Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * pathnode.c
4 : * Routines to manipulate pathlists and create path nodes
5 : *
6 : * Portions Copyright (c) 1996-2023, PostgreSQL Global Development Group
7 : * Portions Copyright (c) 1994, Regents of the University of California
8 : *
9 : *
10 : * IDENTIFICATION
11 : * src/backend/optimizer/util/pathnode.c
12 : *
13 : *-------------------------------------------------------------------------
14 : */
15 : #include "postgres.h"
16 :
17 : #include <math.h>
18 :
19 : #include "foreign/fdwapi.h"
20 : #include "miscadmin.h"
21 : #include "nodes/extensible.h"
22 : #include "nodes/nodeFuncs.h"
23 : #include "optimizer/appendinfo.h"
24 : #include "optimizer/clauses.h"
25 : #include "optimizer/cost.h"
26 : #include "optimizer/optimizer.h"
27 : #include "optimizer/pathnode.h"
28 : #include "optimizer/paths.h"
29 : #include "optimizer/planmain.h"
30 : #include "optimizer/prep.h"
31 : #include "optimizer/restrictinfo.h"
32 : #include "optimizer/tlist.h"
33 : #include "parser/parsetree.h"
34 : #include "utils/lsyscache.h"
35 : #include "utils/memutils.h"
36 : #include "utils/selfuncs.h"
37 :
38 : typedef enum
39 : {
40 : COSTS_EQUAL, /* path costs are fuzzily equal */
41 : COSTS_BETTER1, /* first path is cheaper than second */
42 : COSTS_BETTER2, /* second path is cheaper than first */
43 : COSTS_DIFFERENT /* neither path dominates the other on cost */
44 : } PathCostComparison;
45 :
46 : /*
47 : * STD_FUZZ_FACTOR is the normal fuzz factor for compare_path_costs_fuzzily.
48 : * XXX is it worth making this user-controllable? It provides a tradeoff
49 : * between planner runtime and the accuracy of path cost comparisons.
50 : */
51 : #define STD_FUZZ_FACTOR 1.01
52 :
53 : static List *translate_sub_tlist(List *tlist, int relid);
54 : static int append_total_cost_compare(const ListCell *a, const ListCell *b);
55 : static int append_startup_cost_compare(const ListCell *a, const ListCell *b);
56 : static List *reparameterize_pathlist_by_child(PlannerInfo *root,
57 : List *pathlist,
58 : RelOptInfo *child_rel);
59 :
60 :
61 : /*****************************************************************************
62 : * MISC. PATH UTILITIES
63 : *****************************************************************************/
64 :
65 : /*
66 : * compare_path_costs
67 : * Return -1, 0, or +1 according as path1 is cheaper, the same cost,
68 : * or more expensive than path2 for the specified criterion.
69 : */
70 : int
71 587776 : compare_path_costs(Path *path1, Path *path2, CostSelector criterion)
72 : {
73 587776 : if (criterion == STARTUP_COST)
74 : {
75 301400 : if (path1->startup_cost < path2->startup_cost)
76 185860 : return -1;
77 115540 : if (path1->startup_cost > path2->startup_cost)
78 52452 : return +1;
79 :
80 : /*
81 : * If paths have the same startup cost (not at all unlikely), order
82 : * them by total cost.
83 : */
84 63088 : if (path1->total_cost < path2->total_cost)
85 25024 : return -1;
86 38064 : if (path1->total_cost > path2->total_cost)
87 3936 : return +1;
88 : }
89 : else
90 : {
91 286376 : if (path1->total_cost < path2->total_cost)
92 271842 : return -1;
93 14534 : if (path1->total_cost > path2->total_cost)
94 1002 : return +1;
95 :
96 : /*
97 : * If paths have the same total cost, order them by startup cost.
98 : */
99 13532 : if (path1->startup_cost < path2->startup_cost)
100 44 : return -1;
101 13488 : if (path1->startup_cost > path2->startup_cost)
102 76 : return +1;
103 : }
104 47540 : return 0;
105 : }
106 :
107 : /*
108 : * compare_fractional_path_costs
109 : * Return -1, 0, or +1 according as path1 is cheaper, the same cost,
110 : * or more expensive than path2 for fetching the specified fraction
111 : * of the total tuples.
112 : *
113 : * If fraction is <= 0 or > 1, we interpret it as 1, ie, we select the
114 : * path with the cheaper total_cost.
115 : */
116 : int
117 5848 : compare_fractional_path_costs(Path *path1, Path *path2,
118 : double fraction)
119 : {
120 : Cost cost1,
121 : cost2;
122 :
123 5848 : if (fraction <= 0.0 || fraction >= 1.0)
124 4694 : return compare_path_costs(path1, path2, TOTAL_COST);
125 1154 : cost1 = path1->startup_cost +
126 1154 : fraction * (path1->total_cost - path1->startup_cost);
127 1154 : cost2 = path2->startup_cost +
128 1154 : fraction * (path2->total_cost - path2->startup_cost);
129 1154 : if (cost1 < cost2)
130 690 : return -1;
131 464 : if (cost1 > cost2)
132 464 : return +1;
133 0 : return 0;
134 : }
135 :
136 : /*
137 : * compare_path_costs_fuzzily
138 : * Compare the costs of two paths to see if either can be said to
139 : * dominate the other.
140 : *
141 : * We use fuzzy comparisons so that add_path() can avoid keeping both of
142 : * a pair of paths that really have insignificantly different cost.
143 : *
144 : * The fuzz_factor argument must be 1.0 plus delta, where delta is the
145 : * fraction of the smaller cost that is considered to be a significant
146 : * difference. For example, fuzz_factor = 1.01 makes the fuzziness limit
147 : * be 1% of the smaller cost.
148 : *
149 : * The two paths are said to have "equal" costs if both startup and total
150 : * costs are fuzzily the same. Path1 is said to be better than path2 if
151 : * it has fuzzily better startup cost and fuzzily no worse total cost,
152 : * or if it has fuzzily better total cost and fuzzily no worse startup cost.
153 : * Path2 is better than path1 if the reverse holds. Finally, if one path
154 : * is fuzzily better than the other on startup cost and fuzzily worse on
155 : * total cost, we just say that their costs are "different", since neither
156 : * dominates the other across the whole performance spectrum.
157 : *
158 : * This function also enforces a policy rule that paths for which the relevant
159 : * one of parent->consider_startup and parent->consider_param_startup is false
160 : * cannot survive comparisons solely on the grounds of good startup cost, so
161 : * we never return COSTS_DIFFERENT when that is true for the total-cost loser.
162 : * (But if total costs are fuzzily equal, we compare startup costs anyway,
163 : * in hopes of eliminating one path or the other.)
164 : */
165 : static PathCostComparison
166 2572548 : compare_path_costs_fuzzily(Path *path1, Path *path2, double fuzz_factor)
167 : {
168 : #define CONSIDER_PATH_STARTUP_COST(p) \
169 : ((p)->param_info == NULL ? (p)->parent->consider_startup : (p)->parent->consider_param_startup)
170 :
171 : /*
172 : * Check total cost first since it's more likely to be different; many
173 : * paths have zero startup cost.
174 : */
175 2572548 : if (path1->total_cost > path2->total_cost * fuzz_factor)
176 : {
177 : /* path1 fuzzily worse on total cost */
178 1301996 : if (CONSIDER_PATH_STARTUP_COST(path1) &&
179 44616 : path2->startup_cost > path1->startup_cost * fuzz_factor)
180 : {
181 : /* ... but path2 fuzzily worse on startup, so DIFFERENT */
182 21398 : return COSTS_DIFFERENT;
183 : }
184 : /* else path2 dominates */
185 1280598 : return COSTS_BETTER2;
186 : }
187 1270552 : if (path2->total_cost > path1->total_cost * fuzz_factor)
188 : {
189 : /* path2 fuzzily worse on total cost */
190 685256 : if (CONSIDER_PATH_STARTUP_COST(path2) &&
191 20970 : path1->startup_cost > path2->startup_cost * fuzz_factor)
192 : {
193 : /* ... but path1 fuzzily worse on startup, so DIFFERENT */
194 11882 : return COSTS_DIFFERENT;
195 : }
196 : /* else path1 dominates */
197 673374 : return COSTS_BETTER1;
198 : }
199 : /* fuzzily the same on total cost ... */
200 585296 : if (path1->startup_cost > path2->startup_cost * fuzz_factor)
201 : {
202 : /* ... but path1 fuzzily worse on startup, so path2 wins */
203 231382 : return COSTS_BETTER2;
204 : }
205 353914 : if (path2->startup_cost > path1->startup_cost * fuzz_factor)
206 : {
207 : /* ... but path2 fuzzily worse on startup, so path1 wins */
208 35276 : return COSTS_BETTER1;
209 : }
210 : /* fuzzily the same on both costs */
211 318638 : return COSTS_EQUAL;
212 :
213 : #undef CONSIDER_PATH_STARTUP_COST
214 : }
215 :
216 : /*
217 : * set_cheapest
218 : * Find the minimum-cost paths from among a relation's paths,
219 : * and save them in the rel's cheapest-path fields.
220 : *
221 : * cheapest_total_path is normally the cheapest-total-cost unparameterized
222 : * path; but if there are no unparameterized paths, we assign it to be the
223 : * best (cheapest least-parameterized) parameterized path. However, only
224 : * unparameterized paths are considered candidates for cheapest_startup_path,
225 : * so that will be NULL if there are no unparameterized paths.
226 : *
227 : * The cheapest_parameterized_paths list collects all parameterized paths
228 : * that have survived the add_path() tournament for this relation. (Since
229 : * add_path ignores pathkeys for a parameterized path, these will be paths
230 : * that have best cost or best row count for their parameterization. We
231 : * may also have both a parallel-safe and a non-parallel-safe path in some
232 : * cases for the same parameterization in some cases, but this should be
233 : * relatively rare since, most typically, all paths for the same relation
234 : * will be parallel-safe or none of them will.)
235 : *
236 : * cheapest_parameterized_paths always includes the cheapest-total
237 : * unparameterized path, too, if there is one; the users of that list find
238 : * it more convenient if that's included.
239 : *
240 : * This is normally called only after we've finished constructing the path
241 : * list for the rel node.
242 : */
243 : void
244 1710536 : set_cheapest(RelOptInfo *parent_rel)
245 : {
246 : Path *cheapest_startup_path;
247 : Path *cheapest_total_path;
248 : Path *best_param_path;
249 : List *parameterized_paths;
250 : ListCell *p;
251 :
252 : Assert(IsA(parent_rel, RelOptInfo));
253 :
254 1710536 : if (parent_rel->pathlist == NIL)
255 0 : elog(ERROR, "could not devise a query plan for the given query");
256 :
257 1710536 : cheapest_startup_path = cheapest_total_path = best_param_path = NULL;
258 1710536 : parameterized_paths = NIL;
259 :
260 3739094 : foreach(p, parent_rel->pathlist)
261 : {
262 2028558 : Path *path = (Path *) lfirst(p);
263 : int cmp;
264 :
265 2028558 : if (path->param_info)
266 : {
267 : /* Parameterized path, so add it to parameterized_paths */
268 94706 : parameterized_paths = lappend(parameterized_paths, path);
269 :
270 : /*
271 : * If we have an unparameterized cheapest-total, we no longer care
272 : * about finding the best parameterized path, so move on.
273 : */
274 94706 : if (cheapest_total_path)
275 16840 : continue;
276 :
277 : /*
278 : * Otherwise, track the best parameterized path, which is the one
279 : * with least total cost among those of the minimum
280 : * parameterization.
281 : */
282 77866 : if (best_param_path == NULL)
283 72140 : best_param_path = path;
284 : else
285 : {
286 5726 : switch (bms_subset_compare(PATH_REQ_OUTER(path),
287 5726 : PATH_REQ_OUTER(best_param_path)))
288 : {
289 54 : case BMS_EQUAL:
290 : /* keep the cheaper one */
291 54 : if (compare_path_costs(path, best_param_path,
292 : TOTAL_COST) < 0)
293 0 : best_param_path = path;
294 54 : break;
295 328 : case BMS_SUBSET1:
296 : /* new path is less-parameterized */
297 328 : best_param_path = path;
298 328 : break;
299 4 : case BMS_SUBSET2:
300 : /* old path is less-parameterized, keep it */
301 4 : break;
302 5340 : case BMS_DIFFERENT:
303 :
304 : /*
305 : * This means that neither path has the least possible
306 : * parameterization for the rel. We'll sit on the old
307 : * path until something better comes along.
308 : */
309 5340 : break;
310 : }
311 77866 : }
312 : }
313 : else
314 : {
315 : /* Unparameterized path, so consider it for cheapest slots */
316 1933852 : if (cheapest_total_path == NULL)
317 : {
318 1702086 : cheapest_startup_path = cheapest_total_path = path;
319 1702086 : continue;
320 : }
321 :
322 : /*
323 : * If we find two paths of identical costs, try to keep the
324 : * better-sorted one. The paths might have unrelated sort
325 : * orderings, in which case we can only guess which might be
326 : * better to keep, but if one is superior then we definitely
327 : * should keep that one.
328 : */
329 231766 : cmp = compare_path_costs(cheapest_startup_path, path, STARTUP_COST);
330 231766 : if (cmp > 0 ||
331 182 : (cmp == 0 &&
332 182 : compare_pathkeys(cheapest_startup_path->pathkeys,
333 : path->pathkeys) == PATHKEYS_BETTER2))
334 41722 : cheapest_startup_path = path;
335 :
336 231766 : cmp = compare_path_costs(cheapest_total_path, path, TOTAL_COST);
337 231766 : if (cmp > 0 ||
338 0 : (cmp == 0 &&
339 0 : compare_pathkeys(cheapest_total_path->pathkeys,
340 : path->pathkeys) == PATHKEYS_BETTER2))
341 0 : cheapest_total_path = path;
342 : }
343 : }
344 :
345 : /* Add cheapest unparameterized path, if any, to parameterized_paths */
346 1710536 : if (cheapest_total_path)
347 1702086 : parameterized_paths = lcons(cheapest_total_path, parameterized_paths);
348 :
349 : /*
350 : * If there is no unparameterized path, use the best parameterized path as
351 : * cheapest_total_path (but not as cheapest_startup_path).
352 : */
353 1710536 : if (cheapest_total_path == NULL)
354 8450 : cheapest_total_path = best_param_path;
355 : Assert(cheapest_total_path != NULL);
356 :
357 1710536 : parent_rel->cheapest_startup_path = cheapest_startup_path;
358 1710536 : parent_rel->cheapest_total_path = cheapest_total_path;
359 1710536 : parent_rel->cheapest_unique_path = NULL; /* computed only if needed */
360 1710536 : parent_rel->cheapest_parameterized_paths = parameterized_paths;
361 1710536 : }
362 :
363 : /*
364 : * add_path
365 : * Consider a potential implementation path for the specified parent rel,
366 : * and add it to the rel's pathlist if it is worthy of consideration.
367 : * A path is worthy if it has a better sort order (better pathkeys) or
368 : * cheaper cost (on either dimension), or generates fewer rows, than any
369 : * existing path that has the same or superset parameterization rels.
370 : * We also consider parallel-safe paths more worthy than others.
371 : *
372 : * We also remove from the rel's pathlist any old paths that are dominated
373 : * by new_path --- that is, new_path is cheaper, at least as well ordered,
374 : * generates no more rows, requires no outer rels not required by the old
375 : * path, and is no less parallel-safe.
376 : *
377 : * In most cases, a path with a superset parameterization will generate
378 : * fewer rows (since it has more join clauses to apply), so that those two
379 : * figures of merit move in opposite directions; this means that a path of
380 : * one parameterization can seldom dominate a path of another. But such
381 : * cases do arise, so we make the full set of checks anyway.
382 : *
383 : * There are two policy decisions embedded in this function, along with
384 : * its sibling add_path_precheck. First, we treat all parameterized paths
385 : * as having NIL pathkeys, so that they cannot win comparisons on the
386 : * basis of sort order. This is to reduce the number of parameterized
387 : * paths that are kept; see discussion in src/backend/optimizer/README.
388 : *
389 : * Second, we only consider cheap startup cost to be interesting if
390 : * parent_rel->consider_startup is true for an unparameterized path, or
391 : * parent_rel->consider_param_startup is true for a parameterized one.
392 : * Again, this allows discarding useless paths sooner.
393 : *
394 : * The pathlist is kept sorted by total_cost, with cheaper paths
395 : * at the front. Within this routine, that's simply a speed hack:
396 : * doing it that way makes it more likely that we will reject an inferior
397 : * path after a few comparisons, rather than many comparisons.
398 : * However, add_path_precheck relies on this ordering to exit early
399 : * when possible.
400 : *
401 : * NOTE: discarded Path objects are immediately pfree'd to reduce planner
402 : * memory consumption. We dare not try to free the substructure of a Path,
403 : * since much of it may be shared with other Paths or the query tree itself;
404 : * but just recycling discarded Path nodes is a very useful savings in
405 : * a large join tree. We can recycle the List nodes of pathlist, too.
406 : *
407 : * As noted in optimizer/README, deleting a previously-accepted Path is
408 : * safe because we know that Paths of this rel cannot yet be referenced
409 : * from any other rel, such as a higher-level join. However, in some cases
410 : * it is possible that a Path is referenced by another Path for its own
411 : * rel; we must not delete such a Path, even if it is dominated by the new
412 : * Path. Currently this occurs only for IndexPath objects, which may be
413 : * referenced as children of BitmapHeapPaths as well as being paths in
414 : * their own right. Hence, we don't pfree IndexPaths when rejecting them.
415 : *
416 : * 'parent_rel' is the relation entry to which the path corresponds.
417 : * 'new_path' is a potential path for parent_rel.
418 : *
419 : * Returns nothing, but modifies parent_rel->pathlist.
420 : */
421 : void
422 3027318 : add_path(RelOptInfo *parent_rel, Path *new_path)
423 : {
424 3027318 : bool accept_new = true; /* unless we find a superior old path */
425 3027318 : int insert_at = 0; /* where to insert new item */
426 : List *new_path_pathkeys;
427 : ListCell *p1;
428 :
429 : /*
430 : * This is a convenient place to check for query cancel --- no part of the
431 : * planner goes very long without calling add_path().
432 : */
433 3027318 : CHECK_FOR_INTERRUPTS();
434 :
435 : /* Pretend parameterized paths have no pathkeys, per comment above */
436 3027318 : new_path_pathkeys = new_path->param_info ? NIL : new_path->pathkeys;
437 :
438 : /*
439 : * Loop to check proposed new path against old paths. Note it is possible
440 : * for more than one old path to be tossed out because new_path dominates
441 : * it.
442 : */
443 4373950 : foreach(p1, parent_rel->pathlist)
444 : {
445 2363812 : Path *old_path = (Path *) lfirst(p1);
446 2363812 : bool remove_old = false; /* unless new proves superior */
447 : PathCostComparison costcmp;
448 : PathKeysComparison keyscmp;
449 : BMS_Comparison outercmp;
450 :
451 : /*
452 : * Do a fuzzy cost comparison with standard fuzziness limit.
453 : */
454 2363812 : costcmp = compare_path_costs_fuzzily(new_path, old_path,
455 : STD_FUZZ_FACTOR);
456 :
457 : /*
458 : * If the two paths compare differently for startup and total cost,
459 : * then we want to keep both, and we can skip comparing pathkeys and
460 : * required_outer rels. If they compare the same, proceed with the
461 : * other comparisons. Row count is checked last. (We make the tests
462 : * in this order because the cost comparison is most likely to turn
463 : * out "different", and the pathkeys comparison next most likely. As
464 : * explained above, row count very seldom makes a difference, so even
465 : * though it's cheap to compare there's not much point in checking it
466 : * earlier.)
467 : */
468 2363812 : if (costcmp != COSTS_DIFFERENT)
469 : {
470 : /* Similarly check to see if either dominates on pathkeys */
471 : List *old_path_pathkeys;
472 :
473 2330540 : old_path_pathkeys = old_path->param_info ? NIL : old_path->pathkeys;
474 2330540 : keyscmp = compare_pathkeys(new_path_pathkeys,
475 : old_path_pathkeys);
476 2330540 : if (keyscmp != PATHKEYS_DIFFERENT)
477 : {
478 2233712 : switch (costcmp)
479 : {
480 223360 : case COSTS_EQUAL:
481 223360 : outercmp = bms_subset_compare(PATH_REQ_OUTER(new_path),
482 223360 : PATH_REQ_OUTER(old_path));
483 223360 : if (keyscmp == PATHKEYS_BETTER1)
484 : {
485 1864 : if ((outercmp == BMS_EQUAL ||
486 1864 : outercmp == BMS_SUBSET1) &&
487 1864 : new_path->rows <= old_path->rows &&
488 1856 : new_path->parallel_safe >= old_path->parallel_safe)
489 1856 : remove_old = true; /* new dominates old */
490 : }
491 221496 : else if (keyscmp == PATHKEYS_BETTER2)
492 : {
493 9476 : if ((outercmp == BMS_EQUAL ||
494 9476 : outercmp == BMS_SUBSET2) &&
495 9476 : new_path->rows >= old_path->rows &&
496 9350 : new_path->parallel_safe <= old_path->parallel_safe)
497 9350 : accept_new = false; /* old dominates new */
498 : }
499 : else /* keyscmp == PATHKEYS_EQUAL */
500 : {
501 212020 : if (outercmp == BMS_EQUAL)
502 : {
503 : /*
504 : * Same pathkeys and outer rels, and fuzzily
505 : * the same cost, so keep just one; to decide
506 : * which, first check parallel-safety, then
507 : * rows, then do a fuzzy cost comparison with
508 : * very small fuzz limit. (We used to do an
509 : * exact cost comparison, but that results in
510 : * annoying platform-specific plan variations
511 : * due to roundoff in the cost estimates.) If
512 : * things are still tied, arbitrarily keep
513 : * only the old path. Notice that we will
514 : * keep only the old path even if the
515 : * less-fuzzy comparison decides the startup
516 : * and total costs compare differently.
517 : */
518 208960 : if (new_path->parallel_safe >
519 208960 : old_path->parallel_safe)
520 48 : remove_old = true; /* new dominates old */
521 208912 : else if (new_path->parallel_safe <
522 208912 : old_path->parallel_safe)
523 136 : accept_new = false; /* old dominates new */
524 208776 : else if (new_path->rows < old_path->rows)
525 28 : remove_old = true; /* new dominates old */
526 208748 : else if (new_path->rows > old_path->rows)
527 12 : accept_new = false; /* old dominates new */
528 208736 : else if (compare_path_costs_fuzzily(new_path,
529 : old_path,
530 : 1.0000000001) == COSTS_BETTER1)
531 8778 : remove_old = true; /* new dominates old */
532 : else
533 199958 : accept_new = false; /* old equals or
534 : * dominates new */
535 : }
536 3060 : else if (outercmp == BMS_SUBSET1 &&
537 528 : new_path->rows <= old_path->rows &&
538 512 : new_path->parallel_safe >= old_path->parallel_safe)
539 512 : remove_old = true; /* new dominates old */
540 2548 : else if (outercmp == BMS_SUBSET2 &&
541 2342 : new_path->rows >= old_path->rows &&
542 2326 : new_path->parallel_safe <= old_path->parallel_safe)
543 2326 : accept_new = false; /* old dominates new */
544 : /* else different parameterizations, keep both */
545 : }
546 223360 : break;
547 685354 : case COSTS_BETTER1:
548 685354 : if (keyscmp != PATHKEYS_BETTER2)
549 : {
550 488598 : outercmp = bms_subset_compare(PATH_REQ_OUTER(new_path),
551 488598 : PATH_REQ_OUTER(old_path));
552 488598 : if ((outercmp == BMS_EQUAL ||
553 411290 : outercmp == BMS_SUBSET1) &&
554 411290 : new_path->rows <= old_path->rows &&
555 408628 : new_path->parallel_safe >= old_path->parallel_safe)
556 406104 : remove_old = true; /* new dominates old */
557 : }
558 685354 : break;
559 1324998 : case COSTS_BETTER2:
560 1324998 : if (keyscmp != PATHKEYS_BETTER1)
561 : {
562 915102 : outercmp = bms_subset_compare(PATH_REQ_OUTER(new_path),
563 915102 : PATH_REQ_OUTER(old_path));
564 915102 : if ((outercmp == BMS_EQUAL ||
565 848936 : outercmp == BMS_SUBSET2) &&
566 848936 : new_path->rows >= old_path->rows &&
567 807440 : new_path->parallel_safe <= old_path->parallel_safe)
568 805398 : accept_new = false; /* old dominates new */
569 : }
570 1324998 : break;
571 0 : case COSTS_DIFFERENT:
572 :
573 : /*
574 : * can't get here, but keep this case to keep compiler
575 : * quiet
576 : */
577 0 : break;
578 : }
579 130100 : }
580 : }
581 :
582 : /*
583 : * Remove current element from pathlist if dominated by new.
584 : */
585 2363812 : if (remove_old)
586 : {
587 417326 : parent_rel->pathlist = foreach_delete_current(parent_rel->pathlist,
588 : p1);
589 :
590 : /*
591 : * Delete the data pointed-to by the deleted cell, if possible
592 : */
593 417326 : if (!IsA(old_path, IndexPath))
594 397054 : pfree(old_path);
595 : }
596 : else
597 : {
598 : /* new belongs after this old path if it has cost >= old's */
599 1946486 : if (new_path->total_cost >= old_path->total_cost)
600 1607938 : insert_at = foreach_current_index(p1) + 1;
601 : }
602 :
603 : /*
604 : * If we found an old path that dominates new_path, we can quit
605 : * scanning the pathlist; we will not add new_path, and we assume
606 : * new_path cannot dominate any other elements of the pathlist.
607 : */
608 2363812 : if (!accept_new)
609 1017180 : break;
610 : }
611 :
612 3027318 : if (accept_new)
613 : {
614 : /* Accept the new path: insert it at proper place in pathlist */
615 2010138 : parent_rel->pathlist =
616 2010138 : list_insert_nth(parent_rel->pathlist, insert_at, new_path);
617 : }
618 : else
619 : {
620 : /* Reject and recycle the new path */
621 1017180 : if (!IsA(new_path, IndexPath))
622 956912 : pfree(new_path);
623 : }
624 3027318 : }
625 :
626 : /*
627 : * add_path_precheck
628 : * Check whether a proposed new path could possibly get accepted.
629 : * We assume we know the path's pathkeys and parameterization accurately,
630 : * and have lower bounds for its costs.
631 : *
632 : * Note that we do not know the path's rowcount, since getting an estimate for
633 : * that is too expensive to do before prechecking. We assume here that paths
634 : * of a superset parameterization will generate fewer rows; if that holds,
635 : * then paths with different parameterizations cannot dominate each other
636 : * and so we can simply ignore existing paths of another parameterization.
637 : * (In the infrequent cases where that rule of thumb fails, add_path will
638 : * get rid of the inferior path.)
639 : *
640 : * At the time this is called, we haven't actually built a Path structure,
641 : * so the required information has to be passed piecemeal.
642 : */
643 : bool
644 2991262 : add_path_precheck(RelOptInfo *parent_rel,
645 : Cost startup_cost, Cost total_cost,
646 : List *pathkeys, Relids required_outer)
647 : {
648 : List *new_path_pathkeys;
649 : bool consider_startup;
650 : ListCell *p1;
651 :
652 : /* Pretend parameterized paths have no pathkeys, per add_path policy */
653 2991262 : new_path_pathkeys = required_outer ? NIL : pathkeys;
654 :
655 : /* Decide whether new path's startup cost is interesting */
656 2991262 : consider_startup = required_outer ? parent_rel->consider_param_startup : parent_rel->consider_startup;
657 :
658 3716352 : foreach(p1, parent_rel->pathlist)
659 : {
660 3523652 : Path *old_path = (Path *) lfirst(p1);
661 : PathKeysComparison keyscmp;
662 :
663 : /*
664 : * We are looking for an old_path with the same parameterization (and
665 : * by assumption the same rowcount) that dominates the new path on
666 : * pathkeys as well as both cost metrics. If we find one, we can
667 : * reject the new path.
668 : *
669 : * Cost comparisons here should match compare_path_costs_fuzzily.
670 : */
671 3523652 : if (total_cost > old_path->total_cost * STD_FUZZ_FACTOR)
672 : {
673 : /* new path can win on startup cost only if consider_startup */
674 2481340 : if (startup_cost > old_path->startup_cost * STD_FUZZ_FACTOR ||
675 1101146 : !consider_startup)
676 : {
677 : /* new path loses on cost, so check pathkeys... */
678 : List *old_path_pathkeys;
679 :
680 2459402 : old_path_pathkeys = old_path->param_info ? NIL : old_path->pathkeys;
681 2459402 : keyscmp = compare_pathkeys(new_path_pathkeys,
682 : old_path_pathkeys);
683 2459402 : if (keyscmp == PATHKEYS_EQUAL ||
684 : keyscmp == PATHKEYS_BETTER2)
685 : {
686 : /* new path does not win on pathkeys... */
687 1816132 : if (bms_equal(required_outer, PATH_REQ_OUTER(old_path)))
688 : {
689 : /* Found an old path that dominates the new one */
690 1756250 : return false;
691 : }
692 : }
693 : }
694 : }
695 : else
696 : {
697 : /*
698 : * Since the pathlist is sorted by total_cost, we can stop looking
699 : * once we reach a path with a total_cost larger than the new
700 : * path's.
701 : */
702 1042312 : break;
703 : }
704 : }
705 :
706 1235012 : return true;
707 : }
708 :
709 : /*
710 : * add_partial_path
711 : * Like add_path, our goal here is to consider whether a path is worthy
712 : * of being kept around, but the considerations here are a bit different.
713 : * A partial path is one which can be executed in any number of workers in
714 : * parallel such that each worker will generate a subset of the path's
715 : * overall result.
716 : *
717 : * As in add_path, the partial_pathlist is kept sorted with the cheapest
718 : * total path in front. This is depended on by multiple places, which
719 : * just take the front entry as the cheapest path without searching.
720 : *
721 : * We don't generate parameterized partial paths for several reasons. Most
722 : * importantly, they're not safe to execute, because there's nothing to
723 : * make sure that a parallel scan within the parameterized portion of the
724 : * plan is running with the same value in every worker at the same time.
725 : * Fortunately, it seems unlikely to be worthwhile anyway, because having
726 : * each worker scan the entire outer relation and a subset of the inner
727 : * relation will generally be a terrible plan. The inner (parameterized)
728 : * side of the plan will be small anyway. There could be rare cases where
729 : * this wins big - e.g. if join order constraints put a 1-row relation on
730 : * the outer side of the topmost join with a parameterized plan on the inner
731 : * side - but we'll have to be content not to handle such cases until
732 : * somebody builds an executor infrastructure that can cope with them.
733 : *
734 : * Because we don't consider parameterized paths here, we also don't
735 : * need to consider the row counts as a measure of quality: every path will
736 : * produce the same number of rows. Neither do we need to consider startup
737 : * costs: parallelism is only used for plans that will be run to completion.
738 : * Therefore, this routine is much simpler than add_path: it needs to
739 : * consider only pathkeys and total cost.
740 : *
741 : * As with add_path, we pfree paths that are found to be dominated by
742 : * another partial path; this requires that there be no other references to
743 : * such paths yet. Hence, GatherPaths must not be created for a rel until
744 : * we're done creating all partial paths for it. Unlike add_path, we don't
745 : * take an exception for IndexPaths as partial index paths won't be
746 : * referenced by partial BitmapHeapPaths.
747 : */
748 : void
749 88326 : add_partial_path(RelOptInfo *parent_rel, Path *new_path)
750 : {
751 88326 : bool accept_new = true; /* unless we find a superior old path */
752 88326 : int insert_at = 0; /* where to insert new item */
753 : ListCell *p1;
754 :
755 : /* Check for query cancel. */
756 88326 : CHECK_FOR_INTERRUPTS();
757 :
758 : /* Path to be added must be parallel safe. */
759 : Assert(new_path->parallel_safe);
760 :
761 : /* Relation should be OK for parallelism, too. */
762 : Assert(parent_rel->consider_parallel);
763 :
764 : /*
765 : * As in add_path, throw out any paths which are dominated by the new
766 : * path, but throw out the new path if some existing path dominates it.
767 : */
768 119840 : foreach(p1, parent_rel->partial_pathlist)
769 : {
770 45764 : Path *old_path = (Path *) lfirst(p1);
771 45764 : bool remove_old = false; /* unless new proves superior */
772 : PathKeysComparison keyscmp;
773 :
774 : /* Compare pathkeys. */
775 45764 : keyscmp = compare_pathkeys(new_path->pathkeys, old_path->pathkeys);
776 :
777 : /* Unless pathkeys are incompatible, keep just one of the two paths. */
778 45764 : if (keyscmp != PATHKEYS_DIFFERENT)
779 : {
780 45566 : if (new_path->total_cost > old_path->total_cost * STD_FUZZ_FACTOR)
781 : {
782 : /* New path costs more; keep it only if pathkeys are better. */
783 14664 : if (keyscmp != PATHKEYS_BETTER1)
784 6436 : accept_new = false;
785 : }
786 30902 : else if (old_path->total_cost > new_path->total_cost
787 30902 : * STD_FUZZ_FACTOR)
788 : {
789 : /* Old path costs more; keep it only if pathkeys are better. */
790 22746 : if (keyscmp != PATHKEYS_BETTER2)
791 11698 : remove_old = true;
792 : }
793 8156 : else if (keyscmp == PATHKEYS_BETTER1)
794 : {
795 : /* Costs are about the same, new path has better pathkeys. */
796 0 : remove_old = true;
797 : }
798 8156 : else if (keyscmp == PATHKEYS_BETTER2)
799 : {
800 : /* Costs are about the same, old path has better pathkeys. */
801 1662 : accept_new = false;
802 : }
803 6494 : else if (old_path->total_cost > new_path->total_cost * 1.0000000001)
804 : {
805 : /* Pathkeys are the same, and the old path costs more. */
806 342 : remove_old = true;
807 : }
808 : else
809 : {
810 : /*
811 : * Pathkeys are the same, and new path isn't materially
812 : * cheaper.
813 : */
814 6152 : accept_new = false;
815 : }
816 : }
817 :
818 : /*
819 : * Remove current element from partial_pathlist if dominated by new.
820 : */
821 45764 : if (remove_old)
822 : {
823 12040 : parent_rel->partial_pathlist =
824 12040 : foreach_delete_current(parent_rel->partial_pathlist, p1);
825 12040 : pfree(old_path);
826 : }
827 : else
828 : {
829 : /* new belongs after this old path if it has cost >= old's */
830 33724 : if (new_path->total_cost >= old_path->total_cost)
831 22274 : insert_at = foreach_current_index(p1) + 1;
832 : }
833 :
834 : /*
835 : * If we found an old path that dominates new_path, we can quit
836 : * scanning the partial_pathlist; we will not add new_path, and we
837 : * assume new_path cannot dominate any later path.
838 : */
839 45764 : if (!accept_new)
840 14250 : break;
841 : }
842 :
843 88326 : if (accept_new)
844 : {
845 : /* Accept the new path: insert it at proper place */
846 74076 : parent_rel->partial_pathlist =
847 74076 : list_insert_nth(parent_rel->partial_pathlist, insert_at, new_path);
848 : }
849 : else
850 : {
851 : /* Reject and recycle the new path */
852 14250 : pfree(new_path);
853 : }
854 88326 : }
855 :
856 : /*
857 : * add_partial_path_precheck
858 : * Check whether a proposed new partial path could possibly get accepted.
859 : *
860 : * Unlike add_path_precheck, we can ignore startup cost and parameterization,
861 : * since they don't matter for partial paths (see add_partial_path). But
862 : * we do want to make sure we don't add a partial path if there's already
863 : * a complete path that dominates it, since in that case the proposed path
864 : * is surely a loser.
865 : */
866 : bool
867 60798 : add_partial_path_precheck(RelOptInfo *parent_rel, Cost total_cost,
868 : List *pathkeys)
869 : {
870 : ListCell *p1;
871 :
872 : /*
873 : * Our goal here is twofold. First, we want to find out whether this path
874 : * is clearly inferior to some existing partial path. If so, we want to
875 : * reject it immediately. Second, we want to find out whether this path
876 : * is clearly superior to some existing partial path -- at least, modulo
877 : * final cost computations. If so, we definitely want to consider it.
878 : *
879 : * Unlike add_path(), we always compare pathkeys here. This is because we
880 : * expect partial_pathlist to be very short, and getting a definitive
881 : * answer at this stage avoids the need to call add_path_precheck.
882 : */
883 85848 : foreach(p1, parent_rel->partial_pathlist)
884 : {
885 68582 : Path *old_path = (Path *) lfirst(p1);
886 : PathKeysComparison keyscmp;
887 :
888 68582 : keyscmp = compare_pathkeys(pathkeys, old_path->pathkeys);
889 68582 : if (keyscmp != PATHKEYS_DIFFERENT)
890 : {
891 68390 : if (total_cost > old_path->total_cost * STD_FUZZ_FACTOR &&
892 : keyscmp != PATHKEYS_BETTER1)
893 43532 : return false;
894 34816 : if (old_path->total_cost > total_cost * STD_FUZZ_FACTOR &&
895 : keyscmp != PATHKEYS_BETTER2)
896 9958 : return true;
897 : }
898 : }
899 :
900 : /*
901 : * This path is neither clearly inferior to an existing partial path nor
902 : * clearly good enough that it might replace one. Compare it to
903 : * non-parallel plans. If it loses even before accounting for the cost of
904 : * the Gather node, we should definitely reject it.
905 : *
906 : * Note that we pass the total_cost to add_path_precheck twice. This is
907 : * because it's never advantageous to consider the startup cost of a
908 : * partial path; the resulting plans, if run in parallel, will be run to
909 : * completion.
910 : */
911 17266 : if (!add_path_precheck(parent_rel, total_cost, total_cost, pathkeys,
912 : NULL))
913 784 : return false;
914 :
915 16482 : return true;
916 : }
917 :
918 :
919 : /*****************************************************************************
920 : * PATH NODE CREATION ROUTINES
921 : *****************************************************************************/
922 :
923 : /*
924 : * create_seqscan_path
925 : * Creates a path corresponding to a sequential scan, returning the
926 : * pathnode.
927 : */
928 : Path *
929 335958 : create_seqscan_path(PlannerInfo *root, RelOptInfo *rel,
930 : Relids required_outer, int parallel_workers)
931 : {
932 335958 : Path *pathnode = makeNode(Path);
933 :
934 335958 : pathnode->pathtype = T_SeqScan;
935 335958 : pathnode->parent = rel;
936 335958 : pathnode->pathtarget = rel->reltarget;
937 335958 : pathnode->param_info = get_baserel_parampathinfo(root, rel,
938 : required_outer);
939 335958 : pathnode->parallel_aware = (parallel_workers > 0);
940 335958 : pathnode->parallel_safe = rel->consider_parallel;
941 335958 : pathnode->parallel_workers = parallel_workers;
942 335958 : pathnode->pathkeys = NIL; /* seqscan has unordered result */
943 :
944 335958 : cost_seqscan(pathnode, root, rel, pathnode->param_info);
945 :
946 335958 : return pathnode;
947 : }
948 :
949 : /*
950 : * create_samplescan_path
951 : * Creates a path node for a sampled table scan.
952 : */
953 : Path *
954 252 : create_samplescan_path(PlannerInfo *root, RelOptInfo *rel, Relids required_outer)
955 : {
956 252 : Path *pathnode = makeNode(Path);
957 :
958 252 : pathnode->pathtype = T_SampleScan;
959 252 : pathnode->parent = rel;
960 252 : pathnode->pathtarget = rel->reltarget;
961 252 : pathnode->param_info = get_baserel_parampathinfo(root, rel,
962 : required_outer);
963 252 : pathnode->parallel_aware = false;
964 252 : pathnode->parallel_safe = rel->consider_parallel;
965 252 : pathnode->parallel_workers = 0;
966 252 : pathnode->pathkeys = NIL; /* samplescan has unordered result */
967 :
968 252 : cost_samplescan(pathnode, root, rel, pathnode->param_info);
969 :
970 252 : return pathnode;
971 : }
972 :
973 : /*
974 : * create_index_path
975 : * Creates a path node for an index scan.
976 : *
977 : * 'index' is a usable index.
978 : * 'indexclauses' is a list of IndexClause nodes representing clauses
979 : * to be enforced as qual conditions in the scan.
980 : * 'indexorderbys' is a list of bare expressions (no RestrictInfos)
981 : * to be used as index ordering operators in the scan.
982 : * 'indexorderbycols' is an integer list of index column numbers (zero based)
983 : * the ordering operators can be used with.
984 : * 'pathkeys' describes the ordering of the path.
985 : * 'indexscandir' is either ForwardScanDirection or BackwardScanDirection.
986 : * 'indexonly' is true if an index-only scan is wanted.
987 : * 'required_outer' is the set of outer relids for a parameterized path.
988 : * 'loop_count' is the number of repetitions of the indexscan to factor into
989 : * estimates of caching behavior.
990 : * 'partial_path' is true if constructing a parallel index scan path.
991 : *
992 : * Returns the new path node.
993 : */
994 : IndexPath *
995 520860 : create_index_path(PlannerInfo *root,
996 : IndexOptInfo *index,
997 : List *indexclauses,
998 : List *indexorderbys,
999 : List *indexorderbycols,
1000 : List *pathkeys,
1001 : ScanDirection indexscandir,
1002 : bool indexonly,
1003 : Relids required_outer,
1004 : double loop_count,
1005 : bool partial_path)
1006 : {
1007 520860 : IndexPath *pathnode = makeNode(IndexPath);
1008 520860 : RelOptInfo *rel = index->rel;
1009 :
1010 520860 : pathnode->path.pathtype = indexonly ? T_IndexOnlyScan : T_IndexScan;
1011 520860 : pathnode->path.parent = rel;
1012 520860 : pathnode->path.pathtarget = rel->reltarget;
1013 520860 : pathnode->path.param_info = get_baserel_parampathinfo(root, rel,
1014 : required_outer);
1015 520860 : pathnode->path.parallel_aware = false;
1016 520860 : pathnode->path.parallel_safe = rel->consider_parallel;
1017 520860 : pathnode->path.parallel_workers = 0;
1018 520860 : pathnode->path.pathkeys = pathkeys;
1019 :
1020 520860 : pathnode->indexinfo = index;
1021 520860 : pathnode->indexclauses = indexclauses;
1022 520860 : pathnode->indexorderbys = indexorderbys;
1023 520860 : pathnode->indexorderbycols = indexorderbycols;
1024 520860 : pathnode->indexscandir = indexscandir;
1025 :
1026 520860 : cost_index(pathnode, root, loop_count, partial_path);
1027 :
1028 520860 : return pathnode;
1029 : }
1030 :
1031 : /*
1032 : * create_bitmap_heap_path
1033 : * Creates a path node for a bitmap scan.
1034 : *
1035 : * 'bitmapqual' is a tree of IndexPath, BitmapAndPath, and BitmapOrPath nodes.
1036 : * 'required_outer' is the set of outer relids for a parameterized path.
1037 : * 'loop_count' is the number of repetitions of the indexscan to factor into
1038 : * estimates of caching behavior.
1039 : *
1040 : * loop_count should match the value used when creating the component
1041 : * IndexPaths.
1042 : */
1043 : BitmapHeapPath *
1044 248756 : create_bitmap_heap_path(PlannerInfo *root,
1045 : RelOptInfo *rel,
1046 : Path *bitmapqual,
1047 : Relids required_outer,
1048 : double loop_count,
1049 : int parallel_degree)
1050 : {
1051 248756 : BitmapHeapPath *pathnode = makeNode(BitmapHeapPath);
1052 :
1053 248756 : pathnode->path.pathtype = T_BitmapHeapScan;
1054 248756 : pathnode->path.parent = rel;
1055 248756 : pathnode->path.pathtarget = rel->reltarget;
1056 248756 : pathnode->path.param_info = get_baserel_parampathinfo(root, rel,
1057 : required_outer);
1058 248756 : pathnode->path.parallel_aware = (parallel_degree > 0);
1059 248756 : pathnode->path.parallel_safe = rel->consider_parallel;
1060 248756 : pathnode->path.parallel_workers = parallel_degree;
1061 248756 : pathnode->path.pathkeys = NIL; /* always unordered */
1062 :
1063 248756 : pathnode->bitmapqual = bitmapqual;
1064 :
1065 248756 : cost_bitmap_heap_scan(&pathnode->path, root, rel,
1066 : pathnode->path.param_info,
1067 : bitmapqual, loop_count);
1068 :
1069 248756 : return pathnode;
1070 : }
1071 :
1072 : /*
1073 : * create_bitmap_and_path
1074 : * Creates a path node representing a BitmapAnd.
1075 : */
1076 : BitmapAndPath *
1077 24720 : create_bitmap_and_path(PlannerInfo *root,
1078 : RelOptInfo *rel,
1079 : List *bitmapquals)
1080 : {
1081 24720 : BitmapAndPath *pathnode = makeNode(BitmapAndPath);
1082 24720 : Relids required_outer = NULL;
1083 : ListCell *lc;
1084 :
1085 24720 : pathnode->path.pathtype = T_BitmapAnd;
1086 24720 : pathnode->path.parent = rel;
1087 24720 : pathnode->path.pathtarget = rel->reltarget;
1088 :
1089 : /*
1090 : * Identify the required outer rels as the union of what the child paths
1091 : * depend on. (Alternatively, we could insist that the caller pass this
1092 : * in, but it's more convenient and reliable to compute it here.)
1093 : */
1094 74160 : foreach(lc, bitmapquals)
1095 : {
1096 49440 : Path *bitmapqual = (Path *) lfirst(lc);
1097 :
1098 49440 : required_outer = bms_add_members(required_outer,
1099 49440 : PATH_REQ_OUTER(bitmapqual));
1100 : }
1101 24720 : pathnode->path.param_info = get_baserel_parampathinfo(root, rel,
1102 : required_outer);
1103 :
1104 : /*
1105 : * Currently, a BitmapHeapPath, BitmapAndPath, or BitmapOrPath will be
1106 : * parallel-safe if and only if rel->consider_parallel is set. So, we can
1107 : * set the flag for this path based only on the relation-level flag,
1108 : * without actually iterating over the list of children.
1109 : */
1110 24720 : pathnode->path.parallel_aware = false;
1111 24720 : pathnode->path.parallel_safe = rel->consider_parallel;
1112 24720 : pathnode->path.parallel_workers = 0;
1113 :
1114 24720 : pathnode->path.pathkeys = NIL; /* always unordered */
1115 :
1116 24720 : pathnode->bitmapquals = bitmapquals;
1117 :
1118 : /* this sets bitmapselectivity as well as the regular cost fields: */
1119 24720 : cost_bitmap_and_node(pathnode, root);
1120 :
1121 24720 : return pathnode;
1122 : }
1123 :
1124 : /*
1125 : * create_bitmap_or_path
1126 : * Creates a path node representing a BitmapOr.
1127 : */
1128 : BitmapOrPath *
1129 712 : create_bitmap_or_path(PlannerInfo *root,
1130 : RelOptInfo *rel,
1131 : List *bitmapquals)
1132 : {
1133 712 : BitmapOrPath *pathnode = makeNode(BitmapOrPath);
1134 712 : Relids required_outer = NULL;
1135 : ListCell *lc;
1136 :
1137 712 : pathnode->path.pathtype = T_BitmapOr;
1138 712 : pathnode->path.parent = rel;
1139 712 : pathnode->path.pathtarget = rel->reltarget;
1140 :
1141 : /*
1142 : * Identify the required outer rels as the union of what the child paths
1143 : * depend on. (Alternatively, we could insist that the caller pass this
1144 : * in, but it's more convenient and reliable to compute it here.)
1145 : */
1146 2196 : foreach(lc, bitmapquals)
1147 : {
1148 1484 : Path *bitmapqual = (Path *) lfirst(lc);
1149 :
1150 1484 : required_outer = bms_add_members(required_outer,
1151 1484 : PATH_REQ_OUTER(bitmapqual));
1152 : }
1153 712 : pathnode->path.param_info = get_baserel_parampathinfo(root, rel,
1154 : required_outer);
1155 :
1156 : /*
1157 : * Currently, a BitmapHeapPath, BitmapAndPath, or BitmapOrPath will be
1158 : * parallel-safe if and only if rel->consider_parallel is set. So, we can
1159 : * set the flag for this path based only on the relation-level flag,
1160 : * without actually iterating over the list of children.
1161 : */
1162 712 : pathnode->path.parallel_aware = false;
1163 712 : pathnode->path.parallel_safe = rel->consider_parallel;
1164 712 : pathnode->path.parallel_workers = 0;
1165 :
1166 712 : pathnode->path.pathkeys = NIL; /* always unordered */
1167 :
1168 712 : pathnode->bitmapquals = bitmapquals;
1169 :
1170 : /* this sets bitmapselectivity as well as the regular cost fields: */
1171 712 : cost_bitmap_or_node(pathnode, root);
1172 :
1173 712 : return pathnode;
1174 : }
1175 :
1176 : /*
1177 : * create_tidscan_path
1178 : * Creates a path corresponding to a scan by TID, returning the pathnode.
1179 : */
1180 : TidPath *
1181 756 : create_tidscan_path(PlannerInfo *root, RelOptInfo *rel, List *tidquals,
1182 : Relids required_outer)
1183 : {
1184 756 : TidPath *pathnode = makeNode(TidPath);
1185 :
1186 756 : pathnode->path.pathtype = T_TidScan;
1187 756 : pathnode->path.parent = rel;
1188 756 : pathnode->path.pathtarget = rel->reltarget;
1189 756 : pathnode->path.param_info = get_baserel_parampathinfo(root, rel,
1190 : required_outer);
1191 756 : pathnode->path.parallel_aware = false;
1192 756 : pathnode->path.parallel_safe = rel->consider_parallel;
1193 756 : pathnode->path.parallel_workers = 0;
1194 756 : pathnode->path.pathkeys = NIL; /* always unordered */
1195 :
1196 756 : pathnode->tidquals = tidquals;
1197 :
1198 756 : cost_tidscan(&pathnode->path, root, rel, tidquals,
1199 : pathnode->path.param_info);
1200 :
1201 756 : return pathnode;
1202 : }
1203 :
1204 : /*
1205 : * create_tidrangescan_path
1206 : * Creates a path corresponding to a scan by a range of TIDs, returning
1207 : * the pathnode.
1208 : */
1209 : TidRangePath *
1210 202 : create_tidrangescan_path(PlannerInfo *root, RelOptInfo *rel,
1211 : List *tidrangequals, Relids required_outer)
1212 : {
1213 202 : TidRangePath *pathnode = makeNode(TidRangePath);
1214 :
1215 202 : pathnode->path.pathtype = T_TidRangeScan;
1216 202 : pathnode->path.parent = rel;
1217 202 : pathnode->path.pathtarget = rel->reltarget;
1218 202 : pathnode->path.param_info = get_baserel_parampathinfo(root, rel,
1219 : required_outer);
1220 202 : pathnode->path.parallel_aware = false;
1221 202 : pathnode->path.parallel_safe = rel->consider_parallel;
1222 202 : pathnode->path.parallel_workers = 0;
1223 202 : pathnode->path.pathkeys = NIL; /* always unordered */
1224 :
1225 202 : pathnode->tidrangequals = tidrangequals;
1226 :
1227 202 : cost_tidrangescan(&pathnode->path, root, rel, tidrangequals,
1228 : pathnode->path.param_info);
1229 :
1230 202 : return pathnode;
1231 : }
1232 :
1233 : /*
1234 : * create_append_path
1235 : * Creates a path corresponding to an Append plan, returning the
1236 : * pathnode.
1237 : *
1238 : * Note that we must handle subpaths = NIL, representing a dummy access path.
1239 : * Also, there are callers that pass root = NULL.
1240 : */
1241 : AppendPath *
1242 62696 : create_append_path(PlannerInfo *root,
1243 : RelOptInfo *rel,
1244 : List *subpaths, List *partial_subpaths,
1245 : List *pathkeys, Relids required_outer,
1246 : int parallel_workers, bool parallel_aware,
1247 : double rows)
1248 : {
1249 62696 : AppendPath *pathnode = makeNode(AppendPath);
1250 : ListCell *l;
1251 :
1252 : Assert(!parallel_aware || parallel_workers > 0);
1253 :
1254 62696 : pathnode->path.pathtype = T_Append;
1255 62696 : pathnode->path.parent = rel;
1256 62696 : pathnode->path.pathtarget = rel->reltarget;
1257 :
1258 : /*
1259 : * If this is for a baserel (not a join or non-leaf partition), we prefer
1260 : * to apply get_baserel_parampathinfo to construct a full ParamPathInfo
1261 : * for the path. This supports building a Memoize path atop this path,
1262 : * and if this is a partitioned table the info may be useful for run-time
1263 : * pruning (cf make_partition_pruneinfo()).
1264 : *
1265 : * However, if we don't have "root" then that won't work and we fall back
1266 : * on the simpler get_appendrel_parampathinfo. There's no point in doing
1267 : * the more expensive thing for a dummy path, either.
1268 : */
1269 62696 : if (rel->reloptkind == RELOPT_BASEREL && root && subpaths != NIL)
1270 27670 : pathnode->path.param_info = get_baserel_parampathinfo(root,
1271 : rel,
1272 : required_outer);
1273 : else
1274 35026 : pathnode->path.param_info = get_appendrel_parampathinfo(rel,
1275 : required_outer);
1276 :
1277 62696 : pathnode->path.parallel_aware = parallel_aware;
1278 62696 : pathnode->path.parallel_safe = rel->consider_parallel;
1279 62696 : pathnode->path.parallel_workers = parallel_workers;
1280 62696 : pathnode->path.pathkeys = pathkeys;
1281 :
1282 : /*
1283 : * For parallel append, non-partial paths are sorted by descending total
1284 : * costs. That way, the total time to finish all non-partial paths is
1285 : * minimized. Also, the partial paths are sorted by descending startup
1286 : * costs. There may be some paths that require to do startup work by a
1287 : * single worker. In such case, it's better for workers to choose the
1288 : * expensive ones first, whereas the leader should choose the cheapest
1289 : * startup plan.
1290 : */
1291 62696 : if (pathnode->path.parallel_aware)
1292 : {
1293 : /*
1294 : * We mustn't fiddle with the order of subpaths when the Append has
1295 : * pathkeys. The order they're listed in is critical to keeping the
1296 : * pathkeys valid.
1297 : */
1298 : Assert(pathkeys == NIL);
1299 :
1300 20890 : list_sort(subpaths, append_total_cost_compare);
1301 20890 : list_sort(partial_subpaths, append_startup_cost_compare);
1302 : }
1303 62696 : pathnode->first_partial_path = list_length(subpaths);
1304 62696 : pathnode->subpaths = list_concat(subpaths, partial_subpaths);
1305 :
1306 : /*
1307 : * Apply query-wide LIMIT if known and path is for sole base relation.
1308 : * (Handling this at this low level is a bit klugy.)
1309 : */
1310 62696 : if (root != NULL && bms_equal(rel->relids, root->all_query_rels))
1311 34520 : pathnode->limit_tuples = root->limit_tuples;
1312 : else
1313 28176 : pathnode->limit_tuples = -1.0;
1314 :
1315 207368 : foreach(l, pathnode->subpaths)
1316 : {
1317 144672 : Path *subpath = (Path *) lfirst(l);
1318 :
1319 256480 : pathnode->path.parallel_safe = pathnode->path.parallel_safe &&
1320 111808 : subpath->parallel_safe;
1321 :
1322 : /* All child paths must have same parameterization */
1323 : Assert(bms_equal(PATH_REQ_OUTER(subpath), required_outer));
1324 : }
1325 :
1326 : Assert(!parallel_aware || pathnode->path.parallel_safe);
1327 :
1328 : /*
1329 : * If there's exactly one child path then the output of the Append is
1330 : * necessarily ordered the same as the child's, so we can inherit the
1331 : * child's pathkeys if any, overriding whatever the caller might've said.
1332 : * Furthermore, if the child's parallel awareness matches the Append's,
1333 : * then the Append is a no-op and will be discarded later (in setrefs.c).
1334 : * Then we can inherit the child's size and cost too, effectively charging
1335 : * zero for the Append. Otherwise, we must do the normal costsize
1336 : * calculation.
1337 : */
1338 62696 : if (list_length(pathnode->subpaths) == 1)
1339 : {
1340 20086 : Path *child = (Path *) linitial(pathnode->subpaths);
1341 :
1342 20086 : if (child->parallel_aware == parallel_aware)
1343 : {
1344 19876 : pathnode->path.rows = child->rows;
1345 19876 : pathnode->path.startup_cost = child->startup_cost;
1346 19876 : pathnode->path.total_cost = child->total_cost;
1347 : }
1348 : else
1349 210 : cost_append(pathnode);
1350 : /* Must do this last, else cost_append complains */
1351 20086 : pathnode->path.pathkeys = child->pathkeys;
1352 : }
1353 : else
1354 42610 : cost_append(pathnode);
1355 :
1356 : /* If the caller provided a row estimate, override the computed value. */
1357 62696 : if (rows >= 0)
1358 582 : pathnode->path.rows = rows;
1359 :
1360 62696 : return pathnode;
1361 : }
1362 :
1363 : /*
1364 : * append_total_cost_compare
1365 : * list_sort comparator for sorting append child paths
1366 : * by total_cost descending
1367 : *
1368 : * For equal total costs, we fall back to comparing startup costs; if those
1369 : * are equal too, break ties using bms_compare on the paths' relids.
1370 : * (This is to avoid getting unpredictable results from list_sort.)
1371 : */
1372 : static int
1373 1480 : append_total_cost_compare(const ListCell *a, const ListCell *b)
1374 : {
1375 1480 : Path *path1 = (Path *) lfirst(a);
1376 1480 : Path *path2 = (Path *) lfirst(b);
1377 : int cmp;
1378 :
1379 1480 : cmp = compare_path_costs(path1, path2, TOTAL_COST);
1380 1480 : if (cmp != 0)
1381 1210 : return -cmp;
1382 270 : return bms_compare(path1->parent->relids, path2->parent->relids);
1383 : }
1384 :
1385 : /*
1386 : * append_startup_cost_compare
1387 : * list_sort comparator for sorting append child paths
1388 : * by startup_cost descending
1389 : *
1390 : * For equal startup costs, we fall back to comparing total costs; if those
1391 : * are equal too, break ties using bms_compare on the paths' relids.
1392 : * (This is to avoid getting unpredictable results from list_sort.)
1393 : */
1394 : static int
1395 31914 : append_startup_cost_compare(const ListCell *a, const ListCell *b)
1396 : {
1397 31914 : Path *path1 = (Path *) lfirst(a);
1398 31914 : Path *path2 = (Path *) lfirst(b);
1399 : int cmp;
1400 :
1401 31914 : cmp = compare_path_costs(path1, path2, STARTUP_COST);
1402 31914 : if (cmp != 0)
1403 12360 : return -cmp;
1404 19554 : return bms_compare(path1->parent->relids, path2->parent->relids);
1405 : }
1406 :
1407 : /*
1408 : * create_merge_append_path
1409 : * Creates a path corresponding to a MergeAppend plan, returning the
1410 : * pathnode.
1411 : */
1412 : MergeAppendPath *
1413 3680 : create_merge_append_path(PlannerInfo *root,
1414 : RelOptInfo *rel,
1415 : List *subpaths,
1416 : List *pathkeys,
1417 : Relids required_outer)
1418 : {
1419 3680 : MergeAppendPath *pathnode = makeNode(MergeAppendPath);
1420 : Cost input_startup_cost;
1421 : Cost input_total_cost;
1422 : ListCell *l;
1423 :
1424 3680 : pathnode->path.pathtype = T_MergeAppend;
1425 3680 : pathnode->path.parent = rel;
1426 3680 : pathnode->path.pathtarget = rel->reltarget;
1427 3680 : pathnode->path.param_info = get_appendrel_parampathinfo(rel,
1428 : required_outer);
1429 3680 : pathnode->path.parallel_aware = false;
1430 3680 : pathnode->path.parallel_safe = rel->consider_parallel;
1431 3680 : pathnode->path.parallel_workers = 0;
1432 3680 : pathnode->path.pathkeys = pathkeys;
1433 3680 : pathnode->subpaths = subpaths;
1434 :
1435 : /*
1436 : * Apply query-wide LIMIT if known and path is for sole base relation.
1437 : * (Handling this at this low level is a bit klugy.)
1438 : */
1439 3680 : if (bms_equal(rel->relids, root->all_query_rels))
1440 2018 : pathnode->limit_tuples = root->limit_tuples;
1441 : else
1442 1662 : pathnode->limit_tuples = -1.0;
1443 :
1444 : /*
1445 : * Add up the sizes and costs of the input paths.
1446 : */
1447 3680 : pathnode->path.rows = 0;
1448 3680 : input_startup_cost = 0;
1449 3680 : input_total_cost = 0;
1450 13902 : foreach(l, subpaths)
1451 : {
1452 10222 : Path *subpath = (Path *) lfirst(l);
1453 :
1454 10222 : pathnode->path.rows += subpath->rows;
1455 18408 : pathnode->path.parallel_safe = pathnode->path.parallel_safe &&
1456 8186 : subpath->parallel_safe;
1457 :
1458 10222 : if (pathkeys_contained_in(pathkeys, subpath->pathkeys))
1459 : {
1460 : /* Subpath is adequately ordered, we won't need to sort it */
1461 9942 : input_startup_cost += subpath->startup_cost;
1462 9942 : input_total_cost += subpath->total_cost;
1463 : }
1464 : else
1465 : {
1466 : /* We'll need to insert a Sort node, so include cost for that */
1467 : Path sort_path; /* dummy for result of cost_sort */
1468 :
1469 280 : cost_sort(&sort_path,
1470 : root,
1471 : pathkeys,
1472 : subpath->total_cost,
1473 280 : subpath->parent->tuples,
1474 280 : subpath->pathtarget->width,
1475 : 0.0,
1476 : work_mem,
1477 : pathnode->limit_tuples);
1478 280 : input_startup_cost += sort_path.startup_cost;
1479 280 : input_total_cost += sort_path.total_cost;
1480 : }
1481 :
1482 : /* All child paths must have same parameterization */
1483 : Assert(bms_equal(PATH_REQ_OUTER(subpath), required_outer));
1484 : }
1485 :
1486 : /*
1487 : * Now we can compute total costs of the MergeAppend. If there's exactly
1488 : * one child path and its parallel awareness matches that of the
1489 : * MergeAppend, then the MergeAppend is a no-op and will be discarded
1490 : * later (in setrefs.c); otherwise we do the normal cost calculation.
1491 : */
1492 3680 : if (list_length(subpaths) == 1 &&
1493 110 : ((Path *) linitial(subpaths))->parallel_aware ==
1494 110 : pathnode->path.parallel_aware)
1495 : {
1496 110 : pathnode->path.startup_cost = input_startup_cost;
1497 110 : pathnode->path.total_cost = input_total_cost;
1498 : }
1499 : else
1500 3570 : cost_merge_append(&pathnode->path, root,
1501 : pathkeys, list_length(subpaths),
1502 : input_startup_cost, input_total_cost,
1503 : pathnode->path.rows);
1504 :
1505 3680 : return pathnode;
1506 : }
1507 :
1508 : /*
1509 : * create_group_result_path
1510 : * Creates a path representing a Result-and-nothing-else plan.
1511 : *
1512 : * This is only used for degenerate grouping cases, in which we know we
1513 : * need to produce one result row, possibly filtered by a HAVING qual.
1514 : */
1515 : GroupResultPath *
1516 193730 : create_group_result_path(PlannerInfo *root, RelOptInfo *rel,
1517 : PathTarget *target, List *havingqual)
1518 : {
1519 193730 : GroupResultPath *pathnode = makeNode(GroupResultPath);
1520 :
1521 193730 : pathnode->path.pathtype = T_Result;
1522 193730 : pathnode->path.parent = rel;
1523 193730 : pathnode->path.pathtarget = target;
1524 193730 : pathnode->path.param_info = NULL; /* there are no other rels... */
1525 193730 : pathnode->path.parallel_aware = false;
1526 193730 : pathnode->path.parallel_safe = rel->consider_parallel;
1527 193730 : pathnode->path.parallel_workers = 0;
1528 193730 : pathnode->path.pathkeys = NIL;
1529 193730 : pathnode->quals = havingqual;
1530 :
1531 : /*
1532 : * We can't quite use cost_resultscan() because the quals we want to
1533 : * account for are not baserestrict quals of the rel. Might as well just
1534 : * hack it here.
1535 : */
1536 193730 : pathnode->path.rows = 1;
1537 193730 : pathnode->path.startup_cost = target->cost.startup;
1538 193730 : pathnode->path.total_cost = target->cost.startup +
1539 193730 : cpu_tuple_cost + target->cost.per_tuple;
1540 :
1541 : /*
1542 : * Add cost of qual, if any --- but we ignore its selectivity, since our
1543 : * rowcount estimate should be 1 no matter what the qual is.
1544 : */
1545 193730 : if (havingqual)
1546 : {
1547 : QualCost qual_cost;
1548 :
1549 486 : cost_qual_eval(&qual_cost, havingqual, root);
1550 : /* havingqual is evaluated once at startup */
1551 486 : pathnode->path.startup_cost += qual_cost.startup + qual_cost.per_tuple;
1552 486 : pathnode->path.total_cost += qual_cost.startup + qual_cost.per_tuple;
1553 : }
1554 :
1555 193730 : return pathnode;
1556 : }
1557 :
1558 : /*
1559 : * create_material_path
1560 : * Creates a path corresponding to a Material plan, returning the
1561 : * pathnode.
1562 : */
1563 : MaterialPath *
1564 343604 : create_material_path(RelOptInfo *rel, Path *subpath)
1565 : {
1566 343604 : MaterialPath *pathnode = makeNode(MaterialPath);
1567 :
1568 : Assert(subpath->parent == rel);
1569 :
1570 343604 : pathnode->path.pathtype = T_Material;
1571 343604 : pathnode->path.parent = rel;
1572 343604 : pathnode->path.pathtarget = rel->reltarget;
1573 343604 : pathnode->path.param_info = subpath->param_info;
1574 343604 : pathnode->path.parallel_aware = false;
1575 639156 : pathnode->path.parallel_safe = rel->consider_parallel &&
1576 295552 : subpath->parallel_safe;
1577 343604 : pathnode->path.parallel_workers = subpath->parallel_workers;
1578 343604 : pathnode->path.pathkeys = subpath->pathkeys;
1579 :
1580 343604 : pathnode->subpath = subpath;
1581 :
1582 343604 : cost_material(&pathnode->path,
1583 : subpath->startup_cost,
1584 : subpath->total_cost,
1585 : subpath->rows,
1586 343604 : subpath->pathtarget->width);
1587 :
1588 343604 : return pathnode;
1589 : }
1590 :
1591 : /*
1592 : * create_memoize_path
1593 : * Creates a path corresponding to a Memoize plan, returning the pathnode.
1594 : */
1595 : MemoizePath *
1596 181632 : create_memoize_path(PlannerInfo *root, RelOptInfo *rel, Path *subpath,
1597 : List *param_exprs, List *hash_operators,
1598 : bool singlerow, bool binary_mode, double calls)
1599 : {
1600 181632 : MemoizePath *pathnode = makeNode(MemoizePath);
1601 :
1602 : Assert(subpath->parent == rel);
1603 :
1604 181632 : pathnode->path.pathtype = T_Memoize;
1605 181632 : pathnode->path.parent = rel;
1606 181632 : pathnode->path.pathtarget = rel->reltarget;
1607 181632 : pathnode->path.param_info = subpath->param_info;
1608 181632 : pathnode->path.parallel_aware = false;
1609 350284 : pathnode->path.parallel_safe = rel->consider_parallel &&
1610 168652 : subpath->parallel_safe;
1611 181632 : pathnode->path.parallel_workers = subpath->parallel_workers;
1612 181632 : pathnode->path.pathkeys = subpath->pathkeys;
1613 :
1614 181632 : pathnode->subpath = subpath;
1615 181632 : pathnode->hash_operators = hash_operators;
1616 181632 : pathnode->param_exprs = param_exprs;
1617 181632 : pathnode->singlerow = singlerow;
1618 181632 : pathnode->binary_mode = binary_mode;
1619 181632 : pathnode->calls = calls;
1620 :
1621 : /*
1622 : * For now we set est_entries to 0. cost_memoize_rescan() does all the
1623 : * hard work to determine how many cache entries there are likely to be,
1624 : * so it seems best to leave it up to that function to fill this field in.
1625 : * If left at 0, the executor will make a guess at a good value.
1626 : */
1627 181632 : pathnode->est_entries = 0;
1628 :
1629 : /*
1630 : * Add a small additional charge for caching the first entry. All the
1631 : * harder calculations for rescans are performed in cost_memoize_rescan().
1632 : */
1633 181632 : pathnode->path.startup_cost = subpath->startup_cost + cpu_tuple_cost;
1634 181632 : pathnode->path.total_cost = subpath->total_cost + cpu_tuple_cost;
1635 181632 : pathnode->path.rows = subpath->rows;
1636 :
1637 181632 : return pathnode;
1638 : }
1639 :
1640 : /*
1641 : * create_unique_path
1642 : * Creates a path representing elimination of distinct rows from the
1643 : * input data. Distinct-ness is defined according to the needs of the
1644 : * semijoin represented by sjinfo. If it is not possible to identify
1645 : * how to make the data unique, NULL is returned.
1646 : *
1647 : * If used at all, this is likely to be called repeatedly on the same rel;
1648 : * and the input subpath should always be the same (the cheapest_total path
1649 : * for the rel). So we cache the result.
1650 : */
1651 : UniquePath *
1652 21244 : create_unique_path(PlannerInfo *root, RelOptInfo *rel, Path *subpath,
1653 : SpecialJoinInfo *sjinfo)
1654 : {
1655 : UniquePath *pathnode;
1656 : Path sort_path; /* dummy for result of cost_sort */
1657 : Path agg_path; /* dummy for result of cost_agg */
1658 : MemoryContext oldcontext;
1659 : int numCols;
1660 :
1661 : /* Caller made a mistake if subpath isn't cheapest_total ... */
1662 : Assert(subpath == rel->cheapest_total_path);
1663 : Assert(subpath->parent == rel);
1664 : /* ... or if SpecialJoinInfo is the wrong one */
1665 : Assert(sjinfo->jointype == JOIN_SEMI);
1666 : Assert(bms_equal(rel->relids, sjinfo->syn_righthand));
1667 :
1668 : /* If result already cached, return it */
1669 21244 : if (rel->cheapest_unique_path)
1670 18108 : return (UniquePath *) rel->cheapest_unique_path;
1671 :
1672 : /* If it's not possible to unique-ify, return NULL */
1673 3136 : if (!(sjinfo->semi_can_btree || sjinfo->semi_can_hash))
1674 102 : return NULL;
1675 :
1676 : /*
1677 : * When called during GEQO join planning, we are in a short-lived memory
1678 : * context. We must make sure that the path and any subsidiary data
1679 : * structures created for a baserel survive the GEQO cycle, else the
1680 : * baserel is trashed for future GEQO cycles. On the other hand, when we
1681 : * are creating those for a joinrel during GEQO, we don't want them to
1682 : * clutter the main planning context. Upshot is that the best solution is
1683 : * to explicitly allocate memory in the same context the given RelOptInfo
1684 : * is in.
1685 : */
1686 3034 : oldcontext = MemoryContextSwitchTo(GetMemoryChunkContext(rel));
1687 :
1688 3034 : pathnode = makeNode(UniquePath);
1689 :
1690 3034 : pathnode->path.pathtype = T_Unique;
1691 3034 : pathnode->path.parent = rel;
1692 3034 : pathnode->path.pathtarget = rel->reltarget;
1693 3034 : pathnode->path.param_info = subpath->param_info;
1694 3034 : pathnode->path.parallel_aware = false;
1695 5800 : pathnode->path.parallel_safe = rel->consider_parallel &&
1696 2766 : subpath->parallel_safe;
1697 3034 : pathnode->path.parallel_workers = subpath->parallel_workers;
1698 :
1699 : /*
1700 : * Assume the output is unsorted, since we don't necessarily have pathkeys
1701 : * to represent it. (This might get overridden below.)
1702 : */
1703 3034 : pathnode->path.pathkeys = NIL;
1704 :
1705 3034 : pathnode->subpath = subpath;
1706 3034 : pathnode->in_operators = sjinfo->semi_operators;
1707 3034 : pathnode->uniq_exprs = sjinfo->semi_rhs_exprs;
1708 :
1709 : /*
1710 : * If the input is a relation and it has a unique index that proves the
1711 : * semi_rhs_exprs are unique, then we don't need to do anything. Note
1712 : * that relation_has_unique_index_for automatically considers restriction
1713 : * clauses for the rel, as well.
1714 : */
1715 3702 : if (rel->rtekind == RTE_RELATION && sjinfo->semi_can_btree &&
1716 668 : relation_has_unique_index_for(root, rel, NIL,
1717 : sjinfo->semi_rhs_exprs,
1718 : sjinfo->semi_operators))
1719 : {
1720 0 : pathnode->umethod = UNIQUE_PATH_NOOP;
1721 0 : pathnode->path.rows = rel->rows;
1722 0 : pathnode->path.startup_cost = subpath->startup_cost;
1723 0 : pathnode->path.total_cost = subpath->total_cost;
1724 0 : pathnode->path.pathkeys = subpath->pathkeys;
1725 :
1726 0 : rel->cheapest_unique_path = (Path *) pathnode;
1727 :
1728 0 : MemoryContextSwitchTo(oldcontext);
1729 :
1730 0 : return pathnode;
1731 : }
1732 :
1733 : /*
1734 : * If the input is a subquery whose output must be unique already, then we
1735 : * don't need to do anything. The test for uniqueness has to consider
1736 : * exactly which columns we are extracting; for example "SELECT DISTINCT
1737 : * x,y" doesn't guarantee that x alone is distinct. So we cannot check for
1738 : * this optimization unless semi_rhs_exprs consists only of simple Vars
1739 : * referencing subquery outputs. (Possibly we could do something with
1740 : * expressions in the subquery outputs, too, but for now keep it simple.)
1741 : */
1742 3034 : if (rel->rtekind == RTE_SUBQUERY)
1743 : {
1744 512 : RangeTblEntry *rte = planner_rt_fetch(rel->relid, root);
1745 :
1746 512 : if (query_supports_distinctness(rte->subquery))
1747 : {
1748 : List *sub_tlist_colnos;
1749 :
1750 476 : sub_tlist_colnos = translate_sub_tlist(sjinfo->semi_rhs_exprs,
1751 476 : rel->relid);
1752 :
1753 538 : if (sub_tlist_colnos &&
1754 62 : query_is_distinct_for(rte->subquery,
1755 : sub_tlist_colnos,
1756 : sjinfo->semi_operators))
1757 : {
1758 0 : pathnode->umethod = UNIQUE_PATH_NOOP;
1759 0 : pathnode->path.rows = rel->rows;
1760 0 : pathnode->path.startup_cost = subpath->startup_cost;
1761 0 : pathnode->path.total_cost = subpath->total_cost;
1762 0 : pathnode->path.pathkeys = subpath->pathkeys;
1763 :
1764 0 : rel->cheapest_unique_path = (Path *) pathnode;
1765 :
1766 0 : MemoryContextSwitchTo(oldcontext);
1767 :
1768 0 : return pathnode;
1769 : }
1770 : }
1771 : }
1772 :
1773 : /* Estimate number of output rows */
1774 3034 : pathnode->path.rows = estimate_num_groups(root,
1775 : sjinfo->semi_rhs_exprs,
1776 : rel->rows,
1777 : NULL,
1778 : NULL);
1779 3034 : numCols = list_length(sjinfo->semi_rhs_exprs);
1780 :
1781 3034 : if (sjinfo->semi_can_btree)
1782 : {
1783 : /*
1784 : * Estimate cost for sort+unique implementation
1785 : */
1786 3034 : cost_sort(&sort_path, root, NIL,
1787 : subpath->total_cost,
1788 : rel->rows,
1789 3034 : subpath->pathtarget->width,
1790 : 0.0,
1791 : work_mem,
1792 : -1.0);
1793 :
1794 : /*
1795 : * Charge one cpu_operator_cost per comparison per input tuple. We
1796 : * assume all columns get compared at most of the tuples. (XXX
1797 : * probably this is an overestimate.) This should agree with
1798 : * create_upper_unique_path.
1799 : */
1800 3034 : sort_path.total_cost += cpu_operator_cost * rel->rows * numCols;
1801 : }
1802 :
1803 3034 : if (sjinfo->semi_can_hash)
1804 : {
1805 : /*
1806 : * Estimate the overhead per hashtable entry at 64 bytes (same as in
1807 : * planner.c).
1808 : */
1809 3034 : int hashentrysize = subpath->pathtarget->width + 64;
1810 :
1811 3034 : if (hashentrysize * pathnode->path.rows > get_hash_memory_limit())
1812 : {
1813 : /*
1814 : * We should not try to hash. Hack the SpecialJoinInfo to
1815 : * remember this, in case we come through here again.
1816 : */
1817 0 : sjinfo->semi_can_hash = false;
1818 : }
1819 : else
1820 3034 : cost_agg(&agg_path, root,
1821 : AGG_HASHED, NULL,
1822 : numCols, pathnode->path.rows,
1823 : NIL,
1824 : subpath->startup_cost,
1825 : subpath->total_cost,
1826 : rel->rows,
1827 3034 : subpath->pathtarget->width);
1828 : }
1829 :
1830 3034 : if (sjinfo->semi_can_btree && sjinfo->semi_can_hash)
1831 : {
1832 3034 : if (agg_path.total_cost < sort_path.total_cost)
1833 2958 : pathnode->umethod = UNIQUE_PATH_HASH;
1834 : else
1835 76 : pathnode->umethod = UNIQUE_PATH_SORT;
1836 : }
1837 0 : else if (sjinfo->semi_can_btree)
1838 0 : pathnode->umethod = UNIQUE_PATH_SORT;
1839 0 : else if (sjinfo->semi_can_hash)
1840 0 : pathnode->umethod = UNIQUE_PATH_HASH;
1841 : else
1842 : {
1843 : /* we can get here only if we abandoned hashing above */
1844 0 : MemoryContextSwitchTo(oldcontext);
1845 0 : return NULL;
1846 : }
1847 :
1848 3034 : if (pathnode->umethod == UNIQUE_PATH_HASH)
1849 : {
1850 2958 : pathnode->path.startup_cost = agg_path.startup_cost;
1851 2958 : pathnode->path.total_cost = agg_path.total_cost;
1852 : }
1853 : else
1854 : {
1855 76 : pathnode->path.startup_cost = sort_path.startup_cost;
1856 76 : pathnode->path.total_cost = sort_path.total_cost;
1857 : }
1858 :
1859 3034 : rel->cheapest_unique_path = (Path *) pathnode;
1860 :
1861 3034 : MemoryContextSwitchTo(oldcontext);
1862 :
1863 3034 : return pathnode;
1864 : }
1865 :
1866 : /*
1867 : * create_gather_merge_path
1868 : *
1869 : * Creates a path corresponding to a gather merge scan, returning
1870 : * the pathnode.
1871 : */
1872 : GatherMergePath *
1873 9406 : create_gather_merge_path(PlannerInfo *root, RelOptInfo *rel, Path *subpath,
1874 : PathTarget *target, List *pathkeys,
1875 : Relids required_outer, double *rows)
1876 : {
1877 9406 : GatherMergePath *pathnode = makeNode(GatherMergePath);
1878 9406 : Cost input_startup_cost = 0;
1879 9406 : Cost input_total_cost = 0;
1880 :
1881 : Assert(subpath->parallel_safe);
1882 : Assert(pathkeys);
1883 :
1884 9406 : pathnode->path.pathtype = T_GatherMerge;
1885 9406 : pathnode->path.parent = rel;
1886 9406 : pathnode->path.param_info = get_baserel_parampathinfo(root, rel,
1887 : required_outer);
1888 9406 : pathnode->path.parallel_aware = false;
1889 :
1890 9406 : pathnode->subpath = subpath;
1891 9406 : pathnode->num_workers = subpath->parallel_workers;
1892 9406 : pathnode->path.pathkeys = pathkeys;
1893 9406 : pathnode->path.pathtarget = target ? target : rel->reltarget;
1894 9406 : pathnode->path.rows += subpath->rows;
1895 :
1896 9406 : if (pathkeys_contained_in(pathkeys, subpath->pathkeys))
1897 : {
1898 : /* Subpath is adequately ordered, we won't need to sort it */
1899 9406 : input_startup_cost += subpath->startup_cost;
1900 9406 : input_total_cost += subpath->total_cost;
1901 : }
1902 : else
1903 : {
1904 : /* We'll need to insert a Sort node, so include cost for that */
1905 : Path sort_path; /* dummy for result of cost_sort */
1906 :
1907 0 : cost_sort(&sort_path,
1908 : root,
1909 : pathkeys,
1910 : subpath->total_cost,
1911 : subpath->rows,
1912 0 : subpath->pathtarget->width,
1913 : 0.0,
1914 : work_mem,
1915 : -1);
1916 0 : input_startup_cost += sort_path.startup_cost;
1917 0 : input_total_cost += sort_path.total_cost;
1918 : }
1919 :
1920 9406 : cost_gather_merge(pathnode, root, rel, pathnode->path.param_info,
1921 : input_startup_cost, input_total_cost, rows);
1922 :
1923 9406 : return pathnode;
1924 : }
1925 :
1926 : /*
1927 : * translate_sub_tlist - get subquery column numbers represented by tlist
1928 : *
1929 : * The given targetlist usually contains only Vars referencing the given relid.
1930 : * Extract their varattnos (ie, the column numbers of the subquery) and return
1931 : * as an integer List.
1932 : *
1933 : * If any of the tlist items is not a simple Var, we cannot determine whether
1934 : * the subquery's uniqueness condition (if any) matches ours, so punt and
1935 : * return NIL.
1936 : */
1937 : static List *
1938 476 : translate_sub_tlist(List *tlist, int relid)
1939 : {
1940 476 : List *result = NIL;
1941 : ListCell *l;
1942 :
1943 538 : foreach(l, tlist)
1944 : {
1945 476 : Var *var = (Var *) lfirst(l);
1946 :
1947 476 : if (!var || !IsA(var, Var) ||
1948 62 : var->varno != relid)
1949 414 : return NIL; /* punt */
1950 :
1951 62 : result = lappend_int(result, var->varattno);
1952 : }
1953 62 : return result;
1954 : }
1955 :
1956 : /*
1957 : * create_gather_path
1958 : * Creates a path corresponding to a gather scan, returning the
1959 : * pathnode.
1960 : *
1961 : * 'rows' may optionally be set to override row estimates from other sources.
1962 : */
1963 : GatherPath *
1964 14640 : create_gather_path(PlannerInfo *root, RelOptInfo *rel, Path *subpath,
1965 : PathTarget *target, Relids required_outer, double *rows)
1966 : {
1967 14640 : GatherPath *pathnode = makeNode(GatherPath);
1968 :
1969 : Assert(subpath->parallel_safe);
1970 :
1971 14640 : pathnode->path.pathtype = T_Gather;
1972 14640 : pathnode->path.parent = rel;
1973 14640 : pathnode->path.pathtarget = target;
1974 14640 : pathnode->path.param_info = get_baserel_parampathinfo(root, rel,
1975 : required_outer);
1976 14640 : pathnode->path.parallel_aware = false;
1977 14640 : pathnode->path.parallel_safe = false;
1978 14640 : pathnode->path.parallel_workers = 0;
1979 14640 : pathnode->path.pathkeys = NIL; /* Gather has unordered result */
1980 :
1981 14640 : pathnode->subpath = subpath;
1982 14640 : pathnode->num_workers = subpath->parallel_workers;
1983 14640 : pathnode->single_copy = false;
1984 :
1985 14640 : if (pathnode->num_workers == 0)
1986 : {
1987 0 : pathnode->path.pathkeys = subpath->pathkeys;
1988 0 : pathnode->num_workers = 1;
1989 0 : pathnode->single_copy = true;
1990 : }
1991 :
1992 14640 : cost_gather(pathnode, root, rel, pathnode->path.param_info, rows);
1993 :
1994 14640 : return pathnode;
1995 : }
1996 :
1997 : /*
1998 : * create_subqueryscan_path
1999 : * Creates a path corresponding to a scan of a subquery,
2000 : * returning the pathnode.
2001 : *
2002 : * Caller must pass trivial_pathtarget = true if it believes rel->reltarget to
2003 : * be trivial, ie just a fetch of all the subquery output columns in order.
2004 : * While we could determine that here, the caller can usually do it more
2005 : * efficiently (or at least amortize it over multiple calls).
2006 : */
2007 : SubqueryScanPath *
2008 20490 : create_subqueryscan_path(PlannerInfo *root, RelOptInfo *rel, Path *subpath,
2009 : bool trivial_pathtarget,
2010 : List *pathkeys, Relids required_outer)
2011 : {
2012 20490 : SubqueryScanPath *pathnode = makeNode(SubqueryScanPath);
2013 :
2014 20490 : pathnode->path.pathtype = T_SubqueryScan;
2015 20490 : pathnode->path.parent = rel;
2016 20490 : pathnode->path.pathtarget = rel->reltarget;
2017 20490 : pathnode->path.param_info = get_baserel_parampathinfo(root, rel,
2018 : required_outer);
2019 20490 : pathnode->path.parallel_aware = false;
2020 31596 : pathnode->path.parallel_safe = rel->consider_parallel &&
2021 11106 : subpath->parallel_safe;
2022 20490 : pathnode->path.parallel_workers = subpath->parallel_workers;
2023 20490 : pathnode->path.pathkeys = pathkeys;
2024 20490 : pathnode->subpath = subpath;
2025 :
2026 20490 : cost_subqueryscan(pathnode, root, rel, pathnode->path.param_info,
2027 : trivial_pathtarget);
2028 :
2029 20490 : return pathnode;
2030 : }
2031 :
2032 : /*
2033 : * create_functionscan_path
2034 : * Creates a path corresponding to a sequential scan of a function,
2035 : * returning the pathnode.
2036 : */
2037 : Path *
2038 34858 : create_functionscan_path(PlannerInfo *root, RelOptInfo *rel,
2039 : List *pathkeys, Relids required_outer)
2040 : {
2041 34858 : Path *pathnode = makeNode(Path);
2042 :
2043 34858 : pathnode->pathtype = T_FunctionScan;
2044 34858 : pathnode->parent = rel;
2045 34858 : pathnode->pathtarget = rel->reltarget;
2046 34858 : pathnode->param_info = get_baserel_parampathinfo(root, rel,
2047 : required_outer);
2048 34858 : pathnode->parallel_aware = false;
2049 34858 : pathnode->parallel_safe = rel->consider_parallel;
2050 34858 : pathnode->parallel_workers = 0;
2051 34858 : pathnode->pathkeys = pathkeys;
2052 :
2053 34858 : cost_functionscan(pathnode, root, rel, pathnode->param_info);
2054 :
2055 34858 : return pathnode;
2056 : }
2057 :
2058 : /*
2059 : * create_tablefuncscan_path
2060 : * Creates a path corresponding to a sequential scan of a table function,
2061 : * returning the pathnode.
2062 : */
2063 : Path *
2064 216 : create_tablefuncscan_path(PlannerInfo *root, RelOptInfo *rel,
2065 : Relids required_outer)
2066 : {
2067 216 : Path *pathnode = makeNode(Path);
2068 :
2069 216 : pathnode->pathtype = T_TableFuncScan;
2070 216 : pathnode->parent = rel;
2071 216 : pathnode->pathtarget = rel->reltarget;
2072 216 : pathnode->param_info = get_baserel_parampathinfo(root, rel,
2073 : required_outer);
2074 216 : pathnode->parallel_aware = false;
2075 216 : pathnode->parallel_safe = rel->consider_parallel;
2076 216 : pathnode->parallel_workers = 0;
2077 216 : pathnode->pathkeys = NIL; /* result is always unordered */
2078 :
2079 216 : cost_tablefuncscan(pathnode, root, rel, pathnode->param_info);
2080 :
2081 216 : return pathnode;
2082 : }
2083 :
2084 : /*
2085 : * create_valuesscan_path
2086 : * Creates a path corresponding to a scan of a VALUES list,
2087 : * returning the pathnode.
2088 : */
2089 : Path *
2090 7216 : create_valuesscan_path(PlannerInfo *root, RelOptInfo *rel,
2091 : Relids required_outer)
2092 : {
2093 7216 : Path *pathnode = makeNode(Path);
2094 :
2095 7216 : pathnode->pathtype = T_ValuesScan;
2096 7216 : pathnode->parent = rel;
2097 7216 : pathnode->pathtarget = rel->reltarget;
2098 7216 : pathnode->param_info = get_baserel_parampathinfo(root, rel,
2099 : required_outer);
2100 7216 : pathnode->parallel_aware = false;
2101 7216 : pathnode->parallel_safe = rel->consider_parallel;
2102 7216 : pathnode->parallel_workers = 0;
2103 7216 : pathnode->pathkeys = NIL; /* result is always unordered */
2104 :
2105 7216 : cost_valuesscan(pathnode, root, rel, pathnode->param_info);
2106 :
2107 7216 : return pathnode;
2108 : }
2109 :
2110 : /*
2111 : * create_ctescan_path
2112 : * Creates a path corresponding to a scan of a non-self-reference CTE,
2113 : * returning the pathnode.
2114 : */
2115 : Path *
2116 2474 : create_ctescan_path(PlannerInfo *root, RelOptInfo *rel, Relids required_outer)
2117 : {
2118 2474 : Path *pathnode = makeNode(Path);
2119 :
2120 2474 : pathnode->pathtype = T_CteScan;
2121 2474 : pathnode->parent = rel;
2122 2474 : pathnode->pathtarget = rel->reltarget;
2123 2474 : pathnode->param_info = get_baserel_parampathinfo(root, rel,
2124 : required_outer);
2125 2474 : pathnode->parallel_aware = false;
2126 2474 : pathnode->parallel_safe = rel->consider_parallel;
2127 2474 : pathnode->parallel_workers = 0;
2128 2474 : pathnode->pathkeys = NIL; /* XXX for now, result is always unordered */
2129 :
2130 2474 : cost_ctescan(pathnode, root, rel, pathnode->param_info);
2131 :
2132 2474 : return pathnode;
2133 : }
2134 :
2135 : /*
2136 : * create_namedtuplestorescan_path
2137 : * Creates a path corresponding to a scan of a named tuplestore, returning
2138 : * the pathnode.
2139 : */
2140 : Path *
2141 438 : create_namedtuplestorescan_path(PlannerInfo *root, RelOptInfo *rel,
2142 : Relids required_outer)
2143 : {
2144 438 : Path *pathnode = makeNode(Path);
2145 :
2146 438 : pathnode->pathtype = T_NamedTuplestoreScan;
2147 438 : pathnode->parent = rel;
2148 438 : pathnode->pathtarget = rel->reltarget;
2149 438 : pathnode->param_info = get_baserel_parampathinfo(root, rel,
2150 : required_outer);
2151 438 : pathnode->parallel_aware = false;
2152 438 : pathnode->parallel_safe = rel->consider_parallel;
2153 438 : pathnode->parallel_workers = 0;
2154 438 : pathnode->pathkeys = NIL; /* result is always unordered */
2155 :
2156 438 : cost_namedtuplestorescan(pathnode, root, rel, pathnode->param_info);
2157 :
2158 438 : return pathnode;
2159 : }
2160 :
2161 : /*
2162 : * create_resultscan_path
2163 : * Creates a path corresponding to a scan of an RTE_RESULT relation,
2164 : * returning the pathnode.
2165 : */
2166 : Path *
2167 1370 : create_resultscan_path(PlannerInfo *root, RelOptInfo *rel,
2168 : Relids required_outer)
2169 : {
2170 1370 : Path *pathnode = makeNode(Path);
2171 :
2172 1370 : pathnode->pathtype = T_Result;
2173 1370 : pathnode->parent = rel;
2174 1370 : pathnode->pathtarget = rel->reltarget;
2175 1370 : pathnode->param_info = get_baserel_parampathinfo(root, rel,
2176 : required_outer);
2177 1370 : pathnode->parallel_aware = false;
2178 1370 : pathnode->parallel_safe = rel->consider_parallel;
2179 1370 : pathnode->parallel_workers = 0;
2180 1370 : pathnode->pathkeys = NIL; /* result is always unordered */
2181 :
2182 1370 : cost_resultscan(pathnode, root, rel, pathnode->param_info);
2183 :
2184 1370 : return pathnode;
2185 : }
2186 :
2187 : /*
2188 : * create_worktablescan_path
2189 : * Creates a path corresponding to a scan of a self-reference CTE,
2190 : * returning the pathnode.
2191 : */
2192 : Path *
2193 708 : create_worktablescan_path(PlannerInfo *root, RelOptInfo *rel,
2194 : Relids required_outer)
2195 : {
2196 708 : Path *pathnode = makeNode(Path);
2197 :
2198 708 : pathnode->pathtype = T_WorkTableScan;
2199 708 : pathnode->parent = rel;
2200 708 : pathnode->pathtarget = rel->reltarget;
2201 708 : pathnode->param_info = get_baserel_parampathinfo(root, rel,
2202 : required_outer);
2203 708 : pathnode->parallel_aware = false;
2204 708 : pathnode->parallel_safe = rel->consider_parallel;
2205 708 : pathnode->parallel_workers = 0;
2206 708 : pathnode->pathkeys = NIL; /* result is always unordered */
2207 :
2208 : /* Cost is the same as for a regular CTE scan */
2209 708 : cost_ctescan(pathnode, root, rel, pathnode->param_info);
2210 :
2211 708 : return pathnode;
2212 : }
2213 :
2214 : /*
2215 : * create_foreignscan_path
2216 : * Creates a path corresponding to a scan of a foreign base table,
2217 : * returning the pathnode.
2218 : *
2219 : * This function is never called from core Postgres; rather, it's expected
2220 : * to be called by the GetForeignPaths function of a foreign data wrapper.
2221 : * We make the FDW supply all fields of the path, since we do not have any way
2222 : * to calculate them in core. However, there is a usually-sane default for
2223 : * the pathtarget (rel->reltarget), so we let a NULL for "target" select that.
2224 : */
2225 : ForeignPath *
2226 3204 : create_foreignscan_path(PlannerInfo *root, RelOptInfo *rel,
2227 : PathTarget *target,
2228 : double rows, Cost startup_cost, Cost total_cost,
2229 : List *pathkeys,
2230 : Relids required_outer,
2231 : Path *fdw_outerpath,
2232 : List *fdw_private)
2233 : {
2234 3204 : ForeignPath *pathnode = makeNode(ForeignPath);
2235 :
2236 : /* Historically some FDWs were confused about when to use this */
2237 : Assert(IS_SIMPLE_REL(rel));
2238 :
2239 3204 : pathnode->path.pathtype = T_ForeignScan;
2240 3204 : pathnode->path.parent = rel;
2241 3204 : pathnode->path.pathtarget = target ? target : rel->reltarget;
2242 3204 : pathnode->path.param_info = get_baserel_parampathinfo(root, rel,
2243 : required_outer);
2244 3204 : pathnode->path.parallel_aware = false;
2245 3204 : pathnode->path.parallel_safe = rel->consider_parallel;
2246 3204 : pathnode->path.parallel_workers = 0;
2247 3204 : pathnode->path.rows = rows;
2248 3204 : pathnode->path.startup_cost = startup_cost;
2249 3204 : pathnode->path.total_cost = total_cost;
2250 3204 : pathnode->path.pathkeys = pathkeys;
2251 :
2252 3204 : pathnode->fdw_outerpath = fdw_outerpath;
2253 3204 : pathnode->fdw_private = fdw_private;
2254 :
2255 3204 : return pathnode;
2256 : }
2257 :
2258 : /*
2259 : * create_foreign_join_path
2260 : * Creates a path corresponding to a scan of a foreign join,
2261 : * returning the pathnode.
2262 : *
2263 : * This function is never called from core Postgres; rather, it's expected
2264 : * to be called by the GetForeignJoinPaths function of a foreign data wrapper.
2265 : * We make the FDW supply all fields of the path, since we do not have any way
2266 : * to calculate them in core. However, there is a usually-sane default for
2267 : * the pathtarget (rel->reltarget), so we let a NULL for "target" select that.
2268 : */
2269 : ForeignPath *
2270 738 : create_foreign_join_path(PlannerInfo *root, RelOptInfo *rel,
2271 : PathTarget *target,
2272 : double rows, Cost startup_cost, Cost total_cost,
2273 : List *pathkeys,
2274 : Relids required_outer,
2275 : Path *fdw_outerpath,
2276 : List *fdw_private)
2277 : {
2278 738 : ForeignPath *pathnode = makeNode(ForeignPath);
2279 :
2280 : /*
2281 : * We should use get_joinrel_parampathinfo to handle parameterized paths,
2282 : * but the API of this function doesn't support it, and existing
2283 : * extensions aren't yet trying to build such paths anyway. For the
2284 : * moment just throw an error if someone tries it; eventually we should
2285 : * revisit this.
2286 : */
2287 738 : if (!bms_is_empty(required_outer) || !bms_is_empty(rel->lateral_relids))
2288 0 : elog(ERROR, "parameterized foreign joins are not supported yet");
2289 :
2290 738 : pathnode->path.pathtype = T_ForeignScan;
2291 738 : pathnode->path.parent = rel;
2292 738 : pathnode->path.pathtarget = target ? target : rel->reltarget;
2293 738 : pathnode->path.param_info = NULL; /* XXX see above */
2294 738 : pathnode->path.parallel_aware = false;
2295 738 : pathnode->path.parallel_safe = rel->consider_parallel;
2296 738 : pathnode->path.parallel_workers = 0;
2297 738 : pathnode->path.rows = rows;
2298 738 : pathnode->path.startup_cost = startup_cost;
2299 738 : pathnode->path.total_cost = total_cost;
2300 738 : pathnode->path.pathkeys = pathkeys;
2301 :
2302 738 : pathnode->fdw_outerpath = fdw_outerpath;
2303 738 : pathnode->fdw_private = fdw_private;
2304 :
2305 738 : return pathnode;
2306 : }
2307 :
2308 : /*
2309 : * create_foreign_upper_path
2310 : * Creates a path corresponding to an upper relation that's computed
2311 : * directly by an FDW, returning the pathnode.
2312 : *
2313 : * This function is never called from core Postgres; rather, it's expected to
2314 : * be called by the GetForeignUpperPaths function of a foreign data wrapper.
2315 : * We make the FDW supply all fields of the path, since we do not have any way
2316 : * to calculate them in core. However, there is a usually-sane default for
2317 : * the pathtarget (rel->reltarget), so we let a NULL for "target" select that.
2318 : */
2319 : ForeignPath *
2320 546 : create_foreign_upper_path(PlannerInfo *root, RelOptInfo *rel,
2321 : PathTarget *target,
2322 : double rows, Cost startup_cost, Cost total_cost,
2323 : List *pathkeys,
2324 : Path *fdw_outerpath,
2325 : List *fdw_private)
2326 : {
2327 546 : ForeignPath *pathnode = makeNode(ForeignPath);
2328 :
2329 : /*
2330 : * Upper relations should never have any lateral references, since joining
2331 : * is complete.
2332 : */
2333 : Assert(bms_is_empty(rel->lateral_relids));
2334 :
2335 546 : pathnode->path.pathtype = T_ForeignScan;
2336 546 : pathnode->path.parent = rel;
2337 546 : pathnode->path.pathtarget = target ? target : rel->reltarget;
2338 546 : pathnode->path.param_info = NULL;
2339 546 : pathnode->path.parallel_aware = false;
2340 546 : pathnode->path.parallel_safe = rel->consider_parallel;
2341 546 : pathnode->path.parallel_workers = 0;
2342 546 : pathnode->path.rows = rows;
2343 546 : pathnode->path.startup_cost = startup_cost;
2344 546 : pathnode->path.total_cost = total_cost;
2345 546 : pathnode->path.pathkeys = pathkeys;
2346 :
2347 546 : pathnode->fdw_outerpath = fdw_outerpath;
2348 546 : pathnode->fdw_private = fdw_private;
2349 :
2350 546 : return pathnode;
2351 : }
2352 :
2353 : /*
2354 : * calc_nestloop_required_outer
2355 : * Compute the required_outer set for a nestloop join path
2356 : *
2357 : * Note: result must not share storage with either input
2358 : */
2359 : Relids
2360 1919240 : calc_nestloop_required_outer(Relids outerrelids,
2361 : Relids outer_paramrels,
2362 : Relids innerrelids,
2363 : Relids inner_paramrels)
2364 : {
2365 : Relids required_outer;
2366 :
2367 : /* inner_path can require rels from outer path, but not vice versa */
2368 : Assert(!bms_overlap(outer_paramrels, innerrelids));
2369 : /* easy case if inner path is not parameterized */
2370 1919240 : if (!inner_paramrels)
2371 1301992 : return bms_copy(outer_paramrels);
2372 : /* else, form the union ... */
2373 617248 : required_outer = bms_union(outer_paramrels, inner_paramrels);
2374 : /* ... and remove any mention of now-satisfied outer rels */
2375 617248 : required_outer = bms_del_members(required_outer,
2376 : outerrelids);
2377 617248 : return required_outer;
2378 : }
2379 :
2380 : /*
2381 : * calc_non_nestloop_required_outer
2382 : * Compute the required_outer set for a merge or hash join path
2383 : *
2384 : * Note: result must not share storage with either input
2385 : */
2386 : Relids
2387 1313604 : calc_non_nestloop_required_outer(Path *outer_path, Path *inner_path)
2388 : {
2389 1313604 : Relids outer_paramrels = PATH_REQ_OUTER(outer_path);
2390 1313604 : Relids inner_paramrels = PATH_REQ_OUTER(inner_path);
2391 : Relids required_outer;
2392 :
2393 : /* neither path can require rels from the other */
2394 : Assert(!bms_overlap(outer_paramrels, inner_path->parent->relids));
2395 : Assert(!bms_overlap(inner_paramrels, outer_path->parent->relids));
2396 : /* form the union ... */
2397 1313604 : required_outer = bms_union(outer_paramrels, inner_paramrels);
2398 : /* we do not need an explicit test for empty; bms_union gets it right */
2399 1313604 : return required_outer;
2400 : }
2401 :
2402 : /*
2403 : * create_nestloop_path
2404 : * Creates a pathnode corresponding to a nestloop join between two
2405 : * relations.
2406 : *
2407 : * 'joinrel' is the join relation.
2408 : * 'jointype' is the type of join required
2409 : * 'workspace' is the result from initial_cost_nestloop
2410 : * 'extra' contains various information about the join
2411 : * 'outer_path' is the outer path
2412 : * 'inner_path' is the inner path
2413 : * 'restrict_clauses' are the RestrictInfo nodes to apply at the join
2414 : * 'pathkeys' are the path keys of the new join path
2415 : * 'required_outer' is the set of required outer rels
2416 : *
2417 : * Returns the resulting path node.
2418 : */
2419 : NestPath *
2420 852102 : create_nestloop_path(PlannerInfo *root,
2421 : RelOptInfo *joinrel,
2422 : JoinType jointype,
2423 : JoinCostWorkspace *workspace,
2424 : JoinPathExtraData *extra,
2425 : Path *outer_path,
2426 : Path *inner_path,
2427 : List *restrict_clauses,
2428 : List *pathkeys,
2429 : Relids required_outer)
2430 : {
2431 852102 : NestPath *pathnode = makeNode(NestPath);
2432 852102 : Relids inner_req_outer = PATH_REQ_OUTER(inner_path);
2433 :
2434 : /*
2435 : * If the inner path is parameterized by the outer, we must drop any
2436 : * restrict_clauses that are due to be moved into the inner path. We have
2437 : * to do this now, rather than postpone the work till createplan time,
2438 : * because the restrict_clauses list can affect the size and cost
2439 : * estimates for this path. We detect such clauses by checking for serial
2440 : * number match to clauses already enforced in the inner path.
2441 : */
2442 852102 : if (bms_overlap(inner_req_outer, outer_path->parent->relids))
2443 : {
2444 235408 : Bitmapset *enforced_serials = get_param_path_clause_serials(inner_path);
2445 235408 : List *jclauses = NIL;
2446 : ListCell *lc;
2447 :
2448 503070 : foreach(lc, restrict_clauses)
2449 : {
2450 267662 : RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
2451 :
2452 267662 : if (!bms_is_member(rinfo->rinfo_serial, enforced_serials))
2453 29512 : jclauses = lappend(jclauses, rinfo);
2454 : }
2455 235408 : restrict_clauses = jclauses;
2456 : }
2457 :
2458 852102 : pathnode->jpath.path.pathtype = T_NestLoop;
2459 852102 : pathnode->jpath.path.parent = joinrel;
2460 852102 : pathnode->jpath.path.pathtarget = joinrel->reltarget;
2461 852102 : pathnode->jpath.path.param_info =
2462 852102 : get_joinrel_parampathinfo(root,
2463 : joinrel,
2464 : outer_path,
2465 : inner_path,
2466 : extra->sjinfo,
2467 : required_outer,
2468 : &restrict_clauses);
2469 852102 : pathnode->jpath.path.parallel_aware = false;
2470 2455772 : pathnode->jpath.path.parallel_safe = joinrel->consider_parallel &&
2471 852102 : outer_path->parallel_safe && inner_path->parallel_safe;
2472 : /* This is a foolish way to estimate parallel_workers, but for now... */
2473 852102 : pathnode->jpath.path.parallel_workers = outer_path->parallel_workers;
2474 852102 : pathnode->jpath.path.pathkeys = pathkeys;
2475 852102 : pathnode->jpath.jointype = jointype;
2476 852102 : pathnode->jpath.inner_unique = extra->inner_unique;
2477 852102 : pathnode->jpath.outerjoinpath = outer_path;
2478 852102 : pathnode->jpath.innerjoinpath = inner_path;
2479 852102 : pathnode->jpath.joinrestrictinfo = restrict_clauses;
2480 :
2481 852102 : final_cost_nestloop(root, pathnode, workspace, extra);
2482 :
2483 852102 : return pathnode;
2484 : }
2485 :
2486 : /*
2487 : * create_mergejoin_path
2488 : * Creates a pathnode corresponding to a mergejoin join between
2489 : * two relations
2490 : *
2491 : * 'joinrel' is the join relation
2492 : * 'jointype' is the type of join required
2493 : * 'workspace' is the result from initial_cost_mergejoin
2494 : * 'extra' contains various information about the join
2495 : * 'outer_path' is the outer path
2496 : * 'inner_path' is the inner path
2497 : * 'restrict_clauses' are the RestrictInfo nodes to apply at the join
2498 : * 'pathkeys' are the path keys of the new join path
2499 : * 'required_outer' is the set of required outer rels
2500 : * 'mergeclauses' are the RestrictInfo nodes to use as merge clauses
2501 : * (this should be a subset of the restrict_clauses list)
2502 : * 'outersortkeys' are the sort varkeys for the outer relation
2503 : * 'innersortkeys' are the sort varkeys for the inner relation
2504 : */
2505 : MergePath *
2506 203002 : create_mergejoin_path(PlannerInfo *root,
2507 : RelOptInfo *joinrel,
2508 : JoinType jointype,
2509 : JoinCostWorkspace *workspace,
2510 : JoinPathExtraData *extra,
2511 : Path *outer_path,
2512 : Path *inner_path,
2513 : List *restrict_clauses,
2514 : List *pathkeys,
2515 : Relids required_outer,
2516 : List *mergeclauses,
2517 : List *outersortkeys,
2518 : List *innersortkeys)
2519 : {
2520 203002 : MergePath *pathnode = makeNode(MergePath);
2521 :
2522 203002 : pathnode->jpath.path.pathtype = T_MergeJoin;
2523 203002 : pathnode->jpath.path.parent = joinrel;
2524 203002 : pathnode->jpath.path.pathtarget = joinrel->reltarget;
2525 203002 : pathnode->jpath.path.param_info =
2526 203002 : get_joinrel_parampathinfo(root,
2527 : joinrel,
2528 : outer_path,
2529 : inner_path,
2530 : extra->sjinfo,
2531 : required_outer,
2532 : &restrict_clauses);
2533 203002 : pathnode->jpath.path.parallel_aware = false;
2534 576590 : pathnode->jpath.path.parallel_safe = joinrel->consider_parallel &&
2535 203002 : outer_path->parallel_safe && inner_path->parallel_safe;
2536 : /* This is a foolish way to estimate parallel_workers, but for now... */
2537 203002 : pathnode->jpath.path.parallel_workers = outer_path->parallel_workers;
2538 203002 : pathnode->jpath.path.pathkeys = pathkeys;
2539 203002 : pathnode->jpath.jointype = jointype;
2540 203002 : pathnode->jpath.inner_unique = extra->inner_unique;
2541 203002 : pathnode->jpath.outerjoinpath = outer_path;
2542 203002 : pathnode->jpath.innerjoinpath = inner_path;
2543 203002 : pathnode->jpath.joinrestrictinfo = restrict_clauses;
2544 203002 : pathnode->path_mergeclauses = mergeclauses;
2545 203002 : pathnode->outersortkeys = outersortkeys;
2546 203002 : pathnode->innersortkeys = innersortkeys;
2547 : /* pathnode->skip_mark_restore will be set by final_cost_mergejoin */
2548 : /* pathnode->materialize_inner will be set by final_cost_mergejoin */
2549 :
2550 203002 : final_cost_mergejoin(root, pathnode, workspace, extra);
2551 :
2552 203002 : return pathnode;
2553 : }
2554 :
2555 : /*
2556 : * create_hashjoin_path
2557 : * Creates a pathnode corresponding to a hash join between two relations.
2558 : *
2559 : * 'joinrel' is the join relation
2560 : * 'jointype' is the type of join required
2561 : * 'workspace' is the result from initial_cost_hashjoin
2562 : * 'extra' contains various information about the join
2563 : * 'outer_path' is the cheapest outer path
2564 : * 'inner_path' is the cheapest inner path
2565 : * 'parallel_hash' to select Parallel Hash of inner path (shared hash table)
2566 : * 'restrict_clauses' are the RestrictInfo nodes to apply at the join
2567 : * 'required_outer' is the set of required outer rels
2568 : * 'hashclauses' are the RestrictInfo nodes to use as hash clauses
2569 : * (this should be a subset of the restrict_clauses list)
2570 : */
2571 : HashPath *
2572 189866 : create_hashjoin_path(PlannerInfo *root,
2573 : RelOptInfo *joinrel,
2574 : JoinType jointype,
2575 : JoinCostWorkspace *workspace,
2576 : JoinPathExtraData *extra,
2577 : Path *outer_path,
2578 : Path *inner_path,
2579 : bool parallel_hash,
2580 : List *restrict_clauses,
2581 : Relids required_outer,
2582 : List *hashclauses)
2583 : {
2584 189866 : HashPath *pathnode = makeNode(HashPath);
2585 :
2586 189866 : pathnode->jpath.path.pathtype = T_HashJoin;
2587 189866 : pathnode->jpath.path.parent = joinrel;
2588 189866 : pathnode->jpath.path.pathtarget = joinrel->reltarget;
2589 189866 : pathnode->jpath.path.param_info =
2590 189866 : get_joinrel_parampathinfo(root,
2591 : joinrel,
2592 : outer_path,
2593 : inner_path,
2594 : extra->sjinfo,
2595 : required_outer,
2596 : &restrict_clauses);
2597 189866 : pathnode->jpath.path.parallel_aware =
2598 189866 : joinrel->consider_parallel && parallel_hash;
2599 534714 : pathnode->jpath.path.parallel_safe = joinrel->consider_parallel &&
2600 189866 : outer_path->parallel_safe && inner_path->parallel_safe;
2601 : /* This is a foolish way to estimate parallel_workers, but for now... */
2602 189866 : pathnode->jpath.path.parallel_workers = outer_path->parallel_workers;
2603 :
2604 : /*
2605 : * A hashjoin never has pathkeys, since its output ordering is
2606 : * unpredictable due to possible batching. XXX If the inner relation is
2607 : * small enough, we could instruct the executor that it must not batch,
2608 : * and then we could assume that the output inherits the outer relation's
2609 : * ordering, which might save a sort step. However there is considerable
2610 : * downside if our estimate of the inner relation size is badly off. For
2611 : * the moment we don't risk it. (Note also that if we wanted to take this
2612 : * seriously, joinpath.c would have to consider many more paths for the
2613 : * outer rel than it does now.)
2614 : */
2615 189866 : pathnode->jpath.path.pathkeys = NIL;
2616 189866 : pathnode->jpath.jointype = jointype;
2617 189866 : pathnode->jpath.inner_unique = extra->inner_unique;
2618 189866 : pathnode->jpath.outerjoinpath = outer_path;
2619 189866 : pathnode->jpath.innerjoinpath = inner_path;
2620 189866 : pathnode->jpath.joinrestrictinfo = restrict_clauses;
2621 189866 : pathnode->path_hashclauses = hashclauses;
2622 : /* final_cost_hashjoin will fill in pathnode->num_batches */
2623 :
2624 189866 : final_cost_hashjoin(root, pathnode, workspace, extra);
2625 :
2626 189866 : return pathnode;
2627 : }
2628 :
2629 : /*
2630 : * create_projection_path
2631 : * Creates a pathnode that represents performing a projection.
2632 : *
2633 : * 'rel' is the parent relation associated with the result
2634 : * 'subpath' is the path representing the source of data
2635 : * 'target' is the PathTarget to be computed
2636 : */
2637 : ProjectionPath *
2638 340714 : create_projection_path(PlannerInfo *root,
2639 : RelOptInfo *rel,
2640 : Path *subpath,
2641 : PathTarget *target)
2642 : {
2643 340714 : ProjectionPath *pathnode = makeNode(ProjectionPath);
2644 : PathTarget *oldtarget;
2645 :
2646 : /*
2647 : * We mustn't put a ProjectionPath directly above another; it's useless
2648 : * and will confuse create_projection_plan. Rather than making sure all
2649 : * callers handle that, let's implement it here, by stripping off any
2650 : * ProjectionPath in what we're given. Given this rule, there won't be
2651 : * more than one.
2652 : */
2653 340714 : if (IsA(subpath, ProjectionPath))
2654 : {
2655 12 : ProjectionPath *subpp = (ProjectionPath *) subpath;
2656 :
2657 : Assert(subpp->path.parent == rel);
2658 12 : subpath = subpp->subpath;
2659 : Assert(!IsA(subpath, ProjectionPath));
2660 : }
2661 :
2662 340714 : pathnode->path.pathtype = T_Result;
2663 340714 : pathnode->path.parent = rel;
2664 340714 : pathnode->path.pathtarget = target;
2665 : /* For now, assume we are above any joins, so no parameterization */
2666 340714 : pathnode->path.param_info = NULL;
2667 340714 : pathnode->path.parallel_aware = false;
2668 748462 : pathnode->path.parallel_safe = rel->consider_parallel &&
2669 406292 : subpath->parallel_safe &&
2670 65578 : is_parallel_safe(root, (Node *) target->exprs);
2671 340714 : pathnode->path.parallel_workers = subpath->parallel_workers;
2672 : /* Projection does not change the sort order */
2673 340714 : pathnode->path.pathkeys = subpath->pathkeys;
2674 :
2675 340714 : pathnode->subpath = subpath;
2676 :
2677 : /*
2678 : * We might not need a separate Result node. If the input plan node type
2679 : * can project, we can just tell it to project something else. Or, if it
2680 : * can't project but the desired target has the same expression list as
2681 : * what the input will produce anyway, we can still give it the desired
2682 : * tlist (possibly changing its ressortgroupref labels, but nothing else).
2683 : * Note: in the latter case, create_projection_plan has to recheck our
2684 : * conclusion; see comments therein.
2685 : */
2686 340714 : oldtarget = subpath->pathtarget;
2687 352656 : if (is_projection_capable_path(subpath) ||
2688 11942 : equal(oldtarget->exprs, target->exprs))
2689 : {
2690 : /* No separate Result node needed */
2691 338712 : pathnode->dummypp = true;
2692 :
2693 : /*
2694 : * Set cost of plan as subpath's cost, adjusted for tlist replacement.
2695 : */
2696 338712 : pathnode->path.rows = subpath->rows;
2697 338712 : pathnode->path.startup_cost = subpath->startup_cost +
2698 338712 : (target->cost.startup - oldtarget->cost.startup);
2699 338712 : pathnode->path.total_cost = subpath->total_cost +
2700 338712 : (target->cost.startup - oldtarget->cost.startup) +
2701 338712 : (target->cost.per_tuple - oldtarget->cost.per_tuple) * subpath->rows;
2702 : }
2703 : else
2704 : {
2705 : /* We really do need the Result node */
2706 2002 : pathnode->dummypp = false;
2707 :
2708 : /*
2709 : * The Result node's cost is cpu_tuple_cost per row, plus the cost of
2710 : * evaluating the tlist. There is no qual to worry about.
2711 : */
2712 2002 : pathnode->path.rows = subpath->rows;
2713 2002 : pathnode->path.startup_cost = subpath->startup_cost +
2714 2002 : target->cost.startup;
2715 2002 : pathnode->path.total_cost = subpath->total_cost +
2716 2002 : target->cost.startup +
2717 2002 : (cpu_tuple_cost + target->cost.per_tuple) * subpath->rows;
2718 : }
2719 :
2720 340714 : return pathnode;
2721 : }
2722 :
2723 : /*
2724 : * apply_projection_to_path
2725 : * Add a projection step, or just apply the target directly to given path.
2726 : *
2727 : * This has the same net effect as create_projection_path(), except that if
2728 : * a separate Result plan node isn't needed, we just replace the given path's
2729 : * pathtarget with the desired one. This must be used only when the caller
2730 : * knows that the given path isn't referenced elsewhere and so can be modified
2731 : * in-place.
2732 : *
2733 : * If the input path is a GatherPath or GatherMergePath, we try to push the
2734 : * new target down to its input as well; this is a yet more invasive
2735 : * modification of the input path, which create_projection_path() can't do.
2736 : *
2737 : * Note that we mustn't change the source path's parent link; so when it is
2738 : * add_path'd to "rel" things will be a bit inconsistent. So far that has
2739 : * not caused any trouble.
2740 : *
2741 : * 'rel' is the parent relation associated with the result
2742 : * 'path' is the path representing the source of data
2743 : * 'target' is the PathTarget to be computed
2744 : */
2745 : Path *
2746 22512 : apply_projection_to_path(PlannerInfo *root,
2747 : RelOptInfo *rel,
2748 : Path *path,
2749 : PathTarget *target)
2750 : {
2751 : QualCost oldcost;
2752 :
2753 : /*
2754 : * If given path can't project, we might need a Result node, so make a
2755 : * separate ProjectionPath.
2756 : */
2757 22512 : if (!is_projection_capable_path(path))
2758 10868 : return (Path *) create_projection_path(root, rel, path, target);
2759 :
2760 : /*
2761 : * We can just jam the desired tlist into the existing path, being sure to
2762 : * update its cost estimates appropriately.
2763 : */
2764 11644 : oldcost = path->pathtarget->cost;
2765 11644 : path->pathtarget = target;
2766 :
2767 11644 : path->startup_cost += target->cost.startup - oldcost.startup;
2768 11644 : path->total_cost += target->cost.startup - oldcost.startup +
2769 11644 : (target->cost.per_tuple - oldcost.per_tuple) * path->rows;
2770 :
2771 : /*
2772 : * If the path happens to be a Gather or GatherMerge path, we'd like to
2773 : * arrange for the subpath to return the required target list so that
2774 : * workers can help project. But if there is something that is not
2775 : * parallel-safe in the target expressions, then we can't.
2776 : */
2777 12082 : if ((IsA(path, GatherPath) || IsA(path, GatherMergePath)) &&
2778 438 : is_parallel_safe(root, (Node *) target->exprs))
2779 : {
2780 : /*
2781 : * We always use create_projection_path here, even if the subpath is
2782 : * projection-capable, so as to avoid modifying the subpath in place.
2783 : * It seems unlikely at present that there could be any other
2784 : * references to the subpath, but better safe than sorry.
2785 : *
2786 : * Note that we don't change the parallel path's cost estimates; it
2787 : * might be appropriate to do so, to reflect the fact that the bulk of
2788 : * the target evaluation will happen in workers.
2789 : */
2790 438 : if (IsA(path, GatherPath))
2791 : {
2792 0 : GatherPath *gpath = (GatherPath *) path;
2793 :
2794 0 : gpath->subpath = (Path *)
2795 0 : create_projection_path(root,
2796 0 : gpath->subpath->parent,
2797 : gpath->subpath,
2798 : target);
2799 : }
2800 : else
2801 : {
2802 438 : GatherMergePath *gmpath = (GatherMergePath *) path;
2803 :
2804 438 : gmpath->subpath = (Path *)
2805 438 : create_projection_path(root,
2806 438 : gmpath->subpath->parent,
2807 : gmpath->subpath,
2808 : target);
2809 : }
2810 : }
2811 11206 : else if (path->parallel_safe &&
2812 4616 : !is_parallel_safe(root, (Node *) target->exprs))
2813 : {
2814 : /*
2815 : * We're inserting a parallel-restricted target list into a path
2816 : * currently marked parallel-safe, so we have to mark it as no longer
2817 : * safe.
2818 : */
2819 12 : path->parallel_safe = false;
2820 : }
2821 :
2822 11644 : return path;
2823 : }
2824 :
2825 : /*
2826 : * create_set_projection_path
2827 : * Creates a pathnode that represents performing a projection that
2828 : * includes set-returning functions.
2829 : *
2830 : * 'rel' is the parent relation associated with the result
2831 : * 'subpath' is the path representing the source of data
2832 : * 'target' is the PathTarget to be computed
2833 : */
2834 : ProjectSetPath *
2835 6578 : create_set_projection_path(PlannerInfo *root,
2836 : RelOptInfo *rel,
2837 : Path *subpath,
2838 : PathTarget *target)
2839 : {
2840 6578 : ProjectSetPath *pathnode = makeNode(ProjectSetPath);
2841 : double tlist_rows;
2842 : ListCell *lc;
2843 :
2844 6578 : pathnode->path.pathtype = T_ProjectSet;
2845 6578 : pathnode->path.parent = rel;
2846 6578 : pathnode->path.pathtarget = target;
2847 : /* For now, assume we are above any joins, so no parameterization */
2848 6578 : pathnode->path.param_info = NULL;
2849 6578 : pathnode->path.parallel_aware = false;
2850 14224 : pathnode->path.parallel_safe = rel->consider_parallel &&
2851 7598 : subpath->parallel_safe &&
2852 1020 : is_parallel_safe(root, (Node *) target->exprs);
2853 6578 : pathnode->path.parallel_workers = subpath->parallel_workers;
2854 : /* Projection does not change the sort order XXX? */
2855 6578 : pathnode->path.pathkeys = subpath->pathkeys;
2856 :
2857 6578 : pathnode->subpath = subpath;
2858 :
2859 : /*
2860 : * Estimate number of rows produced by SRFs for each row of input; if
2861 : * there's more than one in this node, use the maximum.
2862 : */
2863 6578 : tlist_rows = 1;
2864 14992 : foreach(lc, target->exprs)
2865 : {
2866 8414 : Node *node = (Node *) lfirst(lc);
2867 : double itemrows;
2868 :
2869 8414 : itemrows = expression_returns_set_rows(root, node);
2870 8414 : if (tlist_rows < itemrows)
2871 6378 : tlist_rows = itemrows;
2872 : }
2873 :
2874 : /*
2875 : * In addition to the cost of evaluating the tlist, charge cpu_tuple_cost
2876 : * per input row, and half of cpu_tuple_cost for each added output row.
2877 : * This is slightly bizarre maybe, but it's what 9.6 did; we may revisit
2878 : * this estimate later.
2879 : */
2880 6578 : pathnode->path.rows = subpath->rows * tlist_rows;
2881 6578 : pathnode->path.startup_cost = subpath->startup_cost +
2882 6578 : target->cost.startup;
2883 6578 : pathnode->path.total_cost = subpath->total_cost +
2884 6578 : target->cost.startup +
2885 6578 : (cpu_tuple_cost + target->cost.per_tuple) * subpath->rows +
2886 6578 : (pathnode->path.rows - subpath->rows) * cpu_tuple_cost / 2;
2887 :
2888 6578 : return pathnode;
2889 : }
2890 :
2891 : /*
2892 : * create_incremental_sort_path
2893 : * Creates a pathnode that represents performing an incremental sort.
2894 : *
2895 : * 'rel' is the parent relation associated with the result
2896 : * 'subpath' is the path representing the source of data
2897 : * 'pathkeys' represents the desired sort order
2898 : * 'presorted_keys' is the number of keys by which the input path is
2899 : * already sorted
2900 : * 'limit_tuples' is the estimated bound on the number of output tuples,
2901 : * or -1 if no LIMIT or couldn't estimate
2902 : */
2903 : IncrementalSortPath *
2904 4058 : create_incremental_sort_path(PlannerInfo *root,
2905 : RelOptInfo *rel,
2906 : Path *subpath,
2907 : List *pathkeys,
2908 : int presorted_keys,
2909 : double limit_tuples)
2910 : {
2911 4058 : IncrementalSortPath *sort = makeNode(IncrementalSortPath);
2912 4058 : SortPath *pathnode = &sort->spath;
2913 :
2914 4058 : pathnode->path.pathtype = T_IncrementalSort;
2915 4058 : pathnode->path.parent = rel;
2916 : /* Sort doesn't project, so use source path's pathtarget */
2917 4058 : pathnode->path.pathtarget = subpath->pathtarget;
2918 : /* For now, assume we are above any joins, so no parameterization */
2919 4058 : pathnode->path.param_info = NULL;
2920 4058 : pathnode->path.parallel_aware = false;
2921 7126 : pathnode->path.parallel_safe = rel->consider_parallel &&
2922 3068 : subpath->parallel_safe;
2923 4058 : pathnode->path.parallel_workers = subpath->parallel_workers;
2924 4058 : pathnode->path.pathkeys = pathkeys;
2925 :
2926 4058 : pathnode->subpath = subpath;
2927 :
2928 4058 : cost_incremental_sort(&pathnode->path,
2929 : root, pathkeys, presorted_keys,
2930 : subpath->startup_cost,
2931 : subpath->total_cost,
2932 : subpath->rows,
2933 4058 : subpath->pathtarget->width,
2934 : 0.0, /* XXX comparison_cost shouldn't be 0? */
2935 : work_mem, limit_tuples);
2936 :
2937 4058 : sort->nPresortedCols = presorted_keys;
2938 :
2939 4058 : return sort;
2940 : }
2941 :
2942 : /*
2943 : * create_sort_path
2944 : * Creates a pathnode that represents performing an explicit sort.
2945 : *
2946 : * 'rel' is the parent relation associated with the result
2947 : * 'subpath' is the path representing the source of data
2948 : * 'pathkeys' represents the desired sort order
2949 : * 'limit_tuples' is the estimated bound on the number of output tuples,
2950 : * or -1 if no LIMIT or couldn't estimate
2951 : */
2952 : SortPath *
2953 59706 : create_sort_path(PlannerInfo *root,
2954 : RelOptInfo *rel,
2955 : Path *subpath,
2956 : List *pathkeys,
2957 : double limit_tuples)
2958 : {
2959 59706 : SortPath *pathnode = makeNode(SortPath);
2960 :
2961 59706 : pathnode->path.pathtype = T_Sort;
2962 59706 : pathnode->path.parent = rel;
2963 : /* Sort doesn't project, so use source path's pathtarget */
2964 59706 : pathnode->path.pathtarget = subpath->pathtarget;
2965 : /* For now, assume we are above any joins, so no parameterization */
2966 59706 : pathnode->path.param_info = NULL;
2967 59706 : pathnode->path.parallel_aware = false;
2968 101310 : pathnode->path.parallel_safe = rel->consider_parallel &&
2969 41604 : subpath->parallel_safe;
2970 59706 : pathnode->path.parallel_workers = subpath->parallel_workers;
2971 59706 : pathnode->path.pathkeys = pathkeys;
2972 :
2973 59706 : pathnode->subpath = subpath;
2974 :
2975 59706 : cost_sort(&pathnode->path, root, pathkeys,
2976 : subpath->total_cost,
2977 : subpath->rows,
2978 59706 : subpath->pathtarget->width,
2979 : 0.0, /* XXX comparison_cost shouldn't be 0? */
2980 : work_mem, limit_tuples);
2981 :
2982 59706 : return pathnode;
2983 : }
2984 :
2985 : /*
2986 : * create_group_path
2987 : * Creates a pathnode that represents performing grouping of presorted input
2988 : *
2989 : * 'rel' is the parent relation associated with the result
2990 : * 'subpath' is the path representing the source of data
2991 : * 'target' is the PathTarget to be computed
2992 : * 'groupClause' is a list of SortGroupClause's representing the grouping
2993 : * 'qual' is the HAVING quals if any
2994 : * 'numGroups' is the estimated number of groups
2995 : */
2996 : GroupPath *
2997 1070 : create_group_path(PlannerInfo *root,
2998 : RelOptInfo *rel,
2999 : Path *subpath,
3000 : List *groupClause,
3001 : List *qual,
3002 : double numGroups)
3003 : {
3004 1070 : GroupPath *pathnode = makeNode(GroupPath);
3005 1070 : PathTarget *target = rel->reltarget;
3006 :
3007 1070 : pathnode->path.pathtype = T_Group;
3008 1070 : pathnode->path.parent = rel;
3009 1070 : pathnode->path.pathtarget = target;
3010 : /* For now, assume we are above any joins, so no parameterization */
3011 1070 : pathnode->path.param_info = NULL;
3012 1070 : pathnode->path.parallel_aware = false;
3013 1758 : pathnode->path.parallel_safe = rel->consider_parallel &&
3014 688 : subpath->parallel_safe;
3015 1070 : pathnode->path.parallel_workers = subpath->parallel_workers;
3016 : /* Group doesn't change sort ordering */
3017 1070 : pathnode->path.pathkeys = subpath->pathkeys;
3018 :
3019 1070 : pathnode->subpath = subpath;
3020 :
3021 1070 : pathnode->groupClause = groupClause;
3022 1070 : pathnode->qual = qual;
3023 :
3024 1070 : cost_group(&pathnode->path, root,
3025 : list_length(groupClause),
3026 : numGroups,
3027 : qual,
3028 : subpath->startup_cost, subpath->total_cost,
3029 : subpath->rows);
3030 :
3031 : /* add tlist eval cost for each output row */
3032 1070 : pathnode->path.startup_cost += target->cost.startup;
3033 1070 : pathnode->path.total_cost += target->cost.startup +
3034 1070 : target->cost.per_tuple * pathnode->path.rows;
3035 :
3036 1070 : return pathnode;
3037 : }
3038 :
3039 : /*
3040 : * create_upper_unique_path
3041 : * Creates a pathnode that represents performing an explicit Unique step
3042 : * on presorted input.
3043 : *
3044 : * This produces a Unique plan node, but the use-case is so different from
3045 : * create_unique_path that it doesn't seem worth trying to merge the two.
3046 : *
3047 : * 'rel' is the parent relation associated with the result
3048 : * 'subpath' is the path representing the source of data
3049 : * 'numCols' is the number of grouping columns
3050 : * 'numGroups' is the estimated number of groups
3051 : *
3052 : * The input path must be sorted on the grouping columns, plus possibly
3053 : * additional columns; so the first numCols pathkeys are the grouping columns
3054 : */
3055 : UpperUniquePath *
3056 3028 : create_upper_unique_path(PlannerInfo *root,
3057 : RelOptInfo *rel,
3058 : Path *subpath,
3059 : int numCols,
3060 : double numGroups)
3061 : {
3062 3028 : UpperUniquePath *pathnode = makeNode(UpperUniquePath);
3063 :
3064 3028 : pathnode->path.pathtype = T_Unique;
3065 3028 : pathnode->path.parent = rel;
3066 : /* Unique doesn't project, so use source path's pathtarget */
3067 3028 : pathnode->path.pathtarget = subpath->pathtarget;
3068 : /* For now, assume we are above any joins, so no parameterization */
3069 3028 : pathnode->path.param_info = NULL;
3070 3028 : pathnode->path.parallel_aware = false;
3071 5578 : pathnode->path.parallel_safe = rel->consider_parallel &&
3072 2550 : subpath->parallel_safe;
3073 3028 : pathnode->path.parallel_workers = subpath->parallel_workers;
3074 : /* Unique doesn't change the input ordering */
3075 3028 : pathnode->path.pathkeys = subpath->pathkeys;
3076 :
3077 3028 : pathnode->subpath = subpath;
3078 3028 : pathnode->numkeys = numCols;
3079 :
3080 : /*
3081 : * Charge one cpu_operator_cost per comparison per input tuple. We assume
3082 : * all columns get compared at most of the tuples. (XXX probably this is
3083 : * an overestimate.)
3084 : */
3085 3028 : pathnode->path.startup_cost = subpath->startup_cost;
3086 3028 : pathnode->path.total_cost = subpath->total_cost +
3087 3028 : cpu_operator_cost * subpath->rows * numCols;
3088 3028 : pathnode->path.rows = numGroups;
3089 :
3090 3028 : return pathnode;
3091 : }
3092 :
3093 : /*
3094 : * create_agg_path
3095 : * Creates a pathnode that represents performing aggregation/grouping
3096 : *
3097 : * 'rel' is the parent relation associated with the result
3098 : * 'subpath' is the path representing the source of data
3099 : * 'target' is the PathTarget to be computed
3100 : * 'aggstrategy' is the Agg node's basic implementation strategy
3101 : * 'aggsplit' is the Agg node's aggregate-splitting mode
3102 : * 'groupClause' is a list of SortGroupClause's representing the grouping
3103 : * 'qual' is the HAVING quals if any
3104 : * 'aggcosts' contains cost info about the aggregate functions to be computed
3105 : * 'numGroups' is the estimated number of groups (1 if not grouping)
3106 : */
3107 : AggPath *
3108 48674 : create_agg_path(PlannerInfo *root,
3109 : RelOptInfo *rel,
3110 : Path *subpath,
3111 : PathTarget *target,
3112 : AggStrategy aggstrategy,
3113 : AggSplit aggsplit,
3114 : List *groupClause,
3115 : List *qual,
3116 : const AggClauseCosts *aggcosts,
3117 : double numGroups)
3118 : {
3119 48674 : AggPath *pathnode = makeNode(AggPath);
3120 :
3121 48674 : pathnode->path.pathtype = T_Agg;
3122 48674 : pathnode->path.parent = rel;
3123 48674 : pathnode->path.pathtarget = target;
3124 : /* For now, assume we are above any joins, so no parameterization */
3125 48674 : pathnode->path.param_info = NULL;
3126 48674 : pathnode->path.parallel_aware = false;
3127 82480 : pathnode->path.parallel_safe = rel->consider_parallel &&
3128 33806 : subpath->parallel_safe;
3129 48674 : pathnode->path.parallel_workers = subpath->parallel_workers;
3130 48674 : if (aggstrategy == AGG_SORTED)
3131 7026 : pathnode->path.pathkeys = subpath->pathkeys; /* preserves order */
3132 : else
3133 41648 : pathnode->path.pathkeys = NIL; /* output is unordered */
3134 48674 : pathnode->subpath = subpath;
3135 :
3136 48674 : pathnode->aggstrategy = aggstrategy;
3137 48674 : pathnode->aggsplit = aggsplit;
3138 48674 : pathnode->numGroups = numGroups;
3139 48674 : pathnode->transitionSpace = aggcosts ? aggcosts->transitionSpace : 0;
3140 48674 : pathnode->groupClause = groupClause;
3141 48674 : pathnode->qual = qual;
3142 :
3143 48674 : cost_agg(&pathnode->path, root,
3144 : aggstrategy, aggcosts,
3145 : list_length(groupClause), numGroups,
3146 : qual,
3147 : subpath->startup_cost, subpath->total_cost,
3148 48674 : subpath->rows, subpath->pathtarget->width);
3149 :
3150 : /* add tlist eval cost for each output row */
3151 48674 : pathnode->path.startup_cost += target->cost.startup;
3152 48674 : pathnode->path.total_cost += target->cost.startup +
3153 48674 : target->cost.per_tuple * pathnode->path.rows;
3154 :
3155 48674 : return pathnode;
3156 : }
3157 :
3158 : /*
3159 : * create_groupingsets_path
3160 : * Creates a pathnode that represents performing GROUPING SETS aggregation
3161 : *
3162 : * GroupingSetsPath represents sorted grouping with one or more grouping sets.
3163 : * The input path's result must be sorted to match the last entry in
3164 : * rollup_groupclauses.
3165 : *
3166 : * 'rel' is the parent relation associated with the result
3167 : * 'subpath' is the path representing the source of data
3168 : * 'target' is the PathTarget to be computed
3169 : * 'having_qual' is the HAVING quals if any
3170 : * 'rollups' is a list of RollupData nodes
3171 : * 'agg_costs' contains cost info about the aggregate functions to be computed
3172 : */
3173 : GroupingSetsPath *
3174 1792 : create_groupingsets_path(PlannerInfo *root,
3175 : RelOptInfo *rel,
3176 : Path *subpath,
3177 : List *having_qual,
3178 : AggStrategy aggstrategy,
3179 : List *rollups,
3180 : const AggClauseCosts *agg_costs)
3181 : {
3182 1792 : GroupingSetsPath *pathnode = makeNode(GroupingSetsPath);
3183 1792 : PathTarget *target = rel->reltarget;
3184 : ListCell *lc;
3185 1792 : bool is_first = true;
3186 1792 : bool is_first_sort = true;
3187 :
3188 : /* The topmost generated Plan node will be an Agg */
3189 1792 : pathnode->path.pathtype = T_Agg;
3190 1792 : pathnode->path.parent = rel;
3191 1792 : pathnode->path.pathtarget = target;
3192 1792 : pathnode->path.param_info = subpath->param_info;
3193 1792 : pathnode->path.parallel_aware = false;
3194 2614 : pathnode->path.parallel_safe = rel->consider_parallel &&
3195 822 : subpath->parallel_safe;
3196 1792 : pathnode->path.parallel_workers = subpath->parallel_workers;
3197 1792 : pathnode->subpath = subpath;
3198 :
3199 : /*
3200 : * Simplify callers by downgrading AGG_SORTED to AGG_PLAIN, and AGG_MIXED
3201 : * to AGG_HASHED, here if possible.
3202 : */
3203 2556 : if (aggstrategy == AGG_SORTED &&
3204 764 : list_length(rollups) == 1 &&
3205 362 : ((RollupData *) linitial(rollups))->groupClause == NIL)
3206 42 : aggstrategy = AGG_PLAIN;
3207 :
3208 2636 : if (aggstrategy == AGG_MIXED &&
3209 844 : list_length(rollups) == 1)
3210 0 : aggstrategy = AGG_HASHED;
3211 :
3212 : /*
3213 : * Output will be in sorted order by group_pathkeys if, and only if, there
3214 : * is a single rollup operation on a non-empty list of grouping
3215 : * expressions.
3216 : */
3217 1792 : if (aggstrategy == AGG_SORTED && list_length(rollups) == 1)
3218 320 : pathnode->path.pathkeys = root->group_pathkeys;
3219 : else
3220 1472 : pathnode->path.pathkeys = NIL;
3221 :
3222 1792 : pathnode->aggstrategy = aggstrategy;
3223 1792 : pathnode->rollups = rollups;
3224 1792 : pathnode->qual = having_qual;
3225 1792 : pathnode->transitionSpace = agg_costs ? agg_costs->transitionSpace : 0;
3226 :
3227 : Assert(rollups != NIL);
3228 : Assert(aggstrategy != AGG_PLAIN || list_length(rollups) == 1);
3229 : Assert(aggstrategy != AGG_MIXED || list_length(rollups) > 1);
3230 :
3231 6396 : foreach(lc, rollups)
3232 : {
3233 4604 : RollupData *rollup = lfirst(lc);
3234 4604 : List *gsets = rollup->gsets;
3235 4604 : int numGroupCols = list_length(linitial(gsets));
3236 :
3237 : /*
3238 : * In AGG_SORTED or AGG_PLAIN mode, the first rollup takes the
3239 : * (already-sorted) input, and following ones do their own sort.
3240 : *
3241 : * In AGG_HASHED mode, there is one rollup for each grouping set.
3242 : *
3243 : * In AGG_MIXED mode, the first rollups are hashed, the first
3244 : * non-hashed one takes the (already-sorted) input, and following ones
3245 : * do their own sort.
3246 : */
3247 4604 : if (is_first)
3248 : {
3249 1792 : cost_agg(&pathnode->path, root,
3250 : aggstrategy,
3251 : agg_costs,
3252 : numGroupCols,
3253 : rollup->numGroups,
3254 : having_qual,
3255 : subpath->startup_cost,
3256 : subpath->total_cost,
3257 : subpath->rows,
3258 1792 : subpath->pathtarget->width);
3259 1792 : is_first = false;
3260 1792 : if (!rollup->is_hashed)
3261 764 : is_first_sort = false;
3262 : }
3263 : else
3264 : {
3265 : Path sort_path; /* dummy for result of cost_sort */
3266 : Path agg_path; /* dummy for result of cost_agg */
3267 :
3268 2812 : if (rollup->is_hashed || is_first_sort)
3269 : {
3270 : /*
3271 : * Account for cost of aggregation, but don't charge input
3272 : * cost again
3273 : */
3274 2158 : cost_agg(&agg_path, root,
3275 2158 : rollup->is_hashed ? AGG_HASHED : AGG_SORTED,
3276 : agg_costs,
3277 : numGroupCols,
3278 : rollup->numGroups,
3279 : having_qual,
3280 : 0.0, 0.0,
3281 : subpath->rows,
3282 2158 : subpath->pathtarget->width);
3283 2158 : if (!rollup->is_hashed)
3284 844 : is_first_sort = false;
3285 : }
3286 : else
3287 : {
3288 : /* Account for cost of sort, but don't charge input cost again */
3289 654 : cost_sort(&sort_path, root, NIL,
3290 : 0.0,
3291 : subpath->rows,
3292 654 : subpath->pathtarget->width,
3293 : 0.0,
3294 : work_mem,
3295 : -1.0);
3296 :
3297 : /* Account for cost of aggregation */
3298 :
3299 654 : cost_agg(&agg_path, root,
3300 : AGG_SORTED,
3301 : agg_costs,
3302 : numGroupCols,
3303 : rollup->numGroups,
3304 : having_qual,
3305 : sort_path.startup_cost,
3306 : sort_path.total_cost,
3307 : sort_path.rows,
3308 654 : subpath->pathtarget->width);
3309 : }
3310 :
3311 2812 : pathnode->path.total_cost += agg_path.total_cost;
3312 2812 : pathnode->path.rows += agg_path.rows;
3313 : }
3314 : }
3315 :
3316 : /* add tlist eval cost for each output row */
3317 1792 : pathnode->path.startup_cost += target->cost.startup;
3318 1792 : pathnode->path.total_cost += target->cost.startup +
3319 1792 : target->cost.per_tuple * pathnode->path.rows;
3320 :
3321 1792 : return pathnode;
3322 : }
3323 :
3324 : /*
3325 : * create_minmaxagg_path
3326 : * Creates a pathnode that represents computation of MIN/MAX aggregates
3327 : *
3328 : * 'rel' is the parent relation associated with the result
3329 : * 'target' is the PathTarget to be computed
3330 : * 'mmaggregates' is a list of MinMaxAggInfo structs
3331 : * 'quals' is the HAVING quals if any
3332 : */
3333 : MinMaxAggPath *
3334 380 : create_minmaxagg_path(PlannerInfo *root,
3335 : RelOptInfo *rel,
3336 : PathTarget *target,
3337 : List *mmaggregates,
3338 : List *quals)
3339 : {
3340 380 : MinMaxAggPath *pathnode = makeNode(MinMaxAggPath);
3341 : Cost initplan_cost;
3342 : ListCell *lc;
3343 :
3344 : /* The topmost generated Plan node will be a Result */
3345 380 : pathnode->path.pathtype = T_Result;
3346 380 : pathnode->path.parent = rel;
3347 380 : pathnode->path.pathtarget = target;
3348 : /* For now, assume we are above any joins, so no parameterization */
3349 380 : pathnode->path.param_info = NULL;
3350 380 : pathnode->path.parallel_aware = false;
3351 : /* A MinMaxAggPath implies use of initplans, so cannot be parallel-safe */
3352 380 : pathnode->path.parallel_safe = false;
3353 380 : pathnode->path.parallel_workers = 0;
3354 : /* Result is one unordered row */
3355 380 : pathnode->path.rows = 1;
3356 380 : pathnode->path.pathkeys = NIL;
3357 :
3358 380 : pathnode->mmaggregates = mmaggregates;
3359 380 : pathnode->quals = quals;
3360 :
3361 : /* Calculate cost of all the initplans ... */
3362 380 : initplan_cost = 0;
3363 796 : foreach(lc, mmaggregates)
3364 : {
3365 416 : MinMaxAggInfo *mminfo = (MinMaxAggInfo *) lfirst(lc);
3366 :
3367 416 : initplan_cost += mminfo->pathcost;
3368 : }
3369 :
3370 : /* add tlist eval cost for each output row, plus cpu_tuple_cost */
3371 380 : pathnode->path.startup_cost = initplan_cost + target->cost.startup;
3372 380 : pathnode->path.total_cost = initplan_cost + target->cost.startup +
3373 380 : target->cost.per_tuple + cpu_tuple_cost;
3374 :
3375 : /*
3376 : * Add cost of qual, if any --- but we ignore its selectivity, since our
3377 : * rowcount estimate should be 1 no matter what the qual is.
3378 : */
3379 380 : if (quals)
3380 : {
3381 : QualCost qual_cost;
3382 :
3383 0 : cost_qual_eval(&qual_cost, quals, root);
3384 0 : pathnode->path.startup_cost += qual_cost.startup;
3385 0 : pathnode->path.total_cost += qual_cost.startup + qual_cost.per_tuple;
3386 : }
3387 :
3388 380 : return pathnode;
3389 : }
3390 :
3391 : /*
3392 : * create_windowagg_path
3393 : * Creates a pathnode that represents computation of window functions
3394 : *
3395 : * 'rel' is the parent relation associated with the result
3396 : * 'subpath' is the path representing the source of data
3397 : * 'target' is the PathTarget to be computed
3398 : * 'windowFuncs' is a list of WindowFunc structs
3399 : * 'winclause' is a WindowClause that is common to all the WindowFuncs
3400 : * 'qual' WindowClause.runconditions from lower-level WindowAggPaths.
3401 : * Must always be NIL when topwindow == false
3402 : * 'topwindow' pass as true only for the top-level WindowAgg. False for all
3403 : * intermediate WindowAggs.
3404 : *
3405 : * The input must be sorted according to the WindowClause's PARTITION keys
3406 : * plus ORDER BY keys.
3407 : */
3408 : WindowAggPath *
3409 2310 : create_windowagg_path(PlannerInfo *root,
3410 : RelOptInfo *rel,
3411 : Path *subpath,
3412 : PathTarget *target,
3413 : List *windowFuncs,
3414 : WindowClause *winclause,
3415 : List *qual,
3416 : bool topwindow)
3417 : {
3418 2310 : WindowAggPath *pathnode = makeNode(WindowAggPath);
3419 :
3420 : /* qual can only be set for the topwindow */
3421 : Assert(qual == NIL || topwindow);
3422 :
3423 2310 : pathnode->path.pathtype = T_WindowAgg;
3424 2310 : pathnode->path.parent = rel;
3425 2310 : pathnode->path.pathtarget = target;
3426 : /* For now, assume we are above any joins, so no parameterization */
3427 2310 : pathnode->path.param_info = NULL;
3428 2310 : pathnode->path.parallel_aware = false;
3429 2310 : pathnode->path.parallel_safe = rel->consider_parallel &&
3430 0 : subpath->parallel_safe;
3431 2310 : pathnode->path.parallel_workers = subpath->parallel_workers;
3432 : /* WindowAgg preserves the input sort order */
3433 2310 : pathnode->path.pathkeys = subpath->pathkeys;
3434 :
3435 2310 : pathnode->subpath = subpath;
3436 2310 : pathnode->winclause = winclause;
3437 2310 : pathnode->qual = qual;
3438 2310 : pathnode->topwindow = topwindow;
3439 :
3440 : /*
3441 : * For costing purposes, assume that there are no redundant partitioning
3442 : * or ordering columns; it's not worth the trouble to deal with that
3443 : * corner case here. So we just pass the unmodified list lengths to
3444 : * cost_windowagg.
3445 : */
3446 2310 : cost_windowagg(&pathnode->path, root,
3447 : windowFuncs,
3448 2310 : list_length(winclause->partitionClause),
3449 2310 : list_length(winclause->orderClause),
3450 : subpath->startup_cost,
3451 : subpath->total_cost,
3452 : subpath->rows);
3453 :
3454 : /* add tlist eval cost for each output row */
3455 2310 : pathnode->path.startup_cost += target->cost.startup;
3456 2310 : pathnode->path.total_cost += target->cost.startup +
3457 2310 : target->cost.per_tuple * pathnode->path.rows;
3458 :
3459 2310 : return pathnode;
3460 : }
3461 :
3462 : /*
3463 : * create_setop_path
3464 : * Creates a pathnode that represents computation of INTERSECT or EXCEPT
3465 : *
3466 : * 'rel' is the parent relation associated with the result
3467 : * 'subpath' is the path representing the source of data
3468 : * 'cmd' is the specific semantics (INTERSECT or EXCEPT, with/without ALL)
3469 : * 'strategy' is the implementation strategy (sorted or hashed)
3470 : * 'distinctList' is a list of SortGroupClause's representing the grouping
3471 : * 'flagColIdx' is the column number where the flag column will be, if any
3472 : * 'firstFlag' is the flag value for the first input relation when hashing;
3473 : * or -1 when sorting
3474 : * 'numGroups' is the estimated number of distinct groups
3475 : * 'outputRows' is the estimated number of output rows
3476 : */
3477 : SetOpPath *
3478 594 : create_setop_path(PlannerInfo *root,
3479 : RelOptInfo *rel,
3480 : Path *subpath,
3481 : SetOpCmd cmd,
3482 : SetOpStrategy strategy,
3483 : List *distinctList,
3484 : AttrNumber flagColIdx,
3485 : int firstFlag,
3486 : double numGroups,
3487 : double outputRows)
3488 : {
3489 594 : SetOpPath *pathnode = makeNode(SetOpPath);
3490 :
3491 594 : pathnode->path.pathtype = T_SetOp;
3492 594 : pathnode->path.parent = rel;
3493 : /* SetOp doesn't project, so use source path's pathtarget */
3494 594 : pathnode->path.pathtarget = subpath->pathtarget;
3495 : /* For now, assume we are above any joins, so no parameterization */
3496 594 : pathnode->path.param_info = NULL;
3497 594 : pathnode->path.parallel_aware = false;
3498 594 : pathnode->path.parallel_safe = rel->consider_parallel &&
3499 0 : subpath->parallel_safe;
3500 594 : pathnode->path.parallel_workers = subpath->parallel_workers;
3501 : /* SetOp preserves the input sort order if in sort mode */
3502 594 : pathnode->path.pathkeys =
3503 594 : (strategy == SETOP_SORTED) ? subpath->pathkeys : NIL;
3504 :
3505 594 : pathnode->subpath = subpath;
3506 594 : pathnode->cmd = cmd;
3507 594 : pathnode->strategy = strategy;
3508 594 : pathnode->distinctList = distinctList;
3509 594 : pathnode->flagColIdx = flagColIdx;
3510 594 : pathnode->firstFlag = firstFlag;
3511 594 : pathnode->numGroups = numGroups;
3512 :
3513 : /*
3514 : * Charge one cpu_operator_cost per comparison per input tuple. We assume
3515 : * all columns get compared at most of the tuples.
3516 : */
3517 594 : pathnode->path.startup_cost = subpath->startup_cost;
3518 1188 : pathnode->path.total_cost = subpath->total_cost +
3519 594 : cpu_operator_cost * subpath->rows * list_length(distinctList);
3520 594 : pathnode->path.rows = outputRows;
3521 :
3522 594 : return pathnode;
3523 : }
3524 :
3525 : /*
3526 : * create_recursiveunion_path
3527 : * Creates a pathnode that represents a recursive UNION node
3528 : *
3529 : * 'rel' is the parent relation associated with the result
3530 : * 'leftpath' is the source of data for the non-recursive term
3531 : * 'rightpath' is the source of data for the recursive term
3532 : * 'target' is the PathTarget to be computed
3533 : * 'distinctList' is a list of SortGroupClause's representing the grouping
3534 : * 'wtParam' is the ID of Param representing work table
3535 : * 'numGroups' is the estimated number of groups
3536 : *
3537 : * For recursive UNION ALL, distinctList is empty and numGroups is zero
3538 : */
3539 : RecursiveUnionPath *
3540 702 : create_recursiveunion_path(PlannerInfo *root,
3541 : RelOptInfo *rel,
3542 : Path *leftpath,
3543 : Path *rightpath,
3544 : PathTarget *target,
3545 : List *distinctList,
3546 : int wtParam,
3547 : double numGroups)
3548 : {
3549 702 : RecursiveUnionPath *pathnode = makeNode(RecursiveUnionPath);
3550 :
3551 702 : pathnode->path.pathtype = T_RecursiveUnion;
3552 702 : pathnode->path.parent = rel;
3553 702 : pathnode->path.pathtarget = target;
3554 : /* For now, assume we are above any joins, so no parameterization */
3555 702 : pathnode->path.param_info = NULL;
3556 702 : pathnode->path.parallel_aware = false;
3557 1404 : pathnode->path.parallel_safe = rel->consider_parallel &&
3558 702 : leftpath->parallel_safe && rightpath->parallel_safe;
3559 : /* Foolish, but we'll do it like joins for now: */
3560 702 : pathnode->path.parallel_workers = leftpath->parallel_workers;
3561 : /* RecursiveUnion result is always unsorted */
3562 702 : pathnode->path.pathkeys = NIL;
3563 :
3564 702 : pathnode->leftpath = leftpath;
3565 702 : pathnode->rightpath = rightpath;
3566 702 : pathnode->distinctList = distinctList;
3567 702 : pathnode->wtParam = wtParam;
3568 702 : pathnode->numGroups = numGroups;
3569 :
3570 702 : cost_recursive_union(&pathnode->path, leftpath, rightpath);
3571 :
3572 702 : return pathnode;
3573 : }
3574 :
3575 : /*
3576 : * create_lockrows_path
3577 : * Creates a pathnode that represents acquiring row locks
3578 : *
3579 : * 'rel' is the parent relation associated with the result
3580 : * 'subpath' is the path representing the source of data
3581 : * 'rowMarks' is a list of PlanRowMark's
3582 : * 'epqParam' is the ID of Param for EvalPlanQual re-eval
3583 : */
3584 : LockRowsPath *
3585 7178 : create_lockrows_path(PlannerInfo *root, RelOptInfo *rel,
3586 : Path *subpath, List *rowMarks, int epqParam)
3587 : {
3588 7178 : LockRowsPath *pathnode = makeNode(LockRowsPath);
3589 :
3590 7178 : pathnode->path.pathtype = T_LockRows;
3591 7178 : pathnode->path.parent = rel;
3592 : /* LockRows doesn't project, so use source path's pathtarget */
3593 7178 : pathnode->path.pathtarget = subpath->pathtarget;
3594 : /* For now, assume we are above any joins, so no parameterization */
3595 7178 : pathnode->path.param_info = NULL;
3596 7178 : pathnode->path.parallel_aware = false;
3597 7178 : pathnode->path.parallel_safe = false;
3598 7178 : pathnode->path.parallel_workers = 0;
3599 7178 : pathnode->path.rows = subpath->rows;
3600 :
3601 : /*
3602 : * The result cannot be assumed sorted, since locking might cause the sort
3603 : * key columns to be replaced with new values.
3604 : */
3605 7178 : pathnode->path.pathkeys = NIL;
3606 :
3607 7178 : pathnode->subpath = subpath;
3608 7178 : pathnode->rowMarks = rowMarks;
3609 7178 : pathnode->epqParam = epqParam;
3610 :
3611 : /*
3612 : * We should charge something extra for the costs of row locking and
3613 : * possible refetches, but it's hard to say how much. For now, use
3614 : * cpu_tuple_cost per row.
3615 : */
3616 7178 : pathnode->path.startup_cost = subpath->startup_cost;
3617 7178 : pathnode->path.total_cost = subpath->total_cost +
3618 7178 : cpu_tuple_cost * subpath->rows;
3619 :
3620 7178 : return pathnode;
3621 : }
3622 :
3623 : /*
3624 : * create_modifytable_path
3625 : * Creates a pathnode that represents performing INSERT/UPDATE/DELETE/MERGE
3626 : * mods
3627 : *
3628 : * 'rel' is the parent relation associated with the result
3629 : * 'subpath' is a Path producing source data
3630 : * 'operation' is the operation type
3631 : * 'canSetTag' is true if we set the command tag/es_processed
3632 : * 'nominalRelation' is the parent RT index for use of EXPLAIN
3633 : * 'rootRelation' is the partitioned table root RT index, or 0 if none
3634 : * 'partColsUpdated' is true if any partitioning columns are being updated,
3635 : * either from the target relation or a descendent partitioned table.
3636 : * 'resultRelations' is an integer list of actual RT indexes of target rel(s)
3637 : * 'updateColnosLists' is a list of UPDATE target column number lists
3638 : * (one sublist per rel); or NIL if not an UPDATE
3639 : * 'withCheckOptionLists' is a list of WCO lists (one per rel)
3640 : * 'returningLists' is a list of RETURNING tlists (one per rel)
3641 : * 'rowMarks' is a list of PlanRowMarks (non-locking only)
3642 : * 'onconflict' is the ON CONFLICT clause, or NULL
3643 : * 'epqParam' is the ID of Param for EvalPlanQual re-eval
3644 : * 'mergeActionLists' is a list of lists of MERGE actions (one per rel)
3645 : */
3646 : ModifyTablePath *
3647 107736 : create_modifytable_path(PlannerInfo *root, RelOptInfo *rel,
3648 : Path *subpath,
3649 : CmdType operation, bool canSetTag,
3650 : Index nominalRelation, Index rootRelation,
3651 : bool partColsUpdated,
3652 : List *resultRelations,
3653 : List *updateColnosLists,
3654 : List *withCheckOptionLists, List *returningLists,
3655 : List *rowMarks, OnConflictExpr *onconflict,
3656 : List *mergeActionLists, int epqParam)
3657 : {
3658 107736 : ModifyTablePath *pathnode = makeNode(ModifyTablePath);
3659 :
3660 : Assert(operation == CMD_MERGE ||
3661 : (operation == CMD_UPDATE ?
3662 : list_length(resultRelations) == list_length(updateColnosLists) :
3663 : updateColnosLists == NIL));
3664 : Assert(withCheckOptionLists == NIL ||
3665 : list_length(resultRelations) == list_length(withCheckOptionLists));
3666 : Assert(returningLists == NIL ||
3667 : list_length(resultRelations) == list_length(returningLists));
3668 :
3669 107736 : pathnode->path.pathtype = T_ModifyTable;
3670 107736 : pathnode->path.parent = rel;
3671 : /* pathtarget is not interesting, just make it minimally valid */
3672 107736 : pathnode->path.pathtarget = rel->reltarget;
3673 : /* For now, assume we are above any joins, so no parameterization */
3674 107736 : pathnode->path.param_info = NULL;
3675 107736 : pathnode->path.parallel_aware = false;
3676 107736 : pathnode->path.parallel_safe = false;
3677 107736 : pathnode->path.parallel_workers = 0;
3678 107736 : pathnode->path.pathkeys = NIL;
3679 :
3680 : /*
3681 : * Compute cost & rowcount as subpath cost & rowcount (if RETURNING)
3682 : *
3683 : * Currently, we don't charge anything extra for the actual table
3684 : * modification work, nor for the WITH CHECK OPTIONS or RETURNING
3685 : * expressions if any. It would only be window dressing, since
3686 : * ModifyTable is always a top-level node and there is no way for the
3687 : * costs to change any higher-level planning choices. But we might want
3688 : * to make it look better sometime.
3689 : */
3690 107736 : pathnode->path.startup_cost = subpath->startup_cost;
3691 107736 : pathnode->path.total_cost = subpath->total_cost;
3692 107736 : if (returningLists != NIL)
3693 : {
3694 2256 : pathnode->path.rows = subpath->rows;
3695 :
3696 : /*
3697 : * Set width to match the subpath output. XXX this is totally wrong:
3698 : * we should return an average of the RETURNING tlist widths. But
3699 : * it's what happened historically, and improving it is a task for
3700 : * another day. (Again, it's mostly window dressing.)
3701 : */
3702 2256 : pathnode->path.pathtarget->width = subpath->pathtarget->width;
3703 : }
3704 : else
3705 : {
3706 105480 : pathnode->path.rows = 0;
3707 105480 : pathnode->path.pathtarget->width = 0;
3708 : }
3709 :
3710 107736 : pathnode->subpath = subpath;
3711 107736 : pathnode->operation = operation;
3712 107736 : pathnode->canSetTag = canSetTag;
3713 107736 : pathnode->nominalRelation = nominalRelation;
3714 107736 : pathnode->rootRelation = rootRelation;
3715 107736 : pathnode->partColsUpdated = partColsUpdated;
3716 107736 : pathnode->resultRelations = resultRelations;
3717 107736 : pathnode->updateColnosLists = updateColnosLists;
3718 107736 : pathnode->withCheckOptionLists = withCheckOptionLists;
3719 107736 : pathnode->returningLists = returningLists;
3720 107736 : pathnode->rowMarks = rowMarks;
3721 107736 : pathnode->onconflict = onconflict;
3722 107736 : pathnode->epqParam = epqParam;
3723 107736 : pathnode->mergeActionLists = mergeActionLists;
3724 :
3725 107736 : return pathnode;
3726 : }
3727 :
3728 : /*
3729 : * create_limit_path
3730 : * Creates a pathnode that represents performing LIMIT/OFFSET
3731 : *
3732 : * In addition to providing the actual OFFSET and LIMIT expressions,
3733 : * the caller must provide estimates of their values for costing purposes.
3734 : * The estimates are as computed by preprocess_limit(), ie, 0 represents
3735 : * the clause not being present, and -1 means it's present but we could
3736 : * not estimate its value.
3737 : *
3738 : * 'rel' is the parent relation associated with the result
3739 : * 'subpath' is the path representing the source of data
3740 : * 'limitOffset' is the actual OFFSET expression, or NULL
3741 : * 'limitCount' is the actual LIMIT expression, or NULL
3742 : * 'offset_est' is the estimated value of the OFFSET expression
3743 : * 'count_est' is the estimated value of the LIMIT expression
3744 : */
3745 : LimitPath *
3746 5712 : create_limit_path(PlannerInfo *root, RelOptInfo *rel,
3747 : Path *subpath,
3748 : Node *limitOffset, Node *limitCount,
3749 : LimitOption limitOption,
3750 : int64 offset_est, int64 count_est)
3751 : {
3752 5712 : LimitPath *pathnode = makeNode(LimitPath);
3753 :
3754 5712 : pathnode->path.pathtype = T_Limit;
3755 5712 : pathnode->path.parent = rel;
3756 : /* Limit doesn't project, so use source path's pathtarget */
3757 5712 : pathnode->path.pathtarget = subpath->pathtarget;
3758 : /* For now, assume we are above any joins, so no parameterization */
3759 5712 : pathnode->path.param_info = NULL;
3760 5712 : pathnode->path.parallel_aware = false;
3761 7722 : pathnode->path.parallel_safe = rel->consider_parallel &&
3762 2010 : subpath->parallel_safe;
3763 5712 : pathnode->path.parallel_workers = subpath->parallel_workers;
3764 5712 : pathnode->path.rows = subpath->rows;
3765 5712 : pathnode->path.startup_cost = subpath->startup_cost;
3766 5712 : pathnode->path.total_cost = subpath->total_cost;
3767 5712 : pathnode->path.pathkeys = subpath->pathkeys;
3768 5712 : pathnode->subpath = subpath;
3769 5712 : pathnode->limitOffset = limitOffset;
3770 5712 : pathnode->limitCount = limitCount;
3771 5712 : pathnode->limitOption = limitOption;
3772 :
3773 : /*
3774 : * Adjust the output rows count and costs according to the offset/limit.
3775 : */
3776 5712 : adjust_limit_rows_costs(&pathnode->path.rows,
3777 : &pathnode->path.startup_cost,
3778 : &pathnode->path.total_cost,
3779 : offset_est, count_est);
3780 :
3781 5712 : return pathnode;
3782 : }
3783 :
3784 : /*
3785 : * adjust_limit_rows_costs
3786 : * Adjust the size and cost estimates for a LimitPath node according to the
3787 : * offset/limit.
3788 : *
3789 : * This is only a cosmetic issue if we are at top level, but if we are
3790 : * building a subquery then it's important to report correct info to the outer
3791 : * planner.
3792 : *
3793 : * When the offset or count couldn't be estimated, use 10% of the estimated
3794 : * number of rows emitted from the subpath.
3795 : *
3796 : * XXX we don't bother to add eval costs of the offset/limit expressions
3797 : * themselves to the path costs. In theory we should, but in most cases those
3798 : * expressions are trivial and it's just not worth the trouble.
3799 : */
3800 : void
3801 5894 : adjust_limit_rows_costs(double *rows, /* in/out parameter */
3802 : Cost *startup_cost, /* in/out parameter */
3803 : Cost *total_cost, /* in/out parameter */
3804 : int64 offset_est,
3805 : int64 count_est)
3806 : {
3807 5894 : double input_rows = *rows;
3808 5894 : Cost input_startup_cost = *startup_cost;
3809 5894 : Cost input_total_cost = *total_cost;
3810 :
3811 5894 : if (offset_est != 0)
3812 : {
3813 : double offset_rows;
3814 :
3815 670 : if (offset_est > 0)
3816 646 : offset_rows = (double) offset_est;
3817 : else
3818 24 : offset_rows = clamp_row_est(input_rows * 0.10);
3819 670 : if (offset_rows > *rows)
3820 26 : offset_rows = *rows;
3821 670 : if (input_rows > 0)
3822 670 : *startup_cost +=
3823 670 : (input_total_cost - input_startup_cost)
3824 670 : * offset_rows / input_rows;
3825 670 : *rows -= offset_rows;
3826 670 : if (*rows < 1)
3827 26 : *rows = 1;
3828 : }
3829 :
3830 5894 : if (count_est != 0)
3831 : {
3832 : double count_rows;
3833 :
3834 5846 : if (count_est > 0)
3835 5840 : count_rows = (double) count_est;
3836 : else
3837 6 : count_rows = clamp_row_est(input_rows * 0.10);
3838 5846 : if (count_rows > *rows)
3839 240 : count_rows = *rows;
3840 5846 : if (input_rows > 0)
3841 5846 : *total_cost = *startup_cost +
3842 5846 : (input_total_cost - input_startup_cost)
3843 5846 : * count_rows / input_rows;
3844 5846 : *rows = count_rows;
3845 5846 : if (*rows < 1)
3846 0 : *rows = 1;
3847 : }
3848 5894 : }
3849 :
3850 :
3851 : /*
3852 : * reparameterize_path
3853 : * Attempt to modify a Path to have greater parameterization
3854 : *
3855 : * We use this to attempt to bring all child paths of an appendrel to the
3856 : * same parameterization level, ensuring that they all enforce the same set
3857 : * of join quals (and thus that that parameterization can be attributed to
3858 : * an append path built from such paths). Currently, only a few path types
3859 : * are supported here, though more could be added at need. We return NULL
3860 : * if we can't reparameterize the given path.
3861 : *
3862 : * Note: we intentionally do not pass created paths to add_path(); it would
3863 : * possibly try to delete them on the grounds of being cost-inferior to the
3864 : * paths they were made from, and we don't want that. Paths made here are
3865 : * not necessarily of general-purpose usefulness, but they can be useful
3866 : * as members of an append path.
3867 : */
3868 : Path *
3869 308 : reparameterize_path(PlannerInfo *root, Path *path,
3870 : Relids required_outer,
3871 : double loop_count)
3872 : {
3873 308 : RelOptInfo *rel = path->parent;
3874 :
3875 : /* Can only increase, not decrease, path's parameterization */
3876 308 : if (!bms_is_subset(PATH_REQ_OUTER(path), required_outer))
3877 0 : return NULL;
3878 308 : switch (path->pathtype)
3879 : {
3880 228 : case T_SeqScan:
3881 228 : return create_seqscan_path(root, rel, required_outer, 0);
3882 0 : case T_SampleScan:
3883 0 : return (Path *) create_samplescan_path(root, rel, required_outer);
3884 0 : case T_IndexScan:
3885 : case T_IndexOnlyScan:
3886 : {
3887 0 : IndexPath *ipath = (IndexPath *) path;
3888 0 : IndexPath *newpath = makeNode(IndexPath);
3889 :
3890 : /*
3891 : * We can't use create_index_path directly, and would not want
3892 : * to because it would re-compute the indexqual conditions
3893 : * which is wasted effort. Instead we hack things a bit:
3894 : * flat-copy the path node, revise its param_info, and redo
3895 : * the cost estimate.
3896 : */
3897 0 : memcpy(newpath, ipath, sizeof(IndexPath));
3898 0 : newpath->path.param_info =
3899 0 : get_baserel_parampathinfo(root, rel, required_outer);
3900 0 : cost_index(newpath, root, loop_count, false);
3901 0 : return (Path *) newpath;
3902 : }
3903 0 : case T_BitmapHeapScan:
3904 : {
3905 0 : BitmapHeapPath *bpath = (BitmapHeapPath *) path;
3906 :
3907 0 : return (Path *) create_bitmap_heap_path(root,
3908 : rel,
3909 : bpath->bitmapqual,
3910 : required_outer,
3911 : loop_count, 0);
3912 : }
3913 0 : case T_SubqueryScan:
3914 : {
3915 0 : SubqueryScanPath *spath = (SubqueryScanPath *) path;
3916 0 : Path *subpath = spath->subpath;
3917 : bool trivial_pathtarget;
3918 :
3919 : /*
3920 : * If existing node has zero extra cost, we must have decided
3921 : * its target is trivial. (The converse is not true, because
3922 : * it might have a trivial target but quals to enforce; but in
3923 : * that case the new node will too, so it doesn't matter
3924 : * whether we get the right answer here.)
3925 : */
3926 0 : trivial_pathtarget =
3927 0 : (subpath->total_cost == spath->path.total_cost);
3928 :
3929 0 : return (Path *) create_subqueryscan_path(root,
3930 : rel,
3931 : subpath,
3932 : trivial_pathtarget,
3933 : spath->path.pathkeys,
3934 : required_outer);
3935 : }
3936 48 : case T_Result:
3937 : /* Supported only for RTE_RESULT scan paths */
3938 48 : if (IsA(path, Path))
3939 48 : return create_resultscan_path(root, rel, required_outer);
3940 0 : break;
3941 0 : case T_Append:
3942 : {
3943 0 : AppendPath *apath = (AppendPath *) path;
3944 0 : List *childpaths = NIL;
3945 0 : List *partialpaths = NIL;
3946 : int i;
3947 : ListCell *lc;
3948 :
3949 : /* Reparameterize the children */
3950 0 : i = 0;
3951 0 : foreach(lc, apath->subpaths)
3952 : {
3953 0 : Path *spath = (Path *) lfirst(lc);
3954 :
3955 0 : spath = reparameterize_path(root, spath,
3956 : required_outer,
3957 : loop_count);
3958 0 : if (spath == NULL)
3959 0 : return NULL;
3960 : /* We have to re-split the regular and partial paths */
3961 0 : if (i < apath->first_partial_path)
3962 0 : childpaths = lappend(childpaths, spath);
3963 : else
3964 0 : partialpaths = lappend(partialpaths, spath);
3965 0 : i++;
3966 : }
3967 0 : return (Path *)
3968 0 : create_append_path(root, rel, childpaths, partialpaths,
3969 : apath->path.pathkeys, required_outer,
3970 : apath->path.parallel_workers,
3971 0 : apath->path.parallel_aware,
3972 : -1);
3973 : }
3974 0 : case T_Material:
3975 : {
3976 0 : MaterialPath *mpath = (MaterialPath *) path;
3977 0 : Path *spath = mpath->subpath;
3978 :
3979 0 : spath = reparameterize_path(root, spath,
3980 : required_outer,
3981 : loop_count);
3982 0 : if (spath == NULL)
3983 0 : return NULL;
3984 0 : return (Path *) create_material_path(rel, spath);
3985 : }
3986 0 : case T_Memoize:
3987 : {
3988 0 : MemoizePath *mpath = (MemoizePath *) path;
3989 0 : Path *spath = mpath->subpath;
3990 :
3991 0 : spath = reparameterize_path(root, spath,
3992 : required_outer,
3993 : loop_count);
3994 0 : if (spath == NULL)
3995 0 : return NULL;
3996 0 : return (Path *) create_memoize_path(root, rel,
3997 : spath,
3998 : mpath->param_exprs,
3999 : mpath->hash_operators,
4000 0 : mpath->singlerow,
4001 0 : mpath->binary_mode,
4002 : mpath->calls);
4003 : }
4004 32 : default:
4005 32 : break;
4006 : }
4007 32 : return NULL;
4008 : }
4009 :
4010 : /*
4011 : * reparameterize_path_by_child
4012 : * Given a path parameterized by the parent of the given child relation,
4013 : * translate the path to be parameterized by the given child relation.
4014 : *
4015 : * The function creates a new path of the same type as the given path, but
4016 : * parameterized by the given child relation. Most fields from the original
4017 : * path can simply be flat-copied, but any expressions must be adjusted to
4018 : * refer to the correct varnos, and any paths must be recursively
4019 : * reparameterized. Other fields that refer to specific relids also need
4020 : * adjustment.
4021 : *
4022 : * The cost, number of rows, width and parallel path properties depend upon
4023 : * path->parent, which does not change during the translation. Hence those
4024 : * members are copied as they are.
4025 : *
4026 : * Currently, only a few path types are supported here, though more could be
4027 : * added at need. We return NULL if we can't reparameterize the given path.
4028 : */
4029 : Path *
4030 8404 : reparameterize_path_by_child(PlannerInfo *root, Path *path,
4031 : RelOptInfo *child_rel)
4032 : {
4033 :
4034 : #define FLAT_COPY_PATH(newnode, node, nodetype) \
4035 : ( (newnode) = makeNode(nodetype), \
4036 : memcpy((newnode), (node), sizeof(nodetype)) )
4037 :
4038 : #define ADJUST_CHILD_ATTRS(node) \
4039 : ((node) = \
4040 : (List *) adjust_appendrel_attrs_multilevel(root, (Node *) (node), \
4041 : child_rel, \
4042 : child_rel->top_parent))
4043 :
4044 : #define REPARAMETERIZE_CHILD_PATH(path) \
4045 : do { \
4046 : (path) = reparameterize_path_by_child(root, (path), child_rel); \
4047 : if ((path) == NULL) \
4048 : return NULL; \
4049 : } while(0)
4050 :
4051 : #define REPARAMETERIZE_CHILD_PATH_LIST(pathlist) \
4052 : do { \
4053 : if ((pathlist) != NIL) \
4054 : { \
4055 : (pathlist) = reparameterize_pathlist_by_child(root, (pathlist), \
4056 : child_rel); \
4057 : if ((pathlist) == NIL) \
4058 : return NULL; \
4059 : } \
4060 : } while(0)
4061 :
4062 : Path *new_path;
4063 : ParamPathInfo *new_ppi;
4064 : ParamPathInfo *old_ppi;
4065 : Relids required_outer;
4066 :
4067 : /*
4068 : * If the path is not parameterized by parent of the given relation, it
4069 : * doesn't need reparameterization.
4070 : */
4071 8404 : if (!path->param_info ||
4072 8320 : !bms_overlap(PATH_REQ_OUTER(path), child_rel->top_parent_relids))
4073 240 : return path;
4074 :
4075 : /*
4076 : * If possible, reparameterize the given path, making a copy.
4077 : *
4078 : * This function is currently only applied to the inner side of a nestloop
4079 : * join that is being partitioned by the partitionwise-join code. Hence,
4080 : * we need only support path types that plausibly arise in that context.
4081 : * (In particular, supporting sorted path types would be a waste of code
4082 : * and cycles: even if we translated them here, they'd just lose in
4083 : * subsequent cost comparisons.) If we do see an unsupported path type,
4084 : * that just means we won't be able to generate a partitionwise-join plan
4085 : * using that path type.
4086 : */
4087 8164 : switch (nodeTag(path))
4088 : {
4089 432 : case T_Path:
4090 432 : FLAT_COPY_PATH(new_path, path, Path);
4091 432 : break;
4092 :
4093 5040 : case T_IndexPath:
4094 : {
4095 : IndexPath *ipath;
4096 :
4097 5040 : FLAT_COPY_PATH(ipath, path, IndexPath);
4098 5040 : ADJUST_CHILD_ATTRS(ipath->indexclauses);
4099 5040 : new_path = (Path *) ipath;
4100 : }
4101 5040 : break;
4102 :
4103 48 : case T_BitmapHeapPath:
4104 : {
4105 : BitmapHeapPath *bhpath;
4106 :
4107 48 : FLAT_COPY_PATH(bhpath, path, BitmapHeapPath);
4108 48 : REPARAMETERIZE_CHILD_PATH(bhpath->bitmapqual);
4109 48 : new_path = (Path *) bhpath;
4110 : }
4111 48 : break;
4112 :
4113 24 : case T_BitmapAndPath:
4114 : {
4115 : BitmapAndPath *bapath;
4116 :
4117 24 : FLAT_COPY_PATH(bapath, path, BitmapAndPath);
4118 24 : REPARAMETERIZE_CHILD_PATH_LIST(bapath->bitmapquals);
4119 24 : new_path = (Path *) bapath;
4120 : }
4121 24 : break;
4122 :
4123 24 : case T_BitmapOrPath:
4124 : {
4125 : BitmapOrPath *bopath;
4126 :
4127 24 : FLAT_COPY_PATH(bopath, path, BitmapOrPath);
4128 24 : REPARAMETERIZE_CHILD_PATH_LIST(bopath->bitmapquals);
4129 24 : new_path = (Path *) bopath;
4130 : }
4131 24 : break;
4132 :
4133 52 : case T_ForeignPath:
4134 : {
4135 : ForeignPath *fpath;
4136 : ReparameterizeForeignPathByChild_function rfpc_func;
4137 :
4138 52 : FLAT_COPY_PATH(fpath, path, ForeignPath);
4139 52 : if (fpath->fdw_outerpath)
4140 0 : REPARAMETERIZE_CHILD_PATH(fpath->fdw_outerpath);
4141 :
4142 : /* Hand over to FDW if needed. */
4143 52 : rfpc_func =
4144 52 : path->parent->fdwroutine->ReparameterizeForeignPathByChild;
4145 52 : if (rfpc_func)
4146 0 : fpath->fdw_private = rfpc_func(root, fpath->fdw_private,
4147 : child_rel);
4148 52 : new_path = (Path *) fpath;
4149 : }
4150 52 : break;
4151 :
4152 0 : case T_CustomPath:
4153 : {
4154 : CustomPath *cpath;
4155 :
4156 0 : FLAT_COPY_PATH(cpath, path, CustomPath);
4157 0 : REPARAMETERIZE_CHILD_PATH_LIST(cpath->custom_paths);
4158 0 : if (cpath->methods &&
4159 0 : cpath->methods->ReparameterizeCustomPathByChild)
4160 0 : cpath->custom_private =
4161 0 : cpath->methods->ReparameterizeCustomPathByChild(root,
4162 : cpath->custom_private,
4163 : child_rel);
4164 0 : new_path = (Path *) cpath;
4165 : }
4166 0 : break;
4167 :
4168 300 : case T_NestPath:
4169 : {
4170 : JoinPath *jpath;
4171 : NestPath *npath;
4172 :
4173 300 : FLAT_COPY_PATH(npath, path, NestPath);
4174 :
4175 300 : jpath = (JoinPath *) npath;
4176 300 : REPARAMETERIZE_CHILD_PATH(jpath->outerjoinpath);
4177 300 : REPARAMETERIZE_CHILD_PATH(jpath->innerjoinpath);
4178 300 : ADJUST_CHILD_ATTRS(jpath->joinrestrictinfo);
4179 300 : new_path = (Path *) npath;
4180 : }
4181 300 : break;
4182 :
4183 36 : case T_MergePath:
4184 : {
4185 : JoinPath *jpath;
4186 : MergePath *mpath;
4187 :
4188 36 : FLAT_COPY_PATH(mpath, path, MergePath);
4189 :
4190 36 : jpath = (JoinPath *) mpath;
4191 36 : REPARAMETERIZE_CHILD_PATH(jpath->outerjoinpath);
4192 36 : REPARAMETERIZE_CHILD_PATH(jpath->innerjoinpath);
4193 36 : ADJUST_CHILD_ATTRS(jpath->joinrestrictinfo);
4194 36 : ADJUST_CHILD_ATTRS(mpath->path_mergeclauses);
4195 36 : new_path = (Path *) mpath;
4196 : }
4197 36 : break;
4198 :
4199 168 : case T_HashPath:
4200 : {
4201 : JoinPath *jpath;
4202 : HashPath *hpath;
4203 :
4204 168 : FLAT_COPY_PATH(hpath, path, HashPath);
4205 :
4206 168 : jpath = (JoinPath *) hpath;
4207 168 : REPARAMETERIZE_CHILD_PATH(jpath->outerjoinpath);
4208 168 : REPARAMETERIZE_CHILD_PATH(jpath->innerjoinpath);
4209 168 : ADJUST_CHILD_ATTRS(jpath->joinrestrictinfo);
4210 168 : ADJUST_CHILD_ATTRS(hpath->path_hashclauses);
4211 168 : new_path = (Path *) hpath;
4212 : }
4213 168 : break;
4214 :
4215 120 : case T_AppendPath:
4216 : {
4217 : AppendPath *apath;
4218 :
4219 120 : FLAT_COPY_PATH(apath, path, AppendPath);
4220 120 : REPARAMETERIZE_CHILD_PATH_LIST(apath->subpaths);
4221 120 : new_path = (Path *) apath;
4222 : }
4223 120 : break;
4224 :
4225 0 : case T_MaterialPath:
4226 : {
4227 : MaterialPath *mpath;
4228 :
4229 0 : FLAT_COPY_PATH(mpath, path, MaterialPath);
4230 0 : REPARAMETERIZE_CHILD_PATH(mpath->subpath);
4231 0 : new_path = (Path *) mpath;
4232 : }
4233 0 : break;
4234 :
4235 1920 : case T_MemoizePath:
4236 : {
4237 : MemoizePath *mpath;
4238 :
4239 1920 : FLAT_COPY_PATH(mpath, path, MemoizePath);
4240 1920 : REPARAMETERIZE_CHILD_PATH(mpath->subpath);
4241 1920 : ADJUST_CHILD_ATTRS(mpath->param_exprs);
4242 1920 : new_path = (Path *) mpath;
4243 : }
4244 1920 : break;
4245 :
4246 0 : case T_GatherPath:
4247 : {
4248 : GatherPath *gpath;
4249 :
4250 0 : FLAT_COPY_PATH(gpath, path, GatherPath);
4251 0 : REPARAMETERIZE_CHILD_PATH(gpath->subpath);
4252 0 : new_path = (Path *) gpath;
4253 : }
4254 0 : break;
4255 :
4256 0 : default:
4257 :
4258 : /* We don't know how to reparameterize this path. */
4259 0 : return NULL;
4260 : }
4261 :
4262 : /*
4263 : * Adjust the parameterization information, which refers to the topmost
4264 : * parent. The topmost parent can be multiple levels away from the given
4265 : * child, hence use multi-level expression adjustment routines.
4266 : */
4267 8164 : old_ppi = new_path->param_info;
4268 : required_outer =
4269 8164 : adjust_child_relids_multilevel(root, old_ppi->ppi_req_outer,
4270 : child_rel,
4271 8164 : child_rel->top_parent);
4272 :
4273 : /* If we already have a PPI for this parameterization, just return it */
4274 8164 : new_ppi = find_param_path_info(new_path->parent, required_outer);
4275 :
4276 : /*
4277 : * If not, build a new one and link it to the list of PPIs. For the same
4278 : * reason as explained in mark_dummy_rel(), allocate new PPI in the same
4279 : * context the given RelOptInfo is in.
4280 : */
4281 8164 : if (new_ppi == NULL)
4282 : {
4283 : MemoryContext oldcontext;
4284 2010 : RelOptInfo *rel = path->parent;
4285 :
4286 2010 : oldcontext = MemoryContextSwitchTo(GetMemoryChunkContext(rel));
4287 :
4288 2010 : new_ppi = makeNode(ParamPathInfo);
4289 2010 : new_ppi->ppi_req_outer = bms_copy(required_outer);
4290 2010 : new_ppi->ppi_rows = old_ppi->ppi_rows;
4291 2010 : new_ppi->ppi_clauses = old_ppi->ppi_clauses;
4292 2010 : ADJUST_CHILD_ATTRS(new_ppi->ppi_clauses);
4293 2010 : new_ppi->ppi_serials = bms_copy(old_ppi->ppi_serials);
4294 2010 : rel->ppilist = lappend(rel->ppilist, new_ppi);
4295 :
4296 2010 : MemoryContextSwitchTo(oldcontext);
4297 : }
4298 8164 : bms_free(required_outer);
4299 :
4300 8164 : new_path->param_info = new_ppi;
4301 :
4302 : /*
4303 : * Adjust the path target if the parent of the outer relation is
4304 : * referenced in the targetlist. This can happen when only the parent of
4305 : * outer relation is laterally referenced in this relation.
4306 : */
4307 8164 : if (bms_overlap(path->parent->lateral_relids,
4308 8164 : child_rel->top_parent_relids))
4309 : {
4310 1152 : new_path->pathtarget = copy_pathtarget(new_path->pathtarget);
4311 1152 : ADJUST_CHILD_ATTRS(new_path->pathtarget->exprs);
4312 : }
4313 :
4314 8164 : return new_path;
4315 : }
4316 :
4317 : /*
4318 : * reparameterize_pathlist_by_child
4319 : * Helper function to reparameterize a list of paths by given child rel.
4320 : */
4321 : static List *
4322 168 : reparameterize_pathlist_by_child(PlannerInfo *root,
4323 : List *pathlist,
4324 : RelOptInfo *child_rel)
4325 : {
4326 : ListCell *lc;
4327 168 : List *result = NIL;
4328 :
4329 504 : foreach(lc, pathlist)
4330 : {
4331 336 : Path *path = reparameterize_path_by_child(root, lfirst(lc),
4332 : child_rel);
4333 :
4334 336 : if (path == NULL)
4335 : {
4336 0 : list_free(result);
4337 0 : return NIL;
4338 : }
4339 :
4340 336 : result = lappend(result, path);
4341 : }
4342 :
4343 168 : return result;
4344 : }
|
