1 /*-------------------------------------------------------------------------
2 *
3 * pathnode.c
4 * Routines to manipulate pathlists and create path nodes
5 *
6 * Portions Copyright (c) 1996-2019, 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 "foreign/fdwapi.h"
21 #include "nodes/extensible.h"
22 #include "nodes/nodeFuncs.h"
23 #include "optimizer/appendinfo.h"
24 #include "optimizer/clauses.h"
25 #include "optimizer/cost.h"
26 #include "optimizer/optimizer.h"
27 #include "optimizer/pathnode.h"
28 #include "optimizer/paths.h"
29 #include "optimizer/planmain.h"
30 #include "optimizer/prep.h"
31 #include "optimizer/restrictinfo.h"
32 #include "optimizer/tlist.h"
33 #include "parser/parsetree.h"
34 #include "utils/lsyscache.h"
35 #include "utils/memutils.h"
36 #include "utils/selfuncs.h"
37
38
39 typedef enum
40 {
41 COSTS_EQUAL, /* path costs are fuzzily equal */
42 COSTS_BETTER1, /* first path is cheaper than second */
43 COSTS_BETTER2, /* second path is cheaper than first */
44 COSTS_DIFFERENT /* neither path dominates the other on cost */
45 } PathCostComparison;
46
47 /*
48 * STD_FUZZ_FACTOR is the normal fuzz factor for compare_path_costs_fuzzily.
49 * XXX is it worth making this user-controllable? It provides a tradeoff
50 * between planner runtime and the accuracy of path cost comparisons.
51 */
52 #define STD_FUZZ_FACTOR 1.01
53
54 static List *translate_sub_tlist(List *tlist, int relid);
55 static int append_total_cost_compare(const void *a, const void *b);
56 static int append_startup_cost_compare(const void *a, const void *b);
57 static List *reparameterize_pathlist_by_child(PlannerInfo *root,
58 List *pathlist,
59 RelOptInfo *child_rel);
60
61
62 /*****************************************************************************
63 * MISC. PATH UTILITIES
64 *****************************************************************************/
65
66 /*
67 * compare_path_costs
68 * Return -1, 0, or +1 according as path1 is cheaper, the same cost,
69 * or more expensive than path2 for the specified criterion.
70 */
71 int
compare_path_costs(Path * path1,Path * path2,CostSelector criterion)72 compare_path_costs(Path *path1, Path *path2, CostSelector criterion)
73 {
74 if (criterion == STARTUP_COST)
75 {
76 if (path1->startup_cost < path2->startup_cost)
77 return -1;
78 if (path1->startup_cost > path2->startup_cost)
79 return +1;
80
81 /*
82 * If paths have the same startup cost (not at all unlikely), order
83 * them by total cost.
84 */
85 if (path1->total_cost < path2->total_cost)
86 return -1;
87 if (path1->total_cost > path2->total_cost)
88 return +1;
89 }
90 else
91 {
92 if (path1->total_cost < path2->total_cost)
93 return -1;
94 if (path1->total_cost > path2->total_cost)
95 return +1;
96
97 /*
98 * If paths have the same total cost, order them by startup cost.
99 */
100 if (path1->startup_cost < path2->startup_cost)
101 return -1;
102 if (path1->startup_cost > path2->startup_cost)
103 return +1;
104 }
105 return 0;
106 }
107
108 /*
109 * compare_path_fractional_costs
110 * Return -1, 0, or +1 according as path1 is cheaper, the same cost,
111 * or more expensive than path2 for fetching the specified fraction
112 * of the total tuples.
113 *
114 * If fraction is <= 0 or > 1, we interpret it as 1, ie, we select the
115 * path with the cheaper total_cost.
116 */
117 int
compare_fractional_path_costs(Path * path1,Path * path2,double fraction)118 compare_fractional_path_costs(Path *path1, Path *path2,
119 double fraction)
120 {
121 Cost cost1,
122 cost2;
123
124 if (fraction <= 0.0 || fraction >= 1.0)
125 return compare_path_costs(path1, path2, TOTAL_COST);
126 cost1 = path1->startup_cost +
127 fraction * (path1->total_cost - path1->startup_cost);
128 cost2 = path2->startup_cost +
129 fraction * (path2->total_cost - path2->startup_cost);
130 if (cost1 < cost2)
131 return -1;
132 if (cost1 > cost2)
133 return +1;
134 return 0;
135 }
136
137 /*
138 * compare_path_costs_fuzzily
139 * Compare the costs of two paths to see if either can be said to
140 * dominate the other.
141 *
142 * We use fuzzy comparisons so that add_path() can avoid keeping both of
143 * a pair of paths that really have insignificantly different cost.
144 *
145 * The fuzz_factor argument must be 1.0 plus delta, where delta is the
146 * fraction of the smaller cost that is considered to be a significant
147 * difference. For example, fuzz_factor = 1.01 makes the fuzziness limit
148 * be 1% of the smaller cost.
149 *
150 * The two paths are said to have "equal" costs if both startup and total
151 * costs are fuzzily the same. Path1 is said to be better than path2 if
152 * it has fuzzily better startup cost and fuzzily no worse total cost,
153 * or if it has fuzzily better total cost and fuzzily no worse startup cost.
154 * Path2 is better than path1 if the reverse holds. Finally, if one path
155 * is fuzzily better than the other on startup cost and fuzzily worse on
156 * total cost, we just say that their costs are "different", since neither
157 * dominates the other across the whole performance spectrum.
158 *
159 * This function also enforces a policy rule that paths for which the relevant
160 * one of parent->consider_startup and parent->consider_param_startup is false
161 * cannot survive comparisons solely on the grounds of good startup cost, so
162 * we never return COSTS_DIFFERENT when that is true for the total-cost loser.
163 * (But if total costs are fuzzily equal, we compare startup costs anyway,
164 * in hopes of eliminating one path or the other.)
165 */
166 static PathCostComparison
compare_path_costs_fuzzily(Path * path1,Path * path2,double fuzz_factor)167 compare_path_costs_fuzzily(Path *path1, Path *path2, double fuzz_factor)
168 {
169 #define CONSIDER_PATH_STARTUP_COST(p) \
170 ((p)->param_info == NULL ? (p)->parent->consider_startup : (p)->parent->consider_param_startup)
171
172 /*
173 * Check total cost first since it's more likely to be different; many
174 * paths have zero startup cost.
175 */
176 if (path1->total_cost > path2->total_cost * fuzz_factor)
177 {
178 /* path1 fuzzily worse on total cost */
179 if (CONSIDER_PATH_STARTUP_COST(path1) &&
180 path2->startup_cost > path1->startup_cost * fuzz_factor)
181 {
182 /* ... but path2 fuzzily worse on startup, so DIFFERENT */
183 return COSTS_DIFFERENT;
184 }
185 /* else path2 dominates */
186 return COSTS_BETTER2;
187 }
188 if (path2->total_cost > path1->total_cost * fuzz_factor)
189 {
190 /* path2 fuzzily worse on total cost */
191 if (CONSIDER_PATH_STARTUP_COST(path2) &&
192 path1->startup_cost > path2->startup_cost * fuzz_factor)
193 {
194 /* ... but path1 fuzzily worse on startup, so DIFFERENT */
195 return COSTS_DIFFERENT;
196 }
197 /* else path1 dominates */
198 return COSTS_BETTER1;
199 }
200 /* fuzzily the same on total cost ... */
201 if (path1->startup_cost > path2->startup_cost * fuzz_factor)
202 {
203 /* ... but path1 fuzzily worse on startup, so path2 wins */
204 return COSTS_BETTER2;
205 }
206 if (path2->startup_cost > path1->startup_cost * fuzz_factor)
207 {
208 /* ... but path2 fuzzily worse on startup, so path1 wins */
209 return COSTS_BETTER1;
210 }
211 /* fuzzily the same on both costs */
212 return COSTS_EQUAL;
213
214 #undef CONSIDER_PATH_STARTUP_COST
215 }
216
217 /*
218 * set_cheapest
219 * Find the minimum-cost paths from among a relation's paths,
220 * and save them in the rel's cheapest-path fields.
221 *
222 * cheapest_total_path is normally the cheapest-total-cost unparameterized
223 * path; but if there are no unparameterized paths, we assign it to be the
224 * best (cheapest least-parameterized) parameterized path. However, only
225 * unparameterized paths are considered candidates for cheapest_startup_path,
226 * so that will be NULL if there are no unparameterized paths.
227 *
228 * The cheapest_parameterized_paths list collects all parameterized paths
229 * that have survived the add_path() tournament for this relation. (Since
230 * add_path ignores pathkeys for a parameterized path, these will be paths
231 * that have best cost or best row count for their parameterization. We
232 * may also have both a parallel-safe and a non-parallel-safe path in some
233 * cases for the same parameterization in some cases, but this should be
234 * relatively rare since, most typically, all paths for the same relation
235 * will be parallel-safe or none of them will.)
236 *
237 * cheapest_parameterized_paths always includes the cheapest-total
238 * unparameterized path, too, if there is one; the users of that list find
239 * it more convenient if that's included.
240 *
241 * This is normally called only after we've finished constructing the path
242 * list for the rel node.
243 */
244 void
set_cheapest(RelOptInfo * parent_rel)245 set_cheapest(RelOptInfo *parent_rel)
246 {
247 Path *cheapest_startup_path;
248 Path *cheapest_total_path;
249 Path *best_param_path;
250 List *parameterized_paths;
251 ListCell *p;
252
253 Assert(IsA(parent_rel, RelOptInfo));
254
255 if (parent_rel->pathlist == NIL)
256 elog(ERROR, "could not devise a query plan for the given query");
257
258 cheapest_startup_path = cheapest_total_path = best_param_path = NULL;
259 parameterized_paths = NIL;
260
261 foreach(p, parent_rel->pathlist)
262 {
263 Path *path = (Path *) lfirst(p);
264 int cmp;
265
266 if (path->param_info)
267 {
268 /* Parameterized path, so add it to parameterized_paths */
269 parameterized_paths = lappend(parameterized_paths, path);
270
271 /*
272 * If we have an unparameterized cheapest-total, we no longer care
273 * about finding the best parameterized path, so move on.
274 */
275 if (cheapest_total_path)
276 continue;
277
278 /*
279 * Otherwise, track the best parameterized path, which is the one
280 * with least total cost among those of the minimum
281 * parameterization.
282 */
283 if (best_param_path == NULL)
284 best_param_path = path;
285 else
286 {
287 switch (bms_subset_compare(PATH_REQ_OUTER(path),
288 PATH_REQ_OUTER(best_param_path)))
289 {
290 case BMS_EQUAL:
291 /* keep the cheaper one */
292 if (compare_path_costs(path, best_param_path,
293 TOTAL_COST) < 0)
294 best_param_path = path;
295 break;
296 case BMS_SUBSET1:
297 /* new path is less-parameterized */
298 best_param_path = path;
299 break;
300 case BMS_SUBSET2:
301 /* old path is less-parameterized, keep it */
302 break;
303 case BMS_DIFFERENT:
304
305 /*
306 * This means that neither path has the least possible
307 * parameterization for the rel. We'll sit on the old
308 * path until something better comes along.
309 */
310 break;
311 }
312 }
313 }
314 else
315 {
316 /* Unparameterized path, so consider it for cheapest slots */
317 if (cheapest_total_path == NULL)
318 {
319 cheapest_startup_path = cheapest_total_path = path;
320 continue;
321 }
322
323 /*
324 * If we find two paths of identical costs, try to keep the
325 * better-sorted one. The paths might have unrelated sort
326 * orderings, in which case we can only guess which might be
327 * better to keep, but if one is superior then we definitely
328 * should keep that one.
329 */
330 cmp = compare_path_costs(cheapest_startup_path, path, STARTUP_COST);
331 if (cmp > 0 ||
332 (cmp == 0 &&
333 compare_pathkeys(cheapest_startup_path->pathkeys,
334 path->pathkeys) == PATHKEYS_BETTER2))
335 cheapest_startup_path = path;
336
337 cmp = compare_path_costs(cheapest_total_path, path, TOTAL_COST);
338 if (cmp > 0 ||
339 (cmp == 0 &&
340 compare_pathkeys(cheapest_total_path->pathkeys,
341 path->pathkeys) == PATHKEYS_BETTER2))
342 cheapest_total_path = path;
343 }
344 }
345
346 /* Add cheapest unparameterized path, if any, to parameterized_paths */
347 if (cheapest_total_path)
348 parameterized_paths = lcons(cheapest_total_path, parameterized_paths);
349
350 /*
351 * If there is no unparameterized path, use the best parameterized path as
352 * cheapest_total_path (but not as cheapest_startup_path).
353 */
354 if (cheapest_total_path == NULL)
355 cheapest_total_path = best_param_path;
356 Assert(cheapest_total_path != NULL);
357
358 parent_rel->cheapest_startup_path = cheapest_startup_path;
359 parent_rel->cheapest_total_path = cheapest_total_path;
360 parent_rel->cheapest_unique_path = NULL; /* computed only if needed */
361 parent_rel->cheapest_parameterized_paths = parameterized_paths;
362 }
363
364 /*
365 * add_path
366 * Consider a potential implementation path for the specified parent rel,
367 * and add it to the rel's pathlist if it is worthy of consideration.
368 * A path is worthy if it has a better sort order (better pathkeys) or
369 * cheaper cost (on either dimension), or generates fewer rows, than any
370 * existing path that has the same or superset parameterization rels.
371 * We also consider parallel-safe paths more worthy than others.
372 *
373 * We also remove from the rel's pathlist any old paths that are dominated
374 * by new_path --- that is, new_path is cheaper, at least as well ordered,
375 * generates no more rows, requires no outer rels not required by the old
376 * path, and is no less parallel-safe.
377 *
378 * In most cases, a path with a superset parameterization will generate
379 * fewer rows (since it has more join clauses to apply), so that those two
380 * figures of merit move in opposite directions; this means that a path of
381 * one parameterization can seldom dominate a path of another. But such
382 * cases do arise, so we make the full set of checks anyway.
383 *
384 * There are two policy decisions embedded in this function, along with
385 * its sibling add_path_precheck. First, we treat all parameterized paths
386 * as having NIL pathkeys, so that they cannot win comparisons on the
387 * basis of sort order. This is to reduce the number of parameterized
388 * paths that are kept; see discussion in src/backend/optimizer/README.
389 *
390 * Second, we only consider cheap startup cost to be interesting if
391 * parent_rel->consider_startup is true for an unparameterized path, or
392 * parent_rel->consider_param_startup is true for a parameterized one.
393 * Again, this allows discarding useless paths sooner.
394 *
395 * The pathlist is kept sorted by total_cost, with cheaper paths
396 * at the front. Within this routine, that's simply a speed hack:
397 * doing it that way makes it more likely that we will reject an inferior
398 * path after a few comparisons, rather than many comparisons.
399 * However, add_path_precheck relies on this ordering to exit early
400 * when possible.
401 *
402 * NOTE: discarded Path objects are immediately pfree'd to reduce planner
403 * memory consumption. We dare not try to free the substructure of a Path,
404 * since much of it may be shared with other Paths or the query tree itself;
405 * but just recycling discarded Path nodes is a very useful savings in
406 * a large join tree. We can recycle the List nodes of pathlist, too.
407 *
408 * As noted in optimizer/README, deleting a previously-accepted Path is
409 * safe because we know that Paths of this rel cannot yet be referenced
410 * from any other rel, such as a higher-level join. However, in some cases
411 * it is possible that a Path is referenced by another Path for its own
412 * rel; we must not delete such a Path, even if it is dominated by the new
413 * Path. Currently this occurs only for IndexPath objects, which may be
414 * referenced as children of BitmapHeapPaths as well as being paths in
415 * their own right. Hence, we don't pfree IndexPaths when rejecting them.
416 *
417 * 'parent_rel' is the relation entry to which the path corresponds.
418 * 'new_path' is a potential path for parent_rel.
419 *
420 * Returns nothing, but modifies parent_rel->pathlist.
421 */
422 void
add_path(RelOptInfo * parent_rel,Path * new_path)423 add_path(RelOptInfo *parent_rel, Path *new_path)
424 {
425 bool accept_new = true; /* unless we find a superior old path */
426 ListCell *insert_after = NULL; /* where to insert new item */
427 List *new_path_pathkeys;
428 ListCell *p1;
429 ListCell *p1_prev;
430 ListCell *p1_next;
431
432 /*
433 * This is a convenient place to check for query cancel --- no part of the
434 * planner goes very long without calling add_path().
435 */
436 CHECK_FOR_INTERRUPTS();
437
438 /* Pretend parameterized paths have no pathkeys, per comment above */
439 new_path_pathkeys = new_path->param_info ? NIL : new_path->pathkeys;
440
441 /*
442 * Loop to check proposed new path against old paths. Note it is possible
443 * for more than one old path to be tossed out because new_path dominates
444 * it.
445 *
446 * We can't use foreach here because the loop body may delete the current
447 * list cell.
448 */
449 p1_prev = NULL;
450 for (p1 = list_head(parent_rel->pathlist); p1 != NULL; p1 = p1_next)
451 {
452 Path *old_path = (Path *) lfirst(p1);
453 bool remove_old = false; /* unless new proves superior */
454 PathCostComparison costcmp;
455 PathKeysComparison keyscmp;
456 BMS_Comparison outercmp;
457
458 p1_next = lnext(p1);
459
460 /*
461 * Do a fuzzy cost comparison with standard fuzziness limit.
462 */
463 costcmp = compare_path_costs_fuzzily(new_path, old_path,
464 STD_FUZZ_FACTOR);
465
466 /*
467 * If the two paths compare differently for startup and total cost,
468 * then we want to keep both, and we can skip comparing pathkeys and
469 * required_outer rels. If they compare the same, proceed with the
470 * other comparisons. Row count is checked last. (We make the tests
471 * in this order because the cost comparison is most likely to turn
472 * out "different", and the pathkeys comparison next most likely. As
473 * explained above, row count very seldom makes a difference, so even
474 * though it's cheap to compare there's not much point in checking it
475 * earlier.)
476 */
477 if (costcmp != COSTS_DIFFERENT)
478 {
479 /* Similarly check to see if either dominates on pathkeys */
480 List *old_path_pathkeys;
481
482 old_path_pathkeys = old_path->param_info ? NIL : old_path->pathkeys;
483 keyscmp = compare_pathkeys(new_path_pathkeys,
484 old_path_pathkeys);
485 if (keyscmp != PATHKEYS_DIFFERENT)
486 {
487 switch (costcmp)
488 {
489 case COSTS_EQUAL:
490 outercmp = bms_subset_compare(PATH_REQ_OUTER(new_path),
491 PATH_REQ_OUTER(old_path));
492 if (keyscmp == PATHKEYS_BETTER1)
493 {
494 if ((outercmp == BMS_EQUAL ||
495 outercmp == BMS_SUBSET1) &&
496 new_path->rows <= old_path->rows &&
497 new_path->parallel_safe >= old_path->parallel_safe)
498 remove_old = true; /* new dominates old */
499 }
500 else if (keyscmp == PATHKEYS_BETTER2)
501 {
502 if ((outercmp == BMS_EQUAL ||
503 outercmp == BMS_SUBSET2) &&
504 new_path->rows >= old_path->rows &&
505 new_path->parallel_safe <= old_path->parallel_safe)
506 accept_new = false; /* old dominates new */
507 }
508 else /* keyscmp == PATHKEYS_EQUAL */
509 {
510 if (outercmp == BMS_EQUAL)
511 {
512 /*
513 * Same pathkeys and outer rels, and fuzzily
514 * the same cost, so keep just one; to decide
515 * which, first check parallel-safety, then
516 * rows, then do a fuzzy cost comparison with
517 * very small fuzz limit. (We used to do an
518 * exact cost comparison, but that results in
519 * annoying platform-specific plan variations
520 * due to roundoff in the cost estimates.) If
521 * things are still tied, arbitrarily keep
522 * only the old path. Notice that we will
523 * keep only the old path even if the
524 * less-fuzzy comparison decides the startup
525 * and total costs compare differently.
526 */
527 if (new_path->parallel_safe >
528 old_path->parallel_safe)
529 remove_old = true; /* new dominates old */
530 else if (new_path->parallel_safe <
531 old_path->parallel_safe)
532 accept_new = false; /* old dominates new */
533 else if (new_path->rows < old_path->rows)
534 remove_old = true; /* new dominates old */
535 else if (new_path->rows > old_path->rows)
536 accept_new = false; /* old dominates new */
537 else if (compare_path_costs_fuzzily(new_path,
538 old_path,
539 1.0000000001) == COSTS_BETTER1)
540 remove_old = true; /* new dominates old */
541 else
542 accept_new = false; /* old equals or
543 * dominates new */
544 }
545 else if (outercmp == BMS_SUBSET1 &&
546 new_path->rows <= old_path->rows &&
547 new_path->parallel_safe >= old_path->parallel_safe)
548 remove_old = true; /* new dominates old */
549 else if (outercmp == BMS_SUBSET2 &&
550 new_path->rows >= old_path->rows &&
551 new_path->parallel_safe <= old_path->parallel_safe)
552 accept_new = false; /* old dominates new */
553 /* else different parameterizations, keep both */
554 }
555 break;
556 case COSTS_BETTER1:
557 if (keyscmp != PATHKEYS_BETTER2)
558 {
559 outercmp = bms_subset_compare(PATH_REQ_OUTER(new_path),
560 PATH_REQ_OUTER(old_path));
561 if ((outercmp == BMS_EQUAL ||
562 outercmp == BMS_SUBSET1) &&
563 new_path->rows <= old_path->rows &&
564 new_path->parallel_safe >= old_path->parallel_safe)
565 remove_old = true; /* new dominates old */
566 }
567 break;
568 case COSTS_BETTER2:
569 if (keyscmp != PATHKEYS_BETTER1)
570 {
571 outercmp = bms_subset_compare(PATH_REQ_OUTER(new_path),
572 PATH_REQ_OUTER(old_path));
573 if ((outercmp == BMS_EQUAL ||
574 outercmp == BMS_SUBSET2) &&
575 new_path->rows >= old_path->rows &&
576 new_path->parallel_safe <= old_path->parallel_safe)
577 accept_new = false; /* old dominates new */
578 }
579 break;
580 case COSTS_DIFFERENT:
581
582 /*
583 * can't get here, but keep this case to keep compiler
584 * quiet
585 */
586 break;
587 }
588 }
589 }
590
591 /*
592 * Remove current element from pathlist if dominated by new.
593 */
594 if (remove_old)
595 {
596 parent_rel->pathlist = list_delete_cell(parent_rel->pathlist,
597 p1, p1_prev);
598
599 /*
600 * Delete the data pointed-to by the deleted cell, if possible
601 */
602 if (!IsA(old_path, IndexPath))
603 pfree(old_path);
604 /* p1_prev does not advance */
605 }
606 else
607 {
608 /* new belongs after this old path if it has cost >= old's */
609 if (new_path->total_cost >= old_path->total_cost)
610 insert_after = p1;
611 /* p1_prev advances */
612 p1_prev = p1;
613 }
614
615 /*
616 * If we found an old path that dominates new_path, we can quit
617 * scanning the pathlist; we will not add new_path, and we assume
618 * new_path cannot dominate any other elements of the pathlist.
619 */
620 if (!accept_new)
621 break;
622 }
623
624 if (accept_new)
625 {
626 /* Accept the new path: insert it at proper place in pathlist */
627 if (insert_after)
628 lappend_cell(parent_rel->pathlist, insert_after, new_path);
629 else
630 parent_rel->pathlist = lcons(new_path, parent_rel->pathlist);
631 }
632 else
633 {
634 /* Reject and recycle the new path */
635 if (!IsA(new_path, IndexPath))
636 pfree(new_path);
637 }
638 }
639
640 /*
641 * add_path_precheck
642 * Check whether a proposed new path could possibly get accepted.
643 * We assume we know the path's pathkeys and parameterization accurately,
644 * and have lower bounds for its costs.
645 *
646 * Note that we do not know the path's rowcount, since getting an estimate for
647 * that is too expensive to do before prechecking. We assume here that paths
648 * of a superset parameterization will generate fewer rows; if that holds,
649 * then paths with different parameterizations cannot dominate each other
650 * and so we can simply ignore existing paths of another parameterization.
651 * (In the infrequent cases where that rule of thumb fails, add_path will
652 * get rid of the inferior path.)
653 *
654 * At the time this is called, we haven't actually built a Path structure,
655 * so the required information has to be passed piecemeal.
656 */
657 bool
add_path_precheck(RelOptInfo * parent_rel,Cost startup_cost,Cost total_cost,List * pathkeys,Relids required_outer)658 add_path_precheck(RelOptInfo *parent_rel,
659 Cost startup_cost, Cost total_cost,
660 List *pathkeys, Relids required_outer)
661 {
662 List *new_path_pathkeys;
663 bool consider_startup;
664 ListCell *p1;
665
666 /* Pretend parameterized paths have no pathkeys, per add_path policy */
667 new_path_pathkeys = required_outer ? NIL : pathkeys;
668
669 /* Decide whether new path's startup cost is interesting */
670 consider_startup = required_outer ? parent_rel->consider_param_startup : parent_rel->consider_startup;
671
672 foreach(p1, parent_rel->pathlist)
673 {
674 Path *old_path = (Path *) lfirst(p1);
675 PathKeysComparison keyscmp;
676
677 /*
678 * We are looking for an old_path with the same parameterization (and
679 * by assumption the same rowcount) that dominates the new path on
680 * pathkeys as well as both cost metrics. If we find one, we can
681 * reject the new path.
682 *
683 * Cost comparisons here should match compare_path_costs_fuzzily.
684 */
685 if (total_cost > old_path->total_cost * STD_FUZZ_FACTOR)
686 {
687 /* new path can win on startup cost only if consider_startup */
688 if (startup_cost > old_path->startup_cost * STD_FUZZ_FACTOR ||
689 !consider_startup)
690 {
691 /* new path loses on cost, so check pathkeys... */
692 List *old_path_pathkeys;
693
694 old_path_pathkeys = old_path->param_info ? NIL : old_path->pathkeys;
695 keyscmp = compare_pathkeys(new_path_pathkeys,
696 old_path_pathkeys);
697 if (keyscmp == PATHKEYS_EQUAL ||
698 keyscmp == PATHKEYS_BETTER2)
699 {
700 /* new path does not win on pathkeys... */
701 if (bms_equal(required_outer, PATH_REQ_OUTER(old_path)))
702 {
703 /* Found an old path that dominates the new one */
704 return false;
705 }
706 }
707 }
708 }
709 else
710 {
711 /*
712 * Since the pathlist is sorted by total_cost, we can stop looking
713 * once we reach a path with a total_cost larger than the new
714 * path's.
715 */
716 break;
717 }
718 }
719
720 return true;
721 }
722
723 /*
724 * add_partial_path
725 * Like add_path, our goal here is to consider whether a path is worthy
726 * of being kept around, but the considerations here are a bit different.
727 * A partial path is one which can be executed in any number of workers in
728 * parallel such that each worker will generate a subset of the path's
729 * overall result.
730 *
731 * As in add_path, the partial_pathlist is kept sorted with the cheapest
732 * total path in front. This is depended on by multiple places, which
733 * just take the front entry as the cheapest path without searching.
734 *
735 * We don't generate parameterized partial paths for several reasons. Most
736 * importantly, they're not safe to execute, because there's nothing to
737 * make sure that a parallel scan within the parameterized portion of the
738 * plan is running with the same value in every worker at the same time.
739 * Fortunately, it seems unlikely to be worthwhile anyway, because having
740 * each worker scan the entire outer relation and a subset of the inner
741 * relation will generally be a terrible plan. The inner (parameterized)
742 * side of the plan will be small anyway. There could be rare cases where
743 * this wins big - e.g. if join order constraints put a 1-row relation on
744 * the outer side of the topmost join with a parameterized plan on the inner
745 * side - but we'll have to be content not to handle such cases until
746 * somebody builds an executor infrastructure that can cope with them.
747 *
748 * Because we don't consider parameterized paths here, we also don't
749 * need to consider the row counts as a measure of quality: every path will
750 * produce the same number of rows. Neither do we need to consider startup
751 * costs: parallelism is only used for plans that will be run to completion.
752 * Therefore, this routine is much simpler than add_path: it needs to
753 * consider only pathkeys and total cost.
754 *
755 * As with add_path, we pfree paths that are found to be dominated by
756 * another partial path; this requires that there be no other references to
757 * such paths yet. Hence, GatherPaths must not be created for a rel until
758 * we're done creating all partial paths for it. Unlike add_path, we don't
759 * take an exception for IndexPaths as partial index paths won't be
760 * referenced by partial BitmapHeapPaths.
761 */
762 void
add_partial_path(RelOptInfo * parent_rel,Path * new_path)763 add_partial_path(RelOptInfo *parent_rel, Path *new_path)
764 {
765 bool accept_new = true; /* unless we find a superior old path */
766 ListCell *insert_after = NULL; /* where to insert new item */
767 ListCell *p1;
768 ListCell *p1_prev;
769 ListCell *p1_next;
770
771 /* Check for query cancel. */
772 CHECK_FOR_INTERRUPTS();
773
774 /* Path to be added must be parallel safe. */
775 Assert(new_path->parallel_safe);
776
777 /* Relation should be OK for parallelism, too. */
778 Assert(parent_rel->consider_parallel);
779
780 /*
781 * As in add_path, throw out any paths which are dominated by the new
782 * path, but throw out the new path if some existing path dominates it.
783 */
784 p1_prev = NULL;
785 for (p1 = list_head(parent_rel->partial_pathlist); p1 != NULL;
786 p1 = p1_next)
787 {
788 Path *old_path = (Path *) lfirst(p1);
789 bool remove_old = false; /* unless new proves superior */
790 PathKeysComparison keyscmp;
791
792 p1_next = lnext(p1);
793
794 /* Compare pathkeys. */
795 keyscmp = compare_pathkeys(new_path->pathkeys, old_path->pathkeys);
796
797 /* Unless pathkeys are incompable, keep just one of the two paths. */
798 if (keyscmp != PATHKEYS_DIFFERENT)
799 {
800 if (new_path->total_cost > old_path->total_cost * STD_FUZZ_FACTOR)
801 {
802 /* New path costs more; keep it only if pathkeys are better. */
803 if (keyscmp != PATHKEYS_BETTER1)
804 accept_new = false;
805 }
806 else if (old_path->total_cost > new_path->total_cost
807 * STD_FUZZ_FACTOR)
808 {
809 /* Old path costs more; keep it only if pathkeys are better. */
810 if (keyscmp != PATHKEYS_BETTER2)
811 remove_old = true;
812 }
813 else if (keyscmp == PATHKEYS_BETTER1)
814 {
815 /* Costs are about the same, new path has better pathkeys. */
816 remove_old = true;
817 }
818 else if (keyscmp == PATHKEYS_BETTER2)
819 {
820 /* Costs are about the same, old path has better pathkeys. */
821 accept_new = false;
822 }
823 else if (old_path->total_cost > new_path->total_cost * 1.0000000001)
824 {
825 /* Pathkeys are the same, and the old path costs more. */
826 remove_old = true;
827 }
828 else
829 {
830 /*
831 * Pathkeys are the same, and new path isn't materially
832 * cheaper.
833 */
834 accept_new = false;
835 }
836 }
837
838 /*
839 * Remove current element from partial_pathlist if dominated by new.
840 */
841 if (remove_old)
842 {
843 parent_rel->partial_pathlist =
844 list_delete_cell(parent_rel->partial_pathlist, p1, p1_prev);
845 pfree(old_path);
846 /* p1_prev does not advance */
847 }
848 else
849 {
850 /* new belongs after this old path if it has cost >= old's */
851 if (new_path->total_cost >= old_path->total_cost)
852 insert_after = p1;
853 /* p1_prev advances */
854 p1_prev = p1;
855 }
856
857 /*
858 * If we found an old path that dominates new_path, we can quit
859 * scanning the partial_pathlist; we will not add new_path, and we
860 * assume new_path cannot dominate any later path.
861 */
862 if (!accept_new)
863 break;
864 }
865
866 if (accept_new)
867 {
868 /* Accept the new path: insert it at proper place */
869 if (insert_after)
870 lappend_cell(parent_rel->partial_pathlist, insert_after, new_path);
871 else
872 parent_rel->partial_pathlist =
873 lcons(new_path, parent_rel->partial_pathlist);
874 }
875 else
876 {
877 /* Reject and recycle the new path */
878 pfree(new_path);
879 }
880 }
881
882 /*
883 * add_partial_path_precheck
884 * Check whether a proposed new partial path could possibly get accepted.
885 *
886 * Unlike add_path_precheck, we can ignore startup cost and parameterization,
887 * since they don't matter for partial paths (see add_partial_path). But
888 * we do want to make sure we don't add a partial path if there's already
889 * a complete path that dominates it, since in that case the proposed path
890 * is surely a loser.
891 */
892 bool
add_partial_path_precheck(RelOptInfo * parent_rel,Cost total_cost,List * pathkeys)893 add_partial_path_precheck(RelOptInfo *parent_rel, Cost total_cost,
894 List *pathkeys)
895 {
896 ListCell *p1;
897
898 /*
899 * Our goal here is twofold. First, we want to find out whether this path
900 * is clearly inferior to some existing partial path. If so, we want to
901 * reject it immediately. Second, we want to find out whether this path
902 * is clearly superior to some existing partial path -- at least, modulo
903 * final cost computations. If so, we definitely want to consider it.
904 *
905 * Unlike add_path(), we always compare pathkeys here. This is because we
906 * expect partial_pathlist to be very short, and getting a definitive
907 * answer at this stage avoids the need to call add_path_precheck.
908 */
909 foreach(p1, parent_rel->partial_pathlist)
910 {
911 Path *old_path = (Path *) lfirst(p1);
912 PathKeysComparison keyscmp;
913
914 keyscmp = compare_pathkeys(pathkeys, old_path->pathkeys);
915 if (keyscmp != PATHKEYS_DIFFERENT)
916 {
917 if (total_cost > old_path->total_cost * STD_FUZZ_FACTOR &&
918 keyscmp != PATHKEYS_BETTER1)
919 return false;
920 if (old_path->total_cost > total_cost * STD_FUZZ_FACTOR &&
921 keyscmp != PATHKEYS_BETTER2)
922 return true;
923 }
924 }
925
926 /*
927 * This path is neither clearly inferior to an existing partial path nor
928 * clearly good enough that it might replace one. Compare it to
929 * non-parallel plans. If it loses even before accounting for the cost of
930 * the Gather node, we should definitely reject it.
931 *
932 * Note that we pass the total_cost to add_path_precheck twice. This is
933 * because it's never advantageous to consider the startup cost of a
934 * partial path; the resulting plans, if run in parallel, will be run to
935 * completion.
936 */
937 if (!add_path_precheck(parent_rel, total_cost, total_cost, pathkeys,
938 NULL))
939 return false;
940
941 return true;
942 }
943
944
945 /*****************************************************************************
946 * PATH NODE CREATION ROUTINES
947 *****************************************************************************/
948
949 /*
950 * create_seqscan_path
951 * Creates a path corresponding to a sequential scan, returning the
952 * pathnode.
953 */
954 Path *
create_seqscan_path(PlannerInfo * root,RelOptInfo * rel,Relids required_outer,int parallel_workers)955 create_seqscan_path(PlannerInfo *root, RelOptInfo *rel,
956 Relids required_outer, int parallel_workers)
957 {
958 Path *pathnode = makeNode(Path);
959
960 pathnode->pathtype = T_SeqScan;
961 pathnode->parent = rel;
962 pathnode->pathtarget = rel->reltarget;
963 pathnode->param_info = get_baserel_parampathinfo(root, rel,
964 required_outer);
965 pathnode->parallel_aware = parallel_workers > 0 ? true : false;
966 pathnode->parallel_safe = rel->consider_parallel;
967 pathnode->parallel_workers = parallel_workers;
968 pathnode->pathkeys = NIL; /* seqscan has unordered result */
969
970 cost_seqscan(pathnode, root, rel, pathnode->param_info);
971
972 return pathnode;
973 }
974
975 /*
976 * create_samplescan_path
977 * Creates a path node for a sampled table scan.
978 */
979 Path *
create_samplescan_path(PlannerInfo * root,RelOptInfo * rel,Relids required_outer)980 create_samplescan_path(PlannerInfo *root, RelOptInfo *rel, Relids required_outer)
981 {
982 Path *pathnode = makeNode(Path);
983
984 pathnode->pathtype = T_SampleScan;
985 pathnode->parent = rel;
986 pathnode->pathtarget = rel->reltarget;
987 pathnode->param_info = get_baserel_parampathinfo(root, rel,
988 required_outer);
989 pathnode->parallel_aware = false;
990 pathnode->parallel_safe = rel->consider_parallel;
991 pathnode->parallel_workers = 0;
992 pathnode->pathkeys = NIL; /* samplescan has unordered result */
993
994 cost_samplescan(pathnode, root, rel, pathnode->param_info);
995
996 return pathnode;
997 }
998
999 /*
1000 * create_index_path
1001 * Creates a path node for an index scan.
1002 *
1003 * 'index' is a usable index.
1004 * 'indexclauses' is a list of IndexClause nodes representing clauses
1005 * to be enforced as qual conditions in the scan.
1006 * 'indexorderbys' is a list of bare expressions (no RestrictInfos)
1007 * to be used as index ordering operators in the scan.
1008 * 'indexorderbycols' is an integer list of index column numbers (zero based)
1009 * the ordering operators can be used with.
1010 * 'pathkeys' describes the ordering of the path.
1011 * 'indexscandir' is ForwardScanDirection or BackwardScanDirection
1012 * for an ordered index, or NoMovementScanDirection for
1013 * an unordered index.
1014 * 'indexonly' is true if an index-only scan is wanted.
1015 * 'required_outer' is the set of outer relids for a parameterized path.
1016 * 'loop_count' is the number of repetitions of the indexscan to factor into
1017 * estimates of caching behavior.
1018 * 'partial_path' is true if constructing a parallel index scan path.
1019 *
1020 * Returns the new path node.
1021 */
1022 IndexPath *
create_index_path(PlannerInfo * root,IndexOptInfo * index,List * indexclauses,List * indexorderbys,List * indexorderbycols,List * pathkeys,ScanDirection indexscandir,bool indexonly,Relids required_outer,double loop_count,bool partial_path)1023 create_index_path(PlannerInfo *root,
1024 IndexOptInfo *index,
1025 List *indexclauses,
1026 List *indexorderbys,
1027 List *indexorderbycols,
1028 List *pathkeys,
1029 ScanDirection indexscandir,
1030 bool indexonly,
1031 Relids required_outer,
1032 double loop_count,
1033 bool partial_path)
1034 {
1035 IndexPath *pathnode = makeNode(IndexPath);
1036 RelOptInfo *rel = index->rel;
1037
1038 pathnode->path.pathtype = indexonly ? T_IndexOnlyScan : T_IndexScan;
1039 pathnode->path.parent = rel;
1040 pathnode->path.pathtarget = rel->reltarget;
1041 pathnode->path.param_info = get_baserel_parampathinfo(root, rel,
1042 required_outer);
1043 pathnode->path.parallel_aware = false;
1044 pathnode->path.parallel_safe = rel->consider_parallel;
1045 pathnode->path.parallel_workers = 0;
1046 pathnode->path.pathkeys = pathkeys;
1047
1048 pathnode->indexinfo = index;
1049 pathnode->indexclauses = indexclauses;
1050 pathnode->indexorderbys = indexorderbys;
1051 pathnode->indexorderbycols = indexorderbycols;
1052 pathnode->indexscandir = indexscandir;
1053
1054 cost_index(pathnode, root, loop_count, partial_path);
1055
1056 return pathnode;
1057 }
1058
1059 /*
1060 * create_bitmap_heap_path
1061 * Creates a path node for a bitmap scan.
1062 *
1063 * 'bitmapqual' is a tree of IndexPath, BitmapAndPath, and BitmapOrPath nodes.
1064 * 'required_outer' is the set of outer relids for a parameterized path.
1065 * 'loop_count' is the number of repetitions of the indexscan to factor into
1066 * estimates of caching behavior.
1067 *
1068 * loop_count should match the value used when creating the component
1069 * IndexPaths.
1070 */
1071 BitmapHeapPath *
create_bitmap_heap_path(PlannerInfo * root,RelOptInfo * rel,Path * bitmapqual,Relids required_outer,double loop_count,int parallel_degree)1072 create_bitmap_heap_path(PlannerInfo *root,
1073 RelOptInfo *rel,
1074 Path *bitmapqual,
1075 Relids required_outer,
1076 double loop_count,
1077 int parallel_degree)
1078 {
1079 BitmapHeapPath *pathnode = makeNode(BitmapHeapPath);
1080
1081 pathnode->path.pathtype = T_BitmapHeapScan;
1082 pathnode->path.parent = rel;
1083 pathnode->path.pathtarget = rel->reltarget;
1084 pathnode->path.param_info = get_baserel_parampathinfo(root, rel,
1085 required_outer);
1086 pathnode->path.parallel_aware = parallel_degree > 0 ? true : false;
1087 pathnode->path.parallel_safe = rel->consider_parallel;
1088 pathnode->path.parallel_workers = parallel_degree;
1089 pathnode->path.pathkeys = NIL; /* always unordered */
1090
1091 pathnode->bitmapqual = bitmapqual;
1092
1093 cost_bitmap_heap_scan(&pathnode->path, root, rel,
1094 pathnode->path.param_info,
1095 bitmapqual, loop_count);
1096
1097 return pathnode;
1098 }
1099
1100 /*
1101 * create_bitmap_and_path
1102 * Creates a path node representing a BitmapAnd.
1103 */
1104 BitmapAndPath *
create_bitmap_and_path(PlannerInfo * root,RelOptInfo * rel,List * bitmapquals)1105 create_bitmap_and_path(PlannerInfo *root,
1106 RelOptInfo *rel,
1107 List *bitmapquals)
1108 {
1109 BitmapAndPath *pathnode = makeNode(BitmapAndPath);
1110 Relids required_outer = NULL;
1111 ListCell *lc;
1112
1113 pathnode->path.pathtype = T_BitmapAnd;
1114 pathnode->path.parent = rel;
1115 pathnode->path.pathtarget = rel->reltarget;
1116
1117 /*
1118 * Identify the required outer rels as the union of what the child paths
1119 * depend on. (Alternatively, we could insist that the caller pass this
1120 * in, but it's more convenient and reliable to compute it here.)
1121 */
1122 foreach(lc, bitmapquals)
1123 {
1124 Path *bitmapqual = (Path *) lfirst(lc);
1125
1126 required_outer = bms_add_members(required_outer,
1127 PATH_REQ_OUTER(bitmapqual));
1128 }
1129 pathnode->path.param_info = get_baserel_parampathinfo(root, rel,
1130 required_outer);
1131
1132 /*
1133 * Currently, a BitmapHeapPath, BitmapAndPath, or BitmapOrPath will be
1134 * parallel-safe if and only if rel->consider_parallel is set. So, we can
1135 * set the flag for this path based only on the relation-level flag,
1136 * without actually iterating over the list of children.
1137 */
1138 pathnode->path.parallel_aware = false;
1139 pathnode->path.parallel_safe = rel->consider_parallel;
1140 pathnode->path.parallel_workers = 0;
1141
1142 pathnode->path.pathkeys = NIL; /* always unordered */
1143
1144 pathnode->bitmapquals = bitmapquals;
1145
1146 /* this sets bitmapselectivity as well as the regular cost fields: */
1147 cost_bitmap_and_node(pathnode, root);
1148
1149 return pathnode;
1150 }
1151
1152 /*
1153 * create_bitmap_or_path
1154 * Creates a path node representing a BitmapOr.
1155 */
1156 BitmapOrPath *
create_bitmap_or_path(PlannerInfo * root,RelOptInfo * rel,List * bitmapquals)1157 create_bitmap_or_path(PlannerInfo *root,
1158 RelOptInfo *rel,
1159 List *bitmapquals)
1160 {
1161 BitmapOrPath *pathnode = makeNode(BitmapOrPath);
1162 Relids required_outer = NULL;
1163 ListCell *lc;
1164
1165 pathnode->path.pathtype = T_BitmapOr;
1166 pathnode->path.parent = rel;
1167 pathnode->path.pathtarget = rel->reltarget;
1168
1169 /*
1170 * Identify the required outer rels as the union of what the child paths
1171 * depend on. (Alternatively, we could insist that the caller pass this
1172 * in, but it's more convenient and reliable to compute it here.)
1173 */
1174 foreach(lc, bitmapquals)
1175 {
1176 Path *bitmapqual = (Path *) lfirst(lc);
1177
1178 required_outer = bms_add_members(required_outer,
1179 PATH_REQ_OUTER(bitmapqual));
1180 }
1181 pathnode->path.param_info = get_baserel_parampathinfo(root, rel,
1182 required_outer);
1183
1184 /*
1185 * Currently, a BitmapHeapPath, BitmapAndPath, or BitmapOrPath will be
1186 * parallel-safe if and only if rel->consider_parallel is set. So, we can
1187 * set the flag for this path based only on the relation-level flag,
1188 * without actually iterating over the list of children.
1189 */
1190 pathnode->path.parallel_aware = false;
1191 pathnode->path.parallel_safe = rel->consider_parallel;
1192 pathnode->path.parallel_workers = 0;
1193
1194 pathnode->path.pathkeys = NIL; /* always unordered */
1195
1196 pathnode->bitmapquals = bitmapquals;
1197
1198 /* this sets bitmapselectivity as well as the regular cost fields: */
1199 cost_bitmap_or_node(pathnode, root);
1200
1201 return pathnode;
1202 }
1203
1204 /*
1205 * create_tidscan_path
1206 * Creates a path corresponding to a scan by TID, returning the pathnode.
1207 */
1208 TidPath *
create_tidscan_path(PlannerInfo * root,RelOptInfo * rel,List * tidquals,Relids required_outer)1209 create_tidscan_path(PlannerInfo *root, RelOptInfo *rel, List *tidquals,
1210 Relids required_outer)
1211 {
1212 TidPath *pathnode = makeNode(TidPath);
1213
1214 pathnode->path.pathtype = T_TidScan;
1215 pathnode->path.parent = rel;
1216 pathnode->path.pathtarget = rel->reltarget;
1217 pathnode->path.param_info = get_baserel_parampathinfo(root, rel,
1218 required_outer);
1219 pathnode->path.parallel_aware = false;
1220 pathnode->path.parallel_safe = rel->consider_parallel;
1221 pathnode->path.parallel_workers = 0;
1222 pathnode->path.pathkeys = NIL; /* always unordered */
1223
1224 pathnode->tidquals = tidquals;
1225
1226 cost_tidscan(&pathnode->path, root, rel, tidquals,
1227 pathnode->path.param_info);
1228
1229 return pathnode;
1230 }
1231
1232 /*
1233 * create_append_path
1234 * Creates a path corresponding to an Append plan, returning the
1235 * pathnode.
1236 *
1237 * Note that we must handle subpaths = NIL, representing a dummy access path.
1238 * Also, there are callers that pass root = NULL.
1239 */
1240 AppendPath *
create_append_path(PlannerInfo * root,RelOptInfo * rel,List * subpaths,List * partial_subpaths,List * pathkeys,Relids required_outer,int parallel_workers,bool parallel_aware,List * partitioned_rels,double rows)1241 create_append_path(PlannerInfo *root,
1242 RelOptInfo *rel,
1243 List *subpaths, List *partial_subpaths,
1244 List *pathkeys, Relids required_outer,
1245 int parallel_workers, bool parallel_aware,
1246 List *partitioned_rels, double rows)
1247 {
1248 AppendPath *pathnode = makeNode(AppendPath);
1249 ListCell *l;
1250
1251 Assert(!parallel_aware || parallel_workers > 0);
1252
1253 pathnode->path.pathtype = T_Append;
1254 pathnode->path.parent = rel;
1255 pathnode->path.pathtarget = rel->reltarget;
1256
1257 /*
1258 * When generating an Append path for a partitioned table, there may be
1259 * parameters that are useful so we can eliminate certain partitions
1260 * during execution. Here we'll go all the way and fully populate the
1261 * parameter info data as we do for normal base relations. However, we
1262 * need only bother doing this for RELOPT_BASEREL rels, as
1263 * RELOPT_OTHER_MEMBER_REL's Append paths are merged into the base rel's
1264 * Append subpaths. It would do no harm to do this, we just avoid it to
1265 * save wasting effort.
1266 */
1267 if (partitioned_rels != NIL && root && rel->reloptkind == RELOPT_BASEREL)
1268 pathnode->path.param_info = get_baserel_parampathinfo(root,
1269 rel,
1270 required_outer);
1271 else
1272 pathnode->path.param_info = get_appendrel_parampathinfo(rel,
1273 required_outer);
1274
1275 pathnode->path.parallel_aware = parallel_aware;
1276 pathnode->path.parallel_safe = rel->consider_parallel;
1277 pathnode->path.parallel_workers = parallel_workers;
1278 pathnode->path.pathkeys = pathkeys;
1279 pathnode->partitioned_rels = list_copy(partitioned_rels);
1280
1281 /*
1282 * For parallel append, non-partial paths are sorted by descending total
1283 * costs. That way, the total time to finish all non-partial paths is
1284 * minimized. Also, the partial paths are sorted by descending startup
1285 * costs. There may be some paths that require to do startup work by a
1286 * single worker. In such case, it's better for workers to choose the
1287 * expensive ones first, whereas the leader should choose the cheapest
1288 * startup plan.
1289 */
1290 if (pathnode->path.parallel_aware)
1291 {
1292 /*
1293 * We mustn't fiddle with the order of subpaths when the Append has
1294 * pathkeys. The order they're listed in is critical to keeping the
1295 * pathkeys valid.
1296 */
1297 Assert(pathkeys == NIL);
1298
1299 subpaths = list_qsort(subpaths, append_total_cost_compare);
1300 partial_subpaths = list_qsort(partial_subpaths,
1301 append_startup_cost_compare);
1302 }
1303 pathnode->first_partial_path = list_length(subpaths);
1304 pathnode->subpaths = list_concat(subpaths, partial_subpaths);
1305
1306 /*
1307 * Apply query-wide LIMIT if known and path is for sole base relation.
1308 * (Handling this at this low level is a bit klugy.)
1309 */
1310 if (root != NULL && bms_equal(rel->relids, root->all_baserels))
1311 pathnode->limit_tuples = root->limit_tuples;
1312 else
1313 pathnode->limit_tuples = -1.0;
1314
1315 foreach(l, pathnode->subpaths)
1316 {
1317 Path *subpath = (Path *) lfirst(l);
1318
1319 pathnode->path.parallel_safe = pathnode->path.parallel_safe &&
1320 subpath->parallel_safe;
1321
1322 /* All child paths must have same parameterization */
1323 Assert(bms_equal(PATH_REQ_OUTER(subpath), required_outer));
1324 }
1325
1326 Assert(!parallel_aware || pathnode->path.parallel_safe);
1327
1328 /*
1329 * If there's exactly one child path, the Append is a no-op and will be
1330 * discarded later (in setrefs.c); therefore, we can inherit the child's
1331 * size and cost, as well as its pathkeys if any (overriding whatever the
1332 * caller might've said). Otherwise, we must do the normal costsize
1333 * calculation.
1334 */
1335 if (list_length(pathnode->subpaths) == 1)
1336 {
1337 Path *child = (Path *) linitial(pathnode->subpaths);
1338
1339 pathnode->path.rows = child->rows;
1340 pathnode->path.startup_cost = child->startup_cost;
1341 pathnode->path.total_cost = child->total_cost;
1342 pathnode->path.pathkeys = child->pathkeys;
1343 }
1344 else
1345 cost_append(pathnode);
1346
1347 /* If the caller provided a row estimate, override the computed value. */
1348 if (rows >= 0)
1349 pathnode->path.rows = rows;
1350
1351 return pathnode;
1352 }
1353
1354 /*
1355 * append_total_cost_compare
1356 * qsort comparator for sorting append child paths by total_cost descending
1357 *
1358 * For equal total costs, we fall back to comparing startup costs; if those
1359 * are equal too, break ties using bms_compare on the paths' relids.
1360 * (This is to avoid getting unpredictable results from qsort.)
1361 */
1362 static int
append_total_cost_compare(const void * a,const void * b)1363 append_total_cost_compare(const void *a, const void *b)
1364 {
1365 Path *path1 = (Path *) lfirst(*(ListCell **) a);
1366 Path *path2 = (Path *) lfirst(*(ListCell **) b);
1367 int cmp;
1368
1369 cmp = compare_path_costs(path1, path2, TOTAL_COST);
1370 if (cmp != 0)
1371 return -cmp;
1372 return bms_compare(path1->parent->relids, path2->parent->relids);
1373 }
1374
1375 /*
1376 * append_startup_cost_compare
1377 * qsort comparator for sorting append child paths by startup_cost descending
1378 *
1379 * For equal startup costs, we fall back to comparing total costs; if those
1380 * are equal too, break ties using bms_compare on the paths' relids.
1381 * (This is to avoid getting unpredictable results from qsort.)
1382 */
1383 static int
append_startup_cost_compare(const void * a,const void * b)1384 append_startup_cost_compare(const void *a, const void *b)
1385 {
1386 Path *path1 = (Path *) lfirst(*(ListCell **) a);
1387 Path *path2 = (Path *) lfirst(*(ListCell **) b);
1388 int cmp;
1389
1390 cmp = compare_path_costs(path1, path2, STARTUP_COST);
1391 if (cmp != 0)
1392 return -cmp;
1393 return bms_compare(path1->parent->relids, path2->parent->relids);
1394 }
1395
1396 /*
1397 * create_merge_append_path
1398 * Creates a path corresponding to a MergeAppend plan, returning the
1399 * pathnode.
1400 */
1401 MergeAppendPath *
create_merge_append_path(PlannerInfo * root,RelOptInfo * rel,List * subpaths,List * pathkeys,Relids required_outer,List * partitioned_rels)1402 create_merge_append_path(PlannerInfo *root,
1403 RelOptInfo *rel,
1404 List *subpaths,
1405 List *pathkeys,
1406 Relids required_outer,
1407 List *partitioned_rels)
1408 {
1409 MergeAppendPath *pathnode = makeNode(MergeAppendPath);
1410 Cost input_startup_cost;
1411 Cost input_total_cost;
1412 ListCell *l;
1413
1414 pathnode->path.pathtype = T_MergeAppend;
1415 pathnode->path.parent = rel;
1416 pathnode->path.pathtarget = rel->reltarget;
1417 pathnode->path.param_info = get_appendrel_parampathinfo(rel,
1418 required_outer);
1419 pathnode->path.parallel_aware = false;
1420 pathnode->path.parallel_safe = rel->consider_parallel;
1421 pathnode->path.parallel_workers = 0;
1422 pathnode->path.pathkeys = pathkeys;
1423 pathnode->partitioned_rels = list_copy(partitioned_rels);
1424 pathnode->subpaths = subpaths;
1425
1426 /*
1427 * Apply query-wide LIMIT if known and path is for sole base relation.
1428 * (Handling this at this low level is a bit klugy.)
1429 */
1430 if (bms_equal(rel->relids, root->all_baserels))
1431 pathnode->limit_tuples = root->limit_tuples;
1432 else
1433 pathnode->limit_tuples = -1.0;
1434
1435 /*
1436 * Add up the sizes and costs of the input paths.
1437 */
1438 pathnode->path.rows = 0;
1439 input_startup_cost = 0;
1440 input_total_cost = 0;
1441 foreach(l, subpaths)
1442 {
1443 Path *subpath = (Path *) lfirst(l);
1444
1445 pathnode->path.rows += subpath->rows;
1446 pathnode->path.parallel_safe = pathnode->path.parallel_safe &&
1447 subpath->parallel_safe;
1448
1449 if (pathkeys_contained_in(pathkeys, subpath->pathkeys))
1450 {
1451 /* Subpath is adequately ordered, we won't need to sort it */
1452 input_startup_cost += subpath->startup_cost;
1453 input_total_cost += subpath->total_cost;
1454 }
1455 else
1456 {
1457 /* We'll need to insert a Sort node, so include cost for that */
1458 Path sort_path; /* dummy for result of cost_sort */
1459
1460 cost_sort(&sort_path,
1461 root,
1462 pathkeys,
1463 subpath->total_cost,
1464 subpath->parent->tuples,
1465 subpath->pathtarget->width,
1466 0.0,
1467 work_mem,
1468 pathnode->limit_tuples);
1469 input_startup_cost += sort_path.startup_cost;
1470 input_total_cost += sort_path.total_cost;
1471 }
1472
1473 /* All child paths must have same parameterization */
1474 Assert(bms_equal(PATH_REQ_OUTER(subpath), required_outer));
1475 }
1476
1477 /*
1478 * Now we can compute total costs of the MergeAppend. If there's exactly
1479 * one child path, the MergeAppend is a no-op and will be discarded later
1480 * (in setrefs.c); otherwise we do the normal cost calculation.
1481 */
1482 if (list_length(subpaths) == 1)
1483 {
1484 pathnode->path.startup_cost = input_startup_cost;
1485 pathnode->path.total_cost = input_total_cost;
1486 }
1487 else
1488 cost_merge_append(&pathnode->path, root,
1489 pathkeys, list_length(subpaths),
1490 input_startup_cost, input_total_cost,
1491 pathnode->path.rows);
1492
1493 return pathnode;
1494 }
1495
1496 /*
1497 * create_group_result_path
1498 * Creates a path representing a Result-and-nothing-else plan.
1499 *
1500 * This is only used for degenerate grouping cases, in which we know we
1501 * need to produce one result row, possibly filtered by a HAVING qual.
1502 */
1503 GroupResultPath *
create_group_result_path(PlannerInfo * root,RelOptInfo * rel,PathTarget * target,List * havingqual)1504 create_group_result_path(PlannerInfo *root, RelOptInfo *rel,
1505 PathTarget *target, List *havingqual)
1506 {
1507 GroupResultPath *pathnode = makeNode(GroupResultPath);
1508
1509 pathnode->path.pathtype = T_Result;
1510 pathnode->path.parent = rel;
1511 pathnode->path.pathtarget = target;
1512 pathnode->path.param_info = NULL; /* there are no other rels... */
1513 pathnode->path.parallel_aware = false;
1514 pathnode->path.parallel_safe = rel->consider_parallel;
1515 pathnode->path.parallel_workers = 0;
1516 pathnode->path.pathkeys = NIL;
1517 pathnode->quals = havingqual;
1518
1519 /*
1520 * We can't quite use cost_resultscan() because the quals we want to
1521 * account for are not baserestrict quals of the rel. Might as well just
1522 * hack it here.
1523 */
1524 pathnode->path.rows = 1;
1525 pathnode->path.startup_cost = target->cost.startup;
1526 pathnode->path.total_cost = target->cost.startup +
1527 cpu_tuple_cost + target->cost.per_tuple;
1528
1529 /*
1530 * Add cost of qual, if any --- but we ignore its selectivity, since our
1531 * rowcount estimate should be 1 no matter what the qual is.
1532 */
1533 if (havingqual)
1534 {
1535 QualCost qual_cost;
1536
1537 cost_qual_eval(&qual_cost, havingqual, root);
1538 /* havingqual is evaluated once at startup */
1539 pathnode->path.startup_cost += qual_cost.startup + qual_cost.per_tuple;
1540 pathnode->path.total_cost += qual_cost.startup + qual_cost.per_tuple;
1541 }
1542
1543 return pathnode;
1544 }
1545
1546 /*
1547 * create_material_path
1548 * Creates a path corresponding to a Material plan, returning the
1549 * pathnode.
1550 */
1551 MaterialPath *
create_material_path(RelOptInfo * rel,Path * subpath)1552 create_material_path(RelOptInfo *rel, Path *subpath)
1553 {
1554 MaterialPath *pathnode = makeNode(MaterialPath);
1555
1556 Assert(subpath->parent == rel);
1557
1558 pathnode->path.pathtype = T_Material;
1559 pathnode->path.parent = rel;
1560 pathnode->path.pathtarget = rel->reltarget;
1561 pathnode->path.param_info = subpath->param_info;
1562 pathnode->path.parallel_aware = false;
1563 pathnode->path.parallel_safe = rel->consider_parallel &&
1564 subpath->parallel_safe;
1565 pathnode->path.parallel_workers = subpath->parallel_workers;
1566 pathnode->path.pathkeys = subpath->pathkeys;
1567
1568 pathnode->subpath = subpath;
1569
1570 cost_material(&pathnode->path,
1571 subpath->startup_cost,
1572 subpath->total_cost,
1573 subpath->rows,
1574 subpath->pathtarget->width);
1575
1576 return pathnode;
1577 }
1578
1579 /*
1580 * create_unique_path
1581 * Creates a path representing elimination of distinct rows from the
1582 * input data. Distinct-ness is defined according to the needs of the
1583 * semijoin represented by sjinfo. If it is not possible to identify
1584 * how to make the data unique, NULL is returned.
1585 *
1586 * If used at all, this is likely to be called repeatedly on the same rel;
1587 * and the input subpath should always be the same (the cheapest_total path
1588 * for the rel). So we cache the result.
1589 */
1590 UniquePath *
create_unique_path(PlannerInfo * root,RelOptInfo * rel,Path * subpath,SpecialJoinInfo * sjinfo)1591 create_unique_path(PlannerInfo *root, RelOptInfo *rel, Path *subpath,
1592 SpecialJoinInfo *sjinfo)
1593 {
1594 UniquePath *pathnode;
1595 Path sort_path; /* dummy for result of cost_sort */
1596 Path agg_path; /* dummy for result of cost_agg */
1597 MemoryContext oldcontext;
1598 int numCols;
1599
1600 /* Caller made a mistake if subpath isn't cheapest_total ... */
1601 Assert(subpath == rel->cheapest_total_path);
1602 Assert(subpath->parent == rel);
1603 /* ... or if SpecialJoinInfo is the wrong one */
1604 Assert(sjinfo->jointype == JOIN_SEMI);
1605 Assert(bms_equal(rel->relids, sjinfo->syn_righthand));
1606
1607 /* If result already cached, return it */
1608 if (rel->cheapest_unique_path)
1609 return (UniquePath *) rel->cheapest_unique_path;
1610
1611 /* If it's not possible to unique-ify, return NULL */
1612 if (!(sjinfo->semi_can_btree || sjinfo->semi_can_hash))
1613 return NULL;
1614
1615 /*
1616 * When called during GEQO join planning, we are in a short-lived memory
1617 * context. We must make sure that the path and any subsidiary data
1618 * structures created for a baserel survive the GEQO cycle, else the
1619 * baserel is trashed for future GEQO cycles. On the other hand, when we
1620 * are creating those for a joinrel during GEQO, we don't want them to
1621 * clutter the main planning context. Upshot is that the best solution is
1622 * to explicitly allocate memory in the same context the given RelOptInfo
1623 * is in.
1624 */
1625 oldcontext = MemoryContextSwitchTo(GetMemoryChunkContext(rel));
1626
1627 pathnode = makeNode(UniquePath);
1628
1629 pathnode->path.pathtype = T_Unique;
1630 pathnode->path.parent = rel;
1631 pathnode->path.pathtarget = rel->reltarget;
1632 pathnode->path.param_info = subpath->param_info;
1633 pathnode->path.parallel_aware = false;
1634 pathnode->path.parallel_safe = rel->consider_parallel &&
1635 subpath->parallel_safe;
1636 pathnode->path.parallel_workers = subpath->parallel_workers;
1637
1638 /*
1639 * Assume the output is unsorted, since we don't necessarily have pathkeys
1640 * to represent it. (This might get overridden below.)
1641 */
1642 pathnode->path.pathkeys = NIL;
1643
1644 pathnode->subpath = subpath;
1645 pathnode->in_operators = sjinfo->semi_operators;
1646 pathnode->uniq_exprs = sjinfo->semi_rhs_exprs;
1647
1648 /*
1649 * If the input is a relation and it has a unique index that proves the
1650 * semi_rhs_exprs are unique, then we don't need to do anything. Note
1651 * that relation_has_unique_index_for automatically considers restriction
1652 * clauses for the rel, as well.
1653 */
1654 if (rel->rtekind == RTE_RELATION && sjinfo->semi_can_btree &&
1655 relation_has_unique_index_for(root, rel, NIL,
1656 sjinfo->semi_rhs_exprs,
1657 sjinfo->semi_operators))
1658 {
1659 pathnode->umethod = UNIQUE_PATH_NOOP;
1660 pathnode->path.rows = rel->rows;
1661 pathnode->path.startup_cost = subpath->startup_cost;
1662 pathnode->path.total_cost = subpath->total_cost;
1663 pathnode->path.pathkeys = subpath->pathkeys;
1664
1665 rel->cheapest_unique_path = (Path *) pathnode;
1666
1667 MemoryContextSwitchTo(oldcontext);
1668
1669 return pathnode;
1670 }
1671
1672 /*
1673 * If the input is a subquery whose output must be unique already, then we
1674 * don't need to do anything. The test for uniqueness has to consider
1675 * exactly which columns we are extracting; for example "SELECT DISTINCT
1676 * x,y" doesn't guarantee that x alone is distinct. So we cannot check for
1677 * this optimization unless semi_rhs_exprs consists only of simple Vars
1678 * referencing subquery outputs. (Possibly we could do something with
1679 * expressions in the subquery outputs, too, but for now keep it simple.)
1680 */
1681 if (rel->rtekind == RTE_SUBQUERY)
1682 {
1683 RangeTblEntry *rte = planner_rt_fetch(rel->relid, root);
1684
1685 if (query_supports_distinctness(rte->subquery))
1686 {
1687 List *sub_tlist_colnos;
1688
1689 sub_tlist_colnos = translate_sub_tlist(sjinfo->semi_rhs_exprs,
1690 rel->relid);
1691
1692 if (sub_tlist_colnos &&
1693 query_is_distinct_for(rte->subquery,
1694 sub_tlist_colnos,
1695 sjinfo->semi_operators))
1696 {
1697 pathnode->umethod = UNIQUE_PATH_NOOP;
1698 pathnode->path.rows = rel->rows;
1699 pathnode->path.startup_cost = subpath->startup_cost;
1700 pathnode->path.total_cost = subpath->total_cost;
1701 pathnode->path.pathkeys = subpath->pathkeys;
1702
1703 rel->cheapest_unique_path = (Path *) pathnode;
1704
1705 MemoryContextSwitchTo(oldcontext);
1706
1707 return pathnode;
1708 }
1709 }
1710 }
1711
1712 /* Estimate number of output rows */
1713 pathnode->path.rows = estimate_num_groups(root,
1714 sjinfo->semi_rhs_exprs,
1715 rel->rows,
1716 NULL);
1717 numCols = list_length(sjinfo->semi_rhs_exprs);
1718
1719 if (sjinfo->semi_can_btree)
1720 {
1721 /*
1722 * Estimate cost for sort+unique implementation
1723 */
1724 cost_sort(&sort_path, root, NIL,
1725 subpath->total_cost,
1726 rel->rows,
1727 subpath->pathtarget->width,
1728 0.0,
1729 work_mem,
1730 -1.0);
1731
1732 /*
1733 * Charge one cpu_operator_cost per comparison per input tuple. We
1734 * assume all columns get compared at most of the tuples. (XXX
1735 * probably this is an overestimate.) This should agree with
1736 * create_upper_unique_path.
1737 */
1738 sort_path.total_cost += cpu_operator_cost * rel->rows * numCols;
1739 }
1740
1741 if (sjinfo->semi_can_hash)
1742 {
1743 /*
1744 * Estimate the overhead per hashtable entry at 64 bytes (same as in
1745 * planner.c).
1746 */
1747 int hashentrysize = subpath->pathtarget->width + 64;
1748
1749 if (hashentrysize * pathnode->path.rows > work_mem * 1024L)
1750 {
1751 /*
1752 * We should not try to hash. Hack the SpecialJoinInfo to
1753 * remember this, in case we come through here again.
1754 */
1755 sjinfo->semi_can_hash = false;
1756 }
1757 else
1758 cost_agg(&agg_path, root,
1759 AGG_HASHED, NULL,
1760 numCols, pathnode->path.rows,
1761 NIL,
1762 subpath->startup_cost,
1763 subpath->total_cost,
1764 rel->rows);
1765 }
1766
1767 if (sjinfo->semi_can_btree && sjinfo->semi_can_hash)
1768 {
1769 if (agg_path.total_cost < sort_path.total_cost)
1770 pathnode->umethod = UNIQUE_PATH_HASH;
1771 else
1772 pathnode->umethod = UNIQUE_PATH_SORT;
1773 }
1774 else if (sjinfo->semi_can_btree)
1775 pathnode->umethod = UNIQUE_PATH_SORT;
1776 else if (sjinfo->semi_can_hash)
1777 pathnode->umethod = UNIQUE_PATH_HASH;
1778 else
1779 {
1780 /* we can get here only if we abandoned hashing above */
1781 MemoryContextSwitchTo(oldcontext);
1782 return NULL;
1783 }
1784
1785 if (pathnode->umethod == UNIQUE_PATH_HASH)
1786 {
1787 pathnode->path.startup_cost = agg_path.startup_cost;
1788 pathnode->path.total_cost = agg_path.total_cost;
1789 }
1790 else
1791 {
1792 pathnode->path.startup_cost = sort_path.startup_cost;
1793 pathnode->path.total_cost = sort_path.total_cost;
1794 }
1795
1796 rel->cheapest_unique_path = (Path *) pathnode;
1797
1798 MemoryContextSwitchTo(oldcontext);
1799
1800 return pathnode;
1801 }
1802
1803 /*
1804 * create_gather_merge_path
1805 *
1806 * Creates a path corresponding to a gather merge scan, returning
1807 * the pathnode.
1808 */
1809 GatherMergePath *
create_gather_merge_path(PlannerInfo * root,RelOptInfo * rel,Path * subpath,PathTarget * target,List * pathkeys,Relids required_outer,double * rows)1810 create_gather_merge_path(PlannerInfo *root, RelOptInfo *rel, Path *subpath,
1811 PathTarget *target, List *pathkeys,
1812 Relids required_outer, double *rows)
1813 {
1814 GatherMergePath *pathnode = makeNode(GatherMergePath);
1815 Cost input_startup_cost = 0;
1816 Cost input_total_cost = 0;
1817
1818 Assert(subpath->parallel_safe);
1819 Assert(pathkeys);
1820
1821 pathnode->path.pathtype = T_GatherMerge;
1822 pathnode->path.parent = rel;
1823 pathnode->path.param_info = get_baserel_parampathinfo(root, rel,
1824 required_outer);
1825 pathnode->path.parallel_aware = false;
1826
1827 pathnode->subpath = subpath;
1828 pathnode->num_workers = subpath->parallel_workers;
1829 pathnode->path.pathkeys = pathkeys;
1830 pathnode->path.pathtarget = target ? target : rel->reltarget;
1831 pathnode->path.rows += subpath->rows;
1832
1833 if (pathkeys_contained_in(pathkeys, subpath->pathkeys))
1834 {
1835 /* Subpath is adequately ordered, we won't need to sort it */
1836 input_startup_cost += subpath->startup_cost;
1837 input_total_cost += subpath->total_cost;
1838 }
1839 else
1840 {
1841 /* We'll need to insert a Sort node, so include cost for that */
1842 Path sort_path; /* dummy for result of cost_sort */
1843
1844 cost_sort(&sort_path,
1845 root,
1846 pathkeys,
1847 subpath->total_cost,
1848 subpath->rows,
1849 subpath->pathtarget->width,
1850 0.0,
1851 work_mem,
1852 -1);
1853 input_startup_cost += sort_path.startup_cost;
1854 input_total_cost += sort_path.total_cost;
1855 }
1856
1857 cost_gather_merge(pathnode, root, rel, pathnode->path.param_info,
1858 input_startup_cost, input_total_cost, rows);
1859
1860 return pathnode;
1861 }
1862
1863 /*
1864 * translate_sub_tlist - get subquery column numbers represented by tlist
1865 *
1866 * The given targetlist usually contains only Vars referencing the given relid.
1867 * Extract their varattnos (ie, the column numbers of the subquery) and return
1868 * as an integer List.
1869 *
1870 * If any of the tlist items is not a simple Var, we cannot determine whether
1871 * the subquery's uniqueness condition (if any) matches ours, so punt and
1872 * return NIL.
1873 */
1874 static List *
translate_sub_tlist(List * tlist,int relid)1875 translate_sub_tlist(List *tlist, int relid)
1876 {
1877 List *result = NIL;
1878 ListCell *l;
1879
1880 foreach(l, tlist)
1881 {
1882 Var *var = (Var *) lfirst(l);
1883
1884 if (!var || !IsA(var, Var) ||
1885 var->varno != relid)
1886 return NIL; /* punt */
1887
1888 result = lappend_int(result, var->varattno);
1889 }
1890 return result;
1891 }
1892
1893 /*
1894 * create_gather_path
1895 * Creates a path corresponding to a gather scan, returning the
1896 * pathnode.
1897 *
1898 * 'rows' may optionally be set to override row estimates from other sources.
1899 */
1900 GatherPath *
create_gather_path(PlannerInfo * root,RelOptInfo * rel,Path * subpath,PathTarget * target,Relids required_outer,double * rows)1901 create_gather_path(PlannerInfo *root, RelOptInfo *rel, Path *subpath,
1902 PathTarget *target, Relids required_outer, double *rows)
1903 {
1904 GatherPath *pathnode = makeNode(GatherPath);
1905
1906 Assert(subpath->parallel_safe);
1907
1908 pathnode->path.pathtype = T_Gather;
1909 pathnode->path.parent = rel;
1910 pathnode->path.pathtarget = target;
1911 pathnode->path.param_info = get_baserel_parampathinfo(root, rel,
1912 required_outer);
1913 pathnode->path.parallel_aware = false;
1914 pathnode->path.parallel_safe = false;
1915 pathnode->path.parallel_workers = 0;
1916 pathnode->path.pathkeys = NIL; /* Gather has unordered result */
1917
1918 pathnode->subpath = subpath;
1919 pathnode->num_workers = subpath->parallel_workers;
1920 pathnode->single_copy = false;
1921
1922 if (pathnode->num_workers == 0)
1923 {
1924 pathnode->path.pathkeys = subpath->pathkeys;
1925 pathnode->num_workers = 1;
1926 pathnode->single_copy = true;
1927 }
1928
1929 cost_gather(pathnode, root, rel, pathnode->path.param_info, rows);
1930
1931 return pathnode;
1932 }
1933
1934 /*
1935 * create_subqueryscan_path
1936 * Creates a path corresponding to a scan of a subquery,
1937 * returning the pathnode.
1938 */
1939 SubqueryScanPath *
create_subqueryscan_path(PlannerInfo * root,RelOptInfo * rel,Path * subpath,List * pathkeys,Relids required_outer)1940 create_subqueryscan_path(PlannerInfo *root, RelOptInfo *rel, Path *subpath,
1941 List *pathkeys, Relids required_outer)
1942 {
1943 SubqueryScanPath *pathnode = makeNode(SubqueryScanPath);
1944
1945 pathnode->path.pathtype = T_SubqueryScan;
1946 pathnode->path.parent = rel;
1947 pathnode->path.pathtarget = rel->reltarget;
1948 pathnode->path.param_info = get_baserel_parampathinfo(root, rel,
1949 required_outer);
1950 pathnode->path.parallel_aware = false;
1951 pathnode->path.parallel_safe = rel->consider_parallel &&
1952 subpath->parallel_safe;
1953 pathnode->path.parallel_workers = subpath->parallel_workers;
1954 pathnode->path.pathkeys = pathkeys;
1955 pathnode->subpath = subpath;
1956
1957 cost_subqueryscan(pathnode, root, rel, pathnode->path.param_info);
1958
1959 return pathnode;
1960 }
1961
1962 /*
1963 * create_functionscan_path
1964 * Creates a path corresponding to a sequential scan of a function,
1965 * returning the pathnode.
1966 */
1967 Path *
create_functionscan_path(PlannerInfo * root,RelOptInfo * rel,List * pathkeys,Relids required_outer)1968 create_functionscan_path(PlannerInfo *root, RelOptInfo *rel,
1969 List *pathkeys, Relids required_outer)
1970 {
1971 Path *pathnode = makeNode(Path);
1972
1973 pathnode->pathtype = T_FunctionScan;
1974 pathnode->parent = rel;
1975 pathnode->pathtarget = rel->reltarget;
1976 pathnode->param_info = get_baserel_parampathinfo(root, rel,
1977 required_outer);
1978 pathnode->parallel_aware = false;
1979 pathnode->parallel_safe = rel->consider_parallel;
1980 pathnode->parallel_workers = 0;
1981 pathnode->pathkeys = pathkeys;
1982
1983 cost_functionscan(pathnode, root, rel, pathnode->param_info);
1984
1985 return pathnode;
1986 }
1987
1988 /*
1989 * create_tablefuncscan_path
1990 * Creates a path corresponding to a sequential scan of a table function,
1991 * returning the pathnode.
1992 */
1993 Path *
create_tablefuncscan_path(PlannerInfo * root,RelOptInfo * rel,Relids required_outer)1994 create_tablefuncscan_path(PlannerInfo *root, RelOptInfo *rel,
1995 Relids required_outer)
1996 {
1997 Path *pathnode = makeNode(Path);
1998
1999 pathnode->pathtype = T_TableFuncScan;
2000 pathnode->parent = rel;
2001 pathnode->pathtarget = rel->reltarget;
2002 pathnode->param_info = get_baserel_parampathinfo(root, rel,
2003 required_outer);
2004 pathnode->parallel_aware = false;
2005 pathnode->parallel_safe = rel->consider_parallel;
2006 pathnode->parallel_workers = 0;
2007 pathnode->pathkeys = NIL; /* result is always unordered */
2008
2009 cost_tablefuncscan(pathnode, root, rel, pathnode->param_info);
2010
2011 return pathnode;
2012 }
2013
2014 /*
2015 * create_valuesscan_path
2016 * Creates a path corresponding to a scan of a VALUES list,
2017 * returning the pathnode.
2018 */
2019 Path *
create_valuesscan_path(PlannerInfo * root,RelOptInfo * rel,Relids required_outer)2020 create_valuesscan_path(PlannerInfo *root, RelOptInfo *rel,
2021 Relids required_outer)
2022 {
2023 Path *pathnode = makeNode(Path);
2024
2025 pathnode->pathtype = T_ValuesScan;
2026 pathnode->parent = rel;
2027 pathnode->pathtarget = rel->reltarget;
2028 pathnode->param_info = get_baserel_parampathinfo(root, rel,
2029 required_outer);
2030 pathnode->parallel_aware = false;
2031 pathnode->parallel_safe = rel->consider_parallel;
2032 pathnode->parallel_workers = 0;
2033 pathnode->pathkeys = NIL; /* result is always unordered */
2034
2035 cost_valuesscan(pathnode, root, rel, pathnode->param_info);
2036
2037 return pathnode;
2038 }
2039
2040 /*
2041 * create_ctescan_path
2042 * Creates a path corresponding to a scan of a non-self-reference CTE,
2043 * returning the pathnode.
2044 */
2045 Path *
create_ctescan_path(PlannerInfo * root,RelOptInfo * rel,Relids required_outer)2046 create_ctescan_path(PlannerInfo *root, RelOptInfo *rel, Relids required_outer)
2047 {
2048 Path *pathnode = makeNode(Path);
2049
2050 pathnode->pathtype = T_CteScan;
2051 pathnode->parent = rel;
2052 pathnode->pathtarget = rel->reltarget;
2053 pathnode->param_info = get_baserel_parampathinfo(root, rel,
2054 required_outer);
2055 pathnode->parallel_aware = false;
2056 pathnode->parallel_safe = rel->consider_parallel;
2057 pathnode->parallel_workers = 0;
2058 pathnode->pathkeys = NIL; /* XXX for now, result is always unordered */
2059
2060 cost_ctescan(pathnode, root, rel, pathnode->param_info);
2061
2062 return pathnode;
2063 }
2064
2065 /*
2066 * create_namedtuplestorescan_path
2067 * Creates a path corresponding to a scan of a named tuplestore, returning
2068 * the pathnode.
2069 */
2070 Path *
create_namedtuplestorescan_path(PlannerInfo * root,RelOptInfo * rel,Relids required_outer)2071 create_namedtuplestorescan_path(PlannerInfo *root, RelOptInfo *rel,
2072 Relids required_outer)
2073 {
2074 Path *pathnode = makeNode(Path);
2075
2076 pathnode->pathtype = T_NamedTuplestoreScan;
2077 pathnode->parent = rel;
2078 pathnode->pathtarget = rel->reltarget;
2079 pathnode->param_info = get_baserel_parampathinfo(root, rel,
2080 required_outer);
2081 pathnode->parallel_aware = false;
2082 pathnode->parallel_safe = rel->consider_parallel;
2083 pathnode->parallel_workers = 0;
2084 pathnode->pathkeys = NIL; /* result is always unordered */
2085
2086 cost_namedtuplestorescan(pathnode, root, rel, pathnode->param_info);
2087
2088 return pathnode;
2089 }
2090
2091 /*
2092 * create_resultscan_path
2093 * Creates a path corresponding to a scan of an RTE_RESULT relation,
2094 * returning the pathnode.
2095 */
2096 Path *
create_resultscan_path(PlannerInfo * root,RelOptInfo * rel,Relids required_outer)2097 create_resultscan_path(PlannerInfo *root, RelOptInfo *rel,
2098 Relids required_outer)
2099 {
2100 Path *pathnode = makeNode(Path);
2101
2102 pathnode->pathtype = T_Result;
2103 pathnode->parent = rel;
2104 pathnode->pathtarget = rel->reltarget;
2105 pathnode->param_info = get_baserel_parampathinfo(root, rel,
2106 required_outer);
2107 pathnode->parallel_aware = false;
2108 pathnode->parallel_safe = rel->consider_parallel;
2109 pathnode->parallel_workers = 0;
2110 pathnode->pathkeys = NIL; /* result is always unordered */
2111
2112 cost_resultscan(pathnode, root, rel, pathnode->param_info);
2113
2114 return pathnode;
2115 }
2116
2117 /*
2118 * create_worktablescan_path
2119 * Creates a path corresponding to a scan of a self-reference CTE,
2120 * returning the pathnode.
2121 */
2122 Path *
create_worktablescan_path(PlannerInfo * root,RelOptInfo * rel,Relids required_outer)2123 create_worktablescan_path(PlannerInfo *root, RelOptInfo *rel,
2124 Relids required_outer)
2125 {
2126 Path *pathnode = makeNode(Path);
2127
2128 pathnode->pathtype = T_WorkTableScan;
2129 pathnode->parent = rel;
2130 pathnode->pathtarget = rel->reltarget;
2131 pathnode->param_info = get_baserel_parampathinfo(root, rel,
2132 required_outer);
2133 pathnode->parallel_aware = false;
2134 pathnode->parallel_safe = rel->consider_parallel;
2135 pathnode->parallel_workers = 0;
2136 pathnode->pathkeys = NIL; /* result is always unordered */
2137
2138 /* Cost is the same as for a regular CTE scan */
2139 cost_ctescan(pathnode, root, rel, pathnode->param_info);
2140
2141 return pathnode;
2142 }
2143
2144 /*
2145 * create_foreignscan_path
2146 * Creates a path corresponding to a scan of a foreign base table,
2147 * returning the pathnode.
2148 *
2149 * This function is never called from core Postgres; rather, it's expected
2150 * to be called by the GetForeignPaths function of a foreign data wrapper.
2151 * We make the FDW supply all fields of the path, since we do not have any way
2152 * to calculate them in core. However, there is a usually-sane default for
2153 * the pathtarget (rel->reltarget), so we let a NULL for "target" select that.
2154 */
2155 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)2156 create_foreignscan_path(PlannerInfo *root, RelOptInfo *rel,
2157 PathTarget *target,
2158 double rows, Cost startup_cost, Cost total_cost,
2159 List *pathkeys,
2160 Relids required_outer,
2161 Path *fdw_outerpath,
2162 List *fdw_private)
2163 {
2164 ForeignPath *pathnode = makeNode(ForeignPath);
2165
2166 /* Historically some FDWs were confused about when to use this */
2167 Assert(IS_SIMPLE_REL(rel));
2168
2169 pathnode->path.pathtype = T_ForeignScan;
2170 pathnode->path.parent = rel;
2171 pathnode->path.pathtarget = target ? target : rel->reltarget;
2172 pathnode->path.param_info = get_baserel_parampathinfo(root, rel,
2173 required_outer);
2174 pathnode->path.parallel_aware = false;
2175 pathnode->path.parallel_safe = rel->consider_parallel;
2176 pathnode->path.parallel_workers = 0;
2177 pathnode->path.rows = rows;
2178 pathnode->path.startup_cost = startup_cost;
2179 pathnode->path.total_cost = total_cost;
2180 pathnode->path.pathkeys = pathkeys;
2181
2182 pathnode->fdw_outerpath = fdw_outerpath;
2183 pathnode->fdw_private = fdw_private;
2184
2185 return pathnode;
2186 }
2187
2188 /*
2189 * create_foreign_join_path
2190 * Creates a path corresponding to a scan of a foreign join,
2191 * returning the pathnode.
2192 *
2193 * This function is never called from core Postgres; rather, it's expected
2194 * to be called by the GetForeignJoinPaths function of a foreign data wrapper.
2195 * We make the FDW supply all fields of the path, since we do not have any way
2196 * to calculate them in core. However, there is a usually-sane default for
2197 * the pathtarget (rel->reltarget), so we let a NULL for "target" select that.
2198 */
2199 ForeignPath *
create_foreign_join_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)2200 create_foreign_join_path(PlannerInfo *root, RelOptInfo *rel,
2201 PathTarget *target,
2202 double rows, Cost startup_cost, Cost total_cost,
2203 List *pathkeys,
2204 Relids required_outer,
2205 Path *fdw_outerpath,
2206 List *fdw_private)
2207 {
2208 ForeignPath *pathnode = makeNode(ForeignPath);
2209
2210 /*
2211 * We should use get_joinrel_parampathinfo to handle parameterized paths,
2212 * but the API of this function doesn't support it, and existing
2213 * extensions aren't yet trying to build such paths anyway. For the
2214 * moment just throw an error if someone tries it; eventually we should
2215 * revisit this.
2216 */
2217 if (!bms_is_empty(required_outer) || !bms_is_empty(rel->lateral_relids))
2218 elog(ERROR, "parameterized foreign joins are not supported yet");
2219
2220 pathnode->path.pathtype = T_ForeignScan;
2221 pathnode->path.parent = rel;
2222 pathnode->path.pathtarget = target ? target : rel->reltarget;
2223 pathnode->path.param_info = NULL; /* XXX see above */
2224 pathnode->path.parallel_aware = false;
2225 pathnode->path.parallel_safe = rel->consider_parallel;
2226 pathnode->path.parallel_workers = 0;
2227 pathnode->path.rows = rows;
2228 pathnode->path.startup_cost = startup_cost;
2229 pathnode->path.total_cost = total_cost;
2230 pathnode->path.pathkeys = pathkeys;
2231
2232 pathnode->fdw_outerpath = fdw_outerpath;
2233 pathnode->fdw_private = fdw_private;
2234
2235 return pathnode;
2236 }
2237
2238 /*
2239 * create_foreign_upper_path
2240 * Creates a path corresponding to an upper relation that's computed
2241 * directly by an FDW, returning the pathnode.
2242 *
2243 * This function is never called from core Postgres; rather, it's expected to
2244 * be called by the GetForeignUpperPaths function of a foreign data wrapper.
2245 * We make the FDW supply all fields of the path, since we do not have any way
2246 * to calculate them in core. However, there is a usually-sane default for
2247 * the pathtarget (rel->reltarget), so we let a NULL for "target" select that.
2248 */
2249 ForeignPath *
create_foreign_upper_path(PlannerInfo * root,RelOptInfo * rel,PathTarget * target,double rows,Cost startup_cost,Cost total_cost,List * pathkeys,Path * fdw_outerpath,List * fdw_private)2250 create_foreign_upper_path(PlannerInfo *root, RelOptInfo *rel,
2251 PathTarget *target,
2252 double rows, Cost startup_cost, Cost total_cost,
2253 List *pathkeys,
2254 Path *fdw_outerpath,
2255 List *fdw_private)
2256 {
2257 ForeignPath *pathnode = makeNode(ForeignPath);
2258
2259 /*
2260 * Upper relations should never have any lateral references, since joining
2261 * is complete.
2262 */
2263 Assert(bms_is_empty(rel->lateral_relids));
2264
2265 pathnode->path.pathtype = T_ForeignScan;
2266 pathnode->path.parent = rel;
2267 pathnode->path.pathtarget = target ? target : rel->reltarget;
2268 pathnode->path.param_info = NULL;
2269 pathnode->path.parallel_aware = false;
2270 pathnode->path.parallel_safe = rel->consider_parallel;
2271 pathnode->path.parallel_workers = 0;
2272 pathnode->path.rows = rows;
2273 pathnode->path.startup_cost = startup_cost;
2274 pathnode->path.total_cost = total_cost;
2275 pathnode->path.pathkeys = pathkeys;
2276
2277 pathnode->fdw_outerpath = fdw_outerpath;
2278 pathnode->fdw_private = fdw_private;
2279
2280 return pathnode;
2281 }
2282
2283 /*
2284 * calc_nestloop_required_outer
2285 * Compute the required_outer set for a nestloop join path
2286 *
2287 * Note: result must not share storage with either input
2288 */
2289 Relids
calc_nestloop_required_outer(Relids outerrelids,Relids outer_paramrels,Relids innerrelids,Relids inner_paramrels)2290 calc_nestloop_required_outer(Relids outerrelids,
2291 Relids outer_paramrels,
2292 Relids innerrelids,
2293 Relids inner_paramrels)
2294 {
2295 Relids required_outer;
2296
2297 /* inner_path can require rels from outer path, but not vice versa */
2298 Assert(!bms_overlap(outer_paramrels, innerrelids));
2299 /* easy case if inner path is not parameterized */
2300 if (!inner_paramrels)
2301 return bms_copy(outer_paramrels);
2302 /* else, form the union ... */
2303 required_outer = bms_union(outer_paramrels, inner_paramrels);
2304 /* ... and remove any mention of now-satisfied outer rels */
2305 required_outer = bms_del_members(required_outer,
2306 outerrelids);
2307 /* maintain invariant that required_outer is exactly NULL if empty */
2308 if (bms_is_empty(required_outer))
2309 {
2310 bms_free(required_outer);
2311 required_outer = NULL;
2312 }
2313 return required_outer;
2314 }
2315
2316 /*
2317 * calc_non_nestloop_required_outer
2318 * Compute the required_outer set for a merge or hash join path
2319 *
2320 * Note: result must not share storage with either input
2321 */
2322 Relids
calc_non_nestloop_required_outer(Path * outer_path,Path * inner_path)2323 calc_non_nestloop_required_outer(Path *outer_path, Path *inner_path)
2324 {
2325 Relids outer_paramrels = PATH_REQ_OUTER(outer_path);
2326 Relids inner_paramrels = PATH_REQ_OUTER(inner_path);
2327 Relids required_outer;
2328
2329 /* neither path can require rels from the other */
2330 Assert(!bms_overlap(outer_paramrels, inner_path->parent->relids));
2331 Assert(!bms_overlap(inner_paramrels, outer_path->parent->relids));
2332 /* form the union ... */
2333 required_outer = bms_union(outer_paramrels, inner_paramrels);
2334 /* we do not need an explicit test for empty; bms_union gets it right */
2335 return required_outer;
2336 }
2337
2338 /*
2339 * create_nestloop_path
2340 * Creates a pathnode corresponding to a nestloop join between two
2341 * relations.
2342 *
2343 * 'joinrel' is the join relation.
2344 * 'jointype' is the type of join required
2345 * 'workspace' is the result from initial_cost_nestloop
2346 * 'extra' contains various information about the join
2347 * 'outer_path' is the outer path
2348 * 'inner_path' is the inner path
2349 * 'restrict_clauses' are the RestrictInfo nodes to apply at the join
2350 * 'pathkeys' are the path keys of the new join path
2351 * 'required_outer' is the set of required outer rels
2352 *
2353 * Returns the resulting path node.
2354 */
2355 NestPath *
create_nestloop_path(PlannerInfo * root,RelOptInfo * joinrel,JoinType jointype,JoinCostWorkspace * workspace,JoinPathExtraData * extra,Path * outer_path,Path * inner_path,List * restrict_clauses,List * pathkeys,Relids required_outer)2356 create_nestloop_path(PlannerInfo *root,
2357 RelOptInfo *joinrel,
2358 JoinType jointype,
2359 JoinCostWorkspace *workspace,
2360 JoinPathExtraData *extra,
2361 Path *outer_path,
2362 Path *inner_path,
2363 List *restrict_clauses,
2364 List *pathkeys,
2365 Relids required_outer)
2366 {
2367 NestPath *pathnode = makeNode(NestPath);
2368 Relids inner_req_outer = PATH_REQ_OUTER(inner_path);
2369
2370 /*
2371 * If the inner path is parameterized by the outer, we must drop any
2372 * restrict_clauses that are due to be moved into the inner path. We have
2373 * to do this now, rather than postpone the work till createplan time,
2374 * because the restrict_clauses list can affect the size and cost
2375 * estimates for this path.
2376 */
2377 if (bms_overlap(inner_req_outer, outer_path->parent->relids))
2378 {
2379 Relids inner_and_outer = bms_union(inner_path->parent->relids,
2380 inner_req_outer);
2381 List *jclauses = NIL;
2382 ListCell *lc;
2383
2384 foreach(lc, restrict_clauses)
2385 {
2386 RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
2387
2388 if (!join_clause_is_movable_into(rinfo,
2389 inner_path->parent->relids,
2390 inner_and_outer))
2391 jclauses = lappend(jclauses, rinfo);
2392 }
2393 restrict_clauses = jclauses;
2394 }
2395
2396 pathnode->path.pathtype = T_NestLoop;
2397 pathnode->path.parent = joinrel;
2398 pathnode->path.pathtarget = joinrel->reltarget;
2399 pathnode->path.param_info =
2400 get_joinrel_parampathinfo(root,
2401 joinrel,
2402 outer_path,
2403 inner_path,
2404 extra->sjinfo,
2405 required_outer,
2406 &restrict_clauses);
2407 pathnode->path.parallel_aware = false;
2408 pathnode->path.parallel_safe = joinrel->consider_parallel &&
2409 outer_path->parallel_safe && inner_path->parallel_safe;
2410 /* This is a foolish way to estimate parallel_workers, but for now... */
2411 pathnode->path.parallel_workers = outer_path->parallel_workers;
2412 pathnode->path.pathkeys = pathkeys;
2413 pathnode->jointype = jointype;
2414 pathnode->inner_unique = extra->inner_unique;
2415 pathnode->outerjoinpath = outer_path;
2416 pathnode->innerjoinpath = inner_path;
2417 pathnode->joinrestrictinfo = restrict_clauses;
2418
2419 final_cost_nestloop(root, pathnode, workspace, extra);
2420
2421 return pathnode;
2422 }
2423
2424 /*
2425 * create_mergejoin_path
2426 * Creates a pathnode corresponding to a mergejoin join between
2427 * two relations
2428 *
2429 * 'joinrel' is the join relation
2430 * 'jointype' is the type of join required
2431 * 'workspace' is the result from initial_cost_mergejoin
2432 * 'extra' contains various information about the join
2433 * 'outer_path' is the outer path
2434 * 'inner_path' is the inner path
2435 * 'restrict_clauses' are the RestrictInfo nodes to apply at the join
2436 * 'pathkeys' are the path keys of the new join path
2437 * 'required_outer' is the set of required outer rels
2438 * 'mergeclauses' are the RestrictInfo nodes to use as merge clauses
2439 * (this should be a subset of the restrict_clauses list)
2440 * 'outersortkeys' are the sort varkeys for the outer relation
2441 * 'innersortkeys' are the sort varkeys for the inner relation
2442 */
2443 MergePath *
create_mergejoin_path(PlannerInfo * root,RelOptInfo * joinrel,JoinType jointype,JoinCostWorkspace * workspace,JoinPathExtraData * extra,Path * outer_path,Path * inner_path,List * restrict_clauses,List * pathkeys,Relids required_outer,List * mergeclauses,List * outersortkeys,List * innersortkeys)2444 create_mergejoin_path(PlannerInfo *root,
2445 RelOptInfo *joinrel,
2446 JoinType jointype,
2447 JoinCostWorkspace *workspace,
2448 JoinPathExtraData *extra,
2449 Path *outer_path,
2450 Path *inner_path,
2451 List *restrict_clauses,
2452 List *pathkeys,
2453 Relids required_outer,
2454 List *mergeclauses,
2455 List *outersortkeys,
2456 List *innersortkeys)
2457 {
2458 MergePath *pathnode = makeNode(MergePath);
2459
2460 pathnode->jpath.path.pathtype = T_MergeJoin;
2461 pathnode->jpath.path.parent = joinrel;
2462 pathnode->jpath.path.pathtarget = joinrel->reltarget;
2463 pathnode->jpath.path.param_info =
2464 get_joinrel_parampathinfo(root,
2465 joinrel,
2466 outer_path,
2467 inner_path,
2468 extra->sjinfo,
2469 required_outer,
2470 &restrict_clauses);
2471 pathnode->jpath.path.parallel_aware = false;
2472 pathnode->jpath.path.parallel_safe = joinrel->consider_parallel &&
2473 outer_path->parallel_safe && inner_path->parallel_safe;
2474 /* This is a foolish way to estimate parallel_workers, but for now... */
2475 pathnode->jpath.path.parallel_workers = outer_path->parallel_workers;
2476 pathnode->jpath.path.pathkeys = pathkeys;
2477 pathnode->jpath.jointype = jointype;
2478 pathnode->jpath.inner_unique = extra->inner_unique;
2479 pathnode->jpath.outerjoinpath = outer_path;
2480 pathnode->jpath.innerjoinpath = inner_path;
2481 pathnode->jpath.joinrestrictinfo = restrict_clauses;
2482 pathnode->path_mergeclauses = mergeclauses;
2483 pathnode->outersortkeys = outersortkeys;
2484 pathnode->innersortkeys = innersortkeys;
2485 /* pathnode->skip_mark_restore will be set by final_cost_mergejoin */
2486 /* pathnode->materialize_inner will be set by final_cost_mergejoin */
2487
2488 final_cost_mergejoin(root, pathnode, workspace, extra);
2489
2490 return pathnode;
2491 }
2492
2493 /*
2494 * create_hashjoin_path
2495 * Creates a pathnode corresponding to a hash join between two relations.
2496 *
2497 * 'joinrel' is the join relation
2498 * 'jointype' is the type of join required
2499 * 'workspace' is the result from initial_cost_hashjoin
2500 * 'extra' contains various information about the join
2501 * 'outer_path' is the cheapest outer path
2502 * 'inner_path' is the cheapest inner path
2503 * 'parallel_hash' to select Parallel Hash of inner path (shared hash table)
2504 * 'restrict_clauses' are the RestrictInfo nodes to apply at the join
2505 * 'required_outer' is the set of required outer rels
2506 * 'hashclauses' are the RestrictInfo nodes to use as hash clauses
2507 * (this should be a subset of the restrict_clauses list)
2508 */
2509 HashPath *
create_hashjoin_path(PlannerInfo * root,RelOptInfo * joinrel,JoinType jointype,JoinCostWorkspace * workspace,JoinPathExtraData * extra,Path * outer_path,Path * inner_path,bool parallel_hash,List * restrict_clauses,Relids required_outer,List * hashclauses)2510 create_hashjoin_path(PlannerInfo *root,
2511 RelOptInfo *joinrel,
2512 JoinType jointype,
2513 JoinCostWorkspace *workspace,
2514 JoinPathExtraData *extra,
2515 Path *outer_path,
2516 Path *inner_path,
2517 bool parallel_hash,
2518 List *restrict_clauses,
2519 Relids required_outer,
2520 List *hashclauses)
2521 {
2522 HashPath *pathnode = makeNode(HashPath);
2523
2524 pathnode->jpath.path.pathtype = T_HashJoin;
2525 pathnode->jpath.path.parent = joinrel;
2526 pathnode->jpath.path.pathtarget = joinrel->reltarget;
2527 pathnode->jpath.path.param_info =
2528 get_joinrel_parampathinfo(root,
2529 joinrel,
2530 outer_path,
2531 inner_path,
2532 extra->sjinfo,
2533 required_outer,
2534 &restrict_clauses);
2535 pathnode->jpath.path.parallel_aware =
2536 joinrel->consider_parallel && parallel_hash;
2537 pathnode->jpath.path.parallel_safe = joinrel->consider_parallel &&
2538 outer_path->parallel_safe && inner_path->parallel_safe;
2539 /* This is a foolish way to estimate parallel_workers, but for now... */
2540 pathnode->jpath.path.parallel_workers = outer_path->parallel_workers;
2541
2542 /*
2543 * A hashjoin never has pathkeys, since its output ordering is
2544 * unpredictable due to possible batching. XXX If the inner relation is
2545 * small enough, we could instruct the executor that it must not batch,
2546 * and then we could assume that the output inherits the outer relation's
2547 * ordering, which might save a sort step. However there is considerable
2548 * downside if our estimate of the inner relation size is badly off. For
2549 * the moment we don't risk it. (Note also that if we wanted to take this
2550 * seriously, joinpath.c would have to consider many more paths for the
2551 * outer rel than it does now.)
2552 */
2553 pathnode->jpath.path.pathkeys = NIL;
2554 pathnode->jpath.jointype = jointype;
2555 pathnode->jpath.inner_unique = extra->inner_unique;
2556 pathnode->jpath.outerjoinpath = outer_path;
2557 pathnode->jpath.innerjoinpath = inner_path;
2558 pathnode->jpath.joinrestrictinfo = restrict_clauses;
2559 pathnode->path_hashclauses = hashclauses;
2560 /* final_cost_hashjoin will fill in pathnode->num_batches */
2561
2562 final_cost_hashjoin(root, pathnode, workspace, extra);
2563
2564 return pathnode;
2565 }
2566
2567 /*
2568 * create_projection_path
2569 * Creates a pathnode that represents performing a projection.
2570 *
2571 * 'rel' is the parent relation associated with the result
2572 * 'subpath' is the path representing the source of data
2573 * 'target' is the PathTarget to be computed
2574 */
2575 ProjectionPath *
create_projection_path(PlannerInfo * root,RelOptInfo * rel,Path * subpath,PathTarget * target)2576 create_projection_path(PlannerInfo *root,
2577 RelOptInfo *rel,
2578 Path *subpath,
2579 PathTarget *target)
2580 {
2581 ProjectionPath *pathnode = makeNode(ProjectionPath);
2582 PathTarget *oldtarget;
2583
2584 /*
2585 * We mustn't put a ProjectionPath directly above another; it's useless
2586 * and will confuse create_projection_plan. Rather than making sure all
2587 * callers handle that, let's implement it here, by stripping off any
2588 * ProjectionPath in what we're given. Given this rule, there won't be
2589 * more than one.
2590 */
2591 if (IsA(subpath, ProjectionPath))
2592 {
2593 ProjectionPath *subpp = (ProjectionPath *) subpath;
2594
2595 Assert(subpp->path.parent == rel);
2596 subpath = subpp->subpath;
2597 Assert(!IsA(subpath, ProjectionPath));
2598 }
2599
2600 pathnode->path.pathtype = T_Result;
2601 pathnode->path.parent = rel;
2602 pathnode->path.pathtarget = target;
2603 /* For now, assume we are above any joins, so no parameterization */
2604 pathnode->path.param_info = NULL;
2605 pathnode->path.parallel_aware = false;
2606 pathnode->path.parallel_safe = rel->consider_parallel &&
2607 subpath->parallel_safe &&
2608 is_parallel_safe(root, (Node *) target->exprs);
2609 pathnode->path.parallel_workers = subpath->parallel_workers;
2610 /* Projection does not change the sort order */
2611 pathnode->path.pathkeys = subpath->pathkeys;
2612
2613 pathnode->subpath = subpath;
2614
2615 /*
2616 * We might not need a separate Result node. If the input plan node type
2617 * can project, we can just tell it to project something else. Or, if it
2618 * can't project but the desired target has the same expression list as
2619 * what the input will produce anyway, we can still give it the desired
2620 * tlist (possibly changing its ressortgroupref labels, but nothing else).
2621 * Note: in the latter case, create_projection_plan has to recheck our
2622 * conclusion; see comments therein.
2623 */
2624 oldtarget = subpath->pathtarget;
2625 if (is_projection_capable_path(subpath) ||
2626 equal(oldtarget->exprs, target->exprs))
2627 {
2628 /* No separate Result node needed */
2629 pathnode->dummypp = true;
2630
2631 /*
2632 * Set cost of plan as subpath's cost, adjusted for tlist replacement.
2633 */
2634 pathnode->path.rows = subpath->rows;
2635 pathnode->path.startup_cost = subpath->startup_cost +
2636 (target->cost.startup - oldtarget->cost.startup);
2637 pathnode->path.total_cost = subpath->total_cost +
2638 (target->cost.startup - oldtarget->cost.startup) +
2639 (target->cost.per_tuple - oldtarget->cost.per_tuple) * subpath->rows;
2640 }
2641 else
2642 {
2643 /* We really do need the Result node */
2644 pathnode->dummypp = false;
2645
2646 /*
2647 * The Result node's cost is cpu_tuple_cost per row, plus the cost of
2648 * evaluating the tlist. There is no qual to worry about.
2649 */
2650 pathnode->path.rows = subpath->rows;
2651 pathnode->path.startup_cost = subpath->startup_cost +
2652 target->cost.startup;
2653 pathnode->path.total_cost = subpath->total_cost +
2654 target->cost.startup +
2655 (cpu_tuple_cost + target->cost.per_tuple) * subpath->rows;
2656 }
2657
2658 return pathnode;
2659 }
2660
2661 /*
2662 * apply_projection_to_path
2663 * Add a projection step, or just apply the target directly to given path.
2664 *
2665 * This has the same net effect as create_projection_path(), except that if
2666 * a separate Result plan node isn't needed, we just replace the given path's
2667 * pathtarget with the desired one. This must be used only when the caller
2668 * knows that the given path isn't referenced elsewhere and so can be modified
2669 * in-place.
2670 *
2671 * If the input path is a GatherPath or GatherMergePath, we try to push the
2672 * new target down to its input as well; this is a yet more invasive
2673 * modification of the input path, which create_projection_path() can't do.
2674 *
2675 * Note that we mustn't change the source path's parent link; so when it is
2676 * add_path'd to "rel" things will be a bit inconsistent. So far that has
2677 * not caused any trouble.
2678 *
2679 * 'rel' is the parent relation associated with the result
2680 * 'path' is the path representing the source of data
2681 * 'target' is the PathTarget to be computed
2682 */
2683 Path *
apply_projection_to_path(PlannerInfo * root,RelOptInfo * rel,Path * path,PathTarget * target)2684 apply_projection_to_path(PlannerInfo *root,
2685 RelOptInfo *rel,
2686 Path *path,
2687 PathTarget *target)
2688 {
2689 QualCost oldcost;
2690
2691 /*
2692 * If given path can't project, we might need a Result node, so make a
2693 * separate ProjectionPath.
2694 */
2695 if (!is_projection_capable_path(path))
2696 return (Path *) create_projection_path(root, rel, path, target);
2697
2698 /*
2699 * We can just jam the desired tlist into the existing path, being sure to
2700 * update its cost estimates appropriately.
2701 */
2702 oldcost = path->pathtarget->cost;
2703 path->pathtarget = target;
2704
2705 path->startup_cost += target->cost.startup - oldcost.startup;
2706 path->total_cost += target->cost.startup - oldcost.startup +
2707 (target->cost.per_tuple - oldcost.per_tuple) * path->rows;
2708
2709 /*
2710 * If the path happens to be a Gather or GatherMerge path, we'd like to
2711 * arrange for the subpath to return the required target list so that
2712 * workers can help project. But if there is something that is not
2713 * parallel-safe in the target expressions, then we can't.
2714 */
2715 if ((IsA(path, GatherPath) ||IsA(path, GatherMergePath)) &&
2716 is_parallel_safe(root, (Node *) target->exprs))
2717 {
2718 /*
2719 * We always use create_projection_path here, even if the subpath is
2720 * projection-capable, so as to avoid modifying the subpath in place.
2721 * It seems unlikely at present that there could be any other
2722 * references to the subpath, but better safe than sorry.
2723 *
2724 * Note that we don't change the parallel path's cost estimates; it
2725 * might be appropriate to do so, to reflect the fact that the bulk of
2726 * the target evaluation will happen in workers.
2727 */
2728 if (IsA(path, GatherPath))
2729 {
2730 GatherPath *gpath = (GatherPath *) path;
2731
2732 gpath->subpath = (Path *)
2733 create_projection_path(root,
2734 gpath->subpath->parent,
2735 gpath->subpath,
2736 target);
2737 }
2738 else
2739 {
2740 GatherMergePath *gmpath = (GatherMergePath *) path;
2741
2742 gmpath->subpath = (Path *)
2743 create_projection_path(root,
2744 gmpath->subpath->parent,
2745 gmpath->subpath,
2746 target);
2747 }
2748 }
2749 else if (path->parallel_safe &&
2750 !is_parallel_safe(root, (Node *) target->exprs))
2751 {
2752 /*
2753 * We're inserting a parallel-restricted target list into a path
2754 * currently marked parallel-safe, so we have to mark it as no longer
2755 * safe.
2756 */
2757 path->parallel_safe = false;
2758 }
2759
2760 return path;
2761 }
2762
2763 /*
2764 * create_set_projection_path
2765 * Creates a pathnode that represents performing a projection that
2766 * includes set-returning functions.
2767 *
2768 * 'rel' is the parent relation associated with the result
2769 * 'subpath' is the path representing the source of data
2770 * 'target' is the PathTarget to be computed
2771 */
2772 ProjectSetPath *
create_set_projection_path(PlannerInfo * root,RelOptInfo * rel,Path * subpath,PathTarget * target)2773 create_set_projection_path(PlannerInfo *root,
2774 RelOptInfo *rel,
2775 Path *subpath,
2776 PathTarget *target)
2777 {
2778 ProjectSetPath *pathnode = makeNode(ProjectSetPath);
2779 double tlist_rows;
2780 ListCell *lc;
2781
2782 pathnode->path.pathtype = T_ProjectSet;
2783 pathnode->path.parent = rel;
2784 pathnode->path.pathtarget = target;
2785 /* For now, assume we are above any joins, so no parameterization */
2786 pathnode->path.param_info = NULL;
2787 pathnode->path.parallel_aware = false;
2788 pathnode->path.parallel_safe = rel->consider_parallel &&
2789 subpath->parallel_safe &&
2790 is_parallel_safe(root, (Node *) target->exprs);
2791 pathnode->path.parallel_workers = subpath->parallel_workers;
2792 /* Projection does not change the sort order XXX? */
2793 pathnode->path.pathkeys = subpath->pathkeys;
2794
2795 pathnode->subpath = subpath;
2796
2797 /*
2798 * Estimate number of rows produced by SRFs for each row of input; if
2799 * there's more than one in this node, use the maximum.
2800 */
2801 tlist_rows = 1;
2802 foreach(lc, target->exprs)
2803 {
2804 Node *node = (Node *) lfirst(lc);
2805 double itemrows;
2806
2807 itemrows = expression_returns_set_rows(root, node);
2808 if (tlist_rows < itemrows)
2809 tlist_rows = itemrows;
2810 }
2811
2812 /*
2813 * In addition to the cost of evaluating the tlist, charge cpu_tuple_cost
2814 * per input row, and half of cpu_tuple_cost for each added output row.
2815 * This is slightly bizarre maybe, but it's what 9.6 did; we may revisit
2816 * this estimate later.
2817 */
2818 pathnode->path.rows = subpath->rows * tlist_rows;
2819 pathnode->path.startup_cost = subpath->startup_cost +
2820 target->cost.startup;
2821 pathnode->path.total_cost = subpath->total_cost +
2822 target->cost.startup +
2823 (cpu_tuple_cost + target->cost.per_tuple) * subpath->rows +
2824 (pathnode->path.rows - subpath->rows) * cpu_tuple_cost / 2;
2825
2826 return pathnode;
2827 }
2828
2829 /*
2830 * create_sort_path
2831 * Creates a pathnode that represents performing an explicit sort.
2832 *
2833 * 'rel' is the parent relation associated with the result
2834 * 'subpath' is the path representing the source of data
2835 * 'pathkeys' represents the desired sort order
2836 * 'limit_tuples' is the estimated bound on the number of output tuples,
2837 * or -1 if no LIMIT or couldn't estimate
2838 */
2839 SortPath *
create_sort_path(PlannerInfo * root,RelOptInfo * rel,Path * subpath,List * pathkeys,double limit_tuples)2840 create_sort_path(PlannerInfo *root,
2841 RelOptInfo *rel,
2842 Path *subpath,
2843 List *pathkeys,
2844 double limit_tuples)
2845 {
2846 SortPath *pathnode = makeNode(SortPath);
2847
2848 pathnode->path.pathtype = T_Sort;
2849 pathnode->path.parent = rel;
2850 /* Sort doesn't project, so use source path's pathtarget */
2851 pathnode->path.pathtarget = subpath->pathtarget;
2852 /* For now, assume we are above any joins, so no parameterization */
2853 pathnode->path.param_info = NULL;
2854 pathnode->path.parallel_aware = false;
2855 pathnode->path.parallel_safe = rel->consider_parallel &&
2856 subpath->parallel_safe;
2857 pathnode->path.parallel_workers = subpath->parallel_workers;
2858 pathnode->path.pathkeys = pathkeys;
2859
2860 pathnode->subpath = subpath;
2861
2862 cost_sort(&pathnode->path, root, pathkeys,
2863 subpath->total_cost,
2864 subpath->rows,
2865 subpath->pathtarget->width,
2866 0.0, /* XXX comparison_cost shouldn't be 0? */
2867 work_mem, limit_tuples);
2868
2869 return pathnode;
2870 }
2871
2872 /*
2873 * create_group_path
2874 * Creates a pathnode that represents performing grouping of presorted input
2875 *
2876 * 'rel' is the parent relation associated with the result
2877 * 'subpath' is the path representing the source of data
2878 * 'target' is the PathTarget to be computed
2879 * 'groupClause' is a list of SortGroupClause's representing the grouping
2880 * 'qual' is the HAVING quals if any
2881 * 'numGroups' is the estimated number of groups
2882 */
2883 GroupPath *
create_group_path(PlannerInfo * root,RelOptInfo * rel,Path * subpath,List * groupClause,List * qual,double numGroups)2884 create_group_path(PlannerInfo *root,
2885 RelOptInfo *rel,
2886 Path *subpath,
2887 List *groupClause,
2888 List *qual,
2889 double numGroups)
2890 {
2891 GroupPath *pathnode = makeNode(GroupPath);
2892 PathTarget *target = rel->reltarget;
2893
2894 pathnode->path.pathtype = T_Group;
2895 pathnode->path.parent = rel;
2896 pathnode->path.pathtarget = target;
2897 /* For now, assume we are above any joins, so no parameterization */
2898 pathnode->path.param_info = NULL;
2899 pathnode->path.parallel_aware = false;
2900 pathnode->path.parallel_safe = rel->consider_parallel &&
2901 subpath->parallel_safe;
2902 pathnode->path.parallel_workers = subpath->parallel_workers;
2903 /* Group doesn't change sort ordering */
2904 pathnode->path.pathkeys = subpath->pathkeys;
2905
2906 pathnode->subpath = subpath;
2907
2908 pathnode->groupClause = groupClause;
2909 pathnode->qual = qual;
2910
2911 cost_group(&pathnode->path, root,
2912 list_length(groupClause),
2913 numGroups,
2914 qual,
2915 subpath->startup_cost, subpath->total_cost,
2916 subpath->rows);
2917
2918 /* add tlist eval cost for each output row */
2919 pathnode->path.startup_cost += target->cost.startup;
2920 pathnode->path.total_cost += target->cost.startup +
2921 target->cost.per_tuple * pathnode->path.rows;
2922
2923 return pathnode;
2924 }
2925
2926 /*
2927 * create_upper_unique_path
2928 * Creates a pathnode that represents performing an explicit Unique step
2929 * on presorted input.
2930 *
2931 * This produces a Unique plan node, but the use-case is so different from
2932 * create_unique_path that it doesn't seem worth trying to merge the two.
2933 *
2934 * 'rel' is the parent relation associated with the result
2935 * 'subpath' is the path representing the source of data
2936 * 'numCols' is the number of grouping columns
2937 * 'numGroups' is the estimated number of groups
2938 *
2939 * The input path must be sorted on the grouping columns, plus possibly
2940 * additional columns; so the first numCols pathkeys are the grouping columns
2941 */
2942 UpperUniquePath *
create_upper_unique_path(PlannerInfo * root,RelOptInfo * rel,Path * subpath,int numCols,double numGroups)2943 create_upper_unique_path(PlannerInfo *root,
2944 RelOptInfo *rel,
2945 Path *subpath,
2946 int numCols,
2947 double numGroups)
2948 {
2949 UpperUniquePath *pathnode = makeNode(UpperUniquePath);
2950
2951 pathnode->path.pathtype = T_Unique;
2952 pathnode->path.parent = rel;
2953 /* Unique doesn't project, so use source path's pathtarget */
2954 pathnode->path.pathtarget = subpath->pathtarget;
2955 /* For now, assume we are above any joins, so no parameterization */
2956 pathnode->path.param_info = NULL;
2957 pathnode->path.parallel_aware = false;
2958 pathnode->path.parallel_safe = rel->consider_parallel &&
2959 subpath->parallel_safe;
2960 pathnode->path.parallel_workers = subpath->parallel_workers;
2961 /* Unique doesn't change the input ordering */
2962 pathnode->path.pathkeys = subpath->pathkeys;
2963
2964 pathnode->subpath = subpath;
2965 pathnode->numkeys = numCols;
2966
2967 /*
2968 * Charge one cpu_operator_cost per comparison per input tuple. We assume
2969 * all columns get compared at most of the tuples. (XXX probably this is
2970 * an overestimate.)
2971 */
2972 pathnode->path.startup_cost = subpath->startup_cost;
2973 pathnode->path.total_cost = subpath->total_cost +
2974 cpu_operator_cost * subpath->rows * numCols;
2975 pathnode->path.rows = numGroups;
2976
2977 return pathnode;
2978 }
2979
2980 /*
2981 * create_agg_path
2982 * Creates a pathnode that represents performing aggregation/grouping
2983 *
2984 * 'rel' is the parent relation associated with the result
2985 * 'subpath' is the path representing the source of data
2986 * 'target' is the PathTarget to be computed
2987 * 'aggstrategy' is the Agg node's basic implementation strategy
2988 * 'aggsplit' is the Agg node's aggregate-splitting mode
2989 * 'groupClause' is a list of SortGroupClause's representing the grouping
2990 * 'qual' is the HAVING quals if any
2991 * 'aggcosts' contains cost info about the aggregate functions to be computed
2992 * 'numGroups' is the estimated number of groups (1 if not grouping)
2993 */
2994 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)2995 create_agg_path(PlannerInfo *root,
2996 RelOptInfo *rel,
2997 Path *subpath,
2998 PathTarget *target,
2999 AggStrategy aggstrategy,
3000 AggSplit aggsplit,
3001 List *groupClause,
3002 List *qual,
3003 const AggClauseCosts *aggcosts,
3004 double numGroups)
3005 {
3006 AggPath *pathnode = makeNode(AggPath);
3007
3008 pathnode->path.pathtype = T_Agg;
3009 pathnode->path.parent = rel;
3010 pathnode->path.pathtarget = target;
3011 /* For now, assume we are above any joins, so no parameterization */
3012 pathnode->path.param_info = NULL;
3013 pathnode->path.parallel_aware = false;
3014 pathnode->path.parallel_safe = rel->consider_parallel &&
3015 subpath->parallel_safe;
3016 pathnode->path.parallel_workers = subpath->parallel_workers;
3017 if (aggstrategy == AGG_SORTED)
3018 pathnode->path.pathkeys = subpath->pathkeys; /* preserves order */
3019 else
3020 pathnode->path.pathkeys = NIL; /* output is unordered */
3021 pathnode->subpath = subpath;
3022
3023 pathnode->aggstrategy = aggstrategy;
3024 pathnode->aggsplit = aggsplit;
3025 pathnode->numGroups = numGroups;
3026 pathnode->groupClause = groupClause;
3027 pathnode->qual = qual;
3028
3029 cost_agg(&pathnode->path, root,
3030 aggstrategy, aggcosts,
3031 list_length(groupClause), numGroups,
3032 qual,
3033 subpath->startup_cost, subpath->total_cost,
3034 subpath->rows);
3035
3036 /* add tlist eval cost for each output row */
3037 pathnode->path.startup_cost += target->cost.startup;
3038 pathnode->path.total_cost += target->cost.startup +
3039 target->cost.per_tuple * pathnode->path.rows;
3040
3041 return pathnode;
3042 }
3043
3044 /*
3045 * create_groupingsets_path
3046 * Creates a pathnode that represents performing GROUPING SETS aggregation
3047 *
3048 * GroupingSetsPath represents sorted grouping with one or more grouping sets.
3049 * The input path's result must be sorted to match the last entry in
3050 * rollup_groupclauses.
3051 *
3052 * 'rel' is the parent relation associated with the result
3053 * 'subpath' is the path representing the source of data
3054 * 'target' is the PathTarget to be computed
3055 * 'having_qual' is the HAVING quals if any
3056 * 'rollups' is a list of RollupData nodes
3057 * 'agg_costs' contains cost info about the aggregate functions to be computed
3058 * 'numGroups' is the estimated total number of groups
3059 */
3060 GroupingSetsPath *
create_groupingsets_path(PlannerInfo * root,RelOptInfo * rel,Path * subpath,List * having_qual,AggStrategy aggstrategy,List * rollups,const AggClauseCosts * agg_costs,double numGroups)3061 create_groupingsets_path(PlannerInfo *root,
3062 RelOptInfo *rel,
3063 Path *subpath,
3064 List *having_qual,
3065 AggStrategy aggstrategy,
3066 List *rollups,
3067 const AggClauseCosts *agg_costs,
3068 double numGroups)
3069 {
3070 GroupingSetsPath *pathnode = makeNode(GroupingSetsPath);
3071 PathTarget *target = rel->reltarget;
3072 ListCell *lc;
3073 bool is_first = true;
3074 bool is_first_sort = true;
3075
3076 /* The topmost generated Plan node will be an Agg */
3077 pathnode->path.pathtype = T_Agg;
3078 pathnode->path.parent = rel;
3079 pathnode->path.pathtarget = target;
3080 pathnode->path.param_info = subpath->param_info;
3081 pathnode->path.parallel_aware = false;
3082 pathnode->path.parallel_safe = rel->consider_parallel &&
3083 subpath->parallel_safe;
3084 pathnode->path.parallel_workers = subpath->parallel_workers;
3085 pathnode->subpath = subpath;
3086
3087 /*
3088 * Simplify callers by downgrading AGG_SORTED to AGG_PLAIN, and AGG_MIXED
3089 * to AGG_HASHED, here if possible.
3090 */
3091 if (aggstrategy == AGG_SORTED &&
3092 list_length(rollups) == 1 &&
3093 ((RollupData *) linitial(rollups))->groupClause == NIL)
3094 aggstrategy = AGG_PLAIN;
3095
3096 if (aggstrategy == AGG_MIXED &&
3097 list_length(rollups) == 1)
3098 aggstrategy = AGG_HASHED;
3099
3100 /*
3101 * Output will be in sorted order by group_pathkeys if, and only if, there
3102 * is a single rollup operation on a non-empty list of grouping
3103 * expressions.
3104 */
3105 if (aggstrategy == AGG_SORTED && list_length(rollups) == 1)
3106 pathnode->path.pathkeys = root->group_pathkeys;
3107 else
3108 pathnode->path.pathkeys = NIL;
3109
3110 pathnode->aggstrategy = aggstrategy;
3111 pathnode->rollups = rollups;
3112 pathnode->qual = having_qual;
3113
3114 Assert(rollups != NIL);
3115 Assert(aggstrategy != AGG_PLAIN || list_length(rollups) == 1);
3116 Assert(aggstrategy != AGG_MIXED || list_length(rollups) > 1);
3117
3118 foreach(lc, rollups)
3119 {
3120 RollupData *rollup = lfirst(lc);
3121 List *gsets = rollup->gsets;
3122 int numGroupCols = list_length(linitial(gsets));
3123
3124 /*
3125 * In AGG_SORTED or AGG_PLAIN mode, the first rollup takes the
3126 * (already-sorted) input, and following ones do their own sort.
3127 *
3128 * In AGG_HASHED mode, there is one rollup for each grouping set.
3129 *
3130 * In AGG_MIXED mode, the first rollups are hashed, the first
3131 * non-hashed one takes the (already-sorted) input, and following ones
3132 * do their own sort.
3133 */
3134 if (is_first)
3135 {
3136 cost_agg(&pathnode->path, root,
3137 aggstrategy,
3138 agg_costs,
3139 numGroupCols,
3140 rollup->numGroups,
3141 having_qual,
3142 subpath->startup_cost,
3143 subpath->total_cost,
3144 subpath->rows);
3145 is_first = false;
3146 if (!rollup->is_hashed)
3147 is_first_sort = false;
3148 }
3149 else
3150 {
3151 Path sort_path; /* dummy for result of cost_sort */
3152 Path agg_path; /* dummy for result of cost_agg */
3153
3154 if (rollup->is_hashed || is_first_sort)
3155 {
3156 /*
3157 * Account for cost of aggregation, but don't charge input
3158 * cost again
3159 */
3160 cost_agg(&agg_path, root,
3161 rollup->is_hashed ? AGG_HASHED : AGG_SORTED,
3162 agg_costs,
3163 numGroupCols,
3164 rollup->numGroups,
3165 having_qual,
3166 0.0, 0.0,
3167 subpath->rows);
3168 if (!rollup->is_hashed)
3169 is_first_sort = false;
3170 }
3171 else
3172 {
3173 /* Account for cost of sort, but don't charge input cost again */
3174 cost_sort(&sort_path, root, NIL,
3175 0.0,
3176 subpath->rows,
3177 subpath->pathtarget->width,
3178 0.0,
3179 work_mem,
3180 -1.0);
3181
3182 /* Account for cost of aggregation */
3183
3184 cost_agg(&agg_path, root,
3185 AGG_SORTED,
3186 agg_costs,
3187 numGroupCols,
3188 rollup->numGroups,
3189 having_qual,
3190 sort_path.startup_cost,
3191 sort_path.total_cost,
3192 sort_path.rows);
3193 }
3194
3195 pathnode->path.total_cost += agg_path.total_cost;
3196 pathnode->path.rows += agg_path.rows;
3197 }
3198 }
3199
3200 /* add tlist eval cost for each output row */
3201 pathnode->path.startup_cost += target->cost.startup;
3202 pathnode->path.total_cost += target->cost.startup +
3203 target->cost.per_tuple * pathnode->path.rows;
3204
3205 return pathnode;
3206 }
3207
3208 /*
3209 * create_minmaxagg_path
3210 * Creates a pathnode that represents computation of MIN/MAX aggregates
3211 *
3212 * 'rel' is the parent relation associated with the result
3213 * 'target' is the PathTarget to be computed
3214 * 'mmaggregates' is a list of MinMaxAggInfo structs
3215 * 'quals' is the HAVING quals if any
3216 */
3217 MinMaxAggPath *
create_minmaxagg_path(PlannerInfo * root,RelOptInfo * rel,PathTarget * target,List * mmaggregates,List * quals)3218 create_minmaxagg_path(PlannerInfo *root,
3219 RelOptInfo *rel,
3220 PathTarget *target,
3221 List *mmaggregates,
3222 List *quals)
3223 {
3224 MinMaxAggPath *pathnode = makeNode(MinMaxAggPath);
3225 Cost initplan_cost;
3226 ListCell *lc;
3227
3228 /* The topmost generated Plan node will be a Result */
3229 pathnode->path.pathtype = T_Result;
3230 pathnode->path.parent = rel;
3231 pathnode->path.pathtarget = target;
3232 /* For now, assume we are above any joins, so no parameterization */
3233 pathnode->path.param_info = NULL;
3234 pathnode->path.parallel_aware = false;
3235 /* A MinMaxAggPath implies use of subplans, so cannot be parallel-safe */
3236 pathnode->path.parallel_safe = false;
3237 pathnode->path.parallel_workers = 0;
3238 /* Result is one unordered row */
3239 pathnode->path.rows = 1;
3240 pathnode->path.pathkeys = NIL;
3241
3242 pathnode->mmaggregates = mmaggregates;
3243 pathnode->quals = quals;
3244
3245 /* Calculate cost of all the initplans ... */
3246 initplan_cost = 0;
3247 foreach(lc, mmaggregates)
3248 {
3249 MinMaxAggInfo *mminfo = (MinMaxAggInfo *) lfirst(lc);
3250
3251 initplan_cost += mminfo->pathcost;
3252 }
3253
3254 /* add tlist eval cost for each output row, plus cpu_tuple_cost */
3255 pathnode->path.startup_cost = initplan_cost + target->cost.startup;
3256 pathnode->path.total_cost = initplan_cost + target->cost.startup +
3257 target->cost.per_tuple + cpu_tuple_cost;
3258
3259 /*
3260 * Add cost of qual, if any --- but we ignore its selectivity, since our
3261 * rowcount estimate should be 1 no matter what the qual is.
3262 */
3263 if (quals)
3264 {
3265 QualCost qual_cost;
3266
3267 cost_qual_eval(&qual_cost, quals, root);
3268 pathnode->path.startup_cost += qual_cost.startup;
3269 pathnode->path.total_cost += qual_cost.startup + qual_cost.per_tuple;
3270 }
3271
3272 return pathnode;
3273 }
3274
3275 /*
3276 * create_windowagg_path
3277 * Creates a pathnode that represents computation of window functions
3278 *
3279 * 'rel' is the parent relation associated with the result
3280 * 'subpath' is the path representing the source of data
3281 * 'target' is the PathTarget to be computed
3282 * 'windowFuncs' is a list of WindowFunc structs
3283 * 'winclause' is a WindowClause that is common to all the WindowFuncs
3284 *
3285 * The input must be sorted according to the WindowClause's PARTITION keys
3286 * plus ORDER BY keys.
3287 */
3288 WindowAggPath *
create_windowagg_path(PlannerInfo * root,RelOptInfo * rel,Path * subpath,PathTarget * target,List * windowFuncs,WindowClause * winclause)3289 create_windowagg_path(PlannerInfo *root,
3290 RelOptInfo *rel,
3291 Path *subpath,
3292 PathTarget *target,
3293 List *windowFuncs,
3294 WindowClause *winclause)
3295 {
3296 WindowAggPath *pathnode = makeNode(WindowAggPath);
3297
3298 pathnode->path.pathtype = T_WindowAgg;
3299 pathnode->path.parent = rel;
3300 pathnode->path.pathtarget = target;
3301 /* For now, assume we are above any joins, so no parameterization */
3302 pathnode->path.param_info = NULL;
3303 pathnode->path.parallel_aware = false;
3304 pathnode->path.parallel_safe = rel->consider_parallel &&
3305 subpath->parallel_safe;
3306 pathnode->path.parallel_workers = subpath->parallel_workers;
3307 /* WindowAgg preserves the input sort order */
3308 pathnode->path.pathkeys = subpath->pathkeys;
3309
3310 pathnode->subpath = subpath;
3311 pathnode->winclause = winclause;
3312
3313 /*
3314 * For costing purposes, assume that there are no redundant partitioning
3315 * or ordering columns; it's not worth the trouble to deal with that
3316 * corner case here. So we just pass the unmodified list lengths to
3317 * cost_windowagg.
3318 */
3319 cost_windowagg(&pathnode->path, root,
3320 windowFuncs,
3321 list_length(winclause->partitionClause),
3322 list_length(winclause->orderClause),
3323 subpath->startup_cost,
3324 subpath->total_cost,
3325 subpath->rows);
3326
3327 /* add tlist eval cost for each output row */
3328 pathnode->path.startup_cost += target->cost.startup;
3329 pathnode->path.total_cost += target->cost.startup +
3330 target->cost.per_tuple * pathnode->path.rows;
3331
3332 return pathnode;
3333 }
3334
3335 /*
3336 * create_setop_path
3337 * Creates a pathnode that represents computation of INTERSECT or EXCEPT
3338 *
3339 * 'rel' is the parent relation associated with the result
3340 * 'subpath' is the path representing the source of data
3341 * 'cmd' is the specific semantics (INTERSECT or EXCEPT, with/without ALL)
3342 * 'strategy' is the implementation strategy (sorted or hashed)
3343 * 'distinctList' is a list of SortGroupClause's representing the grouping
3344 * 'flagColIdx' is the column number where the flag column will be, if any
3345 * 'firstFlag' is the flag value for the first input relation when hashing;
3346 * or -1 when sorting
3347 * 'numGroups' is the estimated number of distinct groups
3348 * 'outputRows' is the estimated number of output rows
3349 */
3350 SetOpPath *
create_setop_path(PlannerInfo * root,RelOptInfo * rel,Path * subpath,SetOpCmd cmd,SetOpStrategy strategy,List * distinctList,AttrNumber flagColIdx,int firstFlag,double numGroups,double outputRows)3351 create_setop_path(PlannerInfo *root,
3352 RelOptInfo *rel,
3353 Path *subpath,
3354 SetOpCmd cmd,
3355 SetOpStrategy strategy,
3356 List *distinctList,
3357 AttrNumber flagColIdx,
3358 int firstFlag,
3359 double numGroups,
3360 double outputRows)
3361 {
3362 SetOpPath *pathnode = makeNode(SetOpPath);
3363
3364 pathnode->path.pathtype = T_SetOp;
3365 pathnode->path.parent = rel;
3366 /* SetOp doesn't project, so use source path's pathtarget */
3367 pathnode->path.pathtarget = subpath->pathtarget;
3368 /* For now, assume we are above any joins, so no parameterization */
3369 pathnode->path.param_info = NULL;
3370 pathnode->path.parallel_aware = false;
3371 pathnode->path.parallel_safe = rel->consider_parallel &&
3372 subpath->parallel_safe;
3373 pathnode->path.parallel_workers = subpath->parallel_workers;
3374 /* SetOp preserves the input sort order if in sort mode */
3375 pathnode->path.pathkeys =
3376 (strategy == SETOP_SORTED) ? subpath->pathkeys : NIL;
3377
3378 pathnode->subpath = subpath;
3379 pathnode->cmd = cmd;
3380 pathnode->strategy = strategy;
3381 pathnode->distinctList = distinctList;
3382 pathnode->flagColIdx = flagColIdx;
3383 pathnode->firstFlag = firstFlag;
3384 pathnode->numGroups = numGroups;
3385
3386 /*
3387 * Charge one cpu_operator_cost per comparison per input tuple. We assume
3388 * all columns get compared at most of the tuples.
3389 */
3390 pathnode->path.startup_cost = subpath->startup_cost;
3391 pathnode->path.total_cost = subpath->total_cost +
3392 cpu_operator_cost * subpath->rows * list_length(distinctList);
3393 pathnode->path.rows = outputRows;
3394
3395 return pathnode;
3396 }
3397
3398 /*
3399 * create_recursiveunion_path
3400 * Creates a pathnode that represents a recursive UNION node
3401 *
3402 * 'rel' is the parent relation associated with the result
3403 * 'leftpath' is the source of data for the non-recursive term
3404 * 'rightpath' is the source of data for the recursive term
3405 * 'target' is the PathTarget to be computed
3406 * 'distinctList' is a list of SortGroupClause's representing the grouping
3407 * 'wtParam' is the ID of Param representing work table
3408 * 'numGroups' is the estimated number of groups
3409 *
3410 * For recursive UNION ALL, distinctList is empty and numGroups is zero
3411 */
3412 RecursiveUnionPath *
create_recursiveunion_path(PlannerInfo * root,RelOptInfo * rel,Path * leftpath,Path * rightpath,PathTarget * target,List * distinctList,int wtParam,double numGroups)3413 create_recursiveunion_path(PlannerInfo *root,
3414 RelOptInfo *rel,
3415 Path *leftpath,
3416 Path *rightpath,
3417 PathTarget *target,
3418 List *distinctList,
3419 int wtParam,
3420 double numGroups)
3421 {
3422 RecursiveUnionPath *pathnode = makeNode(RecursiveUnionPath);
3423
3424 pathnode->path.pathtype = T_RecursiveUnion;
3425 pathnode->path.parent = rel;
3426 pathnode->path.pathtarget = target;
3427 /* For now, assume we are above any joins, so no parameterization */
3428 pathnode->path.param_info = NULL;
3429 pathnode->path.parallel_aware = false;
3430 pathnode->path.parallel_safe = rel->consider_parallel &&
3431 leftpath->parallel_safe && rightpath->parallel_safe;
3432 /* Foolish, but we'll do it like joins for now: */
3433 pathnode->path.parallel_workers = leftpath->parallel_workers;
3434 /* RecursiveUnion result is always unsorted */
3435 pathnode->path.pathkeys = NIL;
3436
3437 pathnode->leftpath = leftpath;
3438 pathnode->rightpath = rightpath;
3439 pathnode->distinctList = distinctList;
3440 pathnode->wtParam = wtParam;
3441 pathnode->numGroups = numGroups;
3442
3443 cost_recursive_union(&pathnode->path, leftpath, rightpath);
3444
3445 return pathnode;
3446 }
3447
3448 /*
3449 * create_lockrows_path
3450 * Creates a pathnode that represents acquiring row locks
3451 *
3452 * 'rel' is the parent relation associated with the result
3453 * 'subpath' is the path representing the source of data
3454 * 'rowMarks' is a list of PlanRowMark's
3455 * 'epqParam' is the ID of Param for EvalPlanQual re-eval
3456 */
3457 LockRowsPath *
create_lockrows_path(PlannerInfo * root,RelOptInfo * rel,Path * subpath,List * rowMarks,int epqParam)3458 create_lockrows_path(PlannerInfo *root, RelOptInfo *rel,
3459 Path *subpath, List *rowMarks, int epqParam)
3460 {
3461 LockRowsPath *pathnode = makeNode(LockRowsPath);
3462
3463 pathnode->path.pathtype = T_LockRows;
3464 pathnode->path.parent = rel;
3465 /* LockRows doesn't project, so use source path's pathtarget */
3466 pathnode->path.pathtarget = subpath->pathtarget;
3467 /* For now, assume we are above any joins, so no parameterization */
3468 pathnode->path.param_info = NULL;
3469 pathnode->path.parallel_aware = false;
3470 pathnode->path.parallel_safe = false;
3471 pathnode->path.parallel_workers = 0;
3472 pathnode->path.rows = subpath->rows;
3473
3474 /*
3475 * The result cannot be assumed sorted, since locking might cause the sort
3476 * key columns to be replaced with new values.
3477 */
3478 pathnode->path.pathkeys = NIL;
3479
3480 pathnode->subpath = subpath;
3481 pathnode->rowMarks = rowMarks;
3482 pathnode->epqParam = epqParam;
3483
3484 /*
3485 * We should charge something extra for the costs of row locking and
3486 * possible refetches, but it's hard to say how much. For now, use
3487 * cpu_tuple_cost per row.
3488 */
3489 pathnode->path.startup_cost = subpath->startup_cost;
3490 pathnode->path.total_cost = subpath->total_cost +
3491 cpu_tuple_cost * subpath->rows;
3492
3493 return pathnode;
3494 }
3495
3496 /*
3497 * create_modifytable_path
3498 * Creates a pathnode that represents performing INSERT/UPDATE/DELETE mods
3499 *
3500 * 'rel' is the parent relation associated with the result
3501 * 'operation' is the operation type
3502 * 'canSetTag' is true if we set the command tag/es_processed
3503 * 'nominalRelation' is the parent RT index for use of EXPLAIN
3504 * 'rootRelation' is the partitioned table root RT index, or 0 if none
3505 * 'partColsUpdated' is true if any partitioning columns are being updated,
3506 * either from the target relation or a descendent partitioned table.
3507 * 'resultRelations' is an integer list of actual RT indexes of target rel(s)
3508 * 'subpaths' is a list of Path(s) producing source data (one per rel)
3509 * 'subroots' is a list of PlannerInfo structs (one per rel)
3510 * 'withCheckOptionLists' is a list of WCO lists (one per rel)
3511 * 'returningLists' is a list of RETURNING tlists (one per rel)
3512 * 'rowMarks' is a list of PlanRowMarks (non-locking only)
3513 * 'onconflict' is the ON CONFLICT clause, or NULL
3514 * 'epqParam' is the ID of Param for EvalPlanQual re-eval
3515 */
3516 ModifyTablePath *
create_modifytable_path(PlannerInfo * root,RelOptInfo * rel,CmdType operation,bool canSetTag,Index nominalRelation,Index rootRelation,bool partColsUpdated,List * resultRelations,List * subpaths,List * subroots,List * withCheckOptionLists,List * returningLists,List * rowMarks,OnConflictExpr * onconflict,int epqParam)3517 create_modifytable_path(PlannerInfo *root, RelOptInfo *rel,
3518 CmdType operation, bool canSetTag,
3519 Index nominalRelation, Index rootRelation,
3520 bool partColsUpdated,
3521 List *resultRelations, List *subpaths,
3522 List *subroots,
3523 List *withCheckOptionLists, List *returningLists,
3524 List *rowMarks, OnConflictExpr *onconflict,
3525 int epqParam)
3526 {
3527 ModifyTablePath *pathnode = makeNode(ModifyTablePath);
3528 double total_size;
3529 ListCell *lc;
3530
3531 Assert(list_length(resultRelations) == list_length(subpaths));
3532 Assert(list_length(resultRelations) == list_length(subroots));
3533 Assert(withCheckOptionLists == NIL ||
3534 list_length(resultRelations) == list_length(withCheckOptionLists));
3535 Assert(returningLists == NIL ||
3536 list_length(resultRelations) == list_length(returningLists));
3537
3538 pathnode->path.pathtype = T_ModifyTable;
3539 pathnode->path.parent = rel;
3540 /* pathtarget is not interesting, just make it minimally valid */
3541 pathnode->path.pathtarget = rel->reltarget;
3542 /* For now, assume we are above any joins, so no parameterization */
3543 pathnode->path.param_info = NULL;
3544 pathnode->path.parallel_aware = false;
3545 pathnode->path.parallel_safe = false;
3546 pathnode->path.parallel_workers = 0;
3547 pathnode->path.pathkeys = NIL;
3548
3549 /*
3550 * Compute cost & rowcount as sum of subpath costs & rowcounts.
3551 *
3552 * Currently, we don't charge anything extra for the actual table
3553 * modification work, nor for the WITH CHECK OPTIONS or RETURNING
3554 * expressions if any. It would only be window dressing, since
3555 * ModifyTable is always a top-level node and there is no way for the
3556 * costs to change any higher-level planning choices. But we might want
3557 * to make it look better sometime.
3558 */
3559 pathnode->path.startup_cost = 0;
3560 pathnode->path.total_cost = 0;
3561 pathnode->path.rows = 0;
3562 total_size = 0;
3563 foreach(lc, subpaths)
3564 {
3565 Path *subpath = (Path *) lfirst(lc);
3566
3567 if (lc == list_head(subpaths)) /* first node? */
3568 pathnode->path.startup_cost = subpath->startup_cost;
3569 pathnode->path.total_cost += subpath->total_cost;
3570 pathnode->path.rows += subpath->rows;
3571 total_size += subpath->pathtarget->width * subpath->rows;
3572 }
3573
3574 /*
3575 * Set width to the average width of the subpath outputs. XXX this is
3576 * totally wrong: we should report zero if no RETURNING, else an average
3577 * of the RETURNING tlist widths. But it's what happened historically,
3578 * and improving it is a task for another day.
3579 */
3580 if (pathnode->path.rows > 0)
3581 total_size /= pathnode->path.rows;
3582 pathnode->path.pathtarget->width = rint(total_size);
3583
3584 pathnode->operation = operation;
3585 pathnode->canSetTag = canSetTag;
3586 pathnode->nominalRelation = nominalRelation;
3587 pathnode->rootRelation = rootRelation;
3588 pathnode->partColsUpdated = partColsUpdated;
3589 pathnode->resultRelations = resultRelations;
3590 pathnode->subpaths = subpaths;
3591 pathnode->subroots = subroots;
3592 pathnode->withCheckOptionLists = withCheckOptionLists;
3593 pathnode->returningLists = returningLists;
3594 pathnode->rowMarks = rowMarks;
3595 pathnode->onconflict = onconflict;
3596 pathnode->epqParam = epqParam;
3597
3598 return pathnode;
3599 }
3600
3601 /*
3602 * create_limit_path
3603 * Creates a pathnode that represents performing LIMIT/OFFSET
3604 *
3605 * In addition to providing the actual OFFSET and LIMIT expressions,
3606 * the caller must provide estimates of their values for costing purposes.
3607 * The estimates are as computed by preprocess_limit(), ie, 0 represents
3608 * the clause not being present, and -1 means it's present but we could
3609 * not estimate its value.
3610 *
3611 * 'rel' is the parent relation associated with the result
3612 * 'subpath' is the path representing the source of data
3613 * 'limitOffset' is the actual OFFSET expression, or NULL
3614 * 'limitCount' is the actual LIMIT expression, or NULL
3615 * 'offset_est' is the estimated value of the OFFSET expression
3616 * 'count_est' is the estimated value of the LIMIT expression
3617 */
3618 LimitPath *
create_limit_path(PlannerInfo * root,RelOptInfo * rel,Path * subpath,Node * limitOffset,Node * limitCount,int64 offset_est,int64 count_est)3619 create_limit_path(PlannerInfo *root, RelOptInfo *rel,
3620 Path *subpath,
3621 Node *limitOffset, Node *limitCount,
3622 int64 offset_est, int64 count_est)
3623 {
3624 LimitPath *pathnode = makeNode(LimitPath);
3625
3626 pathnode->path.pathtype = T_Limit;
3627 pathnode->path.parent = rel;
3628 /* Limit doesn't project, so use source path's pathtarget */
3629 pathnode->path.pathtarget = subpath->pathtarget;
3630 /* For now, assume we are above any joins, so no parameterization */
3631 pathnode->path.param_info = NULL;
3632 pathnode->path.parallel_aware = false;
3633 pathnode->path.parallel_safe = rel->consider_parallel &&
3634 subpath->parallel_safe;
3635 pathnode->path.parallel_workers = subpath->parallel_workers;
3636 pathnode->path.rows = subpath->rows;
3637 pathnode->path.startup_cost = subpath->startup_cost;
3638 pathnode->path.total_cost = subpath->total_cost;
3639 pathnode->path.pathkeys = subpath->pathkeys;
3640 pathnode->subpath = subpath;
3641 pathnode->limitOffset = limitOffset;
3642 pathnode->limitCount = limitCount;
3643
3644 /*
3645 * Adjust the output rows count and costs according to the offset/limit.
3646 */
3647 adjust_limit_rows_costs(&pathnode->path.rows,
3648 &pathnode->path.startup_cost,
3649 &pathnode->path.total_cost,
3650 offset_est, count_est);
3651
3652 return pathnode;
3653 }
3654
3655 /*
3656 * adjust_limit_rows_costs
3657 * Adjust the size and cost estimates for a LimitPath node according to the
3658 * offset/limit.
3659 *
3660 * This is only a cosmetic issue if we are at top level, but if we are
3661 * building a subquery then it's important to report correct info to the outer
3662 * planner.
3663 *
3664 * When the offset or count couldn't be estimated, use 10% of the estimated
3665 * number of rows emitted from the subpath.
3666 *
3667 * XXX we don't bother to add eval costs of the offset/limit expressions
3668 * themselves to the path costs. In theory we should, but in most cases those
3669 * expressions are trivial and it's just not worth the trouble.
3670 */
3671 void
adjust_limit_rows_costs(double * rows,Cost * startup_cost,Cost * total_cost,int64 offset_est,int64 count_est)3672 adjust_limit_rows_costs(double *rows, /* in/out parameter */
3673 Cost *startup_cost, /* in/out parameter */
3674 Cost *total_cost, /* in/out parameter */
3675 int64 offset_est,
3676 int64 count_est)
3677 {
3678 double input_rows = *rows;
3679 Cost input_startup_cost = *startup_cost;
3680 Cost input_total_cost = *total_cost;
3681
3682 if (offset_est != 0)
3683 {
3684 double offset_rows;
3685
3686 if (offset_est > 0)
3687 offset_rows = (double) offset_est;
3688 else
3689 offset_rows = clamp_row_est(input_rows * 0.10);
3690 if (offset_rows > *rows)
3691 offset_rows = *rows;
3692 if (input_rows > 0)
3693 *startup_cost +=
3694 (input_total_cost - input_startup_cost)
3695 * offset_rows / input_rows;
3696 *rows -= offset_rows;
3697 if (*rows < 1)
3698 *rows = 1;
3699 }
3700
3701 if (count_est != 0)
3702 {
3703 double count_rows;
3704
3705 if (count_est > 0)
3706 count_rows = (double) count_est;
3707 else
3708 count_rows = clamp_row_est(input_rows * 0.10);
3709 if (count_rows > *rows)
3710 count_rows = *rows;
3711 if (input_rows > 0)
3712 *total_cost = *startup_cost +
3713 (input_total_cost - input_startup_cost)
3714 * count_rows / input_rows;
3715 *rows = count_rows;
3716 if (*rows < 1)
3717 *rows = 1;
3718 }
3719 }
3720
3721
3722 /*
3723 * reparameterize_path
3724 * Attempt to modify a Path to have greater parameterization
3725 *
3726 * We use this to attempt to bring all child paths of an appendrel to the
3727 * same parameterization level, ensuring that they all enforce the same set
3728 * of join quals (and thus that that parameterization can be attributed to
3729 * an append path built from such paths). Currently, only a few path types
3730 * are supported here, though more could be added at need. We return NULL
3731 * if we can't reparameterize the given path.
3732 *
3733 * Note: we intentionally do not pass created paths to add_path(); it would
3734 * possibly try to delete them on the grounds of being cost-inferior to the
3735 * paths they were made from, and we don't want that. Paths made here are
3736 * not necessarily of general-purpose usefulness, but they can be useful
3737 * as members of an append path.
3738 */
3739 Path *
reparameterize_path(PlannerInfo * root,Path * path,Relids required_outer,double loop_count)3740 reparameterize_path(PlannerInfo *root, Path *path,
3741 Relids required_outer,
3742 double loop_count)
3743 {
3744 RelOptInfo *rel = path->parent;
3745
3746 /* Can only increase, not decrease, path's parameterization */
3747 if (!bms_is_subset(PATH_REQ_OUTER(path), required_outer))
3748 return NULL;
3749 switch (path->pathtype)
3750 {
3751 case T_SeqScan:
3752 return create_seqscan_path(root, rel, required_outer, 0);
3753 case T_SampleScan:
3754 return (Path *) create_samplescan_path(root, rel, required_outer);
3755 case T_IndexScan:
3756 case T_IndexOnlyScan:
3757 {
3758 IndexPath *ipath = (IndexPath *) path;
3759 IndexPath *newpath = makeNode(IndexPath);
3760
3761 /*
3762 * We can't use create_index_path directly, and would not want
3763 * to because it would re-compute the indexqual conditions
3764 * which is wasted effort. Instead we hack things a bit:
3765 * flat-copy the path node, revise its param_info, and redo
3766 * the cost estimate.
3767 */
3768 memcpy(newpath, ipath, sizeof(IndexPath));
3769 newpath->path.param_info =
3770 get_baserel_parampathinfo(root, rel, required_outer);
3771 cost_index(newpath, root, loop_count, false);
3772 return (Path *) newpath;
3773 }
3774 case T_BitmapHeapScan:
3775 {
3776 BitmapHeapPath *bpath = (BitmapHeapPath *) path;
3777
3778 return (Path *) create_bitmap_heap_path(root,
3779 rel,
3780 bpath->bitmapqual,
3781 required_outer,
3782 loop_count, 0);
3783 }
3784 case T_SubqueryScan:
3785 {
3786 SubqueryScanPath *spath = (SubqueryScanPath *) path;
3787
3788 return (Path *) create_subqueryscan_path(root,
3789 rel,
3790 spath->subpath,
3791 spath->path.pathkeys,
3792 required_outer);
3793 }
3794 case T_Result:
3795 /* Supported only for RTE_RESULT scan paths */
3796 if (IsA(path, Path))
3797 return create_resultscan_path(root, rel, required_outer);
3798 break;
3799 case T_Append:
3800 {
3801 AppendPath *apath = (AppendPath *) path;
3802 List *childpaths = NIL;
3803 List *partialpaths = NIL;
3804 int i;
3805 ListCell *lc;
3806
3807 /* Reparameterize the children */
3808 i = 0;
3809 foreach(lc, apath->subpaths)
3810 {
3811 Path *spath = (Path *) lfirst(lc);
3812
3813 spath = reparameterize_path(root, spath,
3814 required_outer,
3815 loop_count);
3816 if (spath == NULL)
3817 return NULL;
3818 /* We have to re-split the regular and partial paths */
3819 if (i < apath->first_partial_path)
3820 childpaths = lappend(childpaths, spath);
3821 else
3822 partialpaths = lappend(partialpaths, spath);
3823 i++;
3824 }
3825 return (Path *)
3826 create_append_path(root, rel, childpaths, partialpaths,
3827 apath->path.pathkeys, required_outer,
3828 apath->path.parallel_workers,
3829 apath->path.parallel_aware,
3830 apath->partitioned_rels,
3831 -1);
3832 }
3833 default:
3834 break;
3835 }
3836 return NULL;
3837 }
3838
3839 /*
3840 * reparameterize_path_by_child
3841 * Given a path parameterized by the parent of the given child relation,
3842 * translate the path to be parameterized by the given child relation.
3843 *
3844 * The function creates a new path of the same type as the given path, but
3845 * parameterized by the given child relation. Most fields from the original
3846 * path can simply be flat-copied, but any expressions must be adjusted to
3847 * refer to the correct varnos, and any paths must be recursively
3848 * reparameterized. Other fields that refer to specific relids also need
3849 * adjustment.
3850 *
3851 * The cost, number of rows, width and parallel path properties depend upon
3852 * path->parent, which does not change during the translation. Hence those
3853 * members are copied as they are.
3854 *
3855 * If the given path can not be reparameterized, the function returns NULL.
3856 */
3857 Path *
reparameterize_path_by_child(PlannerInfo * root,Path * path,RelOptInfo * child_rel)3858 reparameterize_path_by_child(PlannerInfo *root, Path *path,
3859 RelOptInfo *child_rel)
3860 {
3861
3862 #define FLAT_COPY_PATH(newnode, node, nodetype) \
3863 ( (newnode) = makeNode(nodetype), \
3864 memcpy((newnode), (node), sizeof(nodetype)) )
3865
3866 #define ADJUST_CHILD_ATTRS(node) \
3867 ((node) = \
3868 (List *) adjust_appendrel_attrs_multilevel(root, (Node *) (node), \
3869 child_rel->relids, \
3870 child_rel->top_parent_relids))
3871
3872 #define REPARAMETERIZE_CHILD_PATH(path) \
3873 do { \
3874 (path) = reparameterize_path_by_child(root, (path), child_rel); \
3875 if ((path) == NULL) \
3876 return NULL; \
3877 } while(0)
3878
3879 #define REPARAMETERIZE_CHILD_PATH_LIST(pathlist) \
3880 do { \
3881 if ((pathlist) != NIL) \
3882 { \
3883 (pathlist) = reparameterize_pathlist_by_child(root, (pathlist), \
3884 child_rel); \
3885 if ((pathlist) == NIL) \
3886 return NULL; \
3887 } \
3888 } while(0)
3889
3890 Path *new_path;
3891 ParamPathInfo *new_ppi;
3892 ParamPathInfo *old_ppi;
3893 Relids required_outer;
3894
3895 /*
3896 * If the path is not parameterized by parent of the given relation, it
3897 * doesn't need reparameterization.
3898 */
3899 if (!path->param_info ||
3900 !bms_overlap(PATH_REQ_OUTER(path), child_rel->top_parent_relids))
3901 return path;
3902
3903 /*
3904 * If possible, reparameterize the given path, making a copy.
3905 *
3906 * This function is currently only applied to the inner side of a nestloop
3907 * join that is being partitioned by the partitionwise-join code. Hence,
3908 * we need only support path types that plausibly arise in that context.
3909 * (In particular, supporting sorted path types would be a waste of code
3910 * and cycles: even if we translated them here, they'd just lose in
3911 * subsequent cost comparisons.) If we do see an unsupported path type,
3912 * that just means we won't be able to generate a partitionwise-join plan
3913 * using that path type.
3914 */
3915 switch (nodeTag(path))
3916 {
3917 case T_Path:
3918 FLAT_COPY_PATH(new_path, path, Path);
3919 break;
3920
3921 case T_IndexPath:
3922 {
3923 IndexPath *ipath;
3924
3925 FLAT_COPY_PATH(ipath, path, IndexPath);
3926 ADJUST_CHILD_ATTRS(ipath->indexclauses);
3927 new_path = (Path *) ipath;
3928 }
3929 break;
3930
3931 case T_BitmapHeapPath:
3932 {
3933 BitmapHeapPath *bhpath;
3934
3935 FLAT_COPY_PATH(bhpath, path, BitmapHeapPath);
3936 REPARAMETERIZE_CHILD_PATH(bhpath->bitmapqual);
3937 new_path = (Path *) bhpath;
3938 }
3939 break;
3940
3941 case T_BitmapAndPath:
3942 {
3943 BitmapAndPath *bapath;
3944
3945 FLAT_COPY_PATH(bapath, path, BitmapAndPath);
3946 REPARAMETERIZE_CHILD_PATH_LIST(bapath->bitmapquals);
3947 new_path = (Path *) bapath;
3948 }
3949 break;
3950
3951 case T_BitmapOrPath:
3952 {
3953 BitmapOrPath *bopath;
3954
3955 FLAT_COPY_PATH(bopath, path, BitmapOrPath);
3956 REPARAMETERIZE_CHILD_PATH_LIST(bopath->bitmapquals);
3957 new_path = (Path *) bopath;
3958 }
3959 break;
3960
3961 case T_ForeignPath:
3962 {
3963 ForeignPath *fpath;
3964 ReparameterizeForeignPathByChild_function rfpc_func;
3965
3966 FLAT_COPY_PATH(fpath, path, ForeignPath);
3967 if (fpath->fdw_outerpath)
3968 REPARAMETERIZE_CHILD_PATH(fpath->fdw_outerpath);
3969
3970 /* Hand over to FDW if needed. */
3971 rfpc_func =
3972 path->parent->fdwroutine->ReparameterizeForeignPathByChild;
3973 if (rfpc_func)
3974 fpath->fdw_private = rfpc_func(root, fpath->fdw_private,
3975 child_rel);
3976 new_path = (Path *) fpath;
3977 }
3978 break;
3979
3980 case T_CustomPath:
3981 {
3982 CustomPath *cpath;
3983
3984 FLAT_COPY_PATH(cpath, path, CustomPath);
3985 REPARAMETERIZE_CHILD_PATH_LIST(cpath->custom_paths);
3986 if (cpath->methods &&
3987 cpath->methods->ReparameterizeCustomPathByChild)
3988 cpath->custom_private =
3989 cpath->methods->ReparameterizeCustomPathByChild(root,
3990 cpath->custom_private,
3991 child_rel);
3992 new_path = (Path *) cpath;
3993 }
3994 break;
3995
3996 case T_NestPath:
3997 {
3998 JoinPath *jpath;
3999
4000 FLAT_COPY_PATH(jpath, path, NestPath);
4001
4002 REPARAMETERIZE_CHILD_PATH(jpath->outerjoinpath);
4003 REPARAMETERIZE_CHILD_PATH(jpath->innerjoinpath);
4004 ADJUST_CHILD_ATTRS(jpath->joinrestrictinfo);
4005 new_path = (Path *) jpath;
4006 }
4007 break;
4008
4009 case T_MergePath:
4010 {
4011 JoinPath *jpath;
4012 MergePath *mpath;
4013
4014 FLAT_COPY_PATH(mpath, path, MergePath);
4015
4016 jpath = (JoinPath *) mpath;
4017 REPARAMETERIZE_CHILD_PATH(jpath->outerjoinpath);
4018 REPARAMETERIZE_CHILD_PATH(jpath->innerjoinpath);
4019 ADJUST_CHILD_ATTRS(jpath->joinrestrictinfo);
4020 ADJUST_CHILD_ATTRS(mpath->path_mergeclauses);
4021 new_path = (Path *) mpath;
4022 }
4023 break;
4024
4025 case T_HashPath:
4026 {
4027 JoinPath *jpath;
4028 HashPath *hpath;
4029
4030 FLAT_COPY_PATH(hpath, path, HashPath);
4031
4032 jpath = (JoinPath *) hpath;
4033 REPARAMETERIZE_CHILD_PATH(jpath->outerjoinpath);
4034 REPARAMETERIZE_CHILD_PATH(jpath->innerjoinpath);
4035 ADJUST_CHILD_ATTRS(jpath->joinrestrictinfo);
4036 ADJUST_CHILD_ATTRS(hpath->path_hashclauses);
4037 new_path = (Path *) hpath;
4038 }
4039 break;
4040
4041 case T_AppendPath:
4042 {
4043 AppendPath *apath;
4044
4045 FLAT_COPY_PATH(apath, path, AppendPath);
4046 REPARAMETERIZE_CHILD_PATH_LIST(apath->subpaths);
4047 new_path = (Path *) apath;
4048 }
4049 break;
4050
4051 case T_GatherPath:
4052 {
4053 GatherPath *gpath;
4054
4055 FLAT_COPY_PATH(gpath, path, GatherPath);
4056 REPARAMETERIZE_CHILD_PATH(gpath->subpath);
4057 new_path = (Path *) gpath;
4058 }
4059 break;
4060
4061 default:
4062
4063 /* We don't know how to reparameterize this path. */
4064 return NULL;
4065 }
4066
4067 /*
4068 * Adjust the parameterization information, which refers to the topmost
4069 * parent. The topmost parent can be multiple levels away from the given
4070 * child, hence use multi-level expression adjustment routines.
4071 */
4072 old_ppi = new_path->param_info;
4073 required_outer =
4074 adjust_child_relids_multilevel(root, old_ppi->ppi_req_outer,
4075 child_rel->relids,
4076 child_rel->top_parent_relids);
4077
4078 /* If we already have a PPI for this parameterization, just return it */
4079 new_ppi = find_param_path_info(new_path->parent, required_outer);
4080
4081 /*
4082 * If not, build a new one and link it to the list of PPIs. For the same
4083 * reason as explained in mark_dummy_rel(), allocate new PPI in the same
4084 * context the given RelOptInfo is in.
4085 */
4086 if (new_ppi == NULL)
4087 {
4088 MemoryContext oldcontext;
4089 RelOptInfo *rel = path->parent;
4090
4091 oldcontext = MemoryContextSwitchTo(GetMemoryChunkContext(rel));
4092
4093 new_ppi = makeNode(ParamPathInfo);
4094 new_ppi->ppi_req_outer = bms_copy(required_outer);
4095 new_ppi->ppi_rows = old_ppi->ppi_rows;
4096 new_ppi->ppi_clauses = old_ppi->ppi_clauses;
4097 ADJUST_CHILD_ATTRS(new_ppi->ppi_clauses);
4098 rel->ppilist = lappend(rel->ppilist, new_ppi);
4099
4100 MemoryContextSwitchTo(oldcontext);
4101 }
4102 bms_free(required_outer);
4103
4104 new_path->param_info = new_ppi;
4105
4106 /*
4107 * Adjust the path target if the parent of the outer relation is
4108 * referenced in the targetlist. This can happen when only the parent of
4109 * outer relation is laterally referenced in this relation.
4110 */
4111 if (bms_overlap(path->parent->lateral_relids,
4112 child_rel->top_parent_relids))
4113 {
4114 new_path->pathtarget = copy_pathtarget(new_path->pathtarget);
4115 ADJUST_CHILD_ATTRS(new_path->pathtarget->exprs);
4116 }
4117
4118 return new_path;
4119 }
4120
4121 /*
4122 * reparameterize_pathlist_by_child
4123 * Helper function to reparameterize a list of paths by given child rel.
4124 */
4125 static List *
reparameterize_pathlist_by_child(PlannerInfo * root,List * pathlist,RelOptInfo * child_rel)4126 reparameterize_pathlist_by_child(PlannerInfo *root,
4127 List *pathlist,
4128 RelOptInfo *child_rel)
4129 {
4130 ListCell *lc;
4131 List *result = NIL;
4132
4133 foreach(lc, pathlist)
4134 {
4135 Path *path = reparameterize_path_by_child(root, lfirst(lc),
4136 child_rel);
4137
4138 if (path == NULL)
4139 {
4140 list_free(result);
4141 return NIL;
4142 }
4143
4144 result = lappend(result, path);
4145 }
4146
4147 return result;
4148 }
4149