep and its value will
|
|
1946 |
not be inserted into the temporary table. Instead, it's evaluated
|
|
1947 |
when selecting from the temporary table. Therefore, it can't be freed
|
|
1948 |
here even though it's not correlated.
|
|
1949 |
||
1950 |
@todo
|
|
1951 |
Unlock tables even if the join isn't top level select in the tree
|
|
1952 |
*/
|
|
1953 |
void JOIN::join_free() |
|
1954 |
{
|
|
1955 |
Select_Lex_Unit *tmp_unit; |
|
1956 |
Select_Lex *sl; |
|
1957 |
/*
|
|
1958 |
Optimization: if not EXPLAIN and we are done with the JOIN,
|
|
1959 |
free all tables.
|
|
1960 |
*/
|
|
1961 |
bool full= (!select_lex->uncacheable && !session->lex->describe); |
|
1962 |
bool can_unlock= full; |
|
1963 |
||
1964 |
cleanup(full); |
|
1965 |
||
1966 |
for (tmp_unit= select_lex->first_inner_unit(); |
|
1967 |
tmp_unit; |
|
1968 |
tmp_unit= tmp_unit->next_unit()) |
|
1969 |
for (sl= tmp_unit->first_select(); sl; sl= sl->next_select()) |
|
1970 |
{
|
|
1971 |
Item_subselect *subselect= sl->master_unit()->item; |
|
1972 |
bool full_local= full && (!subselect || subselect->is_evaluated()); |
|
1973 |
/*
|
|
1974 |
If this join is evaluated, we can fully clean it up and clean up all
|
|
1975 |
its underlying joins even if they are correlated -- they will not be
|
|
1976 |
used any more anyway.
|
|
1977 |
If this join is not yet evaluated, we still must clean it up to
|
|
1978 |
close its table cursors -- it may never get evaluated, as in case of
|
|
1979 |
... HAVING false OR a IN (SELECT ...))
|
|
1980 |
but all table cursors must be closed before the unlock.
|
|
1981 |
*/
|
|
1982 |
sl->cleanup_all_joins(full_local); |
|
1983 |
/* Can't unlock if at least one JOIN is still needed */
|
|
1984 |
can_unlock= can_unlock && full_local; |
|
1985 |
}
|
|
1986 |
||
1987 |
/*
|
|
1988 |
We are not using tables anymore
|
|
1989 |
Unlock all tables. We may be in an INSERT .... SELECT statement.
|
|
1990 |
*/
|
|
1991 |
if (can_unlock && lock && session->lock && |
|
1992 |
!(select_options & SELECT_NO_UNLOCK) && |
|
1993 |
!select_lex->subquery_in_having && |
|
1994 |
(select_lex == (session->lex->unit.fake_select_lex ? |
|
1995 |
session->lex->unit.fake_select_lex : &session->lex->select_lex))) |
|
1996 |
{
|
|
1997 |
/*
|
|
1998 |
TODO: unlock tables even if the join isn't top level select in the
|
|
1999 |
tree.
|
|
2000 |
*/
|
|
2001 |
mysql_unlock_read_tables(session, lock); // Don't free join->lock |
|
2002 |
lock= 0; |
|
2003 |
}
|
|
2004 |
||
2005 |
return; |
|
2006 |
}
|
|
2007 |
||
2008 |
||
2009 |
/**
|
|
2010 |
Free resources of given join.
|
|
2011 |
||
2012 |
@param fill true if we should free all resources, call with full==1
|
|
2013 |
should be last, before it this function can be called with
|
|
2014 |
full==0
|
|
2015 |
||
2016 |
@note
|
|
2017 |
With subquery this function definitely will be called several times,
|
|
2018 |
but even for simple query it can be called several times.
|
|
2019 |
*/
|
|
2020 |
void JOIN::cleanup(bool full) |
|
2021 |
{
|
|
2022 |
if (table) |
|
2023 |
{
|
|
2024 |
JOIN_TAB *tab,*end; |
|
2025 |
/*
|
|
2026 |
Only a sorted table may be cached. This sorted table is always the
|
|
2027 |
first non const table in join->table
|
|
2028 |
*/
|
|
2029 |
if (tables > const_tables) // Test for not-const tables |
|
2030 |
{
|
|
2031 |
free_io_cache(table[const_tables]); |
|
2032 |
filesort_free_buffers(table[const_tables],full); |
|
2033 |
}
|
|
2034 |
||
2035 |
if (full) |
|
2036 |
{
|
|
2037 |
for (tab= join_tab, end= tab+tables; tab != end; tab++) |
|
1039.2.7
by Jay Pipes
Yet more style and indentation cleanups. |
2038 |
tab->cleanup(); |
1039.2.2
by Jay Pipes
Phase 2 of JOIN refactoring. |
2039 |
table= 0; |
2040 |
}
|
|
2041 |
else
|
|
2042 |
{
|
|
2043 |
for (tab= join_tab, end= tab+tables; tab != end; tab++) |
|
2044 |
{
|
|
1039.2.7
by Jay Pipes
Yet more style and indentation cleanups. |
2045 |
if (tab->table) |
1039.2.2
by Jay Pipes
Phase 2 of JOIN refactoring. |
2046 |
tab->table->file->ha_index_or_rnd_end(); |
2047 |
}
|
|
2048 |
}
|
|
2049 |
cleanup_sj_tmp_tables(this);// |
|
2050 |
}
|
|
2051 |
/*
|
|
2052 |
We are not using tables anymore
|
|
2053 |
Unlock all tables. We may be in an INSERT .... SELECT statement.
|
|
2054 |
*/
|
|
2055 |
if (full) |
|
2056 |
{
|
|
2057 |
if (tmp_join) |
|
2058 |
tmp_table_param.copy_field= 0; |
|
2059 |
group_fields.delete_elements(); |
|
2060 |
/*
|
|
2061 |
We can't call delete_elements() on copy_funcs as this will cause
|
|
2062 |
problems in free_elements() as some of the elements are then deleted.
|
|
2063 |
*/
|
|
2064 |
tmp_table_param.copy_funcs.empty(); |
|
2065 |
/*
|
|
2066 |
If we have tmp_join and 'this' JOIN is not tmp_join and
|
|
2067 |
tmp_table_param.copy_field's of them are equal then we have to remove
|
|
2068 |
pointer to tmp_table_param.copy_field from tmp_join, because it qill
|
|
2069 |
be removed in tmp_table_param.cleanup().
|
|
2070 |
*/
|
|
2071 |
if (tmp_join && |
|
2072 |
tmp_join != this && |
|
2073 |
tmp_join->tmp_table_param.copy_field == |
|
2074 |
tmp_table_param.copy_field) |
|
2075 |
{
|
|
2076 |
tmp_join->tmp_table_param.copy_field= |
|
2077 |
tmp_join->tmp_table_param.save_copy_field= 0; |
|
2078 |
}
|
|
2079 |
tmp_table_param.cleanup(); |
|
2080 |
}
|
|
2081 |
return; |
|
2082 |
}
|
|
2083 |
||
2084 |
/*
|
|
2085 |
used only in JOIN::clear
|
|
2086 |
*/
|
|
2087 |
static void clear_tables(JOIN *join) |
|
2088 |
{
|
|
2089 |
/*
|
|
2090 |
must clear only the non-const tables, as const tables
|
|
2091 |
are not re-calculated.
|
|
2092 |
*/
|
|
1039.2.7
by Jay Pipes
Yet more style and indentation cleanups. |
2093 |
for (uint32_t i= join->const_tables; i < join->tables; i++) |
1039.2.2
by Jay Pipes
Phase 2 of JOIN refactoring. |
2094 |
join->table[i]->mark_as_null_row(); // All fields are NULL |
2095 |
}
|
|
2096 |
||
2097 |
/**
|
|
2098 |
Make an array of pointers to sum_functions to speed up
|
|
2099 |
sum_func calculation.
|
|
2100 |
||
2101 |
@retval
|
|
2102 |
0 ok
|
|
2103 |
@retval
|
|
2104 |
1 Error
|
|
2105 |
*/
|
|
2106 |
bool JOIN::alloc_func_list() |
|
2107 |
{
|
|
2108 |
uint32_t func_count, group_parts; |
|
2109 |
||
2110 |
func_count= tmp_table_param.sum_func_count; |
|
2111 |
/*
|
|
2112 |
If we are using rollup, we need a copy of the summary functions for
|
|
2113 |
each level
|
|
2114 |
*/
|
|
2115 |
if (rollup.state != ROLLUP::STATE_NONE) |
|
2116 |
func_count*= (send_group_parts+1); |
|
2117 |
||
2118 |
group_parts= send_group_parts; |
|
2119 |
/*
|
|
2120 |
If distinct, reserve memory for possible
|
|
2121 |
disctinct->group_by optimization
|
|
2122 |
*/
|
|
2123 |
if (select_distinct) |
|
2124 |
{
|
|
2125 |
group_parts+= fields_list.elements; |
|
2126 |
/*
|
|
2127 |
If the order_st clause is specified then it's possible that
|
|
2128 |
it also will be optimized, so reserve space for it too
|
|
2129 |
*/
|
|
2130 |
if (order) |
|
2131 |
{
|
|
2132 |
order_st *ord; |
|
2133 |
for (ord= order; ord; ord= ord->next) |
|
2134 |
group_parts++; |
|
2135 |
}
|
|
2136 |
}
|
|
2137 |
||
2138 |
/* This must use calloc() as rollup_make_fields depends on this */
|
|
2139 |
sum_funcs= (Item_sum**) session->calloc(sizeof(Item_sum**) * (func_count+1) + |
|
2140 |
sizeof(Item_sum***) * (group_parts+1)); |
|
2141 |
sum_funcs_end= (Item_sum***) (sum_funcs+func_count+1); |
|
2142 |
return(sum_funcs == 0); |
|
2143 |
}
|
|
2144 |
||
2145 |
/**
|
|
2146 |
Initialize 'sum_funcs' array with all Item_sum objects.
|
|
2147 |
||
2148 |
@param field_list All items
|
|
2149 |
@param send_fields Items in select list
|
|
2150 |
@param before_group_by Set to 1 if this is called before GROUP BY handling
|
|
2151 |
@param recompute Set to true if sum_funcs must be recomputed
|
|
2152 |
||
2153 |
@retval
|
|
2154 |
0 ok
|
|
2155 |
@retval
|
|
2156 |
1 error
|
|
2157 |
*/
|
|
1039.2.7
by Jay Pipes
Yet more style and indentation cleanups. |
2158 |
bool JOIN::make_sum_func_list(List<Item> &field_list, |
2159 |
List<Item> &send_fields, |
|
2160 |
bool before_group_by, |
|
2161 |
bool recompute) |
|
1039.2.2
by Jay Pipes
Phase 2 of JOIN refactoring. |
2162 |
{
|
2163 |
List_iterator_fast<Item> it(field_list); |
|
2164 |
Item_sum **func; |
|
2165 |
Item *item; |
|
2166 |
||
2167 |
if (*sum_funcs && !recompute) |
|
2168 |
return(false); /* We have already initialized sum_funcs. */ |
|
2169 |
||
2170 |
func= sum_funcs; |
|
2171 |
while ((item=it++)) |
|
2172 |
{
|
|
2173 |
if (item->type() == Item::SUM_FUNC_ITEM && !item->const_item() && |
|
2174 |
(!((Item_sum*) item)->depended_from() || |
|
2175 |
((Item_sum *)item)->depended_from() == select_lex)) |
|
2176 |
*func++= (Item_sum*) item; |
|
2177 |
}
|
|
2178 |
if (before_group_by && rollup.state == ROLLUP::STATE_INITED) |
|
2179 |
{
|
|
2180 |
rollup.state= ROLLUP::STATE_READY; |
|
2181 |
if (rollup_make_fields(field_list, send_fields, &func)) |
|
2182 |
return(true); // Should never happen |
|
2183 |
}
|
|
2184 |
else if (rollup.state == ROLLUP::STATE_NONE) |
|
2185 |
{
|
|
2186 |
for (uint32_t i=0 ; i <= send_group_parts ;i++) |
|
2187 |
sum_funcs_end[i]= func; |
|
2188 |
}
|
|
2189 |
else if (rollup.state == ROLLUP::STATE_READY) |
|
2190 |
return(false); // Don't put end marker |
|
2191 |
*func=0; // End marker |
|
2192 |
return(false); |
|
2193 |
}
|
|
2194 |
||
2195 |
/** Allocate memory needed for other rollup functions. */
|
|
2196 |
bool JOIN::rollup_init() |
|
2197 |
{
|
|
2198 |
uint32_t i,j; |
|
2199 |
Item **ref_array; |
|
2200 |
||
2201 |
tmp_table_param.quick_group= 0; // Can't create groups in tmp table |
|
2202 |
rollup.state= ROLLUP::STATE_INITED; |
|
2203 |
||
2204 |
/*
|
|
2205 |
Create pointers to the different sum function groups
|
|
2206 |
These are updated by rollup_make_fields()
|
|
2207 |
*/
|
|
2208 |
tmp_table_param.group_parts= send_group_parts; |
|
2209 |
||
2210 |
if (!(rollup.null_items= (Item_null_result**) session->alloc((sizeof(Item*) + |
|
2211 |
sizeof(Item**) + |
|
2212 |
sizeof(List<Item>) + |
|
2213 |
ref_pointer_array_size) |
|
2214 |
* send_group_parts ))) |
|
2215 |
return 1; |
|
2216 |
||
2217 |
rollup.fields= (List<Item>*) (rollup.null_items + send_group_parts); |
|
2218 |
rollup.ref_pointer_arrays= (Item***) (rollup.fields + send_group_parts); |
|
2219 |
ref_array= (Item**) (rollup.ref_pointer_arrays+send_group_parts); |
|
2220 |
||
2221 |
/*
|
|
2222 |
Prepare space for field list for the different levels
|
|
2223 |
These will be filled up in rollup_make_fields()
|
|
2224 |
*/
|
|
2225 |
for (i= 0 ; i < send_group_parts ; i++) |
|
2226 |
{
|
|
2227 |
rollup.null_items[i]= new (session->mem_root) Item_null_result(); |
|
2228 |
List<Item> *rollup_fields= &rollup.fields[i]; |
|
2229 |
rollup_fields->empty(); |
|
2230 |
rollup.ref_pointer_arrays[i]= ref_array; |
|
2231 |
ref_array+= all_fields.elements; |
|
2232 |
}
|
|
2233 |
for (i= 0 ; i < send_group_parts; i++) |
|
2234 |
{
|
|
2235 |
for (j=0 ; j < fields_list.elements ; j++) |
|
2236 |
rollup.fields[i].push_back(rollup.null_items[i]); |
|
2237 |
}
|
|
2238 |
List_iterator<Item> it(all_fields); |
|
2239 |
Item *item; |
|
2240 |
while ((item= it++)) |
|
2241 |
{
|
|
2242 |
order_st *group_tmp; |
|
2243 |
bool found_in_group= 0; |
|
2244 |
||
2245 |
for (group_tmp= group_list; group_tmp; group_tmp= group_tmp->next) |
|
2246 |
{
|
|
2247 |
if (*group_tmp->item == item) |
|
2248 |
{
|
|
2249 |
item->maybe_null= 1; |
|
2250 |
found_in_group= 1; |
|
2251 |
if (item->const_item()) |
|
2252 |
{
|
|
2253 |
/*
|
|
2254 |
For ROLLUP queries each constant item referenced in GROUP BY list
|
|
2255 |
is wrapped up into an Item_func object yielding the same value
|
|
2256 |
as the constant item. The objects of the wrapper class are never
|
|
2257 |
considered as constant items and besides they inherit all
|
|
2258 |
properties of the Item_result_field class.
|
|
2259 |
This wrapping allows us to ensure writing constant items
|
|
2260 |
into temporary tables whenever the result of the ROLLUP
|
|
2261 |
operation has to be written into a temporary table, e.g. when
|
|
2262 |
ROLLUP is used together with DISTINCT in the SELECT list.
|
|
2263 |
Usually when creating temporary tables for a intermidiate
|
|
2264 |
result we do not include fields for constant expressions.
|
|
1039.2.7
by Jay Pipes
Yet more style and indentation cleanups. |
2265 |
*/
|
1039.2.2
by Jay Pipes
Phase 2 of JOIN refactoring. |
2266 |
Item* new_item= new Item_func_rollup_const(item); |
2267 |
if (!new_item) |
|
2268 |
return 1; |
|
2269 |
new_item->fix_fields(session, (Item **) 0); |
|
2270 |
session->change_item_tree(it.ref(), new_item); |
|
2271 |
for (order_st *tmp= group_tmp; tmp; tmp= tmp->next) |
|
2272 |
{
|
|
2273 |
if (*tmp->item == item) |
|
2274 |
session->change_item_tree(tmp->item, new_item); |
|
2275 |
}
|
|
2276 |
}
|
|
2277 |
}
|
|
2278 |
}
|
|
2279 |
if (item->type() == Item::FUNC_ITEM && !found_in_group) |
|
2280 |
{
|
|
2281 |
bool changed= false; |
|
2282 |
if (change_group_ref(session, (Item_func *) item, group_list, &changed)) |
|
2283 |
return 1; |
|
2284 |
/*
|
|
2285 |
We have to prevent creation of a field in a temporary table for
|
|
2286 |
an expression that contains GROUP BY attributes.
|
|
2287 |
Marking the expression item as 'with_sum_func' will ensure this.
|
|
2288 |
*/
|
|
2289 |
if (changed) |
|
2290 |
item->with_sum_func= 1; |
|
2291 |
}
|
|
2292 |
}
|
|
2293 |
return 0; |
|
2294 |
}
|
|
2295 |
||
2296 |
/**
|
|
2297 |
Fill up rollup structures with pointers to fields to use.
|
|
2298 |
||
2299 |
Creates copies of item_sum items for each sum level.
|
|
2300 |
||
2301 |
@param fields_arg List of all fields (hidden and real ones)
|
|
2302 |
@param sel_fields Pointer to selected fields
|
|
2303 |
@param func Store here a pointer to all fields
|
|
2304 |
||
2305 |
@retval
|
|
2306 |
0 if ok;
|
|
2307 |
In this case func is pointing to next not used element.
|
|
2308 |
@retval
|
|
2309 |
1 on error
|
|
2310 |
*/
|
|
1039.2.7
by Jay Pipes
Yet more style and indentation cleanups. |
2311 |
bool JOIN::rollup_make_fields(List<Item> &fields_arg, List<Item> &sel_fields, Item_sum ***func) |
1039.2.2
by Jay Pipes
Phase 2 of JOIN refactoring. |
2312 |
{
|
2313 |
List_iterator_fast<Item> it(fields_arg); |
|
2314 |
Item *first_field= sel_fields.head(); |
|
2315 |
uint32_t level; |
|
2316 |
||
2317 |
/*
|
|
2318 |
Create field lists for the different levels
|
|
2319 |
||
2320 |
The idea here is to have a separate field list for each rollup level to
|
|
2321 |
avoid all runtime checks of which columns should be NULL.
|
|
2322 |
||
2323 |
The list is stored in reverse order to get sum function in such an order
|
|
2324 |
in func that it makes it easy to reset them with init_sum_functions()
|
|
2325 |
||
2326 |
Assuming: SELECT a, b, c SUM(b) FROM t1 GROUP BY a,b WITH ROLLUP
|
|
2327 |
||
2328 |
rollup.fields[0] will contain list where a,b,c is NULL
|
|
2329 |
rollup.fields[1] will contain list where b,c is NULL
|
|
2330 |
...
|
|
2331 |
rollup.ref_pointer_array[#] points to fields for rollup.fields[#]
|
|
2332 |
...
|
|
2333 |
sum_funcs_end[0] points to all sum functions
|
|
2334 |
sum_funcs_end[1] points to all sum functions, except grand totals
|
|
2335 |
...
|
|
2336 |
*/
|
|
2337 |
||
2338 |
for (level=0 ; level < send_group_parts ; level++) |
|
2339 |
{
|
|
2340 |
uint32_t i; |
|
2341 |
uint32_t pos= send_group_parts - level -1; |
|
2342 |
bool real_fields= 0; |
|
2343 |
Item *item; |
|
2344 |
List_iterator<Item> new_it(rollup.fields[pos]); |
|
2345 |
Item **ref_array_start= rollup.ref_pointer_arrays[pos]; |
|
2346 |
order_st *start_group; |
|
2347 |
||
2348 |
/* Point to first hidden field */
|
|
2349 |
Item **ref_array= ref_array_start + fields_arg.elements-1; |
|
2350 |
||
2351 |
/* Remember where the sum functions ends for the previous level */
|
|
2352 |
sum_funcs_end[pos+1]= *func; |
|
2353 |
||
2354 |
/* Find the start of the group for this level */
|
|
1039.2.7
by Jay Pipes
Yet more style and indentation cleanups. |
2355 |
for (i= 0, start_group= group_list ;i++ < pos ;start_group= start_group->next) |
2356 |
{}
|
|
1039.2.2
by Jay Pipes
Phase 2 of JOIN refactoring. |
2357 |
|
2358 |
it.rewind(); |
|
2359 |
while ((item= it++)) |
|
2360 |
{
|
|
2361 |
if (item == first_field) |
|
2362 |
{
|
|
1039.2.7
by Jay Pipes
Yet more style and indentation cleanups. |
2363 |
real_fields= 1; // End of hidden fields |
2364 |
ref_array= ref_array_start; |
|
1039.2.2
by Jay Pipes
Phase 2 of JOIN refactoring. |
2365 |
}
|
2366 |
||
2367 |
if (item->type() == Item::SUM_FUNC_ITEM && !item->const_item() && |
|
2368 |
(!((Item_sum*) item)->depended_from() || |
|
2369 |
((Item_sum *)item)->depended_from() == select_lex)) |
|
2370 |
||
2371 |
{
|
|
1039.2.7
by Jay Pipes
Yet more style and indentation cleanups. |
2372 |
/*
|
2373 |
This is a top level summary function that must be replaced with
|
|
2374 |
a sum function that is reset for this level.
|
|
1039.2.2
by Jay Pipes
Phase 2 of JOIN refactoring. |
2375 |
|
1039.2.7
by Jay Pipes
Yet more style and indentation cleanups. |
2376 |
NOTE: This code creates an object which is not that nice in a
|
2377 |
sub select. Fortunately it's not common to have rollup in
|
|
2378 |
sub selects.
|
|
2379 |
*/
|
|
2380 |
item= item->copy_or_same(session); |
|
2381 |
((Item_sum*) item)->make_unique(); |
|
2382 |
*(*func)= (Item_sum*) item; |
|
2383 |
(*func)++; |
|
1039.2.2
by Jay Pipes
Phase 2 of JOIN refactoring. |
2384 |
}
|
2385 |
else
|
|
2386 |
{
|
|
1039.2.7
by Jay Pipes
Yet more style and indentation cleanups. |
2387 |
/* Check if this is something that is part of this group by */
|
2388 |
order_st *group_tmp; |
|
2389 |
for (group_tmp= start_group, i= pos ; |
|
2390 |
group_tmp ; group_tmp= group_tmp->next, i++) |
|
2391 |
{
|
|
2392 |
if (*group_tmp->item == item) |
|
2393 |
{
|
|
2394 |
/*
|
|
2395 |
This is an element that is used by the GROUP BY and should be
|
|
2396 |
set to NULL in this level
|
|
2397 |
*/
|
|
2398 |
Item_null_result *null_item= new (session->mem_root) Item_null_result(); |
|
2399 |
if (!null_item) |
|
2400 |
return 1; |
|
2401 |
item->maybe_null= 1; // Value will be null sometimes |
|
2402 |
null_item->result_field= item->get_tmp_table_field(); |
|
2403 |
item= null_item; |
|
2404 |
break; |
|
2405 |
}
|
|
2406 |
}
|
|
1039.2.2
by Jay Pipes
Phase 2 of JOIN refactoring. |
2407 |
}
|
2408 |
*ref_array= item; |
|
2409 |
if (real_fields) |
|
2410 |
{
|
|
2411 |
(void) new_it++; // Point to next item |
|
2412 |
new_it.replace(item); // Replace previous |
|
2413 |
ref_array++; |
|
2414 |
}
|
|
2415 |
else
|
|
2416 |
ref_array--; |
|
2417 |
}
|
|
2418 |
}
|
|
2419 |
sum_funcs_end[0]= *func; // Point to last function |
|
2420 |
return 0; |
|
2421 |
}
|
|
2422 |
||
2423 |
/**
|
|
2424 |
Send all rollup levels higher than the current one to the client.
|
|
2425 |
||
2426 |
@b SAMPLE
|
|
2427 |
@code
|
|
2428 |
SELECT a, b, c SUM(b) FROM t1 GROUP BY a,b WITH ROLLUP
|
|
2429 |
@endcode
|
|
2430 |
||
2431 |
@param idx Level we are on:
|
|
2432 |
- 0 = Total sum level
|
|
2433 |
- 1 = First group changed (a)
|
|
2434 |
- 2 = Second group changed (a,b)
|
|
2435 |
||
2436 |
@retval
|
|
2437 |
0 ok
|
|
2438 |
@retval
|
|
2439 |
1 If send_data_failed()
|
|
2440 |
*/
|
|
2441 |
int JOIN::rollup_send_data(uint32_t idx) |
|
2442 |
{
|
|
2443 |
uint32_t i; |
|
2444 |
for (i= send_group_parts ; i-- > idx ; ) |
|
2445 |
{
|
|
2446 |
/* Get reference pointers to sum functions in place */
|
|
2447 |
memcpy(ref_pointer_array, rollup.ref_pointer_arrays[i], |
|
2448 |
ref_pointer_array_size); |
|
2449 |
if ((!having || having->val_int())) |
|
2450 |
{
|
|
2451 |
if (send_records < unit->select_limit_cnt && do_send_rows && |
|
2452 |
result->send_data(rollup.fields[i])) |
|
2453 |
return 1; |
|
2454 |
send_records++; |
|
2455 |
}
|
|
2456 |
}
|
|
2457 |
/* Restore ref_pointer_array */
|
|
2458 |
set_items_ref_array(current_ref_pointer_array); |
|
2459 |
return 0; |
|
2460 |
}
|
|
2461 |
||
2462 |
/**
|
|
2463 |
Write all rollup levels higher than the current one to a temp table.
|
|
2464 |
||
2465 |
@b SAMPLE
|
|
2466 |
@code
|
|
2467 |
SELECT a, b, SUM(c) FROM t1 GROUP BY a,b WITH ROLLUP
|
|
2468 |
@endcode
|
|
2469 |
||
2470 |
@param idx Level we are on:
|
|
2471 |
- 0 = Total sum level
|
|
2472 |
- 1 = First group changed (a)
|
|
2473 |
- 2 = Second group changed (a,b)
|
|
2474 |
@param table reference to temp table
|
|
2475 |
||
2476 |
@retval
|
|
2477 |
0 ok
|
|
2478 |
@retval
|
|
2479 |
1 if write_data_failed()
|
|
2480 |
*/
|
|
2481 |
int JOIN::rollup_write_data(uint32_t idx, Table *table_arg) |
|
2482 |
{
|
|
2483 |
uint32_t i; |
|
2484 |
for (i= send_group_parts ; i-- > idx ; ) |
|
2485 |
{
|
|
2486 |
/* Get reference pointers to sum functions in place */
|
|
2487 |
memcpy(ref_pointer_array, rollup.ref_pointer_arrays[i], |
|
2488 |
ref_pointer_array_size); |
|
2489 |
if ((!having || having->val_int())) |
|
2490 |
{
|
|
2491 |
int write_error; |
|
2492 |
Item *item; |
|
2493 |
List_iterator_fast<Item> it(rollup.fields[i]); |
|
2494 |
while ((item= it++)) |
|
2495 |
{
|
|
2496 |
if (item->type() == Item::NULL_ITEM && item->is_result_field()) |
|
2497 |
item->save_in_result_field(1); |
|
2498 |
}
|
|
2499 |
copy_sum_funcs(sum_funcs_end[i+1], sum_funcs_end[i]); |
|
2500 |
if ((write_error= table_arg->file->ha_write_row(table_arg->record[0]))) |
|
2501 |
{
|
|
2502 |
if (create_myisam_from_heap(session, table_arg, |
|
2503 |
tmp_table_param.start_recinfo, |
|
2504 |
&tmp_table_param.recinfo, |
|
2505 |
write_error, 0)) |
|
2506 |
return 1; |
|
2507 |
}
|
|
2508 |
}
|
|
2509 |
}
|
|
2510 |
/* Restore ref_pointer_array */
|
|
2511 |
set_items_ref_array(current_ref_pointer_array); |
|
2512 |
return 0; |
|
2513 |
}
|
|
2514 |
||
2515 |
/**
|
|
2516 |
clear results if there are not rows found for group
|
|
2517 |
(end_send_group/end_write_group)
|
|
2518 |
*/
|
|
2519 |
void JOIN::clear() |
|
2520 |
{
|
|
2521 |
clear_tables(this); |
|
2522 |
copy_fields(&tmp_table_param); |
|
2523 |
||
2524 |
if (sum_funcs) |
|
2525 |
{
|
|
2526 |
Item_sum *func, **func_ptr= sum_funcs; |
|
2527 |
while ((func= *(func_ptr++))) |
|
2528 |
func->clear(); |
|
2529 |
}
|
|
2530 |
}
|
|
1039.2.7
by Jay Pipes
Yet more style and indentation cleanups. |
2531 |
|
1039.2.2
by Jay Pipes
Phase 2 of JOIN refactoring. |
2532 |
/**
|
2533 |
change select_result object of JOIN.
|
|
2534 |
||
2535 |
@param res new select_result object
|
|
2536 |
||
2537 |
@retval
|
|
2538 |
false OK
|
|
2539 |
@retval
|
|
2540 |
true error
|
|
2541 |
*/
|
|
2542 |
bool JOIN::change_result(select_result *res) |
|
2543 |
{
|
|
2544 |
result= res; |
|
2545 |
if (result->prepare(fields_list, select_lex->master_unit())) |
|
2546 |
{
|
|
2547 |
return(true); |
|
2548 |
}
|
|
2549 |
return(false); |
|
2550 |
}
|
|
2551 |
||
2552 |
/**
|
|
2553 |
Give error if we some tables are done with a full join.
|
|
2554 |
||
2555 |
This is used by multi_table_update and multi_table_delete when running
|
|
2556 |
in safe mode.
|
|
2557 |
||
2558 |
@param join Join condition
|
|
2559 |
||
2560 |
@retval
|
|
2561 |
0 ok
|
|
2562 |
@retval
|
|
2563 |
1 Error (full join used)
|
|
2564 |
*/
|
|
2565 |
bool error_if_full_join(JOIN *join) |
|
2566 |
{
|
|
1039.2.7
by Jay Pipes
Yet more style and indentation cleanups. |
2567 |
for (JOIN_TAB *tab= join->join_tab, *end= join->join_tab+join->tables; tab < end; tab++) |
1039.2.2
by Jay Pipes
Phase 2 of JOIN refactoring. |
2568 |
{
|
2569 |
if (tab->type == JT_ALL && (!tab->select || !tab->select->quick)) |
|
2570 |
{
|
|
2571 |
my_message(ER_UPDATE_WITHOUT_KEY_IN_SAFE_MODE, |
|
2572 |
ER(ER_UPDATE_WITHOUT_KEY_IN_SAFE_MODE), MYF(0)); |
|
2573 |
return(1); |
|
2574 |
}
|
|
2575 |
}
|
|
2576 |
return(0); |
|
2577 |
}
|
|
2578 |
||
1039.2.7
by Jay Pipes
Yet more style and indentation cleanups. |
2579 |
/**
|
2580 |
@brief
|
|
2581 |
|
|
1039.2.2
by Jay Pipes
Phase 2 of JOIN refactoring. |
2582 |
Process one record of the nested loop join.
|
2583 |
||
1039.2.7
by Jay Pipes
Yet more style and indentation cleanups. |
2584 |
@details
|
2585 |
||
2586 |
This function will evaluate parts of WHERE/ON clauses that are
|
|
2587 |
applicable to the partial record on hand and in case of success
|
|
2588 |
submit this record to the next level of the nested loop.
|
|
1039.2.2
by Jay Pipes
Phase 2 of JOIN refactoring. |
2589 |
*/
|
2590 |
enum_nested_loop_state evaluate_join_record(JOIN *join, JOIN_TAB *join_tab, int error) |
|
2591 |
{
|
|
1039.2.7
by Jay Pipes
Yet more style and indentation cleanups. |
2592 |
bool not_used_in_distinct= join_tab->not_used_in_distinct; |
2593 |
ha_rows found_records= join->found_records; |
|
1039.2.2
by Jay Pipes
Phase 2 of JOIN refactoring. |
2594 |
COND *select_cond= join_tab->select_cond; |
2595 |
||
2596 |
if (error > 0 || (join->session->is_error())) // Fatal error |
|
2597 |
return NESTED_LOOP_ERROR; |
|
2598 |
if (error < 0) |
|
2599 |
return NESTED_LOOP_NO_MORE_ROWS; |
|
2600 |
if (join->session->killed) // Aborted by user |
|
2601 |
{
|
|
2602 |
join->session->send_kill_message(); |
|
2603 |
return NESTED_LOOP_KILLED; /* purecov: inspected */ |
|
2604 |
}
|
|
2605 |
if (!select_cond || select_cond->val_int()) |
|
2606 |
{
|
|
2607 |
/*
|
|
2608 |
There is no select condition or the attached pushed down
|
|
2609 |
condition is true => a match is found.
|
|
2610 |
*/
|
|
2611 |
bool found= 1; |
|
2612 |
while (join_tab->first_unmatched && found) |
|
2613 |
{
|
|
2614 |
/*
|
|
2615 |
The while condition is always false if join_tab is not
|
|
2616 |
the last inner join table of an outer join operation.
|
|
2617 |
*/
|
|
2618 |
JOIN_TAB *first_unmatched= join_tab->first_unmatched; |
|
2619 |
/*
|
|
2620 |
Mark that a match for current outer table is found.
|
|
2621 |
This activates push down conditional predicates attached
|
|
2622 |
to the all inner tables of the outer join.
|
|
2623 |
*/
|
|
2624 |
first_unmatched->found= 1; |
|
2625 |
for (JOIN_TAB *tab= first_unmatched; tab <= join_tab; tab++) |
|
2626 |
{
|
|
2627 |
if (tab->table->reginfo.not_exists_optimize) |
|
2628 |
return NESTED_LOOP_NO_MORE_ROWS; |
|
2629 |
/* Check all predicates that has just been activated. */
|
|
2630 |
/*
|
|
2631 |
Actually all predicates non-guarded by first_unmatched->found
|
|
2632 |
will be re-evaluated again. It could be fixed, but, probably,
|
|
2633 |
it's not worth doing now.
|
|
2634 |
*/
|
|
2635 |
if (tab->select_cond && !tab->select_cond->val_int()) |
|
2636 |
{
|
|
2637 |
/* The condition attached to table tab is false */
|
|
2638 |
if (tab == join_tab) |
|
2639 |
found= 0; |
|
2640 |
else
|
|
2641 |
{
|
|
2642 |
/*
|
|
2643 |
Set a return point if rejected predicate is attached
|
|
2644 |
not to the last table of the current nest level.
|
|
2645 |
*/
|
|
2646 |
join->return_tab= tab; |
|
2647 |
return NESTED_LOOP_OK; |
|
2648 |
}
|
|
2649 |
}
|
|
2650 |
}
|
|
2651 |
/*
|
|
2652 |
Check whether join_tab is not the last inner table
|
|
2653 |
for another embedding outer join.
|
|
2654 |
*/
|
|
2655 |
if ((first_unmatched= first_unmatched->first_upper) && |
|
2656 |
first_unmatched->last_inner != join_tab) |
|
2657 |
first_unmatched= 0; |
|
2658 |
join_tab->first_unmatched= first_unmatched; |
|
2659 |
}
|
|
2660 |
||
2661 |
JOIN_TAB *return_tab= join->return_tab; |
|
2662 |
join_tab->found_match= true; |
|
2663 |
if (join_tab->check_weed_out_table) |
|
2664 |
{
|
|
2665 |
int res= do_sj_dups_weedout(join->session, join_tab->check_weed_out_table); |
|
2666 |
if (res == -1) |
|
2667 |
return NESTED_LOOP_ERROR; |
|
2668 |
if (res == 1) |
|
2669 |
return NESTED_LOOP_OK; |
|
2670 |
}
|
|
2671 |
else if (join_tab->do_firstmatch) |
|
2672 |
{
|
|
2673 |
/*
|
|
2674 |
We should return to the join_tab->do_firstmatch after we have
|
|
2675 |
enumerated all the suffixes for current prefix row combination
|
|
2676 |
*/
|
|
2677 |
return_tab= join_tab->do_firstmatch; |
|
2678 |
}
|
|
2679 |
||
2680 |
/*
|
|
2681 |
It was not just a return to lower loop level when one
|
|
2682 |
of the newly activated predicates is evaluated as false
|
|
2683 |
(See above join->return_tab= tab).
|
|
2684 |
*/
|
|
2685 |
join->examined_rows++; |
|
2686 |
join->session->row_count++; |
|
2687 |
||
2688 |
if (found) |
|
2689 |
{
|
|
2690 |
enum enum_nested_loop_state rc; |
|
2691 |
/* A match from join_tab is found for the current partial join. */
|
|
2692 |
rc= (*join_tab->next_select)(join, join_tab+1, 0); |
|
2693 |
if (rc != NESTED_LOOP_OK && rc != NESTED_LOOP_NO_MORE_ROWS) |
|
2694 |
return rc; |
|
2695 |
if (return_tab < join->return_tab) |
|
2696 |
join->return_tab= return_tab; |
|
2697 |
||
2698 |
if (join->return_tab < join_tab) |
|
2699 |
return NESTED_LOOP_OK; |
|
2700 |
/*
|
|
2701 |
Test if this was a SELECT DISTINCT query on a table that
|
|
2702 |
was not in the field list; In this case we can abort if
|
|
2703 |
we found a row, as no new rows can be added to the result.
|
|
2704 |
*/
|
|
2705 |
if (not_used_in_distinct && found_records != join->found_records) |
|
2706 |
return NESTED_LOOP_NO_MORE_ROWS; |
|
2707 |
}
|
|
2708 |
else
|
|
2709 |
join_tab->read_record.file->unlock_row(); |
|
2710 |
}
|
|
2711 |
else
|
|
2712 |
{
|
|
2713 |
/*
|
|
2714 |
The condition pushed down to the table join_tab rejects all rows
|
|
2715 |
with the beginning coinciding with the current partial join.
|
|
2716 |
*/
|
|
2717 |
join->examined_rows++; |
|
2718 |
join->session->row_count++; |
|
2719 |
join_tab->read_record.file->unlock_row(); |
|
2720 |
}
|
|
2721 |
return NESTED_LOOP_OK; |
|
2722 |
}
|
|
2723 |
||
2724 |
/**
|
|
2725 |
@details
|
|
2726 |
Construct a NULL complimented partial join record and feed it to the next
|
|
2727 |
level of the nested loop. This function is used in case we have
|
|
2728 |
an OUTER join and no matching record was found.
|
|
2729 |
*/
|
|
2730 |
enum_nested_loop_state evaluate_null_complemented_join_record(JOIN *join, JOIN_TAB *join_tab) |
|
2731 |
{
|
|
2732 |
/*
|
|
2733 |
The table join_tab is the first inner table of a outer join operation
|
|
2734 |
and no matches has been found for the current outer row.
|
|
2735 |
*/
|
|
2736 |
JOIN_TAB *last_inner_tab= join_tab->last_inner; |
|
2737 |
/* Cache variables for faster loop */
|
|
2738 |
COND *select_cond; |
|
2739 |
for ( ; join_tab <= last_inner_tab ; join_tab++) |
|
2740 |
{
|
|
2741 |
/* Change the the values of guard predicate variables. */
|
|
2742 |
join_tab->found= 1; |
|
2743 |
join_tab->not_null_compl= 0; |
|
2744 |
/* The outer row is complemented by nulls for each inner tables */
|
|
2745 |
join_tab->table->restoreRecordAsDefault(); // Make empty record |
|
2746 |
join_tab->table->mark_as_null_row(); // For group by without error |
|
2747 |
select_cond= join_tab->select_cond; |
|
2748 |
/* Check all attached conditions for inner table rows. */
|
|
2749 |
if (select_cond && !select_cond->val_int()) |
|
2750 |
return NESTED_LOOP_OK; |
|
2751 |
}
|
|
2752 |
join_tab--; |
|
2753 |
/*
|
|
2754 |
The row complemented by nulls might be the first row
|
|
2755 |
of embedding outer joins.
|
|
2756 |
If so, perform the same actions as in the code
|
|
2757 |
for the first regular outer join row above.
|
|
2758 |
*/
|
|
2759 |
for ( ; ; ) |
|
2760 |
{
|
|
2761 |
JOIN_TAB *first_unmatched= join_tab->first_unmatched; |
|
2762 |
if ((first_unmatched= first_unmatched->first_upper) && first_unmatched->last_inner != join_tab) |
|
2763 |
first_unmatched= 0; |
|
2764 |
join_tab->first_unmatched= first_unmatched; |
|
2765 |
if (! first_unmatched) |
|
2766 |
break; |
|
2767 |
first_unmatched->found= 1; |
|
2768 |
for (JOIN_TAB *tab= first_unmatched; tab <= join_tab; tab++) |
|
2769 |
{
|
|
2770 |
if (tab->select_cond && !tab->select_cond->val_int()) |
|
2771 |
{
|
|
2772 |
join->return_tab= tab; |
|
2773 |
return NESTED_LOOP_OK; |
|
2774 |
}
|
|
2775 |
}
|
|
2776 |
}
|
|
2777 |
/*
|
|
2778 |
The row complemented by nulls satisfies all conditions
|
|
2779 |
attached to inner tables.
|
|
2780 |
Send the row complemented by nulls to be joined with the
|
|
2781 |
remaining tables.
|
|
2782 |
*/
|
|
2783 |
return (*join_tab->next_select)(join, join_tab+1, 0); |
|
2784 |
}
|
|
2785 |
||
1039.2.7
by Jay Pipes
Yet more style and indentation cleanups. |
2786 |
enum_nested_loop_state flush_cached_records(JOIN *join, JOIN_TAB *join_tab, bool skip_last) |
1039.2.2
by Jay Pipes
Phase 2 of JOIN refactoring. |
2787 |
{
|
2788 |
enum_nested_loop_state rc= NESTED_LOOP_OK; |
|
2789 |
int error; |
|
2790 |
READ_RECORD *info; |
|
2791 |
||
2792 |
join_tab->table->null_row= 0; |
|
2793 |
if (!join_tab->cache.records) |
|
2794 |
return NESTED_LOOP_OK; /* Nothing to do */ |
|
2795 |
if (skip_last) |
|
2796 |
(void) store_record_in_cache(&join_tab->cache); // Must save this for later |
|
2797 |
if (join_tab->use_quick == 2) |
|
2798 |
{
|
|
2799 |
if (join_tab->select->quick) |
|
2800 |
{ /* Used quick select last. reset it */ |
|
2801 |
delete join_tab->select->quick; |
|
2802 |
join_tab->select->quick=0; |
|
2803 |
}
|
|
2804 |
}
|
|
2805 |
/* read through all records */
|
|
2806 |
if ((error=join_init_read_record(join_tab))) |
|
2807 |
{
|
|
2808 |
reset_cache_write(&join_tab->cache); |
|
2809 |
return error < 0 ? NESTED_LOOP_NO_MORE_ROWS: NESTED_LOOP_ERROR; |
|
2810 |
}
|
|
2811 |
||
2812 |
for (JOIN_TAB *tmp=join->join_tab; tmp != join_tab ; tmp++) |
|
2813 |
{
|
|
2814 |
tmp->status=tmp->table->status; |
|
2815 |
tmp->table->status=0; |
|
2816 |
}
|
|
2817 |
||
2818 |
info= &join_tab->read_record; |
|
2819 |
do
|
|
2820 |
{
|
|
2821 |
if (join->session->killed) |
|
2822 |
{
|
|
2823 |
join->session->send_kill_message(); |
|
2824 |
return NESTED_LOOP_KILLED; // Aborted by user /* purecov: inspected */ |
|
2825 |
}
|
|
2826 |
SQL_SELECT *select=join_tab->select; |
|
2827 |
if (rc == NESTED_LOOP_OK && |
|
2828 |
(!join_tab->cache.select || !join_tab->cache.select->skip_record())) |
|
2829 |
{
|
|
2830 |
uint32_t i; |
|
2831 |
reset_cache_read(&join_tab->cache); |
|
2832 |
for (i=(join_tab->cache.records- (skip_last ? 1 : 0)) ; i-- > 0 ;) |
|
2833 |
{
|
|
2834 |
read_cached_record(join_tab); |
|
2835 |
if (!select || !select->skip_record()) |
|
2836 |
{
|
|
2837 |
int res= 0; |
|
2838 |
if (!join_tab->check_weed_out_table || |
|
2839 |
!(res= do_sj_dups_weedout(join->session, join_tab->check_weed_out_table))) |
|
2840 |
{
|
|
2841 |
rc= (join_tab->next_select)(join,join_tab+1,0); |
|
2842 |
if (rc != NESTED_LOOP_OK && rc != NESTED_LOOP_NO_MORE_ROWS) |
|
2843 |
{
|
|
2844 |
reset_cache_write(&join_tab->cache); |
|
2845 |
return rc; |
|
2846 |
}
|
|
2847 |
}
|
|
2848 |
if (res == -1) |
|
2849 |
return NESTED_LOOP_ERROR; |
|
2850 |
}
|
|
2851 |
}
|
|
2852 |
}
|
|
2853 |
} while (!(error=info->read_record(info))); |
|
2854 |
||
2855 |
if (skip_last) |
|
2856 |
read_cached_record(join_tab); // Restore current record |
|
2857 |
reset_cache_write(&join_tab->cache); |
|
2858 |
if (error > 0) // Fatal error |
|
2859 |
return NESTED_LOOP_ERROR; /* purecov: inspected */ |
|
2860 |
for (JOIN_TAB *tmp2=join->join_tab; tmp2 != join_tab ; tmp2++) |
|
2861 |
tmp2->table->status=tmp2->status; |
|
2862 |
return NESTED_LOOP_OK; |
|
2863 |
}
|
|
2864 |
||
2865 |
/*****************************************************************************
|
|
2866 |
DESCRIPTION
|
|
2867 |
Functions that end one nested loop iteration. Different functions
|
|
2868 |
are used to support GROUP BY clause and to redirect records
|
|
2869 |
to a table (e.g. in case of SELECT into a temporary table) or to the
|
|
2870 |
network client.
|
|
2871 |
||
2872 |
RETURN VALUES
|
|
2873 |
NESTED_LOOP_OK - the record has been successfully handled
|
|
2874 |
NESTED_LOOP_ERROR - a fatal error (like table corruption)
|
|
2875 |
was detected
|
|
2876 |
NESTED_LOOP_KILLED - thread shutdown was requested while processing
|
|
2877 |
the record
|
|
2878 |
NESTED_LOOP_QUERY_LIMIT - the record has been successfully handled;
|
|
2879 |
additionally, the nested loop produced the
|
|
2880 |
number of rows specified in the LIMIT clause
|
|
2881 |
for the query
|
|
2882 |
NESTED_LOOP_CURSOR_LIMIT - the record has been successfully handled;
|
|
2883 |
additionally, there is a cursor and the nested
|
|
2884 |
loop algorithm produced the number of rows
|
|
2885 |
that is specified for current cursor fetch
|
|
2886 |
operation.
|
|
2887 |
All return values except NESTED_LOOP_OK abort the nested loop.
|
|
2888 |
*****************************************************************************/
|
|
2889 |
enum_nested_loop_state end_send(JOIN *join, JOIN_TAB *, bool end_of_records) |
|
2890 |
{
|
|
2891 |
if (! end_of_records) |
|
2892 |
{
|
|
2893 |
int error; |
|
2894 |
if (join->having && join->having->val_int() == 0) |
|
2895 |
return NESTED_LOOP_OK; // Didn't match having |
|
2896 |
error= 0; |
|
2897 |
if (join->do_send_rows) |
|
2898 |
error=join->result->send_data(*join->fields); |
|
2899 |
if (error) |
|
2900 |
return NESTED_LOOP_ERROR; /* purecov: inspected */ |
|
2901 |
if (++join->send_records >= join->unit->select_limit_cnt && join->do_send_rows) |
|
2902 |
{
|
|
2903 |
if (join->select_options & OPTION_FOUND_ROWS) |
|
2904 |
{
|
|
2905 |
JOIN_TAB *jt=join->join_tab; |
|
2906 |
if ((join->tables == 1) && !join->tmp_table && !join->sort_and_group |
|
2907 |
&& !join->send_group_parts && !join->having && !jt->select_cond && |
|
2908 |
!(jt->select && jt->select->quick) && |
|
2909 |
(jt->table->file->ha_table_flags() & HA_STATS_RECORDS_IS_EXACT) && |
|
2910 |
(jt->ref.key < 0)) |
|
2911 |
{
|
|
2912 |
/* Join over all rows in table; Return number of found rows */
|
|
2913 |
Table *table= jt->table; |
|
2914 |
||
2915 |
join->select_options^= OPTION_FOUND_ROWS; |
|
2916 |
if (table->sort.record_pointers || |
|
2917 |
(table->sort.io_cache && my_b_inited(table->sort.io_cache))) |
|
2918 |
{
|
|
2919 |
/* Using filesort */
|
|
2920 |
join->send_records= table->sort.found_records; |
|
2921 |
}
|
|
2922 |
else
|
|
2923 |
{
|
|
2924 |
table->file->info(HA_STATUS_VARIABLE); |
|
2925 |
join->send_records= table->file->stats.records; |
|
2926 |
}
|
|
2927 |
}
|
|
2928 |
else
|
|
2929 |
{
|
|
2930 |
join->do_send_rows= 0; |
|
2931 |
if (join->unit->fake_select_lex) |
|
2932 |
join->unit->fake_select_lex->select_limit= 0; |
|
2933 |
return NESTED_LOOP_OK; |
|
2934 |
}
|
|
2935 |
}
|
|
2936 |
return NESTED_LOOP_QUERY_LIMIT; // Abort nicely |
|
2937 |
}
|
|
2938 |
else if (join->send_records >= join->fetch_limit) |
|
2939 |
{
|
|
2940 |
/*
|
|
2941 |
There is a server side cursor and all rows for
|
|
2942 |
this fetch request are sent.
|
|
2943 |
*/
|
|
2944 |
return NESTED_LOOP_CURSOR_LIMIT; |
|
2945 |
}
|
|
2946 |
}
|
|
2947 |
||
2948 |
return NESTED_LOOP_OK; |
|
2949 |
}
|
|
2950 |
||
2951 |
enum_nested_loop_state end_write(JOIN *join, JOIN_TAB *, bool end_of_records) |
|
2952 |
{
|
|
2953 |
Table *table= join->tmp_table; |
|
2954 |
||
2955 |
if (join->session->killed) // Aborted by user |
|
2956 |
{
|
|
2957 |
join->session->send_kill_message(); |
|
2958 |
return NESTED_LOOP_KILLED; /* purecov: inspected */ |
|
2959 |
}
|
|
2960 |
if (!end_of_records) |
|
2961 |
{
|
|
2962 |
copy_fields(&join->tmp_table_param); |
|
2963 |
copy_funcs(join->tmp_table_param.items_to_copy); |
|
2964 |
if (!join->having || join->having->val_int()) |
|
2965 |
{
|
|
2966 |
int error; |
|
2967 |
join->found_records++; |
|
2968 |
if ((error=table->file->ha_write_row(table->record[0]))) |
|
2969 |
{
|
|
2970 |
if (!table->file->is_fatal_error(error, HA_CHECK_DUP)) |
|
2971 |
goto end; |
|
2972 |
if (create_myisam_from_heap(join->session, table, |
|
2973 |
join->tmp_table_param.start_recinfo, |
|
2974 |
&join->tmp_table_param.recinfo, |
|
2975 |
error, 1)) |
|
2976 |
return NESTED_LOOP_ERROR; // Not a table_is_full error |
|
2977 |
table->s->uniques= 0; // To ensure rows are the same |
|
2978 |
}
|
|
2979 |
if (++join->send_records >= join->tmp_table_param.end_write_records && join->do_send_rows) |
|
2980 |
{
|
|
2981 |
if (!(join->select_options & OPTION_FOUND_ROWS)) |
|
2982 |
return NESTED_LOOP_QUERY_LIMIT; |
|
2983 |
join->do_send_rows= 0; |
|
2984 |
join->unit->select_limit_cnt= HA_POS_ERROR; |
|
2985 |
return NESTED_LOOP_OK; |
|
2986 |
}
|
|
2987 |
}
|
|
2988 |
}
|
|
2989 |
end: |
|
2990 |
return NESTED_LOOP_OK; |
|
2991 |
}
|
|
2992 |
||
2993 |
/** Group by searching after group record and updating it if possible. */
|
|
2994 |
enum_nested_loop_state end_update(JOIN *join, JOIN_TAB *, bool end_of_records) |
|
2995 |
{
|
|
2996 |
Table *table= join->tmp_table; |
|
2997 |
order_st *group; |
|
2998 |
int error; |
|
2999 |
||
3000 |
if (end_of_records) |
|
3001 |
return NESTED_LOOP_OK; |
|
3002 |
if (join->session->killed) // Aborted by user |
|
3003 |
{
|
|
3004 |
join->session->send_kill_message(); |
|
3005 |
return NESTED_LOOP_KILLED; /* purecov: inspected */ |
|
3006 |
}
|
|
3007 |
||
3008 |
join->found_records++; |
|
3009 |
copy_fields(&join->tmp_table_param); // Groups are copied twice. |
|
3010 |
/* Make a key of group index */
|
|
3011 |
for (group=table->group ; group ; group=group->next) |
|
3012 |
{
|
|
3013 |
Item *item= *group->item; |
|
3014 |
item->save_org_in_field(group->field); |
|
3015 |
/* Store in the used key if the field was 0 */
|
|
3016 |
if (item->maybe_null) |
|
3017 |
group->buff[-1]= (char) group->field->is_null(); |
|
3018 |
}
|
|
3019 |
if (!table->file->index_read_map(table->record[1], |
|
3020 |
join->tmp_table_param.group_buff, |
|
3021 |
HA_WHOLE_KEY, |
|
3022 |
HA_READ_KEY_EXACT)) |
|
3023 |
{ /* Update old record */ |
|
3024 |
table->restoreRecord(); |
|
3025 |
update_tmptable_sum_func(join->sum_funcs,table); |
|
3026 |
if ((error= table->file->ha_update_row(table->record[1], |
|
3027 |
table->record[0]))) |
|
3028 |
{
|
|
3029 |
table->file->print_error(error,MYF(0)); /* purecov: inspected */ |
|
3030 |
return NESTED_LOOP_ERROR; /* purecov: inspected */ |
|
3031 |
}
|
|
3032 |
return NESTED_LOOP_OK; |
|
3033 |
}
|
|
3034 |
||
3035 |
/*
|
|
3036 |
Copy null bits from group key to table
|
|
3037 |
We can't copy all data as the key may have different format
|
|
3038 |
as the row data (for example as with VARCHAR keys)
|
|
3039 |
*/
|
|
3040 |
KEY_PART_INFO *key_part; |
|
3041 |
for (group=table->group,key_part=table->key_info[0].key_part; |
|
3042 |
group ; |
|
3043 |
group=group->next,key_part++) |
|
3044 |
{
|
|
3045 |
if (key_part->null_bit) |
|
3046 |
memcpy(table->record[0]+key_part->offset, group->buff, 1); |
|
3047 |
}
|
|
3048 |
init_tmptable_sum_functions(join->sum_funcs); |
|
3049 |
copy_funcs(join->tmp_table_param.items_to_copy); |
|
3050 |
if ((error=table->file->ha_write_row(table->record[0]))) |
|
3051 |
{
|
|
3052 |
if (create_myisam_from_heap(join->session, table, |
|
3053 |
join->tmp_table_param.start_recinfo, |
|
3054 |
&join->tmp_table_param.recinfo, |
|
3055 |
error, 0)) |
|
3056 |
return NESTED_LOOP_ERROR; // Not a table_is_full error |
|
3057 |
/* Change method to update rows */
|
|
3058 |
table->file->ha_index_init(0, 0); |
|
3059 |
join->join_tab[join->tables-1].next_select= end_unique_update; |
|
3060 |
}
|
|
3061 |
join->send_records++; |
|
3062 |
return NESTED_LOOP_OK; |
|
3063 |
}
|
|
3064 |
||
3065 |
/** Like end_update, but this is done with unique constraints instead of keys. */
|
|
3066 |
enum_nested_loop_state end_unique_update(JOIN *join, JOIN_TAB *, bool end_of_records) |
|
3067 |
{
|
|
3068 |
Table *table= join->tmp_table; |
|
3069 |
int error; |
|
3070 |
||
3071 |
if (end_of_records) |
|
3072 |
return NESTED_LOOP_OK; |
|
3073 |
if (join->session->killed) // Aborted by user |
|
3074 |
{
|
|
3075 |
join->session->send_kill_message(); |
|
3076 |
return NESTED_LOOP_KILLED; /* purecov: inspected */ |
|
3077 |
}
|
|
3078 |
||
3079 |
init_tmptable_sum_functions(join->sum_funcs); |
|
3080 |
copy_fields(&join->tmp_table_param); // Groups are copied twice. |
|
3081 |
copy_funcs(join->tmp_table_param.items_to_copy); |
|
3082 |
||
3083 |
if (!(error= table->file->ha_write_row(table->record[0]))) |
|
3084 |
join->send_records++; // New group |
|
3085 |
else
|
|
3086 |
{
|
|
3087 |
if ((int) table->file->get_dup_key(error) < 0) |
|
3088 |
{
|
|
3089 |
table->file->print_error(error,MYF(0)); /* purecov: inspected */ |
|
3090 |
return NESTED_LOOP_ERROR; /* purecov: inspected */ |
|
3091 |
}
|
|
3092 |
if (table->file->rnd_pos(table->record[1],table->file->dup_ref)) |
|
3093 |
{
|
|
3094 |
table->file->print_error(error,MYF(0)); /* purecov: inspected */ |
|
3095 |
return NESTED_LOOP_ERROR; /* purecov: inspected */ |
|
3096 |
}
|
|
3097 |
table->restoreRecord(); |
|
3098 |
update_tmptable_sum_func(join->sum_funcs,table); |
|
3099 |
if ((error= table->file->ha_update_row(table->record[1], |
|
3100 |
table->record[0]))) |
|
3101 |
{
|
|
3102 |
table->file->print_error(error,MYF(0)); /* purecov: inspected */ |
|
3103 |
return NESTED_LOOP_ERROR; /* purecov: inspected */ |
|
3104 |
}
|
|
3105 |
}
|
|
3106 |
return NESTED_LOOP_OK; |
|
3107 |
}
|
|
3108 |
||
3109 |
/**
|
|
3110 |
allocate group fields or take prepared (cached).
|
|
3111 |
||
3112 |
@param main_join join of current select
|
|
3113 |
@param curr_join current join (join of current select or temporary copy
|
|
3114 |
of it)
|
|
3115 |
||
3116 |
@retval
|
|
3117 |
0 ok
|
|
3118 |
@retval
|
|
3119 |
1 failed
|
|
3120 |
*/
|
|
3121 |
static bool make_group_fields(JOIN *main_join, JOIN *curr_join) |
|
3122 |
{
|
|
3123 |
if (main_join->group_fields_cache.elements) |
|
3124 |
{
|
|
3125 |
curr_join->group_fields= main_join->group_fields_cache; |
|
3126 |
curr_join->sort_and_group= 1; |
|
3127 |
}
|
|
3128 |
else
|
|
3129 |
{
|
|
3130 |
if (alloc_group_fields(curr_join, curr_join->group_list)) |
|
3131 |
return (1); |
|
3132 |
main_join->group_fields_cache= curr_join->group_fields; |
|
3133 |
}
|
|
3134 |
return (0); |
|
3135 |
}
|
|
3136 |
||
3137 |
/**
|
|
3138 |
calc how big buffer we need for comparing group entries.
|
|
3139 |
*/
|
|
3140 |
static void calc_group_buffer(JOIN *join,order_st *group) |
|
3141 |
{
|
|
3142 |
uint32_t key_length=0, parts=0, null_parts=0; |
|
3143 |
||
3144 |
if (group) |
|
3145 |
join->group= 1; |
|
3146 |
for (; group ; group=group->next) |
|
3147 |
{
|
|
3148 |
Item *group_item= *group->item; |
|
3149 |
Field *field= group_item->get_tmp_table_field(); |
|
3150 |
if (field) |
|
3151 |
{
|
|
3152 |
enum_field_types type; |
|
3153 |
if ((type= field->type()) == DRIZZLE_TYPE_BLOB) |
|
3154 |
key_length+=MAX_BLOB_WIDTH; // Can't be used as a key |
|
3155 |
else if (type == DRIZZLE_TYPE_VARCHAR) |
|
3156 |
key_length+= field->field_length + HA_KEY_BLOB_LENGTH; |
|
3157 |
else
|
|
3158 |
key_length+= field->pack_length(); |
|
3159 |
}
|
|
3160 |
else
|
|
3161 |
{
|
|
3162 |
switch (group_item->result_type()) { |
|
3163 |
case REAL_RESULT: |
|
3164 |
key_length+= sizeof(double); |
|
3165 |
break; |
|
3166 |
case INT_RESULT: |
|
3167 |
key_length+= sizeof(int64_t); |
|
3168 |
break; |
|
3169 |
case DECIMAL_RESULT: |
|
3170 |
key_length+= my_decimal_get_binary_size(group_item->max_length - |
|
3171 |
(group_item->decimals ? 1 : 0), |
|
3172 |
group_item->decimals); |
|
3173 |
break; |
|
3174 |
case STRING_RESULT: |
|
3175 |
{
|
|
3176 |
enum enum_field_types type= group_item->field_type(); |
|
3177 |
/*
|
|
3178 |
As items represented as DATE/TIME fields in the group buffer
|
|
3179 |
have STRING_RESULT result type, we increase the length
|
|
3180 |
by 8 as maximum pack length of such fields.
|
|
3181 |
*/
|
|
3182 |
if (type == DRIZZLE_TYPE_DATE || |
|
3183 |
type == DRIZZLE_TYPE_DATETIME || |
|
3184 |
type == DRIZZLE_TYPE_TIMESTAMP) |
|
3185 |
{
|
|
3186 |
key_length+= 8; |
|
3187 |
}
|
|
3188 |
else
|
|
3189 |
{
|
|
3190 |
/*
|
|
3191 |
Group strings are taken as varstrings and require an length field.
|
|
3192 |
A field is not yet created by create_tmp_field()
|
|
3193 |
and the sizes should match up.
|
|
3194 |
*/
|
|
3195 |
key_length+= group_item->max_length + HA_KEY_BLOB_LENGTH; |
|
3196 |
}
|
|
3197 |
break; |
|
3198 |
}
|
|
3199 |
default: |
|
3200 |
/* This case should never be choosen */
|
|
3201 |
assert(0); |
|
3202 |
my_error(ER_OUT_OF_RESOURCES, MYF(ME_FATALERROR)); |
|
3203 |
}
|
|
3204 |
}
|
|
3205 |
parts++; |
|
3206 |
if (group_item->maybe_null) |
|
3207 |
null_parts++; |
|
3208 |
}
|
|
3209 |
join->tmp_table_param.group_length=key_length+null_parts; |
|
3210 |
join->tmp_table_param.group_parts=parts; |
|
3211 |
join->tmp_table_param.group_null_parts=null_parts; |
|
3212 |
}
|
|
3213 |
||
3214 |
/**
|
|
3215 |
Get a list of buffers for saveing last group.
|
|
3216 |
||
3217 |
Groups are saved in reverse order for easyer check loop.
|
|
3218 |
*/
|
|
3219 |
static bool alloc_group_fields(JOIN *join,order_st *group) |
|
3220 |
{
|
|
3221 |
if (group) |
|
3222 |
{
|
|
3223 |
for (; group ; group=group->next) |
|
3224 |
{
|
|
3225 |
Cached_item *tmp=new_Cached_item(join->session, *group->item, false); |
|
3226 |
if (!tmp || join->group_fields.push_front(tmp)) |
|
3227 |
return true; |
|
3228 |
}
|
|
3229 |
}
|
|
3230 |
join->sort_and_group=1; /* Mark for do_select */ |
|
3231 |
return false; |
|
3232 |
}
|
|
3233 |
||
3234 |
/**
|
|
3235 |
@todo
|
|
3236 |
- TODO: this function is here only temporarily until 'greedy_search' is
|
|
3237 |
tested and accepted.
|
|
3238 |
||
3239 |
RETURN VALUES
|
|
3240 |
false ok
|
|
3241 |
true Fatal error
|
|
3242 |
*/
|
|
3243 |
static bool find_best(JOIN *join,table_map rest_tables,uint32_t idx,double record_count, double read_time) |
|
3244 |
{
|
|
3245 |
Session *session= join->session; |
|
3246 |
if (session->killed) |
|
3247 |
return(true); |
|
3248 |
if (!rest_tables) |
|
3249 |
{
|
|
3250 |
read_time+=record_count/(double) TIME_FOR_COMPARE; |
|
3251 |
if (join->sort_by_table && |
|
3252 |
join->sort_by_table != |
|
3253 |
join->positions[join->const_tables].table->table) |
|
3254 |
read_time+=record_count; // We have to make a temp table |
|
3255 |
if (read_time < join->best_read) |
|
3256 |
{
|
|
3257 |
memcpy(join->best_positions, join->positions, sizeof(POSITION)*idx); |
|
3258 |
join->best_read= read_time - 0.001; |
|
3259 |
}
|
|
3260 |
return(false); |
|
3261 |
}
|
|
3262 |
if (read_time+record_count/(double) TIME_FOR_COMPARE >= join->best_read) |
|
3263 |
return(false); /* Found better before */ |
|
3264 |
||
3265 |
JOIN_TAB *s; |
|
3266 |
double best_record_count=DBL_MAX,best_read_time=DBL_MAX; |
|
3267 |
for (JOIN_TAB **pos=join->best_ref+idx ; (s=*pos) ; pos++) |
|
3268 |
{
|
|
3269 |
table_map real_table_bit=s->table->map; |
|
3270 |
if ((rest_tables & real_table_bit) && !(rest_tables & s->dependent) && |
|
3271 |
(!idx|| !check_interleaving_with_nj(join->positions[idx-1].table, s))) |
|
3272 |
{
|
|
3273 |
double records, best; |
|
3274 |
advance_sj_state(rest_tables, s); |
|
3275 |
best_access_path(join, s, session, rest_tables, idx, record_count, |
|
3276 |
read_time); |
|
3277 |
records= join->positions[idx].records_read; |
|
3278 |
best= join->positions[idx].read_time; |
|
3279 |
/*
|
|
3280 |
Go to the next level only if there hasn't been a better key on
|
|
3281 |
this level! This will cut down the search for a lot simple cases!
|
|
3282 |
*/
|
|
3283 |
double current_record_count=record_count*records; |
|
3284 |
double current_read_time=read_time+best; |
|
3285 |
if (best_record_count > current_record_count || |
|
3286 |
best_read_time > current_read_time || |
|
3287 |
(idx == join->const_tables && s->table == join->sort_by_table)) |
|
3288 |
{
|
|
3289 |
if (best_record_count >= current_record_count && |
|
3290 |
best_read_time >= current_read_time && |
|
3291 |
(!(s->key_dependent & rest_tables) || records < 2.0)) |
|
3292 |
{
|
|
3293 |
best_record_count=current_record_count; |
|
3294 |
best_read_time=current_read_time; |
|
3295 |
}
|
|
3296 |
std::swap(join->best_ref[idx], *pos); |
|
3297 |
if (find_best(join,rest_tables & ~real_table_bit,idx+1, |
|
3298 |
current_record_count,current_read_time)) |
|
3299 |
return(true); |
|
3300 |
std::swap(join->best_ref[idx], *pos); |
|
3301 |
}
|
|
3302 |
restore_prev_nj_state(s); |
|
3303 |
restore_prev_sj_state(rest_tables, s); |
|
3304 |
if (join->select_options & SELECT_STRAIGHT_JOIN) |
|
3305 |
break; // Don't test all combinations |
|
3306 |
}
|
|
3307 |
}
|
|
3308 |
return(false); |
|
3309 |
}
|
|
3310 |
||
3311 |
static uint32_t cache_record_length(JOIN *join,uint32_t idx) |
|
3312 |
{
|
|
3313 |
uint32_t length=0; |
|
3314 |
JOIN_TAB **pos,**end; |
|
3315 |
Session *session=join->session; |
|
3316 |
||
3317 |
for (pos=join->best_ref+join->const_tables,end=join->best_ref+idx ; |
|
3318 |
pos != end ; |
|
3319 |
pos++) |
|
3320 |
{
|
|
3321 |
JOIN_TAB *join_tab= *pos; |
|
3322 |
if (!join_tab->used_fieldlength) /* Not calced yet */ |
|
3323 |
calc_used_field_length(session, join_tab); |
|
3324 |
length+=join_tab->used_fieldlength; |
|
3325 |
}
|
|
3326 |
return length; |
|
3327 |
}
|
|
3328 |
||
3329 |
/*
|
|
3330 |
Get the number of different row combinations for subset of partial join
|
|
3331 |
||
3332 |
SYNOPSIS
|
|
3333 |
prev_record_reads()
|
|
3334 |
join The join structure
|
|
3335 |
idx Number of tables in the partial join order (i.e. the
|
|
3336 |
partial join order is in join->positions[0..idx-1])
|
|
3337 |
found_ref Bitmap of tables for which we need to find # of distinct
|
|
3338 |
row combinations.
|
|
3339 |
||
3340 |
DESCRIPTION
|
|
3341 |
Given a partial join order (in join->positions[0..idx-1]) and a subset of
|
|
3342 |
tables within that join order (specified in found_ref), find out how many
|
|
3343 |
distinct row combinations of subset tables will be in the result of the
|
|
3344 |
partial join order.
|
|
3345 |
||
3346 |
This is used as follows: Suppose we have a table accessed with a ref-based
|
|
3347 |
method. The ref access depends on current rows of tables in found_ref.
|
|
3348 |
We want to count # of different ref accesses. We assume two ref accesses
|
|
3349 |
will be different if at least one of access parameters is different.
|
|
3350 |
Example: consider a query
|
|
3351 |
||
3352 |
SELECT * FROM t1, t2, t3 WHERE t1.key=c1 AND t2.key=c2 AND t3.key=t1.field
|
|
3353 |
||
3354 |
and a join order:
|
|
3355 |
t1, ref access on t1.key=c1
|
|
3356 |
t2, ref access on t2.key=c2
|
|
3357 |
t3, ref access on t3.key=t1.field
|
|
3358 |
||
3359 |
For t1: n_ref_scans = 1, n_distinct_ref_scans = 1
|
|
3360 |
For t2: n_ref_scans = records_read(t1), n_distinct_ref_scans=1
|
|
3361 |
For t3: n_ref_scans = records_read(t1)*records_read(t2)
|
|
3362 |
n_distinct_ref_scans = #records_read(t1)
|
|
3363 |
||
3364 |
The reason for having this function (at least the latest version of it)
|
|
3365 |
is that we need to account for buffering in join execution.
|
|
3366 |
||
3367 |
An edge-case example: if we have a non-first table in join accessed via
|
|
3368 |
ref(const) or ref(param) where there is a small number of different
|
|
3369 |
values of param, then the access will likely hit the disk cache and will
|
|
3370 |
not require any disk seeks.
|
|
3371 |
||
3372 |
The proper solution would be to assume an LRU disk cache of some size,
|
|
3373 |
calculate probability of cache hits, etc. For now we just count
|
|
3374 |
identical ref accesses as one.
|
|
3375 |
||
3376 |
RETURN
|
|
3377 |
Expected number of row combinations
|
|
3378 |
*/
|
|
3379 |
static double prev_record_reads(JOIN *join, uint32_t idx, table_map found_ref) |
|
3380 |
{
|
|
3381 |
double found=1.0; |
|
3382 |
POSITION *pos_end= join->positions - 1; |
|
3383 |
for (POSITION *pos= join->positions + idx - 1; pos != pos_end; pos--) |
|
3384 |
{
|
|
3385 |
if (pos->table->table->map & found_ref) |
|
3386 |
{
|
|
3387 |
found_ref|= pos->ref_depend_map; |
|
3388 |
/*
|
|
3389 |
For the case of "t1 LEFT JOIN t2 ON ..." where t2 is a const table
|
|
3390 |
with no matching row we will get position[t2].records_read==0.
|
|
3391 |
Actually the size of output is one null-complemented row, therefore
|
|
3392 |
we will use value of 1 whenever we get records_read==0.
|
|
3393 |
||
3394 |
Note
|
|
3395 |
- the above case can't occur if inner part of outer join has more
|
|
3396 |
than one table: table with no matches will not be marked as const.
|
|
3397 |
||
3398 |
- Ideally we should add 1 to records_read for every possible null-
|
|
3399 |
complemented row. We're not doing it because: 1. it will require
|
|
3400 |
non-trivial code and add overhead. 2. The value of records_read
|
|
3401 |
is an inprecise estimate and adding 1 (or, in the worst case,
|
|
3402 |
#max_nested_outer_joins=64-1) will not make it any more precise.
|
|
3403 |
*/
|
|
3404 |
if (pos->records_read > DBL_EPSILON) |
|
3405 |
found*= pos->records_read; |
|
3406 |
}
|
|
3407 |
}
|
|
3408 |
return found; |
|
3409 |
}
|
|
3410 |
||
3411 |
/**
|
|
3412 |
Set up join struct according to best position.
|
|
3413 |
*/
|
|
3414 |
static bool get_best_combination(JOIN *join) |
|
3415 |
{
|
|
3416 |
uint32_t i,tablenr; |
|
3417 |
table_map used_tables; |
|
3418 |
JOIN_TAB *join_tab,*j; |
|
3419 |
KEYUSE *keyuse; |
|
3420 |
uint32_t table_count; |
|
3421 |
Session *session=join->session; |
|
3422 |
||
3423 |
table_count=join->tables; |
|
3424 |
if (!(join->join_tab=join_tab= |
|
3425 |
(JOIN_TAB*) session->alloc(sizeof(JOIN_TAB)*table_count))) |
|
3426 |
return(true); |
|
3427 |
||
3428 |
join->full_join=0; |
|
3429 |
||
3430 |
used_tables= OUTER_REF_TABLE_BIT; // Outer row is already read |
|
3431 |
for (j=join_tab, tablenr=0 ; tablenr < table_count ; tablenr++,j++) |
|
3432 |
{
|
|
3433 |
Table *form; |
|
3434 |
*j= *join->best_positions[tablenr].table; |
|
3435 |
form=join->table[tablenr]=j->table; |
|
3436 |
used_tables|= form->map; |
|
3437 |
form->reginfo.join_tab=j; |
|
3438 |
if (!*j->on_expr_ref) |
|
3439 |
form->reginfo.not_exists_optimize=0; // Only with LEFT JOIN |
|
3440 |
if (j->type == JT_CONST) |
|
3441 |
continue; // Handled in make_join_stat.. |
|
3442 |
||
3443 |
j->ref.key = -1; |
|
3444 |
j->ref.key_parts=0; |
|
3445 |
||
3446 |
if (j->type == JT_SYSTEM) |
|
3447 |
continue; |
|
3448 |
if (j->keys.none() || !(keyuse= join->best_positions[tablenr].key)) |
|
3449 |
{
|
|
3450 |
j->type=JT_ALL; |
|
3451 |
if (tablenr != join->const_tables) |
|
3452 |
join->full_join=1; |
|
3453 |
}
|
|
3454 |
else if (create_ref_for_key(join, j, keyuse, used_tables)) |
|
3455 |
return(true); // Something went wrong |
|
3456 |
}
|
|
3457 |
||
3458 |
for (i=0 ; i < table_count ; i++) |
|
3459 |
join->map2table[join->join_tab[i].table->tablenr]=join->join_tab+i; |
|
3460 |
update_depend_map(join); |
|
3461 |
return(0); |
|
3462 |
}
|
|
3463 |
||
3464 |
/** Save const tables first as used tables. */
|
|
3465 |
static void set_position(JOIN *join,uint32_t idx,JOIN_TAB *table,KEYUSE *key) |
|
3466 |
{
|
|
3467 |
join->positions[idx].table= table; |
|
3468 |
join->positions[idx].key=key; |
|
3469 |
join->positions[idx].records_read=1.0; /* This is a const table */ |
|
3470 |
join->positions[idx].ref_depend_map= 0; |
|
3471 |
||
3472 |
/* Move the const table as down as possible in best_ref */
|
|
3473 |
JOIN_TAB **pos=join->best_ref+idx+1; |
|
3474 |
JOIN_TAB *next=join->best_ref[idx]; |
|
3475 |
for (;next != table ; pos++) |
|
3476 |
{
|
|
3477 |
JOIN_TAB *tmp=pos[0]; |
|
3478 |
pos[0]=next; |
|
3479 |
next=tmp; |
|
3480 |
}
|
|
3481 |
join->best_ref[idx]=table; |
|
3482 |
}
|
|
3483 |
||
3484 |
/**
|
|
3485 |
Selects and invokes a search strategy for an optimal query plan.
|
|
3486 |
||
3487 |
The function checks user-configurable parameters that control the search
|
|
3488 |
strategy for an optimal plan, selects the search method and then invokes
|
|
3489 |
it. Each specific optimization procedure stores the final optimal plan in
|
|
3490 |
the array 'join->best_positions', and the cost of the plan in
|
|
3491 |
'join->best_read'.
|
|
3492 |
||
3493 |
@param join pointer to the structure providing all context info for
|
|
3494 |
the query
|
|
3495 |
@param join_tables set of the tables in the query
|
|
3496 |
||
3497 |
@todo
|
|
3498 |
'MAX_TABLES+2' denotes the old implementation of find_best before
|
|
3499 |
the greedy version. Will be removed when greedy_search is approved.
|
|
3500 |
||
3501 |
@retval
|
|
3502 |
false ok
|
|
3503 |
@retval
|
|
3504 |
true Fatal error
|
|
3505 |
*/
|
|
3506 |
static bool choose_plan(JOIN *join, table_map join_tables) |
|
3507 |
{
|
|
3508 |
uint32_t search_depth= join->session->variables.optimizer_search_depth; |
|
3509 |
uint32_t prune_level= join->session->variables.optimizer_prune_level; |
|
3510 |
bool straight_join= test(join->select_options & SELECT_STRAIGHT_JOIN); |
|
3511 |
||
3512 |
join->cur_embedding_map= 0; |
|
3513 |
reset_nj_counters(join->join_list); |
|
3514 |
/*
|
|
3515 |
if (SELECT_STRAIGHT_JOIN option is set)
|
|
3516 |
reorder tables so dependent tables come after tables they depend
|
|
3517 |
on, otherwise keep tables in the order they were specified in the query
|
|
3518 |
else
|
|
3519 |
Apply heuristic: pre-sort all access plans with respect to the number of
|
|
3520 |
records accessed.
|
|
3521 |
*/
|
|
3522 |
my_qsort(join->best_ref + join->const_tables, |
|
3523 |
join->tables - join->const_tables, sizeof(JOIN_TAB*), |
|
3524 |
straight_join ? join_tab_cmp_straight : join_tab_cmp); |
|
3525 |
join->cur_emb_sj_nests= 0; |
|
3526 |
if (straight_join) |
|
3527 |
{
|
|
3528 |
optimize_straight_join(join, join_tables); |
|
3529 |
}
|
|
3530 |
else
|
|
3531 |
{
|
|
3532 |
if (search_depth == MAX_TABLES+2) |
|
3533 |
{ /* |
|
3534 |
TODO: 'MAX_TABLES+2' denotes the old implementation of find_best before
|
|
3535 |
the greedy version. Will be removed when greedy_search is approved.
|
|
3536 |
*/
|
|
3537 |
join->best_read= DBL_MAX; |
|
3538 |
if (find_best(join, join_tables, join->const_tables, 1.0, 0.0)) |
|
3539 |
return(true); |
|
3540 |
}
|
|
3541 |
else
|
|
3542 |
{
|
|
3543 |
if (search_depth == 0) |
|
3544 |
/* Automatically determine a reasonable value for 'search_depth' */
|
|
3545 |
search_depth= determine_search_depth(join); |
|
3546 |
if (greedy_search(join, join_tables, search_depth, prune_level)) |
|
3547 |
return(true); |
|
3548 |
}
|
|
3549 |
}
|
|
3550 |
||
3551 |
/*
|
|
3552 |
Store the cost of this query into a user variable
|
|
3553 |
Don't update last_query_cost for statements that are not "flat joins" :
|
|
3554 |
i.e. they have subqueries, unions or call stored procedures.
|
|
3555 |
TODO: calculate a correct cost for a query with subqueries and UNIONs.
|
|
3556 |
*/
|
|
3557 |
if (join->session->lex->is_single_level_stmt()) |
|
3558 |
join->session->status_var.last_query_cost= join->best_read; |
|
3559 |
return(false); |
|
3560 |
}
|
|
3561 |
||
3562 |
/**
|
|
3563 |
Find the best access path for an extension of a partial execution
|
|
3564 |
plan and add this path to the plan.
|
|
3565 |
||
3566 |
The function finds the best access path to table 's' from the passed
|
|
3567 |
partial plan where an access path is the general term for any means to
|
|
3568 |
access the data in 's'. An access path may use either an index or a scan,
|
|
3569 |
whichever is cheaper. The input partial plan is passed via the array
|
|
3570 |
'join->positions' of length 'idx'. The chosen access method for 's' and its
|
|
3571 |
cost are stored in 'join->positions[idx]'.
|
|
3572 |
||
3573 |
@param join pointer to the structure providing all context info
|
|
3574 |
for the query
|
|
3575 |
@param s the table to be joined by the function
|
|
3576 |
@param session thread for the connection that submitted the query
|
|
3577 |
@param remaining_tables set of tables not included into the partial plan yet
|
|
3578 |
@param idx the length of the partial plan
|
|
3579 |
@param record_count estimate for the number of records returned by the
|
|
3580 |
partial plan
|
|
3581 |
@param read_time the cost of the partial plan
|
|
3582 |
||
3583 |
@return
|
|
3584 |
None
|
|
3585 |
*/
|
|
3586 |
static void best_access_path(JOIN *join, |
|
3587 |
JOIN_TAB *s, |
|
3588 |
Session *session, |
|
3589 |
table_map remaining_tables, |
|
3590 |
uint32_t idx, |
|
3591 |
double record_count, |
|
3592 |
double) |
|
3593 |
{
|
|
3594 |
KEYUSE *best_key= 0; |
|
3595 |
uint32_t best_max_key_part= 0; |
|
3596 |
bool found_constraint= 0; |
|
3597 |
double best= DBL_MAX; |
|
3598 |
double best_time= DBL_MAX; |
|
3599 |
double records= DBL_MAX; |
|
3600 |
table_map best_ref_depends_map= 0; |
|
3601 |
double tmp; |
|
3602 |
ha_rows rec; |
|
3603 |
uint32_t best_is_sj_inside_out= 0; |
|
3604 |
||
3605 |
if (s->keyuse) |
|
3606 |
{ /* Use key if possible */ |
|
3607 |
Table *table= s->table; |
|
3608 |
KEYUSE *keyuse,*start_key=0; |
|
3609 |
double best_records= DBL_MAX; |
|
3610 |
uint32_t max_key_part=0; |
|
3611 |
uint64_t bound_sj_equalities= 0; |
|
3612 |
bool try_sj_inside_out= false; |
|
3613 |
/*
|
|
3614 |
Discover the bound equalites. We need to do this, if
|
|
3615 |
1. The next table is an SJ-inner table, and
|
|
3616 |
2. It is the first table from that semijoin, and
|
|
3617 |
3. We're not within a semi-join range (i.e. all semi-joins either have
|
|
3618 |
all or none of their tables in join_table_map), except
|
|
3619 |
s->emb_sj_nest (which we've just entered).
|
|
3620 |
3. All correlation references from this sj-nest are bound
|
|
3621 |
*/
|
|
3622 |
if (s->emb_sj_nest && // (1) |
|
3623 |
s->emb_sj_nest->sj_in_exprs < 64 && |
|
3624 |
((remaining_tables & s->emb_sj_nest->sj_inner_tables) == // (2) |
|
3625 |
s->emb_sj_nest->sj_inner_tables) && // (2) |
|
3626 |
join->cur_emb_sj_nests == s->emb_sj_nest->sj_inner_tables && // (3) |
|
3627 |
!(remaining_tables & s->emb_sj_nest->nested_join->sj_corr_tables)) // (4) |
|
3628 |
{
|
|
3629 |
/* This table is an InsideOut scan candidate */
|
|
3630 |
bound_sj_equalities= get_bound_sj_equalities(s->emb_sj_nest, |
|
3631 |
remaining_tables); |
|
3632 |
try_sj_inside_out= true; |
|
3633 |
}
|
|
3634 |
||
3635 |
/* Test how we can use keys */
|
|
3636 |
rec= s->records/MATCHING_ROWS_IN_OTHER_TABLE; // Assumed records/key |
|
3637 |
for (keyuse=s->keyuse ; keyuse->table == table ;) |
|
3638 |
{
|
|
3639 |
key_part_map found_part= 0; |
|
3640 |
table_map found_ref= 0; |
|
3641 |
uint32_t key= keyuse->key; |
|
3642 |
KEY *keyinfo= table->key_info+key; |
|
3643 |
/* Bitmap of keyparts where the ref access is over 'keypart=const': */
|
|
3644 |
key_part_map const_part= 0; |
|
3645 |
/* The or-null keypart in ref-or-null access: */
|
|
3646 |
key_part_map ref_or_null_part= 0; |
|
3647 |
||
3648 |
/* Calculate how many key segments of the current key we can use */
|
|
3649 |
start_key= keyuse; |
|
3650 |
uint64_t handled_sj_equalities=0; |
|
3651 |
key_part_map sj_insideout_map= 0; |
|
3652 |
||
3653 |
do /* For each keypart */ |
|
3654 |
{
|
|
3655 |
uint32_t keypart= keyuse->keypart; |
|
3656 |
table_map best_part_found_ref= 0; |
|
3657 |
double best_prev_record_reads= DBL_MAX; |
|
3658 |
||
3659 |
do /* For each way to access the keypart */ |
|
3660 |
{
|
|
3661 |
||
3662 |
/*
|
|
3663 |
if 1. expression doesn't refer to forward tables
|
|
3664 |
2. we won't get two ref-or-null's
|
|
3665 |
*/
|
|
3666 |
if (!(remaining_tables & keyuse->used_tables) && |
|
3667 |
!(ref_or_null_part && (keyuse->optimize & |
|
3668 |
KEY_OPTIMIZE_REF_OR_NULL))) |
|
3669 |
{
|
|
3670 |
found_part|= keyuse->keypart_map; |
|
3671 |
if (!(keyuse->used_tables & ~join->const_table_map)) |
|
3672 |
const_part|= keyuse->keypart_map; |
|
3673 |
||
3674 |
double tmp2= prev_record_reads(join, idx, (found_ref | |
|
3675 |
keyuse->used_tables)); |
|
3676 |
if (tmp2 < best_prev_record_reads) |
|
3677 |
{
|
|
3678 |
best_part_found_ref= keyuse->used_tables & ~join->const_table_map; |
|
3679 |
best_prev_record_reads= tmp2; |
|
3680 |
}
|
|
3681 |
if (rec > keyuse->ref_table_rows) |
|
3682 |
rec= keyuse->ref_table_rows; |
|
3683 |
/*
|
|
3684 |
If there is one 'key_column IS NULL' expression, we can
|
|
3685 |
use this ref_or_null optimisation of this field
|
|
3686 |
*/
|
|
3687 |
if (keyuse->optimize & KEY_OPTIMIZE_REF_OR_NULL) |
|
3688 |
ref_or_null_part |= keyuse->keypart_map; |
|
3689 |
}
|
|
3690 |
||
3691 |
if (try_sj_inside_out && keyuse->sj_pred_no != UINT_MAX) |
|
3692 |
{
|
|
3693 |
if (!(remaining_tables & keyuse->used_tables)) |
|
3694 |
bound_sj_equalities |= 1UL << keyuse->sj_pred_no; |
|
3695 |
else
|
|
3696 |
{
|
|
3697 |
handled_sj_equalities |= 1UL << keyuse->sj_pred_no; |
|
3698 |
sj_insideout_map |= ((key_part_map)1) << keyuse->keypart; |
|
3699 |
}
|
|
3700 |
}
|
|
3701 |
||
3702 |
keyuse++; |
|
3703 |
} while (keyuse->table == table && keyuse->key == key && |
|
3704 |
keyuse->keypart == keypart); |
|
3705 |
found_ref|= best_part_found_ref; |
|
3706 |
} while (keyuse->table == table && keyuse->key == key); |
|
3707 |
||
3708 |
/*
|
|
3709 |
Assume that that each key matches a proportional part of table.
|
|
3710 |
*/
|
|
3711 |
if (!found_part && !handled_sj_equalities) |
|
3712 |
continue; // Nothing usable found |
|
3713 |
||
3714 |
if (rec < MATCHING_ROWS_IN_OTHER_TABLE) |
|
3715 |
rec= MATCHING_ROWS_IN_OTHER_TABLE; // Fix for small tables |
|
3716 |
||
3717 |
bool sj_inside_out_scan= false; |
|
3718 |
{
|
|
3719 |
found_constraint= 1; |
|
3720 |
/*
|
|
3721 |
Check if InsideOut scan is applicable:
|
|
3722 |
1. All IN-equalities are either "bound" or "handled"
|
|
3723 |
2. Index keyparts are
|
|
3724 |
...
|
|
3725 |
*/
|
|
3726 |
if (try_sj_inside_out && |
|
3727 |
table->covering_keys.test(key) && |
|
3728 |
(handled_sj_equalities | bound_sj_equalities) == // (1) |
|
3729 |
PREV_BITS(uint64_t, s->emb_sj_nest->sj_in_exprs)) // (1) |
|
3730 |
{
|
|
3731 |
uint32_t n_fixed_parts= max_part_bit(found_part); |
|
3732 |
if (n_fixed_parts != keyinfo->key_parts && |
|
3733 |
(PREV_BITS(uint, n_fixed_parts) | sj_insideout_map) == |
|
3734 |
PREV_BITS(uint, keyinfo->key_parts)) |
|
3735 |
{
|
|
3736 |
/*
|
|
3737 |
Not all parts are fixed. Produce bitmap of remaining bits and
|
|
3738 |
check if all of them are covered.
|
|
3739 |
*/
|
|
3740 |
sj_inside_out_scan= true; |
|
3741 |
if (!n_fixed_parts) |
|
3742 |
{
|
|
3743 |
/*
|
|
3744 |
It's a confluent ref scan.
|
|
3745 |
||
3746 |
That is, all found KEYUSE elements refer to IN-equalities,
|
|
3747 |
and there is really no ref access because there is no
|
|
3748 |
t.keypart0 = {bound expression}
|
|
3749 |
||
3750 |
Calculate the cost of complete loose index scan.
|
|
3751 |
*/
|
|
3752 |
records= (double)s->table->file->stats.records; |
|
3753 |
||
3754 |
/* The cost is entire index scan cost (divided by 2) */
|
|
3755 |
best_time= s->table->file->index_only_read_time(key, records); |
|
3756 |
||
3757 |
/* Now figure how many different keys we will get */
|
|
3758 |
ulong rpc; |
|
3759 |
if ((rpc= keyinfo->rec_per_key[keyinfo->key_parts-1])) |
|
3760 |
records= records / rpc; |
|
3761 |
start_key= NULL; |
|
3762 |
}
|
|
3763 |
}
|
|
3764 |
}
|
|
3765 |
||
3766 |
/*
|
|
3767 |
Check if we found full key
|
|
3768 |
*/
|
|
3769 |
if (found_part == PREV_BITS(uint,keyinfo->key_parts) && |
|
3770 |
!ref_or_null_part) |
|
3771 |
{ /* use eq key */ |
|
3772 |
max_key_part= UINT32_MAX; |
|
3773 |
if ((keyinfo->flags & (HA_NOSAME | HA_NULL_PART_KEY)) == HA_NOSAME) |
|
3774 |
{
|
|
3775 |
tmp = prev_record_reads(join, idx, found_ref); |
|
3776 |
records=1.0; |
|
3777 |
}
|
|
3778 |
else
|
|
3779 |
{
|
|
3780 |
if (!found_ref) |
|
3781 |
{ /* We found a const key */ |
|
3782 |
/*
|
|
3783 |
ReuseRangeEstimateForRef-1:
|
|
3784 |
We get here if we've found a ref(const) (c_i are constants):
|
|
3785 |
"(keypart1=c1) AND ... AND (keypartN=cN)" [ref_const_cond]
|
|
3786 |
||
3787 |
If range optimizer was able to construct a "range"
|
|
3788 |
access on this index, then its condition "quick_cond" was
|
|
3789 |
eqivalent to ref_const_cond (*), and we can re-use E(#rows)
|
|
3790 |
from the range optimizer.
|
|
3791 |
||
3792 |
Proof of (*): By properties of range and ref optimizers
|
|
3793 |
quick_cond will be equal or tighther than ref_const_cond.
|
|
3794 |
ref_const_cond already covers "smallest" possible interval -
|
|
3795 |
a singlepoint interval over all keyparts. Therefore,
|
|
3796 |
quick_cond is equivalent to ref_const_cond (if it was an
|
|
3797 |
empty interval we wouldn't have got here).
|
|
3798 |
*/
|
|
3799 |
if (table->quick_keys.test(key)) |
|
3800 |
records= (double) table->quick_rows[key]; |
|
3801 |
else
|
|
3802 |
{
|
|
3803 |
/* quick_range couldn't use key! */
|
|
3804 |
records= (double) s->records/rec; |
|
3805 |
}
|
|
3806 |
}
|
|
3807 |
else
|
|
3808 |
{
|
|
3809 |
if (!(records=keyinfo->rec_per_key[keyinfo->key_parts-1])) |
|
3810 |
{ /* Prefer longer keys */ |
|
3811 |
records= |
|
3812 |
((double) s->records / (double) rec * |
|
3813 |
(1.0 + |
|
3814 |
((double) (table->s->max_key_length-keyinfo->key_length) / |
|
3815 |
(double) table->s->max_key_length))); |
|
3816 |
if (records < 2.0) |
|
3817 |
records=2.0; /* Can't be as good as a unique */ |
|
3818 |
}
|
|
3819 |
/*
|
|
3820 |
ReuseRangeEstimateForRef-2: We get here if we could not reuse
|
|
3821 |
E(#rows) from range optimizer. Make another try:
|
|
3822 |
||
3823 |
If range optimizer produced E(#rows) for a prefix of the ref
|
|
3824 |
access we're considering, and that E(#rows) is lower then our
|
|
3825 |
current estimate, make an adjustment. The criteria of when we
|
|
3826 |
can make an adjustment is a special case of the criteria used
|
|
3827 |
in ReuseRangeEstimateForRef-3.
|
|
3828 |
*/
|
|
3829 |
if (table->quick_keys.test(key) && |
|
3830 |
const_part & (1 << table->quick_key_parts[key]) && |
|
3831 |
table->quick_n_ranges[key] == 1 && |
|
3832 |
records > (double) table->quick_rows[key]) |
|
3833 |
{
|
|
3834 |
records= (double) table->quick_rows[key]; |
|
3835 |
}
|
|
3836 |
}
|
|
3837 |
/* Limit the number of matched rows */
|
|
3838 |
tmp= records; |
|
3839 |
set_if_smaller(tmp, (double) session->variables.max_seeks_for_key); |
|
3840 |
if (table->covering_keys.test(key)) |
|
3841 |
{
|
|
3842 |
/* we can use only index tree */
|
|
3843 |
tmp= record_count * table->file->index_only_read_time(key, tmp); |
|
3844 |
}
|
|
3845 |
else
|
|
1067.4.4
by Nathan Williams
The rest of the files in the drizzled directory were purged of the cmin macro and replace with std::min (except for the definition in globals.h and 1 usage in stacktrace.cc). |
3846 |
tmp= record_count * min(tmp,s->worst_seeks); |
1039.2.2
by Jay Pipes
Phase 2 of JOIN refactoring. |
3847 |
}
|
3848 |
}
|
|
3849 |
else
|
|
3850 |
{
|
|
3851 |
/*
|
|
3852 |
Use as much key-parts as possible and a uniq key is better
|
|
3853 |
than a not unique key
|
|
3854 |
Set tmp to (previous record count) * (records / combination)
|
|
3855 |
*/
|
|
3856 |
if ((found_part & 1) && |
|
3857 |
(!(table->file->index_flags(key, 0, 0) & HA_ONLY_WHOLE_INDEX) || |
|
3858 |
found_part == PREV_BITS(uint,keyinfo->key_parts))) |
|
3859 |
{
|
|
3860 |
max_key_part= max_part_bit(found_part); |
|
3861 |
/*
|
|
3862 |
ReuseRangeEstimateForRef-3:
|
|
3863 |
We're now considering a ref[or_null] access via
|
|
3864 |
(t.keypart1=e1 AND ... AND t.keypartK=eK) [ OR
|
|
3865 |
(same-as-above but with one cond replaced
|
|
3866 |
with "t.keypart_i IS NULL")] (**)
|
|
3867 |
||
3868 |
Try re-using E(#rows) from "range" optimizer:
|
|
3869 |
We can do so if "range" optimizer used the same intervals as
|
|
3870 |
in (**). The intervals used by range optimizer may be not
|
|
3871 |
available at this point (as "range" access might have choosen to
|
|
3872 |
create quick select over another index), so we can't compare
|
|
3873 |
them to (**). We'll make indirect judgements instead.
|
|
3874 |
The sufficient conditions for re-use are:
|
|
3875 |
(C1) All e_i in (**) are constants, i.e. found_ref==false. (if
|
|
3876 |
this is not satisfied we have no way to know which ranges
|
|
3877 |
will be actually scanned by 'ref' until we execute the
|
|
3878 |
join)
|
|
3879 |
(C2) max #key parts in 'range' access == K == max_key_part (this
|
|
3880 |
is apparently a necessary requirement)
|
|
3881 |
||
3882 |
We also have a property that "range optimizer produces equal or
|
|
3883 |
tighter set of scan intervals than ref(const) optimizer". Each
|
|
3884 |
of the intervals in (**) are "tightest possible" intervals when
|
|
3885 |
one limits itself to using keyparts 1..K (which we do in #2).
|
|
3886 |
From here it follows that range access used either one, or
|
|
3887 |
both of the (I1) and (I2) intervals:
|
|
3888 |
||
3889 |
(t.keypart1=c1 AND ... AND t.keypartK=eK) (I1)
|
|
3890 |
(same-as-above but with one cond replaced
|
|
3891 |
with "t.keypart_i IS NULL") (I2)
|
|
3892 |
||
3893 |
The remaining part is to exclude the situation where range
|
|
3894 |
optimizer used one interval while we're considering
|
|
3895 |
ref-or-null and looking for estimate for two intervals. This
|
|
3896 |
is done by last limitation:
|
|
3897 |
||
3898 |
(C3) "range optimizer used (have ref_or_null?2:1) intervals"
|
|
3899 |
*/
|
|
3900 |
if (table->quick_keys.test(key) && !found_ref && //(C1) |
|
3901 |
table->quick_key_parts[key] == max_key_part && //(C2) |
|
3902 |
table->quick_n_ranges[key] == 1+((ref_or_null_part)?1:0)) //(C3) |
|
3903 |
{
|
|
3904 |
tmp= records= (double) table->quick_rows[key]; |
|
3905 |
}
|
|
3906 |
else
|
|
3907 |
{
|
|
3908 |
/* Check if we have statistic about the distribution */
|
|
3909 |
if ((records= keyinfo->rec_per_key[max_key_part-1])) |
|
3910 |
{
|
|
3911 |
/*
|
|
3912 |
Fix for the case where the index statistics is too
|
|
3913 |
optimistic: If
|
|
3914 |
(1) We're considering ref(const) and there is quick select
|
|
3915 |
on the same index,
|
|
3916 |
(2) and that quick select uses more keyparts (i.e. it will
|
|
3917 |
scan equal/smaller interval then this ref(const))
|
|
3918 |
(3) and E(#rows) for quick select is higher then our
|
|
3919 |
estimate,
|
|
3920 |
Then
|
|
3921 |
We'll use E(#rows) from quick select.
|
|
3922 |
||
3923 |
Q: Why do we choose to use 'ref'? Won't quick select be
|
|
3924 |
cheaper in some cases ?
|
|
3925 |
TODO: figure this out and adjust the plan choice if needed.
|
|
3926 |
*/
|
|
3927 |
if (!found_ref && table->quick_keys.test(key) && // (1) |
|
3928 |
table->quick_key_parts[key] > max_key_part && // (2) |
|
3929 |
records < (double)table->quick_rows[key]) // (3) |
|
3930 |
records= (double)table->quick_rows[key]; |
|
3931 |
||
3932 |
tmp= records; |
|
3933 |
}
|
|
3934 |
else
|
|
3935 |
{
|
|
3936 |
/*
|
|
3937 |
Assume that the first key part matches 1% of the file
|
|
3938 |
and that the whole key matches 10 (duplicates) or 1
|
|
3939 |
(unique) records.
|
|
3940 |
Assume also that more key matches proportionally more
|
|
3941 |
records
|
|
3942 |
This gives the formula:
|
|
3943 |
records = (x * (b-a) + a*c-b)/(c-1)
|
|
3944 |
||
3945 |
b = records matched by whole key
|
|
3946 |
a = records matched by first key part (1% of all records?)
|
|
3947 |
c = number of key parts in key
|
|
3948 |
x = used key parts (1 <= x <= c)
|
|
3949 |
*/
|
|
3950 |
double rec_per_key; |
|
3951 |
if (!(rec_per_key=(double) |
|
3952 |
keyinfo->rec_per_key[keyinfo->key_parts-1])) |
|
3953 |
rec_per_key=(double) s->records/rec+1; |
|
3954 |
||
3955 |
if (!s->records) |
|
3956 |
tmp = 0; |
|
3957 |
else if (rec_per_key/(double) s->records >= 0.01) |
|
3958 |
tmp = rec_per_key; |
|
3959 |
else
|
|
3960 |
{
|
|
3961 |
double a=s->records*0.01; |
|
3962 |
if (keyinfo->key_parts > 1) |
|
3963 |
tmp= (max_key_part * (rec_per_key - a) + |
|
3964 |
a*keyinfo->key_parts - rec_per_key)/ |
|
3965 |
(keyinfo->key_parts-1); |
|
3966 |
else
|
|
3967 |
tmp= a; |
|
3968 |
set_if_bigger(tmp,1.0); |
|
3969 |
}
|
|
3970 |
records = (uint32_t) tmp; |
|
3971 |
}
|
|
3972 |
||
3973 |
if (ref_or_null_part) |
|
3974 |
{
|
|
3975 |
/* We need to do two key searches to find key */
|
|
3976 |
tmp *= 2.0; |
|
3977 |
records *= 2.0; |
|
3978 |
}
|
|
3979 |
||
3980 |
/*
|
|
3981 |
ReuseRangeEstimateForRef-4: We get here if we could not reuse
|
|
3982 |
E(#rows) from range optimizer. Make another try:
|
|
3983 |
||
3984 |
If range optimizer produced E(#rows) for a prefix of the ref
|
|
3985 |
access we're considering, and that E(#rows) is lower then our
|
|
3986 |
current estimate, make the adjustment.
|
|
3987 |
||
3988 |
The decision whether we can re-use the estimate from the range
|
|
3989 |
optimizer is the same as in ReuseRangeEstimateForRef-3,
|
|
3990 |
applied to first table->quick_key_parts[key] key parts.
|
|
3991 |
*/
|
|
3992 |
if (table->quick_keys.test(key) && |
|
3993 |
table->quick_key_parts[key] <= max_key_part && |
|
3994 |
const_part & (1 << table->quick_key_parts[key]) && |
|
3995 |
table->quick_n_ranges[key] == 1 + ((ref_or_null_part & |
|
3996 |
const_part) ? 1 : 0) && |
|
3997 |
records > (double) table->quick_rows[key]) |
|
3998 |
{
|
|
3999 |
tmp= records= (double) table->quick_rows[key]; |
|
4000 |
}
|
|
4001 |
}
|
|
4002 |
||
4003 |
/* Limit the number of matched rows */
|
|
4004 |
set_if_smaller(tmp, (double) session->variables.max_seeks_for_key); |
|
4005 |
if (table->covering_keys.test(key)) |
|
4006 |
{
|
|
4007 |
/* we can use only index tree */
|
|
4008 |
tmp= record_count * table->file->index_only_read_time(key, tmp); |
|
4009 |
}
|
|
4010 |
else
|
|
1067.4.4
by Nathan Williams
The rest of the files in the drizzled directory were purged of the cmin macro and replace with std::min (except for the definition in globals.h and 1 usage in stacktrace.cc). |
4011 |
tmp= record_count * min(tmp,s->worst_seeks); |
1039.2.2
by Jay Pipes
Phase 2 of JOIN refactoring. |
4012 |
}
|
4013 |
else
|
|
4014 |
tmp= best_time; // Do nothing |
|
4015 |
}
|
|
4016 |
||
4017 |
if (sj_inside_out_scan && !start_key) |
|
4018 |
{
|
|
4019 |
tmp= tmp/2; |
|
4020 |
if (records) |
|
4021 |
records= records/2; |
|
4022 |
}
|
|
4023 |
||
4024 |
}
|
|
4025 |
if (tmp < best_time - records/(double) TIME_FOR_COMPARE) |
|
4026 |
{
|
|
4027 |
best_time= tmp + records/(double) TIME_FOR_COMPARE; |
|
4028 |
best= tmp; |
|
4029 |
best_records= records; |
|
4030 |
best_key= start_key; |
|
4031 |
best_max_key_part= max_key_part; |
|
4032 |
best_ref_depends_map= found_ref; |
|
4033 |
best_is_sj_inside_out= sj_inside_out_scan; |
|
4034 |
}
|
|
4035 |
}
|
|
4036 |
records= best_records; |
|
4037 |
}
|
|
4038 |
||
4039 |
/*
|
|
4040 |
Don't test table scan if it can't be better.
|
|
4041 |
Prefer key lookup if we would use the same key for scanning.
|
|
4042 |
||
4043 |
Don't do a table scan on InnoDB tables, if we can read the used
|
|
4044 |
parts of the row from any of the used index.
|
|
4045 |
This is because table scans uses index and we would not win
|
|
4046 |
anything by using a table scan.
|
|
4047 |
||
4048 |
A word for word translation of the below if-statement in sergefp's
|
|
4049 |
understanding: we check if we should use table scan if:
|
|
4050 |
(1) The found 'ref' access produces more records than a table scan
|
|
4051 |
(or index scan, or quick select), or 'ref' is more expensive than
|
|
4052 |
any of them.
|
|
4053 |
(2) This doesn't hold: the best way to perform table scan is to to perform
|
|
4054 |
'range' access using index IDX, and the best way to perform 'ref'
|
|
4055 |
access is to use the same index IDX, with the same or more key parts.
|
|
4056 |
(note: it is not clear how this rule is/should be extended to
|
|
4057 |
index_merge quick selects)
|
|
4058 |
(3) See above note about InnoDB.
|
|
4059 |
(4) NOT ("FORCE INDEX(...)" is used for table and there is 'ref' access
|
|
4060 |
path, but there is no quick select)
|
|
4061 |
If the condition in the above brackets holds, then the only possible
|
|
4062 |
"table scan" access method is ALL/index (there is no quick select).
|
|
4063 |
Since we have a 'ref' access path, and FORCE INDEX instructs us to
|
|
4064 |
choose it over ALL/index, there is no need to consider a full table
|
|
4065 |
scan.
|
|
4066 |
*/
|
|
4067 |
if ((records >= s->found_records || best > s->read_time) && // (1) |
|
4068 |
!(s->quick && best_key && s->quick->index == best_key->key && // (2) |
|
4069 |
best_max_key_part >= s->table->quick_key_parts[best_key->key]) &&// (2) |
|
4070 |
!((s->table->file->ha_table_flags() & HA_TABLE_SCAN_ON_INDEX) && // (3) |
|
4071 |
! s->table->covering_keys.none() && best_key && !s->quick) &&// (3) |
|
4072 |
!(s->table->force_index && best_key && !s->quick)) // (4) |
|
4073 |
{ // Check full join |
|
4074 |
ha_rows rnd_records= s->found_records; |
|
4075 |
/*
|
|
4076 |
If there is a filtering condition on the table (i.e. ref analyzer found
|
|
4077 |
at least one "table.keyXpartY= exprZ", where exprZ refers only to tables
|
|
4078 |
preceding this table in the join order we're now considering), then
|
|
4079 |
assume that 25% of the rows will be filtered out by this condition.
|
|
4080 |
||
4081 |
This heuristic is supposed to force tables used in exprZ to be before
|
|
4082 |
this table in join order.
|
|
4083 |
*/
|
|
4084 |
if (found_constraint) |
|
4085 |
rnd_records-= rnd_records/4; |
|
4086 |
||
4087 |
/*
|
|
4088 |
If applicable, get a more accurate estimate. Don't use the two
|
|
4089 |
heuristics at once.
|
|
4090 |
*/
|
|
4091 |
if (s->table->quick_condition_rows != s->found_records) |
|
4092 |
rnd_records= s->table->quick_condition_rows; |
|
4093 |
||
4094 |
/*
|
|
4095 |
Range optimizer never proposes a RANGE if it isn't better
|
|
4096 |
than FULL: so if RANGE is present, it's always preferred to FULL.
|
|
4097 |
Here we estimate its cost.
|
|
4098 |
*/
|
|
4099 |
if (s->quick) |
|
4100 |
{
|
|
4101 |
/*
|
|
4102 |
For each record we:
|
|
4103 |
- read record range through 'quick'
|
|
4104 |
- skip rows which does not satisfy WHERE constraints
|
|
4105 |
TODO:
|
|
4106 |
We take into account possible use of join cache for ALL/index
|
|
4107 |
access (see first else-branch below), but we don't take it into
|
|
4108 |
account here for range/index_merge access. Find out why this is so.
|
|
4109 |
*/
|
|
4110 |
tmp= record_count * |
|
4111 |
(s->quick->read_time + |
|
4112 |
(s->found_records - rnd_records)/(double) TIME_FOR_COMPARE); |
|
4113 |
}
|
|
4114 |
else
|
|
4115 |
{
|
|
4116 |
/* Estimate cost of reading table. */
|
|
4117 |
tmp= s->table->file->scan_time(); |
|
4118 |
if (s->table->map & join->outer_join) // Can't use join cache |
|
4119 |
{
|
|
4120 |
/*
|
|
4121 |
For each record we have to:
|
|
4122 |
- read the whole table record
|
|
4123 |
- skip rows which does not satisfy join condition
|
|
4124 |
*/
|
|
4125 |
tmp= record_count * |
|
4126 |
(tmp + |
|
4127 |
(s->records - rnd_records)/(double) TIME_FOR_COMPARE); |
|
4128 |
}
|
|
4129 |
else
|
|
4130 |
{
|
|
4131 |
/* We read the table as many times as join buffer becomes full. */
|
|
4132 |
tmp*= (1.0 + floor((double) cache_record_length(join,idx) * |
|
4133 |
record_count / |
|
4134 |
(double) session->variables.join_buff_size)); |
|
4135 |
/*
|
|
4136 |
We don't make full cartesian product between rows in the scanned
|
|
4137 |
table and existing records because we skip all rows from the
|
|
4138 |
scanned table, which does not satisfy join condition when
|
|
4139 |
we read the table (see flush_cached_records for details). Here we
|
|
4140 |
take into account cost to read and skip these records.
|
|
4141 |
*/
|
|
4142 |
tmp+= (s->records - rnd_records)/(double) TIME_FOR_COMPARE; |
|
4143 |
}
|
|
4144 |
}
|
|
4145 |
||
4146 |
/*
|
|
4147 |
We estimate the cost of evaluating WHERE clause for found records
|
|
4148 |
as record_count * rnd_records / TIME_FOR_COMPARE. This cost plus
|
|
4149 |
tmp give us total cost of using Table SCAN
|
|
4150 |
*/
|
|
4151 |
if (best == DBL_MAX || |
|
4152 |
(tmp + record_count/(double) TIME_FOR_COMPARE*rnd_records < |
|
4153 |
best + record_count/(double) TIME_FOR_COMPARE*records)) |
|
4154 |
{
|
|
4155 |
/*
|
|
4156 |
If the table has a range (s->quick is set) make_join_select()
|
|
4157 |
will ensure that this will be used
|
|
4158 |
*/
|
|
4159 |
best= tmp; |
|
4160 |
records= rows2double(rnd_records); |
|
4161 |
best_key= 0; |
|
4162 |
/* range/index_merge/ALL/index access method are "independent", so: */
|
|
4163 |
best_ref_depends_map= 0; |
|
4164 |
best_is_sj_inside_out= false; |
|
4165 |
}
|
|
4166 |
}
|
|
4167 |
||
4168 |
/* Update the cost information for the current partial plan */
|
|
4169 |
join->positions[idx].records_read= records; |
|
4170 |
join->positions[idx].read_time= best; |
|
4171 |
join->positions[idx].key= best_key; |
|
4172 |
join->positions[idx].table= s; |
|
4173 |
join->positions[idx].ref_depend_map= best_ref_depends_map; |
|
4174 |
join->positions[idx].use_insideout_scan= best_is_sj_inside_out; |
|
4175 |
||
4176 |
if (!best_key && |
|
4177 |
idx == join->const_tables && |
|
4178 |
s->table == join->sort_by_table && |
|
4179 |
join->unit->select_limit_cnt >= records) |
|
4180 |
join->sort_by_table= (Table*) 1; // Must use temporary table |
|
4181 |
||
4182 |
return; |
|
4183 |
}
|
|
4184 |
||
4185 |
/**
|
|
4186 |
Select the best ways to access the tables in a query without reordering them.
|
|
4187 |
||
4188 |
Find the best access paths for each query table and compute their costs
|
|
4189 |
according to their order in the array 'join->best_ref' (thus without
|
|
4190 |
reordering the join tables). The function calls sequentially
|
|
4191 |
'best_access_path' for each table in the query to select the best table
|
|
4192 |
access method. The final optimal plan is stored in the array
|
|
4193 |
'join->best_positions', and the corresponding cost in 'join->best_read'.
|
|
4194 |
||
4195 |
@param join pointer to the structure providing all context info for
|
|
4196 |
the query
|
|
4197 |
@param join_tables set of the tables in the query
|
|
4198 |
||
4199 |
@note
|
|
4200 |
This function can be applied to:
|
|
4201 |
- queries with STRAIGHT_JOIN
|
|
4202 |
- internally to compute the cost of an arbitrary QEP
|
|
4203 |
@par
|
|
4204 |
Thus 'optimize_straight_join' can be used at any stage of the query
|
|
4205 |
optimization process to finalize a QEP as it is.
|
|
4206 |
*/
|
|
4207 |
static void optimize_straight_join(JOIN *join, table_map join_tables) |
|
4208 |
{
|
|
4209 |
JOIN_TAB *s; |
|
4210 |
uint32_t idx= join->const_tables; |
|
4211 |
double record_count= 1.0; |
|
4212 |
double read_time= 0.0; |
|
4213 |
||
4214 |
for (JOIN_TAB **pos= join->best_ref + idx ; (s= *pos) ; pos++) |
|
4215 |
{
|
|
4216 |
/* Find the best access method from 's' to the current partial plan */
|
|
4217 |
advance_sj_state(join_tables, s); |
|
4218 |
best_access_path(join, s, join->session, join_tables, idx, |
|
4219 |
record_count, read_time); |
|
4220 |
/* compute the cost of the new plan extended with 's' */
|
|
4221 |
record_count*= join->positions[idx].records_read; |
|
4222 |
read_time+= join->positions[idx].read_time; |
|
4223 |
join_tables&= ~(s->table->map); |
|
4224 |
++idx; |
|
4225 |
}
|
|
4226 |
||
4227 |
read_time+= record_count / (double) TIME_FOR_COMPARE; |
|
4228 |
if (join->sort_by_table && |
|
4229 |
join->sort_by_table != join->positions[join->const_tables].table->table) |
|
4230 |
read_time+= record_count; // We have to make a temp table |
|
4231 |
memcpy(join->best_positions, join->positions, sizeof(POSITION)*idx); |
|
4232 |
join->best_read= read_time; |
|
4233 |
}
|
|
4234 |
||
4235 |
/**
|
|
4236 |
Find a good, possibly optimal, query execution plan (QEP) by a greedy search.
|
|
4237 |
||
4238 |
The search procedure uses a hybrid greedy/exhaustive search with controlled
|
|
4239 |
exhaustiveness. The search is performed in N = card(remaining_tables)
|
|
4240 |
steps. Each step evaluates how promising is each of the unoptimized tables,
|
|
4241 |
selects the most promising table, and extends the current partial QEP with
|
|
4242 |
that table. Currenly the most 'promising' table is the one with least
|
|
4243 |
expensive extension.\
|
|
4244 |
||
4245 |
There are two extreme cases:
|
|
4246 |
-# When (card(remaining_tables) < search_depth), the estimate finds the
|
|
4247 |
best complete continuation of the partial QEP. This continuation can be
|
|
4248 |
used directly as a result of the search.
|
|
4249 |
-# When (search_depth == 1) the 'best_extension_by_limited_search'
|
|
4250 |
consideres the extension of the current QEP with each of the remaining
|
|
4251 |
unoptimized tables.
|
|
4252 |
||
4253 |
All other cases are in-between these two extremes. Thus the parameter
|
|
4254 |
'search_depth' controlls the exhaustiveness of the search. The higher the
|
|
4255 |
value, the longer the optimizaton time and possibly the better the
|
|
4256 |
resulting plan. The lower the value, the fewer alternative plans are
|
|
4257 |
estimated, but the more likely to get a bad QEP.
|
|
4258 |
||
4259 |
All intermediate and final results of the procedure are stored in 'join':
|
|
4260 |
- join->positions : modified for every partial QEP that is explored
|
|
4261 |
- join->best_positions: modified for the current best complete QEP
|
|
4262 |
- join->best_read : modified for the current best complete QEP
|
|
4263 |
- join->best_ref : might be partially reordered
|
|
4264 |
||
4265 |
The final optimal plan is stored in 'join->best_positions', and its
|
|
4266 |
corresponding cost in 'join->best_read'.
|
|
4267 |
||
4268 |
@note
|
|
4269 |
The following pseudocode describes the algorithm of 'greedy_search':
|
|
4270 |
||
4271 |
@code
|
|
4272 |
procedure greedy_search
|
|
4273 |
input: remaining_tables
|
|
4274 |
output: pplan;
|
|
4275 |
{
|
|
4276 |
pplan = <>;
|
|
4277 |
do {
|
|
4278 |
(t, a) = best_extension(pplan, remaining_tables);
|
|
4279 |
pplan = concat(pplan, (t, a));
|
|
4280 |
remaining_tables = remaining_tables - t;
|
|
4281 |
} while (remaining_tables != {})
|
|
4282 |
return pplan;
|
|
4283 |
}
|
|
4284 |
||
4285 |
@endcode
|
|
4286 |
where 'best_extension' is a placeholder for a procedure that selects the
|
|
4287 |
most "promising" of all tables in 'remaining_tables'.
|
|
4288 |
Currently this estimate is performed by calling
|
|
4289 |
'best_extension_by_limited_search' to evaluate all extensions of the
|
|
4290 |
current QEP of size 'search_depth', thus the complexity of 'greedy_search'
|
|
4291 |
mainly depends on that of 'best_extension_by_limited_search'.
|
|
4292 |
||
4293 |
@par
|
|
4294 |
If 'best_extension()' == 'best_extension_by_limited_search()', then the
|
|
4295 |
worst-case complexity of this algorithm is <=
|
|
4296 |
O(N*N^search_depth/search_depth). When serch_depth >= N, then the
|
|
4297 |
complexity of greedy_search is O(N!).
|
|
4298 |
||
4299 |
@par
|
|
4300 |
In the future, 'greedy_search' might be extended to support other
|
|
4301 |
implementations of 'best_extension', e.g. some simpler quadratic procedure.
|
|
4302 |
||
4303 |
@param join pointer to the structure providing all context info
|
|
4304 |
for the query
|
|
4305 |
@param remaining_tables set of tables not included into the partial plan yet
|
|
4306 |
@param search_depth controlls the exhaustiveness of the search
|
|
4307 |
@param prune_level the pruning heuristics that should be applied during
|
|
4308 |
search
|
|
4309 |
||
4310 |
@retval
|
|
4311 |
false ok
|
|
4312 |
@retval
|
|
4313 |
true Fatal error
|
|
4314 |
*/
|
|
4315 |
static bool greedy_search(JOIN *join, |
|
4316 |
table_map remaining_tables, |
|
4317 |
uint32_t search_depth, |
|
4318 |
uint32_t prune_level) |
|
4319 |
{
|
|
4320 |
double record_count= 1.0; |
|
4321 |
double read_time= 0.0; |
|
4322 |
uint32_t idx= join->const_tables; // index into 'join->best_ref' |
|
4323 |
uint32_t best_idx; |
|
4324 |
uint32_t size_remain; // cardinality of remaining_tables |
|
4325 |
POSITION best_pos; |
|
4326 |
JOIN_TAB *best_table; // the next plan node to be added to the curr QEP |
|
4327 |
||
4328 |
/* number of tables that remain to be optimized */
|
|
4329 |
size_remain= my_count_bits(remaining_tables); |
|
4330 |
||
4331 |
do { |
|
4332 |
/* Find the extension of the current QEP with the lowest cost */
|
|
4333 |
join->best_read= DBL_MAX; |
|
4334 |
if (best_extension_by_limited_search(join, remaining_tables, idx, record_count, |
|
4335 |
read_time, search_depth, prune_level)) |
|
4336 |
return(true); |
|
4337 |
||
4338 |
if (size_remain <= search_depth) |
|
4339 |
{
|
|
4340 |
/*
|
|
4341 |
'join->best_positions' contains a complete optimal extension of the
|
|
4342 |
current partial QEP.
|
|
4343 |
*/
|
|
4344 |
return(false); |
|
4345 |
}
|
|
4346 |
||
4347 |
/* select the first table in the optimal extension as most promising */
|
|
4348 |
best_pos= join->best_positions[idx]; |
|
4349 |
best_table= best_pos.table; |
|
4350 |
/*
|
|
4351 |
Each subsequent loop of 'best_extension_by_limited_search' uses
|
|
4352 |
'join->positions' for cost estimates, therefore we have to update its
|
|
4353 |
value.
|
|
4354 |
*/
|
|
4355 |
join->positions[idx]= best_pos; |
|
4356 |
||
4357 |
/* find the position of 'best_table' in 'join->best_ref' */
|
|
4358 |
best_idx= idx; |
|
4359 |
JOIN_TAB *pos= join->best_ref[best_idx]; |
|
4360 |
while (pos && best_table != pos) |
|
4361 |
pos= join->best_ref[++best_idx]; |
|
4362 |
assert((pos != NULL)); // should always find 'best_table' |
|
4363 |
/* move 'best_table' at the first free position in the array of joins */
|
|
4364 |
std::swap(join->best_ref[idx], join->best_ref[best_idx]); |
|
4365 |
||
4366 |
/* compute the cost of the new plan extended with 'best_table' */
|
|
4367 |
record_count*= join->positions[idx].records_read; |
|
4368 |
read_time+= join->positions[idx].read_time; |
|
4369 |
||
4370 |
remaining_tables&= ~(best_table->table->map); |
|
4371 |
--size_remain; |
|
4372 |
++idx; |
|
4373 |
} while (true); |
|
4374 |
}
|
|
4375 |
||
4376 |
||
4377 |
/**
|
|
4378 |
Find a good, possibly optimal, query execution plan (QEP) by a possibly
|
|
4379 |
exhaustive search.
|
|
4380 |
||
4381 |
The procedure searches for the optimal ordering of the query tables in set
|
|
4382 |
'remaining_tables' of size N, and the corresponding optimal access paths to
|
|
4383 |
each table. The choice of a table order and an access path for each table
|
|
4384 |
constitutes a query execution plan (QEP) that fully specifies how to
|
|
4385 |
execute the query.
|
|
4386 |
||
4387 |
The maximal size of the found plan is controlled by the parameter
|
|
4388 |
'search_depth'. When search_depth == N, the resulting plan is complete and
|
|
4389 |
can be used directly as a QEP. If search_depth < N, the found plan consists
|
|
4390 |
of only some of the query tables. Such "partial" optimal plans are useful
|
|
4391 |
only as input to query optimization procedures, and cannot be used directly
|
|
4392 |
to execute a query.
|
|
4393 |
||
4394 |
The algorithm begins with an empty partial plan stored in 'join->positions'
|
|
4395 |
and a set of N tables - 'remaining_tables'. Each step of the algorithm
|
|
4396 |
evaluates the cost of the partial plan extended by all access plans for
|
|
4397 |
each of the relations in 'remaining_tables', expands the current partial
|
|
4398 |
plan with the access plan that results in lowest cost of the expanded
|
|
4399 |
partial plan, and removes the corresponding relation from
|
|
4400 |
'remaining_tables'. The algorithm continues until it either constructs a
|
|
4401 |
complete optimal plan, or constructs an optimal plartial plan with size =
|
|
4402 |
search_depth.
|
|
4403 |
||
4404 |
The final optimal plan is stored in 'join->best_positions'. The
|
|
4405 |
corresponding cost of the optimal plan is in 'join->best_read'.
|
|
4406 |
||
4407 |
@note
|
|
4408 |
The procedure uses a recursive depth-first search where the depth of the
|
|
4409 |
recursion (and thus the exhaustiveness of the search) is controlled by the
|
|
4410 |
parameter 'search_depth'.
|
|
4411 |
||
4412 |
@note
|
|
4413 |
The pseudocode below describes the algorithm of
|
|
4414 |
'best_extension_by_limited_search'. The worst-case complexity of this
|
|
4415 |
algorithm is O(N*N^search_depth/search_depth). When serch_depth >= N, then
|
|
4416 |
the complexity of greedy_search is O(N!).
|
|
4417 |
||
4418 |
@code
|
|
4419 |
procedure best_extension_by_limited_search(
|
|
4420 |
pplan in, // in, partial plan of tables-joined-so-far
|
|
4421 |
pplan_cost, // in, cost of pplan
|
|
4422 |
remaining_tables, // in, set of tables not referenced in pplan
|
|
4423 |
best_plan_so_far, // in/out, best plan found so far
|
|
4424 |
best_plan_so_far_cost,// in/out, cost of best_plan_so_far
|
|
4425 |
search_depth) // in, maximum size of the plans being considered
|
|
4426 |
{
|
|
4427 |
for each table T from remaining_tables
|
|
4428 |
{
|
|
4429 |
// Calculate the cost of using table T as above
|
|
4430 |
cost = complex-series-of-calculations;
|
|
4431 |
||
4432 |
// Add the cost to the cost so far.
|
|
4433 |
pplan_cost+= cost;
|
|
4434 |
||
4435 |
if (pplan_cost >= best_plan_so_far_cost)
|
|
4436 |
// pplan_cost already too great, stop search
|
|
4437 |
continue;
|
|
4438 |
||
4439 |
pplan= expand pplan by best_access_method;
|
|
4440 |
remaining_tables= remaining_tables - table T;
|
|
4441 |
if (remaining_tables is not an empty set
|
|
4442 |
and
|
|
4443 |
search_depth > 1)
|
|
4444 |
{
|
|
4445 |
best_extension_by_limited_search(pplan, pplan_cost,
|
|
4446 |
remaining_tables,
|
|
4447 |
best_plan_so_far,
|
|
4448 |
best_plan_so_far_cost,
|
|
4449 |
search_depth - 1);
|
|
4450 |
}
|
|
4451 |
else
|
|
4452 |
{
|
|
4453 |
best_plan_so_far_cost= pplan_cost;
|
|
4454 |
best_plan_so_far= pplan;
|
|
4455 |
}
|
|
4456 |
}
|
|
4457 |
}
|
|
4458 |
@endcode
|
|
4459 |
||
4460 |
@note
|
|
4461 |
When 'best_extension_by_limited_search' is called for the first time,
|
|
4462 |
'join->best_read' must be set to the largest possible value (e.g. DBL_MAX).
|
|
4463 |
The actual implementation provides a way to optionally use pruning
|
|
4464 |
heuristic (controlled by the parameter 'prune_level') to reduce the search
|
|
4465 |
space by skipping some partial plans.
|
|
4466 |
||
4467 |
@note
|
|
4468 |
The parameter 'search_depth' provides control over the recursion
|
|
4469 |
depth, and thus the size of the resulting optimal plan.
|
|
4470 |
||
4471 |
@param join pointer to the structure providing all context info
|
|
4472 |
for the query
|
|
4473 |
@param remaining_tables set of tables not included into the partial plan yet
|
|
4474 |
@param idx length of the partial QEP in 'join->positions';
|
|
4475 |
since a depth-first search is used, also corresponds
|
|
4476 |
to the current depth of the search tree;
|
|
4477 |
also an index in the array 'join->best_ref';
|
|
4478 |
@param record_count estimate for the number of records returned by the
|
|
4479 |
best partial plan
|
|
4480 |
@param read_time the cost of the best partial plan
|
|
4481 |
@param search_depth maximum depth of the recursion and thus size of the
|
|
4482 |
found optimal plan
|
|
4483 |
(0 < search_depth <= join->tables+1).
|
|
4484 |
@param prune_level pruning heuristics that should be applied during
|
|
4485 |
optimization
|
|
4486 |
(values: 0 = EXHAUSTIVE, 1 = PRUNE_BY_TIME_OR_ROWS)
|
|
4487 |
||
4488 |
@retval
|
|
4489 |
false ok
|
|
4490 |
@retval
|
|
4491 |
true Fatal error
|
|
4492 |
*/
|
|
4493 |
static bool best_extension_by_limited_search(JOIN *join, |
|
4494 |
table_map remaining_tables, |
|
4495 |
uint32_t idx, |
|
4496 |
double record_count, |
|
4497 |
double read_time, |
|
4498 |
uint32_t search_depth, |
|
4499 |
uint32_t prune_level) |
|
4500 |
{
|
|
4501 |
Session *session= join->session; |
|
4502 |
if (session->killed) // Abort |
|
4503 |
return(true); |
|
4504 |
||
4505 |
/*
|
|
4506 |
'join' is a partial plan with lower cost than the best plan so far,
|
|
4507 |
so continue expanding it further with the tables in 'remaining_tables'.
|
|
4508 |
*/
|
|
4509 |
JOIN_TAB *s; |
|
4510 |
double best_record_count= DBL_MAX; |
|
4511 |
double best_read_time= DBL_MAX; |
|
4512 |
||
4513 |
for (JOIN_TAB **pos= join->best_ref + idx ; (s= *pos) ; pos++) |
|
4514 |
{
|
|
4515 |
table_map real_table_bit= s->table->map; |
|
4516 |
if ((remaining_tables & real_table_bit) && |
|
4517 |
!(remaining_tables & s->dependent) && |
|
4518 |
(!idx || !check_interleaving_with_nj(join->positions[idx-1].table, s))) |
|
4519 |
{
|
|
4520 |
double current_record_count, current_read_time; |
|
4521 |
advance_sj_state(remaining_tables, s); |
|
4522 |
||
4523 |
/*
|
|
4524 |
psergey-insideout-todo:
|
|
4525 |
when best_access_path() detects it could do an InsideOut scan or
|
|
4526 |
some other scan, have it return an insideout scan and a flag that
|
|
4527 |
requests to "fork" this loop iteration. (Q: how does that behave
|
|
4528 |
when the depth is insufficient??)
|
|
4529 |
*/
|
|
4530 |
/* Find the best access method from 's' to the current partial plan */
|
|
4531 |
best_access_path(join, s, session, remaining_tables, idx, |
|
4532 |
record_count, read_time); |
|
4533 |
/* Compute the cost of extending the plan with 's' */
|
|
4534 |
current_record_count= record_count * join->positions[idx].records_read; |
|
4535 |
current_read_time= read_time + join->positions[idx].read_time; |
|
4536 |
||
4537 |
/* Expand only partial plans with lower cost than the best QEP so far */
|
|
4538 |
if ((current_read_time + |
|
4539 |
current_record_count / (double) TIME_FOR_COMPARE) >= join->best_read) |
|
4540 |
{
|
|
4541 |
restore_prev_nj_state(s); |
|
4542 |
restore_prev_sj_state(remaining_tables, s); |
|
4543 |
continue; |
|
4544 |
}
|
|
4545 |
||
4546 |
/*
|
|
4547 |
Prune some less promising partial plans. This heuristic may miss
|
|
4548 |
the optimal QEPs, thus it results in a non-exhaustive search.
|
|
4549 |
*/
|
|
4550 |
if (prune_level == 1) |
|
4551 |
{
|
|
4552 |
if (best_record_count > current_record_count || |
|
4553 |
best_read_time > current_read_time || |
|
4554 |
(idx == join->const_tables && s->table == join->sort_by_table)) // 's' is the first table in the QEP |
|
4555 |
{
|
|
4556 |
if (best_record_count >= current_record_count && |
|
4557 |
best_read_time >= current_read_time && |
|
4558 |
/* TODO: What is the reasoning behind this condition? */
|
|
4559 |
(!(s->key_dependent & remaining_tables) || |
|
4560 |
join->positions[idx].records_read < 2.0)) |
|
4561 |
{
|
|
4562 |
best_record_count= current_record_count; |
|
4563 |
best_read_time= current_read_time; |
|
4564 |
}
|
|
4565 |
}
|
|
4566 |
else
|
|
4567 |
{
|
|
4568 |
restore_prev_nj_state(s); |
|
4569 |
restore_prev_sj_state(remaining_tables, s); |
|
4570 |
continue; |
|
4571 |
}
|
|
4572 |
}
|
|
4573 |
||
4574 |
if ( (search_depth > 1) && (remaining_tables & ~real_table_bit) ) |
|
4575 |
{ /* Recursively expand the current partial plan */ |
|
4576 |
std::swap(join->best_ref[idx], *pos); |
|
4577 |
if (best_extension_by_limited_search(join, |
|
4578 |
remaining_tables & ~real_table_bit, |
|
4579 |
idx + 1, |
|
4580 |
current_record_count, |
|
4581 |
current_read_time, |
|
4582 |
search_depth - 1, |
|
4583 |
prune_level)) |
|
4584 |
return(true); |
|
4585 |
std::swap(join->best_ref[idx], *pos); |
|
4586 |
}
|
|
4587 |
else
|
|
4588 |
{ /* |
|
4589 |
'join' is either the best partial QEP with 'search_depth' relations,
|
|
4590 |
or the best complete QEP so far, whichever is smaller.
|
|
4591 |
*/
|
|
4592 |
current_read_time+= current_record_count / (double) TIME_FOR_COMPARE; |
|
4593 |
if (join->sort_by_table && |
|
4594 |
join->sort_by_table != |
|
4595 |
join->positions[join->const_tables].table->table) |
|
4596 |
/* We have to make a temp table */
|
|
4597 |
current_read_time+= current_record_count; |
|
4598 |
if ((search_depth == 1) || (current_read_time < join->best_read)) |
|
4599 |
{
|
|
4600 |
memcpy(join->best_positions, join->positions, |
|
4601 |
sizeof(POSITION) * (idx + 1)); |
|
4602 |
join->best_read= current_read_time - 0.001; |
|
4603 |
}
|
|
4604 |
}
|
|
4605 |
restore_prev_nj_state(s); |
|
4606 |
restore_prev_sj_state(remaining_tables, s); |
|
4607 |
}
|
|
4608 |
}
|
|
4609 |
return(false); |
|
4610 |
}
|
|
4611 |
||
4612 |
/**
|
|
4613 |
Heuristic procedure to automatically guess a reasonable degree of
|
|
4614 |
exhaustiveness for the greedy search procedure.
|
|
4615 |
||
4616 |
The procedure estimates the optimization time and selects a search depth
|
|
4617 |
big enough to result in a near-optimal QEP, that doesn't take too long to
|
|
4618 |
find. If the number of tables in the query exceeds some constant, then
|
|
4619 |
search_depth is set to this constant.
|
|
4620 |
||
4621 |
@param join pointer to the structure providing all context info for
|
|
4622 |
the query
|
|
4623 |
||
4624 |
@note
|
|
4625 |
This is an extremely simplistic implementation that serves as a stub for a
|
|
4626 |
more advanced analysis of the join. Ideally the search depth should be
|
|
4627 |
determined by learning from previous query optimizations, because it will
|
|
4628 |
depend on the CPU power (and other factors).
|
|
4629 |
||
4630 |
@todo
|
|
4631 |
this value should be determined dynamically, based on statistics:
|
|
4632 |
uint32_t max_tables_for_exhaustive_opt= 7;
|
|
4633 |
||
4634 |
@todo
|
|
4635 |
this value could be determined by some mapping of the form:
|
|
4636 |
depth : table_count -> [max_tables_for_exhaustive_opt..MAX_EXHAUSTIVE]
|
|
4637 |
||
4638 |
@return
|
|
4639 |
A positive integer that specifies the search depth (and thus the
|
|
4640 |
exhaustiveness) of the depth-first search algorithm used by
|
|
4641 |
'greedy_search'.
|
|
4642 |
*/
|
|
4643 |
static uint32_t determine_search_depth(JOIN *join) |
|
4644 |
{
|
|
4645 |
uint32_t table_count= join->tables - join->const_tables; |
|
4646 |
uint32_t search_depth; |
|
4647 |
/* TODO: this value should be determined dynamically, based on statistics: */
|
|
4648 |
uint32_t max_tables_for_exhaustive_opt= 7; |
|
4649 |
||
4650 |
if (table_count <= max_tables_for_exhaustive_opt) |
|
4651 |
search_depth= table_count+1; // use exhaustive for small number of tables |
|
4652 |
else
|
|
4653 |
/*
|
|
4654 |
TODO: this value could be determined by some mapping of the form:
|
|
4655 |
depth : table_count -> [max_tables_for_exhaustive_opt..MAX_EXHAUSTIVE]
|
|
4656 |
*/
|
|
4657 |
search_depth= max_tables_for_exhaustive_opt; // use greedy search |
|
4658 |
||
4659 |
return search_depth; |
|
4660 |
}
|
|
4661 |
||
4662 |
static bool make_simple_join(JOIN *join,Table *tmp_table) |
|
4663 |
{
|
|
4664 |
Table **tableptr; |
|
4665 |
JOIN_TAB *join_tab; |
|
4666 |
||
4667 |
/*
|
|
4668 |
Reuse Table * and JOIN_TAB if already allocated by a previous call
|
|
4669 |
to this function through JOIN::exec (may happen for sub-queries).
|
|
4670 |
*/
|
|
4671 |
if (!join->table_reexec) |
|
4672 |
{
|
|
4673 |
if (!(join->table_reexec= (Table**) join->session->alloc(sizeof(Table*)))) |
|
4674 |
return(true); /* purecov: inspected */ |
|
4675 |
if (join->tmp_join) |
|
4676 |
join->tmp_join->table_reexec= join->table_reexec; |
|
4677 |
}
|
|
4678 |
if (!join->join_tab_reexec) |
|
4679 |
{
|
|
4680 |
if (!(join->join_tab_reexec= |
|
4681 |
(JOIN_TAB*) join->session->alloc(sizeof(JOIN_TAB)))) |
|
4682 |
return(true); /* purecov: inspected */ |
|
4683 |
if (join->tmp_join) |
|
4684 |
join->tmp_join->join_tab_reexec= join->join_tab_reexec; |
|
4685 |
}
|
|
4686 |
tableptr= join->table_reexec; |
|
4687 |
join_tab= join->join_tab_reexec; |
|
4688 |
||
4689 |
join->join_tab=join_tab; |
|
4690 |
join->table=tableptr; tableptr[0]=tmp_table; |
|
4691 |
join->tables=1; |
|
4692 |
join->const_tables=0; |
|
4693 |
join->const_table_map=0; |
|
4694 |
join->tmp_table_param.field_count= join->tmp_table_param.sum_func_count= |
|
4695 |
join->tmp_table_param.func_count=0; |
|
4696 |
join->tmp_table_param.copy_field=join->tmp_table_param.copy_field_end=0; |
|
4697 |
join->first_record=join->sort_and_group=0; |
|
4698 |
join->send_records=(ha_rows) 0; |
|
4699 |
join->group=0; |
|
4700 |
join->row_limit=join->unit->select_limit_cnt; |
|
4701 |
join->do_send_rows = (join->row_limit) ? 1 : 0; |
|
4702 |
||
4703 |
join_tab->cache.buff=0; /* No caching */ |
|
4704 |
join_tab->table=tmp_table; |
|
4705 |
join_tab->select=0; |
|
4706 |
join_tab->select_cond=0; |
|
4707 |
join_tab->quick=0; |
|
4708 |
join_tab->type= JT_ALL; /* Map through all records */ |
|
4709 |
join_tab->keys.set(); /* test everything in quick */ |
|
4710 |
join_tab->info=0; |
|
4711 |
join_tab->on_expr_ref=0; |
|
4712 |
join_tab->last_inner= 0; |
|
4713 |
join_tab->first_unmatched= 0; |
|
4714 |
join_tab->ref.key = -1; |
|
4715 |
join_tab->not_used_in_distinct=0; |
|
4716 |
join_tab->read_first_record= join_init_read_record; |
|
4717 |
join_tab->join=join; |
|
4718 |
join_tab->ref.key_parts= 0; |
|
4719 |
join_tab->flush_weedout_table= join_tab->check_weed_out_table= NULL; |
|
4720 |
join_tab->do_firstmatch= NULL; |
|
4721 |
memset(&join_tab->read_record, 0, sizeof(join_tab->read_record)); |
|
4722 |
tmp_table->status=0; |
|
4723 |
tmp_table->null_row=0; |
|
4724 |
return(false); |
|
4725 |
}
|
|
4726 |
||
4727 |
/**
|
|
4728 |
Fill in outer join related info for the execution plan structure.
|
|
4729 |
||
4730 |
For each outer join operation left after simplification of the
|
|
4731 |
original query the function set up the following pointers in the linear
|
|
4732 |
structure join->join_tab representing the selected execution plan.
|
|
4733 |
The first inner table t0 for the operation is set to refer to the last
|
|
4734 |
inner table tk through the field t0->last_inner.
|
|
4735 |
Any inner table ti for the operation are set to refer to the first
|
|
4736 |
inner table ti->first_inner.
|
|
4737 |
The first inner table t0 for the operation is set to refer to the
|
|
4738 |
first inner table of the embedding outer join operation, if there is any,
|
|
4739 |
through the field t0->first_upper.
|
|
4740 |
The on expression for the outer join operation is attached to the
|
|
4741 |
corresponding first inner table through the field t0->on_expr_ref.
|
|
4742 |
Here ti are structures of the JOIN_TAB type.
|
|
4743 |
||
4744 |
EXAMPLE. For the query:
|
|
4745 |
@code
|
|
4746 |
SELECT * FROM t1
|
|
4747 |
LEFT JOIN
|
|
4748 |
(t2, t3 LEFT JOIN t4 ON t3.a=t4.a)
|
|
4749 |
ON (t1.a=t2.a AND t1.b=t3.b)
|
|
4750 |
WHERE t1.c > 5,
|
|
4751 |
@endcode
|
|
4752 |
||
4753 |
given the execution plan with the table order t1,t2,t3,t4
|
|
4754 |
is selected, the following references will be set;
|
|
4755 |
t4->last_inner=[t4], t4->first_inner=[t4], t4->first_upper=[t2]
|
|
4756 |
t2->last_inner=[t4], t2->first_inner=t3->first_inner=[t2],
|
|
4757 |
on expression (t1.a=t2.a AND t1.b=t3.b) will be attached to
|
|
4758 |
*t2->on_expr_ref, while t3.a=t4.a will be attached to *t4->on_expr_ref.
|
|
4759 |
||
4760 |
@param join reference to the info fully describing the query
|
|
4761 |
||
4762 |
@note
|
|
4763 |
The function assumes that the simplification procedure has been
|
|
4764 |
already applied to the join query (see simplify_joins).
|
|
4765 |
This function can be called only after the execution plan
|
|
4766 |
has been chosen.
|
|
4767 |
*/
|
|
4768 |
static void make_outerjoin_info(JOIN *join) |
|
4769 |
{
|
|
4770 |
for (uint32_t i=join->const_tables ; i < join->tables ; i++) |
|
4771 |
{
|
|
4772 |
JOIN_TAB *tab=join->join_tab+i; |
|
4773 |
Table *table=tab->table; |
|
4774 |
TableList *tbl= table->pos_in_table_list; |
|
4775 |
TableList *embedding= tbl->embedding; |
|
4776 |
||
4777 |
if (tbl->outer_join) |
|
4778 |
{
|
|
4779 |
/*
|
|
4780 |
Table tab is the only one inner table for outer join.
|
|
4781 |
(Like table t4 for the table reference t3 LEFT JOIN t4 ON t3.a=t4.a
|
|
4782 |
is in the query above.)
|
|
4783 |
*/
|
|
4784 |
tab->last_inner= tab->first_inner= tab; |
|
4785 |
tab->on_expr_ref= &tbl->on_expr; |
|
4786 |
tab->cond_equal= tbl->cond_equal; |
|
4787 |
if (embedding) |
|
4788 |
tab->first_upper= embedding->nested_join->first_nested; |
|
4789 |
}
|
|
4790 |
for ( ; embedding ; embedding= embedding->embedding) |
|
4791 |
{
|
|
4792 |
/* Ignore sj-nests: */
|
|
4793 |
if (!embedding->on_expr) |
|
4794 |
continue; |
|
4795 |
nested_join_st *nested_join= embedding->nested_join; |
|
4796 |
if (!nested_join->counter_) |
|
4797 |
{
|
|
4798 |
/*
|
|
4799 |
Table tab is the first inner table for nested_join.
|
|
4800 |
Save reference to it in the nested join structure.
|
|
4801 |
*/
|
|
4802 |
nested_join->first_nested= tab; |
|
4803 |
tab->on_expr_ref= &embedding->on_expr; |
|
4804 |
tab->cond_equal= tbl->cond_equal; |
|
4805 |
if (embedding->embedding) |
|
4806 |
tab->first_upper= embedding->embedding->nested_join->first_nested; |
|
4807 |
}
|
|
4808 |
if (!tab->first_inner) |
|
4809 |
tab->first_inner= nested_join->first_nested; |
|
4810 |
if (++nested_join->counter_ < nested_join->join_list.elements) |
|
4811 |
break; |
|
4812 |
/* Table tab is the last inner table for nested join. */
|
|
4813 |
nested_join->first_nested->last_inner= tab; |
|
4814 |
}
|
|
4815 |
}
|
|
4816 |
return; |
|
4817 |
}
|
|
4818 |
||
4819 |
static bool make_join_select(JOIN *join,SQL_SELECT *select,COND *cond) |
|
4820 |
{
|
|
4821 |
Session *session= join->session; |
|
4822 |
if (select) |
|
4823 |
{
|
|
4824 |
add_not_null_conds(join); |
|
4825 |
table_map used_tables; |
|
4826 |
if (cond) /* Because of QUICK_GROUP_MIN_MAX_SELECT */ |
|
4827 |
{ /* there may be a select without a cond. */ |
|
4828 |
if (join->tables > 1) |
|
4829 |
cond->update_used_tables(); // Tablenr may have changed |
|
4830 |
if (join->const_tables == join->tables && |
|
4831 |
session->lex->current_select->master_unit() == |
|
4832 |
&session->lex->unit) // not upper level SELECT |
|
4833 |
join->const_table_map|=RAND_TABLE_BIT; |
|
4834 |
{ // Check const tables |
|
4835 |
COND *const_cond= |
|
4836 |
make_cond_for_table(cond, |
|
4837 |
join->const_table_map, |
|
4838 |
(table_map) 0, 1); |
|
4839 |
for (JOIN_TAB *tab= join->join_tab+join->const_tables; |
|
4840 |
tab < join->join_tab+join->tables ; tab++) |
|
4841 |
{
|
|
4842 |
if (*tab->on_expr_ref) |
|
4843 |
{
|
|
4844 |
JOIN_TAB *cond_tab= tab->first_inner; |
|
4845 |
COND *tmp= make_cond_for_table(*tab->on_expr_ref, |
|
4846 |
join->const_table_map, |
|
4847 |
( table_map) 0, 0); |
|
4848 |
if (!tmp) |
|
4849 |
continue; |
|
4850 |
tmp= new Item_func_trig_cond(tmp, &cond_tab->not_null_compl); |
|
4851 |
if (!tmp) |
|
4852 |
return(1); |
|
4853 |
tmp->quick_fix_field(); |
|
4854 |
cond_tab->select_cond= !cond_tab->select_cond ? tmp : |
|
4855 |
new Item_cond_and(cond_tab->select_cond, |
|
4856 |
tmp); |
|
4857 |
if (!cond_tab->select_cond) |
|
4858 |
return(1); |
|
4859 |
cond_tab->select_cond->quick_fix_field(); |
|
4860 |
}
|
|
4861 |
}
|
|
4862 |
if (const_cond && !const_cond->val_int()) |
|
4863 |
{
|
|
4864 |
return(1); // Impossible const condition |
|
4865 |
}
|
|
4866 |
}
|
|
4867 |
}
|
|
4868 |
used_tables=((select->const_tables=join->const_table_map) | |
|
4869 |
OUTER_REF_TABLE_BIT | RAND_TABLE_BIT); |
|
4870 |
for (uint32_t i=join->const_tables ; i < join->tables ; i++) |
|
4871 |
{
|
|
4872 |
JOIN_TAB *tab=join->join_tab+i; |
|
4873 |
/*
|
|
4874 |
first_inner is the X in queries like:
|
|
4875 |
SELECT * FROM t1 LEFT OUTER JOIN (t2 JOIN t3) ON X
|
|
4876 |
*/
|
|
4877 |
JOIN_TAB *first_inner_tab= tab->first_inner; |
|
4878 |
table_map current_map= tab->table->map; |
|
4879 |
bool use_quick_range=0; |
|
4880 |
COND *tmp; |
|
4881 |
||
4882 |
/*
|
|
4883 |
Following force including random expression in last table condition.
|
|
4884 |
It solve problem with select like SELECT * FROM t1 WHERE rand() > 0.5
|
|
4885 |
*/
|
|
4886 |
if (i == join->tables-1) |
|
4887 |
current_map|= OUTER_REF_TABLE_BIT | RAND_TABLE_BIT; |
|
4888 |
used_tables|=current_map; |
|
4889 |
||
4890 |
if (tab->type == JT_REF && tab->quick && |
|
4891 |
(uint32_t) tab->ref.key == tab->quick->index && |
|
4892 |
tab->ref.key_length < tab->quick->max_used_key_length) |
|
4893 |
{
|
|
4894 |
/* Range uses longer key; Use this instead of ref on key */
|
|
4895 |
tab->type=JT_ALL; |
|
4896 |
use_quick_range=1; |
|
4897 |
tab->use_quick=1; |
|
4898 |
tab->ref.key= -1; |
|
4899 |
tab->ref.key_parts=0; // Don't use ref key. |
|
4900 |
join->best_positions[i].records_read= rows2double(tab->quick->records); |
|
4901 |
/*
|
|
4902 |
We will use join cache here : prevent sorting of the first
|
|
4903 |
table only and sort at the end.
|
|
4904 |
*/
|
|
4905 |
if (i != join->const_tables && join->tables > join->const_tables + 1) |
|
4906 |
join->full_join= 1; |
|
4907 |
}
|
|
4908 |
||
4909 |
tmp= NULL; |
|
4910 |
if (cond) |
|
4911 |
tmp= make_cond_for_table(cond,used_tables,current_map, 0); |
|
4912 |
if (cond && !tmp && tab->quick) |
|
4913 |
{ // Outer join |
|
4914 |
if (tab->type != JT_ALL) |
|
4915 |
{
|
|
4916 |
/*
|
|
4917 |
Don't use the quick method
|
|
4918 |
We come here in the case where we have 'key=constant' and
|
|
4919 |
the test is removed by make_cond_for_table()
|
|
4920 |
*/
|
|
4921 |
delete tab->quick; |
|
4922 |
tab->quick= 0; |
|
4923 |
}
|
|
4924 |
else
|
|
4925 |
{
|
|
4926 |
/*
|
|
4927 |
Hack to handle the case where we only refer to a table
|
|
4928 |
in the ON part of an OUTER JOIN. In this case we want the code
|
|
4929 |
below to check if we should use 'quick' instead.
|
|
4930 |
*/
|
|
4931 |
tmp= new Item_int((int64_t) 1,1); // Always true |
|
4932 |
}
|
|
4933 |
||
4934 |
}
|
|
4935 |
if (tmp || !cond || tab->type == JT_REF || tab->type == JT_REF_OR_NULL || |
|
4936 |
tab->type == JT_EQ_REF) |
|
4937 |
{
|
|
4938 |
SQL_SELECT *sel= tab->select= ((SQL_SELECT*) |
|
4939 |
session->memdup((unsigned char*) select, |
|
4940 |
sizeof(*select))); |
|
4941 |
if (!sel) |
|
4942 |
return(1); // End of memory |
|
4943 |
/*
|
|
4944 |
If tab is an inner table of an outer join operation,
|
|
4945 |
add a match guard to the pushed down predicate.
|
|
4946 |
The guard will turn the predicate on only after
|
|
4947 |
the first match for outer tables is encountered.
|
|
4948 |
*/
|
|
4949 |
if (cond && tmp) |
|
4950 |
{
|
|
4951 |
/*
|
|
4952 |
Because of QUICK_GROUP_MIN_MAX_SELECT there may be a select without
|
|
4953 |
a cond, so neutralize the hack above.
|
|
4954 |
*/
|
|
4955 |
if (!(tmp= add_found_match_trig_cond(first_inner_tab, tmp, 0))) |
|
4956 |
return(1); |
|
4957 |
tab->select_cond=sel->cond=tmp; |
|
4958 |
/* Push condition to storage engine if this is enabled
|
|
4959 |
and the condition is not guarded */
|
|
4960 |
tab->table->file->pushed_cond= NULL; |
|
4961 |
if (session->variables.engine_condition_pushdown) |
|
4962 |
{
|
|
4963 |
COND *push_cond= |
|
4964 |
make_cond_for_table(tmp, current_map, current_map, 0); |
|
4965 |
if (push_cond) |
|
4966 |
{
|
|
4967 |
/* Push condition to handler */
|
|
4968 |
if (!tab->table->file->cond_push(push_cond)) |
|
4969 |
tab->table->file->pushed_cond= push_cond; |
|
4970 |
}
|
|
4971 |
}
|
|
4972 |
}
|
|
4973 |
else
|
|
4974 |
tab->select_cond= sel->cond= NULL; |
|
4975 |
||
4976 |
sel->head=tab->table; |
|
4977 |
if (tab->quick) |
|
4978 |
{
|
|
4979 |
/* Use quick key read if it's a constant and it's not used
|
|
4980 |
with key reading */
|
|
4981 |
if (tab->needed_reg.none() && tab->type != JT_EQ_REF |
|
4982 |
&& (tab->type != JT_REF || (uint32_t) tab->ref.key == tab->quick->index)) |
|
4983 |
{
|
|
4984 |
sel->quick=tab->quick; // Use value from get_quick_... |
|
4985 |
sel->quick_keys.reset(); |
|
4986 |
sel->needed_reg.reset(); |
|
4987 |
}
|
|
4988 |
else
|
|
4989 |
{
|
|
4990 |
delete tab->quick; |
|
4991 |
}
|
|
4992 |
tab->quick=0; |
|
4993 |
}
|
|
4994 |
uint32_t ref_key=(uint32_t) sel->head->reginfo.join_tab->ref.key+1; |
|
4995 |
if (i == join->const_tables && ref_key) |
|
4996 |
{
|
|
4997 |
if (tab->const_keys.any() && |
|
4998 |
tab->table->reginfo.impossible_range) |
|
4999 |
return(1); |
|
5000 |
}
|
|
5001 |
else if (tab->type == JT_ALL && ! use_quick_range) |
|
5002 |
{
|
|
5003 |
if (tab->const_keys.any() && |
|
5004 |
tab->table->reginfo.impossible_range) |
|
5005 |
return(1); // Impossible range |
|
5006 |
/*
|
|
5007 |
We plan to scan all rows.
|
|
5008 |
Check again if we should use an index.
|
|
5009 |
We could have used an column from a previous table in
|
|
5010 |
the index if we are using limit and this is the first table
|
|
5011 |
*/
|
|
5012 |
||
5013 |
if ((cond && (!((tab->keys & tab->const_keys) == tab->keys) && i > 0)) || |
|
5014 |
(!tab->const_keys.none() && (i == join->const_tables) && (join->unit->select_limit_cnt < join->best_positions[i].records_read) && ((join->select_options & OPTION_FOUND_ROWS) == false))) |
|
5015 |
{
|
|
5016 |
/* Join with outer join condition */
|
|
5017 |
COND *orig_cond=sel->cond; |
|
5018 |
sel->cond= and_conds(sel->cond, *tab->on_expr_ref); |
|
5019 |
||
5020 |
/*
|
|
5021 |
We can't call sel->cond->fix_fields,
|
|
5022 |
as it will break tab->on_expr if it's AND condition
|
|
5023 |
(fix_fields currently removes extra AND/OR levels).
|
|
5024 |
Yet attributes of the just built condition are not needed.
|
|
5025 |
Thus we call sel->cond->quick_fix_field for safety.
|
|
5026 |
*/
|
|
5027 |
if (sel->cond && !sel->cond->fixed) |
|
5028 |
sel->cond->quick_fix_field(); |
|
5029 |
||
5030 |
if (sel->test_quick_select(session, tab->keys, |
|
5031 |
used_tables & ~ current_map, |
|
5032 |
(join->select_options & |
|
5033 |
OPTION_FOUND_ROWS ? |
|
5034 |
HA_POS_ERROR : |
|
5035 |
join->unit->select_limit_cnt), 0, |
|
5036 |
false) < 0) |
|
5037 |
{
|
|
5038 |
/*
|
|
5039 |
Before reporting "Impossible WHERE" for the whole query
|
|
5040 |
we have to check isn't it only "impossible ON" instead
|
|
5041 |
*/
|
|
5042 |
sel->cond=orig_cond; |
|
5043 |
if (!*tab->on_expr_ref || |
|
5044 |
sel->test_quick_select(session, tab->keys, |
|
5045 |
used_tables & ~ current_map, |
|
5046 |
(join->select_options & |
|
5047 |
OPTION_FOUND_ROWS ? |
|
5048 |
HA_POS_ERROR : |
|
5049 |
join->unit->select_limit_cnt),0, |
|
5050 |
false) < 0) |
|
5051 |
return(1); // Impossible WHERE |
|
5052 |
}
|
|
5053 |
else
|
|
5054 |
sel->cond=orig_cond; |
|
5055 |
||
5056 |
/* Fix for EXPLAIN */
|
|
5057 |
if (sel->quick) |
|
5058 |
join->best_positions[i].records_read= (double)sel->quick->records; |
|
5059 |
}
|
|
5060 |
else
|
|
5061 |
{
|
|
5062 |
sel->needed_reg=tab->needed_reg; |
|
5063 |
sel->quick_keys.reset(); |
|
5064 |
}
|
|
5065 |
if (!((tab->checked_keys & sel->quick_keys) == sel->quick_keys) || |
|
5066 |
!((tab->checked_keys & sel->needed_reg) == sel->needed_reg)) |
|
5067 |
{
|
|
5068 |
tab->keys= sel->quick_keys; |
|
5069 |
tab->keys|= sel->needed_reg; |
|
5070 |
tab->use_quick= (!sel->needed_reg.none() && |
|
5071 |
(select->quick_keys.none() || |
|
5072 |
(select->quick && |
|
5073 |
(select->quick->records >= 100L)))) ? |
|
5074 |
2 : 1; |
|
5075 |
sel->read_tables= used_tables & ~current_map; |
|
5076 |
}
|
|
5077 |
if (i != join->const_tables && tab->use_quick != 2) |
|
5078 |
{ /* Read with cache */ |
|
5079 |
if (cond && |
|
5080 |
(tmp=make_cond_for_table(cond, |
|
5081 |
join->const_table_map | |
|
5082 |
current_map, |
|
5083 |
current_map, 0))) |
|
5084 |
{
|
|
5085 |
tab->cache.select=(SQL_SELECT*) |
|
5086 |
session->memdup((unsigned char*) sel, sizeof(SQL_SELECT)); |
|
5087 |
tab->cache.select->cond=tmp; |
|
5088 |
tab->cache.select->read_tables=join->const_table_map; |
|
5089 |
}
|
|
5090 |
}
|
|
5091 |
}
|
|
5092 |
}
|
|
5093 |
||
5094 |
/*
|
|
5095 |
Push down conditions from all on expressions.
|
|
5096 |
Each of these conditions are guarded by a variable
|
|
5097 |
that turns if off just before null complemented row for
|
|
5098 |
outer joins is formed. Thus, the condition from an
|
|
5099 |
'on expression' are guaranteed not to be checked for
|
|
5100 |
the null complemented row.
|
|
5101 |
*/
|
|
5102 |
||
5103 |
/* First push down constant conditions from on expressions */
|
|
5104 |
for (JOIN_TAB *join_tab= join->join_tab+join->const_tables; |
|
5105 |
join_tab < join->join_tab+join->tables ; join_tab++) |
|
5106 |
{
|
|
5107 |
if (*join_tab->on_expr_ref) |
|
5108 |
{
|
|
5109 |
JOIN_TAB *cond_tab= join_tab->first_inner; |
|
5110 |
tmp= make_cond_for_table(*join_tab->on_expr_ref, |
|
5111 |
join->const_table_map, |
|
5112 |
(table_map) 0, 0); |
|
5113 |
if (!tmp) |
|
5114 |
continue; |
|
5115 |
tmp= new Item_func_trig_cond(tmp, &cond_tab->not_null_compl); |
|
5116 |
if (!tmp) |
|
5117 |
return(1); |
|
5118 |
tmp->quick_fix_field(); |
|
5119 |
cond_tab->select_cond= !cond_tab->select_cond ? tmp : |
|
5120 |
new Item_cond_and(cond_tab->select_cond,tmp); |
|
5121 |
if (!cond_tab->select_cond) |
|
5122 |
return(1); |
|
5123 |
cond_tab->select_cond->quick_fix_field(); |
|
5124 |
}
|
|
5125 |
}
|
|
5126 |
||
5127 |
/* Push down non-constant conditions from on expressions */
|
|
5128 |
JOIN_TAB *last_tab= tab; |
|
5129 |
while (first_inner_tab && first_inner_tab->last_inner == last_tab) |
|
5130 |
{
|
|
5131 |
/*
|
|
5132 |
Table tab is the last inner table of an outer join.
|
|
5133 |
An on expression is always attached to it.
|
|
5134 |
*/
|
|
5135 |
COND *on_expr= *first_inner_tab->on_expr_ref; |
|
5136 |
||
5137 |
table_map used_tables2= (join->const_table_map | |
|
5138 |
OUTER_REF_TABLE_BIT | RAND_TABLE_BIT); |
|
5139 |
for (tab= join->join_tab+join->const_tables; tab <= last_tab ; tab++) |
|
5140 |
{
|
|
5141 |
current_map= tab->table->map; |
|
5142 |
used_tables2|= current_map; |
|
5143 |
COND *tmp_cond= make_cond_for_table(on_expr, used_tables2, |
|
5144 |
current_map, 0); |
|
5145 |
if (tmp_cond) |
|
5146 |
{
|
|
5147 |
JOIN_TAB *cond_tab= tab < first_inner_tab ? first_inner_tab : tab; |
|
5148 |
/*
|
|
5149 |
First add the guards for match variables of
|
|
5150 |
all embedding outer join operations.
|
|
5151 |
*/
|
|
5152 |
if (!(tmp_cond= add_found_match_trig_cond(cond_tab->first_inner, |
|
5153 |
tmp_cond, |
|
5154 |
first_inner_tab))) |
|
5155 |
return(1); |
|
5156 |
/*
|
|
5157 |
Now add the guard turning the predicate off for
|
|
5158 |
the null complemented row.
|
|
5159 |
*/
|
|
5160 |
tmp_cond= new Item_func_trig_cond(tmp_cond, |
|
5161 |
&first_inner_tab-> |
|
5162 |
not_null_compl); |
|
5163 |
if (tmp_cond) |
|
5164 |
tmp_cond->quick_fix_field(); |
|
5165 |
/* Add the predicate to other pushed down predicates */
|
|
5166 |
cond_tab->select_cond= !cond_tab->select_cond ? tmp_cond : |
|
5167 |
new Item_cond_and(cond_tab->select_cond, |
|
5168 |
tmp_cond); |
|
5169 |
if (!cond_tab->select_cond) |
|
5170 |
return(1); |
|
5171 |
cond_tab->select_cond->quick_fix_field(); |
|
5172 |
}
|
|
5173 |
}
|
|
5174 |
first_inner_tab= first_inner_tab->first_upper; |
|
5175 |
}
|
|
5176 |
}
|
|
5177 |
}
|
|
5178 |
return(0); |
|
5179 |
}
|
|
5180 |
||
5181 |
/*
|
|
5182 |
Plan refinement stage: do various set ups for the executioner
|
|
5183 |
||
5184 |
SYNOPSIS
|
|
5185 |
make_join_readinfo()
|
|
5186 |
join Join being processed
|
|
5187 |
options Join's options (checking for SELECT_DESCRIBE,
|
|
5188 |
SELECT_NO_JOIN_CACHE)
|
|
5189 |
no_jbuf_after Don't use join buffering after table with this number.
|
|
5190 |
||
5191 |
DESCRIPTION
|
|
5192 |
Plan refinement stage: do various set ups for the executioner
|
|
5193 |
- set up use of join buffering
|
|
5194 |
- push index conditions
|
|
5195 |
- increment counters
|
|
5196 |
- etc
|
|
5197 |
||
5198 |
RETURN
|
|
5199 |
false - OK
|
|
5200 |
true - Out of memory
|
|
5201 |
*/
|
|
5202 |
static bool make_join_readinfo(JOIN *join, uint64_t options, uint32_t no_jbuf_after) |
|
5203 |
{
|
|
5204 |
uint32_t i; |
|
5205 |
bool statistics= test(!(join->select_options & SELECT_DESCRIBE)); |
|
5206 |
bool sorted= 1; |
|
5207 |
||
5208 |
for (i=join->const_tables ; i < join->tables ; i++) |
|
5209 |
{
|
|
5210 |
JOIN_TAB *tab=join->join_tab+i; |
|
5211 |
Table *table=tab->table; |
|
5212 |
bool using_join_cache; |
|
5213 |
tab->read_record.table= table; |
|
5214 |
tab->read_record.file=table->file; |
|
5215 |
tab->next_select=sub_select; /* normal select */ |
|
5216 |
/*
|
|
5217 |
TODO: don't always instruct first table's ref/range access method to
|
|
5218 |
produce sorted output.
|
|
5219 |
*/
|
|
5220 |
tab->sorted= sorted; |
|
5221 |
sorted= 0; // only first must be sorted |
|
5222 |
if (tab->insideout_match_tab) |
|
5223 |
{
|
|
5224 |
if (!(tab->insideout_buf= (unsigned char*)join->session->alloc(tab->table->key_info |
|
5225 |
[tab->index]. |
|
5226 |
key_length))) |
|
5227 |
return true; |
|
5228 |
}
|
|
5229 |
switch (tab->type) { |
|
5230 |
case JT_SYSTEM: // Only happens with left join |
|
5231 |
table->status=STATUS_NO_RECORD; |
|
5232 |
tab->read_first_record= join_read_system; |
|
5233 |
tab->read_record.read_record= join_no_more_records; |
|
5234 |
break; |
|
5235 |
case JT_CONST: // Only happens with left join |
|
5236 |
table->status=STATUS_NO_RECORD; |
|
5237 |
tab->read_first_record= join_read_const; |
|
5238 |
tab->read_record.read_record= join_no_more_records; |
|
5239 |
if (table->covering_keys.test(tab->ref.key) && |
|
5240 |
!table->no_keyread) |
|
5241 |
{
|
|
5242 |
table->key_read=1; |
|
5243 |
table->file->extra(HA_EXTRA_KEYREAD); |
|
5244 |
}
|
|
5245 |
break; |
|
5246 |
case JT_EQ_REF: |
|
5247 |
table->status=STATUS_NO_RECORD; |
|
5248 |
if (tab->select) |
|
5249 |
{
|
|
5250 |
delete tab->select->quick; |
|
5251 |
tab->select->quick=0; |
|
5252 |
}
|
|
5253 |
delete tab->quick; |
|
5254 |
tab->quick=0; |
|
5255 |
tab->read_first_record= join_read_key; |
|
5256 |
tab->read_record.read_record= join_no_more_records; |
|
5257 |
if (table->covering_keys.test(tab->ref.key) && !table->no_keyread) |
|
5258 |
{
|
|
5259 |
table->key_read=1; |
|
5260 |
table->file->extra(HA_EXTRA_KEYREAD); |
|
5261 |
}
|
|
5262 |
else
|
|
5263 |
push_index_cond(tab, tab->ref.key, true); |
|
5264 |
break; |
|
5265 |
case JT_REF_OR_NULL: |
|
5266 |
case JT_REF: |
|
5267 |
table->status=STATUS_NO_RECORD; |
|
5268 |
if (tab->select) |
|
5269 |
{
|
|
5270 |
delete tab->select->quick; |
|
5271 |
tab->select->quick=0; |
|
5272 |
}
|
|
5273 |
delete tab->quick; |
|
5274 |
tab->quick=0; |
|
5275 |
if (table->covering_keys.test(tab->ref.key) && !table->no_keyread) |
|
5276 |
{
|
|
5277 |
table->key_read=1; |
|
5278 |
table->file->extra(HA_EXTRA_KEYREAD); |
|
5279 |
}
|
|
5280 |
else
|
|
5281 |
push_index_cond(tab, tab->ref.key, true); |
|
5282 |
if (tab->type == JT_REF) |
|
5283 |
{
|
|
5284 |
tab->read_first_record= join_read_always_key; |
|
5285 |
tab->read_record.read_record= tab->insideout_match_tab? |
|
5286 |
join_read_next_same_diff : join_read_next_same; |
|
5287 |
}
|
|
5288 |
else
|
|
5289 |
{
|
|
5290 |
tab->read_first_record= join_read_always_key_or_null; |
|
5291 |
tab->read_record.read_record= join_read_next_same_or_null; |
|
5292 |
}
|
|
5293 |
break; |
|
5294 |
case JT_ALL: |
|
5295 |
/*
|
|
5296 |
If previous table use cache
|
|
5297 |
If the incoming data set is already sorted don't use cache.
|
|
5298 |
*/
|
|
5299 |
table->status=STATUS_NO_RECORD; |
|
5300 |
using_join_cache= false; |
|
5301 |
if (i != join->const_tables && !(options & SELECT_NO_JOIN_CACHE) && |
|
5302 |
tab->use_quick != 2 && !tab->first_inner && i <= no_jbuf_after && |
|
5303 |
!tab->insideout_match_tab) |
|
5304 |
{
|
|
5305 |
if ((options & SELECT_DESCRIBE) || |
|
5306 |
!join_init_cache(join->session,join->join_tab+join->const_tables, |
|
5307 |
i-join->const_tables)) |
|
5308 |
{
|
|
5309 |
using_join_cache= true; |
|
5310 |
tab[-1].next_select=sub_select_cache; /* Patch previous */ |
|
5311 |
}
|
|
5312 |
}
|
|
5313 |
/* These init changes read_record */
|
|
5314 |
if (tab->use_quick == 2) |
|
5315 |
{
|
|
5316 |
join->session->server_status|=SERVER_QUERY_NO_GOOD_INDEX_USED; |
|
5317 |
tab->read_first_record= join_init_quick_read_record; |
|
5318 |
if (statistics) |
|
5319 |
status_var_increment(join->session->status_var.select_range_check_count); |
|
5320 |
}
|
|
5321 |
else
|
|
5322 |
{
|
|
5323 |
tab->read_first_record= join_init_read_record; |
|
5324 |
if (i == join->const_tables) |
|
5325 |
{
|
|
5326 |
if (tab->select && tab->select->quick) |
|
5327 |
{
|
|
5328 |
if (statistics) |
|
5329 |
status_var_increment(join->session->status_var.select_range_count); |
|
5330 |
}
|
|
5331 |
else
|
|
5332 |
{
|
|
5333 |
join->session->server_status|=SERVER_QUERY_NO_INDEX_USED; |
|
5334 |
if (statistics) |
|
5335 |
status_var_increment(join->session->status_var.select_scan_count); |
|
5336 |
}
|
|
5337 |
}
|
|
5338 |
else
|
|
5339 |
{
|
|
5340 |
if (tab->select && tab->select->quick) |
|
5341 |
{
|
|
5342 |
if (statistics) |
|
5343 |
status_var_increment(join->session->status_var.select_full_range_join_count); |
|
5344 |
}
|
|
5345 |
else
|
|
5346 |
{
|
|
5347 |
join->session->server_status|=SERVER_QUERY_NO_INDEX_USED; |
|
5348 |
if (statistics) |
|
5349 |
status_var_increment(join->session->status_var.select_full_join_count); |
|
5350 |
}
|
|
5351 |
}
|
|
5352 |
if (!table->no_keyread) |
|
5353 |
{
|
|
5354 |
if (tab->select && tab->select->quick && |
|
5355 |
tab->select->quick->index != MAX_KEY && //not index_merge |
|
5356 |
table->covering_keys.test(tab->select->quick->index)) |
|
5357 |
{
|
|
5358 |
table->key_read=1; |
|
5359 |
table->file->extra(HA_EXTRA_KEYREAD); |
|
5360 |
}
|
|
5361 |
else if (!table->covering_keys.none() && |
|
5362 |
!(tab->select && tab->select->quick)) |
|
5363 |
{ // Only read index tree |
|
5364 |
if (!tab->insideout_match_tab) |
|
5365 |
{
|
|
5366 |
/*
|
|
5367 |
See bug #26447: "Using the clustered index for a table scan
|
|
5368 |
is always faster than using a secondary index".
|
|
5369 |
*/
|
|
5370 |
if (table->s->primary_key != MAX_KEY && |
|
5371 |
table->file->primary_key_is_clustered()) |
|
5372 |
tab->index= table->s->primary_key; |
|
5373 |
else
|
|
5374 |
tab->index= table->find_shortest_key(&table->covering_keys); |
|
5375 |
}
|
|
5376 |
tab->read_first_record= join_read_first; |
|
5377 |
tab->type=JT_NEXT; // Read with index_first / index_next |
|
5378 |
}
|
|
5379 |
}
|
|
5380 |
if (tab->select && tab->select->quick && |
|
5381 |
tab->select->quick->index != MAX_KEY && ! tab->table->key_read) |
|
5382 |
push_index_cond(tab, tab->select->quick->index, !using_join_cache); |
|
5383 |
}
|
|
5384 |
break; |
|
5385 |
default: |
|
5386 |
break; /* purecov: deadcode */ |
|
5387 |
case JT_UNKNOWN: |
|
5388 |
case JT_MAYBE_REF: |
|
5389 |
abort(); /* purecov: deadcode */ |
|
5390 |
}
|
|
5391 |
}
|
|
5392 |
join->join_tab[join->tables-1].next_select=0; /* Set by do_select */ |
|
5393 |
return(false); |
|
5394 |
}
|
|
5395 |
||
5396 |
/** Update the dependency map for the tables. */
|
|
5397 |
static void update_depend_map(JOIN *join) |
|
5398 |
{
|
|
5399 |
JOIN_TAB *join_tab=join->join_tab, *end=join_tab+join->tables; |
|
5400 |
||
5401 |
for (; join_tab != end ; join_tab++) |
|
5402 |
{
|
|
5403 |
TABLE_REF *ref= &join_tab->ref; |
|
5404 |
table_map depend_map=0; |
|
5405 |
Item **item=ref->items; |
|
5406 |
uint32_t i; |
|
5407 |
for (i=0 ; i < ref->key_parts ; i++,item++) |
|
5408 |
depend_map|=(*item)->used_tables(); |
|
5409 |
ref->depend_map=depend_map & ~OUTER_REF_TABLE_BIT; |
|
5410 |
depend_map&= ~OUTER_REF_TABLE_BIT; |
|
5411 |
for (JOIN_TAB **tab=join->map2table; depend_map; tab++,depend_map>>=1 ) |
|
5412 |
{
|
|
5413 |
if (depend_map & 1) |
|
5414 |
ref->depend_map|=(*tab)->ref.depend_map; |
|
5415 |
}
|
|
5416 |
}
|
|
5417 |
}
|
|
5418 |
||
5419 |
/** Update the dependency map for the sort order. */
|
|
5420 |
static void update_depend_map(JOIN *join, order_st *order) |
|
5421 |
{
|
|
5422 |
for (; order ; order=order->next) |
|
5423 |
{
|
|
5424 |
table_map depend_map; |
|
5425 |
order->item[0]->update_used_tables(); |
|
5426 |
order->depend_map=depend_map=order->item[0]->used_tables(); |
|
5427 |
// Not item_sum(), RAND() and no reference to table outside of sub select
|
|
5428 |
if (!(order->depend_map & (OUTER_REF_TABLE_BIT | RAND_TABLE_BIT)) |
|
5429 |
&& !order->item[0]->with_sum_func) |
|
5430 |
{
|
|
5431 |
for (JOIN_TAB **tab=join->map2table; depend_map; tab++, depend_map>>=1) |
|
5432 |
{
|
|
5433 |
if (depend_map & 1) |
|
5434 |
order->depend_map|=(*tab)->ref.depend_map; |
|
5435 |
}
|
|
5436 |
}
|
|
5437 |
}
|
|
5438 |
}
|
|
5439 |
||
5440 |
/**
|
|
5441 |
Remove all constants and check if order_st only contains simple
|
|
5442 |
expressions.
|
|
5443 |
||
5444 |
simple_order is set to 1 if sort_order only uses fields from head table
|
|
5445 |
and the head table is not a LEFT JOIN table.
|
|
5446 |
||
5447 |
@param join Join handler
|
|
5448 |
@param first_order List of SORT or GROUP order
|
|
5449 |
@param cond WHERE statement
|
|
5450 |
@param change_list Set to 1 if we should remove things from list.
|
|
5451 |
If this is not set, then only simple_order is
|
|
5452 |
calculated.
|
|
5453 |
@param simple_order Set to 1 if we are only using simple expressions
|
|
5454 |
||
5455 |
@return
|
|
5456 |
Returns new sort order
|
|
5457 |
*/
|
|
5458 |
static order_st *remove_constants(JOIN *join,order_st *first_order, COND *cond, bool change_list, bool *simple_order) |
|
5459 |
{
|
|
5460 |
if (join->tables == join->const_tables) |
|
5461 |
return change_list ? 0 : first_order; // No need to sort |
|
5462 |
||
5463 |
order_st *order,**prev_ptr; |
|
5464 |
table_map first_table= join->join_tab[join->const_tables].table->map; |
|
5465 |
table_map not_const_tables= ~join->const_table_map; |
|
5466 |
table_map ref; |
|
5467 |
||
5468 |
prev_ptr= &first_order; |
|
5469 |
*simple_order= *join->join_tab[join->const_tables].on_expr_ref ? 0 : 1; |
|
5470 |
||
5471 |
/* NOTE: A variable of not_const_tables ^ first_table; breaks gcc 2.7 */
|
|
5472 |
||
5473 |
update_depend_map(join, first_order); |
|
5474 |
for (order=first_order; order ; order=order->next) |
|
5475 |
{
|
|
5476 |
table_map order_tables=order->item[0]->used_tables(); |
|
5477 |
if (order->item[0]->with_sum_func) |
|
5478 |
*simple_order=0; // Must do a temp table to sort |
|
5479 |
else if (!(order_tables & not_const_tables)) |
|
5480 |
{
|
|
5481 |
if (order->item[0]->with_subselect) |
|
5482 |
order->item[0]->val_str(&order->item[0]->str_value); |
|
5483 |
continue; // skip const item |
|
5484 |
}
|
|
5485 |
else
|
|
5486 |
{
|
|
5487 |
if (order_tables & (RAND_TABLE_BIT | OUTER_REF_TABLE_BIT)) |
|
5488 |
*simple_order=0; |
|
5489 |
else
|
|
5490 |
{
|
|
5491 |
Item *comp_item=0; |
|
5492 |
if (cond && const_expression_in_where(cond,order->item[0], &comp_item)) |
|
5493 |
{
|
|
5494 |
continue; |
|
5495 |
}
|
|
5496 |
if ((ref=order_tables & (not_const_tables ^ first_table))) |
|
5497 |
{
|
|
5498 |
if (!(order_tables & first_table) && |
|
5499 |
only_eq_ref_tables(join,first_order, ref)) |
|
5500 |
{
|
|
5501 |
continue; |
|
5502 |
}
|
|
5503 |
*simple_order=0; // Must do a temp table to sort |
|
5504 |
}
|
|
5505 |
}
|
|
5506 |
}
|
|
5507 |
if (change_list) |
|
5508 |
*prev_ptr= order; // use this entry |
|
5509 |
prev_ptr= &order->next; |
|
5510 |
}
|
|
5511 |
if (change_list) |
|
5512 |
*prev_ptr=0; |
|
5513 |
if (prev_ptr == &first_order) // Nothing to sort/group |
|
5514 |
*simple_order=1; |
|
5515 |
return(first_order); |
|
5516 |
}
|
|
5517 |
||
5518 |
static int return_zero_rows(JOIN *join, |
|
5519 |
select_result *result, |
|
5520 |
TableList *tables, |
|
5521 |
List<Item> &fields, |
|
5522 |
bool send_row, |
|
5523 |
uint64_t select_options, |
|
5524 |
const char *info, |
|
5525 |
Item *having) |
|
5526 |
{
|
|
5527 |
if (select_options & SELECT_DESCRIBE) |
|
5528 |
{
|
|
5529 |
select_describe(join, false, false, false, info); |
|
5530 |
return(0); |
|
5531 |
}
|
|
5532 |
||
5533 |
join->join_free(); |
|
5534 |
||
5535 |
if (send_row) |
|
5536 |
{
|
|
5537 |
for (TableList *table= tables; table; table= table->next_leaf) |
|
5538 |
table->table->mark_as_null_row(); // All fields are NULL |
|
5539 |
if (having && having->val_int() == 0) |
|
5540 |
send_row=0; |
|
5541 |
}
|
|
5542 |
if (!(result->send_fields(fields, Protocol::SEND_NUM_ROWS | Protocol::SEND_EOF))) |
|
5543 |
{
|
|
5544 |
if (send_row) |
|
5545 |
{
|
|
5546 |
List_iterator_fast<Item> it(fields); |
|
5547 |
Item *item; |
|
5548 |
while ((item= it++)) |
|
5549 |
item->no_rows_in_result(); |
|
5550 |
result->send_data(fields); |
|
5551 |
}
|
|
5552 |
result->send_eof(); // Should be safe |
|
5553 |
}
|
|
5554 |
/* Update results for FOUND_ROWS */
|
|
5555 |
join->session->limit_found_rows= join->session->examined_row_count= 0; |
|
5556 |
return(0); |
|
5557 |
}
|
|
5558 |
||
5559 |
/**
|
|
5560 |
Simplify joins replacing outer joins by inner joins whenever it's
|
|
5561 |
possible.
|
|
5562 |
||
5563 |
The function, during a retrieval of join_list, eliminates those
|
|
5564 |
outer joins that can be converted into inner join, possibly nested.
|
|
5565 |
It also moves the on expressions for the converted outer joins
|
|
5566 |
and from inner joins to conds.
|
|
5567 |
The function also calculates some attributes for nested joins:
|
|
5568 |
- used_tables
|
|
5569 |
- not_null_tables
|
|
5570 |
- dep_tables.
|
|
5571 |
- on_expr_dep_tables
|
|
5572 |
The first two attributes are used to test whether an outer join can
|
|
5573 |
be substituted for an inner join. The third attribute represents the
|
|
5574 |
relation 'to be dependent on' for tables. If table t2 is dependent
|
|
5575 |
on table t1, then in any evaluated execution plan table access to
|
|
5576 |
table t2 must precede access to table t2. This relation is used also
|
|
5577 |
to check whether the query contains invalid cross-references.
|
|
5578 |
The forth attribute is an auxiliary one and is used to calculate
|
|
5579 |
dep_tables.
|
|
5580 |
As the attribute dep_tables qualifies possibles orders of tables in the
|
|
5581 |
execution plan, the dependencies required by the straight join
|
|
5582 |
modifiers are reflected in this attribute as well.
|
|
5583 |
The function also removes all braces that can be removed from the join
|
|
5584 |
expression without changing its meaning.
|
|
5585 |
||
5586 |
@note
|
|
5587 |
An outer join can be replaced by an inner join if the where condition
|
|
5588 |
or the on expression for an embedding nested join contains a conjunctive
|
|
5589 |
predicate rejecting null values for some attribute of the inner tables.
|
|
5590 |
||
5591 |
E.g. in the query:
|
|
5592 |
@code
|
|
5593 |
SELECT * FROM t1 LEFT JOIN t2 ON t2.a=t1.a WHERE t2.b < 5
|
|
5594 |
@endcode
|
|
5595 |
the predicate t2.b < 5 rejects nulls.
|
|
5596 |
The query is converted first to:
|
|
5597 |
@code
|
|
5598 |
SELECT * FROM t1 INNER JOIN t2 ON t2.a=t1.a WHERE t2.b < 5
|
|
5599 |
@endcode
|
|
5600 |
then to the equivalent form:
|
|
5601 |
@code
|
|
5602 |
SELECT * FROM t1, t2 ON t2.a=t1.a WHERE t2.b < 5 AND t2.a=t1.a
|
|
5603 |
@endcode
|
|
5604 |
||
5605 |
||
5606 |
Similarly the following query:
|
|
5607 |
@code
|
|
5608 |
SELECT * from t1 LEFT JOIN (t2, t3) ON t2.a=t1.a t3.b=t1.b
|
|
5609 |
WHERE t2.c < 5
|
|
5610 |
@endcode
|
|
5611 |
is converted to:
|
|
5612 |
@code
|
|
5613 |
SELECT * FROM t1, (t2, t3) WHERE t2.c < 5 AND t2.a=t1.a t3.b=t1.b
|
|
5614 |
||
5615 |
@endcode
|
|
5616 |
||
5617 |
One conversion might trigger another:
|
|
5618 |
@code
|
|
5619 |
SELECT * FROM t1 LEFT JOIN t2 ON t2.a=t1.a
|
|
5620 |
LEFT JOIN t3 ON t3.b=t2.b
|
|
5621 |
WHERE t3 IS NOT NULL =>
|
|
5622 |
SELECT * FROM t1 LEFT JOIN t2 ON t2.a=t1.a, t3
|
|
5623 |
WHERE t3 IS NOT NULL AND t3.b=t2.b =>
|
|
5624 |
SELECT * FROM t1, t2, t3
|
|
5625 |
WHERE t3 IS NOT NULL AND t3.b=t2.b AND t2.a=t1.a
|
|
5626 |
@endcode
|
|
5627 |
||
5628 |
The function removes all unnecessary braces from the expression
|
|
5629 |
produced by the conversions.
|
|
5630 |
E.g.
|
|
5631 |
@code
|
|
5632 |
SELECT * FROM t1, (t2, t3) WHERE t2.c < 5 AND t2.a=t1.a AND t3.b=t1.b
|
|
5633 |
@endcode
|
|
5634 |
finally is converted to:
|
|
5635 |
@code
|
|
5636 |
SELECT * FROM t1, t2, t3 WHERE t2.c < 5 AND t2.a=t1.a AND t3.b=t1.b
|
|
5637 |
||
5638 |
@endcode
|
|
5639 |
||
5640 |
||
5641 |
It also will remove braces from the following queries:
|
|
5642 |
@code
|
|
5643 |
SELECT * from (t1 LEFT JOIN t2 ON t2.a=t1.a) LEFT JOIN t3 ON t3.b=t2.b
|
|
5644 |
SELECT * from (t1, (t2,t3)) WHERE t1.a=t2.a AND t2.b=t3.b.
|
|
5645 |
@endcode
|
|
5646 |
||
5647 |
The benefit of this simplification procedure is that it might return
|
|
5648 |
a query for which the optimizer can evaluate execution plan with more
|
|
5649 |
join orders. With a left join operation the optimizer does not
|
|
5650 |
consider any plan where one of the inner tables is before some of outer
|
|
5651 |
tables.
|
|
5652 |
||
5653 |
IMPLEMENTATION
|
|
5654 |
The function is implemented by a recursive procedure. On the recursive
|
|
5655 |
ascent all attributes are calculated, all outer joins that can be
|
|
5656 |
converted are replaced and then all unnecessary braces are removed.
|
|
5657 |
As join list contains join tables in the reverse order sequential
|
|
5658 |
elimination of outer joins does not require extra recursive calls.
|
|
5659 |
||
5660 |
SEMI-JOIN NOTES
|
|
5661 |
Remove all semi-joins that have are within another semi-join (i.e. have
|
|
5662 |
an "ancestor" semi-join nest)
|
|
5663 |
||
5664 |
EXAMPLES
|
|
5665 |
Here is an example of a join query with invalid cross references:
|
|
5666 |
@code
|
|
5667 |
SELECT * FROM t1 LEFT JOIN t2 ON t2.a=t3.a LEFT JOIN t3 ON t3.b=t1.b
|
|
5668 |
@endcode
|
|
5669 |
||
5670 |
@param join reference to the query info
|
|
5671 |
@param join_list list representation of the join to be converted
|
|
5672 |
@param conds conditions to add on expressions for converted joins
|
|
5673 |
@param top true <=> conds is the where condition
|
|
5674 |
||
5675 |
@return
|
|
5676 |
- The new condition, if success
|
|
5677 |
- 0, otherwise
|
|
5678 |
*/
|
|
5679 |
static COND *simplify_joins(JOIN *join, List<TableList> *join_list, COND *conds, bool top, bool in_sj) |
|
5680 |
{
|
|
5681 |
TableList *table; |
|
5682 |
nested_join_st *nested_join; |
|
5683 |
TableList *prev_table= 0; |
|
5684 |
List_iterator<TableList> li(*join_list); |
|
5685 |
||
5686 |
/*
|
|
5687 |
Try to simplify join operations from join_list.
|
|
5688 |
The most outer join operation is checked for conversion first.
|
|
5689 |
*/
|
|
5690 |
while ((table= li++)) |
|
5691 |
{
|
|
5692 |
table_map used_tables; |
|
5693 |
table_map not_null_tables= (table_map) 0; |
|
5694 |
||
5695 |
if ((nested_join= table->nested_join)) |
|
5696 |
{
|
|
5697 |
/*
|
|
5698 |
If the element of join_list is a nested join apply
|
|
5699 |
the procedure to its nested join list first.
|
|
5700 |
*/
|
|
5701 |
if (table->on_expr) |
|
5702 |
{
|
|
5703 |
Item *expr= table->on_expr; |
|
5704 |
/*
|
|
5705 |
If an on expression E is attached to the table,
|
|
5706 |
check all null rejected predicates in this expression.
|
|
5707 |
If such a predicate over an attribute belonging to
|
|
5708 |
an inner table of an embedded outer join is found,
|
|
5709 |
the outer join is converted to an inner join and
|
|
5710 |
the corresponding on expression is added to E.
|
|
5711 |
*/
|
|
5712 |
expr= simplify_joins(join, &nested_join->join_list, |
|
5713 |
expr, false, in_sj || table->sj_on_expr); |
|
5714 |
||
5715 |
if (!table->prep_on_expr || expr != table->on_expr) |
|
5716 |
{
|
|
5717 |
assert(expr); |
|
5718 |
||
5719 |
table->on_expr= expr; |
|
5720 |
table->prep_on_expr= expr->copy_andor_structure(join->session); |
|
5721 |
}
|
|
5722 |
}
|
|
5723 |
nested_join->used_tables= (table_map) 0; |
|
5724 |
nested_join->not_null_tables=(table_map) 0; |
|
5725 |
conds= simplify_joins(join, &nested_join->join_list, conds, top, in_sj || table->sj_on_expr); |
|
5726 |
used_tables= nested_join->used_tables; |
|
5727 |
not_null_tables= nested_join->not_null_tables; |
|
5728 |
}
|
|
5729 |
else
|
|
5730 |
{
|
|
5731 |
if (!table->prep_on_expr) |
|
5732 |
table->prep_on_expr= table->on_expr; |
|
5733 |
used_tables= table->table->map; |
|
5734 |
if (conds) |
|
5735 |
not_null_tables= conds->not_null_tables(); |
|
5736 |
}
|
|
5737 |
||
5738 |
if (table->embedding) |
|
5739 |
{
|
|
5740 |
table->embedding->nested_join->used_tables|= used_tables; |
|
5741 |
table->embedding->nested_join->not_null_tables|= not_null_tables; |
|
5742 |
}
|
|
5743 |
||
5744 |
if (!table->outer_join || (used_tables & not_null_tables)) |
|
5745 |
{
|
|
5746 |
/*
|
|
5747 |
For some of the inner tables there are conjunctive predicates
|
|
5748 |
that reject nulls => the outer join can be replaced by an inner join.
|
|
5749 |
*/
|
|
5750 |
table->outer_join= 0; |
|
5751 |
if (table->on_expr) |
|
5752 |
{
|
|
5753 |
/* Add ON expression to the WHERE or upper-level ON condition. */
|
|
5754 |
if (conds) |
|
5755 |
{
|
|
5756 |
conds= and_conds(conds, table->on_expr); |
|
5757 |
conds->top_level_item(); |
|
5758 |
/* conds is always a new item as both cond and on_expr existed */
|
|
5759 |
assert(!conds->fixed); |
|
5760 |
conds->fix_fields(join->session, &conds); |
|
5761 |
}
|
|
5762 |
else
|
|
5763 |
conds= table->on_expr; |
|
5764 |
table->prep_on_expr= table->on_expr= 0; |
|
5765 |
}
|
|
5766 |
}
|
|
5767 |
||
5768 |
if (!top) |
|
5769 |
continue; |
|
5770 |
||
5771 |
/*
|
|
5772 |
Only inner tables of non-convertible outer joins
|
|
5773 |
remain with on_expr.
|
|
5774 |
*/
|
|
5775 |
if (table->on_expr) |
|
5776 |
{
|
|
5777 |
table->dep_tables|= table->on_expr->used_tables(); |
|
5778 |
if (table->embedding) |
|
5779 |
{
|
|
5780 |
table->dep_tables&= ~table->embedding->nested_join->used_tables; |
|
5781 |
/*
|
|
5782 |
Embedding table depends on tables used
|
|
5783 |
in embedded on expressions.
|
|
5784 |
*/
|
|
5785 |
table->embedding->on_expr_dep_tables|= table->on_expr->used_tables(); |
|
5786 |
}
|
|
5787 |
else
|
|
5788 |
table->dep_tables&= ~table->table->map; |
|
5789 |
}
|
|
5790 |
||
5791 |
if (prev_table) |
|
5792 |
{
|
|
5793 |
/* The order of tables is reverse: prev_table follows table */
|
|
5794 |
if (prev_table->straight) |
|
5795 |
prev_table->dep_tables|= used_tables; |
|
5796 |
if (prev_table->on_expr) |
|
5797 |
{
|
|
5798 |
prev_table->dep_tables|= table->on_expr_dep_tables; |
|
5799 |
table_map prev_used_tables= prev_table->nested_join ? |
|
5800 |
prev_table->nested_join->used_tables : |
|
5801 |
prev_table->table->map; |
|
5802 |
/*
|
|
5803 |
If on expression contains only references to inner tables
|
|
5804 |
we still make the inner tables dependent on the outer tables.
|
|
5805 |
It would be enough to set dependency only on one outer table
|
|
5806 |
for them. Yet this is really a rare case.
|
|
5807 |
*/
|
|
5808 |
if (!(prev_table->on_expr->used_tables() & ~prev_used_tables)) |
|
5809 |
prev_table->dep_tables|= used_tables; |
|
5810 |
}
|
|
5811 |
}
|
|
5812 |
prev_table= table; |
|
5813 |
}
|
|
5814 |
||
5815 |
/*
|
|
5816 |
Flatten nested joins that can be flattened.
|
|
5817 |
no ON expression and not a semi-join => can be flattened.
|
|
5818 |
*/
|
|
5819 |
li.rewind(); |
|
5820 |
while ((table= li++)) |
|
5821 |
{
|
|
5822 |
nested_join= table->nested_join; |
|
5823 |
if (table->sj_on_expr && !in_sj) |
|
5824 |
{
|
|
5825 |
/*
|
|
5826 |
If this is a semi-join that is not contained within another semi-join,
|
|
5827 |
leave it intact (otherwise it is flattened)
|
|
5828 |
*/
|
|
5829 |
join->select_lex->sj_nests.push_back(table); |
|
5830 |
}
|
|
5831 |
else if (nested_join && !table->on_expr) |
|
5832 |
{
|
|
5833 |
TableList *tbl; |
|
5834 |
List_iterator<TableList> it(nested_join->join_list); |
|
5835 |
while ((tbl= it++)) |
|
5836 |
{
|
|
5837 |
tbl->embedding= table->embedding; |
|
5838 |
tbl->join_list= table->join_list; |
|
5839 |
}
|
|
5840 |
li.replace(nested_join->join_list); |
|
5841 |
}
|
|
5842 |
}
|
|
5843 |
return(conds); |
|
5844 |
}
|
|
5845 |
||
5846 |
static int remove_duplicates(JOIN *join, Table *entry,List<Item> &fields, Item *having) |
|
5847 |
{
|
|
5848 |
int error; |
|
5849 |
uint32_t reclength,offset; |
|
5850 |
uint32_t field_count; |
|
5851 |
Session *session= join->session; |
|
5852 |
||
5853 |
entry->reginfo.lock_type=TL_WRITE; |
|
5854 |
||
5855 |
/* Calculate how many saved fields there is in list */
|
|
5856 |
field_count=0; |
|
5857 |
List_iterator<Item> it(fields); |
|
5858 |
Item *item; |
|
5859 |
while ((item=it++)) |
|
5860 |
{
|
|
5861 |
if (item->get_tmp_table_field() && ! item->const_item()) |
|
5862 |
field_count++; |
|
5863 |
}
|
|
5864 |
||
5865 |
if (!field_count && !(join->select_options & OPTION_FOUND_ROWS) && !having) |
|
5866 |
{ // only const items with no OPTION_FOUND_ROWS |
|
5867 |
join->unit->select_limit_cnt= 1; // Only send first row |
|
5868 |
return(0); |
|
5869 |
}
|
|
5870 |
Field **first_field=entry->field+entry->s->fields - field_count; |
|
5871 |
offset= (field_count ? |
|
5872 |
entry->field[entry->s->fields - field_count]-> |
|
5873 |
offset(entry->record[0]) : 0); |
|
5874 |
reclength= entry->s->reclength-offset; |
|
5875 |
||
5876 |
free_io_cache(entry); // Safety |
|
5877 |
entry->file->info(HA_STATUS_VARIABLE); |
|
5878 |
if (entry->s->db_type() == heap_engine || |
|
5879 |
(!entry->s->blob_fields && |
|
5880 |
((ALIGN_SIZE(reclength) + HASH_OVERHEAD) * entry->file->stats.records < |
|
5881 |
session->variables.sortbuff_size))) |
|
5882 |
error= remove_dup_with_hash_index(join->session, entry, |
|
5883 |
field_count, first_field, |
|
5884 |
reclength, having); |
|
5885 |
else
|
|
5886 |
error= remove_dup_with_compare(join->session, entry, first_field, offset, |
|
5887 |
having); |
|
5888 |
||
5889 |
free_blobs(first_field); |
|
5890 |
return(error); |
|
5891 |
}
|
|
5892 |
||
5893 |
/**
|
|
5894 |
Function to setup clauses without sum functions.
|
|
5895 |
*/
|
|
5896 |
static int setup_without_group(Session *session, |
|
5897 |
Item **ref_pointer_array, |
|
5898 |
TableList *tables, |
|
5899 |
TableList *, |
|
5900 |
List<Item> &fields, |
|
5901 |
List<Item> &all_fields, |
|
5902 |
COND **conds, |
|
5903 |
order_st *order, |
|
5904 |
order_st *group, |
|
5905 |
bool *hidden_group_fields) |
|
5906 |
{
|
|
5907 |
int res; |
|
5908 |
nesting_map save_allow_sum_func=session->lex->allow_sum_func ; |
|
5909 |
||
5910 |
session->lex->allow_sum_func&= ~(1 << session->lex->current_select->nest_level); |
|
5911 |
res= setup_conds(session, tables, conds); |
|
5912 |
||
5913 |
session->lex->allow_sum_func|= 1 << session->lex->current_select->nest_level; |
|
5914 |
res= res || setup_order(session, ref_pointer_array, tables, fields, all_fields, |
|
5915 |
order); |
|
5916 |
session->lex->allow_sum_func&= ~(1 << session->lex->current_select->nest_level); |
|
5917 |
res= res || setup_group(session, ref_pointer_array, tables, fields, all_fields, |
|
5918 |
group, hidden_group_fields); |
|
5919 |
session->lex->allow_sum_func= save_allow_sum_func; |
|
5920 |
return(res); |
|
5921 |
}
|
|
5922 |
||
5923 |
/**
|
|
5924 |
Calculate the best possible join and initialize the join structure.
|
|
5925 |
||
5926 |
@retval
|
|
5927 |
0 ok
|
|
5928 |
@retval
|
|
5929 |
1 Fatal error
|
|
5930 |
*/
|
|
5931 |
static bool make_join_statistics(JOIN *join, TableList *tables, COND *conds, DYNAMIC_ARRAY *keyuse_array) |
|
5932 |
{
|
|
5933 |
int error; |
|
5934 |
Table *table; |
|
5935 |
uint32_t i,table_count,const_count,key; |
|
5936 |
table_map found_const_table_map, all_table_map, found_ref, refs; |
|
5937 |
key_map const_ref, eq_part; |
|
5938 |
Table **table_vector; |
|
5939 |
JOIN_TAB *stat,*stat_end,*s,**stat_ref; |
|
5940 |
KEYUSE *keyuse,*start_keyuse; |
|
5941 |
table_map outer_join=0; |
|
5942 |
SARGABLE_PARAM *sargables= 0; |
|
5943 |
JOIN_TAB *stat_vector[MAX_TABLES+1]; |
|
5944 |
||
5945 |
table_count=join->tables; |
|
5946 |
stat=(JOIN_TAB*) join->session->calloc(sizeof(JOIN_TAB)*table_count); |
|
5947 |
stat_ref=(JOIN_TAB**) join->session->alloc(sizeof(JOIN_TAB*)*MAX_TABLES); |
|
5948 |
table_vector=(Table**) join->session->alloc(sizeof(Table*)*(table_count*2)); |
|
5949 |
if (!stat || !stat_ref || !table_vector) |
|
5950 |
return(1); // Eom /* purecov: inspected */ |
|
5951 |
||
5952 |
join->best_ref=stat_vector; |
|
5953 |
||
5954 |
stat_end=stat+table_count; |
|
5955 |
found_const_table_map= all_table_map=0; |
|
5956 |
const_count=0; |
|
5957 |
||
5958 |
for (s= stat, i= 0; |
|
5959 |
tables; |
|
5960 |
s++, tables= tables->next_leaf, i++) |
|
5961 |
{
|
|
5962 |
TableList *embedding= tables->embedding; |
|
5963 |
stat_vector[i]=s; |
|
5964 |
s->keys.reset(); |
|
5965 |
s->const_keys.reset(); |
|
5966 |
s->checked_keys.reset(); |
|
5967 |
s->needed_reg.reset(); |
|
5968 |
table_vector[i]=s->table=table=tables->table; |
|
5969 |
table->pos_in_table_list= tables; |
|
5970 |
error= table->file->info(HA_STATUS_VARIABLE | HA_STATUS_NO_LOCK); |
|
5971 |
if(error) |
|
5972 |
{
|
|
5973 |
table->file->print_error(error, MYF(0)); |
|
5974 |
return(1); |
|
5975 |
}
|
|
5976 |
table->quick_keys.reset(); |
|
5977 |
table->reginfo.join_tab=s; |
|
5978 |
table->reginfo.not_exists_optimize=0; |
|
5979 |
memset(table->const_key_parts, 0, |
|
5980 |
sizeof(key_part_map)*table->s->keys); |
|
5981 |
all_table_map|= table->map; |
|
5982 |
s->join=join; |
|
5983 |
s->info=0; // For describe |
|
5984 |
||
5985 |
s->dependent= tables->dep_tables; |
|
5986 |
s->key_dependent= 0; |
|
5987 |
if (tables->schema_table) |
|
5988 |
table->file->stats.records= 2; |
|
5989 |
table->quick_condition_rows= table->file->stats.records; |
|
5990 |
||
5991 |
s->on_expr_ref= &tables->on_expr; |
|
5992 |
if (*s->on_expr_ref) |
|
5993 |
{
|
|
5994 |
/* s is the only inner table of an outer join */
|
|
5995 |
if (!table->file->stats.records && !embedding) |
|
5996 |
{ // Empty table |
|
5997 |
s->dependent= 0; // Ignore LEFT JOIN depend. |
|
5998 |
set_position(join,const_count++,s,(KEYUSE*) 0); |
|
5999 |
continue; |
|
6000 |
}
|
|
6001 |
outer_join|= table->map; |
|
6002 |
s->embedding_map= 0; |
|
6003 |
for (;embedding; embedding= embedding->embedding) |
|
6004 |
s->embedding_map|= embedding->nested_join->nj_map; |
|
6005 |
continue; |
|
6006 |
}
|
|
6007 |
if (embedding && !(embedding->sj_on_expr && ! embedding->embedding)) |
|
6008 |
{
|
|
6009 |
/* s belongs to a nested join, maybe to several embedded joins */
|
|
6010 |
s->embedding_map= 0; |
|
6011 |
do
|
|
6012 |
{
|
|
6013 |
nested_join_st *nested_join= embedding->nested_join; |
|
6014 |
s->embedding_map|=nested_join->nj_map; |
|
6015 |
s->dependent|= embedding->dep_tables; |
|
6016 |
embedding= embedding->embedding; |
|
6017 |
outer_join|= nested_join->used_tables; |
|
6018 |
}
|
|
6019 |
while (embedding); |
|
6020 |
continue; |
|
6021 |
}
|
|
6022 |
if ((table->file->stats.records <= 1) && !s->dependent && |
|
6023 |
(table->file->ha_table_flags() & HA_STATS_RECORDS_IS_EXACT) && |
|
6024 |
!join->no_const_tables) |
|
6025 |
{
|
|
6026 |
set_position(join,const_count++,s,(KEYUSE*) 0); |
|
6027 |
}
|
|
6028 |
}
|
|
6029 |
stat_vector[i]=0; |
|
6030 |
join->outer_join=outer_join; |
|
6031 |
||
6032 |
if (join->outer_join) |
|
6033 |
{
|
|
6034 |
/*
|
|
6035 |
Build transitive closure for relation 'to be dependent on'.
|
|
6036 |
This will speed up the plan search for many cases with outer joins,
|
|
6037 |
as well as allow us to catch illegal cross references/
|
|
6038 |
Warshall's algorithm is used to build the transitive closure.
|
|
6039 |
As we use bitmaps to represent the relation the complexity
|
|
6040 |
of the algorithm is O((number of tables)^2).
|
|
6041 |
*/
|
|
6042 |
for (i= 0, s= stat ; i < table_count ; i++, s++) |
|
6043 |
{
|
|
6044 |
for (uint32_t j= 0 ; j < table_count ; j++) |
|
6045 |
{
|
|
6046 |
table= stat[j].table; |
|
6047 |
if (s->dependent & table->map) |
|
6048 |
s->dependent |= table->reginfo.join_tab->dependent; |
|
6049 |
}
|
|
6050 |
if (s->dependent) |
|
6051 |
s->table->maybe_null= 1; |
|
6052 |
}
|
|
6053 |
/* Catch illegal cross references for outer joins */
|
|
6054 |
for (i= 0, s= stat ; i < table_count ; i++, s++) |
|
6055 |
{
|
|
6056 |
if (s->dependent & s->table->map) |
|
6057 |
{
|
|
6058 |
join->tables=0; // Don't use join->table |
|
6059 |
my_message(ER_WRONG_OUTER_JOIN, ER(ER_WRONG_OUTER_JOIN), MYF(0)); |
|
6060 |
return(1); |
|
6061 |
}
|
|
6062 |
s->key_dependent= s->dependent; |
|
6063 |
}
|
|
6064 |
}
|
|
6065 |
||
6066 |
if (conds || outer_join) |
|
6067 |
if (update_ref_and_keys(join->session, keyuse_array, stat, join->tables, |
|
6068 |
conds, join->cond_equal, |
|
6069 |
~outer_join, join->select_lex, &sargables)) |
|
6070 |
return(1); |
|
6071 |
||
6072 |
/* Read tables with 0 or 1 rows (system tables) */
|
|
6073 |
join->const_table_map= 0; |
|
6074 |
||
6075 |
for (POSITION *p_pos=join->positions, *p_end=p_pos+const_count; |
|
6076 |
p_pos < p_end ; |
|
6077 |
p_pos++) |
|
6078 |
{
|
|
6079 |
int tmp; |
|
6080 |
s= p_pos->table; |
|
6081 |
s->type=JT_SYSTEM; |
|
6082 |
join->const_table_map|=s->table->map; |
|
6083 |
if ((tmp=join_read_const_table(s, p_pos))) |
|
6084 |
{
|
|
6085 |
if (tmp > 0) |
|
6086 |
return(1); // Fatal error |
|
6087 |
}
|
|
6088 |
else
|
|
6089 |
found_const_table_map|= s->table->map; |
|
6090 |
}
|
|
6091 |
||
6092 |
/* loop until no more const tables are found */
|
|
6093 |
int ref_changed; |
|
6094 |
do
|
|
6095 |
{
|
|
6096 |
more_const_tables_found: |
|
6097 |
ref_changed = 0; |
|
6098 |
found_ref=0; |
|
6099 |
||
6100 |
/*
|
|
6101 |
We only have to loop from stat_vector + const_count as
|
|
6102 |
set_position() will move all const_tables first in stat_vector
|
|
6103 |
*/
|
|
6104 |
||
6105 |
for (JOIN_TAB **pos=stat_vector+const_count ; (s= *pos) ; pos++) |
|
6106 |
{
|
|
6107 |
table=s->table; |
|
6108 |
||
6109 |
/*
|
|
6110 |
If equi-join condition by a key is null rejecting and after a
|
|
6111 |
substitution of a const table the key value happens to be null
|
|
6112 |
then we can state that there are no matches for this equi-join.
|
|
6113 |
*/
|
|
6114 |
if ((keyuse= s->keyuse) && *s->on_expr_ref && !s->embedding_map) |
|
6115 |
{
|
|
6116 |
/*
|
|
6117 |
When performing an outer join operation if there are no matching rows
|
|
6118 |
for the single row of the outer table all the inner tables are to be
|
|
6119 |
null complemented and thus considered as constant tables.
|
|
6120 |
Here we apply this consideration to the case of outer join operations
|
|
6121 |
with a single inner table only because the case with nested tables
|
|
6122 |
would require a more thorough analysis.
|
|
6123 |
TODO. Apply single row substitution to null complemented inner tables
|
|
6124 |
for nested outer join operations.
|
|
6125 |
*/
|
|
6126 |
while (keyuse->table == table) |
|
6127 |
{
|
|
6128 |
if (!(keyuse->val->used_tables() & ~join->const_table_map) && |
|
6129 |
keyuse->val->is_null() && keyuse->null_rejecting) |
|
6130 |
{
|
|
6131 |
s->type= JT_CONST; |
|
6132 |
table->mark_as_null_row(); |
|
6133 |
found_const_table_map|= table->map; |
|
6134 |
join->const_table_map|= table->map; |
|
6135 |
set_position(join,const_count++,s,(KEYUSE*) 0); |
|
6136 |
goto more_const_tables_found; |
|
6137 |
}
|
|
6138 |
keyuse++; |
|
6139 |
}
|
|
6140 |
}
|
|
6141 |
||
6142 |
if (s->dependent) // If dependent on some table |
|
6143 |
{
|
|
6144 |
// All dep. must be constants
|
|
6145 |
if (s->dependent & ~(found_const_table_map)) |
|
6146 |
continue; |
|
6147 |
if (table->file->stats.records <= 1L && |
|
6148 |
(table->file->ha_table_flags() & HA_STATS_RECORDS_IS_EXACT) && |
|
6149 |
!table->pos_in_table_list->embedding) |
|
6150 |
{ // system table |
|
6151 |
int tmp= 0; |
|
6152 |
s->type=JT_SYSTEM; |
|
6153 |
join->const_table_map|=table->map; |
|
6154 |
set_position(join,const_count++,s,(KEYUSE*) 0); |
|
6155 |
if ((tmp= join_read_const_table(s, join->positions+const_count-1))) |
|
6156 |
{
|
|
6157 |
if (tmp > 0) |
|
6158 |
return(1); // Fatal error |
|
6159 |
}
|
|
6160 |
else
|
|
6161 |
found_const_table_map|= table->map; |
|
6162 |
continue; |
|
6163 |
}
|
|
6164 |
}
|
|
6165 |
/* check if table can be read by key or table only uses const refs */
|
|
6166 |
if ((keyuse=s->keyuse)) |
|
6167 |
{
|
|
6168 |
s->type= JT_REF; |
|
6169 |
while (keyuse->table == table) |
|
6170 |
{
|
|
6171 |
start_keyuse=keyuse; |
|
6172 |
key=keyuse->key; |
|
6173 |
s->keys.set(key); // QQ: remove this ? |
|
6174 |
||
6175 |
refs=0; |
|
6176 |
const_ref.reset(); |
|
6177 |
eq_part.reset(); |
|
6178 |
do
|
|
6179 |
{
|
|
6180 |
if (keyuse->val->type() != Item::NULL_ITEM && !keyuse->optimize) |
|
6181 |
{
|
|
6182 |
if (!((~found_const_table_map) & keyuse->used_tables)) |
|
6183 |
const_ref.set(keyuse->keypart); |
|
6184 |
else
|
|
6185 |
refs|=keyuse->used_tables; |
|
6186 |
eq_part.set(keyuse->keypart); |
|
6187 |
}
|
|
6188 |
keyuse++; |
|
6189 |
} while (keyuse->table == table && keyuse->key == key); |
|
6190 |
||
6191 |
if (is_keymap_prefix(eq_part, table->key_info[key].key_parts) && |
|
6192 |
!table->pos_in_table_list->embedding) |
|
6193 |
{
|
|
6194 |
if ((table->key_info[key].flags & (HA_NOSAME)) == HA_NOSAME) |
|
6195 |
{
|
|
6196 |
if (const_ref == eq_part) |
|
6197 |
{ // Found everything for ref. |
|
6198 |
int tmp; |
|
6199 |
ref_changed = 1; |
|
6200 |
s->type= JT_CONST; |
|
6201 |
join->const_table_map|= table->map; |
|
6202 |
set_position(join,const_count++,s,start_keyuse); |
|
6203 |
if (create_ref_for_key(join, s, start_keyuse, found_const_table_map)) |
|
6204 |
return(1); |
|
6205 |
if ((tmp=join_read_const_table(s, join->positions+const_count-1))) |
|
6206 |
{
|
|
6207 |
if (tmp > 0) |
|
6208 |
return(1); // Fatal error |
|
6209 |
}
|
|
6210 |
else
|
|
6211 |
found_const_table_map|= table->map; |
|
6212 |
break; |
|
6213 |
}
|
|
6214 |
else
|
|
6215 |
found_ref|= refs; // Table is const if all refs are const |
|
6216 |
}
|
|
6217 |
else if (const_ref == eq_part) |
|
6218 |
s->const_keys.set(key); |
|
6219 |
}
|
|
6220 |
}
|
|
6221 |
}
|
|
6222 |
}
|
|
6223 |
} while (join->const_table_map & found_ref && ref_changed); |
|
6224 |
||
6225 |
/*
|
|
6226 |
Update info on indexes that can be used for search lookups as
|
|
6227 |
reading const tables may has added new sargable predicates.
|
|
6228 |
*/
|
|
6229 |
if (const_count && sargables) |
|
6230 |
{
|
|
6231 |
for( ; sargables->field ; sargables++) |
|
6232 |
{
|
|
6233 |
Field *field= sargables->field; |
|
6234 |
JOIN_TAB *join_tab= field->table->reginfo.join_tab; |
|
6235 |
key_map possible_keys= field->key_start; |
|
6236 |
possible_keys&= field->table->keys_in_use_for_query; |
|
6237 |
bool is_const= 1; |
|
6238 |
for (uint32_t j=0; j < sargables->num_values; j++) |
|
6239 |
is_const&= sargables->arg_value[j]->const_item(); |
|
6240 |
if (is_const) |
|
6241 |
join_tab[0].const_keys|= possible_keys; |
|
6242 |
}
|
|
6243 |
}
|
|
6244 |
||
6245 |
if (pull_out_semijoin_tables(join)) |
|
6246 |
return(true); |
|
6247 |
||
6248 |
/* Calc how many (possible) matched records in each table */
|
|
6249 |
||
6250 |
for (s=stat ; s < stat_end ; s++) |
|
6251 |
{
|
|
6252 |
if (s->type == JT_SYSTEM || s->type == JT_CONST) |
|
6253 |
{
|
|
6254 |
/* Only one matching row */
|
|
6255 |
s->found_records=s->records=s->read_time=1; s->worst_seeks=1.0; |
|
6256 |
continue; |
|
6257 |
}
|
|
6258 |
/* Approximate found rows and time to read them */
|
|
6259 |
s->found_records=s->records=s->table->file->stats.records; |
|
6260 |
s->read_time=(ha_rows) s->table->file->scan_time(); |
|
6261 |
||
6262 |
/*
|
|
6263 |
Set a max range of how many seeks we can expect when using keys
|
|
6264 |
This is can't be to high as otherwise we are likely to use
|
|
6265 |
table scan.
|
|
6266 |
*/
|
|
1067.4.4
by Nathan Williams
The rest of the files in the drizzled directory were purged of the cmin macro and replace with std::min (except for the definition in globals.h and 1 usage in stacktrace.cc). |
6267 |
s->worst_seeks= min((double) s->found_records / 10, |
6268 |
(double) s->read_time*3); |
|
1039.2.2
by Jay Pipes
Phase 2 of JOIN refactoring. |
6269 |
if (s->worst_seeks < 2.0) // Fix for small tables |
6270 |
s->worst_seeks=2.0; |
|
6271 |
||
6272 |
/*
|
|
6273 |
Add to stat->const_keys those indexes for which all group fields or
|
|
6274 |
all select distinct fields participate in one index.
|
|
6275 |
*/
|
|
6276 |
add_group_and_distinct_keys(join, s); |
|
6277 |
||
6278 |
if (s->const_keys.any() && |
|
6279 |
!s->table->pos_in_table_list->embedding) |
|
6280 |
{
|
|
6281 |
ha_rows records; |
|
6282 |
SQL_SELECT *select; |
|
6283 |
select= make_select(s->table, found_const_table_map, found_const_table_map, *s->on_expr_ref ? *s->on_expr_ref : conds, 1, &error); |
|
6284 |
if (! select) |
|
6285 |
return(1); |
|
6286 |
records= get_quick_record_count(join->session, select, s->table, &s->const_keys, join->row_limit); |
|
6287 |
s->quick=select->quick; |
|
6288 |
s->needed_reg=select->needed_reg; |
|
6289 |
select->quick=0; |
|
6290 |
if (records == 0 && s->table->reginfo.impossible_range) |
|
6291 |
{
|
|
6292 |
/*
|
|
6293 |
Impossible WHERE or ON expression
|
|
6294 |
In case of ON, we mark that the we match one empty NULL row.
|
|
6295 |
In case of WHERE, don't set found_const_table_map to get the
|
|
6296 |
caller to abort with a zero row result.
|
|
6297 |
*/
|
|
6298 |
join->const_table_map|= s->table->map; |
|
6299 |
set_position(join,const_count++,s,(KEYUSE*) 0); |
|
6300 |
s->type= JT_CONST; |
|
6301 |
if (*s->on_expr_ref) |
|
6302 |
{
|
|
6303 |
/* Generate empty row */
|
|
6304 |
s->info= "Impossible ON condition"; |
|
6305 |
found_const_table_map|= s->table->map; |
|
6306 |
s->type= JT_CONST; |
|
6307 |
s->table->mark_as_null_row(); // All fields are NULL |
|
6308 |
}
|
|
6309 |
}
|
|
6310 |
if (records != HA_POS_ERROR) |
|
6311 |
{
|
|
6312 |
s->found_records=records; |
|
6313 |
s->read_time= (ha_rows) (s->quick ? s->quick->read_time : 0.0); |
|
6314 |
}
|
|
6315 |
delete select; |
|
6316 |
}
|
|
6317 |
}
|
|
6318 |
||
6319 |
join->join_tab=stat; |
|
6320 |
join->map2table=stat_ref; |
|
6321 |
join->table= join->all_tables=table_vector; |
|
6322 |
join->const_tables=const_count; |
|
6323 |
join->found_const_table_map=found_const_table_map; |
|
6324 |
||
6325 |
/* Find an optimal join order of the non-constant tables. */
|
|
6326 |
if (join->const_tables != join->tables) |
|
6327 |
{
|
|
6328 |
optimize_keyuse(join, keyuse_array); |
|
6329 |
if (choose_plan(join, all_table_map & ~join->const_table_map)) |
|
6330 |
return(true); |
|
6331 |
}
|
|
6332 |
else
|
|
6333 |
{
|
|
6334 |
memcpy(join->best_positions, join->positions, sizeof(POSITION)*join->const_tables); |
|
6335 |
join->best_read= 1.0; |
|
6336 |
}
|
|
6337 |
/* Generate an execution plan from the found optimal join order. */
|
|
6338 |
return (join->session->killed || get_best_combination(join)); |
|
6339 |
}
|
|
6340 |
||
6341 |
/**
|
|
6342 |
Assign each nested join structure a bit in nested_join_map.
|
|
6343 |
||
6344 |
Assign each nested join structure (except "confluent" ones - those that
|
|
6345 |
embed only one element) a bit in nested_join_map.
|
|
6346 |
||
6347 |
@param join Join being processed
|
|
6348 |
@param join_list List of tables
|
|
6349 |
@param first_unused Number of first unused bit in nested_join_map before the
|
|
6350 |
call
|
|
6351 |
||
6352 |
@note
|
|
6353 |
This function is called after simplify_joins(), when there are no
|
|
6354 |
redundant nested joins, #non_confluent_nested_joins <= #tables_in_join so
|
|
6355 |
we will not run out of bits in nested_join_map.
|
|
6356 |
||
6357 |
@return
|
|
6358 |
First unused bit in nested_join_map after the call.
|
|
6359 |
*/
|
|
6360 |
static uint32_t build_bitmap_for_nested_joins(List<TableList> *join_list, uint32_t first_unused) |
|
6361 |
{
|
|
6362 |
List_iterator<TableList> li(*join_list); |
|
6363 |
TableList *table; |
|
6364 |
while ((table= li++)) |
|
6365 |
{
|
|
6366 |
nested_join_st *nested_join; |
|
6367 |
if ((nested_join= table->nested_join)) |
|
6368 |
{
|
|
6369 |
/*
|
|
6370 |
It is guaranteed by simplify_joins() function that a nested join
|
|
6371 |
that has only one child is either
|
|
6372 |
- a single-table view (the child is the underlying table), or
|
|
6373 |
- a single-table semi-join nest
|
|
6374 |
||
6375 |
We don't assign bits to such sj-nests because
|
|
6376 |
1. it is redundant (a "sequence" of one table cannot be interleaved
|
|
6377 |
with anything)
|
|
6378 |
2. we could run out bits in nested_join_map otherwise.
|
|
6379 |
*/
|
|
6380 |
if (nested_join->join_list.elements != 1) |
|
6381 |
{
|
|
6382 |
/* Don't assign bits to sj-nests */
|
|
6383 |
if (table->on_expr) |
|
6384 |
nested_join->nj_map= (nested_join_map) 1 << first_unused++; |
|
6385 |
first_unused= build_bitmap_for_nested_joins(&nested_join->join_list, |
|
6386 |
first_unused); |
|
6387 |
}
|
|
6388 |
}
|
|
6389 |
}
|
|
6390 |
return(first_unused); |
|
6391 |
}
|
|
6392 |
||
6393 |
||
6394 |
/**
|
|
6395 |
Return table number if there is only one table in sort order
|
|
6396 |
and group and order is compatible, else return 0.
|
|
6397 |
*/
|
|
6398 |
static Table *get_sort_by_table(order_st *a,order_st *b,TableList *tables) |
|
6399 |
{
|
|
6400 |
table_map map= (table_map) 0; |
|
6401 |
||
6402 |
if (!a) |
|
6403 |
a= b; // Only one need to be given |
|
6404 |
else if (!b) |
|
6405 |
b= a; |
|
6406 |
||
6407 |
for (; a && b; a=a->next,b=b->next) |
|
6408 |
{
|
|
6409 |
if (!(*a->item)->eq(*b->item,1)) |
|
6410 |
return (Table *) NULL; |
|
6411 |
map|= a->item[0]->used_tables(); |
|
6412 |
}
|
|
6413 |
if (!map || (map & (RAND_TABLE_BIT | OUTER_REF_TABLE_BIT))) |
|
6414 |
return (Table *) NULL; |
|
6415 |
||
6416 |
for (; !(map & tables->table->map); tables= tables->next_leaf) {}; |
|
6417 |
if (map != tables->table->map) |
|
6418 |
return (Table *) NULL; // More than one table |
|
6419 |
return tables->table; |
|
6420 |
}
|
|
6421 |
||
6422 |
/**
|
|
6423 |
Set nested_join_st::counter=0 in all nested joins in passed list.
|
|
6424 |
||
6425 |
Recursively set nested_join_st::counter=0 for all nested joins contained in
|
|
6426 |
the passed join_list.
|
|
6427 |
||
6428 |
@param join_list List of nested joins to process. It may also contain base
|
|
6429 |
tables which will be ignored.
|
|
6430 |
*/
|
|
6431 |
static void reset_nj_counters(List<TableList> *join_list) |
|
6432 |
{
|
|
6433 |
List_iterator<TableList> li(*join_list); |
|
6434 |
TableList *table; |
|
6435 |
while ((table= li++)) |
|
6436 |
{
|
|
6437 |
nested_join_st *nested_join; |
|
6438 |
if ((nested_join= table->nested_join)) |
|
6439 |
{
|
|
6440 |
nested_join->counter_= 0; |
|
6441 |
reset_nj_counters(&nested_join->join_list); |
|
6442 |
}
|
|
6443 |
}
|
|
6444 |
return; |
|
6445 |
}
|
|
6446 |
||
6447 |
/**
|
|
6448 |
Return 1 if second is a subpart of first argument.
|
|
6449 |
||
6450 |
If first parts has different direction, change it to second part
|
|
6451 |
(group is sorted like order)
|
|
6452 |
*/
|
|
6453 |
static bool test_if_subpart(order_st *a,order_st *b) |
|
6454 |
{
|
|
6455 |
for (; a && b; a=a->next,b=b->next) |
|
6456 |
{
|
|
6457 |
if ((*a->item)->eq(*b->item,1)) |
|
6458 |
a->asc=b->asc; |
|
6459 |
else
|
|
6460 |
return 0; |
|
6461 |
}
|
|
6462 |
return test(!b); |
|
6463 |
}
|
|
6464 |
||
6465 |
/**
|
|
6466 |
Nested joins perspective: Remove the last table from the join order.
|
|
6467 |
||
6468 |
Remove the last table from the partial join order and update the nested
|
|
6469 |
joins counters and join->cur_embedding_map. It is ok to call this
|
|
6470 |
function for the first table in join order (for which
|
|
6471 |
check_interleaving_with_nj has not been called)
|
|
6472 |
||
6473 |
@param last join table to remove, it is assumed to be the last in current
|
|
6474 |
partial join order.
|
|
6475 |
*/
|
|
6476 |
static void restore_prev_nj_state(JOIN_TAB *last) |
|
6477 |
{
|
|
6478 |
TableList *last_emb= last->table->pos_in_table_list->embedding; |
|
6479 |
JOIN *join= last->join; |
|
6480 |
while (last_emb) |
|
6481 |
{
|
|
6482 |
if (last_emb->on_expr) |
|
6483 |
{
|
|
6484 |
if (!(--last_emb->nested_join->counter_)) |
|
6485 |
join->cur_embedding_map&= ~last_emb->nested_join->nj_map; |
|
6486 |
else if (last_emb->nested_join->join_list.elements-1 == |
|
6487 |
last_emb->nested_join->counter_) |
|
6488 |
join->cur_embedding_map|= last_emb->nested_join->nj_map; |
|
6489 |
else
|
|
6490 |
break; |
|
6491 |
}
|
|
6492 |
last_emb= last_emb->embedding; |
|
6493 |
}
|
|
6494 |
}
|
|
6495 |
||
6496 |
/**
|
|
6497 |
Determine if the set is already ordered for order_st BY, so it can
|
|
6498 |
disable join cache because it will change the ordering of the results.
|
|
6499 |
Code handles sort table that is at any location (not only first after
|
|
6500 |
the const tables) despite the fact that it's currently prohibited.
|
|
6501 |
We must disable join cache if the first non-const table alone is
|
|
6502 |
ordered. If there is a temp table the ordering is done as a last
|
|
6503 |
operation and doesn't prevent join cache usage.
|
|
6504 |
*/
|
|
6505 |
static uint32_t make_join_orderinfo(JOIN *join) |
|
6506 |
{
|
|
6507 |
uint32_t i; |
|
6508 |
if (join->need_tmp) |
|
6509 |
return join->tables; |
|
6510 |
||
6511 |
for (i=join->const_tables ; i < join->tables ; i++) |
|
6512 |
{
|
|
6513 |
JOIN_TAB *tab= join->join_tab+i; |
|
6514 |
Table *table= tab->table; |
|
6515 |
if ((table == join->sort_by_table && |
|
6516 |
(!join->order || join->skip_sort_order)) || |
|
6517 |
(join->sort_by_table == (Table *) 1 && i != join->const_tables)) |
|
6518 |
{
|
|
6519 |
break; |
|
6520 |
}
|
|
6521 |
}
|
|
6522 |
return i; |
|
6523 |
}
|
|
6524 |
||
6525 |
/**
|
|
6526 |
Setup the strategies to eliminate semi-join duplicates.
|
|
6527 |
||
6528 |
SYNOPSIS
|
|
6529 |
setup_semijoin_dups_elimination()
|
|
6530 |
join Join to process
|
|
6531 |
options Join options (needed to see if join buffering will be
|
|
6532 |
used or not)
|
|
6533 |
no_jbuf_after Another bit of information re where join buffering will
|
|
6534 |
be used.
|
|
6535 |
||
6536 |
DESCRIPTION
|
|
6537 |
Setup the strategies to eliminate semi-join duplicates. ATM there are 3
|
|
6538 |
strategies:
|
|
6539 |
||
6540 |
1. DuplicateWeedout (use of temptable to remove duplicates based on rowids
|
|
6541 |
of row combinations)
|
|
6542 |
2. FirstMatch (pick only the 1st matching row combination of inner tables)
|
|
6543 |
3. InsideOut (scanning the sj-inner table in a way that groups duplicates
|
|
6544 |
together and picking the 1st one)
|
|
6545 |
||
6546 |
The join order has "duplicate-generating ranges", and every range is
|
|
6547 |
served by one strategy or a combination of FirstMatch with with some
|
|
6548 |
other strategy.
|
|
6549 |
||
6550 |
"Duplicate-generating range" is defined as a range within the join order
|
|
6551 |
that contains all of the inner tables of a semi-join. All ranges must be
|
|
6552 |
disjoint, if tables of several semi-joins are interleaved, then the ranges
|
|
6553 |
are joined together, which is equivalent to converting
|
|
6554 |
SELECT ... WHERE oe1 IN (SELECT ie1 ...) AND oe2 IN (SELECT ie2 )
|
|
6555 |
to
|
|
6556 |
SELECT ... WHERE (oe1, oe2) IN (SELECT ie1, ie2 ... ...)
|
|
6557 |
.
|
|
6558 |
||
6559 |
Applicability conditions are as follows:
|
|
6560 |
||
6561 |
DuplicateWeedout strategy
|
|
6562 |
~~~~~~~~~~~~~~~~~~~~~~~~~
|
|
6563 |
||
6564 |
(ot|nt)* [ it ((it|ot|nt)* (it|ot))] (nt)*
|
|
6565 |
+------+ +=========================+ +---+
|
|
6566 |
(1) (2) (3)
|
|
6567 |
||
6568 |
(1) - Prefix of OuterTables (those that participate in
|
|
6569 |
IN-equality and/or are correlated with subquery) and outer
|
|
6570 |
Noncorrelated Tables.
|
|
6571 |
(2) - The handled range. The range starts with the first sj-inner
|
|
6572 |
table, and covers all sj-inner and outer tables
|
|
6573 |
Within the range, Inner, Outer, outer Noncorrelated tables
|
|
6574 |
may follow in any order.
|
|
6575 |
(3) - The suffix of outer Noncorrelated tables.
|
|
6576 |
||
6577 |
FirstMatch strategy
|
|
6578 |
~~~~~~~~~~~~~~~~~~~
|
|
6579 |
||
6580 |
(ot|nt)* [ it ((it|nt)* it) ] (nt)*
|
|
6581 |
+------+ +==================+ +---+
|
|
6582 |
(1) (2) (3)
|
|
6583 |
||
6584 |
(1) - Prefix of outer and non-correlated tables
|
|
6585 |
(2) - The handled range, which may contain only inner and
|
|
6586 |
non-correlated tables.
|
|
6587 |
(3) - The suffix of outer Noncorrelated tables.
|
|
6588 |
||
6589 |
InsideOut strategy
|
|
6590 |
~~~~~~~~~~~~~~~~~~
|
|
6591 |
||
6592 |
(ot|ct|nt) [ insideout_tbl (ot|nt|it)* it ] (ot|nt)*
|
|
6593 |
+--------+ +===========+ +=============+ +------+
|
|
6594 |
(1) (2) (3) (4)
|
|
6595 |
||
6596 |
(1) - Prefix that may contain any outer tables. The prefix must contain
|
|
6597 |
all the non-trivially correlated outer tables. (non-trivially means
|
|
6598 |
that the correlation is not just through the IN-equality).
|
|
6599 |
||
6600 |
(2) - Inner table for which the InsideOut scan is performed.
|
|
6601 |
||
6602 |
(3) - The remainder of the duplicate-generating range. It is served by
|
|
6603 |
application of FirstMatch strategy, with the exception that
|
|
6604 |
outer IN-correlated tables are considered to be non-correlated.
|
|
6605 |
||
6606 |
(4) - THe suffix of outer and outer non-correlated tables.
|
|
6607 |
||
6608 |
If several strategies are applicable, their relative priorities are:
|
|
6609 |
1. InsideOut
|
|
6610 |
2. FirstMatch
|
|
6611 |
3. DuplicateWeedout
|
|
6612 |
||
6613 |
This function walks over the join order and sets up the strategies by
|
|
6614 |
setting appropriate members in join_tab structures.
|
|
6615 |
||
6616 |
RETURN
|
|
6617 |
false OK
|
|
6618 |
true Out of memory error
|
|
6619 |
*/
|
|
6620 |
static int setup_semijoin_dups_elimination(JOIN *join, uint64_t options, uint32_t no_jbuf_after) |
|
6621 |
{
|
|
6622 |
table_map cur_map= join->const_table_map | PSEUDO_TABLE_BITS; |
|
6623 |
struct { |
|
6624 |
/*
|
|
6625 |
0 - invalid (EOF marker),
|
|
6626 |
1 - InsideOut,
|
|
6627 |
2 - Temptable (maybe confluent),
|
|
6628 |
3 - Temptable with join buffering
|
|
6629 |
*/
|
|
6630 |
uint32_t strategy; |
|
6631 |
uint32_t start_idx; /* Left range bound */ |
|
6632 |
uint32_t end_idx; /* Right range bound */ |
|
6633 |
/*
|
|
6634 |
For Temptable strategy: Bitmap of all outer and correlated tables from
|
|
6635 |
all involved join nests.
|
|
6636 |
*/
|
|
6637 |
table_map outer_tables; |
|
6638 |
} dups_ranges [MAX_TABLES]; |
|
6639 |
||
6640 |
TableList *emb_insideout_nest= NULL; |
|
6641 |
table_map emb_sj_map= 0; /* A bitmap of sj-nests (that is, their sj-inner |
|
6642 |
tables) whose ranges we're in */
|
|
6643 |
table_map emb_outer_tables= 0; /* sj-outer tables for those sj-nests */ |
|
6644 |
table_map range_start_map= 0; /* table_map at current range start */ |
|
6645 |
bool dealing_with_jbuf= false; /* true <=> table within cur range uses join buf */ |
|
6646 |
int cur_range= 0; |
|
6647 |
uint32_t i; |
|
6648 |
||
6649 |
/*
|
|
6650 |
First pass: locate the duplicate-generating ranges and pick the strategies.
|
|
6651 |
*/
|
|
6652 |
for (i=join->const_tables ; i < join->tables ; i++) |
|
6653 |
{
|
|
6654 |
JOIN_TAB *tab=join->join_tab+i; |
|
6655 |
Table *table=tab->table; |
|
6656 |
cur_map |= table->map; |
|
6657 |
||
6658 |
if (tab->emb_sj_nest) // Encountered an sj-inner table |
|
6659 |
{
|
|
6660 |
if (!emb_sj_map) |
|
6661 |
{
|
|
6662 |
dups_ranges[cur_range].start_idx= i; |
|
6663 |
range_start_map= cur_map & ~table->map; |
|
6664 |
/*
|
|
6665 |
Remember if this is a possible start of range that is covered by
|
|
6666 |
the InsideOut strategy (the reason that it is not covered could
|
|
6667 |
be that it overlaps with anther semi-join's range. we don't
|
|
6668 |
support InsideOut for joined ranges)
|
|
6669 |
*/
|
|
6670 |
if (join->best_positions[i].use_insideout_scan) |
|
6671 |
emb_insideout_nest= tab->emb_sj_nest; |
|
6672 |
}
|
|
6673 |
||
6674 |
emb_sj_map |= tab->emb_sj_nest->sj_inner_tables; |
|
6675 |
emb_outer_tables |= tab->emb_sj_nest->nested_join->sj_depends_on; |
|
6676 |
||
6677 |
if (tab->emb_sj_nest != emb_insideout_nest) |
|
6678 |
{
|
|
6679 |
/*
|
|
6680 |
Two different semi-joins interleave. This cannot be handled by
|
|
6681 |
InsideOut strategy.
|
|
6682 |
*/
|
|
6683 |
emb_insideout_nest= NULL; |
|
6684 |
}
|
|
6685 |
}
|
|
6686 |
||
6687 |
if (emb_sj_map) /* We're in duplicate-generating range */ |
|
6688 |
{
|
|
6689 |
if (i != join->const_tables && !(options & SELECT_NO_JOIN_CACHE) && |
|
6690 |
tab->type == JT_ALL && tab->use_quick != 2 && !tab->first_inner && |
|
6691 |
i <= no_jbuf_after && !dealing_with_jbuf) |
|
6692 |
{
|
|
6693 |
/*
|
|
6694 |
This table uses join buffering, which makes use of FirstMatch or
|
|
6695 |
InsideOut strategies impossible for the current and (we assume)
|
|
6696 |
preceding duplicate-producing ranges.
|
|
6697 |
That is, for the join order:
|
|
6698 |
||
6699 |
x x [ x x] x [x x x] x [x x X* x] x
|
|
6700 |
| | | | | \
|
|
6701 |
+-----+ +-----+ | join buffering use
|
|
6702 |
r1 r2 we're here
|
|
6703 |
||
6704 |
we'll have to remove r1 and r2 and use duplicate-elimination
|
|
6705 |
strategy that spans all the tables, starting from the very 1st
|
|
6706 |
one.
|
|
6707 |
*/
|
|
6708 |
dealing_with_jbuf= true; |
|
6709 |
emb_insideout_nest= false; |
|
6710 |
||
6711 |
/*
|
|
6712 |
Absorb all preceding duplicate-eliminating ranges. Their strategies
|
|
6713 |
do not matter:
|
|
6714 |
*/
|
|
6715 |
for (int prev_range= 0; prev_range < cur_range; prev_range++) |
|
6716 |
{
|
|
6717 |
dups_ranges[cur_range].outer_tables |= |
|
6718 |
dups_ranges[prev_range].outer_tables; |
|
6719 |
}
|
|
6720 |
dups_ranges[0].start_idx= 0; /* Will need to start from the 1st table */ |
|
6721 |
dups_ranges[0].outer_tables= dups_ranges[cur_range].outer_tables; |
|
6722 |
cur_range= 0; |
|
6723 |
}
|
|
6724 |
||
6725 |
/*
|
|
6726 |
Check if we are at the end of duplicate-producing range. We are if
|
|
6727 |
||
6728 |
1. It's an InsideOut range (which presumes all correlated tables are
|
|
6729 |
in the prefix), and all inner tables are in the join order prefix,
|
|
6730 |
or
|
|
6731 |
2. It's a DuplicateElimination range (possibly covering several
|
|
6732 |
SJ-nests), and all inner, outer, and correlated tables of all
|
|
6733 |
sj-nests are in the join order prefix.
|
|
6734 |
*/
|
|
6735 |
bool end_of_range= false; |
|
6736 |
if (emb_insideout_nest && |
|
6737 |
bitmap_covers(cur_map, emb_insideout_nest->sj_inner_tables)) |
|
6738 |
{
|
|
6739 |
/* Save that this range is handled with InsideOut: */
|
|
6740 |
dups_ranges[cur_range].strategy= 1; |
|
6741 |
end_of_range= true; |
|
6742 |
}
|
|
6743 |
else if (bitmap_covers(cur_map, emb_outer_tables | emb_sj_map)) |
|
6744 |
{
|
|
6745 |
/*
|
|
6746 |
This is a complete range to be handled with either DuplicateWeedout
|
|
6747 |
or FirstMatch
|
|
6748 |
*/
|
|
6749 |
dups_ranges[cur_range].strategy= dealing_with_jbuf? 3 : 2; |
|
6750 |
/*
|
|
6751 |
This will hold tables from within the range that need to be put
|
|
6752 |
into the join buffer before we can use the FirstMatch on its tail.
|
|
6753 |
*/
|
|
6754 |
dups_ranges[cur_range].outer_tables= emb_outer_tables & |
|
6755 |
~range_start_map; |
|
6756 |
end_of_range= true; |
|
6757 |
}
|
|
6758 |
||
6759 |
if (end_of_range) |
|
6760 |
{
|
|
6761 |
dups_ranges[cur_range].end_idx= i+1; |
|
6762 |
emb_sj_map= emb_outer_tables= 0; |
|
6763 |
emb_insideout_nest= NULL; |
|
6764 |
dealing_with_jbuf= false; |
|
6765 |
dups_ranges[++cur_range].strategy= 0; |
|
6766 |
}
|
|
6767 |
}
|
|
6768 |
}
|
|
6769 |
||
6770 |
Session *session= join->session; |
|
6771 |
SJ_TMP_TABLE **next_sjtbl_ptr= &join->sj_tmp_tables; |
|
6772 |
/*
|
|
6773 |
Second pass: setup the chosen strategies
|
|
6774 |
*/
|
|
6775 |
for (int j= 0; j < cur_range; j++) |
|
6776 |
{
|
|
6777 |
JOIN_TAB *tab=join->join_tab + dups_ranges[j].start_idx; |
|
6778 |
JOIN_TAB *jump_to; |
|
6779 |
if (dups_ranges[j].strategy == 1) // InsideOut strategy |
|
6780 |
{
|
|
6781 |
tab->insideout_match_tab= join->join_tab + dups_ranges[j].end_idx - 1; |
|
6782 |
jump_to= tab++; |
|
6783 |
}
|
|
6784 |
else // DuplicateWeedout strategy |
|
6785 |
{
|
|
6786 |
SJ_TMP_TABLE::TAB sjtabs[MAX_TABLES]; |
|
6787 |
table_map weed_cur_map= join->const_table_map | PSEUDO_TABLE_BITS; |
|
6788 |
uint32_t jt_rowid_offset= 0; // # tuple bytes are already occupied (w/o NULL bytes) |
|
6789 |
uint32_t jt_null_bits= 0; // # null bits in tuple bytes |
|
6790 |
SJ_TMP_TABLE::TAB *last_tab= sjtabs; |
|
6791 |
uint32_t rowid_keep_flags= JOIN_TAB::CALL_POSITION | JOIN_TAB::KEEP_ROWID; |
|
6792 |
JOIN_TAB *last_outer_tab= tab - 1; |
|
6793 |
/*
|
|
6794 |
Walk through the range and remember
|
|
6795 |
- tables that need their rowids to be put into temptable
|
|
6796 |
- the last outer table
|
|
6797 |
*/
|
|
6798 |
for (; tab < join->join_tab + dups_ranges[j].end_idx; tab++) |
|
6799 |
{
|
|
6800 |
if (sj_table_is_included(join, tab)) |
|
6801 |
{
|
|
6802 |
last_tab->join_tab= tab; |
|
6803 |
last_tab->rowid_offset= jt_rowid_offset; |
|
6804 |
jt_rowid_offset += tab->table->file->ref_length; |
|
6805 |
if (tab->table->maybe_null) |
|
6806 |
{
|
|
6807 |
last_tab->null_byte= jt_null_bits / 8; |
|
6808 |
last_tab->null_bit= jt_null_bits++; |
|
6809 |
}
|
|
6810 |
last_tab++; |
|
6811 |
tab->table->prepare_for_position(); |
|
6812 |
tab->rowid_keep_flags= rowid_keep_flags; |
|
6813 |
}
|
|
6814 |
weed_cur_map |= tab->table->map; |
|
6815 |
if (!tab->emb_sj_nest && bitmap_covers(weed_cur_map, |
|
6816 |
dups_ranges[j].outer_tables)) |
|
6817 |
last_outer_tab= tab; |
|
6818 |
}
|
|
6819 |
||
6820 |
if (jt_rowid_offset) /* Temptable has at least one rowid */ |
|
6821 |
{
|
|
6822 |
SJ_TMP_TABLE *sjtbl; |
|
6823 |
uint32_t tabs_size= (last_tab - sjtabs) * sizeof(SJ_TMP_TABLE::TAB); |
|
6824 |
if (!(sjtbl= (SJ_TMP_TABLE*)session->alloc(sizeof(SJ_TMP_TABLE))) || |
|
6825 |
!(sjtbl->tabs= (SJ_TMP_TABLE::TAB*) session->alloc(tabs_size))) |
|
6826 |
return(true); |
|
6827 |
memcpy(sjtbl->tabs, sjtabs, tabs_size); |
|
6828 |
sjtbl->tabs_end= sjtbl->tabs + (last_tab - sjtabs); |
|
6829 |
sjtbl->rowid_len= jt_rowid_offset; |
|
6830 |
sjtbl->null_bits= jt_null_bits; |
|
6831 |
sjtbl->null_bytes= (jt_null_bits + 7)/8; |
|
6832 |
||
6833 |
*next_sjtbl_ptr= sjtbl; |
|
6834 |
next_sjtbl_ptr= &(sjtbl->next); |
|
6835 |
sjtbl->next= NULL; |
|
6836 |
||
6837 |
sjtbl->tmp_table= |
|
6838 |
create_duplicate_weedout_tmp_table(session, |
|
6839 |
sjtbl->rowid_len + |
|
6840 |
sjtbl->null_bytes, |
|
6841 |
sjtbl); |
|
6842 |
||
6843 |
join->join_tab[dups_ranges[j].start_idx].flush_weedout_table= sjtbl; |
|
6844 |
join->join_tab[dups_ranges[j].end_idx - 1].check_weed_out_table= sjtbl; |
|
6845 |
}
|
|
6846 |
tab= last_outer_tab + 1; |
|
6847 |
jump_to= last_outer_tab; |
|
6848 |
}
|
|
6849 |
||
6850 |
/* Create the FirstMatch tail */
|
|
6851 |
for (; tab < join->join_tab + dups_ranges[j].end_idx; tab++) |
|
6852 |
{
|
|
6853 |
if (tab->emb_sj_nest) |
|
6854 |
tab->do_firstmatch= jump_to; |
|
6855 |
else
|
|
6856 |
jump_to= tab; |
|
6857 |
}
|
|
6858 |
}
|
|
6859 |
return(false); |
|
6860 |
}
|
|
6861 |
||
6862 |
static void cleanup_sj_tmp_tables(JOIN *join) |
|
6863 |
{
|
|
6864 |
for (SJ_TMP_TABLE *sj_tbl= join->sj_tmp_tables; sj_tbl; |
|
6865 |
sj_tbl= sj_tbl->next) |
|
6866 |
{
|
|
6867 |
if (sj_tbl->tmp_table) |
|
6868 |
{
|
|
6869 |
sj_tbl->tmp_table->free_tmp_table(join->session); |
|
6870 |
}
|
|
6871 |
}
|
|
6872 |
join->sj_tmp_tables= NULL; |
|
6873 |
}
|
|
6874 |
||
6875 |
/**
|
|
6876 |
Create a condition for a const reference and add this to the
|
|
6877 |
currenct select for the table.
|
|
6878 |
*/
|
|
6879 |
static bool add_ref_to_table_cond(Session *session, JOIN_TAB *join_tab) |
|
6880 |
{
|
|
6881 |
if (!join_tab->ref.key_parts) |
|
6882 |
return(false); |
|
6883 |
||
6884 |
Item_cond_and *cond=new Item_cond_and(); |
|
6885 |
Table *table=join_tab->table; |
|
6886 |
int error; |
|
6887 |
if (!cond) |
|
6888 |
return(true); |
|
6889 |
||
6890 |
for (uint32_t i=0 ; i < join_tab->ref.key_parts ; i++) |
|
6891 |
{
|
|
6892 |
Field *field=table->field[table->key_info[join_tab->ref.key].key_part[i]. |
|
6893 |
fieldnr-1]; |
|
6894 |
Item *value=join_tab->ref.items[i]; |
|
6895 |
cond->add(new Item_func_equal(new Item_field(field), value)); |
|
6896 |
}
|
|
6897 |
if (session->is_fatal_error) |
|
6898 |
return(true); |
|
6899 |
||
6900 |
if (!cond->fixed) |
|
6901 |
cond->fix_fields(session, (Item**)&cond); |
|
6902 |
if (join_tab->select) |
|
6903 |
{
|
|
6904 |
error=(int) cond->add(join_tab->select->cond); |
|
6905 |
join_tab->select_cond=join_tab->select->cond=cond; |
|
6906 |
}
|
|
6907 |
else if ((join_tab->select= make_select(join_tab->table, 0, 0, cond, 0, |
|
6908 |
&error))) |
|
6909 |
join_tab->select_cond=cond; |
|
6910 |
||
6911 |
return(error ? true : false); |
|
6912 |
}
|
|
6913 |
||
6914 |
/**
|
|
6915 |
@brief Replaces an expression destructively inside the expression tree of
|
|
6916 |
the WHERE clase.
|
|
6917 |
||
6918 |
@note Because of current requirements for semijoin flattening, we do not
|
|
6919 |
need to recurse here, hence this function will only examine the top-level
|
|
6920 |
AND conditions. (see JOIN::prepare, comment above the line
|
|
6921 |
'if (do_materialize)'
|
|
6922 |
||
6923 |
@param join The top-level query.
|
|
6924 |
@param old_cond The expression to be replaced.
|
|
6925 |
@param new_cond The expression to be substituted.
|
|
6926 |
@param do_fix_fields If true, Item::fix_fields(Session*, Item**) is called for
|
|
6927 |
the new expression.
|
|
6928 |
@return <code>true</code> if there was an error, <code>false</code> if
|
|
6929 |
successful.
|
|
6930 |
*/
|
|
6931 |
static bool replace_where_subcondition(JOIN *join, Item *old_cond, |
|
6932 |
Item *new_cond, bool do_fix_fields) |
|
6933 |
{
|
|
6934 |
if (join->conds == old_cond) { |
|
6935 |
join->conds= new_cond; |
|
6936 |
if (do_fix_fields) |
|
6937 |
new_cond->fix_fields(join->session, &join->conds); |
|
6938 |
return false; |
|
6939 |
}
|
|
6940 |
||
6941 |
if (join->conds->type() == Item::COND_ITEM) { |
|
6942 |
List_iterator<Item> li(*((Item_cond*)join->conds)->argument_list()); |
|
6943 |
Item *item; |
|
6944 |
while ((item= li++)) |
|
6945 |
if (item == old_cond) |
|
6946 |
{
|
|
6947 |
li.replace(new_cond); |
|
6948 |
if (do_fix_fields) |
|
6949 |
new_cond->fix_fields(join->session, li.ref()); |
|
6950 |
return false; |
|
6951 |
}
|
|
6952 |
}
|
|
6953 |
||
6954 |
return true; |
|
6955 |
}
|
|
6956 |
||
6957 |
/*
|
|
6958 |
Pull tables out of semi-join nests, if possible
|
|
6959 |
||
6960 |
SYNOPSIS
|
|
6961 |
pull_out_semijoin_tables()
|
|
6962 |
join The join where to do the semi-join flattening
|
|
6963 |
||
6964 |
DESCRIPTION
|
|
6965 |
Try to pull tables out of semi-join nests.
|
|
6966 |
||
6967 |
PRECONDITIONS
|
|
6968 |
When this function is called, the join may have several semi-join nests
|
|
6969 |
(possibly within different semi-join nests), but it is guaranteed that
|
|
6970 |
one semi-join nest does not contain another.
|
|
6971 |
||
6972 |
ACTION
|
|
6973 |
A table can be pulled out of the semi-join nest if
|
|
6974 |
- It is a constant table
|
|
6975 |
- It is accessed
|
|
6976 |
||
6977 |
POSTCONDITIONS
|
|
6978 |
* Pulled out tables have JOIN_TAB::emb_sj_nest == NULL (like the outer
|
|
6979 |
tables)
|
|
6980 |
* Tables that were not pulled out have JOIN_TAB::emb_sj_nest.
|
|
6981 |
* Semi-join nests TableList::sj_inner_tables
|
|
6982 |
||
6983 |
This operation is (and should be) performed at each PS execution since
|
|
6984 |
tables may become/cease to be constant across PS reexecutions.
|
|
6985 |
||
6986 |
RETURN
|
|
6987 |
0 - OK
|
|
6988 |
1 - Out of memory error
|
|
6989 |
*/
|
|
6990 |
static int pull_out_semijoin_tables(JOIN *join) |
|
6991 |
{
|
|
6992 |
TableList *sj_nest; |
|
6993 |
List_iterator<TableList> sj_list_it(join->select_lex->sj_nests); |
|
6994 |
||
6995 |
/* Try pulling out of the each of the semi-joins */
|
|
6996 |
while ((sj_nest= sj_list_it++)) |
|
6997 |
{
|
|
6998 |
/* Action #1: Mark the constant tables to be pulled out */
|
|
6999 |
table_map pulled_tables= 0; |
|
7000 |
||
7001 |
List_iterator<TableList> child_li(sj_nest->nested_join->join_list); |
|
7002 |
TableList *tbl; |
|
7003 |
while ((tbl= child_li++)) |
|
7004 |
{
|
|
7005 |
if (tbl->table) |
|
7006 |
{
|
|
7007 |
tbl->table->reginfo.join_tab->emb_sj_nest= sj_nest; |
|
7008 |
if (tbl->table->map & join->const_table_map) |
|
7009 |
{
|
|
7010 |
pulled_tables |= tbl->table->map; |
|
7011 |
}
|
|
7012 |
}
|
|
7013 |
}
|
|
7014 |
||
7015 |
/*
|
|
7016 |
Action #2: Find which tables we can pull out based on
|
|
7017 |
update_ref_and_keys() data. Note that pulling one table out can allow
|
|
7018 |
us to pull out some other tables too.
|
|
7019 |
*/
|
|
7020 |
bool pulled_a_table; |
|
7021 |
do
|
|
7022 |
{
|
|
7023 |
pulled_a_table= false; |
|
7024 |
child_li.rewind(); |
|
7025 |
while ((tbl= child_li++)) |
|
7026 |
{
|
|
7027 |
if (tbl->table && !(pulled_tables & tbl->table->map)) |
|
7028 |
{
|
|
7029 |
if (find_eq_ref_candidate(tbl->table, |
|
7030 |
sj_nest->nested_join->used_tables & |
|
7031 |
~pulled_tables)) |
|
7032 |
{
|
|
7033 |
pulled_a_table= true; |
|
7034 |
pulled_tables |= tbl->table->map; |
|
7035 |
}
|
|
7036 |
}
|
|
7037 |
}
|
|
7038 |
} while (pulled_a_table); |
|
7039 |
||
7040 |
child_li.rewind(); |
|
7041 |
if ((sj_nest)->nested_join->used_tables == pulled_tables) |
|
7042 |
{
|
|
7043 |
(sj_nest)->sj_inner_tables= 0; |
|
7044 |
while ((tbl= child_li++)) |
|
7045 |
{
|
|
7046 |
if (tbl->table) |
|
7047 |
tbl->table->reginfo.join_tab->emb_sj_nest= NULL; |
|
7048 |
}
|
|
7049 |
}
|
|
7050 |
else
|
|
7051 |
{
|
|
7052 |
/* Record the bitmap of inner tables, mark the inner tables */
|
|
7053 |
table_map inner_tables=(sj_nest)->nested_join->used_tables & |
|
7054 |
~pulled_tables; |
|
7055 |
(sj_nest)->sj_inner_tables= inner_tables; |
|
7056 |
while ((tbl= child_li++)) |
|
7057 |
{
|
|
7058 |
if (tbl->table) |
|
7059 |
{
|
|
7060 |
if (inner_tables & tbl->table->map) |
|
7061 |
tbl->table->reginfo.join_tab->emb_sj_nest= (sj_nest); |
|
7062 |
else
|
|
7063 |
tbl->table->reginfo.join_tab->emb_sj_nest= NULL; |
|
7064 |
}
|
|
7065 |
}
|
|
7066 |
}
|
|
7067 |
}
|
|
7068 |
return(0); |
|
7069 |
}
|
|
7070 |
||
7071 |
/*
|
|
7072 |
SemiJoinDuplicateElimination: Weed out duplicate row combinations
|
|
7073 |
||
7074 |
SYNPOSIS
|
|
7075 |
do_sj_dups_weedout()
|
|
7076 |
||
7077 |
RETURN
|
|
7078 |
-1 Error
|
|
7079 |
1 The row combination is a duplicate (discard it)
|
|
7080 |
0 The row combination is not a duplicate (continue)
|
|
7081 |
*/
|
|
7082 |
static int do_sj_dups_weedout(Session *session, SJ_TMP_TABLE *sjtbl) |
|
7083 |
{
|
|
7084 |
int error; |
|
7085 |
SJ_TMP_TABLE::TAB *tab= sjtbl->tabs; |
|
7086 |
SJ_TMP_TABLE::TAB *tab_end= sjtbl->tabs_end; |
|
7087 |
unsigned char *ptr= sjtbl->tmp_table->record[0] + 1; |
|
7088 |
unsigned char *nulls_ptr= ptr; |
|
7089 |
||
7090 |
/* Put the the rowids tuple into table->record[0]: */
|
|
7091 |
||
7092 |
// 1. Store the length
|
|
7093 |
if (((Field_varstring*)(sjtbl->tmp_table->field[0]))->length_bytes == 1) |
|
7094 |
{
|
|
7095 |
*ptr= (unsigned char)(sjtbl->rowid_len + sjtbl->null_bytes); |
|
7096 |
ptr++; |
|
7097 |
}
|
|
7098 |
else
|
|
7099 |
{
|
|
7100 |
int2store(ptr, sjtbl->rowid_len + sjtbl->null_bytes); |
|
7101 |
ptr += 2; |
|
7102 |
}
|
|
7103 |
||
7104 |
// 2. Zero the null bytes
|
|
7105 |
if (sjtbl->null_bytes) |
|
7106 |
{
|
|
7107 |
memset(ptr, 0, sjtbl->null_bytes); |
|
7108 |
ptr += sjtbl->null_bytes; |
|
7109 |
}
|
|
7110 |
||
7111 |
// 3. Put the rowids
|
|
7112 |
for (uint32_t i=0; tab != tab_end; tab++, i++) |
|
7113 |
{
|
|
7114 |
handler *h= tab->join_tab->table->file; |
|
7115 |
if (tab->join_tab->table->maybe_null && tab->join_tab->table->null_row) |
|
7116 |
{
|
|
7117 |
/* It's a NULL-complemented row */
|
|
7118 |
*(nulls_ptr + tab->null_byte) |= tab->null_bit; |
|
7119 |
memset(ptr + tab->rowid_offset, 0, h->ref_length); |
|
7120 |
}
|
|
7121 |
else
|
|
7122 |
{
|
|
7123 |
/* Copy the rowid value */
|
|
7124 |
if (tab->join_tab->rowid_keep_flags & JOIN_TAB::CALL_POSITION) |
|
7125 |
h->position(tab->join_tab->table->record[0]); |
|
7126 |
memcpy(ptr + tab->rowid_offset, h->ref, h->ref_length); |
|
7127 |
}
|
|
7128 |
}
|
|
7129 |
||
7130 |
error= sjtbl->tmp_table->file->ha_write_row(sjtbl->tmp_table->record[0]); |
|
7131 |
if (error) |
|
7132 |
{
|
|
7133 |
/* create_myisam_from_heap will generate error if needed */
|
|
7134 |
if (sjtbl->tmp_table->file->is_fatal_error(error, HA_CHECK_DUP) && |
|
7135 |
create_myisam_from_heap(session, sjtbl->tmp_table, sjtbl->start_recinfo, |
|
7136 |
&sjtbl->recinfo, error, 1)) |
|
7137 |
return -1; |
|
7138 |
//return (error == HA_ERR_FOUND_DUPP_KEY || error== HA_ERR_FOUND_DUPP_UNIQUE) ? 1: -1;
|
|
7139 |
return 1; |
|
7140 |
}
|
|
7141 |
return 0; |
|
7142 |
}
|
|
7143 |
||
7144 |
static void free_blobs(Field **ptr) |
|
7145 |
{
|
|
7146 |
for (; *ptr ; ptr++) |
|
7147 |
{
|
|
7148 |
if ((*ptr)->flags & BLOB_FLAG) |
|
7149 |
((Field_blob *) (*ptr))->free(); |
|
7150 |
}
|
|
7151 |
}
|
|
7152 |
||
7153 |
static bool bitmap_covers(const table_map x, const table_map y) |
|
7154 |
{
|
|
7155 |
return !test(y & ~x); |
|
7156 |
}
|
|
7157 |
||
7158 |
/*
|
|
7159 |
Check if the table's rowid is included in the temptable
|
|
7160 |
||
7161 |
SYNOPSIS
|
|
7162 |
sj_table_is_included()
|
|
7163 |
join The join
|
|
7164 |
join_tab The table to be checked
|
|
7165 |
||
7166 |
DESCRIPTION
|
|
7167 |
SemiJoinDuplicateElimination: check the table's rowid should be included
|
|
7168 |
in the temptable. This is so if
|
|
7169 |
||
7170 |
1. The table is not embedded within some semi-join nest
|
|
7171 |
2. The has been pulled out of a semi-join nest, or
|
|
7172 |
||
7173 |
3. The table is functionally dependent on some previous table
|
|
7174 |
||
7175 |
[4. This is also true for constant tables that can't be
|
|
7176 |
NULL-complemented but this function is not called for such tables]
|
|
7177 |
||
7178 |
RETURN
|
|
7179 |
true - Include table's rowid
|
|
7180 |
false - Don't
|
|
7181 |
*/
|
|
7182 |
static bool sj_table_is_included(JOIN *join, JOIN_TAB *join_tab) |
|
7183 |
{
|
|
7184 |
if (join_tab->emb_sj_nest) |
|
7185 |
return false; |
|
7186 |
||
7187 |
/* Check if this table is functionally dependent on the tables that
|
|
7188 |
are within the same outer join nest
|
|
7189 |
*/
|
|
7190 |
TableList *embedding= join_tab->table->pos_in_table_list->embedding; |
|
7191 |
if (join_tab->type == JT_EQ_REF) |
|
7192 |
{
|
|
7193 |
Table_map_iterator it(join_tab->ref.depend_map & ~PSEUDO_TABLE_BITS); |
|
7194 |
uint32_t idx; |
|
7195 |
while ((idx= it.next_bit())!=Table_map_iterator::BITMAP_END) |
|
7196 |
{
|
|
7197 |
JOIN_TAB *ref_tab= join->join_tab + idx; |
|
7198 |
if (embedding == ref_tab->table->pos_in_table_list->embedding) |
|
7199 |
return true; |
|
7200 |
}
|
|
7201 |
/* Ok, functionally dependent */
|
|
7202 |
return false; |
|
7203 |
}
|
|
7204 |
/* Not functionally dependent => need to include*/
|
|
7205 |
return true; |
|
7206 |
}
|
|
7207 |
||
7208 |
/**
|
|
7209 |
@} (end of group Query_Optimizer)
|
|
7210 |
*/
|