1 /*-------------------------------------------------------------------------
2 *
3 * pathnode.c
4 * Routines to manipulate pathlists and create path nodes
5 *
6 * Portions Copyright (c) 1996-2016, 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 "miscadmin.h"
20 #include "nodes/nodeFuncs.h"
21 #include "optimizer/clauses.h"
22 #include "optimizer/cost.h"
23 #include "optimizer/pathnode.h"
24 #include "optimizer/paths.h"
25 #include "optimizer/planmain.h"
26 #include "optimizer/restrictinfo.h"
27 #include "optimizer/var.h"
28 #include "parser/parsetree.h"
29 #include "utils/lsyscache.h"
30 #include "utils/selfuncs.h"
31
32
33 typedef enum
34 {
35 COSTS_EQUAL, /* path costs are fuzzily equal */
36 COSTS_BETTER1, /* first path is cheaper than second */
37 COSTS_BETTER2, /* second path is cheaper than first */
38 COSTS_DIFFERENT /* neither path dominates the other on cost */
39 } PathCostComparison;
40
41 /*
42 * STD_FUZZ_FACTOR is the normal fuzz factor for compare_path_costs_fuzzily.
43 * XXX is it worth making this user-controllable? It provides a tradeoff
44 * between planner runtime and the accuracy of path cost comparisons.
45 */
46 #define STD_FUZZ_FACTOR 1.01
47
48 static List *translate_sub_tlist(List *tlist, int relid);
49
50
51 /*****************************************************************************
52 * MISC. PATH UTILITIES
53 *****************************************************************************/
54
55 /*
56 * compare_path_costs
57 * Return -1, 0, or +1 according as path1 is cheaper, the same cost,
58 * or more expensive than path2 for the specified criterion.
59 */
60 int
compare_path_costs(Path * path1,Path * path2,CostSelector criterion)61 compare_path_costs(Path *path1, Path *path2, CostSelector criterion)
62 {
63 if (criterion == STARTUP_COST)
64 {
65 if (path1->startup_cost < path2->startup_cost)
66 return -1;
67 if (path1->startup_cost > path2->startup_cost)
68 return +1;
69
70 /*
71 * If paths have the same startup cost (not at all unlikely), order
72 * them by total cost.
73 */
74 if (path1->total_cost < path2->total_cost)
75 return -1;
76 if (path1->total_cost > path2->total_cost)
77 return +1;
78 }
79 else
80 {
81 if (path1->total_cost < path2->total_cost)
82 return -1;
83 if (path1->total_cost > path2->total_cost)
84 return +1;
85
86 /*
87 * If paths have the same total cost, order them by startup cost.
88 */
89 if (path1->startup_cost < path2->startup_cost)
90 return -1;
91 if (path1->startup_cost > path2->startup_cost)
92 return +1;
93 }
94 return 0;
95 }
96
97 /*
98 * compare_path_fractional_costs
99 * Return -1, 0, or +1 according as path1 is cheaper, the same cost,
100 * or more expensive than path2 for fetching the specified fraction
101 * of the total tuples.
102 *
103 * If fraction is <= 0 or > 1, we interpret it as 1, ie, we select the
104 * path with the cheaper total_cost.
105 */
106 int
compare_fractional_path_costs(Path * path1,Path * path2,double fraction)107 compare_fractional_path_costs(Path *path1, Path *path2,
108 double fraction)
109 {
110 Cost cost1,
111 cost2;
112
113 if (fraction <= 0.0 || fraction >= 1.0)
114 return compare_path_costs(path1, path2, TOTAL_COST);
115 cost1 = path1->startup_cost +
116 fraction * (path1->total_cost - path1->startup_cost);
117 cost2 = path2->startup_cost +
118 fraction * (path2->total_cost - path2->startup_cost);
119 if (cost1 < cost2)
120 return -1;
121 if (cost1 > cost2)
122 return +1;
123 return 0;
124 }
125
126 /*
127 * compare_path_costs_fuzzily
128 * Compare the costs of two paths to see if either can be said to
129 * dominate the other.
130 *
131 * We use fuzzy comparisons so that add_path() can avoid keeping both of
132 * a pair of paths that really have insignificantly different cost.
133 *
134 * The fuzz_factor argument must be 1.0 plus delta, where delta is the
135 * fraction of the smaller cost that is considered to be a significant
136 * difference. For example, fuzz_factor = 1.01 makes the fuzziness limit
137 * be 1% of the smaller cost.
138 *
139 * The two paths are said to have "equal" costs if both startup and total
140 * costs are fuzzily the same. Path1 is said to be better than path2 if
141 * it has fuzzily better startup cost and fuzzily no worse total cost,
142 * or if it has fuzzily better total cost and fuzzily no worse startup cost.
143 * Path2 is better than path1 if the reverse holds. Finally, if one path
144 * is fuzzily better than the other on startup cost and fuzzily worse on
145 * total cost, we just say that their costs are "different", since neither
146 * dominates the other across the whole performance spectrum.
147 *
148 * This function also enforces a policy rule that paths for which the relevant
149 * one of parent->consider_startup and parent->consider_param_startup is false
150 * cannot survive comparisons solely on the grounds of good startup cost, so
151 * we never return COSTS_DIFFERENT when that is true for the total-cost loser.
152 * (But if total costs are fuzzily equal, we compare startup costs anyway,
153 * in hopes of eliminating one path or the other.)
154 */
155 static PathCostComparison
compare_path_costs_fuzzily(Path * path1,Path * path2,double fuzz_factor)156 compare_path_costs_fuzzily(Path *path1, Path *path2, double fuzz_factor)
157 {
158 #define CONSIDER_PATH_STARTUP_COST(p) \
159 ((p)->param_info == NULL ? (p)->parent->consider_startup : (p)->parent->consider_param_startup)
160
161 /*
162 * Check total cost first since it's more likely to be different; many
163 * paths have zero startup cost.
164 */
165 if (path1->total_cost > path2->total_cost * fuzz_factor)
166 {
167 /* path1 fuzzily worse on total cost */
168 if (CONSIDER_PATH_STARTUP_COST(path1) &&
169 path2->startup_cost > path1->startup_cost * fuzz_factor)
170 {
171 /* ... but path2 fuzzily worse on startup, so DIFFERENT */
172 return COSTS_DIFFERENT;
173 }
174 /* else path2 dominates */
175 return COSTS_BETTER2;
176 }
177 if (path2->total_cost > path1->total_cost * fuzz_factor)
178 {
179 /* path2 fuzzily worse on total cost */
180 if (CONSIDER_PATH_STARTUP_COST(path2) &&
181 path1->startup_cost > path2->startup_cost * fuzz_factor)
182 {
183 /* ... but path1 fuzzily worse on startup, so DIFFERENT */
184 return COSTS_DIFFERENT;
185 }
186 /* else path1 dominates */
187 return COSTS_BETTER1;
188 }
189 /* fuzzily the same on total cost ... */
190 if (path1->startup_cost > path2->startup_cost * fuzz_factor)
191 {
192 /* ... but path1 fuzzily worse on startup, so path2 wins */
193 return COSTS_BETTER2;
194 }
195 if (path2->startup_cost > path1->startup_cost * fuzz_factor)
196 {
197 /* ... but path2 fuzzily worse on startup, so path1 wins */
198 return COSTS_BETTER1;
199 }
200 /* fuzzily the same on both costs */
201 return COSTS_EQUAL;
202
203 #undef CONSIDER_PATH_STARTUP_COST
204 }
205
206 /*
207 * set_cheapest
208 * Find the minimum-cost paths from among a relation's paths,
209 * and save them in the rel's cheapest-path fields.
210 *
211 * cheapest_total_path is normally the cheapest-total-cost unparameterized
212 * path; but if there are no unparameterized paths, we assign it to be the
213 * best (cheapest least-parameterized) parameterized path. However, only
214 * unparameterized paths are considered candidates for cheapest_startup_path,
215 * so that will be NULL if there are no unparameterized paths.
216 *
217 * The cheapest_parameterized_paths list collects all parameterized paths
218 * that have survived the add_path() tournament for this relation. (Since
219 * add_path ignores pathkeys for a parameterized path, these will be paths
220 * that have best cost or best row count for their parameterization. We
221 * may also have both a parallel-safe and a non-parallel-safe path in some
222 * cases for the same parameterization in some cases, but this should be
223 * relatively rare since, most typically, all paths for the same relation
224 * will be parallel-safe or none of them will.)
225 *
226 * cheapest_parameterized_paths always includes the cheapest-total
227 * unparameterized path, too, if there is one; the users of that list find
228 * it more convenient if that's included.
229 *
230 * This is normally called only after we've finished constructing the path
231 * list for the rel node.
232 */
233 void
set_cheapest(RelOptInfo * parent_rel)234 set_cheapest(RelOptInfo *parent_rel)
235 {
236 Path *cheapest_startup_path;
237 Path *cheapest_total_path;
238 Path *best_param_path;
239 List *parameterized_paths;
240 ListCell *p;
241
242 Assert(IsA(parent_rel, RelOptInfo));
243
244 if (parent_rel->pathlist == NIL)
245 elog(ERROR, "could not devise a query plan for the given query");
246
247 cheapest_startup_path = cheapest_total_path = best_param_path = NULL;
248 parameterized_paths = NIL;
249
250 foreach(p, parent_rel->pathlist)
251 {
252 Path *path = (Path *) lfirst(p);
253 int cmp;
254
255 if (path->param_info)
256 {
257 /* Parameterized path, so add it to parameterized_paths */
258 parameterized_paths = lappend(parameterized_paths, path);
259
260 /*
261 * If we have an unparameterized cheapest-total, we no longer care
262 * about finding the best parameterized path, so move on.
263 */
264 if (cheapest_total_path)
265 continue;
266
267 /*
268 * Otherwise, track the best parameterized path, which is the one
269 * with least total cost among those of the minimum
270 * parameterization.
271 */
272 if (best_param_path == NULL)
273 best_param_path = path;
274 else
275 {
276 switch (bms_subset_compare(PATH_REQ_OUTER(path),
277 PATH_REQ_OUTER(best_param_path)))
278 {
279 case BMS_EQUAL:
280 /* keep the cheaper one */
281 if (compare_path_costs(path, best_param_path,
282 TOTAL_COST) < 0)
283 best_param_path = path;
284 break;
285 case BMS_SUBSET1:
286 /* new path is less-parameterized */
287 best_param_path = path;
288 break;
289 case BMS_SUBSET2:
290 /* old path is less-parameterized, keep it */
291 break;
292 case BMS_DIFFERENT:
293
294 /*
295 * This means that neither path has the least possible
296 * parameterization for the rel. We'll sit on the old
297 * path until something better comes along.
298 */
299 break;
300 }
301 }
302 }
303 else
304 {
305 /* Unparameterized path, so consider it for cheapest slots */
306 if (cheapest_total_path == NULL)
307 {
308 cheapest_startup_path = cheapest_total_path = path;
309 continue;
310 }
311
312 /*
313 * If we find two paths of identical costs, try to keep the
314 * better-sorted one. The paths might have unrelated sort
315 * orderings, in which case we can only guess which might be
316 * better to keep, but if one is superior then we definitely
317 * should keep that one.
318 */
319 cmp = compare_path_costs(cheapest_startup_path, path, STARTUP_COST);
320 if (cmp > 0 ||
321 (cmp == 0 &&
322 compare_pathkeys(cheapest_startup_path->pathkeys,
323 path->pathkeys) == PATHKEYS_BETTER2))
324 cheapest_startup_path = path;
325
326 cmp = compare_path_costs(cheapest_total_path, path, TOTAL_COST);
327 if (cmp > 0 ||
328 (cmp == 0 &&
329 compare_pathkeys(cheapest_total_path->pathkeys,
330 path->pathkeys) == PATHKEYS_BETTER2))
331 cheapest_total_path = path;
332 }
333 }
334
335 /* Add cheapest unparameterized path, if any, to parameterized_paths */
336 if (cheapest_total_path)
337 parameterized_paths = lcons(cheapest_total_path, parameterized_paths);
338
339 /*
340 * If there is no unparameterized path, use the best parameterized path as
341 * cheapest_total_path (but not as cheapest_startup_path).
342 */
343 if (cheapest_total_path == NULL)
344 cheapest_total_path = best_param_path;
345 Assert(cheapest_total_path != NULL);
346
347 parent_rel->cheapest_startup_path = cheapest_startup_path;
348 parent_rel->cheapest_total_path = cheapest_total_path;
349 parent_rel->cheapest_unique_path = NULL; /* computed only if needed */
350 parent_rel->cheapest_parameterized_paths = parameterized_paths;
351 }
352
353 /*
354 * add_path
355 * Consider a potential implementation path for the specified parent rel,
356 * and add it to the rel's pathlist if it is worthy of consideration.
357 * A path is worthy if it has a better sort order (better pathkeys) or
358 * cheaper cost (on either dimension), or generates fewer rows, than any
359 * existing path that has the same or superset parameterization rels.
360 * We also consider parallel-safe paths more worthy than others.
361 *
362 * We also remove from the rel's pathlist any old paths that are dominated
363 * by new_path --- that is, new_path is cheaper, at least as well ordered,
364 * generates no more rows, requires no outer rels not required by the old
365 * path, and is no less parallel-safe.
366 *
367 * In most cases, a path with a superset parameterization will generate
368 * fewer rows (since it has more join clauses to apply), so that those two
369 * figures of merit move in opposite directions; this means that a path of
370 * one parameterization can seldom dominate a path of another. But such
371 * cases do arise, so we make the full set of checks anyway.
372 *
373 * There are two policy decisions embedded in this function, along with
374 * its sibling add_path_precheck. First, we treat all parameterized paths
375 * as having NIL pathkeys, so that they cannot win comparisons on the
376 * basis of sort order. This is to reduce the number of parameterized
377 * paths that are kept; see discussion in src/backend/optimizer/README.
378 *
379 * Second, we only consider cheap startup cost to be interesting if
380 * parent_rel->consider_startup is true for an unparameterized path, or
381 * parent_rel->consider_param_startup is true for a parameterized one.
382 * Again, this allows discarding useless paths sooner.
383 *
384 * The pathlist is kept sorted by total_cost, with cheaper paths
385 * at the front. Within this routine, that's simply a speed hack:
386 * doing it that way makes it more likely that we will reject an inferior
387 * path after a few comparisons, rather than many comparisons.
388 * However, add_path_precheck relies on this ordering to exit early
389 * when possible.
390 *
391 * NOTE: discarded Path objects are immediately pfree'd to reduce planner
392 * memory consumption. We dare not try to free the substructure of a Path,
393 * since much of it may be shared with other Paths or the query tree itself;
394 * but just recycling discarded Path nodes is a very useful savings in
395 * a large join tree. We can recycle the List nodes of pathlist, too.
396 *
397 * As noted in optimizer/README, deleting a previously-accepted Path is
398 * safe because we know that Paths of this rel cannot yet be referenced
399 * from any other rel, such as a higher-level join. However, in some cases
400 * it is possible that a Path is referenced by another Path for its own
401 * rel; we must not delete such a Path, even if it is dominated by the new
402 * Path. Currently this occurs only for IndexPath objects, which may be
403 * referenced as children of BitmapHeapPaths as well as being paths in
404 * their own right. Hence, we don't pfree IndexPaths when rejecting them.
405 *
406 * 'parent_rel' is the relation entry to which the path corresponds.
407 * 'new_path' is a potential path for parent_rel.
408 *
409 * Returns nothing, but modifies parent_rel->pathlist.
410 */
411 void
add_path(RelOptInfo * parent_rel,Path * new_path)412 add_path(RelOptInfo *parent_rel, Path *new_path)
413 {
414 bool accept_new = true; /* unless we find a superior old path */
415 ListCell *insert_after = NULL; /* where to insert new item */
416 List *new_path_pathkeys;
417 ListCell *p1;
418 ListCell *p1_prev;
419 ListCell *p1_next;
420
421 /*
422 * This is a convenient place to check for query cancel --- no part of the
423 * planner goes very long without calling add_path().
424 */
425 CHECK_FOR_INTERRUPTS();
426
427 /* Pretend parameterized paths have no pathkeys, per comment above */
428 new_path_pathkeys = new_path->param_info ? NIL : new_path->pathkeys;
429
430 /*
431 * Loop to check proposed new path against old paths. Note it is possible
432 * for more than one old path to be tossed out because new_path dominates
433 * it.
434 *
435 * We can't use foreach here because the loop body may delete the current
436 * list cell.
437 */
438 p1_prev = NULL;
439 for (p1 = list_head(parent_rel->pathlist); p1 != NULL; p1 = p1_next)
440 {
441 Path *old_path = (Path *) lfirst(p1);
442 bool remove_old = false; /* unless new proves superior */
443 PathCostComparison costcmp;
444 PathKeysComparison keyscmp;
445 BMS_Comparison outercmp;
446
447 p1_next = lnext(p1);
448
449 /*
450 * Do a fuzzy cost comparison with standard fuzziness limit.
451 */
452 costcmp = compare_path_costs_fuzzily(new_path, old_path,
453 STD_FUZZ_FACTOR);
454
455 /*
456 * If the two paths compare differently for startup and total cost,
457 * then we want to keep both, and we can skip comparing pathkeys and
458 * required_outer rels. If they compare the same, proceed with the
459 * other comparisons. Row count is checked last. (We make the tests
460 * in this order because the cost comparison is most likely to turn
461 * out "different", and the pathkeys comparison next most likely. As
462 * explained above, row count very seldom makes a difference, so even
463 * though it's cheap to compare there's not much point in checking it
464 * earlier.)
465 */
466 if (costcmp != COSTS_DIFFERENT)
467 {
468 /* Similarly check to see if either dominates on pathkeys */
469 List *old_path_pathkeys;
470
471 old_path_pathkeys = old_path->param_info ? NIL : old_path->pathkeys;
472 keyscmp = compare_pathkeys(new_path_pathkeys,
473 old_path_pathkeys);
474 if (keyscmp != PATHKEYS_DIFFERENT)
475 {
476 switch (costcmp)
477 {
478 case COSTS_EQUAL:
479 outercmp = bms_subset_compare(PATH_REQ_OUTER(new_path),
480 PATH_REQ_OUTER(old_path));
481 if (keyscmp == PATHKEYS_BETTER1)
482 {
483 if ((outercmp == BMS_EQUAL ||
484 outercmp == BMS_SUBSET1) &&
485 new_path->rows <= old_path->rows &&
486 new_path->parallel_safe >= old_path->parallel_safe)
487 remove_old = true; /* new dominates old */
488 }
489 else if (keyscmp == PATHKEYS_BETTER2)
490 {
491 if ((outercmp == BMS_EQUAL ||
492 outercmp == BMS_SUBSET2) &&
493 new_path->rows >= old_path->rows &&
494 new_path->parallel_safe <= old_path->parallel_safe)
495 accept_new = false; /* old dominates new */
496 }
497 else /* keyscmp == PATHKEYS_EQUAL */
498 {
499 if (outercmp == BMS_EQUAL)
500 {
501 /*
502 * Same pathkeys and outer rels, and fuzzily
503 * the same cost, so keep just one; to decide
504 * which, first check parallel-safety, then
505 * rows, then do a fuzzy cost comparison with
506 * very small fuzz limit. (We used to do an
507 * exact cost comparison, but that results in
508 * annoying platform-specific plan variations
509 * due to roundoff in the cost estimates.) If
510 * things are still tied, arbitrarily keep
511 * only the old path. Notice that we will
512 * keep only the old path even if the
513 * less-fuzzy comparison decides the startup
514 * and total costs compare differently.
515 */
516 if (new_path->parallel_safe >
517 old_path->parallel_safe)
518 remove_old = true; /* new dominates old */
519 else if (new_path->parallel_safe <
520 old_path->parallel_safe)
521 accept_new = false; /* old dominates new */
522 else if (new_path->rows < old_path->rows)
523 remove_old = true; /* new dominates old */
524 else if (new_path->rows > old_path->rows)
525 accept_new = false; /* old dominates new */
526 else if (compare_path_costs_fuzzily(new_path,
527 old_path,
528 1.0000000001) == COSTS_BETTER1)
529 remove_old = true; /* new dominates old */
530 else
531 accept_new = false; /* old equals or
532 * dominates new */
533 }
534 else if (outercmp == BMS_SUBSET1 &&
535 new_path->rows <= old_path->rows &&
536 new_path->parallel_safe >= old_path->parallel_safe)
537 remove_old = true; /* new dominates old */
538 else if (outercmp == BMS_SUBSET2 &&
539 new_path->rows >= old_path->rows &&
540 new_path->parallel_safe <= old_path->parallel_safe)
541 accept_new = false; /* old dominates new */
542 /* else different parameterizations, keep both */
543 }
544 break;
545 case COSTS_BETTER1:
546 if (keyscmp != PATHKEYS_BETTER2)
547 {
548 outercmp = bms_subset_compare(PATH_REQ_OUTER(new_path),
549 PATH_REQ_OUTER(old_path));
550 if ((outercmp == BMS_EQUAL ||
551 outercmp == BMS_SUBSET1) &&
552 new_path->rows <= old_path->rows &&
553 new_path->parallel_safe >= old_path->parallel_safe)
554 remove_old = true; /* new dominates old */
555 }
556 break;
557 case COSTS_BETTER2:
558 if (keyscmp != PATHKEYS_BETTER1)
559 {
560 outercmp = bms_subset_compare(PATH_REQ_OUTER(new_path),
561 PATH_REQ_OUTER(old_path));
562 if ((outercmp == BMS_EQUAL ||
563 outercmp == BMS_SUBSET2) &&
564 new_path->rows >= old_path->rows &&
565 new_path->parallel_safe <= old_path->parallel_safe)
566 accept_new = false; /* old dominates new */
567 }
568 break;
569 case COSTS_DIFFERENT:
570
571 /*
572 * can't get here, but keep this case to keep compiler
573 * quiet
574 */
575 break;
576 }
577 }
578 }
579
580 /*
581 * Remove current element from pathlist if dominated by new.
582 */
583 if (remove_old)
584 {
585 parent_rel->pathlist = list_delete_cell(parent_rel->pathlist,
586 p1, p1_prev);
587
588 /*
589 * Delete the data pointed-to by the deleted cell, if possible
590 */
591 if (!IsA(old_path, IndexPath))
592 pfree(old_path);
593 /* p1_prev does not advance */
594 }
595 else
596 {
597 /* new belongs after this old path if it has cost >= old's */
598 if (new_path->total_cost >= old_path->total_cost)
599 insert_after = p1;
600 /* p1_prev advances */
601 p1_prev = p1;
602 }
603
604 /*
605 * If we found an old path that dominates new_path, we can quit
606 * scanning the pathlist; we will not add new_path, and we assume
607 * new_path cannot dominate any other elements of the pathlist.
608 */
609 if (!accept_new)
610 break;
611 }
612
613 if (accept_new)
614 {
615 /* Accept the new path: insert it at proper place in pathlist */
616 if (insert_after)
617 lappend_cell(parent_rel->pathlist, insert_after, new_path);
618 else
619 parent_rel->pathlist = lcons(new_path, parent_rel->pathlist);
620 }
621 else
622 {
623 /* Reject and recycle the new path */
624 if (!IsA(new_path, IndexPath))
625 pfree(new_path);
626 }
627 }
628
629 /*
630 * add_path_precheck
631 * Check whether a proposed new path could possibly get accepted.
632 * We assume we know the path's pathkeys and parameterization accurately,
633 * and have lower bounds for its costs.
634 *
635 * Note that we do not know the path's rowcount, since getting an estimate for
636 * that is too expensive to do before prechecking. We assume here that paths
637 * of a superset parameterization will generate fewer rows; if that holds,
638 * then paths with different parameterizations cannot dominate each other
639 * and so we can simply ignore existing paths of another parameterization.
640 * (In the infrequent cases where that rule of thumb fails, add_path will
641 * get rid of the inferior path.)
642 *
643 * At the time this is called, we haven't actually built a Path structure,
644 * so the required information has to be passed piecemeal.
645 */
646 bool
add_path_precheck(RelOptInfo * parent_rel,Cost startup_cost,Cost total_cost,List * pathkeys,Relids required_outer)647 add_path_precheck(RelOptInfo *parent_rel,
648 Cost startup_cost, Cost total_cost,
649 List *pathkeys, Relids required_outer)
650 {
651 List *new_path_pathkeys;
652 bool consider_startup;
653 ListCell *p1;
654
655 /* Pretend parameterized paths have no pathkeys, per add_path policy */
656 new_path_pathkeys = required_outer ? NIL : pathkeys;
657
658 /* Decide whether new path's startup cost is interesting */
659 consider_startup = required_outer ? parent_rel->consider_param_startup : parent_rel->consider_startup;
660
661 foreach(p1, parent_rel->pathlist)
662 {
663 Path *old_path = (Path *) lfirst(p1);
664 PathKeysComparison keyscmp;
665
666 /*
667 * We are looking for an old_path with the same parameterization (and
668 * by assumption the same rowcount) that dominates the new path on
669 * pathkeys as well as both cost metrics. If we find one, we can
670 * reject the new path.
671 *
672 * Cost comparisons here should match compare_path_costs_fuzzily.
673 */
674 if (total_cost > old_path->total_cost * STD_FUZZ_FACTOR)
675 {
676 /* new path can win on startup cost only if consider_startup */
677 if (startup_cost > old_path->startup_cost * STD_FUZZ_FACTOR ||
678 !consider_startup)
679 {
680 /* new path loses on cost, so check pathkeys... */
681 List *old_path_pathkeys;
682
683 old_path_pathkeys = old_path->param_info ? NIL : old_path->pathkeys;
684 keyscmp = compare_pathkeys(new_path_pathkeys,
685 old_path_pathkeys);
686 if (keyscmp == PATHKEYS_EQUAL ||
687 keyscmp == PATHKEYS_BETTER2)
688 {
689 /* new path does not win on pathkeys... */
690 if (bms_equal(required_outer, PATH_REQ_OUTER(old_path)))
691 {
692 /* Found an old path that dominates the new one */
693 return false;
694 }
695 }
696 }
697 }
698 else
699 {
700 /*
701 * Since the pathlist is sorted by total_cost, we can stop looking
702 * once we reach a path with a total_cost larger than the new
703 * path's.
704 */
705 break;
706 }
707 }
708
709 return true;
710 }
711
712 /*
713 * add_partial_path
714 * Like add_path, our goal here is to consider whether a path is worthy
715 * of being kept around, but the considerations here are a bit different.
716 * A partial path is one which can be executed in any number of workers in
717 * parallel such that each worker will generate a subset of the path's
718 * overall result.
719 *
720 * As in add_path, the partial_pathlist is kept sorted with the cheapest
721 * total path in front. This is depended on by multiple places, which
722 * just take the front entry as the cheapest path without searching.
723 *
724 * We don't generate parameterized partial paths for several reasons. Most
725 * importantly, they're not safe to execute, because there's nothing to
726 * make sure that a parallel scan within the parameterized portion of the
727 * plan is running with the same value in every worker at the same time.
728 * Fortunately, it seems unlikely to be worthwhile anyway, because having
729 * each worker scan the entire outer relation and a subset of the inner
730 * relation will generally be a terrible plan. The inner (parameterized)
731 * side of the plan will be small anyway. There could be rare cases where
732 * this wins big - e.g. if join order constraints put a 1-row relation on
733 * the outer side of the topmost join with a parameterized plan on the inner
734 * side - but we'll have to be content not to handle such cases until
735 * somebody builds an executor infrastructure that can cope with them.
736 *
737 * Because we don't consider parameterized paths here, we also don't
738 * need to consider the row counts as a measure of quality: every path will
739 * produce the same number of rows. Neither do we need to consider startup
740 * costs: parallelism is only used for plans that will be run to completion.
741 * Therefore, this routine is much simpler than add_path: it needs to
742 * consider only pathkeys and total cost.
743 *
744 * As with add_path, we pfree paths that are found to be dominated by
745 * another partial path; this requires that there be no other references to
746 * such paths yet. Hence, GatherPaths must not be created for a rel until
747 * we're done creating all partial paths for it. We do not currently build
748 * partial indexscan paths, so there is no need for an exception for
749 * IndexPaths here; for safety, we instead Assert that a path to be freed
750 * isn't an IndexPath.
751 */
752 void
add_partial_path(RelOptInfo * parent_rel,Path * new_path)753 add_partial_path(RelOptInfo *parent_rel, Path *new_path)
754 {
755 bool accept_new = true; /* unless we find a superior old path */
756 ListCell *insert_after = NULL; /* where to insert new item */
757 ListCell *p1;
758 ListCell *p1_prev;
759 ListCell *p1_next;
760
761 /* Check for query cancel. */
762 CHECK_FOR_INTERRUPTS();
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 p1_prev = NULL;
769 for (p1 = list_head(parent_rel->partial_pathlist); p1 != NULL;
770 p1 = p1_next)
771 {
772 Path *old_path = (Path *) lfirst(p1);
773 bool remove_old = false; /* unless new proves superior */
774 PathKeysComparison keyscmp;
775
776 p1_next = lnext(p1);
777
778 /* Compare pathkeys. */
779 keyscmp = compare_pathkeys(new_path->pathkeys, old_path->pathkeys);
780
781 /* Unless pathkeys are incompable, keep just one of the two paths. */
782 if (keyscmp != PATHKEYS_DIFFERENT)
783 {
784 if (new_path->total_cost > old_path->total_cost * STD_FUZZ_FACTOR)
785 {
786 /* New path costs more; keep it only if pathkeys are better. */
787 if (keyscmp != PATHKEYS_BETTER1)
788 accept_new = false;
789 }
790 else if (old_path->total_cost > new_path->total_cost
791 * STD_FUZZ_FACTOR)
792 {
793 /* Old path costs more; keep it only if pathkeys are better. */
794 if (keyscmp != PATHKEYS_BETTER2)
795 remove_old = true;
796 }
797 else if (keyscmp == PATHKEYS_BETTER1)
798 {
799 /* Costs are about the same, new path has better pathkeys. */
800 remove_old = true;
801 }
802 else if (keyscmp == PATHKEYS_BETTER2)
803 {
804 /* Costs are about the same, old path has better pathkeys. */
805 accept_new = false;
806 }
807 else if (old_path->total_cost > new_path->total_cost * 1.0000000001)
808 {
809 /* Pathkeys are the same, and the old path costs more. */
810 remove_old = true;
811 }
812 else
813 {
814 /*
815 * Pathkeys are the same, and new path isn't materially
816 * cheaper.
817 */
818 accept_new = false;
819 }
820 }
821
822 /*
823 * Remove current element from partial_pathlist if dominated by new.
824 */
825 if (remove_old)
826 {
827 parent_rel->partial_pathlist =
828 list_delete_cell(parent_rel->partial_pathlist, p1, p1_prev);
829 /* we should not see IndexPaths here, so always safe to delete */
830 Assert(!IsA(old_path, IndexPath));
831 pfree(old_path);
832 /* p1_prev does not advance */
833 }
834 else
835 {
836 /* new belongs after this old path if it has cost >= old's */
837 if (new_path->total_cost >= old_path->total_cost)
838 insert_after = p1;
839 /* p1_prev advances */
840 p1_prev = p1;
841 }
842
843 /*
844 * If we found an old path that dominates new_path, we can quit
845 * scanning the partial_pathlist; we will not add new_path, and we
846 * assume new_path cannot dominate any later path.
847 */
848 if (!accept_new)
849 break;
850 }
851
852 if (accept_new)
853 {
854 /* Accept the new path: insert it at proper place */
855 if (insert_after)
856 lappend_cell(parent_rel->partial_pathlist, insert_after, new_path);
857 else
858 parent_rel->partial_pathlist =
859 lcons(new_path, parent_rel->partial_pathlist);
860 }
861 else
862 {
863 /* we should not see IndexPaths here, so always safe to delete */
864 Assert(!IsA(new_path, IndexPath));
865 /* Reject and recycle the new path */
866 pfree(new_path);
867 }
868 }
869
870 /*
871 * add_partial_path_precheck
872 * Check whether a proposed new partial path could possibly get accepted.
873 *
874 * Unlike add_path_precheck, we can ignore startup cost and parameterization,
875 * since they don't matter for partial paths (see add_partial_path). But
876 * we do want to make sure we don't add a partial path if there's already
877 * a complete path that dominates it, since in that case the proposed path
878 * is surely a loser.
879 */
880 bool
add_partial_path_precheck(RelOptInfo * parent_rel,Cost total_cost,List * pathkeys)881 add_partial_path_precheck(RelOptInfo *parent_rel, Cost total_cost,
882 List *pathkeys)
883 {
884 ListCell *p1;
885
886 /*
887 * Our goal here is twofold. First, we want to find out whether this path
888 * is clearly inferior to some existing partial path. If so, we want to
889 * reject it immediately. Second, we want to find out whether this path
890 * is clearly superior to some existing partial path -- at least, modulo
891 * final cost computations. If so, we definitely want to consider it.
892 *
893 * Unlike add_path(), we always compare pathkeys here. This is because we
894 * expect partial_pathlist to be very short, and getting a definitive
895 * answer at this stage avoids the need to call add_path_precheck.
896 */
897 foreach(p1, parent_rel->partial_pathlist)
898 {
899 Path *old_path = (Path *) lfirst(p1);
900 PathKeysComparison keyscmp;
901
902 keyscmp = compare_pathkeys(pathkeys, old_path->pathkeys);
903 if (keyscmp != PATHKEYS_DIFFERENT)
904 {
905 if (total_cost > old_path->total_cost * STD_FUZZ_FACTOR &&
906 keyscmp != PATHKEYS_BETTER1)
907 return false;
908 if (old_path->total_cost > total_cost * STD_FUZZ_FACTOR &&
909 keyscmp != PATHKEYS_BETTER2)
910 return true;
911 }
912 }
913
914 /*
915 * This path is neither clearly inferior to an existing partial path nor
916 * clearly good enough that it might replace one. Compare it to
917 * non-parallel plans. If it loses even before accounting for the cost of
918 * the Gather node, we should definitely reject it.
919 *
920 * Note that we pass the total_cost to add_path_precheck twice. This is
921 * because it's never advantageous to consider the startup cost of a
922 * partial path; the resulting plans, if run in parallel, will be run to
923 * completion.
924 */
925 if (!add_path_precheck(parent_rel, total_cost, total_cost, pathkeys,
926 NULL))
927 return false;
928
929 return true;
930 }
931
932
933 /*****************************************************************************
934 * PATH NODE CREATION ROUTINES
935 *****************************************************************************/
936
937 /*
938 * create_seqscan_path
939 * Creates a path corresponding to a sequential scan, returning the
940 * pathnode.
941 */
942 Path *
create_seqscan_path(PlannerInfo * root,RelOptInfo * rel,Relids required_outer,int parallel_workers)943 create_seqscan_path(PlannerInfo *root, RelOptInfo *rel,
944 Relids required_outer, int parallel_workers)
945 {
946 Path *pathnode = makeNode(Path);
947
948 pathnode->pathtype = T_SeqScan;
949 pathnode->parent = rel;
950 pathnode->pathtarget = rel->reltarget;
951 pathnode->param_info = get_baserel_parampathinfo(root, rel,
952 required_outer);
953 pathnode->parallel_aware = parallel_workers > 0 ? true : false;
954 pathnode->parallel_safe = rel->consider_parallel;
955 pathnode->parallel_workers = parallel_workers;
956 pathnode->pathkeys = NIL; /* seqscan has unordered result */
957
958 cost_seqscan(pathnode, root, rel, pathnode->param_info);
959
960 return pathnode;
961 }
962
963 /*
964 * create_samplescan_path
965 * Creates a path node for a sampled table scan.
966 */
967 Path *
create_samplescan_path(PlannerInfo * root,RelOptInfo * rel,Relids required_outer)968 create_samplescan_path(PlannerInfo *root, RelOptInfo *rel, Relids required_outer)
969 {
970 Path *pathnode = makeNode(Path);
971
972 pathnode->pathtype = T_SampleScan;
973 pathnode->parent = rel;
974 pathnode->pathtarget = rel->reltarget;
975 pathnode->param_info = get_baserel_parampathinfo(root, rel,
976 required_outer);
977 pathnode->parallel_aware = false;
978 pathnode->parallel_safe = rel->consider_parallel;
979 pathnode->parallel_workers = 0;
980 pathnode->pathkeys = NIL; /* samplescan has unordered result */
981
982 cost_samplescan(pathnode, root, rel, pathnode->param_info);
983
984 return pathnode;
985 }
986
987 /*
988 * create_index_path
989 * Creates a path node for an index scan.
990 *
991 * 'index' is a usable index.
992 * 'indexclauses' is a list of RestrictInfo nodes representing clauses
993 * to be used as index qual conditions in the scan.
994 * 'indexclausecols' is an integer list of index column numbers (zero based)
995 * the indexclauses can be used with.
996 * 'indexorderbys' is a list of bare expressions (no RestrictInfos)
997 * to be used as index ordering operators in the scan.
998 * 'indexorderbycols' is an integer list of index column numbers (zero based)
999 * the ordering operators can be used with.
1000 * 'pathkeys' describes the ordering of the path.
1001 * 'indexscandir' is ForwardScanDirection or BackwardScanDirection
1002 * for an ordered index, or NoMovementScanDirection for
1003 * an unordered index.
1004 * 'indexonly' is true if an index-only scan is wanted.
1005 * 'required_outer' is the set of outer relids for a parameterized path.
1006 * 'loop_count' is the number of repetitions of the indexscan to factor into
1007 * estimates of caching behavior.
1008 *
1009 * Returns the new path node.
1010 */
1011 IndexPath *
create_index_path(PlannerInfo * root,IndexOptInfo * index,List * indexclauses,List * indexclausecols,List * indexorderbys,List * indexorderbycols,List * pathkeys,ScanDirection indexscandir,bool indexonly,Relids required_outer,double loop_count)1012 create_index_path(PlannerInfo *root,
1013 IndexOptInfo *index,
1014 List *indexclauses,
1015 List *indexclausecols,
1016 List *indexorderbys,
1017 List *indexorderbycols,
1018 List *pathkeys,
1019 ScanDirection indexscandir,
1020 bool indexonly,
1021 Relids required_outer,
1022 double loop_count)
1023 {
1024 IndexPath *pathnode = makeNode(IndexPath);
1025 RelOptInfo *rel = index->rel;
1026 List *indexquals,
1027 *indexqualcols;
1028
1029 pathnode->path.pathtype = indexonly ? T_IndexOnlyScan : T_IndexScan;
1030 pathnode->path.parent = rel;
1031 pathnode->path.pathtarget = rel->reltarget;
1032 pathnode->path.param_info = get_baserel_parampathinfo(root, rel,
1033 required_outer);
1034 pathnode->path.parallel_aware = false;
1035 pathnode->path.parallel_safe = rel->consider_parallel;
1036 pathnode->path.parallel_workers = 0;
1037 pathnode->path.pathkeys = pathkeys;
1038
1039 /* Convert clauses to indexquals the executor can handle */
1040 expand_indexqual_conditions(index, indexclauses, indexclausecols,
1041 &indexquals, &indexqualcols);
1042
1043 /* Fill in the pathnode */
1044 pathnode->indexinfo = index;
1045 pathnode->indexclauses = indexclauses;
1046 pathnode->indexquals = indexquals;
1047 pathnode->indexqualcols = indexqualcols;
1048 pathnode->indexorderbys = indexorderbys;
1049 pathnode->indexorderbycols = indexorderbycols;
1050 pathnode->indexscandir = indexscandir;
1051
1052 cost_index(pathnode, root, loop_count);
1053
1054 return pathnode;
1055 }
1056
1057 /*
1058 * create_bitmap_heap_path
1059 * Creates a path node for a bitmap scan.
1060 *
1061 * 'bitmapqual' is a tree of IndexPath, BitmapAndPath, and BitmapOrPath nodes.
1062 * 'required_outer' is the set of outer relids for a parameterized path.
1063 * 'loop_count' is the number of repetitions of the indexscan to factor into
1064 * estimates of caching behavior.
1065 *
1066 * loop_count should match the value used when creating the component
1067 * IndexPaths.
1068 */
1069 BitmapHeapPath *
create_bitmap_heap_path(PlannerInfo * root,RelOptInfo * rel,Path * bitmapqual,Relids required_outer,double loop_count)1070 create_bitmap_heap_path(PlannerInfo *root,
1071 RelOptInfo *rel,
1072 Path *bitmapqual,
1073 Relids required_outer,
1074 double loop_count)
1075 {
1076 BitmapHeapPath *pathnode = makeNode(BitmapHeapPath);
1077
1078 pathnode->path.pathtype = T_BitmapHeapScan;
1079 pathnode->path.parent = rel;
1080 pathnode->path.pathtarget = rel->reltarget;
1081 pathnode->path.param_info = get_baserel_parampathinfo(root, rel,
1082 required_outer);
1083 pathnode->path.parallel_aware = false;
1084 pathnode->path.parallel_safe = rel->consider_parallel;
1085 pathnode->path.parallel_workers = 0;
1086 pathnode->path.pathkeys = NIL; /* always unordered */
1087
1088 pathnode->bitmapqual = bitmapqual;
1089
1090 cost_bitmap_heap_scan(&pathnode->path, root, rel,
1091 pathnode->path.param_info,
1092 bitmapqual, loop_count);
1093
1094 return pathnode;
1095 }
1096
1097 /*
1098 * create_bitmap_and_path
1099 * Creates a path node representing a BitmapAnd.
1100 */
1101 BitmapAndPath *
create_bitmap_and_path(PlannerInfo * root,RelOptInfo * rel,List * bitmapquals)1102 create_bitmap_and_path(PlannerInfo *root,
1103 RelOptInfo *rel,
1104 List *bitmapquals)
1105 {
1106 BitmapAndPath *pathnode = makeNode(BitmapAndPath);
1107
1108 pathnode->path.pathtype = T_BitmapAnd;
1109 pathnode->path.parent = rel;
1110 pathnode->path.pathtarget = rel->reltarget;
1111 pathnode->path.param_info = NULL; /* not used in bitmap trees */
1112
1113 /*
1114 * Currently, a BitmapHeapPath, BitmapAndPath, or BitmapOrPath will be
1115 * parallel-safe if and only if rel->consider_parallel is set. So, we can
1116 * set the flag for this path based only on the relation-level flag,
1117 * without actually iterating over the list of children.
1118 */
1119 pathnode->path.parallel_aware = false;
1120 pathnode->path.parallel_safe = rel->consider_parallel;
1121 pathnode->path.parallel_workers = 0;
1122
1123 pathnode->path.pathkeys = NIL; /* always unordered */
1124
1125 pathnode->bitmapquals = bitmapquals;
1126
1127 /* this sets bitmapselectivity as well as the regular cost fields: */
1128 cost_bitmap_and_node(pathnode, root);
1129
1130 return pathnode;
1131 }
1132
1133 /*
1134 * create_bitmap_or_path
1135 * Creates a path node representing a BitmapOr.
1136 */
1137 BitmapOrPath *
create_bitmap_or_path(PlannerInfo * root,RelOptInfo * rel,List * bitmapquals)1138 create_bitmap_or_path(PlannerInfo *root,
1139 RelOptInfo *rel,
1140 List *bitmapquals)
1141 {
1142 BitmapOrPath *pathnode = makeNode(BitmapOrPath);
1143
1144 pathnode->path.pathtype = T_BitmapOr;
1145 pathnode->path.parent = rel;
1146 pathnode->path.pathtarget = rel->reltarget;
1147 pathnode->path.param_info = NULL; /* not used in bitmap trees */
1148
1149 /*
1150 * Currently, a BitmapHeapPath, BitmapAndPath, or BitmapOrPath will be
1151 * parallel-safe if and only if rel->consider_parallel is set. So, we can
1152 * set the flag for this path based only on the relation-level flag,
1153 * without actually iterating over the list of children.
1154 */
1155 pathnode->path.parallel_aware = false;
1156 pathnode->path.parallel_safe = rel->consider_parallel;
1157 pathnode->path.parallel_workers = 0;
1158
1159 pathnode->path.pathkeys = NIL; /* always unordered */
1160
1161 pathnode->bitmapquals = bitmapquals;
1162
1163 /* this sets bitmapselectivity as well as the regular cost fields: */
1164 cost_bitmap_or_node(pathnode, root);
1165
1166 return pathnode;
1167 }
1168
1169 /*
1170 * create_tidscan_path
1171 * Creates a path corresponding to a scan by TID, returning the pathnode.
1172 */
1173 TidPath *
create_tidscan_path(PlannerInfo * root,RelOptInfo * rel,List * tidquals,Relids required_outer)1174 create_tidscan_path(PlannerInfo *root, RelOptInfo *rel, List *tidquals,
1175 Relids required_outer)
1176 {
1177 TidPath *pathnode = makeNode(TidPath);
1178
1179 pathnode->path.pathtype = T_TidScan;
1180 pathnode->path.parent = rel;
1181 pathnode->path.pathtarget = rel->reltarget;
1182 pathnode->path.param_info = get_baserel_parampathinfo(root, rel,
1183 required_outer);
1184 pathnode->path.parallel_aware = false;
1185 pathnode->path.parallel_safe = rel->consider_parallel;
1186 pathnode->path.parallel_workers = 0;
1187 pathnode->path.pathkeys = NIL; /* always unordered */
1188
1189 pathnode->tidquals = tidquals;
1190
1191 cost_tidscan(&pathnode->path, root, rel, tidquals,
1192 pathnode->path.param_info);
1193
1194 return pathnode;
1195 }
1196
1197 /*
1198 * create_append_path
1199 * Creates a path corresponding to an Append plan, returning the
1200 * pathnode.
1201 *
1202 * Note that we must handle subpaths = NIL, representing a dummy access path.
1203 */
1204 AppendPath *
create_append_path(RelOptInfo * rel,List * subpaths,Relids required_outer,int parallel_workers)1205 create_append_path(RelOptInfo *rel, List *subpaths, Relids required_outer,
1206 int parallel_workers)
1207 {
1208 AppendPath *pathnode = makeNode(AppendPath);
1209 ListCell *l;
1210
1211 pathnode->path.pathtype = T_Append;
1212 pathnode->path.parent = rel;
1213 pathnode->path.pathtarget = rel->reltarget;
1214 pathnode->path.param_info = get_appendrel_parampathinfo(rel,
1215 required_outer);
1216 pathnode->path.parallel_aware = false;
1217 pathnode->path.parallel_safe = rel->consider_parallel;
1218 pathnode->path.parallel_workers = parallel_workers;
1219 pathnode->path.pathkeys = NIL; /* result is always considered
1220 * unsorted */
1221 pathnode->subpaths = subpaths;
1222
1223 /*
1224 * We don't bother with inventing a cost_append(), but just do it here.
1225 *
1226 * Compute rows and costs as sums of subplan rows and costs. We charge
1227 * nothing extra for the Append itself, which perhaps is too optimistic,
1228 * but since it doesn't do any selection or projection, it is a pretty
1229 * cheap node.
1230 */
1231 pathnode->path.rows = 0;
1232 pathnode->path.startup_cost = 0;
1233 pathnode->path.total_cost = 0;
1234 foreach(l, subpaths)
1235 {
1236 Path *subpath = (Path *) lfirst(l);
1237
1238 pathnode->path.rows += subpath->rows;
1239
1240 if (l == list_head(subpaths)) /* first node? */
1241 pathnode->path.startup_cost = subpath->startup_cost;
1242 pathnode->path.total_cost += subpath->total_cost;
1243 pathnode->path.parallel_safe = pathnode->path.parallel_safe &&
1244 subpath->parallel_safe;
1245
1246 /* All child paths must have same parameterization */
1247 Assert(bms_equal(PATH_REQ_OUTER(subpath), required_outer));
1248 }
1249
1250 return pathnode;
1251 }
1252
1253 /*
1254 * create_merge_append_path
1255 * Creates a path corresponding to a MergeAppend plan, returning the
1256 * pathnode.
1257 */
1258 MergeAppendPath *
create_merge_append_path(PlannerInfo * root,RelOptInfo * rel,List * subpaths,List * pathkeys,Relids required_outer)1259 create_merge_append_path(PlannerInfo *root,
1260 RelOptInfo *rel,
1261 List *subpaths,
1262 List *pathkeys,
1263 Relids required_outer)
1264 {
1265 MergeAppendPath *pathnode = makeNode(MergeAppendPath);
1266 Cost input_startup_cost;
1267 Cost input_total_cost;
1268 ListCell *l;
1269
1270 pathnode->path.pathtype = T_MergeAppend;
1271 pathnode->path.parent = rel;
1272 pathnode->path.pathtarget = rel->reltarget;
1273 pathnode->path.param_info = get_appendrel_parampathinfo(rel,
1274 required_outer);
1275 pathnode->path.parallel_aware = false;
1276 pathnode->path.parallel_safe = rel->consider_parallel;
1277 pathnode->path.parallel_workers = 0;
1278 pathnode->path.pathkeys = pathkeys;
1279 pathnode->subpaths = subpaths;
1280
1281 /*
1282 * Apply query-wide LIMIT if known and path is for sole base relation.
1283 * (Handling this at this low level is a bit klugy.)
1284 */
1285 if (bms_equal(rel->relids, root->all_baserels))
1286 pathnode->limit_tuples = root->limit_tuples;
1287 else
1288 pathnode->limit_tuples = -1.0;
1289
1290 /*
1291 * Add up the sizes and costs of the input paths.
1292 */
1293 pathnode->path.rows = 0;
1294 input_startup_cost = 0;
1295 input_total_cost = 0;
1296 foreach(l, subpaths)
1297 {
1298 Path *subpath = (Path *) lfirst(l);
1299
1300 pathnode->path.rows += subpath->rows;
1301 pathnode->path.parallel_safe = pathnode->path.parallel_safe &&
1302 subpath->parallel_safe;
1303
1304 if (pathkeys_contained_in(pathkeys, subpath->pathkeys))
1305 {
1306 /* Subpath is adequately ordered, we won't need to sort it */
1307 input_startup_cost += subpath->startup_cost;
1308 input_total_cost += subpath->total_cost;
1309 }
1310 else
1311 {
1312 /* We'll need to insert a Sort node, so include cost for that */
1313 Path sort_path; /* dummy for result of cost_sort */
1314
1315 cost_sort(&sort_path,
1316 root,
1317 pathkeys,
1318 subpath->total_cost,
1319 subpath->parent->tuples,
1320 subpath->pathtarget->width,
1321 0.0,
1322 work_mem,
1323 pathnode->limit_tuples);
1324 input_startup_cost += sort_path.startup_cost;
1325 input_total_cost += sort_path.total_cost;
1326 }
1327
1328 /* All child paths must have same parameterization */
1329 Assert(bms_equal(PATH_REQ_OUTER(subpath), required_outer));
1330 }
1331
1332 /* Now we can compute total costs of the MergeAppend */
1333 cost_merge_append(&pathnode->path, root,
1334 pathkeys, list_length(subpaths),
1335 input_startup_cost, input_total_cost,
1336 rel->tuples);
1337
1338 return pathnode;
1339 }
1340
1341 /*
1342 * create_result_path
1343 * Creates a path representing a Result-and-nothing-else plan.
1344 *
1345 * This is only used for degenerate cases, such as a query with an empty
1346 * jointree.
1347 */
1348 ResultPath *
create_result_path(PlannerInfo * root,RelOptInfo * rel,PathTarget * target,List * resconstantqual)1349 create_result_path(PlannerInfo *root, RelOptInfo *rel,
1350 PathTarget *target, List *resconstantqual)
1351 {
1352 ResultPath *pathnode = makeNode(ResultPath);
1353
1354 pathnode->path.pathtype = T_Result;
1355 pathnode->path.parent = rel;
1356 pathnode->path.pathtarget = target;
1357 pathnode->path.param_info = NULL; /* there are no other rels... */
1358 pathnode->path.parallel_aware = false;
1359 pathnode->path.parallel_safe = rel->consider_parallel;
1360 pathnode->path.parallel_workers = 0;
1361 pathnode->path.pathkeys = NIL;
1362 pathnode->quals = resconstantqual;
1363
1364 /* Hardly worth defining a cost_result() function ... just do it */
1365 pathnode->path.rows = 1;
1366 pathnode->path.startup_cost = target->cost.startup;
1367 pathnode->path.total_cost = target->cost.startup +
1368 cpu_tuple_cost + target->cost.per_tuple;
1369 if (resconstantqual)
1370 {
1371 QualCost qual_cost;
1372
1373 cost_qual_eval(&qual_cost, resconstantqual, root);
1374 /* resconstantqual is evaluated once at startup */
1375 pathnode->path.startup_cost += qual_cost.startup + qual_cost.per_tuple;
1376 pathnode->path.total_cost += qual_cost.startup + qual_cost.per_tuple;
1377 }
1378
1379 return pathnode;
1380 }
1381
1382 /*
1383 * create_material_path
1384 * Creates a path corresponding to a Material plan, returning the
1385 * pathnode.
1386 */
1387 MaterialPath *
create_material_path(RelOptInfo * rel,Path * subpath)1388 create_material_path(RelOptInfo *rel, Path *subpath)
1389 {
1390 MaterialPath *pathnode = makeNode(MaterialPath);
1391
1392 Assert(subpath->parent == rel);
1393
1394 pathnode->path.pathtype = T_Material;
1395 pathnode->path.parent = rel;
1396 pathnode->path.pathtarget = rel->reltarget;
1397 pathnode->path.param_info = subpath->param_info;
1398 pathnode->path.parallel_aware = false;
1399 pathnode->path.parallel_safe = rel->consider_parallel &&
1400 subpath->parallel_safe;
1401 pathnode->path.parallel_workers = subpath->parallel_workers;
1402 pathnode->path.pathkeys = subpath->pathkeys;
1403
1404 pathnode->subpath = subpath;
1405
1406 cost_material(&pathnode->path,
1407 subpath->startup_cost,
1408 subpath->total_cost,
1409 subpath->rows,
1410 subpath->pathtarget->width);
1411
1412 return pathnode;
1413 }
1414
1415 /*
1416 * create_unique_path
1417 * Creates a path representing elimination of distinct rows from the
1418 * input data. Distinct-ness is defined according to the needs of the
1419 * semijoin represented by sjinfo. If it is not possible to identify
1420 * how to make the data unique, NULL is returned.
1421 *
1422 * If used at all, this is likely to be called repeatedly on the same rel;
1423 * and the input subpath should always be the same (the cheapest_total path
1424 * for the rel). So we cache the result.
1425 */
1426 UniquePath *
create_unique_path(PlannerInfo * root,RelOptInfo * rel,Path * subpath,SpecialJoinInfo * sjinfo)1427 create_unique_path(PlannerInfo *root, RelOptInfo *rel, Path *subpath,
1428 SpecialJoinInfo *sjinfo)
1429 {
1430 UniquePath *pathnode;
1431 Path sort_path; /* dummy for result of cost_sort */
1432 Path agg_path; /* dummy for result of cost_agg */
1433 MemoryContext oldcontext;
1434 int numCols;
1435
1436 /* Caller made a mistake if subpath isn't cheapest_total ... */
1437 Assert(subpath == rel->cheapest_total_path);
1438 Assert(subpath->parent == rel);
1439 /* ... or if SpecialJoinInfo is the wrong one */
1440 Assert(sjinfo->jointype == JOIN_SEMI);
1441 Assert(bms_equal(rel->relids, sjinfo->syn_righthand));
1442
1443 /* If result already cached, return it */
1444 if (rel->cheapest_unique_path)
1445 return (UniquePath *) rel->cheapest_unique_path;
1446
1447 /* If it's not possible to unique-ify, return NULL */
1448 if (!(sjinfo->semi_can_btree || sjinfo->semi_can_hash))
1449 return NULL;
1450
1451 /*
1452 * We must ensure path struct and subsidiary data are allocated in main
1453 * planning context; otherwise GEQO memory management causes trouble.
1454 */
1455 oldcontext = MemoryContextSwitchTo(root->planner_cxt);
1456
1457 pathnode = makeNode(UniquePath);
1458
1459 pathnode->path.pathtype = T_Unique;
1460 pathnode->path.parent = rel;
1461 pathnode->path.pathtarget = rel->reltarget;
1462 pathnode->path.param_info = subpath->param_info;
1463 pathnode->path.parallel_aware = false;
1464 pathnode->path.parallel_safe = rel->consider_parallel &&
1465 subpath->parallel_safe;
1466 pathnode->path.parallel_workers = subpath->parallel_workers;
1467
1468 /*
1469 * Assume the output is unsorted, since we don't necessarily have pathkeys
1470 * to represent it. (This might get overridden below.)
1471 */
1472 pathnode->path.pathkeys = NIL;
1473
1474 pathnode->subpath = subpath;
1475 pathnode->in_operators = sjinfo->semi_operators;
1476 pathnode->uniq_exprs = sjinfo->semi_rhs_exprs;
1477
1478 /*
1479 * If the input is a relation and it has a unique index that proves the
1480 * semi_rhs_exprs are unique, then we don't need to do anything. Note
1481 * that relation_has_unique_index_for automatically considers restriction
1482 * clauses for the rel, as well.
1483 */
1484 if (rel->rtekind == RTE_RELATION && sjinfo->semi_can_btree &&
1485 relation_has_unique_index_for(root, rel, NIL,
1486 sjinfo->semi_rhs_exprs,
1487 sjinfo->semi_operators))
1488 {
1489 pathnode->umethod = UNIQUE_PATH_NOOP;
1490 pathnode->path.rows = rel->rows;
1491 pathnode->path.startup_cost = subpath->startup_cost;
1492 pathnode->path.total_cost = subpath->total_cost;
1493 pathnode->path.pathkeys = subpath->pathkeys;
1494
1495 rel->cheapest_unique_path = (Path *) pathnode;
1496
1497 MemoryContextSwitchTo(oldcontext);
1498
1499 return pathnode;
1500 }
1501
1502 /*
1503 * If the input is a subquery whose output must be unique already, then we
1504 * don't need to do anything. The test for uniqueness has to consider
1505 * exactly which columns we are extracting; for example "SELECT DISTINCT
1506 * x,y" doesn't guarantee that x alone is distinct. So we cannot check for
1507 * this optimization unless semi_rhs_exprs consists only of simple Vars
1508 * referencing subquery outputs. (Possibly we could do something with
1509 * expressions in the subquery outputs, too, but for now keep it simple.)
1510 */
1511 if (rel->rtekind == RTE_SUBQUERY)
1512 {
1513 RangeTblEntry *rte = planner_rt_fetch(rel->relid, root);
1514
1515 if (query_supports_distinctness(rte->subquery))
1516 {
1517 List *sub_tlist_colnos;
1518
1519 sub_tlist_colnos = translate_sub_tlist(sjinfo->semi_rhs_exprs,
1520 rel->relid);
1521
1522 if (sub_tlist_colnos &&
1523 query_is_distinct_for(rte->subquery,
1524 sub_tlist_colnos,
1525 sjinfo->semi_operators))
1526 {
1527 pathnode->umethod = UNIQUE_PATH_NOOP;
1528 pathnode->path.rows = rel->rows;
1529 pathnode->path.startup_cost = subpath->startup_cost;
1530 pathnode->path.total_cost = subpath->total_cost;
1531 pathnode->path.pathkeys = subpath->pathkeys;
1532
1533 rel->cheapest_unique_path = (Path *) pathnode;
1534
1535 MemoryContextSwitchTo(oldcontext);
1536
1537 return pathnode;
1538 }
1539 }
1540 }
1541
1542 /* Estimate number of output rows */
1543 pathnode->path.rows = estimate_num_groups(root,
1544 sjinfo->semi_rhs_exprs,
1545 rel->rows,
1546 NULL);
1547 numCols = list_length(sjinfo->semi_rhs_exprs);
1548
1549 if (sjinfo->semi_can_btree)
1550 {
1551 /*
1552 * Estimate cost for sort+unique implementation
1553 */
1554 cost_sort(&sort_path, root, NIL,
1555 subpath->total_cost,
1556 rel->rows,
1557 subpath->pathtarget->width,
1558 0.0,
1559 work_mem,
1560 -1.0);
1561
1562 /*
1563 * Charge one cpu_operator_cost per comparison per input tuple. We
1564 * assume all columns get compared at most of the tuples. (XXX
1565 * probably this is an overestimate.) This should agree with
1566 * create_upper_unique_path.
1567 */
1568 sort_path.total_cost += cpu_operator_cost * rel->rows * numCols;
1569 }
1570
1571 if (sjinfo->semi_can_hash)
1572 {
1573 /*
1574 * Estimate the overhead per hashtable entry at 64 bytes (same as in
1575 * planner.c).
1576 */
1577 int hashentrysize = subpath->pathtarget->width + 64;
1578
1579 if (hashentrysize * pathnode->path.rows > work_mem * 1024L)
1580 {
1581 /*
1582 * We should not try to hash. Hack the SpecialJoinInfo to
1583 * remember this, in case we come through here again.
1584 */
1585 sjinfo->semi_can_hash = false;
1586 }
1587 else
1588 cost_agg(&agg_path, root,
1589 AGG_HASHED, NULL,
1590 numCols, pathnode->path.rows,
1591 subpath->startup_cost,
1592 subpath->total_cost,
1593 rel->rows);
1594 }
1595
1596 if (sjinfo->semi_can_btree && sjinfo->semi_can_hash)
1597 {
1598 if (agg_path.total_cost < sort_path.total_cost)
1599 pathnode->umethod = UNIQUE_PATH_HASH;
1600 else
1601 pathnode->umethod = UNIQUE_PATH_SORT;
1602 }
1603 else if (sjinfo->semi_can_btree)
1604 pathnode->umethod = UNIQUE_PATH_SORT;
1605 else if (sjinfo->semi_can_hash)
1606 pathnode->umethod = UNIQUE_PATH_HASH;
1607 else
1608 {
1609 /* we can get here only if we abandoned hashing above */
1610 MemoryContextSwitchTo(oldcontext);
1611 return NULL;
1612 }
1613
1614 if (pathnode->umethod == UNIQUE_PATH_HASH)
1615 {
1616 pathnode->path.startup_cost = agg_path.startup_cost;
1617 pathnode->path.total_cost = agg_path.total_cost;
1618 }
1619 else
1620 {
1621 pathnode->path.startup_cost = sort_path.startup_cost;
1622 pathnode->path.total_cost = sort_path.total_cost;
1623 }
1624
1625 rel->cheapest_unique_path = (Path *) pathnode;
1626
1627 MemoryContextSwitchTo(oldcontext);
1628
1629 return pathnode;
1630 }
1631
1632 /*
1633 * translate_sub_tlist - get subquery column numbers represented by tlist
1634 *
1635 * The given targetlist usually contains only Vars referencing the given relid.
1636 * Extract their varattnos (ie, the column numbers of the subquery) and return
1637 * as an integer List.
1638 *
1639 * If any of the tlist items is not a simple Var, we cannot determine whether
1640 * the subquery's uniqueness condition (if any) matches ours, so punt and
1641 * return NIL.
1642 */
1643 static List *
translate_sub_tlist(List * tlist,int relid)1644 translate_sub_tlist(List *tlist, int relid)
1645 {
1646 List *result = NIL;
1647 ListCell *l;
1648
1649 foreach(l, tlist)
1650 {
1651 Var *var = (Var *) lfirst(l);
1652
1653 if (!var || !IsA(var, Var) ||
1654 var->varno != relid)
1655 return NIL; /* punt */
1656
1657 result = lappend_int(result, var->varattno);
1658 }
1659 return result;
1660 }
1661
1662 /*
1663 * create_gather_path
1664 * Creates a path corresponding to a gather scan, returning the
1665 * pathnode.
1666 *
1667 * 'rows' may optionally be set to override row estimates from other sources.
1668 */
1669 GatherPath *
create_gather_path(PlannerInfo * root,RelOptInfo * rel,Path * subpath,PathTarget * target,Relids required_outer,double * rows)1670 create_gather_path(PlannerInfo *root, RelOptInfo *rel, Path *subpath,
1671 PathTarget *target, Relids required_outer, double *rows)
1672 {
1673 GatherPath *pathnode = makeNode(GatherPath);
1674
1675 Assert(subpath->parallel_safe);
1676
1677 pathnode->path.pathtype = T_Gather;
1678 pathnode->path.parent = rel;
1679 pathnode->path.pathtarget = target;
1680 pathnode->path.param_info = get_baserel_parampathinfo(root, rel,
1681 required_outer);
1682 pathnode->path.parallel_aware = false;
1683 pathnode->path.parallel_safe = false;
1684 pathnode->path.parallel_workers = 0;
1685 pathnode->path.pathkeys = NIL; /* Gather has unordered result */
1686
1687 pathnode->subpath = subpath;
1688 pathnode->num_workers = subpath->parallel_workers;
1689 pathnode->single_copy = false;
1690
1691 if (pathnode->num_workers == 0)
1692 {
1693 pathnode->path.pathkeys = subpath->pathkeys;
1694 pathnode->num_workers = 1;
1695 pathnode->single_copy = true;
1696 }
1697
1698 cost_gather(pathnode, root, rel, pathnode->path.param_info, rows);
1699
1700 return pathnode;
1701 }
1702
1703 /*
1704 * create_subqueryscan_path
1705 * Creates a path corresponding to a scan of a subquery,
1706 * returning the pathnode.
1707 */
1708 SubqueryScanPath *
create_subqueryscan_path(PlannerInfo * root,RelOptInfo * rel,Path * subpath,List * pathkeys,Relids required_outer)1709 create_subqueryscan_path(PlannerInfo *root, RelOptInfo *rel, Path *subpath,
1710 List *pathkeys, Relids required_outer)
1711 {
1712 SubqueryScanPath *pathnode = makeNode(SubqueryScanPath);
1713
1714 pathnode->path.pathtype = T_SubqueryScan;
1715 pathnode->path.parent = rel;
1716 pathnode->path.pathtarget = rel->reltarget;
1717 pathnode->path.param_info = get_baserel_parampathinfo(root, rel,
1718 required_outer);
1719 pathnode->path.parallel_aware = false;
1720 pathnode->path.parallel_safe = rel->consider_parallel &&
1721 subpath->parallel_safe;
1722 pathnode->path.parallel_workers = subpath->parallel_workers;
1723 pathnode->path.pathkeys = pathkeys;
1724 pathnode->subpath = subpath;
1725
1726 cost_subqueryscan(pathnode, root, rel, pathnode->path.param_info);
1727
1728 return pathnode;
1729 }
1730
1731 /*
1732 * create_functionscan_path
1733 * Creates a path corresponding to a sequential scan of a function,
1734 * returning the pathnode.
1735 */
1736 Path *
create_functionscan_path(PlannerInfo * root,RelOptInfo * rel,List * pathkeys,Relids required_outer)1737 create_functionscan_path(PlannerInfo *root, RelOptInfo *rel,
1738 List *pathkeys, Relids required_outer)
1739 {
1740 Path *pathnode = makeNode(Path);
1741
1742 pathnode->pathtype = T_FunctionScan;
1743 pathnode->parent = rel;
1744 pathnode->pathtarget = rel->reltarget;
1745 pathnode->param_info = get_baserel_parampathinfo(root, rel,
1746 required_outer);
1747 pathnode->parallel_aware = false;
1748 pathnode->parallel_safe = rel->consider_parallel;
1749 pathnode->parallel_workers = 0;
1750 pathnode->pathkeys = pathkeys;
1751
1752 cost_functionscan(pathnode, root, rel, pathnode->param_info);
1753
1754 return pathnode;
1755 }
1756
1757 /*
1758 * create_valuesscan_path
1759 * Creates a path corresponding to a scan of a VALUES list,
1760 * returning the pathnode.
1761 */
1762 Path *
create_valuesscan_path(PlannerInfo * root,RelOptInfo * rel,Relids required_outer)1763 create_valuesscan_path(PlannerInfo *root, RelOptInfo *rel,
1764 Relids required_outer)
1765 {
1766 Path *pathnode = makeNode(Path);
1767
1768 pathnode->pathtype = T_ValuesScan;
1769 pathnode->parent = rel;
1770 pathnode->pathtarget = rel->reltarget;
1771 pathnode->param_info = get_baserel_parampathinfo(root, rel,
1772 required_outer);
1773 pathnode->parallel_aware = false;
1774 pathnode->parallel_safe = rel->consider_parallel;
1775 pathnode->parallel_workers = 0;
1776 pathnode->pathkeys = NIL; /* result is always unordered */
1777
1778 cost_valuesscan(pathnode, root, rel, pathnode->param_info);
1779
1780 return pathnode;
1781 }
1782
1783 /*
1784 * create_ctescan_path
1785 * Creates a path corresponding to a scan of a non-self-reference CTE,
1786 * returning the pathnode.
1787 */
1788 Path *
create_ctescan_path(PlannerInfo * root,RelOptInfo * rel,Relids required_outer)1789 create_ctescan_path(PlannerInfo *root, RelOptInfo *rel, Relids required_outer)
1790 {
1791 Path *pathnode = makeNode(Path);
1792
1793 pathnode->pathtype = T_CteScan;
1794 pathnode->parent = rel;
1795 pathnode->pathtarget = rel->reltarget;
1796 pathnode->param_info = get_baserel_parampathinfo(root, rel,
1797 required_outer);
1798 pathnode->parallel_aware = false;
1799 pathnode->parallel_safe = rel->consider_parallel;
1800 pathnode->parallel_workers = 0;
1801 pathnode->pathkeys = NIL; /* XXX for now, result is always unordered */
1802
1803 cost_ctescan(pathnode, root, rel, pathnode->param_info);
1804
1805 return pathnode;
1806 }
1807
1808 /*
1809 * create_worktablescan_path
1810 * Creates a path corresponding to a scan of a self-reference CTE,
1811 * returning the pathnode.
1812 */
1813 Path *
create_worktablescan_path(PlannerInfo * root,RelOptInfo * rel,Relids required_outer)1814 create_worktablescan_path(PlannerInfo *root, RelOptInfo *rel,
1815 Relids required_outer)
1816 {
1817 Path *pathnode = makeNode(Path);
1818
1819 pathnode->pathtype = T_WorkTableScan;
1820 pathnode->parent = rel;
1821 pathnode->pathtarget = rel->reltarget;
1822 pathnode->param_info = get_baserel_parampathinfo(root, rel,
1823 required_outer);
1824 pathnode->parallel_aware = false;
1825 pathnode->parallel_safe = rel->consider_parallel;
1826 pathnode->parallel_workers = 0;
1827 pathnode->pathkeys = NIL; /* result is always unordered */
1828
1829 /* Cost is the same as for a regular CTE scan */
1830 cost_ctescan(pathnode, root, rel, pathnode->param_info);
1831
1832 return pathnode;
1833 }
1834
1835 /*
1836 * create_foreignscan_path
1837 * Creates a path corresponding to a scan of a foreign table, foreign join,
1838 * or foreign upper-relation processing, returning the pathnode.
1839 *
1840 * This function is never called from core Postgres; rather, it's expected
1841 * to be called by the GetForeignPaths, GetForeignJoinPaths, or
1842 * GetForeignUpperPaths function of a foreign data wrapper. We make the FDW
1843 * supply all fields of the path, since we do not have any way to calculate
1844 * them in core. However, there is a usually-sane default for the pathtarget
1845 * (rel->reltarget), so we let a NULL for "target" select that.
1846 */
1847 ForeignPath *
create_foreignscan_path(PlannerInfo * root,RelOptInfo * rel,PathTarget * target,double rows,Cost startup_cost,Cost total_cost,List * pathkeys,Relids required_outer,Path * fdw_outerpath,List * fdw_private)1848 create_foreignscan_path(PlannerInfo *root, RelOptInfo *rel,
1849 PathTarget *target,
1850 double rows, Cost startup_cost, Cost total_cost,
1851 List *pathkeys,
1852 Relids required_outer,
1853 Path *fdw_outerpath,
1854 List *fdw_private)
1855 {
1856 ForeignPath *pathnode = makeNode(ForeignPath);
1857
1858 /*
1859 * Since the path's required_outer should always include all the rel's
1860 * lateral_relids, forcibly add those if necessary. This is a bit of a
1861 * hack, but up till early 2019 the contrib FDWs failed to ensure that,
1862 * and it's likely that the same error has propagated into many external
1863 * FDWs. Don't risk modifying the passed-in relid set here.
1864 */
1865 if (rel->lateral_relids && !bms_is_subset(rel->lateral_relids,
1866 required_outer))
1867 required_outer = bms_union(required_outer, rel->lateral_relids);
1868
1869 /*
1870 * Although this function is only designed to be used for scans of
1871 * baserels, before v12 postgres_fdw abused it to make paths for join and
1872 * upper rels. It will work for such cases as long as required_outer is
1873 * empty (otherwise get_baserel_parampathinfo does the wrong thing), which
1874 * fortunately is the expected case for now.
1875 */
1876 if (!bms_is_empty(required_outer) &&
1877 !(rel->reloptkind == RELOPT_BASEREL ||
1878 rel->reloptkind == RELOPT_OTHER_MEMBER_REL))
1879 elog(ERROR, "parameterized foreign joins are not supported yet");
1880
1881 pathnode->path.pathtype = T_ForeignScan;
1882 pathnode->path.parent = rel;
1883 pathnode->path.pathtarget = target ? target : rel->reltarget;
1884 pathnode->path.param_info = get_baserel_parampathinfo(root, rel,
1885 required_outer);
1886 pathnode->path.parallel_aware = false;
1887 pathnode->path.parallel_safe = rel->consider_parallel;
1888 pathnode->path.parallel_workers = 0;
1889 pathnode->path.rows = rows;
1890 pathnode->path.startup_cost = startup_cost;
1891 pathnode->path.total_cost = total_cost;
1892 pathnode->path.pathkeys = pathkeys;
1893
1894 pathnode->fdw_outerpath = fdw_outerpath;
1895 pathnode->fdw_private = fdw_private;
1896
1897 return pathnode;
1898 }
1899
1900 /*
1901 * calc_nestloop_required_outer
1902 * Compute the required_outer set for a nestloop join path
1903 *
1904 * Note: result must not share storage with either input
1905 */
1906 Relids
calc_nestloop_required_outer(Path * outer_path,Path * inner_path)1907 calc_nestloop_required_outer(Path *outer_path, Path *inner_path)
1908 {
1909 Relids outer_paramrels = PATH_REQ_OUTER(outer_path);
1910 Relids inner_paramrels = PATH_REQ_OUTER(inner_path);
1911 Relids required_outer;
1912
1913 /* inner_path can require rels from outer path, but not vice versa */
1914 Assert(!bms_overlap(outer_paramrels, inner_path->parent->relids));
1915 /* easy case if inner path is not parameterized */
1916 if (!inner_paramrels)
1917 return bms_copy(outer_paramrels);
1918 /* else, form the union ... */
1919 required_outer = bms_union(outer_paramrels, inner_paramrels);
1920 /* ... and remove any mention of now-satisfied outer rels */
1921 required_outer = bms_del_members(required_outer,
1922 outer_path->parent->relids);
1923 /* maintain invariant that required_outer is exactly NULL if empty */
1924 if (bms_is_empty(required_outer))
1925 {
1926 bms_free(required_outer);
1927 required_outer = NULL;
1928 }
1929 return required_outer;
1930 }
1931
1932 /*
1933 * calc_non_nestloop_required_outer
1934 * Compute the required_outer set for a merge or hash join path
1935 *
1936 * Note: result must not share storage with either input
1937 */
1938 Relids
calc_non_nestloop_required_outer(Path * outer_path,Path * inner_path)1939 calc_non_nestloop_required_outer(Path *outer_path, Path *inner_path)
1940 {
1941 Relids outer_paramrels = PATH_REQ_OUTER(outer_path);
1942 Relids inner_paramrels = PATH_REQ_OUTER(inner_path);
1943 Relids required_outer;
1944
1945 /* neither path can require rels from the other */
1946 Assert(!bms_overlap(outer_paramrels, inner_path->parent->relids));
1947 Assert(!bms_overlap(inner_paramrels, outer_path->parent->relids));
1948 /* form the union ... */
1949 required_outer = bms_union(outer_paramrels, inner_paramrels);
1950 /* we do not need an explicit test for empty; bms_union gets it right */
1951 return required_outer;
1952 }
1953
1954 /*
1955 * create_nestloop_path
1956 * Creates a pathnode corresponding to a nestloop join between two
1957 * relations.
1958 *
1959 * 'joinrel' is the join relation.
1960 * 'jointype' is the type of join required
1961 * 'workspace' is the result from initial_cost_nestloop
1962 * 'sjinfo' is extra info about the join for selectivity estimation
1963 * 'semifactors' contains valid data if jointype is SEMI or ANTI
1964 * 'outer_path' is the outer path
1965 * 'inner_path' is the inner path
1966 * 'restrict_clauses' are the RestrictInfo nodes to apply at the join
1967 * 'pathkeys' are the path keys of the new join path
1968 * 'required_outer' is the set of required outer rels
1969 *
1970 * Returns the resulting path node.
1971 */
1972 NestPath *
create_nestloop_path(PlannerInfo * root,RelOptInfo * joinrel,JoinType jointype,JoinCostWorkspace * workspace,SpecialJoinInfo * sjinfo,SemiAntiJoinFactors * semifactors,Path * outer_path,Path * inner_path,List * restrict_clauses,List * pathkeys,Relids required_outer)1973 create_nestloop_path(PlannerInfo *root,
1974 RelOptInfo *joinrel,
1975 JoinType jointype,
1976 JoinCostWorkspace *workspace,
1977 SpecialJoinInfo *sjinfo,
1978 SemiAntiJoinFactors *semifactors,
1979 Path *outer_path,
1980 Path *inner_path,
1981 List *restrict_clauses,
1982 List *pathkeys,
1983 Relids required_outer)
1984 {
1985 NestPath *pathnode = makeNode(NestPath);
1986 Relids inner_req_outer = PATH_REQ_OUTER(inner_path);
1987
1988 /*
1989 * If the inner path is parameterized by the outer, we must drop any
1990 * restrict_clauses that are due to be moved into the inner path. We have
1991 * to do this now, rather than postpone the work till createplan time,
1992 * because the restrict_clauses list can affect the size and cost
1993 * estimates for this path.
1994 */
1995 if (bms_overlap(inner_req_outer, outer_path->parent->relids))
1996 {
1997 Relids inner_and_outer = bms_union(inner_path->parent->relids,
1998 inner_req_outer);
1999 List *jclauses = NIL;
2000 ListCell *lc;
2001
2002 foreach(lc, restrict_clauses)
2003 {
2004 RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
2005
2006 if (!join_clause_is_movable_into(rinfo,
2007 inner_path->parent->relids,
2008 inner_and_outer))
2009 jclauses = lappend(jclauses, rinfo);
2010 }
2011 restrict_clauses = jclauses;
2012 }
2013
2014 pathnode->path.pathtype = T_NestLoop;
2015 pathnode->path.parent = joinrel;
2016 pathnode->path.pathtarget = joinrel->reltarget;
2017 pathnode->path.param_info =
2018 get_joinrel_parampathinfo(root,
2019 joinrel,
2020 outer_path,
2021 inner_path,
2022 sjinfo,
2023 required_outer,
2024 &restrict_clauses);
2025 pathnode->path.parallel_aware = false;
2026 pathnode->path.parallel_safe = joinrel->consider_parallel &&
2027 outer_path->parallel_safe && inner_path->parallel_safe;
2028 /* This is a foolish way to estimate parallel_workers, but for now... */
2029 pathnode->path.parallel_workers = outer_path->parallel_workers;
2030 pathnode->path.pathkeys = pathkeys;
2031 pathnode->jointype = jointype;
2032 pathnode->outerjoinpath = outer_path;
2033 pathnode->innerjoinpath = inner_path;
2034 pathnode->joinrestrictinfo = restrict_clauses;
2035
2036 final_cost_nestloop(root, pathnode, workspace, sjinfo, semifactors);
2037
2038 return pathnode;
2039 }
2040
2041 /*
2042 * create_mergejoin_path
2043 * Creates a pathnode corresponding to a mergejoin join between
2044 * two relations
2045 *
2046 * 'joinrel' is the join relation
2047 * 'jointype' is the type of join required
2048 * 'workspace' is the result from initial_cost_mergejoin
2049 * 'sjinfo' is extra info about the join for selectivity estimation
2050 * 'outer_path' is the outer path
2051 * 'inner_path' is the inner path
2052 * 'restrict_clauses' are the RestrictInfo nodes to apply at the join
2053 * 'pathkeys' are the path keys of the new join path
2054 * 'required_outer' is the set of required outer rels
2055 * 'mergeclauses' are the RestrictInfo nodes to use as merge clauses
2056 * (this should be a subset of the restrict_clauses list)
2057 * 'outersortkeys' are the sort varkeys for the outer relation
2058 * 'innersortkeys' are the sort varkeys for the inner relation
2059 */
2060 MergePath *
create_mergejoin_path(PlannerInfo * root,RelOptInfo * joinrel,JoinType jointype,JoinCostWorkspace * workspace,SpecialJoinInfo * sjinfo,Path * outer_path,Path * inner_path,List * restrict_clauses,List * pathkeys,Relids required_outer,List * mergeclauses,List * outersortkeys,List * innersortkeys)2061 create_mergejoin_path(PlannerInfo *root,
2062 RelOptInfo *joinrel,
2063 JoinType jointype,
2064 JoinCostWorkspace *workspace,
2065 SpecialJoinInfo *sjinfo,
2066 Path *outer_path,
2067 Path *inner_path,
2068 List *restrict_clauses,
2069 List *pathkeys,
2070 Relids required_outer,
2071 List *mergeclauses,
2072 List *outersortkeys,
2073 List *innersortkeys)
2074 {
2075 MergePath *pathnode = makeNode(MergePath);
2076
2077 pathnode->jpath.path.pathtype = T_MergeJoin;
2078 pathnode->jpath.path.parent = joinrel;
2079 pathnode->jpath.path.pathtarget = joinrel->reltarget;
2080 pathnode->jpath.path.param_info =
2081 get_joinrel_parampathinfo(root,
2082 joinrel,
2083 outer_path,
2084 inner_path,
2085 sjinfo,
2086 required_outer,
2087 &restrict_clauses);
2088 pathnode->jpath.path.parallel_aware = false;
2089 pathnode->jpath.path.parallel_safe = joinrel->consider_parallel &&
2090 outer_path->parallel_safe && inner_path->parallel_safe;
2091 /* This is a foolish way to estimate parallel_workers, but for now... */
2092 pathnode->jpath.path.parallel_workers = outer_path->parallel_workers;
2093 pathnode->jpath.path.pathkeys = pathkeys;
2094 pathnode->jpath.jointype = jointype;
2095 pathnode->jpath.outerjoinpath = outer_path;
2096 pathnode->jpath.innerjoinpath = inner_path;
2097 pathnode->jpath.joinrestrictinfo = restrict_clauses;
2098 pathnode->path_mergeclauses = mergeclauses;
2099 pathnode->outersortkeys = outersortkeys;
2100 pathnode->innersortkeys = innersortkeys;
2101 /* pathnode->materialize_inner will be set by final_cost_mergejoin */
2102
2103 final_cost_mergejoin(root, pathnode, workspace, sjinfo);
2104
2105 return pathnode;
2106 }
2107
2108 /*
2109 * create_hashjoin_path
2110 * Creates a pathnode corresponding to a hash join between two relations.
2111 *
2112 * 'joinrel' is the join relation
2113 * 'jointype' is the type of join required
2114 * 'workspace' is the result from initial_cost_hashjoin
2115 * 'sjinfo' is extra info about the join for selectivity estimation
2116 * 'semifactors' contains valid data if jointype is SEMI or ANTI
2117 * 'outer_path' is the cheapest outer path
2118 * 'inner_path' is the cheapest inner path
2119 * 'restrict_clauses' are the RestrictInfo nodes to apply at the join
2120 * 'required_outer' is the set of required outer rels
2121 * 'hashclauses' are the RestrictInfo nodes to use as hash clauses
2122 * (this should be a subset of the restrict_clauses list)
2123 */
2124 HashPath *
create_hashjoin_path(PlannerInfo * root,RelOptInfo * joinrel,JoinType jointype,JoinCostWorkspace * workspace,SpecialJoinInfo * sjinfo,SemiAntiJoinFactors * semifactors,Path * outer_path,Path * inner_path,List * restrict_clauses,Relids required_outer,List * hashclauses)2125 create_hashjoin_path(PlannerInfo *root,
2126 RelOptInfo *joinrel,
2127 JoinType jointype,
2128 JoinCostWorkspace *workspace,
2129 SpecialJoinInfo *sjinfo,
2130 SemiAntiJoinFactors *semifactors,
2131 Path *outer_path,
2132 Path *inner_path,
2133 List *restrict_clauses,
2134 Relids required_outer,
2135 List *hashclauses)
2136 {
2137 HashPath *pathnode = makeNode(HashPath);
2138
2139 pathnode->jpath.path.pathtype = T_HashJoin;
2140 pathnode->jpath.path.parent = joinrel;
2141 pathnode->jpath.path.pathtarget = joinrel->reltarget;
2142 pathnode->jpath.path.param_info =
2143 get_joinrel_parampathinfo(root,
2144 joinrel,
2145 outer_path,
2146 inner_path,
2147 sjinfo,
2148 required_outer,
2149 &restrict_clauses);
2150 pathnode->jpath.path.parallel_aware = false;
2151 pathnode->jpath.path.parallel_safe = joinrel->consider_parallel &&
2152 outer_path->parallel_safe && inner_path->parallel_safe;
2153 /* This is a foolish way to estimate parallel_workers, but for now... */
2154 pathnode->jpath.path.parallel_workers = outer_path->parallel_workers;
2155
2156 /*
2157 * A hashjoin never has pathkeys, since its output ordering is
2158 * unpredictable due to possible batching. XXX If the inner relation is
2159 * small enough, we could instruct the executor that it must not batch,
2160 * and then we could assume that the output inherits the outer relation's
2161 * ordering, which might save a sort step. However there is considerable
2162 * downside if our estimate of the inner relation size is badly off. For
2163 * the moment we don't risk it. (Note also that if we wanted to take this
2164 * seriously, joinpath.c would have to consider many more paths for the
2165 * outer rel than it does now.)
2166 */
2167 pathnode->jpath.path.pathkeys = NIL;
2168 pathnode->jpath.jointype = jointype;
2169 pathnode->jpath.outerjoinpath = outer_path;
2170 pathnode->jpath.innerjoinpath = inner_path;
2171 pathnode->jpath.joinrestrictinfo = restrict_clauses;
2172 pathnode->path_hashclauses = hashclauses;
2173 /* final_cost_hashjoin will fill in pathnode->num_batches */
2174
2175 final_cost_hashjoin(root, pathnode, workspace, sjinfo, semifactors);
2176
2177 return pathnode;
2178 }
2179
2180 /*
2181 * create_projection_path
2182 * Creates a pathnode that represents performing a projection.
2183 *
2184 * 'rel' is the parent relation associated with the result
2185 * 'subpath' is the path representing the source of data
2186 * 'target' is the PathTarget to be computed
2187 */
2188 ProjectionPath *
create_projection_path(PlannerInfo * root,RelOptInfo * rel,Path * subpath,PathTarget * target)2189 create_projection_path(PlannerInfo *root,
2190 RelOptInfo *rel,
2191 Path *subpath,
2192 PathTarget *target)
2193 {
2194 ProjectionPath *pathnode = makeNode(ProjectionPath);
2195 PathTarget *oldtarget = subpath->pathtarget;
2196
2197 pathnode->path.pathtype = T_Result;
2198 pathnode->path.parent = rel;
2199 pathnode->path.pathtarget = target;
2200 /* For now, assume we are above any joins, so no parameterization */
2201 pathnode->path.param_info = NULL;
2202 pathnode->path.parallel_aware = false;
2203 pathnode->path.parallel_safe = rel->consider_parallel &&
2204 subpath->parallel_safe &&
2205 !has_parallel_hazard((Node *) target->exprs, false);
2206 pathnode->path.parallel_workers = subpath->parallel_workers;
2207 /* Projection does not change the sort order */
2208 pathnode->path.pathkeys = subpath->pathkeys;
2209
2210 pathnode->subpath = subpath;
2211
2212 /*
2213 * We might not need a separate Result node. If the input plan node type
2214 * can project, we can just tell it to project something else. Or, if it
2215 * can't project but the desired target has the same expression list as
2216 * what the input will produce anyway, we can still give it the desired
2217 * tlist (possibly changing its ressortgroupref labels, but nothing else).
2218 * Note: in the latter case, create_projection_plan has to recheck our
2219 * conclusion; see comments therein.
2220 */
2221 if (is_projection_capable_path(subpath) ||
2222 equal(oldtarget->exprs, target->exprs))
2223 {
2224 /* No separate Result node needed */
2225 pathnode->dummypp = true;
2226
2227 /*
2228 * Set cost of plan as subpath's cost, adjusted for tlist replacement.
2229 */
2230 pathnode->path.rows = subpath->rows;
2231 pathnode->path.startup_cost = subpath->startup_cost +
2232 (target->cost.startup - oldtarget->cost.startup);
2233 pathnode->path.total_cost = subpath->total_cost +
2234 (target->cost.startup - oldtarget->cost.startup) +
2235 (target->cost.per_tuple - oldtarget->cost.per_tuple) * subpath->rows;
2236 }
2237 else
2238 {
2239 /* We really do need the Result node */
2240 pathnode->dummypp = false;
2241
2242 /*
2243 * The Result node's cost is cpu_tuple_cost per row, plus the cost of
2244 * evaluating the tlist. There is no qual to worry about.
2245 */
2246 pathnode->path.rows = subpath->rows;
2247 pathnode->path.startup_cost = subpath->startup_cost +
2248 target->cost.startup;
2249 pathnode->path.total_cost = subpath->total_cost +
2250 target->cost.startup +
2251 (cpu_tuple_cost + target->cost.per_tuple) * subpath->rows;
2252 }
2253
2254 return pathnode;
2255 }
2256
2257 /*
2258 * apply_projection_to_path
2259 * Add a projection step, or just apply the target directly to given path.
2260 *
2261 * This has the same net effect as create_projection_path(), except that if
2262 * a separate Result plan node isn't needed, we just replace the given path's
2263 * pathtarget with the desired one. This must be used only when the caller
2264 * knows that the given path isn't referenced elsewhere and so can be modified
2265 * in-place.
2266 *
2267 * If the input path is a GatherPath, we try to push the new target down to
2268 * its input as well; this is a yet more invasive modification of the input
2269 * path, which create_projection_path() can't do.
2270 *
2271 * Note that we mustn't change the source path's parent link; so when it is
2272 * add_path'd to "rel" things will be a bit inconsistent. So far that has
2273 * not caused any trouble.
2274 *
2275 * 'rel' is the parent relation associated with the result
2276 * 'path' is the path representing the source of data
2277 * 'target' is the PathTarget to be computed
2278 */
2279 Path *
apply_projection_to_path(PlannerInfo * root,RelOptInfo * rel,Path * path,PathTarget * target)2280 apply_projection_to_path(PlannerInfo *root,
2281 RelOptInfo *rel,
2282 Path *path,
2283 PathTarget *target)
2284 {
2285 QualCost oldcost;
2286
2287 /*
2288 * If given path can't project, we might need a Result node, so make a
2289 * separate ProjectionPath.
2290 */
2291 if (!is_projection_capable_path(path))
2292 return (Path *) create_projection_path(root, rel, path, target);
2293
2294 /*
2295 * We can just jam the desired tlist into the existing path, being sure to
2296 * update its cost estimates appropriately.
2297 */
2298 oldcost = path->pathtarget->cost;
2299 path->pathtarget = target;
2300
2301 path->startup_cost += target->cost.startup - oldcost.startup;
2302 path->total_cost += target->cost.startup - oldcost.startup +
2303 (target->cost.per_tuple - oldcost.per_tuple) * path->rows;
2304
2305 /*
2306 * If the path happens to be a Gather path, we'd like to arrange for the
2307 * subpath to return the required target list so that workers can help
2308 * project. But if there is something that is not parallel-safe in the
2309 * target expressions, then we can't.
2310 */
2311 if (IsA(path, GatherPath) &&
2312 !has_parallel_hazard((Node *) target->exprs, false))
2313 {
2314 GatherPath *gpath = (GatherPath *) path;
2315
2316 /*
2317 * We always use create_projection_path here, even if the subpath is
2318 * projection-capable, so as to avoid modifying the subpath in place.
2319 * It seems unlikely at present that there could be any other
2320 * references to the subpath, but better safe than sorry.
2321 *
2322 * Note that we don't change the GatherPath's cost estimates; it might
2323 * be appropriate to do so, to reflect the fact that the bulk of the
2324 * target evaluation will happen in workers.
2325 */
2326 gpath->subpath = (Path *)
2327 create_projection_path(root,
2328 gpath->subpath->parent,
2329 gpath->subpath,
2330 target);
2331 }
2332 else if (path->parallel_safe &&
2333 has_parallel_hazard((Node *) target->exprs, false))
2334 {
2335 /*
2336 * We're inserting a parallel-restricted target list into a path
2337 * currently marked parallel-safe, so we have to mark it as no longer
2338 * safe.
2339 */
2340 path->parallel_safe = false;
2341 }
2342
2343 return path;
2344 }
2345
2346 /*
2347 * create_sort_path
2348 * Creates a pathnode that represents performing an explicit sort.
2349 *
2350 * 'rel' is the parent relation associated with the result
2351 * 'subpath' is the path representing the source of data
2352 * 'pathkeys' represents the desired sort order
2353 * 'limit_tuples' is the estimated bound on the number of output tuples,
2354 * or -1 if no LIMIT or couldn't estimate
2355 */
2356 SortPath *
create_sort_path(PlannerInfo * root,RelOptInfo * rel,Path * subpath,List * pathkeys,double limit_tuples)2357 create_sort_path(PlannerInfo *root,
2358 RelOptInfo *rel,
2359 Path *subpath,
2360 List *pathkeys,
2361 double limit_tuples)
2362 {
2363 SortPath *pathnode = makeNode(SortPath);
2364
2365 pathnode->path.pathtype = T_Sort;
2366 pathnode->path.parent = rel;
2367 /* Sort doesn't project, so use source path's pathtarget */
2368 pathnode->path.pathtarget = subpath->pathtarget;
2369 /* For now, assume we are above any joins, so no parameterization */
2370 pathnode->path.param_info = NULL;
2371 pathnode->path.parallel_aware = false;
2372 pathnode->path.parallel_safe = rel->consider_parallel &&
2373 subpath->parallel_safe;
2374 pathnode->path.parallel_workers = subpath->parallel_workers;
2375 pathnode->path.pathkeys = pathkeys;
2376
2377 pathnode->subpath = subpath;
2378
2379 cost_sort(&pathnode->path, root, pathkeys,
2380 subpath->total_cost,
2381 subpath->rows,
2382 subpath->pathtarget->width,
2383 0.0, /* XXX comparison_cost shouldn't be 0? */
2384 work_mem, limit_tuples);
2385
2386 return pathnode;
2387 }
2388
2389 /*
2390 * create_group_path
2391 * Creates a pathnode that represents performing grouping of presorted input
2392 *
2393 * 'rel' is the parent relation associated with the result
2394 * 'subpath' is the path representing the source of data
2395 * 'target' is the PathTarget to be computed
2396 * 'groupClause' is a list of SortGroupClause's representing the grouping
2397 * 'qual' is the HAVING quals if any
2398 * 'numGroups' is the estimated number of groups
2399 */
2400 GroupPath *
create_group_path(PlannerInfo * root,RelOptInfo * rel,Path * subpath,PathTarget * target,List * groupClause,List * qual,double numGroups)2401 create_group_path(PlannerInfo *root,
2402 RelOptInfo *rel,
2403 Path *subpath,
2404 PathTarget *target,
2405 List *groupClause,
2406 List *qual,
2407 double numGroups)
2408 {
2409 GroupPath *pathnode = makeNode(GroupPath);
2410
2411 pathnode->path.pathtype = T_Group;
2412 pathnode->path.parent = rel;
2413 pathnode->path.pathtarget = target;
2414 /* For now, assume we are above any joins, so no parameterization */
2415 pathnode->path.param_info = NULL;
2416 pathnode->path.parallel_aware = false;
2417 pathnode->path.parallel_safe = rel->consider_parallel &&
2418 subpath->parallel_safe;
2419 pathnode->path.parallel_workers = subpath->parallel_workers;
2420 /* Group doesn't change sort ordering */
2421 pathnode->path.pathkeys = subpath->pathkeys;
2422
2423 pathnode->subpath = subpath;
2424
2425 pathnode->groupClause = groupClause;
2426 pathnode->qual = qual;
2427
2428 cost_group(&pathnode->path, root,
2429 list_length(groupClause),
2430 numGroups,
2431 subpath->startup_cost, subpath->total_cost,
2432 subpath->rows);
2433
2434 /* add tlist eval cost for each output row */
2435 pathnode->path.startup_cost += target->cost.startup;
2436 pathnode->path.total_cost += target->cost.startup +
2437 target->cost.per_tuple * pathnode->path.rows;
2438
2439 return pathnode;
2440 }
2441
2442 /*
2443 * create_upper_unique_path
2444 * Creates a pathnode that represents performing an explicit Unique step
2445 * on presorted input.
2446 *
2447 * This produces a Unique plan node, but the use-case is so different from
2448 * create_unique_path that it doesn't seem worth trying to merge the two.
2449 *
2450 * 'rel' is the parent relation associated with the result
2451 * 'subpath' is the path representing the source of data
2452 * 'numCols' is the number of grouping columns
2453 * 'numGroups' is the estimated number of groups
2454 *
2455 * The input path must be sorted on the grouping columns, plus possibly
2456 * additional columns; so the first numCols pathkeys are the grouping columns
2457 */
2458 UpperUniquePath *
create_upper_unique_path(PlannerInfo * root,RelOptInfo * rel,Path * subpath,int numCols,double numGroups)2459 create_upper_unique_path(PlannerInfo *root,
2460 RelOptInfo *rel,
2461 Path *subpath,
2462 int numCols,
2463 double numGroups)
2464 {
2465 UpperUniquePath *pathnode = makeNode(UpperUniquePath);
2466
2467 pathnode->path.pathtype = T_Unique;
2468 pathnode->path.parent = rel;
2469 /* Unique doesn't project, so use source path's pathtarget */
2470 pathnode->path.pathtarget = subpath->pathtarget;
2471 /* For now, assume we are above any joins, so no parameterization */
2472 pathnode->path.param_info = NULL;
2473 pathnode->path.parallel_aware = false;
2474 pathnode->path.parallel_safe = rel->consider_parallel &&
2475 subpath->parallel_safe;
2476 pathnode->path.parallel_workers = subpath->parallel_workers;
2477 /* Unique doesn't change the input ordering */
2478 pathnode->path.pathkeys = subpath->pathkeys;
2479
2480 pathnode->subpath = subpath;
2481 pathnode->numkeys = numCols;
2482
2483 /*
2484 * Charge one cpu_operator_cost per comparison per input tuple. We assume
2485 * all columns get compared at most of the tuples. (XXX probably this is
2486 * an overestimate.)
2487 */
2488 pathnode->path.startup_cost = subpath->startup_cost;
2489 pathnode->path.total_cost = subpath->total_cost +
2490 cpu_operator_cost * subpath->rows * numCols;
2491 pathnode->path.rows = numGroups;
2492
2493 return pathnode;
2494 }
2495
2496 /*
2497 * create_agg_path
2498 * Creates a pathnode that represents performing aggregation/grouping
2499 *
2500 * 'rel' is the parent relation associated with the result
2501 * 'subpath' is the path representing the source of data
2502 * 'target' is the PathTarget to be computed
2503 * 'aggstrategy' is the Agg node's basic implementation strategy
2504 * 'aggsplit' is the Agg node's aggregate-splitting mode
2505 * 'groupClause' is a list of SortGroupClause's representing the grouping
2506 * 'qual' is the HAVING quals if any
2507 * 'aggcosts' contains cost info about the aggregate functions to be computed
2508 * 'numGroups' is the estimated number of groups (1 if not grouping)
2509 */
2510 AggPath *
create_agg_path(PlannerInfo * root,RelOptInfo * rel,Path * subpath,PathTarget * target,AggStrategy aggstrategy,AggSplit aggsplit,List * groupClause,List * qual,const AggClauseCosts * aggcosts,double numGroups)2511 create_agg_path(PlannerInfo *root,
2512 RelOptInfo *rel,
2513 Path *subpath,
2514 PathTarget *target,
2515 AggStrategy aggstrategy,
2516 AggSplit aggsplit,
2517 List *groupClause,
2518 List *qual,
2519 const AggClauseCosts *aggcosts,
2520 double numGroups)
2521 {
2522 AggPath *pathnode = makeNode(AggPath);
2523
2524 pathnode->path.pathtype = T_Agg;
2525 pathnode->path.parent = rel;
2526 pathnode->path.pathtarget = target;
2527 /* For now, assume we are above any joins, so no parameterization */
2528 pathnode->path.param_info = NULL;
2529 pathnode->path.parallel_aware = false;
2530 pathnode->path.parallel_safe = rel->consider_parallel &&
2531 subpath->parallel_safe;
2532 pathnode->path.parallel_workers = subpath->parallel_workers;
2533 if (aggstrategy == AGG_SORTED)
2534 pathnode->path.pathkeys = subpath->pathkeys; /* preserves order */
2535 else
2536 pathnode->path.pathkeys = NIL; /* output is unordered */
2537 pathnode->subpath = subpath;
2538
2539 pathnode->aggstrategy = aggstrategy;
2540 pathnode->aggsplit = aggsplit;
2541 pathnode->numGroups = numGroups;
2542 pathnode->groupClause = groupClause;
2543 pathnode->qual = qual;
2544
2545 cost_agg(&pathnode->path, root,
2546 aggstrategy, aggcosts,
2547 list_length(groupClause), numGroups,
2548 subpath->startup_cost, subpath->total_cost,
2549 subpath->rows);
2550
2551 /* add tlist eval cost for each output row */
2552 pathnode->path.startup_cost += target->cost.startup;
2553 pathnode->path.total_cost += target->cost.startup +
2554 target->cost.per_tuple * pathnode->path.rows;
2555
2556 return pathnode;
2557 }
2558
2559 /*
2560 * create_groupingsets_path
2561 * Creates a pathnode that represents performing GROUPING SETS aggregation
2562 *
2563 * GroupingSetsPath represents sorted grouping with one or more grouping sets.
2564 * The input path's result must be sorted to match the last entry in
2565 * rollup_groupclauses.
2566 *
2567 * 'rel' is the parent relation associated with the result
2568 * 'subpath' is the path representing the source of data
2569 * 'target' is the PathTarget to be computed
2570 * 'having_qual' is the HAVING quals if any
2571 * 'rollup_lists' is a list of grouping sets
2572 * 'rollup_groupclauses' is a list of grouping clauses for grouping sets
2573 * 'agg_costs' contains cost info about the aggregate functions to be computed
2574 * 'numGroups' is the estimated number of groups
2575 */
2576 GroupingSetsPath *
create_groupingsets_path(PlannerInfo * root,RelOptInfo * rel,Path * subpath,PathTarget * target,List * having_qual,List * rollup_lists,List * rollup_groupclauses,const AggClauseCosts * agg_costs,double numGroups)2577 create_groupingsets_path(PlannerInfo *root,
2578 RelOptInfo *rel,
2579 Path *subpath,
2580 PathTarget *target,
2581 List *having_qual,
2582 List *rollup_lists,
2583 List *rollup_groupclauses,
2584 const AggClauseCosts *agg_costs,
2585 double numGroups)
2586 {
2587 GroupingSetsPath *pathnode = makeNode(GroupingSetsPath);
2588 int numGroupCols;
2589
2590 /* The topmost generated Plan node will be an Agg */
2591 pathnode->path.pathtype = T_Agg;
2592 pathnode->path.parent = rel;
2593 pathnode->path.pathtarget = target;
2594 pathnode->path.param_info = subpath->param_info;
2595 pathnode->path.parallel_aware = false;
2596 pathnode->path.parallel_safe = rel->consider_parallel &&
2597 subpath->parallel_safe;
2598 pathnode->path.parallel_workers = subpath->parallel_workers;
2599 pathnode->subpath = subpath;
2600
2601 /*
2602 * Output will be in sorted order by group_pathkeys if, and only if, there
2603 * is a single rollup operation on a non-empty list of grouping
2604 * expressions.
2605 */
2606 if (list_length(rollup_groupclauses) == 1 &&
2607 ((List *) linitial(rollup_groupclauses)) != NIL)
2608 pathnode->path.pathkeys = root->group_pathkeys;
2609 else
2610 pathnode->path.pathkeys = NIL;
2611
2612 pathnode->rollup_groupclauses = rollup_groupclauses;
2613 pathnode->rollup_lists = rollup_lists;
2614 pathnode->qual = having_qual;
2615
2616 Assert(rollup_lists != NIL);
2617 Assert(list_length(rollup_lists) == list_length(rollup_groupclauses));
2618
2619 /* Account for cost of the topmost Agg node */
2620 numGroupCols = list_length((List *) linitial((List *) llast(rollup_lists)));
2621
2622 cost_agg(&pathnode->path, root,
2623 (numGroupCols > 0) ? AGG_SORTED : AGG_PLAIN,
2624 agg_costs,
2625 numGroupCols,
2626 numGroups,
2627 subpath->startup_cost,
2628 subpath->total_cost,
2629 subpath->rows);
2630
2631 /*
2632 * Add in the costs and output rows of the additional sorting/aggregation
2633 * steps, if any. Only total costs count, since the extra sorts aren't
2634 * run on startup.
2635 */
2636 if (list_length(rollup_lists) > 1)
2637 {
2638 ListCell *lc;
2639
2640 foreach(lc, rollup_lists)
2641 {
2642 List *gsets = (List *) lfirst(lc);
2643 Path sort_path; /* dummy for result of cost_sort */
2644 Path agg_path; /* dummy for result of cost_agg */
2645
2646 /* We must iterate over all but the last rollup_lists element */
2647 if (lnext(lc) == NULL)
2648 break;
2649
2650 /* Account for cost of sort, but don't charge input cost again */
2651 cost_sort(&sort_path, root, NIL,
2652 0.0,
2653 subpath->rows,
2654 subpath->pathtarget->width,
2655 0.0,
2656 work_mem,
2657 -1.0);
2658
2659 /* Account for cost of aggregation */
2660 numGroupCols = list_length((List *) linitial(gsets));
2661
2662 cost_agg(&agg_path, root,
2663 AGG_SORTED,
2664 agg_costs,
2665 numGroupCols,
2666 numGroups, /* XXX surely not right for all steps? */
2667 sort_path.startup_cost,
2668 sort_path.total_cost,
2669 sort_path.rows);
2670
2671 pathnode->path.total_cost += agg_path.total_cost;
2672 pathnode->path.rows += agg_path.rows;
2673 }
2674 }
2675
2676 /* add tlist eval cost for each output row */
2677 pathnode->path.startup_cost += target->cost.startup;
2678 pathnode->path.total_cost += target->cost.startup +
2679 target->cost.per_tuple * pathnode->path.rows;
2680
2681 return pathnode;
2682 }
2683
2684 /*
2685 * create_minmaxagg_path
2686 * Creates a pathnode that represents computation of MIN/MAX aggregates
2687 *
2688 * 'rel' is the parent relation associated with the result
2689 * 'target' is the PathTarget to be computed
2690 * 'mmaggregates' is a list of MinMaxAggInfo structs
2691 * 'quals' is the HAVING quals if any
2692 */
2693 MinMaxAggPath *
create_minmaxagg_path(PlannerInfo * root,RelOptInfo * rel,PathTarget * target,List * mmaggregates,List * quals)2694 create_minmaxagg_path(PlannerInfo *root,
2695 RelOptInfo *rel,
2696 PathTarget *target,
2697 List *mmaggregates,
2698 List *quals)
2699 {
2700 MinMaxAggPath *pathnode = makeNode(MinMaxAggPath);
2701 Cost initplan_cost;
2702 ListCell *lc;
2703
2704 /* The topmost generated Plan node will be a Result */
2705 pathnode->path.pathtype = T_Result;
2706 pathnode->path.parent = rel;
2707 pathnode->path.pathtarget = target;
2708 /* For now, assume we are above any joins, so no parameterization */
2709 pathnode->path.param_info = NULL;
2710 pathnode->path.parallel_aware = false;
2711 /* A MinMaxAggPath implies use of subplans, so cannot be parallel-safe */
2712 pathnode->path.parallel_safe = false;
2713 pathnode->path.parallel_workers = 0;
2714 /* Result is one unordered row */
2715 pathnode->path.rows = 1;
2716 pathnode->path.pathkeys = NIL;
2717
2718 pathnode->mmaggregates = mmaggregates;
2719 pathnode->quals = quals;
2720
2721 /* Calculate cost of all the initplans ... */
2722 initplan_cost = 0;
2723 foreach(lc, mmaggregates)
2724 {
2725 MinMaxAggInfo *mminfo = (MinMaxAggInfo *) lfirst(lc);
2726
2727 initplan_cost += mminfo->pathcost;
2728 }
2729
2730 /* add tlist eval cost for each output row, plus cpu_tuple_cost */
2731 pathnode->path.startup_cost = initplan_cost + target->cost.startup;
2732 pathnode->path.total_cost = initplan_cost + target->cost.startup +
2733 target->cost.per_tuple + cpu_tuple_cost;
2734
2735 return pathnode;
2736 }
2737
2738 /*
2739 * create_windowagg_path
2740 * Creates a pathnode that represents computation of window functions
2741 *
2742 * 'rel' is the parent relation associated with the result
2743 * 'subpath' is the path representing the source of data
2744 * 'target' is the PathTarget to be computed
2745 * 'windowFuncs' is a list of WindowFunc structs
2746 * 'winclause' is a WindowClause that is common to all the WindowFuncs
2747 * 'winpathkeys' is the pathkeys for the PARTITION keys + ORDER keys
2748 *
2749 * The actual sort order of the input must match winpathkeys, but might
2750 * have additional keys after those.
2751 */
2752 WindowAggPath *
create_windowagg_path(PlannerInfo * root,RelOptInfo * rel,Path * subpath,PathTarget * target,List * windowFuncs,WindowClause * winclause,List * winpathkeys)2753 create_windowagg_path(PlannerInfo *root,
2754 RelOptInfo *rel,
2755 Path *subpath,
2756 PathTarget *target,
2757 List *windowFuncs,
2758 WindowClause *winclause,
2759 List *winpathkeys)
2760 {
2761 WindowAggPath *pathnode = makeNode(WindowAggPath);
2762
2763 pathnode->path.pathtype = T_WindowAgg;
2764 pathnode->path.parent = rel;
2765 pathnode->path.pathtarget = target;
2766 /* For now, assume we are above any joins, so no parameterization */
2767 pathnode->path.param_info = NULL;
2768 pathnode->path.parallel_aware = false;
2769 pathnode->path.parallel_safe = rel->consider_parallel &&
2770 subpath->parallel_safe;
2771 pathnode->path.parallel_workers = subpath->parallel_workers;
2772 /* WindowAgg preserves the input sort order */
2773 pathnode->path.pathkeys = subpath->pathkeys;
2774
2775 pathnode->subpath = subpath;
2776 pathnode->winclause = winclause;
2777 pathnode->winpathkeys = winpathkeys;
2778
2779 /*
2780 * For costing purposes, assume that there are no redundant partitioning
2781 * or ordering columns; it's not worth the trouble to deal with that
2782 * corner case here. So we just pass the unmodified list lengths to
2783 * cost_windowagg.
2784 */
2785 cost_windowagg(&pathnode->path, root,
2786 windowFuncs,
2787 list_length(winclause->partitionClause),
2788 list_length(winclause->orderClause),
2789 subpath->startup_cost,
2790 subpath->total_cost,
2791 subpath->rows);
2792
2793 /* add tlist eval cost for each output row */
2794 pathnode->path.startup_cost += target->cost.startup;
2795 pathnode->path.total_cost += target->cost.startup +
2796 target->cost.per_tuple * pathnode->path.rows;
2797
2798 return pathnode;
2799 }
2800
2801 /*
2802 * create_setop_path
2803 * Creates a pathnode that represents computation of INTERSECT or EXCEPT
2804 *
2805 * 'rel' is the parent relation associated with the result
2806 * 'subpath' is the path representing the source of data
2807 * 'cmd' is the specific semantics (INTERSECT or EXCEPT, with/without ALL)
2808 * 'strategy' is the implementation strategy (sorted or hashed)
2809 * 'distinctList' is a list of SortGroupClause's representing the grouping
2810 * 'flagColIdx' is the column number where the flag column will be, if any
2811 * 'firstFlag' is the flag value for the first input relation when hashing;
2812 * or -1 when sorting
2813 * 'numGroups' is the estimated number of distinct groups
2814 * 'outputRows' is the estimated number of output rows
2815 */
2816 SetOpPath *
create_setop_path(PlannerInfo * root,RelOptInfo * rel,Path * subpath,SetOpCmd cmd,SetOpStrategy strategy,List * distinctList,AttrNumber flagColIdx,int firstFlag,double numGroups,double outputRows)2817 create_setop_path(PlannerInfo *root,
2818 RelOptInfo *rel,
2819 Path *subpath,
2820 SetOpCmd cmd,
2821 SetOpStrategy strategy,
2822 List *distinctList,
2823 AttrNumber flagColIdx,
2824 int firstFlag,
2825 double numGroups,
2826 double outputRows)
2827 {
2828 SetOpPath *pathnode = makeNode(SetOpPath);
2829
2830 pathnode->path.pathtype = T_SetOp;
2831 pathnode->path.parent = rel;
2832 /* SetOp doesn't project, so use source path's pathtarget */
2833 pathnode->path.pathtarget = subpath->pathtarget;
2834 /* For now, assume we are above any joins, so no parameterization */
2835 pathnode->path.param_info = NULL;
2836 pathnode->path.parallel_aware = false;
2837 pathnode->path.parallel_safe = rel->consider_parallel &&
2838 subpath->parallel_safe;
2839 pathnode->path.parallel_workers = subpath->parallel_workers;
2840 /* SetOp preserves the input sort order if in sort mode */
2841 pathnode->path.pathkeys =
2842 (strategy == SETOP_SORTED) ? subpath->pathkeys : NIL;
2843
2844 pathnode->subpath = subpath;
2845 pathnode->cmd = cmd;
2846 pathnode->strategy = strategy;
2847 pathnode->distinctList = distinctList;
2848 pathnode->flagColIdx = flagColIdx;
2849 pathnode->firstFlag = firstFlag;
2850 pathnode->numGroups = numGroups;
2851
2852 /*
2853 * Charge one cpu_operator_cost per comparison per input tuple. We assume
2854 * all columns get compared at most of the tuples.
2855 */
2856 pathnode->path.startup_cost = subpath->startup_cost;
2857 pathnode->path.total_cost = subpath->total_cost +
2858 cpu_operator_cost * subpath->rows * list_length(distinctList);
2859 pathnode->path.rows = outputRows;
2860
2861 return pathnode;
2862 }
2863
2864 /*
2865 * create_recursiveunion_path
2866 * Creates a pathnode that represents a recursive UNION node
2867 *
2868 * 'rel' is the parent relation associated with the result
2869 * 'leftpath' is the source of data for the non-recursive term
2870 * 'rightpath' is the source of data for the recursive term
2871 * 'target' is the PathTarget to be computed
2872 * 'distinctList' is a list of SortGroupClause's representing the grouping
2873 * 'wtParam' is the ID of Param representing work table
2874 * 'numGroups' is the estimated number of groups
2875 *
2876 * For recursive UNION ALL, distinctList is empty and numGroups is zero
2877 */
2878 RecursiveUnionPath *
create_recursiveunion_path(PlannerInfo * root,RelOptInfo * rel,Path * leftpath,Path * rightpath,PathTarget * target,List * distinctList,int wtParam,double numGroups)2879 create_recursiveunion_path(PlannerInfo *root,
2880 RelOptInfo *rel,
2881 Path *leftpath,
2882 Path *rightpath,
2883 PathTarget *target,
2884 List *distinctList,
2885 int wtParam,
2886 double numGroups)
2887 {
2888 RecursiveUnionPath *pathnode = makeNode(RecursiveUnionPath);
2889
2890 pathnode->path.pathtype = T_RecursiveUnion;
2891 pathnode->path.parent = rel;
2892 pathnode->path.pathtarget = target;
2893 /* For now, assume we are above any joins, so no parameterization */
2894 pathnode->path.param_info = NULL;
2895 pathnode->path.parallel_aware = false;
2896 pathnode->path.parallel_safe = rel->consider_parallel &&
2897 leftpath->parallel_safe && rightpath->parallel_safe;
2898 /* Foolish, but we'll do it like joins for now: */
2899 pathnode->path.parallel_workers = leftpath->parallel_workers;
2900 /* RecursiveUnion result is always unsorted */
2901 pathnode->path.pathkeys = NIL;
2902
2903 pathnode->leftpath = leftpath;
2904 pathnode->rightpath = rightpath;
2905 pathnode->distinctList = distinctList;
2906 pathnode->wtParam = wtParam;
2907 pathnode->numGroups = numGroups;
2908
2909 cost_recursive_union(&pathnode->path, leftpath, rightpath);
2910
2911 return pathnode;
2912 }
2913
2914 /*
2915 * create_lockrows_path
2916 * Creates a pathnode that represents acquiring row locks
2917 *
2918 * 'rel' is the parent relation associated with the result
2919 * 'subpath' is the path representing the source of data
2920 * 'rowMarks' is a list of PlanRowMark's
2921 * 'epqParam' is the ID of Param for EvalPlanQual re-eval
2922 */
2923 LockRowsPath *
create_lockrows_path(PlannerInfo * root,RelOptInfo * rel,Path * subpath,List * rowMarks,int epqParam)2924 create_lockrows_path(PlannerInfo *root, RelOptInfo *rel,
2925 Path *subpath, List *rowMarks, int epqParam)
2926 {
2927 LockRowsPath *pathnode = makeNode(LockRowsPath);
2928
2929 pathnode->path.pathtype = T_LockRows;
2930 pathnode->path.parent = rel;
2931 /* LockRows doesn't project, so use source path's pathtarget */
2932 pathnode->path.pathtarget = subpath->pathtarget;
2933 /* For now, assume we are above any joins, so no parameterization */
2934 pathnode->path.param_info = NULL;
2935 pathnode->path.parallel_aware = false;
2936 pathnode->path.parallel_safe = false;
2937 pathnode->path.parallel_workers = 0;
2938 pathnode->path.rows = subpath->rows;
2939
2940 /*
2941 * The result cannot be assumed sorted, since locking might cause the sort
2942 * key columns to be replaced with new values.
2943 */
2944 pathnode->path.pathkeys = NIL;
2945
2946 pathnode->subpath = subpath;
2947 pathnode->rowMarks = rowMarks;
2948 pathnode->epqParam = epqParam;
2949
2950 /*
2951 * We should charge something extra for the costs of row locking and
2952 * possible refetches, but it's hard to say how much. For now, use
2953 * cpu_tuple_cost per row.
2954 */
2955 pathnode->path.startup_cost = subpath->startup_cost;
2956 pathnode->path.total_cost = subpath->total_cost +
2957 cpu_tuple_cost * subpath->rows;
2958
2959 return pathnode;
2960 }
2961
2962 /*
2963 * create_modifytable_path
2964 * Creates a pathnode that represents performing INSERT/UPDATE/DELETE mods
2965 *
2966 * 'rel' is the parent relation associated with the result
2967 * 'operation' is the operation type
2968 * 'canSetTag' is true if we set the command tag/es_processed
2969 * 'nominalRelation' is the parent RT index for use of EXPLAIN
2970 * 'resultRelations' is an integer list of actual RT indexes of target rel(s)
2971 * 'subpaths' is a list of Path(s) producing source data (one per rel)
2972 * 'subroots' is a list of PlannerInfo structs (one per rel)
2973 * 'withCheckOptionLists' is a list of WCO lists (one per rel)
2974 * 'returningLists' is a list of RETURNING tlists (one per rel)
2975 * 'rowMarks' is a list of PlanRowMarks (non-locking only)
2976 * 'onconflict' is the ON CONFLICT clause, or NULL
2977 * 'epqParam' is the ID of Param for EvalPlanQual re-eval
2978 */
2979 ModifyTablePath *
create_modifytable_path(PlannerInfo * root,RelOptInfo * rel,CmdType operation,bool canSetTag,Index nominalRelation,List * resultRelations,List * subpaths,List * subroots,List * withCheckOptionLists,List * returningLists,List * rowMarks,OnConflictExpr * onconflict,int epqParam)2980 create_modifytable_path(PlannerInfo *root, RelOptInfo *rel,
2981 CmdType operation, bool canSetTag,
2982 Index nominalRelation,
2983 List *resultRelations, List *subpaths,
2984 List *subroots,
2985 List *withCheckOptionLists, List *returningLists,
2986 List *rowMarks, OnConflictExpr *onconflict,
2987 int epqParam)
2988 {
2989 ModifyTablePath *pathnode = makeNode(ModifyTablePath);
2990 double total_size;
2991 ListCell *lc;
2992
2993 Assert(list_length(resultRelations) == list_length(subpaths));
2994 Assert(list_length(resultRelations) == list_length(subroots));
2995 Assert(withCheckOptionLists == NIL ||
2996 list_length(resultRelations) == list_length(withCheckOptionLists));
2997 Assert(returningLists == NIL ||
2998 list_length(resultRelations) == list_length(returningLists));
2999
3000 pathnode->path.pathtype = T_ModifyTable;
3001 pathnode->path.parent = rel;
3002 /* pathtarget is not interesting, just make it minimally valid */
3003 pathnode->path.pathtarget = rel->reltarget;
3004 /* For now, assume we are above any joins, so no parameterization */
3005 pathnode->path.param_info = NULL;
3006 pathnode->path.parallel_aware = false;
3007 pathnode->path.parallel_safe = false;
3008 pathnode->path.parallel_workers = 0;
3009 pathnode->path.pathkeys = NIL;
3010
3011 /*
3012 * Compute cost & rowcount as sum of subpath costs & rowcounts.
3013 *
3014 * Currently, we don't charge anything extra for the actual table
3015 * modification work, nor for the WITH CHECK OPTIONS or RETURNING
3016 * expressions if any. It would only be window dressing, since
3017 * ModifyTable is always a top-level node and there is no way for the
3018 * costs to change any higher-level planning choices. But we might want
3019 * to make it look better sometime.
3020 */
3021 pathnode->path.startup_cost = 0;
3022 pathnode->path.total_cost = 0;
3023 pathnode->path.rows = 0;
3024 total_size = 0;
3025 foreach(lc, subpaths)
3026 {
3027 Path *subpath = (Path *) lfirst(lc);
3028
3029 if (lc == list_head(subpaths)) /* first node? */
3030 pathnode->path.startup_cost = subpath->startup_cost;
3031 pathnode->path.total_cost += subpath->total_cost;
3032 pathnode->path.rows += subpath->rows;
3033 total_size += subpath->pathtarget->width * subpath->rows;
3034 }
3035
3036 /*
3037 * Set width to the average width of the subpath outputs. XXX this is
3038 * totally wrong: we should report zero if no RETURNING, else an average
3039 * of the RETURNING tlist widths. But it's what happened historically,
3040 * and improving it is a task for another day.
3041 */
3042 if (pathnode->path.rows > 0)
3043 total_size /= pathnode->path.rows;
3044 pathnode->path.pathtarget->width = rint(total_size);
3045
3046 pathnode->operation = operation;
3047 pathnode->canSetTag = canSetTag;
3048 pathnode->nominalRelation = nominalRelation;
3049 pathnode->resultRelations = resultRelations;
3050 pathnode->subpaths = subpaths;
3051 pathnode->subroots = subroots;
3052 pathnode->withCheckOptionLists = withCheckOptionLists;
3053 pathnode->returningLists = returningLists;
3054 pathnode->rowMarks = rowMarks;
3055 pathnode->onconflict = onconflict;
3056 pathnode->epqParam = epqParam;
3057
3058 return pathnode;
3059 }
3060
3061 /*
3062 * create_limit_path
3063 * Creates a pathnode that represents performing LIMIT/OFFSET
3064 *
3065 * In addition to providing the actual OFFSET and LIMIT expressions,
3066 * the caller must provide estimates of their values for costing purposes.
3067 * The estimates are as computed by preprocess_limit(), ie, 0 represents
3068 * the clause not being present, and -1 means it's present but we could
3069 * not estimate its value.
3070 *
3071 * 'rel' is the parent relation associated with the result
3072 * 'subpath' is the path representing the source of data
3073 * 'limitOffset' is the actual OFFSET expression, or NULL
3074 * 'limitCount' is the actual LIMIT expression, or NULL
3075 * 'offset_est' is the estimated value of the OFFSET expression
3076 * 'count_est' is the estimated value of the LIMIT expression
3077 */
3078 LimitPath *
create_limit_path(PlannerInfo * root,RelOptInfo * rel,Path * subpath,Node * limitOffset,Node * limitCount,int64 offset_est,int64 count_est)3079 create_limit_path(PlannerInfo *root, RelOptInfo *rel,
3080 Path *subpath,
3081 Node *limitOffset, Node *limitCount,
3082 int64 offset_est, int64 count_est)
3083 {
3084 LimitPath *pathnode = makeNode(LimitPath);
3085
3086 pathnode->path.pathtype = T_Limit;
3087 pathnode->path.parent = rel;
3088 /* Limit doesn't project, so use source path's pathtarget */
3089 pathnode->path.pathtarget = subpath->pathtarget;
3090 /* For now, assume we are above any joins, so no parameterization */
3091 pathnode->path.param_info = NULL;
3092 pathnode->path.parallel_aware = false;
3093 pathnode->path.parallel_safe = rel->consider_parallel &&
3094 subpath->parallel_safe;
3095 pathnode->path.parallel_workers = subpath->parallel_workers;
3096 pathnode->path.rows = subpath->rows;
3097 pathnode->path.startup_cost = subpath->startup_cost;
3098 pathnode->path.total_cost = subpath->total_cost;
3099 pathnode->path.pathkeys = subpath->pathkeys;
3100 pathnode->subpath = subpath;
3101 pathnode->limitOffset = limitOffset;
3102 pathnode->limitCount = limitCount;
3103
3104 /*
3105 * Adjust the output rows count and costs according to the offset/limit.
3106 * This is only a cosmetic issue if we are at top level, but if we are
3107 * building a subquery then it's important to report correct info to the
3108 * outer planner.
3109 *
3110 * When the offset or count couldn't be estimated, use 10% of the
3111 * estimated number of rows emitted from the subpath.
3112 *
3113 * XXX we don't bother to add eval costs of the offset/limit expressions
3114 * themselves to the path costs. In theory we should, but in most cases
3115 * those expressions are trivial and it's just not worth the trouble.
3116 */
3117 if (offset_est != 0)
3118 {
3119 double offset_rows;
3120
3121 if (offset_est > 0)
3122 offset_rows = (double) offset_est;
3123 else
3124 offset_rows = clamp_row_est(subpath->rows * 0.10);
3125 if (offset_rows > pathnode->path.rows)
3126 offset_rows = pathnode->path.rows;
3127 if (subpath->rows > 0)
3128 pathnode->path.startup_cost +=
3129 (subpath->total_cost - subpath->startup_cost)
3130 * offset_rows / subpath->rows;
3131 pathnode->path.rows -= offset_rows;
3132 if (pathnode->path.rows < 1)
3133 pathnode->path.rows = 1;
3134 }
3135
3136 if (count_est != 0)
3137 {
3138 double count_rows;
3139
3140 if (count_est > 0)
3141 count_rows = (double) count_est;
3142 else
3143 count_rows = clamp_row_est(subpath->rows * 0.10);
3144 if (count_rows > pathnode->path.rows)
3145 count_rows = pathnode->path.rows;
3146 if (subpath->rows > 0)
3147 pathnode->path.total_cost = pathnode->path.startup_cost +
3148 (subpath->total_cost - subpath->startup_cost)
3149 * count_rows / subpath->rows;
3150 pathnode->path.rows = count_rows;
3151 if (pathnode->path.rows < 1)
3152 pathnode->path.rows = 1;
3153 }
3154
3155 return pathnode;
3156 }
3157
3158
3159 /*
3160 * reparameterize_path
3161 * Attempt to modify a Path to have greater parameterization
3162 *
3163 * We use this to attempt to bring all child paths of an appendrel to the
3164 * same parameterization level, ensuring that they all enforce the same set
3165 * of join quals (and thus that that parameterization can be attributed to
3166 * an append path built from such paths). Currently, only a few path types
3167 * are supported here, though more could be added at need. We return NULL
3168 * if we can't reparameterize the given path.
3169 *
3170 * Note: we intentionally do not pass created paths to add_path(); it would
3171 * possibly try to delete them on the grounds of being cost-inferior to the
3172 * paths they were made from, and we don't want that. Paths made here are
3173 * not necessarily of general-purpose usefulness, but they can be useful
3174 * as members of an append path.
3175 */
3176 Path *
reparameterize_path(PlannerInfo * root,Path * path,Relids required_outer,double loop_count)3177 reparameterize_path(PlannerInfo *root, Path *path,
3178 Relids required_outer,
3179 double loop_count)
3180 {
3181 RelOptInfo *rel = path->parent;
3182
3183 /* Can only increase, not decrease, path's parameterization */
3184 if (!bms_is_subset(PATH_REQ_OUTER(path), required_outer))
3185 return NULL;
3186 switch (path->pathtype)
3187 {
3188 case T_SeqScan:
3189 return create_seqscan_path(root, rel, required_outer, 0);
3190 case T_SampleScan:
3191 return (Path *) create_samplescan_path(root, rel, required_outer);
3192 case T_IndexScan:
3193 case T_IndexOnlyScan:
3194 {
3195 IndexPath *ipath = (IndexPath *) path;
3196 IndexPath *newpath = makeNode(IndexPath);
3197
3198 /*
3199 * We can't use create_index_path directly, and would not want
3200 * to because it would re-compute the indexqual conditions
3201 * which is wasted effort. Instead we hack things a bit:
3202 * flat-copy the path node, revise its param_info, and redo
3203 * the cost estimate.
3204 */
3205 memcpy(newpath, ipath, sizeof(IndexPath));
3206 newpath->path.param_info =
3207 get_baserel_parampathinfo(root, rel, required_outer);
3208 cost_index(newpath, root, loop_count);
3209 return (Path *) newpath;
3210 }
3211 case T_BitmapHeapScan:
3212 {
3213 BitmapHeapPath *bpath = (BitmapHeapPath *) path;
3214
3215 return (Path *) create_bitmap_heap_path(root,
3216 rel,
3217 bpath->bitmapqual,
3218 required_outer,
3219 loop_count);
3220 }
3221 case T_SubqueryScan:
3222 {
3223 SubqueryScanPath *spath = (SubqueryScanPath *) path;
3224
3225 return (Path *) create_subqueryscan_path(root,
3226 rel,
3227 spath->subpath,
3228 spath->path.pathkeys,
3229 required_outer);
3230 }
3231 case T_Append:
3232 {
3233 AppendPath *apath = (AppendPath *) path;
3234 List *childpaths = NIL;
3235 ListCell *lc;
3236
3237 /* Reparameterize the children */
3238 foreach(lc, apath->subpaths)
3239 {
3240 Path *spath = (Path *) lfirst(lc);
3241
3242 spath = reparameterize_path(root, spath,
3243 required_outer,
3244 loop_count);
3245 if (spath == NULL)
3246 return NULL;
3247 childpaths = lappend(childpaths, spath);
3248 }
3249 return (Path *)
3250 create_append_path(rel, childpaths,
3251 required_outer,
3252 apath->path.parallel_workers);
3253 }
3254 default:
3255 break;
3256 }
3257 return NULL;
3258 }
3259