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 |
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 |
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 |
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 |
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 |
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 |
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 |
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 |
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 * |
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 * |
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 * |
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 * |
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 * |
1105 | create_bitmap_and_path(PlannerInfo *root, |
1106 | RelOptInfo *rel, |
1107 | List *bitmapquals) |
1108 | { |
1109 | BitmapAndPath *pathnode = makeNode(BitmapAndPath); |
1110 | |
1111 | pathnode->path.pathtype = T_BitmapAnd; |
1112 | pathnode->path.parent = rel; |
1113 | pathnode->path.pathtarget = rel->reltarget; |
1114 | pathnode->path.param_info = NULL; /* not used in bitmap trees */ |
1115 | |
1116 | /* |
1117 | * Currently, a BitmapHeapPath, BitmapAndPath, or BitmapOrPath will be |
1118 | * parallel-safe if and only if rel->consider_parallel is set. So, we can |
1119 | * set the flag for this path based only on the relation-level flag, |
1120 | * without actually iterating over the list of children. |
1121 | */ |
1122 | pathnode->path.parallel_aware = false; |
1123 | pathnode->path.parallel_safe = rel->consider_parallel; |
1124 | pathnode->path.parallel_workers = 0; |
1125 | |
1126 | pathnode->path.pathkeys = NIL; /* always unordered */ |
1127 | |
1128 | pathnode->bitmapquals = bitmapquals; |
1129 | |
1130 | /* this sets bitmapselectivity as well as the regular cost fields: */ |
1131 | cost_bitmap_and_node(pathnode, root); |
1132 | |
1133 | return pathnode; |
1134 | } |
1135 | |
1136 | /* |
1137 | * create_bitmap_or_path |
1138 | * Creates a path node representing a BitmapOr. |
1139 | */ |
1140 | BitmapOrPath * |
1141 | create_bitmap_or_path(PlannerInfo *root, |
1142 | RelOptInfo *rel, |
1143 | List *bitmapquals) |
1144 | { |
1145 | BitmapOrPath *pathnode = makeNode(BitmapOrPath); |
1146 | |
1147 | pathnode->path.pathtype = T_BitmapOr; |
1148 | pathnode->path.parent = rel; |
1149 | pathnode->path.pathtarget = rel->reltarget; |
1150 | pathnode->path.param_info = NULL; /* not used in bitmap trees */ |
1151 | |
1152 | /* |
1153 | * Currently, a BitmapHeapPath, BitmapAndPath, or BitmapOrPath will be |
1154 | * parallel-safe if and only if rel->consider_parallel is set. So, we can |
1155 | * set the flag for this path based only on the relation-level flag, |
1156 | * without actually iterating over the list of children. |
1157 | */ |
1158 | pathnode->path.parallel_aware = false; |
1159 | pathnode->path.parallel_safe = rel->consider_parallel; |
1160 | pathnode->path.parallel_workers = 0; |
1161 | |
1162 | pathnode->path.pathkeys = NIL; /* always unordered */ |
1163 | |
1164 | pathnode->bitmapquals = bitmapquals; |
1165 | |
1166 | /* this sets bitmapselectivity as well as the regular cost fields: */ |
1167 | cost_bitmap_or_node(pathnode, root); |
1168 | |
1169 | return pathnode; |
1170 | } |
1171 | |
1172 | /* |
1173 | * create_tidscan_path |
1174 | * Creates a path corresponding to a scan by TID, returning the pathnode. |
1175 | */ |
1176 | TidPath * |
1177 | create_tidscan_path(PlannerInfo *root, RelOptInfo *rel, List *tidquals, |
1178 | Relids required_outer) |
1179 | { |
1180 | TidPath *pathnode = makeNode(TidPath); |
1181 | |
1182 | pathnode->path.pathtype = T_TidScan; |
1183 | pathnode->path.parent = rel; |
1184 | pathnode->path.pathtarget = rel->reltarget; |
1185 | pathnode->path.param_info = get_baserel_parampathinfo(root, rel, |
1186 | required_outer); |
1187 | pathnode->path.parallel_aware = false; |
1188 | pathnode->path.parallel_safe = rel->consider_parallel; |
1189 | pathnode->path.parallel_workers = 0; |
1190 | pathnode->path.pathkeys = NIL; /* always unordered */ |
1191 | |
1192 | pathnode->tidquals = tidquals; |
1193 | |
1194 | cost_tidscan(&pathnode->path, root, rel, tidquals, |
1195 | pathnode->path.param_info); |
1196 | |
1197 | return pathnode; |
1198 | } |
1199 | |
1200 | /* |
1201 | * create_append_path |
1202 | * Creates a path corresponding to an Append plan, returning the |
1203 | * pathnode. |
1204 | * |
1205 | * Note that we must handle subpaths = NIL, representing a dummy access path. |
1206 | * Also, there are callers that pass root = NULL. |
1207 | */ |
1208 | AppendPath * |
1209 | create_append_path(PlannerInfo *root, |
1210 | RelOptInfo *rel, |
1211 | List *subpaths, List *partial_subpaths, |
1212 | List *pathkeys, Relids required_outer, |
1213 | int parallel_workers, bool parallel_aware, |
1214 | List *partitioned_rels, double rows) |
1215 | { |
1216 | AppendPath *pathnode = makeNode(AppendPath); |
1217 | ListCell *l; |
1218 | |
1219 | Assert(!parallel_aware || parallel_workers > 0); |
1220 | |
1221 | pathnode->path.pathtype = T_Append; |
1222 | pathnode->path.parent = rel; |
1223 | pathnode->path.pathtarget = rel->reltarget; |
1224 | |
1225 | /* |
1226 | * When generating an Append path for a partitioned table, there may be |
1227 | * parameters that are useful so we can eliminate certain partitions |
1228 | * during execution. Here we'll go all the way and fully populate the |
1229 | * parameter info data as we do for normal base relations. However, we |
1230 | * need only bother doing this for RELOPT_BASEREL rels, as |
1231 | * RELOPT_OTHER_MEMBER_REL's Append paths are merged into the base rel's |
1232 | * Append subpaths. It would do no harm to do this, we just avoid it to |
1233 | * save wasting effort. |
1234 | */ |
1235 | if (partitioned_rels != NIL && root && rel->reloptkind == RELOPT_BASEREL) |
1236 | pathnode->path.param_info = get_baserel_parampathinfo(root, |
1237 | rel, |
1238 | required_outer); |
1239 | else |
1240 | pathnode->path.param_info = get_appendrel_parampathinfo(rel, |
1241 | required_outer); |
1242 | |
1243 | pathnode->path.parallel_aware = parallel_aware; |
1244 | pathnode->path.parallel_safe = rel->consider_parallel; |
1245 | pathnode->path.parallel_workers = parallel_workers; |
1246 | pathnode->path.pathkeys = pathkeys; |
1247 | pathnode->partitioned_rels = list_copy(partitioned_rels); |
1248 | |
1249 | /* |
1250 | * For parallel append, non-partial paths are sorted by descending total |
1251 | * costs. That way, the total time to finish all non-partial paths is |
1252 | * minimized. Also, the partial paths are sorted by descending startup |
1253 | * costs. There may be some paths that require to do startup work by a |
1254 | * single worker. In such case, it's better for workers to choose the |
1255 | * expensive ones first, whereas the leader should choose the cheapest |
1256 | * startup plan. |
1257 | */ |
1258 | if (pathnode->path.parallel_aware) |
1259 | { |
1260 | /* |
1261 | * We mustn't fiddle with the order of subpaths when the Append has |
1262 | * pathkeys. The order they're listed in is critical to keeping the |
1263 | * pathkeys valid. |
1264 | */ |
1265 | Assert(pathkeys == NIL); |
1266 | |
1267 | subpaths = list_qsort(subpaths, append_total_cost_compare); |
1268 | partial_subpaths = list_qsort(partial_subpaths, |
1269 | append_startup_cost_compare); |
1270 | } |
1271 | pathnode->first_partial_path = list_length(subpaths); |
1272 | pathnode->subpaths = list_concat(subpaths, partial_subpaths); |
1273 | |
1274 | /* |
1275 | * Apply query-wide LIMIT if known and path is for sole base relation. |
1276 | * (Handling this at this low level is a bit klugy.) |
1277 | */ |
1278 | if (root != NULL && bms_equal(rel->relids, root->all_baserels)) |
1279 | pathnode->limit_tuples = root->limit_tuples; |
1280 | else |
1281 | pathnode->limit_tuples = -1.0; |
1282 | |
1283 | foreach(l, pathnode->subpaths) |
1284 | { |
1285 | Path *subpath = (Path *) lfirst(l); |
1286 | |
1287 | pathnode->path.parallel_safe = pathnode->path.parallel_safe && |
1288 | subpath->parallel_safe; |
1289 | |
1290 | /* All child paths must have same parameterization */ |
1291 | Assert(bms_equal(PATH_REQ_OUTER(subpath), required_outer)); |
1292 | } |
1293 | |
1294 | Assert(!parallel_aware || pathnode->path.parallel_safe); |
1295 | |
1296 | /* |
1297 | * If there's exactly one child path, the Append is a no-op and will be |
1298 | * discarded later (in setrefs.c); therefore, we can inherit the child's |
1299 | * size and cost, as well as its pathkeys if any (overriding whatever the |
1300 | * caller might've said). Otherwise, we must do the normal costsize |
1301 | * calculation. |
1302 | */ |
1303 | if (list_length(pathnode->subpaths) == 1) |
1304 | { |
1305 | Path *child = (Path *) linitial(pathnode->subpaths); |
1306 | |
1307 | pathnode->path.rows = child->rows; |
1308 | pathnode->path.startup_cost = child->startup_cost; |
1309 | pathnode->path.total_cost = child->total_cost; |
1310 | pathnode->path.pathkeys = child->pathkeys; |
1311 | } |
1312 | else |
1313 | cost_append(pathnode); |
1314 | |
1315 | /* If the caller provided a row estimate, override the computed value. */ |
1316 | if (rows >= 0) |
1317 | pathnode->path.rows = rows; |
1318 | |
1319 | return pathnode; |
1320 | } |
1321 | |
1322 | /* |
1323 | * append_total_cost_compare |
1324 | * qsort comparator for sorting append child paths by total_cost descending |
1325 | * |
1326 | * For equal total costs, we fall back to comparing startup costs; if those |
1327 | * are equal too, break ties using bms_compare on the paths' relids. |
1328 | * (This is to avoid getting unpredictable results from qsort.) |
1329 | */ |
1330 | static int |
1331 | append_total_cost_compare(const void *a, const void *b) |
1332 | { |
1333 | Path *path1 = (Path *) lfirst(*(ListCell **) a); |
1334 | Path *path2 = (Path *) lfirst(*(ListCell **) b); |
1335 | int cmp; |
1336 | |
1337 | cmp = compare_path_costs(path1, path2, TOTAL_COST); |
1338 | if (cmp != 0) |
1339 | return -cmp; |
1340 | return bms_compare(path1->parent->relids, path2->parent->relids); |
1341 | } |
1342 | |
1343 | /* |
1344 | * append_startup_cost_compare |
1345 | * qsort comparator for sorting append child paths by startup_cost descending |
1346 | * |
1347 | * For equal startup costs, we fall back to comparing total costs; if those |
1348 | * are equal too, break ties using bms_compare on the paths' relids. |
1349 | * (This is to avoid getting unpredictable results from qsort.) |
1350 | */ |
1351 | static int |
1352 | append_startup_cost_compare(const void *a, const void *b) |
1353 | { |
1354 | Path *path1 = (Path *) lfirst(*(ListCell **) a); |
1355 | Path *path2 = (Path *) lfirst(*(ListCell **) b); |
1356 | int cmp; |
1357 | |
1358 | cmp = compare_path_costs(path1, path2, STARTUP_COST); |
1359 | if (cmp != 0) |
1360 | return -cmp; |
1361 | return bms_compare(path1->parent->relids, path2->parent->relids); |
1362 | } |
1363 | |
1364 | /* |
1365 | * create_merge_append_path |
1366 | * Creates a path corresponding to a MergeAppend plan, returning the |
1367 | * pathnode. |
1368 | */ |
1369 | MergeAppendPath * |
1370 | create_merge_append_path(PlannerInfo *root, |
1371 | RelOptInfo *rel, |
1372 | List *subpaths, |
1373 | List *pathkeys, |
1374 | Relids required_outer, |
1375 | List *partitioned_rels) |
1376 | { |
1377 | MergeAppendPath *pathnode = makeNode(MergeAppendPath); |
1378 | Cost input_startup_cost; |
1379 | Cost input_total_cost; |
1380 | ListCell *l; |
1381 | |
1382 | pathnode->path.pathtype = T_MergeAppend; |
1383 | pathnode->path.parent = rel; |
1384 | pathnode->path.pathtarget = rel->reltarget; |
1385 | pathnode->path.param_info = get_appendrel_parampathinfo(rel, |
1386 | required_outer); |
1387 | pathnode->path.parallel_aware = false; |
1388 | pathnode->path.parallel_safe = rel->consider_parallel; |
1389 | pathnode->path.parallel_workers = 0; |
1390 | pathnode->path.pathkeys = pathkeys; |
1391 | pathnode->partitioned_rels = list_copy(partitioned_rels); |
1392 | pathnode->subpaths = subpaths; |
1393 | |
1394 | /* |
1395 | * Apply query-wide LIMIT if known and path is for sole base relation. |
1396 | * (Handling this at this low level is a bit klugy.) |
1397 | */ |
1398 | if (bms_equal(rel->relids, root->all_baserels)) |
1399 | pathnode->limit_tuples = root->limit_tuples; |
1400 | else |
1401 | pathnode->limit_tuples = -1.0; |
1402 | |
1403 | /* |
1404 | * Add up the sizes and costs of the input paths. |
1405 | */ |
1406 | pathnode->path.rows = 0; |
1407 | input_startup_cost = 0; |
1408 | input_total_cost = 0; |
1409 | foreach(l, subpaths) |
1410 | { |
1411 | Path *subpath = (Path *) lfirst(l); |
1412 | |
1413 | pathnode->path.rows += subpath->rows; |
1414 | pathnode->path.parallel_safe = pathnode->path.parallel_safe && |
1415 | subpath->parallel_safe; |
1416 | |
1417 | if (pathkeys_contained_in(pathkeys, subpath->pathkeys)) |
1418 | { |
1419 | /* Subpath is adequately ordered, we won't need to sort it */ |
1420 | input_startup_cost += subpath->startup_cost; |
1421 | input_total_cost += subpath->total_cost; |
1422 | } |
1423 | else |
1424 | { |
1425 | /* We'll need to insert a Sort node, so include cost for that */ |
1426 | Path sort_path; /* dummy for result of cost_sort */ |
1427 | |
1428 | cost_sort(&sort_path, |
1429 | root, |
1430 | pathkeys, |
1431 | subpath->total_cost, |
1432 | subpath->parent->tuples, |
1433 | subpath->pathtarget->width, |
1434 | 0.0, |
1435 | work_mem, |
1436 | pathnode->limit_tuples); |
1437 | input_startup_cost += sort_path.startup_cost; |
1438 | input_total_cost += sort_path.total_cost; |
1439 | } |
1440 | |
1441 | /* All child paths must have same parameterization */ |
1442 | Assert(bms_equal(PATH_REQ_OUTER(subpath), required_outer)); |
1443 | } |
1444 | |
1445 | /* |
1446 | * Now we can compute total costs of the MergeAppend. If there's exactly |
1447 | * one child path, the MergeAppend is a no-op and will be discarded later |
1448 | * (in setrefs.c); otherwise we do the normal cost calculation. |
1449 | */ |
1450 | if (list_length(subpaths) == 1) |
1451 | { |
1452 | pathnode->path.startup_cost = input_startup_cost; |
1453 | pathnode->path.total_cost = input_total_cost; |
1454 | } |
1455 | else |
1456 | cost_merge_append(&pathnode->path, root, |
1457 | pathkeys, list_length(subpaths), |
1458 | input_startup_cost, input_total_cost, |
1459 | pathnode->path.rows); |
1460 | |
1461 | return pathnode; |
1462 | } |
1463 | |
1464 | /* |
1465 | * create_group_result_path |
1466 | * Creates a path representing a Result-and-nothing-else plan. |
1467 | * |
1468 | * This is only used for degenerate grouping cases, in which we know we |
1469 | * need to produce one result row, possibly filtered by a HAVING qual. |
1470 | */ |
1471 | GroupResultPath * |
1472 | create_group_result_path(PlannerInfo *root, RelOptInfo *rel, |
1473 | PathTarget *target, List *havingqual) |
1474 | { |
1475 | GroupResultPath *pathnode = makeNode(GroupResultPath); |
1476 | |
1477 | pathnode->path.pathtype = T_Result; |
1478 | pathnode->path.parent = rel; |
1479 | pathnode->path.pathtarget = target; |
1480 | pathnode->path.param_info = NULL; /* there are no other rels... */ |
1481 | pathnode->path.parallel_aware = false; |
1482 | pathnode->path.parallel_safe = rel->consider_parallel; |
1483 | pathnode->path.parallel_workers = 0; |
1484 | pathnode->path.pathkeys = NIL; |
1485 | pathnode->quals = havingqual; |
1486 | |
1487 | /* |
1488 | * We can't quite use cost_resultscan() because the quals we want to |
1489 | * account for are not baserestrict quals of the rel. Might as well just |
1490 | * hack it here. |
1491 | */ |
1492 | pathnode->path.rows = 1; |
1493 | pathnode->path.startup_cost = target->cost.startup; |
1494 | pathnode->path.total_cost = target->cost.startup + |
1495 | cpu_tuple_cost + target->cost.per_tuple; |
1496 | |
1497 | /* |
1498 | * Add cost of qual, if any --- but we ignore its selectivity, since our |
1499 | * rowcount estimate should be 1 no matter what the qual is. |
1500 | */ |
1501 | if (havingqual) |
1502 | { |
1503 | QualCost qual_cost; |
1504 | |
1505 | cost_qual_eval(&qual_cost, havingqual, root); |
1506 | /* havingqual is evaluated once at startup */ |
1507 | pathnode->path.startup_cost += qual_cost.startup + qual_cost.per_tuple; |
1508 | pathnode->path.total_cost += qual_cost.startup + qual_cost.per_tuple; |
1509 | } |
1510 | |
1511 | return pathnode; |
1512 | } |
1513 | |
1514 | /* |
1515 | * create_material_path |
1516 | * Creates a path corresponding to a Material plan, returning the |
1517 | * pathnode. |
1518 | */ |
1519 | MaterialPath * |
1520 | create_material_path(RelOptInfo *rel, Path *subpath) |
1521 | { |
1522 | MaterialPath *pathnode = makeNode(MaterialPath); |
1523 | |
1524 | Assert(subpath->parent == rel); |
1525 | |
1526 | pathnode->path.pathtype = T_Material; |
1527 | pathnode->path.parent = rel; |
1528 | pathnode->path.pathtarget = rel->reltarget; |
1529 | pathnode->path.param_info = subpath->param_info; |
1530 | pathnode->path.parallel_aware = false; |
1531 | pathnode->path.parallel_safe = rel->consider_parallel && |
1532 | subpath->parallel_safe; |
1533 | pathnode->path.parallel_workers = subpath->parallel_workers; |
1534 | pathnode->path.pathkeys = subpath->pathkeys; |
1535 | |
1536 | pathnode->subpath = subpath; |
1537 | |
1538 | cost_material(&pathnode->path, |
1539 | subpath->startup_cost, |
1540 | subpath->total_cost, |
1541 | subpath->rows, |
1542 | subpath->pathtarget->width); |
1543 | |
1544 | return pathnode; |
1545 | } |
1546 | |
1547 | /* |
1548 | * create_unique_path |
1549 | * Creates a path representing elimination of distinct rows from the |
1550 | * input data. Distinct-ness is defined according to the needs of the |
1551 | * semijoin represented by sjinfo. If it is not possible to identify |
1552 | * how to make the data unique, NULL is returned. |
1553 | * |
1554 | * If used at all, this is likely to be called repeatedly on the same rel; |
1555 | * and the input subpath should always be the same (the cheapest_total path |
1556 | * for the rel). So we cache the result. |
1557 | */ |
1558 | UniquePath * |
1559 | create_unique_path(PlannerInfo *root, RelOptInfo *rel, Path *subpath, |
1560 | SpecialJoinInfo *sjinfo) |
1561 | { |
1562 | UniquePath *pathnode; |
1563 | Path sort_path; /* dummy for result of cost_sort */ |
1564 | Path agg_path; /* dummy for result of cost_agg */ |
1565 | MemoryContext oldcontext; |
1566 | int numCols; |
1567 | |
1568 | /* Caller made a mistake if subpath isn't cheapest_total ... */ |
1569 | Assert(subpath == rel->cheapest_total_path); |
1570 | Assert(subpath->parent == rel); |
1571 | /* ... or if SpecialJoinInfo is the wrong one */ |
1572 | Assert(sjinfo->jointype == JOIN_SEMI); |
1573 | Assert(bms_equal(rel->relids, sjinfo->syn_righthand)); |
1574 | |
1575 | /* If result already cached, return it */ |
1576 | if (rel->cheapest_unique_path) |
1577 | return (UniquePath *) rel->cheapest_unique_path; |
1578 | |
1579 | /* If it's not possible to unique-ify, return NULL */ |
1580 | if (!(sjinfo->semi_can_btree || sjinfo->semi_can_hash)) |
1581 | return NULL; |
1582 | |
1583 | /* |
1584 | * When called during GEQO join planning, we are in a short-lived memory |
1585 | * context. We must make sure that the path and any subsidiary data |
1586 | * structures created for a baserel survive the GEQO cycle, else the |
1587 | * baserel is trashed for future GEQO cycles. On the other hand, when we |
1588 | * are creating those for a joinrel during GEQO, we don't want them to |
1589 | * clutter the main planning context. Upshot is that the best solution is |
1590 | * to explicitly allocate memory in the same context the given RelOptInfo |
1591 | * is in. |
1592 | */ |
1593 | oldcontext = MemoryContextSwitchTo(GetMemoryChunkContext(rel)); |
1594 | |
1595 | pathnode = makeNode(UniquePath); |
1596 | |
1597 | pathnode->path.pathtype = T_Unique; |
1598 | pathnode->path.parent = rel; |
1599 | pathnode->path.pathtarget = rel->reltarget; |
1600 | pathnode->path.param_info = subpath->param_info; |
1601 | pathnode->path.parallel_aware = false; |
1602 | pathnode->path.parallel_safe = rel->consider_parallel && |
1603 | subpath->parallel_safe; |
1604 | pathnode->path.parallel_workers = subpath->parallel_workers; |
1605 | |
1606 | /* |
1607 | * Assume the output is unsorted, since we don't necessarily have pathkeys |
1608 | * to represent it. (This might get overridden below.) |
1609 | */ |
1610 | pathnode->path.pathkeys = NIL; |
1611 | |
1612 | pathnode->subpath = subpath; |
1613 | pathnode->in_operators = sjinfo->semi_operators; |
1614 | pathnode->uniq_exprs = sjinfo->semi_rhs_exprs; |
1615 | |
1616 | /* |
1617 | * If the input is a relation and it has a unique index that proves the |
1618 | * semi_rhs_exprs are unique, then we don't need to do anything. Note |
1619 | * that relation_has_unique_index_for automatically considers restriction |
1620 | * clauses for the rel, as well. |
1621 | */ |
1622 | if (rel->rtekind == RTE_RELATION && sjinfo->semi_can_btree && |
1623 | relation_has_unique_index_for(root, rel, NIL, |
1624 | sjinfo->semi_rhs_exprs, |
1625 | sjinfo->semi_operators)) |
1626 | { |
1627 | pathnode->umethod = UNIQUE_PATH_NOOP; |
1628 | pathnode->path.rows = rel->rows; |
1629 | pathnode->path.startup_cost = subpath->startup_cost; |
1630 | pathnode->path.total_cost = subpath->total_cost; |
1631 | pathnode->path.pathkeys = subpath->pathkeys; |
1632 | |
1633 | rel->cheapest_unique_path = (Path *) pathnode; |
1634 | |
1635 | MemoryContextSwitchTo(oldcontext); |
1636 | |
1637 | return pathnode; |
1638 | } |
1639 | |
1640 | /* |
1641 | * If the input is a subquery whose output must be unique already, then we |
1642 | * don't need to do anything. The test for uniqueness has to consider |
1643 | * exactly which columns we are extracting; for example "SELECT DISTINCT |
1644 | * x,y" doesn't guarantee that x alone is distinct. So we cannot check for |
1645 | * this optimization unless semi_rhs_exprs consists only of simple Vars |
1646 | * referencing subquery outputs. (Possibly we could do something with |
1647 | * expressions in the subquery outputs, too, but for now keep it simple.) |
1648 | */ |
1649 | if (rel->rtekind == RTE_SUBQUERY) |
1650 | { |
1651 | RangeTblEntry *rte = planner_rt_fetch(rel->relid, root); |
1652 | |
1653 | if (query_supports_distinctness(rte->subquery)) |
1654 | { |
1655 | List *sub_tlist_colnos; |
1656 | |
1657 | sub_tlist_colnos = translate_sub_tlist(sjinfo->semi_rhs_exprs, |
1658 | rel->relid); |
1659 | |
1660 | if (sub_tlist_colnos && |
1661 | query_is_distinct_for(rte->subquery, |
1662 | sub_tlist_colnos, |
1663 | sjinfo->semi_operators)) |
1664 | { |
1665 | pathnode->umethod = UNIQUE_PATH_NOOP; |
1666 | pathnode->path.rows = rel->rows; |
1667 | pathnode->path.startup_cost = subpath->startup_cost; |
1668 | pathnode->path.total_cost = subpath->total_cost; |
1669 | pathnode->path.pathkeys = subpath->pathkeys; |
1670 | |
1671 | rel->cheapest_unique_path = (Path *) pathnode; |
1672 | |
1673 | MemoryContextSwitchTo(oldcontext); |
1674 | |
1675 | return pathnode; |
1676 | } |
1677 | } |
1678 | } |
1679 | |
1680 | /* Estimate number of output rows */ |
1681 | pathnode->path.rows = estimate_num_groups(root, |
1682 | sjinfo->semi_rhs_exprs, |
1683 | rel->rows, |
1684 | NULL); |
1685 | numCols = list_length(sjinfo->semi_rhs_exprs); |
1686 | |
1687 | if (sjinfo->semi_can_btree) |
1688 | { |
1689 | /* |
1690 | * Estimate cost for sort+unique implementation |
1691 | */ |
1692 | cost_sort(&sort_path, root, NIL, |
1693 | subpath->total_cost, |
1694 | rel->rows, |
1695 | subpath->pathtarget->width, |
1696 | 0.0, |
1697 | work_mem, |
1698 | -1.0); |
1699 | |
1700 | /* |
1701 | * Charge one cpu_operator_cost per comparison per input tuple. We |
1702 | * assume all columns get compared at most of the tuples. (XXX |
1703 | * probably this is an overestimate.) This should agree with |
1704 | * create_upper_unique_path. |
1705 | */ |
1706 | sort_path.total_cost += cpu_operator_cost * rel->rows * numCols; |
1707 | } |
1708 | |
1709 | if (sjinfo->semi_can_hash) |
1710 | { |
1711 | /* |
1712 | * Estimate the overhead per hashtable entry at 64 bytes (same as in |
1713 | * planner.c). |
1714 | */ |
1715 | int hashentrysize = subpath->pathtarget->width + 64; |
1716 | |
1717 | if (hashentrysize * pathnode->path.rows > work_mem * 1024L) |
1718 | { |
1719 | /* |
1720 | * We should not try to hash. Hack the SpecialJoinInfo to |
1721 | * remember this, in case we come through here again. |
1722 | */ |
1723 | sjinfo->semi_can_hash = false; |
1724 | } |
1725 | else |
1726 | cost_agg(&agg_path, root, |
1727 | AGG_HASHED, NULL, |
1728 | numCols, pathnode->path.rows, |
1729 | NIL, |
1730 | subpath->startup_cost, |
1731 | subpath->total_cost, |
1732 | rel->rows); |
1733 | } |
1734 | |
1735 | if (sjinfo->semi_can_btree && sjinfo->semi_can_hash) |
1736 | { |
1737 | if (agg_path.total_cost < sort_path.total_cost) |
1738 | pathnode->umethod = UNIQUE_PATH_HASH; |
1739 | else |
1740 | pathnode->umethod = UNIQUE_PATH_SORT; |
1741 | } |
1742 | else if (sjinfo->semi_can_btree) |
1743 | pathnode->umethod = UNIQUE_PATH_SORT; |
1744 | else if (sjinfo->semi_can_hash) |
1745 | pathnode->umethod = UNIQUE_PATH_HASH; |
1746 | else |
1747 | { |
1748 | /* we can get here only if we abandoned hashing above */ |
1749 | MemoryContextSwitchTo(oldcontext); |
1750 | return NULL; |
1751 | } |
1752 | |
1753 | if (pathnode->umethod == UNIQUE_PATH_HASH) |
1754 | { |
1755 | pathnode->path.startup_cost = agg_path.startup_cost; |
1756 | pathnode->path.total_cost = agg_path.total_cost; |
1757 | } |
1758 | else |
1759 | { |
1760 | pathnode->path.startup_cost = sort_path.startup_cost; |
1761 | pathnode->path.total_cost = sort_path.total_cost; |
1762 | } |
1763 | |
1764 | rel->cheapest_unique_path = (Path *) pathnode; |
1765 | |
1766 | MemoryContextSwitchTo(oldcontext); |
1767 | |
1768 | return pathnode; |
1769 | } |
1770 | |
1771 | /* |
1772 | * create_gather_merge_path |
1773 | * |
1774 | * Creates a path corresponding to a gather merge scan, returning |
1775 | * the pathnode. |
1776 | */ |
1777 | GatherMergePath * |
1778 | create_gather_merge_path(PlannerInfo *root, RelOptInfo *rel, Path *subpath, |
1779 | PathTarget *target, List *pathkeys, |
1780 | Relids required_outer, double *rows) |
1781 | { |
1782 | GatherMergePath *pathnode = makeNode(GatherMergePath); |
1783 | Cost input_startup_cost = 0; |
1784 | Cost input_total_cost = 0; |
1785 | |
1786 | Assert(subpath->parallel_safe); |
1787 | Assert(pathkeys); |
1788 | |
1789 | pathnode->path.pathtype = T_GatherMerge; |
1790 | pathnode->path.parent = rel; |
1791 | pathnode->path.param_info = get_baserel_parampathinfo(root, rel, |
1792 | required_outer); |
1793 | pathnode->path.parallel_aware = false; |
1794 | |
1795 | pathnode->subpath = subpath; |
1796 | pathnode->num_workers = subpath->parallel_workers; |
1797 | pathnode->path.pathkeys = pathkeys; |
1798 | pathnode->path.pathtarget = target ? target : rel->reltarget; |
1799 | pathnode->path.rows += subpath->rows; |
1800 | |
1801 | if (pathkeys_contained_in(pathkeys, subpath->pathkeys)) |
1802 | { |
1803 | /* Subpath is adequately ordered, we won't need to sort it */ |
1804 | input_startup_cost += subpath->startup_cost; |
1805 | input_total_cost += subpath->total_cost; |
1806 | } |
1807 | else |
1808 | { |
1809 | /* We'll need to insert a Sort node, so include cost for that */ |
1810 | Path sort_path; /* dummy for result of cost_sort */ |
1811 | |
1812 | cost_sort(&sort_path, |
1813 | root, |
1814 | pathkeys, |
1815 | subpath->total_cost, |
1816 | subpath->rows, |
1817 | subpath->pathtarget->width, |
1818 | 0.0, |
1819 | work_mem, |
1820 | -1); |
1821 | input_startup_cost += sort_path.startup_cost; |
1822 | input_total_cost += sort_path.total_cost; |
1823 | } |
1824 | |
1825 | cost_gather_merge(pathnode, root, rel, pathnode->path.param_info, |
1826 | input_startup_cost, input_total_cost, rows); |
1827 | |
1828 | return pathnode; |
1829 | } |
1830 | |
1831 | /* |
1832 | * translate_sub_tlist - get subquery column numbers represented by tlist |
1833 | * |
1834 | * The given targetlist usually contains only Vars referencing the given relid. |
1835 | * Extract their varattnos (ie, the column numbers of the subquery) and return |
1836 | * as an integer List. |
1837 | * |
1838 | * If any of the tlist items is not a simple Var, we cannot determine whether |
1839 | * the subquery's uniqueness condition (if any) matches ours, so punt and |
1840 | * return NIL. |
1841 | */ |
1842 | static List * |
1843 | translate_sub_tlist(List *tlist, int relid) |
1844 | { |
1845 | List *result = NIL; |
1846 | ListCell *l; |
1847 | |
1848 | foreach(l, tlist) |
1849 | { |
1850 | Var *var = (Var *) lfirst(l); |
1851 | |
1852 | if (!var || !IsA(var, Var) || |
1853 | var->varno != relid) |
1854 | return NIL; /* punt */ |
1855 | |
1856 | result = lappend_int(result, var->varattno); |
1857 | } |
1858 | return result; |
1859 | } |
1860 | |
1861 | /* |
1862 | * create_gather_path |
1863 | * Creates a path corresponding to a gather scan, returning the |
1864 | * pathnode. |
1865 | * |
1866 | * 'rows' may optionally be set to override row estimates from other sources. |
1867 | */ |
1868 | GatherPath * |
1869 | create_gather_path(PlannerInfo *root, RelOptInfo *rel, Path *subpath, |
1870 | PathTarget *target, Relids required_outer, double *rows) |
1871 | { |
1872 | GatherPath *pathnode = makeNode(GatherPath); |
1873 | |
1874 | Assert(subpath->parallel_safe); |
1875 | |
1876 | pathnode->path.pathtype = T_Gather; |
1877 | pathnode->path.parent = rel; |
1878 | pathnode->path.pathtarget = target; |
1879 | pathnode->path.param_info = get_baserel_parampathinfo(root, rel, |
1880 | required_outer); |
1881 | pathnode->path.parallel_aware = false; |
1882 | pathnode->path.parallel_safe = false; |
1883 | pathnode->path.parallel_workers = 0; |
1884 | pathnode->path.pathkeys = NIL; /* Gather has unordered result */ |
1885 | |
1886 | pathnode->subpath = subpath; |
1887 | pathnode->num_workers = subpath->parallel_workers; |
1888 | pathnode->single_copy = false; |
1889 | |
1890 | if (pathnode->num_workers == 0) |
1891 | { |
1892 | pathnode->path.pathkeys = subpath->pathkeys; |
1893 | pathnode->num_workers = 1; |
1894 | pathnode->single_copy = true; |
1895 | } |
1896 | |
1897 | cost_gather(pathnode, root, rel, pathnode->path.param_info, rows); |
1898 | |
1899 | return pathnode; |
1900 | } |
1901 | |
1902 | /* |
1903 | * create_subqueryscan_path |
1904 | * Creates a path corresponding to a scan of a subquery, |
1905 | * returning the pathnode. |
1906 | */ |
1907 | SubqueryScanPath * |
1908 | create_subqueryscan_path(PlannerInfo *root, RelOptInfo *rel, Path *subpath, |
1909 | List *pathkeys, Relids required_outer) |
1910 | { |
1911 | SubqueryScanPath *pathnode = makeNode(SubqueryScanPath); |
1912 | |
1913 | pathnode->path.pathtype = T_SubqueryScan; |
1914 | pathnode->path.parent = rel; |
1915 | pathnode->path.pathtarget = rel->reltarget; |
1916 | pathnode->path.param_info = get_baserel_parampathinfo(root, rel, |
1917 | required_outer); |
1918 | pathnode->path.parallel_aware = false; |
1919 | pathnode->path.parallel_safe = rel->consider_parallel && |
1920 | subpath->parallel_safe; |
1921 | pathnode->path.parallel_workers = subpath->parallel_workers; |
1922 | pathnode->path.pathkeys = pathkeys; |
1923 | pathnode->subpath = subpath; |
1924 | |
1925 | cost_subqueryscan(pathnode, root, rel, pathnode->path.param_info); |
1926 | |
1927 | return pathnode; |
1928 | } |
1929 | |
1930 | /* |
1931 | * create_functionscan_path |
1932 | * Creates a path corresponding to a sequential scan of a function, |
1933 | * returning the pathnode. |
1934 | */ |
1935 | Path * |
1936 | create_functionscan_path(PlannerInfo *root, RelOptInfo *rel, |
1937 | List *pathkeys, Relids required_outer) |
1938 | { |
1939 | Path *pathnode = makeNode(Path); |
1940 | |
1941 | pathnode->pathtype = T_FunctionScan; |
1942 | pathnode->parent = rel; |
1943 | pathnode->pathtarget = rel->reltarget; |
1944 | pathnode->param_info = get_baserel_parampathinfo(root, rel, |
1945 | required_outer); |
1946 | pathnode->parallel_aware = false; |
1947 | pathnode->parallel_safe = rel->consider_parallel; |
1948 | pathnode->parallel_workers = 0; |
1949 | pathnode->pathkeys = pathkeys; |
1950 | |
1951 | cost_functionscan(pathnode, root, rel, pathnode->param_info); |
1952 | |
1953 | return pathnode; |
1954 | } |
1955 | |
1956 | /* |
1957 | * create_tablefuncscan_path |
1958 | * Creates a path corresponding to a sequential scan of a table function, |
1959 | * returning the pathnode. |
1960 | */ |
1961 | Path * |
1962 | create_tablefuncscan_path(PlannerInfo *root, RelOptInfo *rel, |
1963 | Relids required_outer) |
1964 | { |
1965 | Path *pathnode = makeNode(Path); |
1966 | |
1967 | pathnode->pathtype = T_TableFuncScan; |
1968 | pathnode->parent = rel; |
1969 | pathnode->pathtarget = rel->reltarget; |
1970 | pathnode->param_info = get_baserel_parampathinfo(root, rel, |
1971 | required_outer); |
1972 | pathnode->parallel_aware = false; |
1973 | pathnode->parallel_safe = rel->consider_parallel; |
1974 | pathnode->parallel_workers = 0; |
1975 | pathnode->pathkeys = NIL; /* result is always unordered */ |
1976 | |
1977 | cost_tablefuncscan(pathnode, root, rel, pathnode->param_info); |
1978 | |
1979 | return pathnode; |
1980 | } |
1981 | |
1982 | /* |
1983 | * create_valuesscan_path |
1984 | * Creates a path corresponding to a scan of a VALUES list, |
1985 | * returning the pathnode. |
1986 | */ |
1987 | Path * |
1988 | create_valuesscan_path(PlannerInfo *root, RelOptInfo *rel, |
1989 | Relids required_outer) |
1990 | { |
1991 | Path *pathnode = makeNode(Path); |
1992 | |
1993 | pathnode->pathtype = T_ValuesScan; |
1994 | pathnode->parent = rel; |
1995 | pathnode->pathtarget = rel->reltarget; |
1996 | pathnode->param_info = get_baserel_parampathinfo(root, rel, |
1997 | required_outer); |
1998 | pathnode->parallel_aware = false; |
1999 | pathnode->parallel_safe = rel->consider_parallel; |
2000 | pathnode->parallel_workers = 0; |
2001 | pathnode->pathkeys = NIL; /* result is always unordered */ |
2002 | |
2003 | cost_valuesscan(pathnode, root, rel, pathnode->param_info); |
2004 | |
2005 | return pathnode; |
2006 | } |
2007 | |
2008 | /* |
2009 | * create_ctescan_path |
2010 | * Creates a path corresponding to a scan of a non-self-reference CTE, |
2011 | * returning the pathnode. |
2012 | */ |
2013 | Path * |
2014 | create_ctescan_path(PlannerInfo *root, RelOptInfo *rel, Relids required_outer) |
2015 | { |
2016 | Path *pathnode = makeNode(Path); |
2017 | |
2018 | pathnode->pathtype = T_CteScan; |
2019 | pathnode->parent = rel; |
2020 | pathnode->pathtarget = rel->reltarget; |
2021 | pathnode->param_info = get_baserel_parampathinfo(root, rel, |
2022 | required_outer); |
2023 | pathnode->parallel_aware = false; |
2024 | pathnode->parallel_safe = rel->consider_parallel; |
2025 | pathnode->parallel_workers = 0; |
2026 | pathnode->pathkeys = NIL; /* XXX for now, result is always unordered */ |
2027 | |
2028 | cost_ctescan(pathnode, root, rel, pathnode->param_info); |
2029 | |
2030 | return pathnode; |
2031 | } |
2032 | |
2033 | /* |
2034 | * create_namedtuplestorescan_path |
2035 | * Creates a path corresponding to a scan of a named tuplestore, returning |
2036 | * the pathnode. |
2037 | */ |
2038 | Path * |
2039 | create_namedtuplestorescan_path(PlannerInfo *root, RelOptInfo *rel, |
2040 | Relids required_outer) |
2041 | { |
2042 | Path *pathnode = makeNode(Path); |
2043 | |
2044 | pathnode->pathtype = T_NamedTuplestoreScan; |
2045 | pathnode->parent = rel; |
2046 | pathnode->pathtarget = rel->reltarget; |
2047 | pathnode->param_info = get_baserel_parampathinfo(root, rel, |
2048 | required_outer); |
2049 | pathnode->parallel_aware = false; |
2050 | pathnode->parallel_safe = rel->consider_parallel; |
2051 | pathnode->parallel_workers = 0; |
2052 | pathnode->pathkeys = NIL; /* result is always unordered */ |
2053 | |
2054 | cost_namedtuplestorescan(pathnode, root, rel, pathnode->param_info); |
2055 | |
2056 | return pathnode; |
2057 | } |
2058 | |
2059 | /* |
2060 | * create_resultscan_path |
2061 | * Creates a path corresponding to a scan of an RTE_RESULT relation, |
2062 | * returning the pathnode. |
2063 | */ |
2064 | Path * |
2065 | create_resultscan_path(PlannerInfo *root, RelOptInfo *rel, |
2066 | Relids required_outer) |
2067 | { |
2068 | Path *pathnode = makeNode(Path); |
2069 | |
2070 | pathnode->pathtype = T_Result; |
2071 | pathnode->parent = rel; |
2072 | pathnode->pathtarget = rel->reltarget; |
2073 | pathnode->param_info = get_baserel_parampathinfo(root, rel, |
2074 | required_outer); |
2075 | pathnode->parallel_aware = false; |
2076 | pathnode->parallel_safe = rel->consider_parallel; |
2077 | pathnode->parallel_workers = 0; |
2078 | pathnode->pathkeys = NIL; /* result is always unordered */ |
2079 | |
2080 | cost_resultscan(pathnode, root, rel, pathnode->param_info); |
2081 | |
2082 | return pathnode; |
2083 | } |
2084 | |
2085 | /* |
2086 | * create_worktablescan_path |
2087 | * Creates a path corresponding to a scan of a self-reference CTE, |
2088 | * returning the pathnode. |
2089 | */ |
2090 | Path * |
2091 | create_worktablescan_path(PlannerInfo *root, RelOptInfo *rel, |
2092 | Relids required_outer) |
2093 | { |
2094 | Path *pathnode = makeNode(Path); |
2095 | |
2096 | pathnode->pathtype = T_WorkTableScan; |
2097 | pathnode->parent = rel; |
2098 | pathnode->pathtarget = rel->reltarget; |
2099 | pathnode->param_info = get_baserel_parampathinfo(root, rel, |
2100 | required_outer); |
2101 | pathnode->parallel_aware = false; |
2102 | pathnode->parallel_safe = rel->consider_parallel; |
2103 | pathnode->parallel_workers = 0; |
2104 | pathnode->pathkeys = NIL; /* result is always unordered */ |
2105 | |
2106 | /* Cost is the same as for a regular CTE scan */ |
2107 | cost_ctescan(pathnode, root, rel, pathnode->param_info); |
2108 | |
2109 | return pathnode; |
2110 | } |
2111 | |
2112 | /* |
2113 | * create_foreignscan_path |
2114 | * Creates a path corresponding to a scan of a foreign base table, |
2115 | * returning the pathnode. |
2116 | * |
2117 | * This function is never called from core Postgres; rather, it's expected |
2118 | * to be called by the GetForeignPaths function of a foreign data wrapper. |
2119 | * We make the FDW supply all fields of the path, since we do not have any way |
2120 | * to calculate them in core. However, there is a usually-sane default for |
2121 | * the pathtarget (rel->reltarget), so we let a NULL for "target" select that. |
2122 | */ |
2123 | ForeignPath * |
2124 | create_foreignscan_path(PlannerInfo *root, RelOptInfo *rel, |
2125 | PathTarget *target, |
2126 | double rows, Cost startup_cost, Cost total_cost, |
2127 | List *pathkeys, |
2128 | Relids required_outer, |
2129 | Path *fdw_outerpath, |
2130 | List *fdw_private) |
2131 | { |
2132 | ForeignPath *pathnode = makeNode(ForeignPath); |
2133 | |
2134 | /* Historically some FDWs were confused about when to use this */ |
2135 | Assert(IS_SIMPLE_REL(rel)); |
2136 | |
2137 | pathnode->path.pathtype = T_ForeignScan; |
2138 | pathnode->path.parent = rel; |
2139 | pathnode->path.pathtarget = target ? target : rel->reltarget; |
2140 | pathnode->path.param_info = get_baserel_parampathinfo(root, rel, |
2141 | required_outer); |
2142 | pathnode->path.parallel_aware = false; |
2143 | pathnode->path.parallel_safe = rel->consider_parallel; |
2144 | pathnode->path.parallel_workers = 0; |
2145 | pathnode->path.rows = rows; |
2146 | pathnode->path.startup_cost = startup_cost; |
2147 | pathnode->path.total_cost = total_cost; |
2148 | pathnode->path.pathkeys = pathkeys; |
2149 | |
2150 | pathnode->fdw_outerpath = fdw_outerpath; |
2151 | pathnode->fdw_private = fdw_private; |
2152 | |
2153 | return pathnode; |
2154 | } |
2155 | |
2156 | /* |
2157 | * create_foreign_join_path |
2158 | * Creates a path corresponding to a scan of a foreign join, |
2159 | * returning the pathnode. |
2160 | * |
2161 | * This function is never called from core Postgres; rather, it's expected |
2162 | * to be called by the GetForeignJoinPaths function of a foreign data wrapper. |
2163 | * We make the FDW supply all fields of the path, since we do not have any way |
2164 | * to calculate them in core. However, there is a usually-sane default for |
2165 | * the pathtarget (rel->reltarget), so we let a NULL for "target" select that. |
2166 | */ |
2167 | ForeignPath * |
2168 | create_foreign_join_path(PlannerInfo *root, RelOptInfo *rel, |
2169 | PathTarget *target, |
2170 | double rows, Cost startup_cost, Cost total_cost, |
2171 | List *pathkeys, |
2172 | Relids required_outer, |
2173 | Path *fdw_outerpath, |
2174 | List *fdw_private) |
2175 | { |
2176 | ForeignPath *pathnode = makeNode(ForeignPath); |
2177 | |
2178 | /* |
2179 | * We should use get_joinrel_parampathinfo to handle parameterized paths, |
2180 | * but the API of this function doesn't support it, and existing |
2181 | * extensions aren't yet trying to build such paths anyway. For the |
2182 | * moment just throw an error if someone tries it; eventually we should |
2183 | * revisit this. |
2184 | */ |
2185 | if (!bms_is_empty(required_outer) || !bms_is_empty(rel->lateral_relids)) |
2186 | elog(ERROR, "parameterized foreign joins are not supported yet" ); |
2187 | |
2188 | pathnode->path.pathtype = T_ForeignScan; |
2189 | pathnode->path.parent = rel; |
2190 | pathnode->path.pathtarget = target ? target : rel->reltarget; |
2191 | pathnode->path.param_info = NULL; /* XXX see above */ |
2192 | pathnode->path.parallel_aware = false; |
2193 | pathnode->path.parallel_safe = rel->consider_parallel; |
2194 | pathnode->path.parallel_workers = 0; |
2195 | pathnode->path.rows = rows; |
2196 | pathnode->path.startup_cost = startup_cost; |
2197 | pathnode->path.total_cost = total_cost; |
2198 | pathnode->path.pathkeys = pathkeys; |
2199 | |
2200 | pathnode->fdw_outerpath = fdw_outerpath; |
2201 | pathnode->fdw_private = fdw_private; |
2202 | |
2203 | return pathnode; |
2204 | } |
2205 | |
2206 | /* |
2207 | * create_foreign_upper_path |
2208 | * Creates a path corresponding to an upper relation that's computed |
2209 | * directly by an FDW, returning the pathnode. |
2210 | * |
2211 | * This function is never called from core Postgres; rather, it's expected to |
2212 | * be called by the GetForeignUpperPaths function of a foreign data wrapper. |
2213 | * We make the FDW supply all fields of the path, since we do not have any way |
2214 | * to calculate them in core. However, there is a usually-sane default for |
2215 | * the pathtarget (rel->reltarget), so we let a NULL for "target" select that. |
2216 | */ |
2217 | ForeignPath * |
2218 | create_foreign_upper_path(PlannerInfo *root, RelOptInfo *rel, |
2219 | PathTarget *target, |
2220 | double rows, Cost startup_cost, Cost total_cost, |
2221 | List *pathkeys, |
2222 | Path *fdw_outerpath, |
2223 | List *fdw_private) |
2224 | { |
2225 | ForeignPath *pathnode = makeNode(ForeignPath); |
2226 | |
2227 | /* |
2228 | * Upper relations should never have any lateral references, since joining |
2229 | * is complete. |
2230 | */ |
2231 | Assert(bms_is_empty(rel->lateral_relids)); |
2232 | |
2233 | pathnode->path.pathtype = T_ForeignScan; |
2234 | pathnode->path.parent = rel; |
2235 | pathnode->path.pathtarget = target ? target : rel->reltarget; |
2236 | pathnode->path.param_info = NULL; |
2237 | pathnode->path.parallel_aware = false; |
2238 | pathnode->path.parallel_safe = rel->consider_parallel; |
2239 | pathnode->path.parallel_workers = 0; |
2240 | pathnode->path.rows = rows; |
2241 | pathnode->path.startup_cost = startup_cost; |
2242 | pathnode->path.total_cost = total_cost; |
2243 | pathnode->path.pathkeys = pathkeys; |
2244 | |
2245 | pathnode->fdw_outerpath = fdw_outerpath; |
2246 | pathnode->fdw_private = fdw_private; |
2247 | |
2248 | return pathnode; |
2249 | } |
2250 | |
2251 | /* |
2252 | * calc_nestloop_required_outer |
2253 | * Compute the required_outer set for a nestloop join path |
2254 | * |
2255 | * Note: result must not share storage with either input |
2256 | */ |
2257 | Relids |
2258 | calc_nestloop_required_outer(Relids outerrelids, |
2259 | Relids outer_paramrels, |
2260 | Relids innerrelids, |
2261 | Relids inner_paramrels) |
2262 | { |
2263 | Relids required_outer; |
2264 | |
2265 | /* inner_path can require rels from outer path, but not vice versa */ |
2266 | Assert(!bms_overlap(outer_paramrels, innerrelids)); |
2267 | /* easy case if inner path is not parameterized */ |
2268 | if (!inner_paramrels) |
2269 | return bms_copy(outer_paramrels); |
2270 | /* else, form the union ... */ |
2271 | required_outer = bms_union(outer_paramrels, inner_paramrels); |
2272 | /* ... and remove any mention of now-satisfied outer rels */ |
2273 | required_outer = bms_del_members(required_outer, |
2274 | outerrelids); |
2275 | /* maintain invariant that required_outer is exactly NULL if empty */ |
2276 | if (bms_is_empty(required_outer)) |
2277 | { |
2278 | bms_free(required_outer); |
2279 | required_outer = NULL; |
2280 | } |
2281 | return required_outer; |
2282 | } |
2283 | |
2284 | /* |
2285 | * calc_non_nestloop_required_outer |
2286 | * Compute the required_outer set for a merge or hash join path |
2287 | * |
2288 | * Note: result must not share storage with either input |
2289 | */ |
2290 | Relids |
2291 | calc_non_nestloop_required_outer(Path *outer_path, Path *inner_path) |
2292 | { |
2293 | Relids outer_paramrels = PATH_REQ_OUTER(outer_path); |
2294 | Relids inner_paramrels = PATH_REQ_OUTER(inner_path); |
2295 | Relids required_outer; |
2296 | |
2297 | /* neither path can require rels from the other */ |
2298 | Assert(!bms_overlap(outer_paramrels, inner_path->parent->relids)); |
2299 | Assert(!bms_overlap(inner_paramrels, outer_path->parent->relids)); |
2300 | /* form the union ... */ |
2301 | required_outer = bms_union(outer_paramrels, inner_paramrels); |
2302 | /* we do not need an explicit test for empty; bms_union gets it right */ |
2303 | return required_outer; |
2304 | } |
2305 | |
2306 | /* |
2307 | * create_nestloop_path |
2308 | * Creates a pathnode corresponding to a nestloop join between two |
2309 | * relations. |
2310 | * |
2311 | * 'joinrel' is the join relation. |
2312 | * 'jointype' is the type of join required |
2313 | * 'workspace' is the result from initial_cost_nestloop |
2314 | * 'extra' contains various information about the join |
2315 | * 'outer_path' is the outer path |
2316 | * 'inner_path' is the inner path |
2317 | * 'restrict_clauses' are the RestrictInfo nodes to apply at the join |
2318 | * 'pathkeys' are the path keys of the new join path |
2319 | * 'required_outer' is the set of required outer rels |
2320 | * |
2321 | * Returns the resulting path node. |
2322 | */ |
2323 | NestPath * |
2324 | create_nestloop_path(PlannerInfo *root, |
2325 | RelOptInfo *joinrel, |
2326 | JoinType jointype, |
2327 | JoinCostWorkspace *workspace, |
2328 | JoinPathExtraData *, |
2329 | Path *outer_path, |
2330 | Path *inner_path, |
2331 | List *restrict_clauses, |
2332 | List *pathkeys, |
2333 | Relids required_outer) |
2334 | { |
2335 | NestPath *pathnode = makeNode(NestPath); |
2336 | Relids inner_req_outer = PATH_REQ_OUTER(inner_path); |
2337 | |
2338 | /* |
2339 | * If the inner path is parameterized by the outer, we must drop any |
2340 | * restrict_clauses that are due to be moved into the inner path. We have |
2341 | * to do this now, rather than postpone the work till createplan time, |
2342 | * because the restrict_clauses list can affect the size and cost |
2343 | * estimates for this path. |
2344 | */ |
2345 | if (bms_overlap(inner_req_outer, outer_path->parent->relids)) |
2346 | { |
2347 | Relids inner_and_outer = bms_union(inner_path->parent->relids, |
2348 | inner_req_outer); |
2349 | List *jclauses = NIL; |
2350 | ListCell *lc; |
2351 | |
2352 | foreach(lc, restrict_clauses) |
2353 | { |
2354 | RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc); |
2355 | |
2356 | if (!join_clause_is_movable_into(rinfo, |
2357 | inner_path->parent->relids, |
2358 | inner_and_outer)) |
2359 | jclauses = lappend(jclauses, rinfo); |
2360 | } |
2361 | restrict_clauses = jclauses; |
2362 | } |
2363 | |
2364 | pathnode->path.pathtype = T_NestLoop; |
2365 | pathnode->path.parent = joinrel; |
2366 | pathnode->path.pathtarget = joinrel->reltarget; |
2367 | pathnode->path.param_info = |
2368 | get_joinrel_parampathinfo(root, |
2369 | joinrel, |
2370 | outer_path, |
2371 | inner_path, |
2372 | extra->sjinfo, |
2373 | required_outer, |
2374 | &restrict_clauses); |
2375 | pathnode->path.parallel_aware = false; |
2376 | pathnode->path.parallel_safe = joinrel->consider_parallel && |
2377 | outer_path->parallel_safe && inner_path->parallel_safe; |
2378 | /* This is a foolish way to estimate parallel_workers, but for now... */ |
2379 | pathnode->path.parallel_workers = outer_path->parallel_workers; |
2380 | pathnode->path.pathkeys = pathkeys; |
2381 | pathnode->jointype = jointype; |
2382 | pathnode->inner_unique = extra->inner_unique; |
2383 | pathnode->outerjoinpath = outer_path; |
2384 | pathnode->innerjoinpath = inner_path; |
2385 | pathnode->joinrestrictinfo = restrict_clauses; |
2386 | |
2387 | final_cost_nestloop(root, pathnode, workspace, extra); |
2388 | |
2389 | return pathnode; |
2390 | } |
2391 | |
2392 | /* |
2393 | * create_mergejoin_path |
2394 | * Creates a pathnode corresponding to a mergejoin join between |
2395 | * two relations |
2396 | * |
2397 | * 'joinrel' is the join relation |
2398 | * 'jointype' is the type of join required |
2399 | * 'workspace' is the result from initial_cost_mergejoin |
2400 | * 'extra' contains various information about the join |
2401 | * 'outer_path' is the outer path |
2402 | * 'inner_path' is the inner path |
2403 | * 'restrict_clauses' are the RestrictInfo nodes to apply at the join |
2404 | * 'pathkeys' are the path keys of the new join path |
2405 | * 'required_outer' is the set of required outer rels |
2406 | * 'mergeclauses' are the RestrictInfo nodes to use as merge clauses |
2407 | * (this should be a subset of the restrict_clauses list) |
2408 | * 'outersortkeys' are the sort varkeys for the outer relation |
2409 | * 'innersortkeys' are the sort varkeys for the inner relation |
2410 | */ |
2411 | MergePath * |
2412 | create_mergejoin_path(PlannerInfo *root, |
2413 | RelOptInfo *joinrel, |
2414 | JoinType jointype, |
2415 | JoinCostWorkspace *workspace, |
2416 | JoinPathExtraData *, |
2417 | Path *outer_path, |
2418 | Path *inner_path, |
2419 | List *restrict_clauses, |
2420 | List *pathkeys, |
2421 | Relids required_outer, |
2422 | List *mergeclauses, |
2423 | List *outersortkeys, |
2424 | List *innersortkeys) |
2425 | { |
2426 | MergePath *pathnode = makeNode(MergePath); |
2427 | |
2428 | pathnode->jpath.path.pathtype = T_MergeJoin; |
2429 | pathnode->jpath.path.parent = joinrel; |
2430 | pathnode->jpath.path.pathtarget = joinrel->reltarget; |
2431 | pathnode->jpath.path.param_info = |
2432 | get_joinrel_parampathinfo(root, |
2433 | joinrel, |
2434 | outer_path, |
2435 | inner_path, |
2436 | extra->sjinfo, |
2437 | required_outer, |
2438 | &restrict_clauses); |
2439 | pathnode->jpath.path.parallel_aware = false; |
2440 | pathnode->jpath.path.parallel_safe = joinrel->consider_parallel && |
2441 | outer_path->parallel_safe && inner_path->parallel_safe; |
2442 | /* This is a foolish way to estimate parallel_workers, but for now... */ |
2443 | pathnode->jpath.path.parallel_workers = outer_path->parallel_workers; |
2444 | pathnode->jpath.path.pathkeys = pathkeys; |
2445 | pathnode->jpath.jointype = jointype; |
2446 | pathnode->jpath.inner_unique = extra->inner_unique; |
2447 | pathnode->jpath.outerjoinpath = outer_path; |
2448 | pathnode->jpath.innerjoinpath = inner_path; |
2449 | pathnode->jpath.joinrestrictinfo = restrict_clauses; |
2450 | pathnode->path_mergeclauses = mergeclauses; |
2451 | pathnode->outersortkeys = outersortkeys; |
2452 | pathnode->innersortkeys = innersortkeys; |
2453 | /* pathnode->skip_mark_restore will be set by final_cost_mergejoin */ |
2454 | /* pathnode->materialize_inner will be set by final_cost_mergejoin */ |
2455 | |
2456 | final_cost_mergejoin(root, pathnode, workspace, extra); |
2457 | |
2458 | return pathnode; |
2459 | } |
2460 | |
2461 | /* |
2462 | * create_hashjoin_path |
2463 | * Creates a pathnode corresponding to a hash join between two relations. |
2464 | * |
2465 | * 'joinrel' is the join relation |
2466 | * 'jointype' is the type of join required |
2467 | * 'workspace' is the result from initial_cost_hashjoin |
2468 | * 'extra' contains various information about the join |
2469 | * 'outer_path' is the cheapest outer path |
2470 | * 'inner_path' is the cheapest inner path |
2471 | * 'parallel_hash' to select Parallel Hash of inner path (shared hash table) |
2472 | * 'restrict_clauses' are the RestrictInfo nodes to apply at the join |
2473 | * 'required_outer' is the set of required outer rels |
2474 | * 'hashclauses' are the RestrictInfo nodes to use as hash clauses |
2475 | * (this should be a subset of the restrict_clauses list) |
2476 | */ |
2477 | HashPath * |
2478 | create_hashjoin_path(PlannerInfo *root, |
2479 | RelOptInfo *joinrel, |
2480 | JoinType jointype, |
2481 | JoinCostWorkspace *workspace, |
2482 | JoinPathExtraData *, |
2483 | Path *outer_path, |
2484 | Path *inner_path, |
2485 | bool parallel_hash, |
2486 | List *restrict_clauses, |
2487 | Relids required_outer, |
2488 | List *hashclauses) |
2489 | { |
2490 | HashPath *pathnode = makeNode(HashPath); |
2491 | |
2492 | pathnode->jpath.path.pathtype = T_HashJoin; |
2493 | pathnode->jpath.path.parent = joinrel; |
2494 | pathnode->jpath.path.pathtarget = joinrel->reltarget; |
2495 | pathnode->jpath.path.param_info = |
2496 | get_joinrel_parampathinfo(root, |
2497 | joinrel, |
2498 | outer_path, |
2499 | inner_path, |
2500 | extra->sjinfo, |
2501 | required_outer, |
2502 | &restrict_clauses); |
2503 | pathnode->jpath.path.parallel_aware = |
2504 | joinrel->consider_parallel && parallel_hash; |
2505 | pathnode->jpath.path.parallel_safe = joinrel->consider_parallel && |
2506 | outer_path->parallel_safe && inner_path->parallel_safe; |
2507 | /* This is a foolish way to estimate parallel_workers, but for now... */ |
2508 | pathnode->jpath.path.parallel_workers = outer_path->parallel_workers; |
2509 | |
2510 | /* |
2511 | * A hashjoin never has pathkeys, since its output ordering is |
2512 | * unpredictable due to possible batching. XXX If the inner relation is |
2513 | * small enough, we could instruct the executor that it must not batch, |
2514 | * and then we could assume that the output inherits the outer relation's |
2515 | * ordering, which might save a sort step. However there is considerable |
2516 | * downside if our estimate of the inner relation size is badly off. For |
2517 | * the moment we don't risk it. (Note also that if we wanted to take this |
2518 | * seriously, joinpath.c would have to consider many more paths for the |
2519 | * outer rel than it does now.) |
2520 | */ |
2521 | pathnode->jpath.path.pathkeys = NIL; |
2522 | pathnode->jpath.jointype = jointype; |
2523 | pathnode->jpath.inner_unique = extra->inner_unique; |
2524 | pathnode->jpath.outerjoinpath = outer_path; |
2525 | pathnode->jpath.innerjoinpath = inner_path; |
2526 | pathnode->jpath.joinrestrictinfo = restrict_clauses; |
2527 | pathnode->path_hashclauses = hashclauses; |
2528 | /* final_cost_hashjoin will fill in pathnode->num_batches */ |
2529 | |
2530 | final_cost_hashjoin(root, pathnode, workspace, extra); |
2531 | |
2532 | return pathnode; |
2533 | } |
2534 | |
2535 | /* |
2536 | * create_projection_path |
2537 | * Creates a pathnode that represents performing a projection. |
2538 | * |
2539 | * 'rel' is the parent relation associated with the result |
2540 | * 'subpath' is the path representing the source of data |
2541 | * 'target' is the PathTarget to be computed |
2542 | */ |
2543 | ProjectionPath * |
2544 | create_projection_path(PlannerInfo *root, |
2545 | RelOptInfo *rel, |
2546 | Path *subpath, |
2547 | PathTarget *target) |
2548 | { |
2549 | ProjectionPath *pathnode = makeNode(ProjectionPath); |
2550 | PathTarget *oldtarget = subpath->pathtarget; |
2551 | |
2552 | pathnode->path.pathtype = T_Result; |
2553 | pathnode->path.parent = rel; |
2554 | pathnode->path.pathtarget = target; |
2555 | /* For now, assume we are above any joins, so no parameterization */ |
2556 | pathnode->path.param_info = NULL; |
2557 | pathnode->path.parallel_aware = false; |
2558 | pathnode->path.parallel_safe = rel->consider_parallel && |
2559 | subpath->parallel_safe && |
2560 | is_parallel_safe(root, (Node *) target->exprs); |
2561 | pathnode->path.parallel_workers = subpath->parallel_workers; |
2562 | /* Projection does not change the sort order */ |
2563 | pathnode->path.pathkeys = subpath->pathkeys; |
2564 | |
2565 | pathnode->subpath = subpath; |
2566 | |
2567 | /* |
2568 | * We might not need a separate Result node. If the input plan node type |
2569 | * can project, we can just tell it to project something else. Or, if it |
2570 | * can't project but the desired target has the same expression list as |
2571 | * what the input will produce anyway, we can still give it the desired |
2572 | * tlist (possibly changing its ressortgroupref labels, but nothing else). |
2573 | * Note: in the latter case, create_projection_plan has to recheck our |
2574 | * conclusion; see comments therein. |
2575 | */ |
2576 | if (is_projection_capable_path(subpath) || |
2577 | equal(oldtarget->exprs, target->exprs)) |
2578 | { |
2579 | /* No separate Result node needed */ |
2580 | pathnode->dummypp = true; |
2581 | |
2582 | /* |
2583 | * Set cost of plan as subpath's cost, adjusted for tlist replacement. |
2584 | */ |
2585 | pathnode->path.rows = subpath->rows; |
2586 | pathnode->path.startup_cost = subpath->startup_cost + |
2587 | (target->cost.startup - oldtarget->cost.startup); |
2588 | pathnode->path.total_cost = subpath->total_cost + |
2589 | (target->cost.startup - oldtarget->cost.startup) + |
2590 | (target->cost.per_tuple - oldtarget->cost.per_tuple) * subpath->rows; |
2591 | } |
2592 | else |
2593 | { |
2594 | /* We really do need the Result node */ |
2595 | pathnode->dummypp = false; |
2596 | |
2597 | /* |
2598 | * The Result node's cost is cpu_tuple_cost per row, plus the cost of |
2599 | * evaluating the tlist. There is no qual to worry about. |
2600 | */ |
2601 | pathnode->path.rows = subpath->rows; |
2602 | pathnode->path.startup_cost = subpath->startup_cost + |
2603 | target->cost.startup; |
2604 | pathnode->path.total_cost = subpath->total_cost + |
2605 | target->cost.startup + |
2606 | (cpu_tuple_cost + target->cost.per_tuple) * subpath->rows; |
2607 | } |
2608 | |
2609 | return pathnode; |
2610 | } |
2611 | |
2612 | /* |
2613 | * apply_projection_to_path |
2614 | * Add a projection step, or just apply the target directly to given path. |
2615 | * |
2616 | * This has the same net effect as create_projection_path(), except that if |
2617 | * a separate Result plan node isn't needed, we just replace the given path's |
2618 | * pathtarget with the desired one. This must be used only when the caller |
2619 | * knows that the given path isn't referenced elsewhere and so can be modified |
2620 | * in-place. |
2621 | * |
2622 | * If the input path is a GatherPath or GatherMergePath, we try to push the |
2623 | * new target down to its input as well; this is a yet more invasive |
2624 | * modification of the input path, which create_projection_path() can't do. |
2625 | * |
2626 | * Note that we mustn't change the source path's parent link; so when it is |
2627 | * add_path'd to "rel" things will be a bit inconsistent. So far that has |
2628 | * not caused any trouble. |
2629 | * |
2630 | * 'rel' is the parent relation associated with the result |
2631 | * 'path' is the path representing the source of data |
2632 | * 'target' is the PathTarget to be computed |
2633 | */ |
2634 | Path * |
2635 | apply_projection_to_path(PlannerInfo *root, |
2636 | RelOptInfo *rel, |
2637 | Path *path, |
2638 | PathTarget *target) |
2639 | { |
2640 | QualCost oldcost; |
2641 | |
2642 | /* |
2643 | * If given path can't project, we might need a Result node, so make a |
2644 | * separate ProjectionPath. |
2645 | */ |
2646 | if (!is_projection_capable_path(path)) |
2647 | return (Path *) create_projection_path(root, rel, path, target); |
2648 | |
2649 | /* |
2650 | * We can just jam the desired tlist into the existing path, being sure to |
2651 | * update its cost estimates appropriately. |
2652 | */ |
2653 | oldcost = path->pathtarget->cost; |
2654 | path->pathtarget = target; |
2655 | |
2656 | path->startup_cost += target->cost.startup - oldcost.startup; |
2657 | path->total_cost += target->cost.startup - oldcost.startup + |
2658 | (target->cost.per_tuple - oldcost.per_tuple) * path->rows; |
2659 | |
2660 | /* |
2661 | * If the path happens to be a Gather or GatherMerge path, we'd like to |
2662 | * arrange for the subpath to return the required target list so that |
2663 | * workers can help project. But if there is something that is not |
2664 | * parallel-safe in the target expressions, then we can't. |
2665 | */ |
2666 | if ((IsA(path, GatherPath) ||IsA(path, GatherMergePath)) && |
2667 | is_parallel_safe(root, (Node *) target->exprs)) |
2668 | { |
2669 | /* |
2670 | * We always use create_projection_path here, even if the subpath is |
2671 | * projection-capable, so as to avoid modifying the subpath in place. |
2672 | * It seems unlikely at present that there could be any other |
2673 | * references to the subpath, but better safe than sorry. |
2674 | * |
2675 | * Note that we don't change the parallel path's cost estimates; it |
2676 | * might be appropriate to do so, to reflect the fact that the bulk of |
2677 | * the target evaluation will happen in workers. |
2678 | */ |
2679 | if (IsA(path, GatherPath)) |
2680 | { |
2681 | GatherPath *gpath = (GatherPath *) path; |
2682 | |
2683 | gpath->subpath = (Path *) |
2684 | create_projection_path(root, |
2685 | gpath->subpath->parent, |
2686 | gpath->subpath, |
2687 | target); |
2688 | } |
2689 | else |
2690 | { |
2691 | GatherMergePath *gmpath = (GatherMergePath *) path; |
2692 | |
2693 | gmpath->subpath = (Path *) |
2694 | create_projection_path(root, |
2695 | gmpath->subpath->parent, |
2696 | gmpath->subpath, |
2697 | target); |
2698 | } |
2699 | } |
2700 | else if (path->parallel_safe && |
2701 | !is_parallel_safe(root, (Node *) target->exprs)) |
2702 | { |
2703 | /* |
2704 | * We're inserting a parallel-restricted target list into a path |
2705 | * currently marked parallel-safe, so we have to mark it as no longer |
2706 | * safe. |
2707 | */ |
2708 | path->parallel_safe = false; |
2709 | } |
2710 | |
2711 | return path; |
2712 | } |
2713 | |
2714 | /* |
2715 | * create_set_projection_path |
2716 | * Creates a pathnode that represents performing a projection that |
2717 | * includes set-returning functions. |
2718 | * |
2719 | * 'rel' is the parent relation associated with the result |
2720 | * 'subpath' is the path representing the source of data |
2721 | * 'target' is the PathTarget to be computed |
2722 | */ |
2723 | ProjectSetPath * |
2724 | create_set_projection_path(PlannerInfo *root, |
2725 | RelOptInfo *rel, |
2726 | Path *subpath, |
2727 | PathTarget *target) |
2728 | { |
2729 | ProjectSetPath *pathnode = makeNode(ProjectSetPath); |
2730 | double tlist_rows; |
2731 | ListCell *lc; |
2732 | |
2733 | pathnode->path.pathtype = T_ProjectSet; |
2734 | pathnode->path.parent = rel; |
2735 | pathnode->path.pathtarget = target; |
2736 | /* For now, assume we are above any joins, so no parameterization */ |
2737 | pathnode->path.param_info = NULL; |
2738 | pathnode->path.parallel_aware = false; |
2739 | pathnode->path.parallel_safe = rel->consider_parallel && |
2740 | subpath->parallel_safe && |
2741 | is_parallel_safe(root, (Node *) target->exprs); |
2742 | pathnode->path.parallel_workers = subpath->parallel_workers; |
2743 | /* Projection does not change the sort order XXX? */ |
2744 | pathnode->path.pathkeys = subpath->pathkeys; |
2745 | |
2746 | pathnode->subpath = subpath; |
2747 | |
2748 | /* |
2749 | * Estimate number of rows produced by SRFs for each row of input; if |
2750 | * there's more than one in this node, use the maximum. |
2751 | */ |
2752 | tlist_rows = 1; |
2753 | foreach(lc, target->exprs) |
2754 | { |
2755 | Node *node = (Node *) lfirst(lc); |
2756 | double itemrows; |
2757 | |
2758 | itemrows = expression_returns_set_rows(root, node); |
2759 | if (tlist_rows < itemrows) |
2760 | tlist_rows = itemrows; |
2761 | } |
2762 | |
2763 | /* |
2764 | * In addition to the cost of evaluating the tlist, charge cpu_tuple_cost |
2765 | * per input row, and half of cpu_tuple_cost for each added output row. |
2766 | * This is slightly bizarre maybe, but it's what 9.6 did; we may revisit |
2767 | * this estimate later. |
2768 | */ |
2769 | pathnode->path.rows = subpath->rows * tlist_rows; |
2770 | pathnode->path.startup_cost = subpath->startup_cost + |
2771 | target->cost.startup; |
2772 | pathnode->path.total_cost = subpath->total_cost + |
2773 | target->cost.startup + |
2774 | (cpu_tuple_cost + target->cost.per_tuple) * subpath->rows + |
2775 | (pathnode->path.rows - subpath->rows) * cpu_tuple_cost / 2; |
2776 | |
2777 | return pathnode; |
2778 | } |
2779 | |
2780 | /* |
2781 | * create_sort_path |
2782 | * Creates a pathnode that represents performing an explicit sort. |
2783 | * |
2784 | * 'rel' is the parent relation associated with the result |
2785 | * 'subpath' is the path representing the source of data |
2786 | * 'pathkeys' represents the desired sort order |
2787 | * 'limit_tuples' is the estimated bound on the number of output tuples, |
2788 | * or -1 if no LIMIT or couldn't estimate |
2789 | */ |
2790 | SortPath * |
2791 | create_sort_path(PlannerInfo *root, |
2792 | RelOptInfo *rel, |
2793 | Path *subpath, |
2794 | List *pathkeys, |
2795 | double limit_tuples) |
2796 | { |
2797 | SortPath *pathnode = makeNode(SortPath); |
2798 | |
2799 | pathnode->path.pathtype = T_Sort; |
2800 | pathnode->path.parent = rel; |
2801 | /* Sort doesn't project, so use source path's pathtarget */ |
2802 | pathnode->path.pathtarget = subpath->pathtarget; |
2803 | /* For now, assume we are above any joins, so no parameterization */ |
2804 | pathnode->path.param_info = NULL; |
2805 | pathnode->path.parallel_aware = false; |
2806 | pathnode->path.parallel_safe = rel->consider_parallel && |
2807 | subpath->parallel_safe; |
2808 | pathnode->path.parallel_workers = subpath->parallel_workers; |
2809 | pathnode->path.pathkeys = pathkeys; |
2810 | |
2811 | pathnode->subpath = subpath; |
2812 | |
2813 | cost_sort(&pathnode->path, root, pathkeys, |
2814 | subpath->total_cost, |
2815 | subpath->rows, |
2816 | subpath->pathtarget->width, |
2817 | 0.0, /* XXX comparison_cost shouldn't be 0? */ |
2818 | work_mem, limit_tuples); |
2819 | |
2820 | return pathnode; |
2821 | } |
2822 | |
2823 | /* |
2824 | * create_group_path |
2825 | * Creates a pathnode that represents performing grouping of presorted input |
2826 | * |
2827 | * 'rel' is the parent relation associated with the result |
2828 | * 'subpath' is the path representing the source of data |
2829 | * 'target' is the PathTarget to be computed |
2830 | * 'groupClause' is a list of SortGroupClause's representing the grouping |
2831 | * 'qual' is the HAVING quals if any |
2832 | * 'numGroups' is the estimated number of groups |
2833 | */ |
2834 | GroupPath * |
2835 | create_group_path(PlannerInfo *root, |
2836 | RelOptInfo *rel, |
2837 | Path *subpath, |
2838 | List *groupClause, |
2839 | List *qual, |
2840 | double numGroups) |
2841 | { |
2842 | GroupPath *pathnode = makeNode(GroupPath); |
2843 | PathTarget *target = rel->reltarget; |
2844 | |
2845 | pathnode->path.pathtype = T_Group; |
2846 | pathnode->path.parent = rel; |
2847 | pathnode->path.pathtarget = target; |
2848 | /* For now, assume we are above any joins, so no parameterization */ |
2849 | pathnode->path.param_info = NULL; |
2850 | pathnode->path.parallel_aware = false; |
2851 | pathnode->path.parallel_safe = rel->consider_parallel && |
2852 | subpath->parallel_safe; |
2853 | pathnode->path.parallel_workers = subpath->parallel_workers; |
2854 | /* Group doesn't change sort ordering */ |
2855 | pathnode->path.pathkeys = subpath->pathkeys; |
2856 | |
2857 | pathnode->subpath = subpath; |
2858 | |
2859 | pathnode->groupClause = groupClause; |
2860 | pathnode->qual = qual; |
2861 | |
2862 | cost_group(&pathnode->path, root, |
2863 | list_length(groupClause), |
2864 | numGroups, |
2865 | qual, |
2866 | subpath->startup_cost, subpath->total_cost, |
2867 | subpath->rows); |
2868 | |
2869 | /* add tlist eval cost for each output row */ |
2870 | pathnode->path.startup_cost += target->cost.startup; |
2871 | pathnode->path.total_cost += target->cost.startup + |
2872 | target->cost.per_tuple * pathnode->path.rows; |
2873 | |
2874 | return pathnode; |
2875 | } |
2876 | |
2877 | /* |
2878 | * create_upper_unique_path |
2879 | * Creates a pathnode that represents performing an explicit Unique step |
2880 | * on presorted input. |
2881 | * |
2882 | * This produces a Unique plan node, but the use-case is so different from |
2883 | * create_unique_path that it doesn't seem worth trying to merge the two. |
2884 | * |
2885 | * 'rel' is the parent relation associated with the result |
2886 | * 'subpath' is the path representing the source of data |
2887 | * 'numCols' is the number of grouping columns |
2888 | * 'numGroups' is the estimated number of groups |
2889 | * |
2890 | * The input path must be sorted on the grouping columns, plus possibly |
2891 | * additional columns; so the first numCols pathkeys are the grouping columns |
2892 | */ |
2893 | UpperUniquePath * |
2894 | create_upper_unique_path(PlannerInfo *root, |
2895 | RelOptInfo *rel, |
2896 | Path *subpath, |
2897 | int numCols, |
2898 | double numGroups) |
2899 | { |
2900 | UpperUniquePath *pathnode = makeNode(UpperUniquePath); |
2901 | |
2902 | pathnode->path.pathtype = T_Unique; |
2903 | pathnode->path.parent = rel; |
2904 | /* Unique doesn't project, so use source path's pathtarget */ |
2905 | pathnode->path.pathtarget = subpath->pathtarget; |
2906 | /* For now, assume we are above any joins, so no parameterization */ |
2907 | pathnode->path.param_info = NULL; |
2908 | pathnode->path.parallel_aware = false; |
2909 | pathnode->path.parallel_safe = rel->consider_parallel && |
2910 | subpath->parallel_safe; |
2911 | pathnode->path.parallel_workers = subpath->parallel_workers; |
2912 | /* Unique doesn't change the input ordering */ |
2913 | pathnode->path.pathkeys = subpath->pathkeys; |
2914 | |
2915 | pathnode->subpath = subpath; |
2916 | pathnode->numkeys = numCols; |
2917 | |
2918 | /* |
2919 | * Charge one cpu_operator_cost per comparison per input tuple. We assume |
2920 | * all columns get compared at most of the tuples. (XXX probably this is |
2921 | * an overestimate.) |
2922 | */ |
2923 | pathnode->path.startup_cost = subpath->startup_cost; |
2924 | pathnode->path.total_cost = subpath->total_cost + |
2925 | cpu_operator_cost * subpath->rows * numCols; |
2926 | pathnode->path.rows = numGroups; |
2927 | |
2928 | return pathnode; |
2929 | } |
2930 | |
2931 | /* |
2932 | * create_agg_path |
2933 | * Creates a pathnode that represents performing aggregation/grouping |
2934 | * |
2935 | * 'rel' is the parent relation associated with the result |
2936 | * 'subpath' is the path representing the source of data |
2937 | * 'target' is the PathTarget to be computed |
2938 | * 'aggstrategy' is the Agg node's basic implementation strategy |
2939 | * 'aggsplit' is the Agg node's aggregate-splitting mode |
2940 | * 'groupClause' is a list of SortGroupClause's representing the grouping |
2941 | * 'qual' is the HAVING quals if any |
2942 | * 'aggcosts' contains cost info about the aggregate functions to be computed |
2943 | * 'numGroups' is the estimated number of groups (1 if not grouping) |
2944 | */ |
2945 | AggPath * |
2946 | create_agg_path(PlannerInfo *root, |
2947 | RelOptInfo *rel, |
2948 | Path *subpath, |
2949 | PathTarget *target, |
2950 | AggStrategy aggstrategy, |
2951 | AggSplit aggsplit, |
2952 | List *groupClause, |
2953 | List *qual, |
2954 | const AggClauseCosts *aggcosts, |
2955 | double numGroups) |
2956 | { |
2957 | AggPath *pathnode = makeNode(AggPath); |
2958 | |
2959 | pathnode->path.pathtype = T_Agg; |
2960 | pathnode->path.parent = rel; |
2961 | pathnode->path.pathtarget = target; |
2962 | /* For now, assume we are above any joins, so no parameterization */ |
2963 | pathnode->path.param_info = NULL; |
2964 | pathnode->path.parallel_aware = false; |
2965 | pathnode->path.parallel_safe = rel->consider_parallel && |
2966 | subpath->parallel_safe; |
2967 | pathnode->path.parallel_workers = subpath->parallel_workers; |
2968 | if (aggstrategy == AGG_SORTED) |
2969 | pathnode->path.pathkeys = subpath->pathkeys; /* preserves order */ |
2970 | else |
2971 | pathnode->path.pathkeys = NIL; /* output is unordered */ |
2972 | pathnode->subpath = subpath; |
2973 | |
2974 | pathnode->aggstrategy = aggstrategy; |
2975 | pathnode->aggsplit = aggsplit; |
2976 | pathnode->numGroups = numGroups; |
2977 | pathnode->groupClause = groupClause; |
2978 | pathnode->qual = qual; |
2979 | |
2980 | cost_agg(&pathnode->path, root, |
2981 | aggstrategy, aggcosts, |
2982 | list_length(groupClause), numGroups, |
2983 | qual, |
2984 | subpath->startup_cost, subpath->total_cost, |
2985 | subpath->rows); |
2986 | |
2987 | /* add tlist eval cost for each output row */ |
2988 | pathnode->path.startup_cost += target->cost.startup; |
2989 | pathnode->path.total_cost += target->cost.startup + |
2990 | target->cost.per_tuple * pathnode->path.rows; |
2991 | |
2992 | return pathnode; |
2993 | } |
2994 | |
2995 | /* |
2996 | * create_groupingsets_path |
2997 | * Creates a pathnode that represents performing GROUPING SETS aggregation |
2998 | * |
2999 | * GroupingSetsPath represents sorted grouping with one or more grouping sets. |
3000 | * The input path's result must be sorted to match the last entry in |
3001 | * rollup_groupclauses. |
3002 | * |
3003 | * 'rel' is the parent relation associated with the result |
3004 | * 'subpath' is the path representing the source of data |
3005 | * 'target' is the PathTarget to be computed |
3006 | * 'having_qual' is the HAVING quals if any |
3007 | * 'rollups' is a list of RollupData nodes |
3008 | * 'agg_costs' contains cost info about the aggregate functions to be computed |
3009 | * 'numGroups' is the estimated total number of groups |
3010 | */ |
3011 | GroupingSetsPath * |
3012 | create_groupingsets_path(PlannerInfo *root, |
3013 | RelOptInfo *rel, |
3014 | Path *subpath, |
3015 | List *having_qual, |
3016 | AggStrategy aggstrategy, |
3017 | List *rollups, |
3018 | const AggClauseCosts *agg_costs, |
3019 | double numGroups) |
3020 | { |
3021 | GroupingSetsPath *pathnode = makeNode(GroupingSetsPath); |
3022 | PathTarget *target = rel->reltarget; |
3023 | ListCell *lc; |
3024 | bool is_first = true; |
3025 | bool is_first_sort = true; |
3026 | |
3027 | /* The topmost generated Plan node will be an Agg */ |
3028 | pathnode->path.pathtype = T_Agg; |
3029 | pathnode->path.parent = rel; |
3030 | pathnode->path.pathtarget = target; |
3031 | pathnode->path.param_info = subpath->param_info; |
3032 | pathnode->path.parallel_aware = false; |
3033 | pathnode->path.parallel_safe = rel->consider_parallel && |
3034 | subpath->parallel_safe; |
3035 | pathnode->path.parallel_workers = subpath->parallel_workers; |
3036 | pathnode->subpath = subpath; |
3037 | |
3038 | /* |
3039 | * Simplify callers by downgrading AGG_SORTED to AGG_PLAIN, and AGG_MIXED |
3040 | * to AGG_HASHED, here if possible. |
3041 | */ |
3042 | if (aggstrategy == AGG_SORTED && |
3043 | list_length(rollups) == 1 && |
3044 | ((RollupData *) linitial(rollups))->groupClause == NIL) |
3045 | aggstrategy = AGG_PLAIN; |
3046 | |
3047 | if (aggstrategy == AGG_MIXED && |
3048 | list_length(rollups) == 1) |
3049 | aggstrategy = AGG_HASHED; |
3050 | |
3051 | /* |
3052 | * Output will be in sorted order by group_pathkeys if, and only if, there |
3053 | * is a single rollup operation on a non-empty list of grouping |
3054 | * expressions. |
3055 | */ |
3056 | if (aggstrategy == AGG_SORTED && list_length(rollups) == 1) |
3057 | pathnode->path.pathkeys = root->group_pathkeys; |
3058 | else |
3059 | pathnode->path.pathkeys = NIL; |
3060 | |
3061 | pathnode->aggstrategy = aggstrategy; |
3062 | pathnode->rollups = rollups; |
3063 | pathnode->qual = having_qual; |
3064 | |
3065 | Assert(rollups != NIL); |
3066 | Assert(aggstrategy != AGG_PLAIN || list_length(rollups) == 1); |
3067 | Assert(aggstrategy != AGG_MIXED || list_length(rollups) > 1); |
3068 | |
3069 | foreach(lc, rollups) |
3070 | { |
3071 | RollupData *rollup = lfirst(lc); |
3072 | List *gsets = rollup->gsets; |
3073 | int numGroupCols = list_length(linitial(gsets)); |
3074 | |
3075 | /* |
3076 | * In AGG_SORTED or AGG_PLAIN mode, the first rollup takes the |
3077 | * (already-sorted) input, and following ones do their own sort. |
3078 | * |
3079 | * In AGG_HASHED mode, there is one rollup for each grouping set. |
3080 | * |
3081 | * In AGG_MIXED mode, the first rollups are hashed, the first |
3082 | * non-hashed one takes the (already-sorted) input, and following ones |
3083 | * do their own sort. |
3084 | */ |
3085 | if (is_first) |
3086 | { |
3087 | cost_agg(&pathnode->path, root, |
3088 | aggstrategy, |
3089 | agg_costs, |
3090 | numGroupCols, |
3091 | rollup->numGroups, |
3092 | having_qual, |
3093 | subpath->startup_cost, |
3094 | subpath->total_cost, |
3095 | subpath->rows); |
3096 | is_first = false; |
3097 | if (!rollup->is_hashed) |
3098 | is_first_sort = false; |
3099 | } |
3100 | else |
3101 | { |
3102 | Path sort_path; /* dummy for result of cost_sort */ |
3103 | Path agg_path; /* dummy for result of cost_agg */ |
3104 | |
3105 | if (rollup->is_hashed || is_first_sort) |
3106 | { |
3107 | /* |
3108 | * Account for cost of aggregation, but don't charge input |
3109 | * cost again |
3110 | */ |
3111 | cost_agg(&agg_path, root, |
3112 | rollup->is_hashed ? AGG_HASHED : AGG_SORTED, |
3113 | agg_costs, |
3114 | numGroupCols, |
3115 | rollup->numGroups, |
3116 | having_qual, |
3117 | 0.0, 0.0, |
3118 | subpath->rows); |
3119 | if (!rollup->is_hashed) |
3120 | is_first_sort = false; |
3121 | } |
3122 | else |
3123 | { |
3124 | /* Account for cost of sort, but don't charge input cost again */ |
3125 | cost_sort(&sort_path, root, NIL, |
3126 | 0.0, |
3127 | subpath->rows, |
3128 | subpath->pathtarget->width, |
3129 | 0.0, |
3130 | work_mem, |
3131 | -1.0); |
3132 | |
3133 | /* Account for cost of aggregation */ |
3134 | |
3135 | cost_agg(&agg_path, root, |
3136 | AGG_SORTED, |
3137 | agg_costs, |
3138 | numGroupCols, |
3139 | rollup->numGroups, |
3140 | having_qual, |
3141 | sort_path.startup_cost, |
3142 | sort_path.total_cost, |
3143 | sort_path.rows); |
3144 | } |
3145 | |
3146 | pathnode->path.total_cost += agg_path.total_cost; |
3147 | pathnode->path.rows += agg_path.rows; |
3148 | } |
3149 | } |
3150 | |
3151 | /* add tlist eval cost for each output row */ |
3152 | pathnode->path.startup_cost += target->cost.startup; |
3153 | pathnode->path.total_cost += target->cost.startup + |
3154 | target->cost.per_tuple * pathnode->path.rows; |
3155 | |
3156 | return pathnode; |
3157 | } |
3158 | |
3159 | /* |
3160 | * create_minmaxagg_path |
3161 | * Creates a pathnode that represents computation of MIN/MAX aggregates |
3162 | * |
3163 | * 'rel' is the parent relation associated with the result |
3164 | * 'target' is the PathTarget to be computed |
3165 | * 'mmaggregates' is a list of MinMaxAggInfo structs |
3166 | * 'quals' is the HAVING quals if any |
3167 | */ |
3168 | MinMaxAggPath * |
3169 | create_minmaxagg_path(PlannerInfo *root, |
3170 | RelOptInfo *rel, |
3171 | PathTarget *target, |
3172 | List *mmaggregates, |
3173 | List *quals) |
3174 | { |
3175 | MinMaxAggPath *pathnode = makeNode(MinMaxAggPath); |
3176 | Cost initplan_cost; |
3177 | ListCell *lc; |
3178 | |
3179 | /* The topmost generated Plan node will be a Result */ |
3180 | pathnode->path.pathtype = T_Result; |
3181 | pathnode->path.parent = rel; |
3182 | pathnode->path.pathtarget = target; |
3183 | /* For now, assume we are above any joins, so no parameterization */ |
3184 | pathnode->path.param_info = NULL; |
3185 | pathnode->path.parallel_aware = false; |
3186 | /* A MinMaxAggPath implies use of subplans, so cannot be parallel-safe */ |
3187 | pathnode->path.parallel_safe = false; |
3188 | pathnode->path.parallel_workers = 0; |
3189 | /* Result is one unordered row */ |
3190 | pathnode->path.rows = 1; |
3191 | pathnode->path.pathkeys = NIL; |
3192 | |
3193 | pathnode->mmaggregates = mmaggregates; |
3194 | pathnode->quals = quals; |
3195 | |
3196 | /* Calculate cost of all the initplans ... */ |
3197 | initplan_cost = 0; |
3198 | foreach(lc, mmaggregates) |
3199 | { |
3200 | MinMaxAggInfo *mminfo = (MinMaxAggInfo *) lfirst(lc); |
3201 | |
3202 | initplan_cost += mminfo->pathcost; |
3203 | } |
3204 | |
3205 | /* add tlist eval cost for each output row, plus cpu_tuple_cost */ |
3206 | pathnode->path.startup_cost = initplan_cost + target->cost.startup; |
3207 | pathnode->path.total_cost = initplan_cost + target->cost.startup + |
3208 | target->cost.per_tuple + cpu_tuple_cost; |
3209 | |
3210 | /* |
3211 | * Add cost of qual, if any --- but we ignore its selectivity, since our |
3212 | * rowcount estimate should be 1 no matter what the qual is. |
3213 | */ |
3214 | if (quals) |
3215 | { |
3216 | QualCost qual_cost; |
3217 | |
3218 | cost_qual_eval(&qual_cost, quals, root); |
3219 | pathnode->path.startup_cost += qual_cost.startup; |
3220 | pathnode->path.total_cost += qual_cost.startup + qual_cost.per_tuple; |
3221 | } |
3222 | |
3223 | return pathnode; |
3224 | } |
3225 | |
3226 | /* |
3227 | * create_windowagg_path |
3228 | * Creates a pathnode that represents computation of window functions |
3229 | * |
3230 | * 'rel' is the parent relation associated with the result |
3231 | * 'subpath' is the path representing the source of data |
3232 | * 'target' is the PathTarget to be computed |
3233 | * 'windowFuncs' is a list of WindowFunc structs |
3234 | * 'winclause' is a WindowClause that is common to all the WindowFuncs |
3235 | * |
3236 | * The input must be sorted according to the WindowClause's PARTITION keys |
3237 | * plus ORDER BY keys. |
3238 | */ |
3239 | WindowAggPath * |
3240 | create_windowagg_path(PlannerInfo *root, |
3241 | RelOptInfo *rel, |
3242 | Path *subpath, |
3243 | PathTarget *target, |
3244 | List *windowFuncs, |
3245 | WindowClause *winclause) |
3246 | { |
3247 | WindowAggPath *pathnode = makeNode(WindowAggPath); |
3248 | |
3249 | pathnode->path.pathtype = T_WindowAgg; |
3250 | pathnode->path.parent = rel; |
3251 | pathnode->path.pathtarget = target; |
3252 | /* For now, assume we are above any joins, so no parameterization */ |
3253 | pathnode->path.param_info = NULL; |
3254 | pathnode->path.parallel_aware = false; |
3255 | pathnode->path.parallel_safe = rel->consider_parallel && |
3256 | subpath->parallel_safe; |
3257 | pathnode->path.parallel_workers = subpath->parallel_workers; |
3258 | /* WindowAgg preserves the input sort order */ |
3259 | pathnode->path.pathkeys = subpath->pathkeys; |
3260 | |
3261 | pathnode->subpath = subpath; |
3262 | pathnode->winclause = winclause; |
3263 | |
3264 | /* |
3265 | * For costing purposes, assume that there are no redundant partitioning |
3266 | * or ordering columns; it's not worth the trouble to deal with that |
3267 | * corner case here. So we just pass the unmodified list lengths to |
3268 | * cost_windowagg. |
3269 | */ |
3270 | cost_windowagg(&pathnode->path, root, |
3271 | windowFuncs, |
3272 | list_length(winclause->partitionClause), |
3273 | list_length(winclause->orderClause), |
3274 | subpath->startup_cost, |
3275 | subpath->total_cost, |
3276 | subpath->rows); |
3277 | |
3278 | /* add tlist eval cost for each output row */ |
3279 | pathnode->path.startup_cost += target->cost.startup; |
3280 | pathnode->path.total_cost += target->cost.startup + |
3281 | target->cost.per_tuple * pathnode->path.rows; |
3282 | |
3283 | return pathnode; |
3284 | } |
3285 | |
3286 | /* |
3287 | * create_setop_path |
3288 | * Creates a pathnode that represents computation of INTERSECT or EXCEPT |
3289 | * |
3290 | * 'rel' is the parent relation associated with the result |
3291 | * 'subpath' is the path representing the source of data |
3292 | * 'cmd' is the specific semantics (INTERSECT or EXCEPT, with/without ALL) |
3293 | * 'strategy' is the implementation strategy (sorted or hashed) |
3294 | * 'distinctList' is a list of SortGroupClause's representing the grouping |
3295 | * 'flagColIdx' is the column number where the flag column will be, if any |
3296 | * 'firstFlag' is the flag value for the first input relation when hashing; |
3297 | * or -1 when sorting |
3298 | * 'numGroups' is the estimated number of distinct groups |
3299 | * 'outputRows' is the estimated number of output rows |
3300 | */ |
3301 | SetOpPath * |
3302 | create_setop_path(PlannerInfo *root, |
3303 | RelOptInfo *rel, |
3304 | Path *subpath, |
3305 | SetOpCmd cmd, |
3306 | SetOpStrategy strategy, |
3307 | List *distinctList, |
3308 | AttrNumber flagColIdx, |
3309 | int firstFlag, |
3310 | double numGroups, |
3311 | double outputRows) |
3312 | { |
3313 | SetOpPath *pathnode = makeNode(SetOpPath); |
3314 | |
3315 | pathnode->path.pathtype = T_SetOp; |
3316 | pathnode->path.parent = rel; |
3317 | /* SetOp doesn't project, so use source path's pathtarget */ |
3318 | pathnode->path.pathtarget = subpath->pathtarget; |
3319 | /* For now, assume we are above any joins, so no parameterization */ |
3320 | pathnode->path.param_info = NULL; |
3321 | pathnode->path.parallel_aware = false; |
3322 | pathnode->path.parallel_safe = rel->consider_parallel && |
3323 | subpath->parallel_safe; |
3324 | pathnode->path.parallel_workers = subpath->parallel_workers; |
3325 | /* SetOp preserves the input sort order if in sort mode */ |
3326 | pathnode->path.pathkeys = |
3327 | (strategy == SETOP_SORTED) ? subpath->pathkeys : NIL; |
3328 | |
3329 | pathnode->subpath = subpath; |
3330 | pathnode->cmd = cmd; |
3331 | pathnode->strategy = strategy; |
3332 | pathnode->distinctList = distinctList; |
3333 | pathnode->flagColIdx = flagColIdx; |
3334 | pathnode->firstFlag = firstFlag; |
3335 | pathnode->numGroups = numGroups; |
3336 | |
3337 | /* |
3338 | * Charge one cpu_operator_cost per comparison per input tuple. We assume |
3339 | * all columns get compared at most of the tuples. |
3340 | */ |
3341 | pathnode->path.startup_cost = subpath->startup_cost; |
3342 | pathnode->path.total_cost = subpath->total_cost + |
3343 | cpu_operator_cost * subpath->rows * list_length(distinctList); |
3344 | pathnode->path.rows = outputRows; |
3345 | |
3346 | return pathnode; |
3347 | } |
3348 | |
3349 | /* |
3350 | * create_recursiveunion_path |
3351 | * Creates a pathnode that represents a recursive UNION node |
3352 | * |
3353 | * 'rel' is the parent relation associated with the result |
3354 | * 'leftpath' is the source of data for the non-recursive term |
3355 | * 'rightpath' is the source of data for the recursive term |
3356 | * 'target' is the PathTarget to be computed |
3357 | * 'distinctList' is a list of SortGroupClause's representing the grouping |
3358 | * 'wtParam' is the ID of Param representing work table |
3359 | * 'numGroups' is the estimated number of groups |
3360 | * |
3361 | * For recursive UNION ALL, distinctList is empty and numGroups is zero |
3362 | */ |
3363 | RecursiveUnionPath * |
3364 | create_recursiveunion_path(PlannerInfo *root, |
3365 | RelOptInfo *rel, |
3366 | Path *leftpath, |
3367 | Path *rightpath, |
3368 | PathTarget *target, |
3369 | List *distinctList, |
3370 | int wtParam, |
3371 | double numGroups) |
3372 | { |
3373 | RecursiveUnionPath *pathnode = makeNode(RecursiveUnionPath); |
3374 | |
3375 | pathnode->path.pathtype = T_RecursiveUnion; |
3376 | pathnode->path.parent = rel; |
3377 | pathnode->path.pathtarget = target; |
3378 | /* For now, assume we are above any joins, so no parameterization */ |
3379 | pathnode->path.param_info = NULL; |
3380 | pathnode->path.parallel_aware = false; |
3381 | pathnode->path.parallel_safe = rel->consider_parallel && |
3382 | leftpath->parallel_safe && rightpath->parallel_safe; |
3383 | /* Foolish, but we'll do it like joins for now: */ |
3384 | pathnode->path.parallel_workers = leftpath->parallel_workers; |
3385 | /* RecursiveUnion result is always unsorted */ |
3386 | pathnode->path.pathkeys = NIL; |
3387 | |
3388 | pathnode->leftpath = leftpath; |
3389 | pathnode->rightpath = rightpath; |
3390 | pathnode->distinctList = distinctList; |
3391 | pathnode->wtParam = wtParam; |
3392 | pathnode->numGroups = numGroups; |
3393 | |
3394 | cost_recursive_union(&pathnode->path, leftpath, rightpath); |
3395 | |
3396 | return pathnode; |
3397 | } |
3398 | |
3399 | /* |
3400 | * create_lockrows_path |
3401 | * Creates a pathnode that represents acquiring row locks |
3402 | * |
3403 | * 'rel' is the parent relation associated with the result |
3404 | * 'subpath' is the path representing the source of data |
3405 | * 'rowMarks' is a list of PlanRowMark's |
3406 | * 'epqParam' is the ID of Param for EvalPlanQual re-eval |
3407 | */ |
3408 | LockRowsPath * |
3409 | create_lockrows_path(PlannerInfo *root, RelOptInfo *rel, |
3410 | Path *subpath, List *rowMarks, int epqParam) |
3411 | { |
3412 | LockRowsPath *pathnode = makeNode(LockRowsPath); |
3413 | |
3414 | pathnode->path.pathtype = T_LockRows; |
3415 | pathnode->path.parent = rel; |
3416 | /* LockRows doesn't project, so use source path's pathtarget */ |
3417 | pathnode->path.pathtarget = subpath->pathtarget; |
3418 | /* For now, assume we are above any joins, so no parameterization */ |
3419 | pathnode->path.param_info = NULL; |
3420 | pathnode->path.parallel_aware = false; |
3421 | pathnode->path.parallel_safe = false; |
3422 | pathnode->path.parallel_workers = 0; |
3423 | pathnode->path.rows = subpath->rows; |
3424 | |
3425 | /* |
3426 | * The result cannot be assumed sorted, since locking might cause the sort |
3427 | * key columns to be replaced with new values. |
3428 | */ |
3429 | pathnode->path.pathkeys = NIL; |
3430 | |
3431 | pathnode->subpath = subpath; |
3432 | pathnode->rowMarks = rowMarks; |
3433 | pathnode->epqParam = epqParam; |
3434 | |
3435 | /* |
3436 | * We should charge something extra for the costs of row locking and |
3437 | * possible refetches, but it's hard to say how much. For now, use |
3438 | * cpu_tuple_cost per row. |
3439 | */ |
3440 | pathnode->path.startup_cost = subpath->startup_cost; |
3441 | pathnode->path.total_cost = subpath->total_cost + |
3442 | cpu_tuple_cost * subpath->rows; |
3443 | |
3444 | return pathnode; |
3445 | } |
3446 | |
3447 | /* |
3448 | * create_modifytable_path |
3449 | * Creates a pathnode that represents performing INSERT/UPDATE/DELETE mods |
3450 | * |
3451 | * 'rel' is the parent relation associated with the result |
3452 | * 'operation' is the operation type |
3453 | * 'canSetTag' is true if we set the command tag/es_processed |
3454 | * 'nominalRelation' is the parent RT index for use of EXPLAIN |
3455 | * 'rootRelation' is the partitioned table root RT index, or 0 if none |
3456 | * 'partColsUpdated' is true if any partitioning columns are being updated, |
3457 | * either from the target relation or a descendent partitioned table. |
3458 | * 'resultRelations' is an integer list of actual RT indexes of target rel(s) |
3459 | * 'subpaths' is a list of Path(s) producing source data (one per rel) |
3460 | * 'subroots' is a list of PlannerInfo structs (one per rel) |
3461 | * 'withCheckOptionLists' is a list of WCO lists (one per rel) |
3462 | * 'returningLists' is a list of RETURNING tlists (one per rel) |
3463 | * 'rowMarks' is a list of PlanRowMarks (non-locking only) |
3464 | * 'onconflict' is the ON CONFLICT clause, or NULL |
3465 | * 'epqParam' is the ID of Param for EvalPlanQual re-eval |
3466 | */ |
3467 | ModifyTablePath * |
3468 | create_modifytable_path(PlannerInfo *root, RelOptInfo *rel, |
3469 | CmdType operation, bool canSetTag, |
3470 | Index nominalRelation, Index rootRelation, |
3471 | bool partColsUpdated, |
3472 | List *resultRelations, List *subpaths, |
3473 | List *subroots, |
3474 | List *withCheckOptionLists, List *returningLists, |
3475 | List *rowMarks, OnConflictExpr *onconflict, |
3476 | int epqParam) |
3477 | { |
3478 | ModifyTablePath *pathnode = makeNode(ModifyTablePath); |
3479 | double total_size; |
3480 | ListCell *lc; |
3481 | |
3482 | Assert(list_length(resultRelations) == list_length(subpaths)); |
3483 | Assert(list_length(resultRelations) == list_length(subroots)); |
3484 | Assert(withCheckOptionLists == NIL || |
3485 | list_length(resultRelations) == list_length(withCheckOptionLists)); |
3486 | Assert(returningLists == NIL || |
3487 | list_length(resultRelations) == list_length(returningLists)); |
3488 | |
3489 | pathnode->path.pathtype = T_ModifyTable; |
3490 | pathnode->path.parent = rel; |
3491 | /* pathtarget is not interesting, just make it minimally valid */ |
3492 | pathnode->path.pathtarget = rel->reltarget; |
3493 | /* For now, assume we are above any joins, so no parameterization */ |
3494 | pathnode->path.param_info = NULL; |
3495 | pathnode->path.parallel_aware = false; |
3496 | pathnode->path.parallel_safe = false; |
3497 | pathnode->path.parallel_workers = 0; |
3498 | pathnode->path.pathkeys = NIL; |
3499 | |
3500 | /* |
3501 | * Compute cost & rowcount as sum of subpath costs & rowcounts. |
3502 | * |
3503 | * Currently, we don't charge anything extra for the actual table |
3504 | * modification work, nor for the WITH CHECK OPTIONS or RETURNING |
3505 | * expressions if any. It would only be window dressing, since |
3506 | * ModifyTable is always a top-level node and there is no way for the |
3507 | * costs to change any higher-level planning choices. But we might want |
3508 | * to make it look better sometime. |
3509 | */ |
3510 | pathnode->path.startup_cost = 0; |
3511 | pathnode->path.total_cost = 0; |
3512 | pathnode->path.rows = 0; |
3513 | total_size = 0; |
3514 | foreach(lc, subpaths) |
3515 | { |
3516 | Path *subpath = (Path *) lfirst(lc); |
3517 | |
3518 | if (lc == list_head(subpaths)) /* first node? */ |
3519 | pathnode->path.startup_cost = subpath->startup_cost; |
3520 | pathnode->path.total_cost += subpath->total_cost; |
3521 | pathnode->path.rows += subpath->rows; |
3522 | total_size += subpath->pathtarget->width * subpath->rows; |
3523 | } |
3524 | |
3525 | /* |
3526 | * Set width to the average width of the subpath outputs. XXX this is |
3527 | * totally wrong: we should report zero if no RETURNING, else an average |
3528 | * of the RETURNING tlist widths. But it's what happened historically, |
3529 | * and improving it is a task for another day. |
3530 | */ |
3531 | if (pathnode->path.rows > 0) |
3532 | total_size /= pathnode->path.rows; |
3533 | pathnode->path.pathtarget->width = rint(total_size); |
3534 | |
3535 | pathnode->operation = operation; |
3536 | pathnode->canSetTag = canSetTag; |
3537 | pathnode->nominalRelation = nominalRelation; |
3538 | pathnode->rootRelation = rootRelation; |
3539 | pathnode->partColsUpdated = partColsUpdated; |
3540 | pathnode->resultRelations = resultRelations; |
3541 | pathnode->subpaths = subpaths; |
3542 | pathnode->subroots = subroots; |
3543 | pathnode->withCheckOptionLists = withCheckOptionLists; |
3544 | pathnode->returningLists = returningLists; |
3545 | pathnode->rowMarks = rowMarks; |
3546 | pathnode->onconflict = onconflict; |
3547 | pathnode->epqParam = epqParam; |
3548 | |
3549 | return pathnode; |
3550 | } |
3551 | |
3552 | /* |
3553 | * create_limit_path |
3554 | * Creates a pathnode that represents performing LIMIT/OFFSET |
3555 | * |
3556 | * In addition to providing the actual OFFSET and LIMIT expressions, |
3557 | * the caller must provide estimates of their values for costing purposes. |
3558 | * The estimates are as computed by preprocess_limit(), ie, 0 represents |
3559 | * the clause not being present, and -1 means it's present but we could |
3560 | * not estimate its value. |
3561 | * |
3562 | * 'rel' is the parent relation associated with the result |
3563 | * 'subpath' is the path representing the source of data |
3564 | * 'limitOffset' is the actual OFFSET expression, or NULL |
3565 | * 'limitCount' is the actual LIMIT expression, or NULL |
3566 | * 'offset_est' is the estimated value of the OFFSET expression |
3567 | * 'count_est' is the estimated value of the LIMIT expression |
3568 | */ |
3569 | LimitPath * |
3570 | create_limit_path(PlannerInfo *root, RelOptInfo *rel, |
3571 | Path *subpath, |
3572 | Node *limitOffset, Node *limitCount, |
3573 | int64 offset_est, int64 count_est) |
3574 | { |
3575 | LimitPath *pathnode = makeNode(LimitPath); |
3576 | |
3577 | pathnode->path.pathtype = T_Limit; |
3578 | pathnode->path.parent = rel; |
3579 | /* Limit doesn't project, so use source path's pathtarget */ |
3580 | pathnode->path.pathtarget = subpath->pathtarget; |
3581 | /* For now, assume we are above any joins, so no parameterization */ |
3582 | pathnode->path.param_info = NULL; |
3583 | pathnode->path.parallel_aware = false; |
3584 | pathnode->path.parallel_safe = rel->consider_parallel && |
3585 | subpath->parallel_safe; |
3586 | pathnode->path.parallel_workers = subpath->parallel_workers; |
3587 | pathnode->path.rows = subpath->rows; |
3588 | pathnode->path.startup_cost = subpath->startup_cost; |
3589 | pathnode->path.total_cost = subpath->total_cost; |
3590 | pathnode->path.pathkeys = subpath->pathkeys; |
3591 | pathnode->subpath = subpath; |
3592 | pathnode->limitOffset = limitOffset; |
3593 | pathnode->limitCount = limitCount; |
3594 | |
3595 | /* |
3596 | * Adjust the output rows count and costs according to the offset/limit. |
3597 | */ |
3598 | adjust_limit_rows_costs(&pathnode->path.rows, |
3599 | &pathnode->path.startup_cost, |
3600 | &pathnode->path.total_cost, |
3601 | offset_est, count_est); |
3602 | |
3603 | return pathnode; |
3604 | } |
3605 | |
3606 | /* |
3607 | * adjust_limit_rows_costs |
3608 | * Adjust the size and cost estimates for a LimitPath node according to the |
3609 | * offset/limit. |
3610 | * |
3611 | * This is only a cosmetic issue if we are at top level, but if we are |
3612 | * building a subquery then it's important to report correct info to the outer |
3613 | * planner. |
3614 | * |
3615 | * When the offset or count couldn't be estimated, use 10% of the estimated |
3616 | * number of rows emitted from the subpath. |
3617 | * |
3618 | * XXX we don't bother to add eval costs of the offset/limit expressions |
3619 | * themselves to the path costs. In theory we should, but in most cases those |
3620 | * expressions are trivial and it's just not worth the trouble. |
3621 | */ |
3622 | void |
3623 | adjust_limit_rows_costs(double *rows, /* in/out parameter */ |
3624 | Cost *startup_cost, /* in/out parameter */ |
3625 | Cost *total_cost, /* in/out parameter */ |
3626 | int64 offset_est, |
3627 | int64 count_est) |
3628 | { |
3629 | double input_rows = *rows; |
3630 | Cost input_startup_cost = *startup_cost; |
3631 | Cost input_total_cost = *total_cost; |
3632 | |
3633 | if (offset_est != 0) |
3634 | { |
3635 | double offset_rows; |
3636 | |
3637 | if (offset_est > 0) |
3638 | offset_rows = (double) offset_est; |
3639 | else |
3640 | offset_rows = clamp_row_est(input_rows * 0.10); |
3641 | if (offset_rows > *rows) |
3642 | offset_rows = *rows; |
3643 | if (input_rows > 0) |
3644 | *startup_cost += |
3645 | (input_total_cost - input_startup_cost) |
3646 | * offset_rows / input_rows; |
3647 | *rows -= offset_rows; |
3648 | if (*rows < 1) |
3649 | *rows = 1; |
3650 | } |
3651 | |
3652 | if (count_est != 0) |
3653 | { |
3654 | double count_rows; |
3655 | |
3656 | if (count_est > 0) |
3657 | count_rows = (double) count_est; |
3658 | else |
3659 | count_rows = clamp_row_est(input_rows * 0.10); |
3660 | if (count_rows > *rows) |
3661 | count_rows = *rows; |
3662 | if (input_rows > 0) |
3663 | *total_cost = *startup_cost + |
3664 | (input_total_cost - input_startup_cost) |
3665 | * count_rows / input_rows; |
3666 | *rows = count_rows; |
3667 | if (*rows < 1) |
3668 | *rows = 1; |
3669 | } |
3670 | } |
3671 | |
3672 | |
3673 | /* |
3674 | * reparameterize_path |
3675 | * Attempt to modify a Path to have greater parameterization |
3676 | * |
3677 | * We use this to attempt to bring all child paths of an appendrel to the |
3678 | * same parameterization level, ensuring that they all enforce the same set |
3679 | * of join quals (and thus that that parameterization can be attributed to |
3680 | * an append path built from such paths). Currently, only a few path types |
3681 | * are supported here, though more could be added at need. We return NULL |
3682 | * if we can't reparameterize the given path. |
3683 | * |
3684 | * Note: we intentionally do not pass created paths to add_path(); it would |
3685 | * possibly try to delete them on the grounds of being cost-inferior to the |
3686 | * paths they were made from, and we don't want that. Paths made here are |
3687 | * not necessarily of general-purpose usefulness, but they can be useful |
3688 | * as members of an append path. |
3689 | */ |
3690 | Path * |
3691 | reparameterize_path(PlannerInfo *root, Path *path, |
3692 | Relids required_outer, |
3693 | double loop_count) |
3694 | { |
3695 | RelOptInfo *rel = path->parent; |
3696 | |
3697 | /* Can only increase, not decrease, path's parameterization */ |
3698 | if (!bms_is_subset(PATH_REQ_OUTER(path), required_outer)) |
3699 | return NULL; |
3700 | switch (path->pathtype) |
3701 | { |
3702 | case T_SeqScan: |
3703 | return create_seqscan_path(root, rel, required_outer, 0); |
3704 | case T_SampleScan: |
3705 | return (Path *) create_samplescan_path(root, rel, required_outer); |
3706 | case T_IndexScan: |
3707 | case T_IndexOnlyScan: |
3708 | { |
3709 | IndexPath *ipath = (IndexPath *) path; |
3710 | IndexPath *newpath = makeNode(IndexPath); |
3711 | |
3712 | /* |
3713 | * We can't use create_index_path directly, and would not want |
3714 | * to because it would re-compute the indexqual conditions |
3715 | * which is wasted effort. Instead we hack things a bit: |
3716 | * flat-copy the path node, revise its param_info, and redo |
3717 | * the cost estimate. |
3718 | */ |
3719 | memcpy(newpath, ipath, sizeof(IndexPath)); |
3720 | newpath->path.param_info = |
3721 | get_baserel_parampathinfo(root, rel, required_outer); |
3722 | cost_index(newpath, root, loop_count, false); |
3723 | return (Path *) newpath; |
3724 | } |
3725 | case T_BitmapHeapScan: |
3726 | { |
3727 | BitmapHeapPath *bpath = (BitmapHeapPath *) path; |
3728 | |
3729 | return (Path *) create_bitmap_heap_path(root, |
3730 | rel, |
3731 | bpath->bitmapqual, |
3732 | required_outer, |
3733 | loop_count, 0); |
3734 | } |
3735 | case T_SubqueryScan: |
3736 | { |
3737 | SubqueryScanPath *spath = (SubqueryScanPath *) path; |
3738 | |
3739 | return (Path *) create_subqueryscan_path(root, |
3740 | rel, |
3741 | spath->subpath, |
3742 | spath->path.pathkeys, |
3743 | required_outer); |
3744 | } |
3745 | case T_Result: |
3746 | /* Supported only for RTE_RESULT scan paths */ |
3747 | if (IsA(path, Path)) |
3748 | return create_resultscan_path(root, rel, required_outer); |
3749 | break; |
3750 | case T_Append: |
3751 | { |
3752 | AppendPath *apath = (AppendPath *) path; |
3753 | List *childpaths = NIL; |
3754 | List *partialpaths = NIL; |
3755 | int i; |
3756 | ListCell *lc; |
3757 | |
3758 | /* Reparameterize the children */ |
3759 | i = 0; |
3760 | foreach(lc, apath->subpaths) |
3761 | { |
3762 | Path *spath = (Path *) lfirst(lc); |
3763 | |
3764 | spath = reparameterize_path(root, spath, |
3765 | required_outer, |
3766 | loop_count); |
3767 | if (spath == NULL) |
3768 | return NULL; |
3769 | /* We have to re-split the regular and partial paths */ |
3770 | if (i < apath->first_partial_path) |
3771 | childpaths = lappend(childpaths, spath); |
3772 | else |
3773 | partialpaths = lappend(partialpaths, spath); |
3774 | i++; |
3775 | } |
3776 | return (Path *) |
3777 | create_append_path(root, rel, childpaths, partialpaths, |
3778 | apath->path.pathkeys, required_outer, |
3779 | apath->path.parallel_workers, |
3780 | apath->path.parallel_aware, |
3781 | apath->partitioned_rels, |
3782 | -1); |
3783 | } |
3784 | default: |
3785 | break; |
3786 | } |
3787 | return NULL; |
3788 | } |
3789 | |
3790 | /* |
3791 | * reparameterize_path_by_child |
3792 | * Given a path parameterized by the parent of the given child relation, |
3793 | * translate the path to be parameterized by the given child relation. |
3794 | * |
3795 | * The function creates a new path of the same type as the given path, but |
3796 | * parameterized by the given child relation. Most fields from the original |
3797 | * path can simply be flat-copied, but any expressions must be adjusted to |
3798 | * refer to the correct varnos, and any paths must be recursively |
3799 | * reparameterized. Other fields that refer to specific relids also need |
3800 | * adjustment. |
3801 | * |
3802 | * The cost, number of rows, width and parallel path properties depend upon |
3803 | * path->parent, which does not change during the translation. Hence those |
3804 | * members are copied as they are. |
3805 | * |
3806 | * If the given path can not be reparameterized, the function returns NULL. |
3807 | */ |
3808 | Path * |
3809 | reparameterize_path_by_child(PlannerInfo *root, Path *path, |
3810 | RelOptInfo *child_rel) |
3811 | { |
3812 | |
3813 | #define FLAT_COPY_PATH(newnode, node, nodetype) \ |
3814 | ( (newnode) = makeNode(nodetype), \ |
3815 | memcpy((newnode), (node), sizeof(nodetype)) ) |
3816 | |
3817 | #define ADJUST_CHILD_ATTRS(node) \ |
3818 | ((node) = \ |
3819 | (List *) adjust_appendrel_attrs_multilevel(root, (Node *) (node), \ |
3820 | child_rel->relids, \ |
3821 | child_rel->top_parent_relids)) |
3822 | |
3823 | #define REPARAMETERIZE_CHILD_PATH(path) \ |
3824 | do { \ |
3825 | (path) = reparameterize_path_by_child(root, (path), child_rel); \ |
3826 | if ((path) == NULL) \ |
3827 | return NULL; \ |
3828 | } while(0); |
3829 | |
3830 | #define REPARAMETERIZE_CHILD_PATH_LIST(pathlist) \ |
3831 | do { \ |
3832 | if ((pathlist) != NIL) \ |
3833 | { \ |
3834 | (pathlist) = reparameterize_pathlist_by_child(root, (pathlist), \ |
3835 | child_rel); \ |
3836 | if ((pathlist) == NIL) \ |
3837 | return NULL; \ |
3838 | } \ |
3839 | } while(0); |
3840 | |
3841 | Path *new_path; |
3842 | ParamPathInfo *new_ppi; |
3843 | ParamPathInfo *old_ppi; |
3844 | Relids required_outer; |
3845 | |
3846 | /* |
3847 | * If the path is not parameterized by parent of the given relation, it |
3848 | * doesn't need reparameterization. |
3849 | */ |
3850 | if (!path->param_info || |
3851 | !bms_overlap(PATH_REQ_OUTER(path), child_rel->top_parent_relids)) |
3852 | return path; |
3853 | |
3854 | /* Reparameterize a copy of given path. */ |
3855 | switch (nodeTag(path)) |
3856 | { |
3857 | case T_Path: |
3858 | FLAT_COPY_PATH(new_path, path, Path); |
3859 | break; |
3860 | |
3861 | case T_IndexPath: |
3862 | { |
3863 | IndexPath *ipath; |
3864 | |
3865 | FLAT_COPY_PATH(ipath, path, IndexPath); |
3866 | ADJUST_CHILD_ATTRS(ipath->indexclauses); |
3867 | new_path = (Path *) ipath; |
3868 | } |
3869 | break; |
3870 | |
3871 | case T_BitmapHeapPath: |
3872 | { |
3873 | BitmapHeapPath *bhpath; |
3874 | |
3875 | FLAT_COPY_PATH(bhpath, path, BitmapHeapPath); |
3876 | REPARAMETERIZE_CHILD_PATH(bhpath->bitmapqual); |
3877 | new_path = (Path *) bhpath; |
3878 | } |
3879 | break; |
3880 | |
3881 | case T_BitmapAndPath: |
3882 | { |
3883 | BitmapAndPath *bapath; |
3884 | |
3885 | FLAT_COPY_PATH(bapath, path, BitmapAndPath); |
3886 | REPARAMETERIZE_CHILD_PATH_LIST(bapath->bitmapquals); |
3887 | new_path = (Path *) bapath; |
3888 | } |
3889 | break; |
3890 | |
3891 | case T_BitmapOrPath: |
3892 | { |
3893 | BitmapOrPath *bopath; |
3894 | |
3895 | FLAT_COPY_PATH(bopath, path, BitmapOrPath); |
3896 | REPARAMETERIZE_CHILD_PATH_LIST(bopath->bitmapquals); |
3897 | new_path = (Path *) bopath; |
3898 | } |
3899 | break; |
3900 | |
3901 | case T_TidPath: |
3902 | { |
3903 | TidPath *tpath; |
3904 | |
3905 | FLAT_COPY_PATH(tpath, path, TidPath); |
3906 | ADJUST_CHILD_ATTRS(tpath->tidquals); |
3907 | new_path = (Path *) tpath; |
3908 | } |
3909 | break; |
3910 | |
3911 | case T_ForeignPath: |
3912 | { |
3913 | ForeignPath *fpath; |
3914 | ReparameterizeForeignPathByChild_function rfpc_func; |
3915 | |
3916 | FLAT_COPY_PATH(fpath, path, ForeignPath); |
3917 | if (fpath->fdw_outerpath) |
3918 | REPARAMETERIZE_CHILD_PATH(fpath->fdw_outerpath); |
3919 | |
3920 | /* Hand over to FDW if needed. */ |
3921 | rfpc_func = |
3922 | path->parent->fdwroutine->ReparameterizeForeignPathByChild; |
3923 | if (rfpc_func) |
3924 | fpath->fdw_private = rfpc_func(root, fpath->fdw_private, |
3925 | child_rel); |
3926 | new_path = (Path *) fpath; |
3927 | } |
3928 | break; |
3929 | |
3930 | case T_CustomPath: |
3931 | { |
3932 | CustomPath *cpath; |
3933 | |
3934 | FLAT_COPY_PATH(cpath, path, CustomPath); |
3935 | REPARAMETERIZE_CHILD_PATH_LIST(cpath->custom_paths); |
3936 | if (cpath->methods && |
3937 | cpath->methods->ReparameterizeCustomPathByChild) |
3938 | cpath->custom_private = |
3939 | cpath->methods->ReparameterizeCustomPathByChild(root, |
3940 | cpath->custom_private, |
3941 | child_rel); |
3942 | new_path = (Path *) cpath; |
3943 | } |
3944 | break; |
3945 | |
3946 | case T_NestPath: |
3947 | { |
3948 | JoinPath *jpath; |
3949 | |
3950 | FLAT_COPY_PATH(jpath, path, NestPath); |
3951 | |
3952 | REPARAMETERIZE_CHILD_PATH(jpath->outerjoinpath); |
3953 | REPARAMETERIZE_CHILD_PATH(jpath->innerjoinpath); |
3954 | ADJUST_CHILD_ATTRS(jpath->joinrestrictinfo); |
3955 | new_path = (Path *) jpath; |
3956 | } |
3957 | break; |
3958 | |
3959 | case T_MergePath: |
3960 | { |
3961 | JoinPath *jpath; |
3962 | MergePath *mpath; |
3963 | |
3964 | FLAT_COPY_PATH(mpath, path, MergePath); |
3965 | |
3966 | jpath = (JoinPath *) mpath; |
3967 | REPARAMETERIZE_CHILD_PATH(jpath->outerjoinpath); |
3968 | REPARAMETERIZE_CHILD_PATH(jpath->innerjoinpath); |
3969 | ADJUST_CHILD_ATTRS(jpath->joinrestrictinfo); |
3970 | ADJUST_CHILD_ATTRS(mpath->path_mergeclauses); |
3971 | new_path = (Path *) mpath; |
3972 | } |
3973 | break; |
3974 | |
3975 | case T_HashPath: |
3976 | { |
3977 | JoinPath *jpath; |
3978 | HashPath *hpath; |
3979 | |
3980 | FLAT_COPY_PATH(hpath, path, HashPath); |
3981 | |
3982 | jpath = (JoinPath *) hpath; |
3983 | REPARAMETERIZE_CHILD_PATH(jpath->outerjoinpath); |
3984 | REPARAMETERIZE_CHILD_PATH(jpath->innerjoinpath); |
3985 | ADJUST_CHILD_ATTRS(jpath->joinrestrictinfo); |
3986 | ADJUST_CHILD_ATTRS(hpath->path_hashclauses); |
3987 | new_path = (Path *) hpath; |
3988 | } |
3989 | break; |
3990 | |
3991 | case T_AppendPath: |
3992 | { |
3993 | AppendPath *apath; |
3994 | |
3995 | FLAT_COPY_PATH(apath, path, AppendPath); |
3996 | REPARAMETERIZE_CHILD_PATH_LIST(apath->subpaths); |
3997 | new_path = (Path *) apath; |
3998 | } |
3999 | break; |
4000 | |
4001 | case T_MergeAppendPath: |
4002 | { |
4003 | MergeAppendPath *mapath; |
4004 | |
4005 | FLAT_COPY_PATH(mapath, path, MergeAppendPath); |
4006 | REPARAMETERIZE_CHILD_PATH_LIST(mapath->subpaths); |
4007 | new_path = (Path *) mapath; |
4008 | } |
4009 | break; |
4010 | |
4011 | case T_MaterialPath: |
4012 | { |
4013 | MaterialPath *mpath; |
4014 | |
4015 | FLAT_COPY_PATH(mpath, path, MaterialPath); |
4016 | REPARAMETERIZE_CHILD_PATH(mpath->subpath); |
4017 | new_path = (Path *) mpath; |
4018 | } |
4019 | break; |
4020 | |
4021 | case T_UniquePath: |
4022 | { |
4023 | UniquePath *upath; |
4024 | |
4025 | FLAT_COPY_PATH(upath, path, UniquePath); |
4026 | REPARAMETERIZE_CHILD_PATH(upath->subpath); |
4027 | ADJUST_CHILD_ATTRS(upath->uniq_exprs); |
4028 | new_path = (Path *) upath; |
4029 | } |
4030 | break; |
4031 | |
4032 | case T_GatherPath: |
4033 | { |
4034 | GatherPath *gpath; |
4035 | |
4036 | FLAT_COPY_PATH(gpath, path, GatherPath); |
4037 | REPARAMETERIZE_CHILD_PATH(gpath->subpath); |
4038 | new_path = (Path *) gpath; |
4039 | } |
4040 | break; |
4041 | |
4042 | case T_GatherMergePath: |
4043 | { |
4044 | GatherMergePath *gmpath; |
4045 | |
4046 | FLAT_COPY_PATH(gmpath, path, GatherMergePath); |
4047 | REPARAMETERIZE_CHILD_PATH(gmpath->subpath); |
4048 | new_path = (Path *) gmpath; |
4049 | } |
4050 | break; |
4051 | |
4052 | default: |
4053 | |
4054 | /* We don't know how to reparameterize this path. */ |
4055 | return NULL; |
4056 | } |
4057 | |
4058 | /* |
4059 | * Adjust the parameterization information, which refers to the topmost |
4060 | * parent. The topmost parent can be multiple levels away from the given |
4061 | * child, hence use multi-level expression adjustment routines. |
4062 | */ |
4063 | old_ppi = new_path->param_info; |
4064 | required_outer = |
4065 | adjust_child_relids_multilevel(root, old_ppi->ppi_req_outer, |
4066 | child_rel->relids, |
4067 | child_rel->top_parent_relids); |
4068 | |
4069 | /* If we already have a PPI for this parameterization, just return it */ |
4070 | new_ppi = find_param_path_info(new_path->parent, required_outer); |
4071 | |
4072 | /* |
4073 | * If not, build a new one and link it to the list of PPIs. For the same |
4074 | * reason as explained in mark_dummy_rel(), allocate new PPI in the same |
4075 | * context the given RelOptInfo is in. |
4076 | */ |
4077 | if (new_ppi == NULL) |
4078 | { |
4079 | MemoryContext oldcontext; |
4080 | RelOptInfo *rel = path->parent; |
4081 | |
4082 | oldcontext = MemoryContextSwitchTo(GetMemoryChunkContext(rel)); |
4083 | |
4084 | new_ppi = makeNode(ParamPathInfo); |
4085 | new_ppi->ppi_req_outer = bms_copy(required_outer); |
4086 | new_ppi->ppi_rows = old_ppi->ppi_rows; |
4087 | new_ppi->ppi_clauses = old_ppi->ppi_clauses; |
4088 | ADJUST_CHILD_ATTRS(new_ppi->ppi_clauses); |
4089 | rel->ppilist = lappend(rel->ppilist, new_ppi); |
4090 | |
4091 | MemoryContextSwitchTo(oldcontext); |
4092 | } |
4093 | bms_free(required_outer); |
4094 | |
4095 | new_path->param_info = new_ppi; |
4096 | |
4097 | /* |
4098 | * Adjust the path target if the parent of the outer relation is |
4099 | * referenced in the targetlist. This can happen when only the parent of |
4100 | * outer relation is laterally referenced in this relation. |
4101 | */ |
4102 | if (bms_overlap(path->parent->lateral_relids, |
4103 | child_rel->top_parent_relids)) |
4104 | { |
4105 | new_path->pathtarget = copy_pathtarget(new_path->pathtarget); |
4106 | ADJUST_CHILD_ATTRS(new_path->pathtarget->exprs); |
4107 | } |
4108 | |
4109 | return new_path; |
4110 | } |
4111 | |
4112 | /* |
4113 | * reparameterize_pathlist_by_child |
4114 | * Helper function to reparameterize a list of paths by given child rel. |
4115 | */ |
4116 | static List * |
4117 | reparameterize_pathlist_by_child(PlannerInfo *root, |
4118 | List *pathlist, |
4119 | RelOptInfo *child_rel) |
4120 | { |
4121 | ListCell *lc; |
4122 | List *result = NIL; |
4123 | |
4124 | foreach(lc, pathlist) |
4125 | { |
4126 | Path *path = reparameterize_path_by_child(root, lfirst(lc), |
4127 | child_rel); |
4128 | |
4129 | if (path == NULL) |
4130 | { |
4131 | list_free(result); |
4132 | return NIL; |
4133 | } |
4134 | |
4135 | result = lappend(result, path); |
4136 | } |
4137 | |
4138 | return result; |
4139 | } |
4140 | |