Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * subselect.c
4 : * Planning routines for subselects.
5 : *
6 : * This module deals with SubLinks and CTEs, but not subquery RTEs (i.e.,
7 : * not sub-SELECT-in-FROM cases).
8 : *
9 : * Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group
10 : * Portions Copyright (c) 1994, Regents of the University of California
11 : *
12 : * IDENTIFICATION
13 : * src/backend/optimizer/plan/subselect.c
14 : *
15 : *-------------------------------------------------------------------------
16 : */
17 : #include "postgres.h"
18 :
19 : #include "access/htup_details.h"
20 : #include "catalog/pg_operator.h"
21 : #include "catalog/pg_type.h"
22 : #include "executor/executor.h"
23 : #include "miscadmin.h"
24 : #include "nodes/makefuncs.h"
25 : #include "nodes/nodeFuncs.h"
26 : #include "optimizer/clauses.h"
27 : #include "optimizer/cost.h"
28 : #include "optimizer/optimizer.h"
29 : #include "optimizer/paramassign.h"
30 : #include "optimizer/pathnode.h"
31 : #include "optimizer/planmain.h"
32 : #include "optimizer/planner.h"
33 : #include "optimizer/prep.h"
34 : #include "optimizer/subselect.h"
35 : #include "parser/parse_relation.h"
36 : #include "rewrite/rewriteManip.h"
37 : #include "utils/builtins.h"
38 : #include "utils/lsyscache.h"
39 : #include "utils/syscache.h"
40 :
41 :
42 : typedef struct convert_testexpr_context
43 : {
44 : PlannerInfo *root;
45 : List *subst_nodes; /* Nodes to substitute for Params */
46 : } convert_testexpr_context;
47 :
48 : typedef struct process_sublinks_context
49 : {
50 : PlannerInfo *root;
51 : bool isTopQual;
52 : } process_sublinks_context;
53 :
54 : typedef struct finalize_primnode_context
55 : {
56 : PlannerInfo *root;
57 : Bitmapset *paramids; /* Non-local PARAM_EXEC paramids found */
58 : } finalize_primnode_context;
59 :
60 : typedef struct inline_cte_walker_context
61 : {
62 : const char *ctename; /* name and relative level of target CTE */
63 : int levelsup;
64 : Query *ctequery; /* query to substitute */
65 : } inline_cte_walker_context;
66 :
67 :
68 : static Node *build_subplan(PlannerInfo *root, Plan *plan, Path *path,
69 : PlannerInfo *subroot, List *plan_params,
70 : SubLinkType subLinkType, int subLinkId,
71 : Node *testexpr, List *testexpr_paramids,
72 : bool unknownEqFalse);
73 : static List *generate_subquery_params(PlannerInfo *root, List *tlist,
74 : List **paramIds);
75 : static List *generate_subquery_vars(PlannerInfo *root, List *tlist,
76 : Index varno);
77 : static Node *convert_testexpr(PlannerInfo *root,
78 : Node *testexpr,
79 : List *subst_nodes);
80 : static Node *convert_testexpr_mutator(Node *node,
81 : convert_testexpr_context *context);
82 : static bool subplan_is_hashable(Plan *plan);
83 : static bool subpath_is_hashable(Path *path);
84 : static bool testexpr_is_hashable(Node *testexpr, List *param_ids);
85 : static bool test_opexpr_is_hashable(OpExpr *testexpr, List *param_ids);
86 : static bool hash_ok_operator(OpExpr *expr);
87 : static bool contain_dml(Node *node);
88 : static bool contain_dml_walker(Node *node, void *context);
89 : static bool contain_outer_selfref(Node *node);
90 : static bool contain_outer_selfref_walker(Node *node, Index *depth);
91 : static void inline_cte(PlannerInfo *root, CommonTableExpr *cte);
92 : static bool inline_cte_walker(Node *node, inline_cte_walker_context *context);
93 : static bool simplify_EXISTS_query(PlannerInfo *root, Query *query);
94 : static Query *convert_EXISTS_to_ANY(PlannerInfo *root, Query *subselect,
95 : Node **testexpr, List **paramIds);
96 : static Node *replace_correlation_vars_mutator(Node *node, PlannerInfo *root);
97 : static Node *process_sublinks_mutator(Node *node,
98 : process_sublinks_context *context);
99 : static Bitmapset *finalize_plan(PlannerInfo *root,
100 : Plan *plan,
101 : int gather_param,
102 : Bitmapset *valid_params,
103 : Bitmapset *scan_params);
104 : static bool finalize_primnode(Node *node, finalize_primnode_context *context);
105 : static bool finalize_agg_primnode(Node *node, finalize_primnode_context *context);
106 :
107 :
108 : /*
109 : * Get the datatype/typmod/collation of the first column of the plan's output.
110 : *
111 : * This information is stored for ARRAY_SUBLINK execution and for
112 : * exprType()/exprTypmod()/exprCollation(), which have no way to get at the
113 : * plan associated with a SubPlan node. We really only need the info for
114 : * EXPR_SUBLINK and ARRAY_SUBLINK subplans, but for consistency we save it
115 : * always.
116 : */
117 : static void
118 45952 : get_first_col_type(Plan *plan, Oid *coltype, int32 *coltypmod,
119 : Oid *colcollation)
120 : {
121 : /* In cases such as EXISTS, tlist might be empty; arbitrarily use VOID */
122 45952 : if (plan->targetlist)
123 : {
124 43032 : TargetEntry *tent = linitial_node(TargetEntry, plan->targetlist);
125 :
126 43032 : if (!tent->resjunk)
127 : {
128 43032 : *coltype = exprType((Node *) tent->expr);
129 43032 : *coltypmod = exprTypmod((Node *) tent->expr);
130 43032 : *colcollation = exprCollation((Node *) tent->expr);
131 43032 : return;
132 : }
133 : }
134 2920 : *coltype = VOIDOID;
135 2920 : *coltypmod = -1;
136 2920 : *colcollation = InvalidOid;
137 : }
138 :
139 : /*
140 : * Convert a SubLink (as created by the parser) into a SubPlan.
141 : *
142 : * We are given the SubLink's contained query, type, ID, and testexpr. We are
143 : * also told if this expression appears at top level of a WHERE/HAVING qual.
144 : *
145 : * Note: we assume that the testexpr has been AND/OR flattened (actually,
146 : * it's been through eval_const_expressions), but not converted to
147 : * implicit-AND form; and any SubLinks in it should already have been
148 : * converted to SubPlans. The subquery is as yet untouched, however.
149 : *
150 : * The result is whatever we need to substitute in place of the SubLink node
151 : * in the executable expression. If we're going to do the subplan as a
152 : * regular subplan, this will be the constructed SubPlan node. If we're going
153 : * to do the subplan as an InitPlan, the SubPlan node instead goes into
154 : * root->init_plans, and what we return here is an expression tree
155 : * representing the InitPlan's result: usually just a Param node representing
156 : * a single scalar result, but possibly a row comparison tree containing
157 : * multiple Param nodes, or for a MULTIEXPR subquery a simple NULL constant
158 : * (since the real output Params are elsewhere in the tree, and the MULTIEXPR
159 : * subquery itself is in a resjunk tlist entry whose value is uninteresting).
160 : */
161 : static Node *
162 41270 : make_subplan(PlannerInfo *root, Query *orig_subquery,
163 : SubLinkType subLinkType, int subLinkId,
164 : Node *testexpr, bool isTopQual)
165 : {
166 : Query *subquery;
167 41270 : bool simple_exists = false;
168 : double tuple_fraction;
169 : PlannerInfo *subroot;
170 : RelOptInfo *final_rel;
171 : Path *best_path;
172 : Plan *plan;
173 : List *plan_params;
174 : Node *result;
175 :
176 : /*
177 : * Copy the source Query node. This is a quick and dirty kluge to resolve
178 : * the fact that the parser can generate trees with multiple links to the
179 : * same sub-Query node, but the planner wants to scribble on the Query.
180 : * Try to clean this up when we do querytree redesign...
181 : */
182 41270 : subquery = copyObject(orig_subquery);
183 :
184 : /*
185 : * If it's an EXISTS subplan, we might be able to simplify it.
186 : */
187 41270 : if (subLinkType == EXISTS_SUBLINK)
188 2668 : simple_exists = simplify_EXISTS_query(root, subquery);
189 :
190 : /*
191 : * For an EXISTS subplan, tell lower-level planner to expect that only the
192 : * first tuple will be retrieved. For ALL and ANY subplans, we will be
193 : * able to stop evaluating if the test condition fails or matches, so very
194 : * often not all the tuples will be retrieved; for lack of a better idea,
195 : * specify 50% retrieval. For EXPR, MULTIEXPR, and ROWCOMPARE subplans,
196 : * use default behavior (we're only expecting one row out, anyway).
197 : *
198 : * NOTE: if you change these numbers, also change cost_subplan() in
199 : * path/costsize.c.
200 : *
201 : * XXX If an ANY subplan is uncorrelated, build_subplan may decide to hash
202 : * its output. In that case it would've been better to specify full
203 : * retrieval. At present, however, we can only check hashability after
204 : * we've made the subplan :-(. (Determining whether it'll fit in hash_mem
205 : * is the really hard part.) Therefore, we don't want to be too
206 : * optimistic about the percentage of tuples retrieved, for fear of
207 : * selecting a plan that's bad for the materialization case.
208 : */
209 41270 : if (subLinkType == EXISTS_SUBLINK)
210 2668 : tuple_fraction = 1.0; /* just like a LIMIT 1 */
211 38602 : else if (subLinkType == ALL_SUBLINK ||
212 : subLinkType == ANY_SUBLINK)
213 550 : tuple_fraction = 0.5; /* 50% */
214 : else
215 38052 : tuple_fraction = 0.0; /* default behavior */
216 :
217 : /* plan_params should not be in use in current query level */
218 : Assert(root->plan_params == NIL);
219 :
220 : /* Generate Paths for the subquery */
221 41270 : subroot = subquery_planner(root->glob, subquery, root, false,
222 : tuple_fraction, NULL);
223 :
224 : /* Isolate the params needed by this specific subplan */
225 41270 : plan_params = root->plan_params;
226 41270 : root->plan_params = NIL;
227 :
228 : /*
229 : * Select best Path and turn it into a Plan. At least for now, there
230 : * seems no reason to postpone doing that.
231 : */
232 41270 : final_rel = fetch_upper_rel(subroot, UPPERREL_FINAL, NULL);
233 41270 : best_path = get_cheapest_fractional_path(final_rel, tuple_fraction);
234 :
235 41270 : plan = create_plan(subroot, best_path);
236 :
237 : /* And convert to SubPlan or InitPlan format. */
238 41270 : result = build_subplan(root, plan, best_path,
239 : subroot, plan_params,
240 : subLinkType, subLinkId,
241 : testexpr, NIL, isTopQual);
242 :
243 : /*
244 : * If it's a correlated EXISTS with an unimportant targetlist, we might be
245 : * able to transform it to the equivalent of an IN and then implement it
246 : * by hashing. We don't have enough information yet to tell which way is
247 : * likely to be better (it depends on the expected number of executions of
248 : * the EXISTS qual, and we are much too early in planning the outer query
249 : * to be able to guess that). So we generate both plans, if possible, and
250 : * leave it to setrefs.c to decide which to use.
251 : */
252 41270 : if (simple_exists && IsA(result, SubPlan))
253 : {
254 : Node *newtestexpr;
255 : List *paramIds;
256 :
257 : /* Make a second copy of the original subquery */
258 2336 : subquery = copyObject(orig_subquery);
259 : /* and re-simplify */
260 2336 : simple_exists = simplify_EXISTS_query(root, subquery);
261 : Assert(simple_exists);
262 : /* See if it can be converted to an ANY query */
263 2336 : subquery = convert_EXISTS_to_ANY(root, subquery,
264 : &newtestexpr, ¶mIds);
265 2336 : if (subquery)
266 : {
267 : /* Generate Paths for the ANY subquery; we'll need all rows */
268 1774 : subroot = subquery_planner(root->glob, subquery, root, false, 0.0,
269 : NULL);
270 :
271 : /* Isolate the params needed by this specific subplan */
272 1774 : plan_params = root->plan_params;
273 1774 : root->plan_params = NIL;
274 :
275 : /* Select best Path */
276 1774 : final_rel = fetch_upper_rel(subroot, UPPERREL_FINAL, NULL);
277 1774 : best_path = final_rel->cheapest_total_path;
278 :
279 : /* Now we can check if it'll fit in hash_mem */
280 1774 : if (subpath_is_hashable(best_path))
281 : {
282 : SubPlan *hashplan;
283 : AlternativeSubPlan *asplan;
284 :
285 : /* OK, finish planning the ANY subquery */
286 1768 : plan = create_plan(subroot, best_path);
287 :
288 : /* ... and convert to SubPlan format */
289 1768 : hashplan = castNode(SubPlan,
290 : build_subplan(root, plan, best_path,
291 : subroot, plan_params,
292 : ANY_SUBLINK, 0,
293 : newtestexpr,
294 : paramIds,
295 : true));
296 : /* Check we got what we expected */
297 : Assert(hashplan->parParam == NIL);
298 : Assert(hashplan->useHashTable);
299 :
300 : /* Leave it to setrefs.c to decide which plan to use */
301 1768 : asplan = makeNode(AlternativeSubPlan);
302 1768 : asplan->subplans = list_make2(result, hashplan);
303 1768 : result = (Node *) asplan;
304 1768 : root->hasAlternativeSubPlans = true;
305 : }
306 : }
307 : }
308 :
309 41270 : return result;
310 : }
311 :
312 : /*
313 : * Build a SubPlan node given the raw inputs --- subroutine for make_subplan
314 : *
315 : * Returns either the SubPlan, or a replacement expression if we decide to
316 : * make it an InitPlan, as explained in the comments for make_subplan.
317 : */
318 : static Node *
319 43038 : build_subplan(PlannerInfo *root, Plan *plan, Path *path,
320 : PlannerInfo *subroot, List *plan_params,
321 : SubLinkType subLinkType, int subLinkId,
322 : Node *testexpr, List *testexpr_paramids,
323 : bool unknownEqFalse)
324 : {
325 : Node *result;
326 : SubPlan *splan;
327 : bool isInitPlan;
328 : ListCell *lc;
329 :
330 : /*
331 : * Initialize the SubPlan node. Note plan_id, plan_name, and cost fields
332 : * are set further down.
333 : */
334 43038 : splan = makeNode(SubPlan);
335 43038 : splan->subLinkType = subLinkType;
336 43038 : splan->testexpr = NULL;
337 43038 : splan->paramIds = NIL;
338 43038 : get_first_col_type(plan, &splan->firstColType, &splan->firstColTypmod,
339 : &splan->firstColCollation);
340 43038 : splan->useHashTable = false;
341 43038 : splan->unknownEqFalse = unknownEqFalse;
342 43038 : splan->parallel_safe = plan->parallel_safe;
343 43038 : splan->setParam = NIL;
344 43038 : splan->parParam = NIL;
345 43038 : splan->args = NIL;
346 :
347 : /*
348 : * Make parParam and args lists of param IDs and expressions that current
349 : * query level will pass to this child plan.
350 : */
351 89472 : foreach(lc, plan_params)
352 : {
353 46434 : PlannerParamItem *pitem = (PlannerParamItem *) lfirst(lc);
354 46434 : Node *arg = pitem->item;
355 :
356 : /*
357 : * The Var, PlaceHolderVar, Aggref, GroupingFunc, or ReturningExpr has
358 : * already been adjusted to have the correct varlevelsup, phlevelsup,
359 : * agglevelsup, or retlevelsup.
360 : *
361 : * If it's a PlaceHolderVar, Aggref, GroupingFunc, or ReturningExpr,
362 : * its arguments might contain SubLinks, which have not yet been
363 : * processed (see the comments for SS_replace_correlation_vars). Do
364 : * that now.
365 : */
366 46434 : if (IsA(arg, PlaceHolderVar) ||
367 46422 : IsA(arg, Aggref) ||
368 46370 : IsA(arg, GroupingFunc) ||
369 46306 : IsA(arg, ReturningExpr))
370 146 : arg = SS_process_sublinks(root, arg, false);
371 :
372 46434 : splan->parParam = lappend_int(splan->parParam, pitem->paramId);
373 46434 : splan->args = lappend(splan->args, arg);
374 : }
375 :
376 : /*
377 : * Un-correlated or undirect correlated plans of EXISTS, EXPR, ARRAY,
378 : * ROWCOMPARE, or MULTIEXPR types can be used as initPlans. For EXISTS,
379 : * EXPR, or ARRAY, we return a Param referring to the result of evaluating
380 : * the initPlan. For ROWCOMPARE, we must modify the testexpr tree to
381 : * contain PARAM_EXEC Params instead of the PARAM_SUBLINK Params emitted
382 : * by the parser, and then return that tree. For MULTIEXPR, we return a
383 : * null constant: the resjunk targetlist item containing the SubLink does
384 : * not need to return anything useful, since the referencing Params are
385 : * elsewhere.
386 : */
387 43038 : if (splan->parParam == NIL && subLinkType == EXISTS_SUBLINK)
388 300 : {
389 : Param *prm;
390 :
391 : Assert(testexpr == NULL);
392 300 : prm = generate_new_exec_param(root, BOOLOID, -1, InvalidOid);
393 300 : splan->setParam = list_make1_int(prm->paramid);
394 300 : isInitPlan = true;
395 300 : result = (Node *) prm;
396 : }
397 42738 : else if (splan->parParam == NIL && subLinkType == EXPR_SUBLINK)
398 10374 : {
399 10374 : TargetEntry *te = linitial(plan->targetlist);
400 : Param *prm;
401 :
402 : Assert(!te->resjunk);
403 : Assert(testexpr == NULL);
404 10374 : prm = generate_new_exec_param(root,
405 10374 : exprType((Node *) te->expr),
406 10374 : exprTypmod((Node *) te->expr),
407 10374 : exprCollation((Node *) te->expr));
408 10374 : splan->setParam = list_make1_int(prm->paramid);
409 10374 : isInitPlan = true;
410 10374 : result = (Node *) prm;
411 : }
412 32364 : else if (splan->parParam == NIL && subLinkType == ARRAY_SUBLINK)
413 122 : {
414 122 : TargetEntry *te = linitial(plan->targetlist);
415 : Oid arraytype;
416 : Param *prm;
417 :
418 : Assert(!te->resjunk);
419 : Assert(testexpr == NULL);
420 122 : arraytype = get_promoted_array_type(exprType((Node *) te->expr));
421 122 : if (!OidIsValid(arraytype))
422 0 : elog(ERROR, "could not find array type for datatype %s",
423 : format_type_be(exprType((Node *) te->expr)));
424 122 : prm = generate_new_exec_param(root,
425 : arraytype,
426 122 : exprTypmod((Node *) te->expr),
427 122 : exprCollation((Node *) te->expr));
428 122 : splan->setParam = list_make1_int(prm->paramid);
429 122 : isInitPlan = true;
430 122 : result = (Node *) prm;
431 : }
432 32242 : else if (splan->parParam == NIL && subLinkType == ROWCOMPARE_SUBLINK)
433 18 : {
434 : /* Adjust the Params */
435 : List *params;
436 :
437 : Assert(testexpr != NULL);
438 18 : params = generate_subquery_params(root,
439 : plan->targetlist,
440 : &splan->paramIds);
441 18 : result = convert_testexpr(root,
442 : testexpr,
443 : params);
444 18 : splan->setParam = list_copy(splan->paramIds);
445 18 : isInitPlan = true;
446 :
447 : /*
448 : * The executable expression is returned to become part of the outer
449 : * plan's expression tree; it is not kept in the initplan node.
450 : */
451 : }
452 32224 : else if (subLinkType == MULTIEXPR_SUBLINK)
453 : {
454 : /*
455 : * Whether it's an initplan or not, it needs to set a PARAM_EXEC Param
456 : * for each output column.
457 : */
458 : List *params;
459 :
460 : Assert(testexpr == NULL);
461 132 : params = generate_subquery_params(root,
462 : plan->targetlist,
463 : &splan->setParam);
464 :
465 : /*
466 : * Save the list of replacement Params in the n'th cell of
467 : * root->multiexpr_params; setrefs.c will use it to replace
468 : * PARAM_MULTIEXPR Params.
469 : */
470 264 : while (list_length(root->multiexpr_params) < subLinkId)
471 132 : root->multiexpr_params = lappend(root->multiexpr_params, NIL);
472 132 : lc = list_nth_cell(root->multiexpr_params, subLinkId - 1);
473 : Assert(lfirst(lc) == NIL);
474 132 : lfirst(lc) = params;
475 :
476 : /* It can be an initplan if there are no parParams. */
477 132 : if (splan->parParam == NIL)
478 : {
479 30 : isInitPlan = true;
480 30 : result = (Node *) makeNullConst(RECORDOID, -1, InvalidOid);
481 : }
482 : else
483 : {
484 102 : isInitPlan = false;
485 102 : result = (Node *) splan;
486 : }
487 : }
488 : else
489 : {
490 : /*
491 : * Adjust the Params in the testexpr, unless caller already took care
492 : * of it (as indicated by passing a list of Param IDs).
493 : */
494 32092 : if (testexpr && testexpr_paramids == NIL)
495 562 : {
496 : List *params;
497 :
498 562 : params = generate_subquery_params(root,
499 : plan->targetlist,
500 : &splan->paramIds);
501 562 : splan->testexpr = convert_testexpr(root,
502 : testexpr,
503 : params);
504 : }
505 : else
506 : {
507 31530 : splan->testexpr = testexpr;
508 31530 : splan->paramIds = testexpr_paramids;
509 : }
510 :
511 : /*
512 : * We can't convert subplans of ALL_SUBLINK or ANY_SUBLINK types to
513 : * initPlans, even when they are uncorrelated or undirect correlated,
514 : * because we need to scan the output of the subplan for each outer
515 : * tuple. But if it's a not-direct-correlated IN (= ANY) test, we
516 : * might be able to use a hashtable to avoid comparing all the tuples.
517 : */
518 32092 : if (subLinkType == ANY_SUBLINK &&
519 4498 : splan->parParam == NIL &&
520 4396 : subplan_is_hashable(plan) &&
521 2198 : testexpr_is_hashable(splan->testexpr, splan->paramIds))
522 2174 : splan->useHashTable = true;
523 :
524 : /*
525 : * Otherwise, we have the option to tack a Material node onto the top
526 : * of the subplan, to reduce the cost of reading it repeatedly. This
527 : * is pointless for a direct-correlated subplan, since we'd have to
528 : * recompute its results each time anyway. For uncorrelated/undirect
529 : * correlated subplans, we add Material unless the subplan's top plan
530 : * node would materialize its output anyway. Also, if enable_material
531 : * is false, then the user does not want us to materialize anything
532 : * unnecessarily, so we don't.
533 : */
534 29918 : else if (splan->parParam == NIL && enable_material &&
535 42 : !ExecMaterializesOutput(nodeTag(plan)))
536 42 : plan = materialize_finished_plan(plan);
537 :
538 32092 : result = (Node *) splan;
539 32092 : isInitPlan = false;
540 : }
541 :
542 : /*
543 : * Add the subplan, its path, and its PlannerInfo to the global lists.
544 : */
545 43038 : root->glob->subplans = lappend(root->glob->subplans, plan);
546 43038 : root->glob->subpaths = lappend(root->glob->subpaths, path);
547 43038 : root->glob->subroots = lappend(root->glob->subroots, subroot);
548 43038 : splan->plan_id = list_length(root->glob->subplans);
549 :
550 43038 : if (isInitPlan)
551 10844 : root->init_plans = lappend(root->init_plans, splan);
552 :
553 : /*
554 : * A parameterless subplan (not initplan) should be prepared to handle
555 : * REWIND efficiently. If it has direct parameters then there's no point
556 : * since it'll be reset on each scan anyway; and if it's an initplan then
557 : * there's no point since it won't get re-run without parameter changes
558 : * anyway. The input of a hashed subplan doesn't need REWIND either.
559 : */
560 43038 : if (splan->parParam == NIL && !isInitPlan && !splan->useHashTable)
561 42 : root->glob->rewindPlanIDs = bms_add_member(root->glob->rewindPlanIDs,
562 : splan->plan_id);
563 :
564 : /* Label the subplan for EXPLAIN purposes */
565 43038 : splan->plan_name = psprintf("%s %d",
566 : isInitPlan ? "InitPlan" : "SubPlan",
567 : splan->plan_id);
568 :
569 : /* Lastly, fill in the cost estimates for use later */
570 43038 : cost_subplan(root, splan, plan);
571 :
572 43038 : return result;
573 : }
574 :
575 : /*
576 : * generate_subquery_params: build a list of Params representing the output
577 : * columns of a sublink's sub-select, given the sub-select's targetlist.
578 : *
579 : * We also return an integer list of the paramids of the Params.
580 : */
581 : static List *
582 712 : generate_subquery_params(PlannerInfo *root, List *tlist, List **paramIds)
583 : {
584 : List *result;
585 : List *ids;
586 : ListCell *lc;
587 :
588 712 : result = ids = NIL;
589 1672 : foreach(lc, tlist)
590 : {
591 960 : TargetEntry *tent = (TargetEntry *) lfirst(lc);
592 : Param *param;
593 :
594 960 : if (tent->resjunk)
595 6 : continue;
596 :
597 954 : param = generate_new_exec_param(root,
598 954 : exprType((Node *) tent->expr),
599 954 : exprTypmod((Node *) tent->expr),
600 954 : exprCollation((Node *) tent->expr));
601 954 : result = lappend(result, param);
602 954 : ids = lappend_int(ids, param->paramid);
603 : }
604 :
605 712 : *paramIds = ids;
606 712 : return result;
607 : }
608 :
609 : /*
610 : * generate_subquery_vars: build a list of Vars representing the output
611 : * columns of a sublink's sub-select, given the sub-select's targetlist.
612 : * The Vars have the specified varno (RTE index).
613 : */
614 : static List *
615 4336 : generate_subquery_vars(PlannerInfo *root, List *tlist, Index varno)
616 : {
617 : List *result;
618 : ListCell *lc;
619 :
620 4336 : result = NIL;
621 8726 : foreach(lc, tlist)
622 : {
623 4390 : TargetEntry *tent = (TargetEntry *) lfirst(lc);
624 : Var *var;
625 :
626 4390 : if (tent->resjunk)
627 0 : continue;
628 :
629 4390 : var = makeVarFromTargetEntry(varno, tent);
630 4390 : result = lappend(result, var);
631 : }
632 :
633 4336 : return result;
634 : }
635 :
636 : /*
637 : * convert_testexpr: convert the testexpr given by the parser into
638 : * actually executable form. This entails replacing PARAM_SUBLINK Params
639 : * with Params or Vars representing the results of the sub-select. The
640 : * nodes to be substituted are passed in as the List result from
641 : * generate_subquery_params or generate_subquery_vars.
642 : */
643 : static Node *
644 5192 : convert_testexpr(PlannerInfo *root,
645 : Node *testexpr,
646 : List *subst_nodes)
647 : {
648 : convert_testexpr_context context;
649 :
650 5192 : context.root = root;
651 5192 : context.subst_nodes = subst_nodes;
652 5192 : return convert_testexpr_mutator(testexpr, &context);
653 : }
654 :
655 : static Node *
656 24702 : convert_testexpr_mutator(Node *node,
657 : convert_testexpr_context *context)
658 : {
659 24702 : if (node == NULL)
660 46 : return NULL;
661 24656 : if (IsA(node, Param))
662 : {
663 5376 : Param *param = (Param *) node;
664 :
665 5376 : if (param->paramkind == PARAM_SUBLINK)
666 : {
667 10740 : if (param->paramid <= 0 ||
668 5370 : param->paramid > list_length(context->subst_nodes))
669 0 : elog(ERROR, "unexpected PARAM_SUBLINK ID: %d", param->paramid);
670 :
671 : /*
672 : * We copy the list item to avoid having doubly-linked
673 : * substructure in the modified parse tree. This is probably
674 : * unnecessary when it's a Param, but be safe.
675 : */
676 5370 : return (Node *) copyObject(list_nth(context->subst_nodes,
677 : param->paramid - 1));
678 : }
679 : }
680 19286 : if (IsA(node, SubLink))
681 : {
682 : /*
683 : * If we come across a nested SubLink, it is neither necessary nor
684 : * correct to recurse into it: any PARAM_SUBLINKs we might find inside
685 : * belong to the inner SubLink not the outer. So just return it as-is.
686 : *
687 : * This reasoning depends on the assumption that nothing will pull
688 : * subexpressions into or out of the testexpr field of a SubLink, at
689 : * least not without replacing PARAM_SUBLINKs first. If we did want
690 : * to do that we'd need to rethink the parser-output representation
691 : * altogether, since currently PARAM_SUBLINKs are only unique per
692 : * SubLink not globally across the query. The whole point of
693 : * replacing them with Vars or PARAM_EXEC nodes is to make them
694 : * globally unique before they escape from the SubLink's testexpr.
695 : *
696 : * Note: this can't happen when called during SS_process_sublinks,
697 : * because that recursively processes inner SubLinks first. It can
698 : * happen when called from convert_ANY_sublink_to_join, though.
699 : */
700 12 : return node;
701 : }
702 19274 : return expression_tree_mutator(node, convert_testexpr_mutator, context);
703 : }
704 :
705 : /*
706 : * subplan_is_hashable: can we implement an ANY subplan by hashing?
707 : *
708 : * This is not responsible for checking whether the combining testexpr
709 : * is suitable for hashing. We only look at the subquery itself.
710 : */
711 : static bool
712 2198 : subplan_is_hashable(Plan *plan)
713 : {
714 : double subquery_size;
715 :
716 : /*
717 : * The estimated size of the subquery result must fit in hash_mem. (Note:
718 : * we use heap tuple overhead here even though the tuples will actually be
719 : * stored as MinimalTuples; this provides some fudge factor for hashtable
720 : * overhead.)
721 : */
722 2198 : subquery_size = plan->plan_rows *
723 2198 : (MAXALIGN(plan->plan_width) + MAXALIGN(SizeofHeapTupleHeader));
724 2198 : if (subquery_size > get_hash_memory_limit())
725 0 : return false;
726 :
727 2198 : return true;
728 : }
729 :
730 : /*
731 : * subpath_is_hashable: can we implement an ANY subplan by hashing?
732 : *
733 : * Identical to subplan_is_hashable, but work from a Path for the subplan.
734 : */
735 : static bool
736 1774 : subpath_is_hashable(Path *path)
737 : {
738 : double subquery_size;
739 :
740 : /*
741 : * The estimated size of the subquery result must fit in hash_mem. (Note:
742 : * we use heap tuple overhead here even though the tuples will actually be
743 : * stored as MinimalTuples; this provides some fudge factor for hashtable
744 : * overhead.)
745 : */
746 1774 : subquery_size = path->rows *
747 1774 : (MAXALIGN(path->pathtarget->width) + MAXALIGN(SizeofHeapTupleHeader));
748 1774 : if (subquery_size > get_hash_memory_limit())
749 6 : return false;
750 :
751 1768 : return true;
752 : }
753 :
754 : /*
755 : * testexpr_is_hashable: is an ANY SubLink's test expression hashable?
756 : *
757 : * To identify LHS vs RHS of the hash expression, we must be given the
758 : * list of output Param IDs of the SubLink's subquery.
759 : */
760 : static bool
761 2198 : testexpr_is_hashable(Node *testexpr, List *param_ids)
762 : {
763 : /*
764 : * The testexpr must be a single OpExpr, or an AND-clause containing only
765 : * OpExprs, each of which satisfy test_opexpr_is_hashable().
766 : */
767 2198 : if (testexpr && IsA(testexpr, OpExpr))
768 : {
769 1332 : if (test_opexpr_is_hashable((OpExpr *) testexpr, param_ids))
770 1308 : return true;
771 : }
772 866 : else if (is_andclause(testexpr))
773 : {
774 : ListCell *l;
775 :
776 2598 : foreach(l, ((BoolExpr *) testexpr)->args)
777 : {
778 1732 : Node *andarg = (Node *) lfirst(l);
779 :
780 1732 : if (!IsA(andarg, OpExpr))
781 0 : return false;
782 1732 : if (!test_opexpr_is_hashable((OpExpr *) andarg, param_ids))
783 0 : return false;
784 : }
785 866 : return true;
786 : }
787 :
788 24 : return false;
789 : }
790 :
791 : static bool
792 3064 : test_opexpr_is_hashable(OpExpr *testexpr, List *param_ids)
793 : {
794 : /*
795 : * The combining operator must be hashable and strict. The need for
796 : * hashability is obvious, since we want to use hashing. Without
797 : * strictness, behavior in the presence of nulls is too unpredictable. We
798 : * actually must assume even more than plain strictness: it can't yield
799 : * NULL for non-null inputs, either (see nodeSubplan.c). However, hash
800 : * indexes and hash joins assume that too.
801 : */
802 3064 : if (!hash_ok_operator(testexpr))
803 12 : return false;
804 :
805 : /*
806 : * The left and right inputs must belong to the outer and inner queries
807 : * respectively; hence Params that will be supplied by the subquery must
808 : * not appear in the LHS, and Vars of the outer query must not appear in
809 : * the RHS. (Ordinarily, this must be true because of the way that the
810 : * parser builds an ANY SubLink's testexpr ... but inlining of functions
811 : * could have changed the expression's structure, so we have to check.
812 : * Such cases do not occur often enough to be worth trying to optimize, so
813 : * we don't worry about trying to commute the clause or anything like
814 : * that; we just need to be sure not to build an invalid plan.)
815 : */
816 3052 : if (list_length(testexpr->args) != 2)
817 0 : return false;
818 3052 : if (contain_exec_param((Node *) linitial(testexpr->args), param_ids))
819 12 : return false;
820 3040 : if (contain_var_clause((Node *) lsecond(testexpr->args)))
821 0 : return false;
822 3040 : return true;
823 : }
824 :
825 : /*
826 : * Check expression is hashable + strict
827 : *
828 : * We could use op_hashjoinable() and op_strict(), but do it like this to
829 : * avoid a redundant cache lookup.
830 : */
831 : static bool
832 9446 : hash_ok_operator(OpExpr *expr)
833 : {
834 9446 : Oid opid = expr->opno;
835 :
836 : /* quick out if not a binary operator */
837 9446 : if (list_length(expr->args) != 2)
838 0 : return false;
839 9446 : if (opid == ARRAY_EQ_OP ||
840 : opid == RECORD_EQ_OP)
841 : {
842 : /* these are strict, but must check input type to ensure hashable */
843 12 : Node *leftarg = linitial(expr->args);
844 :
845 12 : return op_hashjoinable(opid, exprType(leftarg));
846 : }
847 : else
848 : {
849 : /* else must look up the operator properties */
850 : HeapTuple tup;
851 : Form_pg_operator optup;
852 :
853 9434 : tup = SearchSysCache1(OPEROID, ObjectIdGetDatum(opid));
854 9434 : if (!HeapTupleIsValid(tup))
855 0 : elog(ERROR, "cache lookup failed for operator %u", opid);
856 9434 : optup = (Form_pg_operator) GETSTRUCT(tup);
857 9434 : if (!optup->oprcanhash || !func_strict(optup->oprcode))
858 : {
859 992 : ReleaseSysCache(tup);
860 992 : return false;
861 : }
862 8442 : ReleaseSysCache(tup);
863 8442 : return true;
864 : }
865 : }
866 :
867 :
868 : /*
869 : * SS_process_ctes: process a query's WITH list
870 : *
871 : * Consider each CTE in the WITH list and either ignore it (if it's an
872 : * unreferenced SELECT), "inline" it to create a regular sub-SELECT-in-FROM,
873 : * or convert it to an initplan.
874 : *
875 : * A side effect is to fill in root->cte_plan_ids with a list that
876 : * parallels root->parse->cteList and provides the subplan ID for
877 : * each CTE's initplan, or a dummy ID (-1) if we didn't make an initplan.
878 : */
879 : void
880 2910 : SS_process_ctes(PlannerInfo *root)
881 : {
882 : ListCell *lc;
883 :
884 : Assert(root->cte_plan_ids == NIL);
885 :
886 6960 : foreach(lc, root->parse->cteList)
887 : {
888 4056 : CommonTableExpr *cte = (CommonTableExpr *) lfirst(lc);
889 4056 : CmdType cmdType = ((Query *) cte->ctequery)->commandType;
890 : Query *subquery;
891 : PlannerInfo *subroot;
892 : RelOptInfo *final_rel;
893 : Path *best_path;
894 : Plan *plan;
895 : SubPlan *splan;
896 : int paramid;
897 :
898 : /*
899 : * Ignore SELECT CTEs that are not actually referenced anywhere.
900 : */
901 4056 : if (cte->cterefcount == 0 && cmdType == CMD_SELECT)
902 : {
903 : /* Make a dummy entry in cte_plan_ids */
904 52 : root->cte_plan_ids = lappend_int(root->cte_plan_ids, -1);
905 1536 : continue;
906 : }
907 :
908 : /*
909 : * Consider inlining the CTE (creating RTE_SUBQUERY RTE(s)) instead of
910 : * implementing it as a separately-planned CTE.
911 : *
912 : * We cannot inline if any of these conditions hold:
913 : *
914 : * 1. The user said not to (the CTEMaterializeAlways option).
915 : *
916 : * 2. The CTE is recursive.
917 : *
918 : * 3. The CTE has side-effects; this includes either not being a plain
919 : * SELECT, or containing volatile functions. Inlining might change
920 : * the side-effects, which would be bad.
921 : *
922 : * 4. The CTE is multiply-referenced and contains a self-reference to
923 : * a recursive CTE outside itself. Inlining would result in multiple
924 : * recursive self-references, which we don't support.
925 : *
926 : * Otherwise, we have an option whether to inline or not. That should
927 : * always be a win if there's just a single reference, but if the CTE
928 : * is multiply-referenced then it's unclear: inlining adds duplicate
929 : * computations, but the ability to absorb restrictions from the outer
930 : * query level could outweigh that. We do not have nearly enough
931 : * information at this point to tell whether that's true, so we let
932 : * the user express a preference. Our default behavior is to inline
933 : * only singly-referenced CTEs, but a CTE marked CTEMaterializeNever
934 : * will be inlined even if multiply referenced.
935 : *
936 : * Note: we check for volatile functions last, because that's more
937 : * expensive than the other tests needed.
938 : */
939 4004 : if ((cte->ctematerialized == CTEMaterializeNever ||
940 3956 : (cte->ctematerialized == CTEMaterializeDefault &&
941 3754 : cte->cterefcount == 1)) &&
942 2796 : !cte->cterecursive &&
943 1576 : cmdType == CMD_SELECT &&
944 1576 : !contain_dml(cte->ctequery) &&
945 1568 : (cte->cterefcount <= 1 ||
946 36 : !contain_outer_selfref(cte->ctequery)) &&
947 1556 : !contain_volatile_functions(cte->ctequery))
948 : {
949 1484 : inline_cte(root, cte);
950 : /* Make a dummy entry in cte_plan_ids */
951 1484 : root->cte_plan_ids = lappend_int(root->cte_plan_ids, -1);
952 1484 : continue;
953 : }
954 :
955 : /*
956 : * Copy the source Query node. Probably not necessary, but let's keep
957 : * this similar to make_subplan.
958 : */
959 2520 : subquery = (Query *) copyObject(cte->ctequery);
960 :
961 : /* plan_params should not be in use in current query level */
962 : Assert(root->plan_params == NIL);
963 :
964 : /*
965 : * Generate Paths for the CTE query. Always plan for full retrieval
966 : * --- we don't have enough info to predict otherwise.
967 : */
968 2520 : subroot = subquery_planner(root->glob, subquery, root,
969 2520 : cte->cterecursive, 0.0, NULL);
970 :
971 : /*
972 : * Since the current query level doesn't yet contain any RTEs, it
973 : * should not be possible for the CTE to have requested parameters of
974 : * this level.
975 : */
976 2514 : if (root->plan_params)
977 0 : elog(ERROR, "unexpected outer reference in CTE query");
978 :
979 : /*
980 : * Select best Path and turn it into a Plan. At least for now, there
981 : * seems no reason to postpone doing that.
982 : */
983 2514 : final_rel = fetch_upper_rel(subroot, UPPERREL_FINAL, NULL);
984 2514 : best_path = final_rel->cheapest_total_path;
985 :
986 2514 : plan = create_plan(subroot, best_path);
987 :
988 : /*
989 : * Make a SubPlan node for it. This is just enough unlike
990 : * build_subplan that we can't share code.
991 : *
992 : * Note plan_id, plan_name, and cost fields are set further down.
993 : */
994 2514 : splan = makeNode(SubPlan);
995 2514 : splan->subLinkType = CTE_SUBLINK;
996 2514 : splan->testexpr = NULL;
997 2514 : splan->paramIds = NIL;
998 2514 : get_first_col_type(plan, &splan->firstColType, &splan->firstColTypmod,
999 : &splan->firstColCollation);
1000 2514 : splan->useHashTable = false;
1001 2514 : splan->unknownEqFalse = false;
1002 :
1003 : /*
1004 : * CTE scans are not considered for parallelism (cf
1005 : * set_rel_consider_parallel).
1006 : */
1007 2514 : splan->parallel_safe = false;
1008 2514 : splan->setParam = NIL;
1009 2514 : splan->parParam = NIL;
1010 2514 : splan->args = NIL;
1011 :
1012 : /*
1013 : * The node can't have any inputs (since it's an initplan), so the
1014 : * parParam and args lists remain empty. (It could contain references
1015 : * to earlier CTEs' output param IDs, but CTE outputs are not
1016 : * propagated via the args list.)
1017 : */
1018 :
1019 : /*
1020 : * Assign a param ID to represent the CTE's output. No ordinary
1021 : * "evaluation" of this param slot ever happens, but we use the param
1022 : * ID for setParam/chgParam signaling just as if the CTE plan were
1023 : * returning a simple scalar output. (Also, the executor abuses the
1024 : * ParamExecData slot for this param ID for communication among
1025 : * multiple CteScan nodes that might be scanning this CTE.)
1026 : */
1027 2514 : paramid = assign_special_exec_param(root);
1028 2514 : splan->setParam = list_make1_int(paramid);
1029 :
1030 : /*
1031 : * Add the subplan, its path, and its PlannerInfo to the global lists.
1032 : */
1033 2514 : root->glob->subplans = lappend(root->glob->subplans, plan);
1034 2514 : root->glob->subpaths = lappend(root->glob->subpaths, best_path);
1035 2514 : root->glob->subroots = lappend(root->glob->subroots, subroot);
1036 2514 : splan->plan_id = list_length(root->glob->subplans);
1037 :
1038 2514 : root->init_plans = lappend(root->init_plans, splan);
1039 :
1040 2514 : root->cte_plan_ids = lappend_int(root->cte_plan_ids, splan->plan_id);
1041 :
1042 : /* Label the subplan for EXPLAIN purposes */
1043 2514 : splan->plan_name = psprintf("CTE %s", cte->ctename);
1044 :
1045 : /* Lastly, fill in the cost estimates for use later */
1046 2514 : cost_subplan(root, splan, plan);
1047 : }
1048 2904 : }
1049 :
1050 : /*
1051 : * contain_dml: is any subquery not a plain SELECT?
1052 : *
1053 : * We reject SELECT FOR UPDATE/SHARE as well as INSERT etc.
1054 : */
1055 : static bool
1056 1576 : contain_dml(Node *node)
1057 : {
1058 1576 : return contain_dml_walker(node, NULL);
1059 : }
1060 :
1061 : static bool
1062 105790 : contain_dml_walker(Node *node, void *context)
1063 : {
1064 105790 : if (node == NULL)
1065 36480 : return false;
1066 69310 : if (IsA(node, Query))
1067 : {
1068 2870 : Query *query = (Query *) node;
1069 :
1070 2870 : if (query->commandType != CMD_SELECT ||
1071 2870 : query->rowMarks != NIL)
1072 8 : return true;
1073 :
1074 2862 : return query_tree_walker(query, contain_dml_walker, context, 0);
1075 : }
1076 66440 : return expression_tree_walker(node, contain_dml_walker, context);
1077 : }
1078 :
1079 : /*
1080 : * contain_outer_selfref: is there an external recursive self-reference?
1081 : */
1082 : static bool
1083 36 : contain_outer_selfref(Node *node)
1084 : {
1085 36 : Index depth = 0;
1086 :
1087 : /*
1088 : * We should be starting with a Query, so that depth will be 1 while
1089 : * examining its immediate contents.
1090 : */
1091 : Assert(IsA(node, Query));
1092 :
1093 36 : return contain_outer_selfref_walker(node, &depth);
1094 : }
1095 :
1096 : static bool
1097 810 : contain_outer_selfref_walker(Node *node, Index *depth)
1098 : {
1099 810 : if (node == NULL)
1100 486 : return false;
1101 324 : if (IsA(node, RangeTblEntry))
1102 : {
1103 30 : RangeTblEntry *rte = (RangeTblEntry *) node;
1104 :
1105 : /*
1106 : * Check for a self-reference to a CTE that's above the Query that our
1107 : * search started at.
1108 : */
1109 30 : if (rte->rtekind == RTE_CTE &&
1110 12 : rte->self_reference &&
1111 12 : rte->ctelevelsup >= *depth)
1112 12 : return true;
1113 18 : return false; /* allow range_table_walker to continue */
1114 : }
1115 294 : if (IsA(node, Query))
1116 : {
1117 : /* Recurse into subquery, tracking nesting depth properly */
1118 42 : Query *query = (Query *) node;
1119 : bool result;
1120 :
1121 42 : (*depth)++;
1122 :
1123 42 : result = query_tree_walker(query, contain_outer_selfref_walker,
1124 : depth, QTW_EXAMINE_RTES_BEFORE);
1125 :
1126 42 : (*depth)--;
1127 :
1128 42 : return result;
1129 : }
1130 252 : return expression_tree_walker(node, contain_outer_selfref_walker, depth);
1131 : }
1132 :
1133 : /*
1134 : * inline_cte: convert RTE_CTE references to given CTE into RTE_SUBQUERYs
1135 : */
1136 : static void
1137 1484 : inline_cte(PlannerInfo *root, CommonTableExpr *cte)
1138 : {
1139 : struct inline_cte_walker_context context;
1140 :
1141 1484 : context.ctename = cte->ctename;
1142 : /* Start at levelsup = -1 because we'll immediately increment it */
1143 1484 : context.levelsup = -1;
1144 1484 : context.ctequery = castNode(Query, cte->ctequery);
1145 :
1146 1484 : (void) inline_cte_walker((Node *) root->parse, &context);
1147 1484 : }
1148 :
1149 : static bool
1150 540442 : inline_cte_walker(Node *node, inline_cte_walker_context *context)
1151 : {
1152 540442 : if (node == NULL)
1153 145852 : return false;
1154 394590 : if (IsA(node, Query))
1155 : {
1156 10686 : Query *query = (Query *) node;
1157 :
1158 10686 : context->levelsup++;
1159 :
1160 : /*
1161 : * Visit the query's RTE nodes after their contents; otherwise
1162 : * query_tree_walker would descend into the newly inlined CTE query,
1163 : * which we don't want.
1164 : */
1165 10686 : (void) query_tree_walker(query, inline_cte_walker, context,
1166 : QTW_EXAMINE_RTES_AFTER);
1167 :
1168 10686 : context->levelsup--;
1169 :
1170 10686 : return false;
1171 : }
1172 383904 : else if (IsA(node, RangeTblEntry))
1173 : {
1174 20038 : RangeTblEntry *rte = (RangeTblEntry *) node;
1175 :
1176 20038 : if (rte->rtekind == RTE_CTE &&
1177 5916 : strcmp(rte->ctename, context->ctename) == 0 &&
1178 1514 : rte->ctelevelsup == context->levelsup)
1179 : {
1180 : /*
1181 : * Found a reference to replace. Generate a copy of the CTE query
1182 : * with appropriate level adjustment for outer references (e.g.,
1183 : * to other CTEs).
1184 : */
1185 1508 : Query *newquery = copyObject(context->ctequery);
1186 :
1187 1508 : if (context->levelsup > 0)
1188 946 : IncrementVarSublevelsUp((Node *) newquery, context->levelsup, 1);
1189 :
1190 : /*
1191 : * Convert the RTE_CTE RTE into a RTE_SUBQUERY.
1192 : *
1193 : * Historically, a FOR UPDATE clause has been treated as extending
1194 : * into views and subqueries, but not into CTEs. We preserve this
1195 : * distinction by not trying to push rowmarks into the new
1196 : * subquery.
1197 : */
1198 1508 : rte->rtekind = RTE_SUBQUERY;
1199 1508 : rte->subquery = newquery;
1200 1508 : rte->security_barrier = false;
1201 :
1202 : /* Zero out CTE-specific fields */
1203 1508 : rte->ctename = NULL;
1204 1508 : rte->ctelevelsup = 0;
1205 1508 : rte->self_reference = false;
1206 1508 : rte->coltypes = NIL;
1207 1508 : rte->coltypmods = NIL;
1208 1508 : rte->colcollations = NIL;
1209 : }
1210 :
1211 20038 : return false;
1212 : }
1213 :
1214 363866 : return expression_tree_walker(node, inline_cte_walker, context);
1215 : }
1216 :
1217 : /*
1218 : * Attempt to transform 'testexpr' over the VALUES subquery into
1219 : * a ScalarArrayOpExpr. We currently support the transformation only when
1220 : * it ends up with a constant array. Otherwise, the evaluation of non-hashed
1221 : * SAOP might be slower than the corresponding Hash Join with VALUES.
1222 : *
1223 : * Return transformed ScalarArrayOpExpr or NULL if transformation isn't
1224 : * allowed.
1225 : */
1226 : ScalarArrayOpExpr *
1227 4526 : convert_VALUES_to_ANY(PlannerInfo *root, Node *testexpr, Query *values)
1228 : {
1229 : RangeTblEntry *rte;
1230 : Node *leftop;
1231 : Node *rightop;
1232 : Oid opno;
1233 : ListCell *lc;
1234 : Oid inputcollid;
1235 4526 : List *exprs = NIL;
1236 :
1237 : /*
1238 : * Check we have a binary operator over a single-column subquery with no
1239 : * joins and no LIMIT/OFFSET/ORDER BY clauses.
1240 : */
1241 8970 : if (!IsA(testexpr, OpExpr) ||
1242 8888 : list_length(((OpExpr *) testexpr)->args) != 2 ||
1243 4444 : list_length(values->targetList) > 1 ||
1244 4444 : values->limitCount != NULL ||
1245 4432 : values->limitOffset != NULL ||
1246 8810 : values->sortClause != NIL ||
1247 4402 : list_length(values->rtable) != 1)
1248 3634 : return NULL;
1249 :
1250 892 : rte = linitial_node(RangeTblEntry, values->rtable);
1251 892 : leftop = linitial(((OpExpr *) testexpr)->args);
1252 892 : rightop = lsecond(((OpExpr *) testexpr)->args);
1253 892 : opno = ((OpExpr *) testexpr)->opno;
1254 892 : inputcollid = ((OpExpr *) testexpr)->inputcollid;
1255 :
1256 : /*
1257 : * Also, check that only RTE corresponds to VALUES; the list of values has
1258 : * at least two items and no volatile functions.
1259 : */
1260 1024 : if (rte->rtekind != RTE_VALUES ||
1261 252 : list_length(rte->values_lists) < 2 ||
1262 120 : contain_volatile_functions((Node *) rte->values_lists))
1263 772 : return NULL;
1264 :
1265 360 : foreach(lc, rte->values_lists)
1266 : {
1267 276 : List *elem = lfirst(lc);
1268 276 : Node *value = linitial(elem);
1269 :
1270 : /*
1271 : * Prepare an evaluation of the right side of the operator with
1272 : * substitution of the given value.
1273 : */
1274 276 : value = convert_testexpr(root, rightop, list_make1(value));
1275 :
1276 : /*
1277 : * Try to evaluate constant expressions. We could get Const as a
1278 : * result.
1279 : */
1280 276 : value = eval_const_expressions(root, value);
1281 :
1282 : /*
1283 : * As we only support constant output arrays, all the items must also
1284 : * be constant.
1285 : */
1286 276 : if (!IsA(value, Const))
1287 36 : return NULL;
1288 :
1289 240 : exprs = lappend(exprs, value);
1290 : }
1291 :
1292 : /* Finally, build ScalarArrayOpExpr at the top of the 'exprs' list. */
1293 84 : return make_SAOP_expr(opno, leftop, exprType(rightop),
1294 84 : linitial_oid(rte->colcollations), inputcollid,
1295 : exprs, false);
1296 : }
1297 :
1298 : /*
1299 : * convert_ANY_sublink_to_join: try to convert an ANY SubLink to a join
1300 : *
1301 : * The caller has found an ANY SubLink at the top level of one of the query's
1302 : * qual clauses, but has not checked the properties of the SubLink further.
1303 : * Decide whether it is appropriate to process this SubLink in join style.
1304 : * If so, form a JoinExpr and return it. Return NULL if the SubLink cannot
1305 : * be converted to a join.
1306 : *
1307 : * The only non-obvious input parameter is available_rels: this is the set
1308 : * of query rels that can safely be referenced in the sublink expression.
1309 : * (We must restrict this to avoid changing the semantics when a sublink
1310 : * is present in an outer join's ON qual.) The conversion must fail if
1311 : * the converted qual would reference any but these parent-query relids.
1312 : *
1313 : * On success, the returned JoinExpr has larg = NULL and rarg = the jointree
1314 : * item representing the pulled-up subquery. The caller must set larg to
1315 : * represent the relation(s) on the lefthand side of the new join, and insert
1316 : * the JoinExpr into the upper query's jointree at an appropriate place
1317 : * (typically, where the lefthand relation(s) had been). Note that the
1318 : * passed-in SubLink must also be removed from its original position in the
1319 : * query quals, since the quals of the returned JoinExpr replace it.
1320 : * (Notionally, we replace the SubLink with a constant TRUE, then elide the
1321 : * redundant constant from the qual.)
1322 : *
1323 : * On success, the caller is also responsible for recursively applying
1324 : * pull_up_sublinks processing to the rarg and quals of the returned JoinExpr.
1325 : * (On failure, there is no need to do anything, since pull_up_sublinks will
1326 : * be applied when we recursively plan the sub-select.)
1327 : *
1328 : * Side effects of a successful conversion include adding the SubLink's
1329 : * subselect to the query's rangetable, so that it can be referenced in
1330 : * the JoinExpr's rarg.
1331 : */
1332 : JoinExpr *
1333 4448 : convert_ANY_sublink_to_join(PlannerInfo *root, SubLink *sublink,
1334 : Relids available_rels)
1335 : {
1336 : JoinExpr *result;
1337 4448 : Query *parse = root->parse;
1338 4448 : Query *subselect = (Query *) sublink->subselect;
1339 : Relids upper_varnos;
1340 : int rtindex;
1341 : ParseNamespaceItem *nsitem;
1342 : RangeTblEntry *rte;
1343 : RangeTblRef *rtr;
1344 : List *subquery_vars;
1345 : Node *quals;
1346 : ParseState *pstate;
1347 : Relids sub_ref_outer_relids;
1348 : bool use_lateral;
1349 :
1350 : Assert(sublink->subLinkType == ANY_SUBLINK);
1351 :
1352 : /*
1353 : * If the sub-select contains any Vars of the parent query, we treat it as
1354 : * LATERAL. (Vars from higher levels don't matter here.)
1355 : */
1356 4448 : sub_ref_outer_relids = pull_varnos_of_level(NULL, (Node *) subselect, 1);
1357 4448 : use_lateral = !bms_is_empty(sub_ref_outer_relids);
1358 :
1359 : /*
1360 : * Can't convert if the sub-select contains parent-level Vars of relations
1361 : * not in available_rels.
1362 : */
1363 4448 : if (!bms_is_subset(sub_ref_outer_relids, available_rels))
1364 12 : return NULL;
1365 :
1366 : /*
1367 : * The test expression must contain some Vars of the parent query, else
1368 : * it's not gonna be a join. (Note that it won't have Vars referring to
1369 : * the subquery, rather Params.)
1370 : */
1371 4436 : upper_varnos = pull_varnos(root, sublink->testexpr);
1372 4436 : if (bms_is_empty(upper_varnos))
1373 12 : return NULL;
1374 :
1375 : /*
1376 : * However, it can't refer to anything outside available_rels.
1377 : */
1378 4424 : if (!bms_is_subset(upper_varnos, available_rels))
1379 24 : return NULL;
1380 :
1381 : /*
1382 : * The combining operators and left-hand expressions mustn't be volatile.
1383 : */
1384 4400 : if (contain_volatile_functions(sublink->testexpr))
1385 64 : return NULL;
1386 :
1387 : /* Create a dummy ParseState for addRangeTableEntryForSubquery */
1388 4336 : pstate = make_parsestate(NULL);
1389 :
1390 : /*
1391 : * Okay, pull up the sub-select into upper range table.
1392 : *
1393 : * We rely here on the assumption that the outer query has no references
1394 : * to the inner (necessarily true, other than the Vars that we build
1395 : * below). Therefore this is a lot easier than what pull_up_subqueries has
1396 : * to go through.
1397 : */
1398 4336 : nsitem = addRangeTableEntryForSubquery(pstate,
1399 : subselect,
1400 : makeAlias("ANY_subquery", NIL),
1401 : use_lateral,
1402 : false);
1403 4336 : rte = nsitem->p_rte;
1404 4336 : parse->rtable = lappend(parse->rtable, rte);
1405 4336 : rtindex = list_length(parse->rtable);
1406 :
1407 : /*
1408 : * Form a RangeTblRef for the pulled-up sub-select.
1409 : */
1410 4336 : rtr = makeNode(RangeTblRef);
1411 4336 : rtr->rtindex = rtindex;
1412 :
1413 : /*
1414 : * Build a list of Vars representing the subselect outputs.
1415 : */
1416 4336 : subquery_vars = generate_subquery_vars(root,
1417 : subselect->targetList,
1418 : rtindex);
1419 :
1420 : /*
1421 : * Build the new join's qual expression, replacing Params with these Vars.
1422 : */
1423 4336 : quals = convert_testexpr(root, sublink->testexpr, subquery_vars);
1424 :
1425 : /*
1426 : * And finally, build the JoinExpr node.
1427 : */
1428 4336 : result = makeNode(JoinExpr);
1429 4336 : result->jointype = JOIN_SEMI;
1430 4336 : result->isNatural = false;
1431 4336 : result->larg = NULL; /* caller must fill this in */
1432 4336 : result->rarg = (Node *) rtr;
1433 4336 : result->usingClause = NIL;
1434 4336 : result->join_using_alias = NULL;
1435 4336 : result->quals = quals;
1436 4336 : result->alias = NULL;
1437 4336 : result->rtindex = 0; /* we don't need an RTE for it */
1438 :
1439 4336 : return result;
1440 : }
1441 :
1442 : /*
1443 : * convert_EXISTS_sublink_to_join: try to convert an EXISTS SubLink to a join
1444 : *
1445 : * The API of this function is identical to convert_ANY_sublink_to_join's,
1446 : * except that we also support the case where the caller has found NOT EXISTS,
1447 : * so we need an additional input parameter "under_not".
1448 : */
1449 : JoinExpr *
1450 3846 : convert_EXISTS_sublink_to_join(PlannerInfo *root, SubLink *sublink,
1451 : bool under_not, Relids available_rels)
1452 : {
1453 : JoinExpr *result;
1454 3846 : Query *parse = root->parse;
1455 3846 : Query *subselect = (Query *) sublink->subselect;
1456 : Node *whereClause;
1457 : PlannerInfo subroot;
1458 : int rtoffset;
1459 : int varno;
1460 : Relids clause_varnos;
1461 : Relids upper_varnos;
1462 :
1463 : Assert(sublink->subLinkType == EXISTS_SUBLINK);
1464 :
1465 : /*
1466 : * Can't flatten if it contains WITH. (We could arrange to pull up the
1467 : * WITH into the parent query's cteList, but that risks changing the
1468 : * semantics, since a WITH ought to be executed once per associated query
1469 : * call.) Note that convert_ANY_sublink_to_join doesn't have to reject
1470 : * this case, since it just produces a subquery RTE that doesn't have to
1471 : * get flattened into the parent query.
1472 : */
1473 3846 : if (subselect->cteList)
1474 0 : return NULL;
1475 :
1476 : /*
1477 : * Copy the subquery so we can modify it safely (see comments in
1478 : * make_subplan).
1479 : */
1480 3846 : subselect = copyObject(subselect);
1481 :
1482 : /*
1483 : * See if the subquery can be simplified based on the knowledge that it's
1484 : * being used in EXISTS(). If we aren't able to get rid of its
1485 : * targetlist, we have to fail, because the pullup operation leaves us
1486 : * with noplace to evaluate the targetlist.
1487 : */
1488 3846 : if (!simplify_EXISTS_query(root, subselect))
1489 32 : return NULL;
1490 :
1491 : /*
1492 : * Separate out the WHERE clause. (We could theoretically also remove
1493 : * top-level plain JOIN/ON clauses, but it's probably not worth the
1494 : * trouble.)
1495 : */
1496 3814 : whereClause = subselect->jointree->quals;
1497 3814 : subselect->jointree->quals = NULL;
1498 :
1499 : /*
1500 : * The rest of the sub-select must not refer to any Vars of the parent
1501 : * query. (Vars of higher levels should be okay, though.)
1502 : */
1503 3814 : if (contain_vars_of_level((Node *) subselect, 1))
1504 0 : return NULL;
1505 :
1506 : /*
1507 : * On the other hand, the WHERE clause must contain some Vars of the
1508 : * parent query, else it's not gonna be a join.
1509 : */
1510 3814 : if (!contain_vars_of_level(whereClause, 1))
1511 98 : return NULL;
1512 :
1513 : /*
1514 : * We don't risk optimizing if the WHERE clause is volatile, either.
1515 : */
1516 3716 : if (contain_volatile_functions(whereClause))
1517 0 : return NULL;
1518 :
1519 : /*
1520 : * Scan the rangetable for relation RTEs and retrieve the necessary
1521 : * catalog information for each relation. Using this information, clear
1522 : * the inh flag for any relation that has no children, collect not-null
1523 : * attribute numbers for any relation that has column not-null
1524 : * constraints, and expand virtual generated columns for any relation that
1525 : * contains them.
1526 : *
1527 : * Note: we construct up an entirely dummy PlannerInfo for use here. This
1528 : * is fine because only the "glob" and "parse" links will be used in this
1529 : * case.
1530 : *
1531 : * Note: we temporarily assign back the WHERE clause so that any virtual
1532 : * generated column references within it can be expanded. It should be
1533 : * separated out again afterward.
1534 : */
1535 330724 : MemSet(&subroot, 0, sizeof(subroot));
1536 3716 : subroot.type = T_PlannerInfo;
1537 3716 : subroot.glob = root->glob;
1538 3716 : subroot.parse = subselect;
1539 3716 : subselect->jointree->quals = whereClause;
1540 3716 : subselect = preprocess_relation_rtes(&subroot);
1541 :
1542 : /*
1543 : * Now separate out the WHERE clause again.
1544 : */
1545 3716 : whereClause = subselect->jointree->quals;
1546 3716 : subselect->jointree->quals = NULL;
1547 :
1548 : /*
1549 : * The subquery must have a nonempty jointree, but we can make it so.
1550 : */
1551 3716 : replace_empty_jointree(subselect);
1552 :
1553 : /*
1554 : * Prepare to pull up the sub-select into top range table.
1555 : *
1556 : * We rely here on the assumption that the outer query has no references
1557 : * to the inner (necessarily true). Therefore this is a lot easier than
1558 : * what pull_up_subqueries has to go through.
1559 : *
1560 : * In fact, it's even easier than what convert_ANY_sublink_to_join has to
1561 : * do. The machinations of simplify_EXISTS_query ensured that there is
1562 : * nothing interesting in the subquery except an rtable and jointree, and
1563 : * even the jointree FromExpr no longer has quals. So we can just append
1564 : * the rtable to our own and use the FromExpr in our jointree. But first,
1565 : * adjust all level-zero varnos in the subquery to account for the rtable
1566 : * merger.
1567 : */
1568 3716 : rtoffset = list_length(parse->rtable);
1569 3716 : OffsetVarNodes((Node *) subselect, rtoffset, 0);
1570 3716 : OffsetVarNodes(whereClause, rtoffset, 0);
1571 :
1572 : /*
1573 : * Upper-level vars in subquery will now be one level closer to their
1574 : * parent than before; in particular, anything that had been level 1
1575 : * becomes level zero.
1576 : */
1577 3716 : IncrementVarSublevelsUp((Node *) subselect, -1, 1);
1578 3716 : IncrementVarSublevelsUp(whereClause, -1, 1);
1579 :
1580 : /*
1581 : * Now that the WHERE clause is adjusted to match the parent query
1582 : * environment, we can easily identify all the level-zero rels it uses.
1583 : * The ones <= rtoffset belong to the upper query; the ones > rtoffset do
1584 : * not.
1585 : */
1586 3716 : clause_varnos = pull_varnos(root, whereClause);
1587 3716 : upper_varnos = NULL;
1588 3716 : varno = -1;
1589 11184 : while ((varno = bms_next_member(clause_varnos, varno)) >= 0)
1590 : {
1591 7468 : if (varno <= rtoffset)
1592 3740 : upper_varnos = bms_add_member(upper_varnos, varno);
1593 : }
1594 3716 : bms_free(clause_varnos);
1595 : Assert(!bms_is_empty(upper_varnos));
1596 :
1597 : /*
1598 : * Now that we've got the set of upper-level varnos, we can make the last
1599 : * check: only available_rels can be referenced.
1600 : */
1601 3716 : if (!bms_is_subset(upper_varnos, available_rels))
1602 44 : return NULL;
1603 :
1604 : /*
1605 : * Now we can attach the modified subquery rtable to the parent. This also
1606 : * adds subquery's RTEPermissionInfos into the upper query.
1607 : */
1608 3672 : CombineRangeTables(&parse->rtable, &parse->rteperminfos,
1609 : subselect->rtable, subselect->rteperminfos);
1610 :
1611 : /*
1612 : * And finally, build the JoinExpr node.
1613 : */
1614 3672 : result = makeNode(JoinExpr);
1615 3672 : result->jointype = under_not ? JOIN_ANTI : JOIN_SEMI;
1616 3672 : result->isNatural = false;
1617 3672 : result->larg = NULL; /* caller must fill this in */
1618 : /* flatten out the FromExpr node if it's useless */
1619 3672 : if (list_length(subselect->jointree->fromlist) == 1)
1620 3654 : result->rarg = (Node *) linitial(subselect->jointree->fromlist);
1621 : else
1622 18 : result->rarg = (Node *) subselect->jointree;
1623 3672 : result->usingClause = NIL;
1624 3672 : result->join_using_alias = NULL;
1625 3672 : result->quals = whereClause;
1626 3672 : result->alias = NULL;
1627 3672 : result->rtindex = 0; /* we don't need an RTE for it */
1628 :
1629 3672 : return result;
1630 : }
1631 :
1632 : /*
1633 : * simplify_EXISTS_query: remove any useless stuff in an EXISTS's subquery
1634 : *
1635 : * The only thing that matters about an EXISTS query is whether it returns
1636 : * zero or more than zero rows. Therefore, we can remove certain SQL features
1637 : * that won't affect that. The only part that is really likely to matter in
1638 : * typical usage is simplifying the targetlist: it's a common habit to write
1639 : * "SELECT * FROM" even though there is no need to evaluate any columns.
1640 : *
1641 : * Note: by suppressing the targetlist we could cause an observable behavioral
1642 : * change, namely that any errors that might occur in evaluating the tlist
1643 : * won't occur, nor will other side-effects of volatile functions. This seems
1644 : * unlikely to bother anyone in practice.
1645 : *
1646 : * Returns true if was able to discard the targetlist, else false.
1647 : */
1648 : static bool
1649 8850 : simplify_EXISTS_query(PlannerInfo *root, Query *query)
1650 : {
1651 : ListCell *lc;
1652 :
1653 : /*
1654 : * We don't try to simplify at all if the query uses set operations,
1655 : * aggregates, grouping sets, SRFs, modifying CTEs, HAVING, OFFSET, or FOR
1656 : * UPDATE/SHARE; none of these seem likely in normal usage and their
1657 : * possible effects are complex. (Note: we could ignore an "OFFSET 0"
1658 : * clause, but that traditionally is used as an optimization fence, so we
1659 : * don't.)
1660 : */
1661 8850 : if (query->commandType != CMD_SELECT ||
1662 8850 : query->setOperations ||
1663 8850 : query->hasAggs ||
1664 8850 : query->groupingSets ||
1665 8850 : query->hasWindowFuncs ||
1666 8850 : query->hasTargetSRFs ||
1667 8850 : query->hasModifyingCTE ||
1668 8850 : query->havingQual ||
1669 8850 : query->limitOffset ||
1670 8826 : query->rowMarks)
1671 52 : return false;
1672 :
1673 : /*
1674 : * LIMIT with a constant positive (or NULL) value doesn't affect the
1675 : * semantics of EXISTS, so let's ignore such clauses. This is worth doing
1676 : * because people accustomed to certain other DBMSes may be in the habit
1677 : * of writing EXISTS(SELECT ... LIMIT 1) as an optimization. If there's a
1678 : * LIMIT with anything else as argument, though, we can't simplify.
1679 : */
1680 8798 : if (query->limitCount)
1681 : {
1682 : /*
1683 : * The LIMIT clause has not yet been through eval_const_expressions,
1684 : * so we have to apply that here. It might seem like this is a waste
1685 : * of cycles, since the only case plausibly worth worrying about is
1686 : * "LIMIT 1" ... but what we'll actually see is "LIMIT int8(1::int4)",
1687 : * so we have to fold constants or we're not going to recognize it.
1688 : */
1689 24 : Node *node = eval_const_expressions(root, query->limitCount);
1690 : Const *limit;
1691 :
1692 : /* Might as well update the query if we simplified the clause. */
1693 24 : query->limitCount = node;
1694 :
1695 24 : if (!IsA(node, Const))
1696 0 : return false;
1697 :
1698 24 : limit = (Const *) node;
1699 : Assert(limit->consttype == INT8OID);
1700 24 : if (!limit->constisnull && DatumGetInt64(limit->constvalue) <= 0)
1701 12 : return false;
1702 :
1703 : /* Whether or not the targetlist is safe, we can drop the LIMIT. */
1704 12 : query->limitCount = NULL;
1705 : }
1706 :
1707 : /*
1708 : * Otherwise, we can throw away the targetlist, as well as any GROUP,
1709 : * WINDOW, DISTINCT, and ORDER BY clauses; none of those clauses will
1710 : * change a nonzero-rows result to zero rows or vice versa. (Furthermore,
1711 : * since our parsetree representation of these clauses depends on the
1712 : * targetlist, we'd better throw them away if we drop the targetlist.)
1713 : */
1714 8786 : query->targetList = NIL;
1715 8786 : query->groupClause = NIL;
1716 8786 : query->windowClause = NIL;
1717 8786 : query->distinctClause = NIL;
1718 8786 : query->sortClause = NIL;
1719 8786 : query->hasDistinctOn = false;
1720 :
1721 : /*
1722 : * Since we have thrown away the GROUP BY clauses, we'd better remove the
1723 : * RTE_GROUP RTE and clear the hasGroupRTE flag.
1724 : */
1725 17950 : foreach(lc, query->rtable)
1726 : {
1727 9170 : RangeTblEntry *rte = lfirst_node(RangeTblEntry, lc);
1728 :
1729 : /*
1730 : * Remove the RTE_GROUP RTE and clear the hasGroupRTE flag. (Since
1731 : * we'll exit the foreach loop immediately, we don't bother with
1732 : * foreach_delete_current.)
1733 : */
1734 9170 : if (rte->rtekind == RTE_GROUP)
1735 : {
1736 : Assert(query->hasGroupRTE);
1737 6 : query->rtable = list_delete_cell(query->rtable, lc);
1738 6 : query->hasGroupRTE = false;
1739 6 : break;
1740 : }
1741 : }
1742 :
1743 8786 : return true;
1744 : }
1745 :
1746 : /*
1747 : * convert_EXISTS_to_ANY: try to convert EXISTS to a hashable ANY sublink
1748 : *
1749 : * The subselect is expected to be a fresh copy that we can munge up,
1750 : * and to have been successfully passed through simplify_EXISTS_query.
1751 : *
1752 : * On success, the modified subselect is returned, and we store a suitable
1753 : * upper-level test expression at *testexpr, plus a list of the subselect's
1754 : * output Params at *paramIds. (The test expression is already Param-ified
1755 : * and hence need not go through convert_testexpr, which is why we have to
1756 : * deal with the Param IDs specially.)
1757 : *
1758 : * On failure, returns NULL.
1759 : */
1760 : static Query *
1761 2336 : convert_EXISTS_to_ANY(PlannerInfo *root, Query *subselect,
1762 : Node **testexpr, List **paramIds)
1763 : {
1764 : Node *whereClause;
1765 : PlannerInfo subroot;
1766 : List *leftargs,
1767 : *rightargs,
1768 : *opids,
1769 : *opcollations,
1770 : *newWhere,
1771 : *tlist,
1772 : *testlist,
1773 : *paramids;
1774 : ListCell *lc,
1775 : *rc,
1776 : *oc,
1777 : *cc;
1778 : AttrNumber resno;
1779 :
1780 : /*
1781 : * Query must not require a targetlist, since we have to insert a new one.
1782 : * Caller should have dealt with the case already.
1783 : */
1784 : Assert(subselect->targetList == NIL);
1785 :
1786 : /*
1787 : * Separate out the WHERE clause. (We could theoretically also remove
1788 : * top-level plain JOIN/ON clauses, but it's probably not worth the
1789 : * trouble.)
1790 : */
1791 2336 : whereClause = subselect->jointree->quals;
1792 2336 : subselect->jointree->quals = NULL;
1793 :
1794 : /*
1795 : * The rest of the sub-select must not refer to any Vars of the parent
1796 : * query. (Vars of higher levels should be okay, though.)
1797 : *
1798 : * Note: we need not check for Aggrefs separately because we know the
1799 : * sub-select is as yet unoptimized; any uplevel Aggref must therefore
1800 : * contain an uplevel Var reference. This is not the case below ...
1801 : */
1802 2336 : if (contain_vars_of_level((Node *) subselect, 1))
1803 6 : return NULL;
1804 :
1805 : /*
1806 : * We don't risk optimizing if the WHERE clause is volatile, either.
1807 : */
1808 2330 : if (contain_volatile_functions(whereClause))
1809 0 : return NULL;
1810 :
1811 : /*
1812 : * Clean up the WHERE clause by doing const-simplification etc on it.
1813 : * Aside from simplifying the processing we're about to do, this is
1814 : * important for being able to pull chunks of the WHERE clause up into the
1815 : * parent query. Since we are invoked partway through the parent's
1816 : * preprocess_expression() work, earlier steps of preprocess_expression()
1817 : * wouldn't get applied to the pulled-up stuff unless we do them here. For
1818 : * the parts of the WHERE clause that get put back into the child query,
1819 : * this work is partially duplicative, but it shouldn't hurt.
1820 : *
1821 : * Note: we do not run flatten_join_alias_vars. This is OK because any
1822 : * parent aliases were flattened already, and we're not going to pull any
1823 : * child Vars (of any description) into the parent.
1824 : *
1825 : * Note: we construct up an entirely dummy PlannerInfo to pass to
1826 : * eval_const_expressions. This is fine because only the "glob" and
1827 : * "parse" links are used by eval_const_expressions.
1828 : */
1829 207370 : MemSet(&subroot, 0, sizeof(subroot));
1830 2330 : subroot.type = T_PlannerInfo;
1831 2330 : subroot.glob = root->glob;
1832 2330 : subroot.parse = subselect;
1833 2330 : whereClause = eval_const_expressions(&subroot, whereClause);
1834 2330 : whereClause = (Node *) canonicalize_qual((Expr *) whereClause, false);
1835 2330 : whereClause = (Node *) make_ands_implicit((Expr *) whereClause);
1836 :
1837 : /*
1838 : * We now have a flattened implicit-AND list of clauses, which we try to
1839 : * break apart into "outervar = innervar" hash clauses. Anything that
1840 : * can't be broken apart just goes back into the newWhere list. Note that
1841 : * we aren't trying hard yet to ensure that we have only outer or only
1842 : * inner on each side; we'll check that if we get to the end.
1843 : */
1844 2330 : leftargs = rightargs = opids = opcollations = newWhere = NIL;
1845 8964 : foreach(lc, (List *) whereClause)
1846 : {
1847 6634 : OpExpr *expr = (OpExpr *) lfirst(lc);
1848 :
1849 10854 : if (IsA(expr, OpExpr) &&
1850 4220 : hash_ok_operator(expr))
1851 : {
1852 3228 : Node *leftarg = (Node *) linitial(expr->args);
1853 3228 : Node *rightarg = (Node *) lsecond(expr->args);
1854 :
1855 3228 : if (contain_vars_of_level(leftarg, 1))
1856 : {
1857 502 : leftargs = lappend(leftargs, leftarg);
1858 502 : rightargs = lappend(rightargs, rightarg);
1859 502 : opids = lappend_oid(opids, expr->opno);
1860 502 : opcollations = lappend_oid(opcollations, expr->inputcollid);
1861 502 : continue;
1862 : }
1863 2726 : if (contain_vars_of_level(rightarg, 1))
1864 : {
1865 : /*
1866 : * We must commute the clause to put the outer var on the
1867 : * left, because the hashing code in nodeSubplan.c expects
1868 : * that. This probably shouldn't ever fail, since hashable
1869 : * operators ought to have commutators, but be paranoid.
1870 : */
1871 2162 : expr->opno = get_commutator(expr->opno);
1872 2162 : if (OidIsValid(expr->opno) && hash_ok_operator(expr))
1873 : {
1874 2162 : leftargs = lappend(leftargs, rightarg);
1875 2162 : rightargs = lappend(rightargs, leftarg);
1876 2162 : opids = lappend_oid(opids, expr->opno);
1877 2162 : opcollations = lappend_oid(opcollations, expr->inputcollid);
1878 2162 : continue;
1879 : }
1880 : /* If no commutator, no chance to optimize the WHERE clause */
1881 0 : return NULL;
1882 : }
1883 : }
1884 : /* Couldn't handle it as a hash clause */
1885 3970 : newWhere = lappend(newWhere, expr);
1886 : }
1887 :
1888 : /*
1889 : * If we didn't find anything we could convert, fail.
1890 : */
1891 2330 : if (leftargs == NIL)
1892 466 : return NULL;
1893 :
1894 : /*
1895 : * There mustn't be any parent Vars or Aggs in the stuff that we intend to
1896 : * put back into the child query. Note: you might think we don't need to
1897 : * check for Aggs separately, because an uplevel Agg must contain an
1898 : * uplevel Var in its argument. But it is possible that the uplevel Var
1899 : * got optimized away by eval_const_expressions. Consider
1900 : *
1901 : * SUM(CASE WHEN false THEN uplevelvar ELSE 0 END)
1902 : */
1903 3638 : if (contain_vars_of_level((Node *) newWhere, 1) ||
1904 1774 : contain_vars_of_level((Node *) rightargs, 1))
1905 90 : return NULL;
1906 1816 : if (root->parse->hasAggs &&
1907 84 : (contain_aggs_of_level((Node *) newWhere, 1) ||
1908 42 : contain_aggs_of_level((Node *) rightargs, 1)))
1909 0 : return NULL;
1910 :
1911 : /*
1912 : * And there can't be any child Vars in the stuff we intend to pull up.
1913 : * (Note: we'd need to check for child Aggs too, except we know the child
1914 : * has no aggs at all because of simplify_EXISTS_query's check. The same
1915 : * goes for window functions.)
1916 : */
1917 1774 : if (contain_vars_of_level((Node *) leftargs, 0))
1918 0 : return NULL;
1919 :
1920 : /*
1921 : * Also reject sublinks in the stuff we intend to pull up. (It might be
1922 : * possible to support this, but doesn't seem worth the complication.)
1923 : */
1924 1774 : if (contain_subplans((Node *) leftargs))
1925 0 : return NULL;
1926 :
1927 : /*
1928 : * Okay, adjust the sublevelsup in the stuff we're pulling up.
1929 : */
1930 1774 : IncrementVarSublevelsUp((Node *) leftargs, -1, 1);
1931 :
1932 : /*
1933 : * Put back any child-level-only WHERE clauses.
1934 : */
1935 1774 : if (newWhere)
1936 1570 : subselect->jointree->quals = (Node *) make_ands_explicit(newWhere);
1937 :
1938 : /*
1939 : * Build a new targetlist for the child that emits the expressions we
1940 : * need. Concurrently, build a testexpr for the parent using Params to
1941 : * reference the child outputs. (Since we generate Params directly here,
1942 : * there will be no need to convert the testexpr in build_subplan.)
1943 : */
1944 1774 : tlist = testlist = paramids = NIL;
1945 1774 : resno = 1;
1946 4348 : forfour(lc, leftargs, rc, rightargs, oc, opids, cc, opcollations)
1947 : {
1948 2574 : Node *leftarg = (Node *) lfirst(lc);
1949 2574 : Node *rightarg = (Node *) lfirst(rc);
1950 2574 : Oid opid = lfirst_oid(oc);
1951 2574 : Oid opcollation = lfirst_oid(cc);
1952 : Param *param;
1953 :
1954 2574 : param = generate_new_exec_param(root,
1955 : exprType(rightarg),
1956 : exprTypmod(rightarg),
1957 : exprCollation(rightarg));
1958 2574 : tlist = lappend(tlist,
1959 2574 : makeTargetEntry((Expr *) rightarg,
1960 2574 : resno++,
1961 : NULL,
1962 : false));
1963 2574 : testlist = lappend(testlist,
1964 2574 : make_opclause(opid, BOOLOID, false,
1965 : (Expr *) leftarg, (Expr *) param,
1966 : InvalidOid, opcollation));
1967 2574 : paramids = lappend_int(paramids, param->paramid);
1968 : }
1969 :
1970 : /* Put everything where it should go, and we're done */
1971 1774 : subselect->targetList = tlist;
1972 1774 : *testexpr = (Node *) make_ands_explicit(testlist);
1973 1774 : *paramIds = paramids;
1974 :
1975 1774 : return subselect;
1976 : }
1977 :
1978 :
1979 : /*
1980 : * Replace correlation vars (uplevel vars) with Params.
1981 : *
1982 : * Uplevel PlaceHolderVars, aggregates, GROUPING() expressions,
1983 : * MergeSupportFuncs, and ReturningExprs are replaced, too.
1984 : *
1985 : * Note: it is critical that this runs immediately after SS_process_sublinks.
1986 : * Since we do not recurse into the arguments of uplevel PHVs and aggregates,
1987 : * they will get copied to the appropriate subplan args list in the parent
1988 : * query with uplevel vars not replaced by Params, but only adjusted in level
1989 : * (see replace_outer_placeholdervar and replace_outer_agg). That's exactly
1990 : * what we want for the vars of the parent level --- but if a PHV's or
1991 : * aggregate's argument contains any further-up variables, they have to be
1992 : * replaced with Params in their turn. That will happen when the parent level
1993 : * runs SS_replace_correlation_vars. Therefore it must do so after expanding
1994 : * its sublinks to subplans. And we don't want any steps in between, else
1995 : * those steps would never get applied to the argument expressions, either in
1996 : * the parent or the child level.
1997 : *
1998 : * Another fairly tricky thing going on here is the handling of SubLinks in
1999 : * the arguments of uplevel PHVs/aggregates. Those are not touched inside the
2000 : * intermediate query level, either. Instead, SS_process_sublinks recurses on
2001 : * them after copying the PHV or Aggref expression into the parent plan level
2002 : * (this is actually taken care of in build_subplan).
2003 : */
2004 : Node *
2005 161262 : SS_replace_correlation_vars(PlannerInfo *root, Node *expr)
2006 : {
2007 : /* No setup needed for tree walk, so away we go */
2008 161262 : return replace_correlation_vars_mutator(expr, root);
2009 : }
2010 :
2011 : static Node *
2012 1402048 : replace_correlation_vars_mutator(Node *node, PlannerInfo *root)
2013 : {
2014 1402048 : if (node == NULL)
2015 63792 : return NULL;
2016 1338256 : if (IsA(node, Var))
2017 : {
2018 363364 : if (((Var *) node)->varlevelsup > 0)
2019 55182 : return (Node *) replace_outer_var(root, (Var *) node);
2020 : }
2021 1283074 : if (IsA(node, PlaceHolderVar))
2022 : {
2023 102 : if (((PlaceHolderVar *) node)->phlevelsup > 0)
2024 60 : return (Node *) replace_outer_placeholdervar(root,
2025 : (PlaceHolderVar *) node);
2026 : }
2027 1283014 : if (IsA(node, Aggref))
2028 : {
2029 9072 : if (((Aggref *) node)->agglevelsup > 0)
2030 52 : return (Node *) replace_outer_agg(root, (Aggref *) node);
2031 : }
2032 1282962 : if (IsA(node, GroupingFunc))
2033 : {
2034 90 : if (((GroupingFunc *) node)->agglevelsup > 0)
2035 64 : return (Node *) replace_outer_grouping(root, (GroupingFunc *) node);
2036 : }
2037 1282898 : if (IsA(node, MergeSupportFunc))
2038 : {
2039 36 : if (root->parse->commandType != CMD_MERGE)
2040 6 : return (Node *) replace_outer_merge_support(root,
2041 : (MergeSupportFunc *) node);
2042 : }
2043 1282892 : if (IsA(node, ReturningExpr))
2044 : {
2045 18 : if (((ReturningExpr *) node)->retlevelsup > 0)
2046 18 : return (Node *) replace_outer_returning(root,
2047 : (ReturningExpr *) node);
2048 : }
2049 1282874 : return expression_tree_mutator(node, replace_correlation_vars_mutator, root);
2050 : }
2051 :
2052 : /*
2053 : * Expand SubLinks to SubPlans in the given expression.
2054 : *
2055 : * The isQual argument tells whether or not this expression is a WHERE/HAVING
2056 : * qualifier expression. If it is, any sublinks appearing at top level need
2057 : * not distinguish FALSE from UNKNOWN return values.
2058 : */
2059 : Node *
2060 109546 : SS_process_sublinks(PlannerInfo *root, Node *expr, bool isQual)
2061 : {
2062 : process_sublinks_context context;
2063 :
2064 109546 : context.root = root;
2065 109546 : context.isTopQual = isQual;
2066 109546 : return process_sublinks_mutator(expr, &context);
2067 : }
2068 :
2069 : static Node *
2070 1457742 : process_sublinks_mutator(Node *node, process_sublinks_context *context)
2071 : {
2072 : process_sublinks_context locContext;
2073 :
2074 1457742 : locContext.root = context->root;
2075 :
2076 1457742 : if (node == NULL)
2077 62870 : return NULL;
2078 1394872 : if (IsA(node, SubLink))
2079 : {
2080 41270 : SubLink *sublink = (SubLink *) node;
2081 : Node *testexpr;
2082 :
2083 : /*
2084 : * First, recursively process the lefthand-side expressions, if any.
2085 : * They're not top-level anymore.
2086 : */
2087 41270 : locContext.isTopQual = false;
2088 41270 : testexpr = process_sublinks_mutator(sublink->testexpr, &locContext);
2089 :
2090 : /*
2091 : * Now build the SubPlan node and make the expr to return.
2092 : */
2093 41270 : return make_subplan(context->root,
2094 41270 : (Query *) sublink->subselect,
2095 : sublink->subLinkType,
2096 : sublink->subLinkId,
2097 : testexpr,
2098 41270 : context->isTopQual);
2099 : }
2100 :
2101 : /*
2102 : * Don't recurse into the arguments of an outer PHV, Aggref, GroupingFunc,
2103 : * or ReturningExpr here. Any SubLinks in the arguments have to be dealt
2104 : * with at the outer query level; they'll be handled when build_subplan
2105 : * collects the PHV, Aggref, GroupingFunc, or ReturningExpr into the
2106 : * arguments to be passed down to the current subplan.
2107 : */
2108 1353602 : if (IsA(node, PlaceHolderVar))
2109 : {
2110 234 : if (((PlaceHolderVar *) node)->phlevelsup > 0)
2111 12 : return node;
2112 : }
2113 1353368 : else if (IsA(node, Aggref))
2114 : {
2115 586 : if (((Aggref *) node)->agglevelsup > 0)
2116 18 : return node;
2117 : }
2118 1352782 : else if (IsA(node, GroupingFunc))
2119 : {
2120 160 : if (((GroupingFunc *) node)->agglevelsup > 0)
2121 36 : return node;
2122 : }
2123 1352622 : else if (IsA(node, ReturningExpr))
2124 : {
2125 198 : if (((ReturningExpr *) node)->retlevelsup > 0)
2126 6 : return node;
2127 : }
2128 :
2129 : /*
2130 : * We should never see a SubPlan expression in the input (since this is
2131 : * the very routine that creates 'em to begin with). We shouldn't find
2132 : * ourselves invoked directly on a Query, either.
2133 : */
2134 : Assert(!IsA(node, SubPlan));
2135 : Assert(!IsA(node, AlternativeSubPlan));
2136 : Assert(!IsA(node, Query));
2137 :
2138 : /*
2139 : * Because make_subplan() could return an AND or OR clause, we have to
2140 : * take steps to preserve AND/OR flatness of a qual. We assume the input
2141 : * has been AND/OR flattened and so we need no recursion here.
2142 : *
2143 : * (Due to the coding here, we will not get called on the List subnodes of
2144 : * an AND; and the input is *not* yet in implicit-AND format. So no check
2145 : * is needed for a bare List.)
2146 : *
2147 : * Anywhere within the top-level AND/OR clause structure, we can tell
2148 : * make_subplan() that NULL and FALSE are interchangeable. So isTopQual
2149 : * propagates down in both cases. (Note that this is unlike the meaning
2150 : * of "top level qual" used in most other places in Postgres.)
2151 : */
2152 1353530 : if (is_andclause(node))
2153 : {
2154 21508 : List *newargs = NIL;
2155 : ListCell *l;
2156 :
2157 : /* Still at qual top-level */
2158 21508 : locContext.isTopQual = context->isTopQual;
2159 :
2160 78356 : foreach(l, ((BoolExpr *) node)->args)
2161 : {
2162 : Node *newarg;
2163 :
2164 56848 : newarg = process_sublinks_mutator(lfirst(l), &locContext);
2165 56848 : if (is_andclause(newarg))
2166 0 : newargs = list_concat(newargs, ((BoolExpr *) newarg)->args);
2167 : else
2168 56848 : newargs = lappend(newargs, newarg);
2169 : }
2170 21508 : return (Node *) make_andclause(newargs);
2171 : }
2172 :
2173 1332022 : if (is_orclause(node))
2174 : {
2175 2552 : List *newargs = NIL;
2176 : ListCell *l;
2177 :
2178 : /* Still at qual top-level */
2179 2552 : locContext.isTopQual = context->isTopQual;
2180 :
2181 9136 : foreach(l, ((BoolExpr *) node)->args)
2182 : {
2183 : Node *newarg;
2184 :
2185 6584 : newarg = process_sublinks_mutator(lfirst(l), &locContext);
2186 6584 : if (is_orclause(newarg))
2187 0 : newargs = list_concat(newargs, ((BoolExpr *) newarg)->args);
2188 : else
2189 6584 : newargs = lappend(newargs, newarg);
2190 : }
2191 2552 : return (Node *) make_orclause(newargs);
2192 : }
2193 :
2194 : /*
2195 : * If we recurse down through anything other than an AND or OR node, we
2196 : * are definitely not at top qual level anymore.
2197 : */
2198 1329470 : locContext.isTopQual = false;
2199 :
2200 1329470 : return expression_tree_mutator(node,
2201 : process_sublinks_mutator,
2202 : &locContext);
2203 : }
2204 :
2205 : /*
2206 : * SS_identify_outer_params - identify the Params available from outer levels
2207 : *
2208 : * This must be run after SS_replace_correlation_vars and SS_process_sublinks
2209 : * processing is complete in a given query level as well as all of its
2210 : * descendant levels (which means it's most practical to do it at the end of
2211 : * processing the query level). We compute the set of paramIds that outer
2212 : * levels will make available to this level+descendants, and record it in
2213 : * root->outer_params for use while computing extParam/allParam sets in final
2214 : * plan cleanup. (We can't just compute it then, because the upper levels'
2215 : * plan_params lists are transient and will be gone by then.)
2216 : */
2217 : void
2218 524374 : SS_identify_outer_params(PlannerInfo *root)
2219 : {
2220 : Bitmapset *outer_params;
2221 : PlannerInfo *proot;
2222 : ListCell *l;
2223 :
2224 : /*
2225 : * If no parameters have been assigned anywhere in the tree, we certainly
2226 : * don't need to do anything here.
2227 : */
2228 524374 : if (root->glob->paramExecTypes == NIL)
2229 352570 : return;
2230 :
2231 : /*
2232 : * Scan all query levels above this one to see which parameters are due to
2233 : * be available from them, either because lower query levels have
2234 : * requested them (via plan_params) or because they will be available from
2235 : * initPlans of those levels.
2236 : */
2237 171804 : outer_params = NULL;
2238 230250 : for (proot = root->parent_root; proot != NULL; proot = proot->parent_root)
2239 : {
2240 : /*
2241 : * Include ordinary Var/PHV/Aggref/GroupingFunc/ReturningExpr params.
2242 : */
2243 106262 : foreach(l, proot->plan_params)
2244 : {
2245 47816 : PlannerParamItem *pitem = (PlannerParamItem *) lfirst(l);
2246 :
2247 47816 : outer_params = bms_add_member(outer_params, pitem->paramId);
2248 : }
2249 : /* Include any outputs of outer-level initPlans */
2250 64788 : foreach(l, proot->init_plans)
2251 : {
2252 6342 : SubPlan *initsubplan = (SubPlan *) lfirst(l);
2253 : ListCell *l2;
2254 :
2255 12684 : foreach(l2, initsubplan->setParam)
2256 : {
2257 6342 : outer_params = bms_add_member(outer_params, lfirst_int(l2));
2258 : }
2259 : }
2260 : /* Include worktable ID, if a recursive query is being planned */
2261 58446 : if (proot->wt_param_id >= 0)
2262 3278 : outer_params = bms_add_member(outer_params, proot->wt_param_id);
2263 : }
2264 171804 : root->outer_params = outer_params;
2265 : }
2266 :
2267 : /*
2268 : * SS_charge_for_initplans - account for initplans in Path costs & parallelism
2269 : *
2270 : * If any initPlans have been created in the current query level, they will
2271 : * get attached to the Plan tree created from whichever Path we select from
2272 : * the given rel. Increment all that rel's Paths' costs to account for them,
2273 : * and if any of the initPlans are parallel-unsafe, mark all the rel's Paths
2274 : * parallel-unsafe as well.
2275 : *
2276 : * This is separate from SS_attach_initplans because we might conditionally
2277 : * create more initPlans during create_plan(), depending on which Path we
2278 : * select. However, Paths that would generate such initPlans are expected
2279 : * to have included their cost and parallel-safety effects already.
2280 : */
2281 : void
2282 524374 : SS_charge_for_initplans(PlannerInfo *root, RelOptInfo *final_rel)
2283 : {
2284 : Cost initplan_cost;
2285 : bool unsafe_initplans;
2286 : ListCell *lc;
2287 :
2288 : /* Nothing to do if no initPlans */
2289 524374 : if (root->init_plans == NIL)
2290 512218 : return;
2291 :
2292 : /*
2293 : * Compute the cost increment just once, since it will be the same for all
2294 : * Paths. Also check for parallel-unsafe initPlans.
2295 : */
2296 12156 : SS_compute_initplan_cost(root->init_plans,
2297 : &initplan_cost, &unsafe_initplans);
2298 :
2299 : /*
2300 : * Now adjust the costs and parallel_safe flags.
2301 : */
2302 24482 : foreach(lc, final_rel->pathlist)
2303 : {
2304 12326 : Path *path = (Path *) lfirst(lc);
2305 :
2306 12326 : path->startup_cost += initplan_cost;
2307 12326 : path->total_cost += initplan_cost;
2308 12326 : if (unsafe_initplans)
2309 6738 : path->parallel_safe = false;
2310 : }
2311 :
2312 : /*
2313 : * Adjust partial paths' costs too, or forget them entirely if we must
2314 : * consider the rel parallel-unsafe.
2315 : */
2316 12156 : if (unsafe_initplans)
2317 : {
2318 6662 : final_rel->partial_pathlist = NIL;
2319 6662 : final_rel->consider_parallel = false;
2320 : }
2321 : else
2322 : {
2323 5506 : foreach(lc, final_rel->partial_pathlist)
2324 : {
2325 12 : Path *path = (Path *) lfirst(lc);
2326 :
2327 12 : path->startup_cost += initplan_cost;
2328 12 : path->total_cost += initplan_cost;
2329 : }
2330 : }
2331 :
2332 : /* We needn't do set_cheapest() here, caller will do it */
2333 : }
2334 :
2335 : /*
2336 : * SS_compute_initplan_cost - count up the cost delta for some initplans
2337 : *
2338 : * The total cost returned in *initplan_cost_p should be added to both the
2339 : * startup and total costs of the plan node the initplans get attached to.
2340 : * We also report whether any of the initplans are not parallel-safe.
2341 : *
2342 : * The primary user of this is SS_charge_for_initplans, but it's also
2343 : * used in adjusting costs when we move initplans to another plan node.
2344 : */
2345 : void
2346 12394 : SS_compute_initplan_cost(List *init_plans,
2347 : Cost *initplan_cost_p,
2348 : bool *unsafe_initplans_p)
2349 : {
2350 : Cost initplan_cost;
2351 : bool unsafe_initplans;
2352 : ListCell *lc;
2353 :
2354 : /*
2355 : * We assume each initPlan gets run once during top plan startup. This is
2356 : * a conservative overestimate, since in fact an initPlan might be
2357 : * executed later than plan startup, or even not at all.
2358 : */
2359 12394 : initplan_cost = 0;
2360 12394 : unsafe_initplans = false;
2361 25812 : foreach(lc, init_plans)
2362 : {
2363 13418 : SubPlan *initsubplan = lfirst_node(SubPlan, lc);
2364 :
2365 13418 : initplan_cost += initsubplan->startup_cost + initsubplan->per_call_cost;
2366 13418 : if (!initsubplan->parallel_safe)
2367 7638 : unsafe_initplans = true;
2368 : }
2369 12394 : *initplan_cost_p = initplan_cost;
2370 12394 : *unsafe_initplans_p = unsafe_initplans;
2371 12394 : }
2372 :
2373 : /*
2374 : * SS_attach_initplans - attach initplans to topmost plan node
2375 : *
2376 : * Attach any initplans created in the current query level to the specified
2377 : * plan node, which should normally be the topmost node for the query level.
2378 : * (In principle the initPlans could go in any node at or above where they're
2379 : * referenced; but there seems no reason to put them any lower than the
2380 : * topmost node, so we don't bother to track exactly where they came from.)
2381 : *
2382 : * We do not touch the plan node's cost or parallel_safe flag. The initplans
2383 : * must have been accounted for in SS_charge_for_initplans, or by any later
2384 : * code that adds initplans via SS_make_initplan_from_plan.
2385 : */
2386 : void
2387 523104 : SS_attach_initplans(PlannerInfo *root, Plan *plan)
2388 : {
2389 523104 : plan->initPlan = root->init_plans;
2390 523104 : }
2391 :
2392 : /*
2393 : * SS_finalize_plan - do final parameter processing for a completed Plan.
2394 : *
2395 : * This recursively computes the extParam and allParam sets for every Plan
2396 : * node in the given plan tree. (Oh, and RangeTblFunction.funcparams too.)
2397 : *
2398 : * We assume that SS_finalize_plan has already been run on any initplans or
2399 : * subplans the plan tree could reference.
2400 : */
2401 : void
2402 199064 : SS_finalize_plan(PlannerInfo *root, Plan *plan)
2403 : {
2404 : /* No setup needed, just recurse through plan tree. */
2405 199064 : (void) finalize_plan(root, plan, -1, root->outer_params, NULL);
2406 199064 : }
2407 :
2408 : /*
2409 : * Recursive processing of all nodes in the plan tree
2410 : *
2411 : * gather_param is the rescan_param of an ancestral Gather/GatherMerge,
2412 : * or -1 if there is none.
2413 : *
2414 : * valid_params is the set of param IDs supplied by outer plan levels
2415 : * that are valid to reference in this plan node or its children.
2416 : *
2417 : * scan_params is a set of param IDs to force scan plan nodes to reference.
2418 : * This is for EvalPlanQual support, and is always NULL at the top of the
2419 : * recursion.
2420 : *
2421 : * The return value is the computed allParam set for the given Plan node.
2422 : * This is just an internal notational convenience: we can add a child
2423 : * plan's allParams to the set of param IDs of interest to this level
2424 : * in the same statement that recurses to that child.
2425 : *
2426 : * Do not scribble on caller's values of valid_params or scan_params!
2427 : *
2428 : * Note: although we attempt to deal with initPlans anywhere in the tree, the
2429 : * logic is not really right. The problem is that a plan node might return an
2430 : * output Param of its initPlan as a targetlist item, in which case it's valid
2431 : * for the parent plan level to reference that same Param; the parent's usage
2432 : * will be converted into a Var referencing the child plan node by setrefs.c.
2433 : * But this function would see the parent's reference as out of scope and
2434 : * complain about it. For now, this does not matter because the planner only
2435 : * attaches initPlans to the topmost plan node in a query level, so the case
2436 : * doesn't arise. If we ever merge this processing into setrefs.c, maybe it
2437 : * can be handled more cleanly.
2438 : */
2439 : static Bitmapset *
2440 1502224 : finalize_plan(PlannerInfo *root, Plan *plan,
2441 : int gather_param,
2442 : Bitmapset *valid_params,
2443 : Bitmapset *scan_params)
2444 : {
2445 : finalize_primnode_context context;
2446 : int locally_added_param;
2447 : Bitmapset *nestloop_params;
2448 : Bitmapset *initExtParam;
2449 : Bitmapset *initSetParam;
2450 : Bitmapset *child_params;
2451 : ListCell *l;
2452 :
2453 1502224 : if (plan == NULL)
2454 873584 : return NULL;
2455 :
2456 628640 : context.root = root;
2457 628640 : context.paramids = NULL; /* initialize set to empty */
2458 628640 : locally_added_param = -1; /* there isn't one */
2459 628640 : nestloop_params = NULL; /* there aren't any */
2460 :
2461 : /*
2462 : * Examine any initPlans to determine the set of external params they
2463 : * reference and the set of output params they supply. (We assume
2464 : * SS_finalize_plan was run on them already.)
2465 : */
2466 628640 : initExtParam = initSetParam = NULL;
2467 642398 : foreach(l, plan->initPlan)
2468 : {
2469 13758 : SubPlan *initsubplan = (SubPlan *) lfirst(l);
2470 13758 : Plan *initplan = planner_subplan_get_plan(root, initsubplan);
2471 : ListCell *l2;
2472 :
2473 13758 : initExtParam = bms_add_members(initExtParam, initplan->extParam);
2474 27564 : foreach(l2, initsubplan->setParam)
2475 : {
2476 13806 : initSetParam = bms_add_member(initSetParam, lfirst_int(l2));
2477 : }
2478 : }
2479 :
2480 : /* Any setParams are validly referenceable in this node and children */
2481 628640 : if (initSetParam)
2482 12508 : valid_params = bms_union(valid_params, initSetParam);
2483 :
2484 : /*
2485 : * When we call finalize_primnode, context.paramids sets are automatically
2486 : * merged together. But when recursing to self, we have to do it the hard
2487 : * way. We want the paramids set to include params in subplans as well as
2488 : * at this level.
2489 : */
2490 :
2491 : /* Find params in targetlist and qual */
2492 628640 : finalize_primnode((Node *) plan->targetlist, &context);
2493 628640 : finalize_primnode((Node *) plan->qual, &context);
2494 :
2495 : /*
2496 : * If it's a parallel-aware scan node, mark it as dependent on the parent
2497 : * Gather/GatherMerge's rescan Param.
2498 : */
2499 628640 : if (plan->parallel_aware)
2500 : {
2501 2540 : if (gather_param < 0)
2502 0 : elog(ERROR, "parallel-aware plan node is not below a Gather");
2503 2540 : context.paramids =
2504 2540 : bms_add_member(context.paramids, gather_param);
2505 : }
2506 :
2507 : /* Check additional node-type-specific fields */
2508 628640 : switch (nodeTag(plan))
2509 : {
2510 71386 : case T_Result:
2511 71386 : finalize_primnode(((Result *) plan)->resconstantqual,
2512 : &context);
2513 71386 : break;
2514 :
2515 94932 : case T_SeqScan:
2516 94932 : context.paramids = bms_add_members(context.paramids, scan_params);
2517 94932 : break;
2518 :
2519 104 : case T_SampleScan:
2520 104 : finalize_primnode((Node *) ((SampleScan *) plan)->tablesample,
2521 : &context);
2522 104 : context.paramids = bms_add_members(context.paramids, scan_params);
2523 104 : break;
2524 :
2525 99816 : case T_IndexScan:
2526 99816 : finalize_primnode((Node *) ((IndexScan *) plan)->indexqual,
2527 : &context);
2528 99816 : finalize_primnode((Node *) ((IndexScan *) plan)->indexorderby,
2529 : &context);
2530 :
2531 : /*
2532 : * we need not look at indexqualorig, since it will have the same
2533 : * param references as indexqual. Likewise, we can ignore
2534 : * indexorderbyorig.
2535 : */
2536 99816 : context.paramids = bms_add_members(context.paramids, scan_params);
2537 99816 : break;
2538 :
2539 7568 : case T_IndexOnlyScan:
2540 7568 : finalize_primnode((Node *) ((IndexOnlyScan *) plan)->indexqual,
2541 : &context);
2542 7568 : finalize_primnode((Node *) ((IndexOnlyScan *) plan)->recheckqual,
2543 : &context);
2544 7568 : finalize_primnode((Node *) ((IndexOnlyScan *) plan)->indexorderby,
2545 : &context);
2546 :
2547 : /*
2548 : * we need not look at indextlist, since it cannot contain Params.
2549 : */
2550 7568 : context.paramids = bms_add_members(context.paramids, scan_params);
2551 7568 : break;
2552 :
2553 8340 : case T_BitmapIndexScan:
2554 8340 : finalize_primnode((Node *) ((BitmapIndexScan *) plan)->indexqual,
2555 : &context);
2556 :
2557 : /*
2558 : * we need not look at indexqualorig, since it will have the same
2559 : * param references as indexqual.
2560 : */
2561 8340 : break;
2562 :
2563 7974 : case T_BitmapHeapScan:
2564 7974 : finalize_primnode((Node *) ((BitmapHeapScan *) plan)->bitmapqualorig,
2565 : &context);
2566 7974 : context.paramids = bms_add_members(context.paramids, scan_params);
2567 7974 : break;
2568 :
2569 604 : case T_TidScan:
2570 604 : finalize_primnode((Node *) ((TidScan *) plan)->tidquals,
2571 : &context);
2572 604 : context.paramids = bms_add_members(context.paramids, scan_params);
2573 604 : break;
2574 :
2575 34 : case T_TidRangeScan:
2576 34 : finalize_primnode((Node *) ((TidRangeScan *) plan)->tidrangequals,
2577 : &context);
2578 34 : context.paramids = bms_add_members(context.paramids, scan_params);
2579 34 : break;
2580 :
2581 19190 : case T_SubqueryScan:
2582 : {
2583 19190 : SubqueryScan *sscan = (SubqueryScan *) plan;
2584 : RelOptInfo *rel;
2585 : Bitmapset *subquery_params;
2586 :
2587 : /* We must run finalize_plan on the subquery */
2588 19190 : rel = find_base_rel(root, sscan->scan.scanrelid);
2589 19190 : subquery_params = rel->subroot->outer_params;
2590 19190 : if (gather_param >= 0)
2591 24 : subquery_params = bms_add_member(bms_copy(subquery_params),
2592 : gather_param);
2593 19190 : finalize_plan(rel->subroot, sscan->subplan, gather_param,
2594 : subquery_params, NULL);
2595 :
2596 : /* Now we can add its extParams to the parent's params */
2597 38380 : context.paramids = bms_add_members(context.paramids,
2598 19190 : sscan->subplan->extParam);
2599 : /* We need scan_params too, though */
2600 19190 : context.paramids = bms_add_members(context.paramids,
2601 : scan_params);
2602 : }
2603 19190 : break;
2604 :
2605 24400 : case T_FunctionScan:
2606 : {
2607 24400 : FunctionScan *fscan = (FunctionScan *) plan;
2608 : ListCell *lc;
2609 :
2610 : /*
2611 : * Call finalize_primnode independently on each function
2612 : * expression, so that we can record which params are
2613 : * referenced in each, in order to decide which need
2614 : * re-evaluating during rescan.
2615 : */
2616 49004 : foreach(lc, fscan->functions)
2617 : {
2618 24604 : RangeTblFunction *rtfunc = (RangeTblFunction *) lfirst(lc);
2619 : finalize_primnode_context funccontext;
2620 :
2621 24604 : funccontext = context;
2622 24604 : funccontext.paramids = NULL;
2623 :
2624 24604 : finalize_primnode(rtfunc->funcexpr, &funccontext);
2625 :
2626 : /* remember results for execution */
2627 24604 : rtfunc->funcparams = funccontext.paramids;
2628 :
2629 : /* add the function's params to the overall set */
2630 24604 : context.paramids = bms_add_members(context.paramids,
2631 24604 : funccontext.paramids);
2632 : }
2633 :
2634 24400 : context.paramids = bms_add_members(context.paramids,
2635 : scan_params);
2636 : }
2637 24400 : break;
2638 :
2639 234 : case T_TableFuncScan:
2640 234 : finalize_primnode((Node *) ((TableFuncScan *) plan)->tablefunc,
2641 : &context);
2642 234 : context.paramids = bms_add_members(context.paramids, scan_params);
2643 234 : break;
2644 :
2645 5744 : case T_ValuesScan:
2646 5744 : finalize_primnode((Node *) ((ValuesScan *) plan)->values_lists,
2647 : &context);
2648 5744 : context.paramids = bms_add_members(context.paramids, scan_params);
2649 5744 : break;
2650 :
2651 4146 : case T_CteScan:
2652 : {
2653 : /*
2654 : * You might think we should add the node's cteParam to
2655 : * paramids, but we shouldn't because that param is just a
2656 : * linkage mechanism for multiple CteScan nodes for the same
2657 : * CTE; it is never used for changed-param signaling. What we
2658 : * have to do instead is to find the referenced CTE plan and
2659 : * incorporate its external paramids, so that the correct
2660 : * things will happen if the CTE references outer-level
2661 : * variables. See test cases for bug #4902. (We assume
2662 : * SS_finalize_plan was run on the CTE plan already.)
2663 : */
2664 4146 : int plan_id = ((CteScan *) plan)->ctePlanId;
2665 : Plan *cteplan;
2666 :
2667 : /* so, do this ... */
2668 4146 : if (plan_id < 1 || plan_id > list_length(root->glob->subplans))
2669 0 : elog(ERROR, "could not find plan for CteScan referencing plan ID %d",
2670 : plan_id);
2671 4146 : cteplan = (Plan *) list_nth(root->glob->subplans, plan_id - 1);
2672 4146 : context.paramids =
2673 4146 : bms_add_members(context.paramids, cteplan->extParam);
2674 :
2675 : #ifdef NOT_USED
2676 : /* ... but not this */
2677 : context.paramids =
2678 : bms_add_member(context.paramids,
2679 : ((CteScan *) plan)->cteParam);
2680 : #endif
2681 :
2682 4146 : context.paramids = bms_add_members(context.paramids,
2683 : scan_params);
2684 : }
2685 4146 : break;
2686 :
2687 1012 : case T_WorkTableScan:
2688 1012 : context.paramids =
2689 1012 : bms_add_member(context.paramids,
2690 : ((WorkTableScan *) plan)->wtParam);
2691 1012 : context.paramids = bms_add_members(context.paramids, scan_params);
2692 1012 : break;
2693 :
2694 394 : case T_NamedTuplestoreScan:
2695 394 : context.paramids = bms_add_members(context.paramids, scan_params);
2696 394 : break;
2697 :
2698 804 : case T_ForeignScan:
2699 : {
2700 804 : ForeignScan *fscan = (ForeignScan *) plan;
2701 :
2702 804 : finalize_primnode((Node *) fscan->fdw_exprs,
2703 : &context);
2704 804 : finalize_primnode((Node *) fscan->fdw_recheck_quals,
2705 : &context);
2706 :
2707 : /* We assume fdw_scan_tlist cannot contain Params */
2708 804 : context.paramids = bms_add_members(context.paramids,
2709 : scan_params);
2710 : }
2711 804 : break;
2712 :
2713 0 : case T_CustomScan:
2714 : {
2715 0 : CustomScan *cscan = (CustomScan *) plan;
2716 : ListCell *lc;
2717 :
2718 0 : finalize_primnode((Node *) cscan->custom_exprs,
2719 : &context);
2720 : /* We assume custom_scan_tlist cannot contain Params */
2721 0 : context.paramids =
2722 0 : bms_add_members(context.paramids, scan_params);
2723 :
2724 : /* child nodes if any */
2725 0 : foreach(lc, cscan->custom_plans)
2726 : {
2727 0 : context.paramids =
2728 0 : bms_add_members(context.paramids,
2729 0 : finalize_plan(root,
2730 0 : (Plan *) lfirst(lc),
2731 : gather_param,
2732 : valid_params,
2733 : scan_params));
2734 : }
2735 : }
2736 0 : break;
2737 :
2738 86416 : case T_ModifyTable:
2739 : {
2740 86416 : ModifyTable *mtplan = (ModifyTable *) plan;
2741 :
2742 : /* Force descendant scan nodes to reference epqParam */
2743 86416 : locally_added_param = mtplan->epqParam;
2744 86416 : valid_params = bms_add_member(bms_copy(valid_params),
2745 : locally_added_param);
2746 86416 : scan_params = bms_add_member(bms_copy(scan_params),
2747 : locally_added_param);
2748 86416 : finalize_primnode((Node *) mtplan->returningLists,
2749 : &context);
2750 86416 : finalize_primnode((Node *) mtplan->onConflictSet,
2751 : &context);
2752 86416 : finalize_primnode((Node *) mtplan->onConflictWhere,
2753 : &context);
2754 : /* exclRelTlist contains only Vars, doesn't need examination */
2755 : }
2756 86416 : break;
2757 :
2758 10386 : case T_Append:
2759 : {
2760 35984 : foreach(l, ((Append *) plan)->appendplans)
2761 : {
2762 25598 : context.paramids =
2763 25598 : bms_add_members(context.paramids,
2764 25598 : finalize_plan(root,
2765 25598 : (Plan *) lfirst(l),
2766 : gather_param,
2767 : valid_params,
2768 : scan_params));
2769 : }
2770 : }
2771 10386 : break;
2772 :
2773 114 : case T_MergeAppend:
2774 : {
2775 474 : foreach(l, ((MergeAppend *) plan)->mergeplans)
2776 : {
2777 360 : context.paramids =
2778 360 : bms_add_members(context.paramids,
2779 360 : finalize_plan(root,
2780 360 : (Plan *) lfirst(l),
2781 : gather_param,
2782 : valid_params,
2783 : scan_params));
2784 : }
2785 : }
2786 114 : break;
2787 :
2788 154 : case T_BitmapAnd:
2789 : {
2790 462 : foreach(l, ((BitmapAnd *) plan)->bitmapplans)
2791 : {
2792 308 : context.paramids =
2793 308 : bms_add_members(context.paramids,
2794 308 : finalize_plan(root,
2795 308 : (Plan *) lfirst(l),
2796 : gather_param,
2797 : valid_params,
2798 : scan_params));
2799 : }
2800 : }
2801 154 : break;
2802 :
2803 212 : case T_BitmapOr:
2804 : {
2805 636 : foreach(l, ((BitmapOr *) plan)->bitmapplans)
2806 : {
2807 424 : context.paramids =
2808 424 : bms_add_members(context.paramids,
2809 424 : finalize_plan(root,
2810 424 : (Plan *) lfirst(l),
2811 : gather_param,
2812 : valid_params,
2813 : scan_params));
2814 : }
2815 : }
2816 212 : break;
2817 :
2818 68464 : case T_NestLoop:
2819 : {
2820 68464 : finalize_primnode((Node *) ((Join *) plan)->joinqual,
2821 : &context);
2822 : /* collect set of params that will be passed to right child */
2823 121032 : foreach(l, ((NestLoop *) plan)->nestParams)
2824 : {
2825 52568 : NestLoopParam *nlp = (NestLoopParam *) lfirst(l);
2826 :
2827 52568 : nestloop_params = bms_add_member(nestloop_params,
2828 : nlp->paramno);
2829 : }
2830 : }
2831 68464 : break;
2832 :
2833 4482 : case T_MergeJoin:
2834 4482 : finalize_primnode((Node *) ((Join *) plan)->joinqual,
2835 : &context);
2836 4482 : finalize_primnode((Node *) ((MergeJoin *) plan)->mergeclauses,
2837 : &context);
2838 4482 : break;
2839 :
2840 20046 : case T_HashJoin:
2841 20046 : finalize_primnode((Node *) ((Join *) plan)->joinqual,
2842 : &context);
2843 20046 : finalize_primnode((Node *) ((HashJoin *) plan)->hashclauses,
2844 : &context);
2845 20046 : break;
2846 :
2847 20046 : case T_Hash:
2848 20046 : finalize_primnode((Node *) ((Hash *) plan)->hashkeys,
2849 : &context);
2850 20046 : break;
2851 :
2852 2278 : case T_Limit:
2853 2278 : finalize_primnode(((Limit *) plan)->limitOffset,
2854 : &context);
2855 2278 : finalize_primnode(((Limit *) plan)->limitCount,
2856 : &context);
2857 2278 : break;
2858 :
2859 1012 : case T_RecursiveUnion:
2860 : /* child nodes are allowed to reference wtParam */
2861 1012 : locally_added_param = ((RecursiveUnion *) plan)->wtParam;
2862 1012 : valid_params = bms_add_member(bms_copy(valid_params),
2863 : locally_added_param);
2864 : /* wtParam does *not* get added to scan_params */
2865 1012 : break;
2866 :
2867 7734 : case T_LockRows:
2868 : /* Force descendant scan nodes to reference epqParam */
2869 7734 : locally_added_param = ((LockRows *) plan)->epqParam;
2870 7734 : valid_params = bms_add_member(bms_copy(valid_params),
2871 : locally_added_param);
2872 7734 : scan_params = bms_add_member(bms_copy(scan_params),
2873 : locally_added_param);
2874 7734 : break;
2875 :
2876 10704 : case T_Agg:
2877 : {
2878 10704 : Agg *agg = (Agg *) plan;
2879 :
2880 : /*
2881 : * AGG_HASHED plans need to know which Params are referenced
2882 : * in aggregate calls. Do a separate scan to identify them.
2883 : */
2884 10704 : if (agg->aggstrategy == AGG_HASHED)
2885 : {
2886 : finalize_primnode_context aggcontext;
2887 :
2888 1362 : aggcontext.root = root;
2889 1362 : aggcontext.paramids = NULL;
2890 1362 : finalize_agg_primnode((Node *) agg->plan.targetlist,
2891 : &aggcontext);
2892 1362 : finalize_agg_primnode((Node *) agg->plan.qual,
2893 : &aggcontext);
2894 1362 : agg->aggParams = aggcontext.paramids;
2895 : }
2896 : }
2897 10704 : break;
2898 :
2899 122 : case T_WindowAgg:
2900 122 : finalize_primnode(((WindowAgg *) plan)->startOffset,
2901 : &context);
2902 122 : finalize_primnode(((WindowAgg *) plan)->endOffset,
2903 : &context);
2904 122 : break;
2905 :
2906 956 : case T_Gather:
2907 : /* child nodes are allowed to reference rescan_param, if any */
2908 956 : locally_added_param = ((Gather *) plan)->rescan_param;
2909 956 : if (locally_added_param >= 0)
2910 : {
2911 950 : valid_params = bms_add_member(bms_copy(valid_params),
2912 : locally_added_param);
2913 :
2914 : /*
2915 : * We currently don't support nested Gathers. The issue so
2916 : * far as this function is concerned would be how to identify
2917 : * which child nodes depend on which Gather.
2918 : */
2919 : Assert(gather_param < 0);
2920 : /* Pass down rescan_param to child parallel-aware nodes */
2921 950 : gather_param = locally_added_param;
2922 : }
2923 : /* rescan_param does *not* get added to scan_params */
2924 956 : break;
2925 :
2926 330 : case T_GatherMerge:
2927 : /* child nodes are allowed to reference rescan_param, if any */
2928 330 : locally_added_param = ((GatherMerge *) plan)->rescan_param;
2929 330 : if (locally_added_param >= 0)
2930 : {
2931 330 : valid_params = bms_add_member(bms_copy(valid_params),
2932 : locally_added_param);
2933 :
2934 : /*
2935 : * We currently don't support nested Gathers. The issue so
2936 : * far as this function is concerned would be how to identify
2937 : * which child nodes depend on which Gather.
2938 : */
2939 : Assert(gather_param < 0);
2940 : /* Pass down rescan_param to child parallel-aware nodes */
2941 330 : gather_param = locally_added_param;
2942 : }
2943 : /* rescan_param does *not* get added to scan_params */
2944 330 : break;
2945 :
2946 1924 : case T_Memoize:
2947 1924 : finalize_primnode((Node *) ((Memoize *) plan)->param_exprs,
2948 : &context);
2949 1924 : break;
2950 :
2951 46578 : case T_ProjectSet:
2952 : case T_Material:
2953 : case T_Sort:
2954 : case T_IncrementalSort:
2955 : case T_Unique:
2956 : case T_SetOp:
2957 : case T_Group:
2958 : /* no node-type-specific fields need fixing */
2959 46578 : break;
2960 :
2961 0 : default:
2962 0 : elog(ERROR, "unrecognized node type: %d",
2963 : (int) nodeTag(plan));
2964 : }
2965 :
2966 : /* Process left and right child plans, if any */
2967 628640 : child_params = finalize_plan(root,
2968 628640 : plan->lefttree,
2969 : gather_param,
2970 : valid_params,
2971 : scan_params);
2972 628640 : context.paramids = bms_add_members(context.paramids, child_params);
2973 :
2974 628640 : if (nestloop_params)
2975 : {
2976 : /* right child can reference nestloop_params as well as valid_params */
2977 47198 : child_params = finalize_plan(root,
2978 47198 : plan->righttree,
2979 : gather_param,
2980 : bms_union(nestloop_params, valid_params),
2981 : scan_params);
2982 : /* ... and they don't count as parameters used at my level */
2983 47198 : child_params = bms_difference(child_params, nestloop_params);
2984 47198 : bms_free(nestloop_params);
2985 : }
2986 : else
2987 : {
2988 : /* easy case */
2989 581442 : child_params = finalize_plan(root,
2990 581442 : plan->righttree,
2991 : gather_param,
2992 : valid_params,
2993 : scan_params);
2994 : }
2995 628640 : context.paramids = bms_add_members(context.paramids, child_params);
2996 :
2997 : /*
2998 : * Any locally generated parameter doesn't count towards its generating
2999 : * plan node's external dependencies. (Note: if we changed valid_params
3000 : * and/or scan_params, we leak those bitmapsets; not worth the notational
3001 : * trouble to clean them up.)
3002 : */
3003 628640 : if (locally_added_param >= 0)
3004 : {
3005 96442 : context.paramids = bms_del_member(context.paramids,
3006 : locally_added_param);
3007 : }
3008 :
3009 : /* Now we have all the paramids referenced in this node and children */
3010 :
3011 628640 : if (!bms_is_subset(context.paramids, valid_params))
3012 0 : elog(ERROR, "plan should not reference subplan's variable");
3013 :
3014 : /*
3015 : * The plan node's allParam and extParam fields should include all its
3016 : * referenced paramids, plus contributions from any child initPlans.
3017 : * However, any setParams of the initPlans should not be present in the
3018 : * parent node's extParams, only in its allParams. (It's possible that
3019 : * some initPlans have extParams that are setParams of other initPlans.)
3020 : */
3021 :
3022 : /* allParam must include initplans' extParams and setParams */
3023 628640 : plan->allParam = bms_union(context.paramids, initExtParam);
3024 628640 : plan->allParam = bms_add_members(plan->allParam, initSetParam);
3025 : /* extParam must include any initplan extParams */
3026 628640 : plan->extParam = bms_union(context.paramids, initExtParam);
3027 : /* but not any initplan setParams */
3028 628640 : plan->extParam = bms_del_members(plan->extParam, initSetParam);
3029 :
3030 628640 : return plan->allParam;
3031 : }
3032 :
3033 : /*
3034 : * finalize_primnode: add IDs of all PARAM_EXEC params that appear (or will
3035 : * appear) in the given expression tree to the result set.
3036 : */
3037 : static bool
3038 10689552 : finalize_primnode(Node *node, finalize_primnode_context *context)
3039 : {
3040 10689552 : if (node == NULL)
3041 1244256 : return false;
3042 9445296 : if (IsA(node, Param))
3043 : {
3044 133208 : if (((Param *) node)->paramkind == PARAM_EXEC)
3045 : {
3046 130972 : int paramid = ((Param *) node)->paramid;
3047 :
3048 130972 : context->paramids = bms_add_member(context->paramids, paramid);
3049 : }
3050 133208 : return false; /* no more to do here */
3051 : }
3052 9312088 : else if (IsA(node, Aggref))
3053 : {
3054 : /*
3055 : * Check to see if the aggregate will be replaced by a Param
3056 : * referencing a subquery output during setrefs.c. If so, we must
3057 : * account for that Param here. (For various reasons, it's not
3058 : * convenient to perform that substitution earlier than setrefs.c, nor
3059 : * to perform this processing after setrefs.c. Thus we need a wart
3060 : * here.)
3061 : */
3062 13910 : Aggref *aggref = (Aggref *) node;
3063 : Param *aggparam;
3064 :
3065 13910 : aggparam = find_minmax_agg_replacement_param(context->root, aggref);
3066 13910 : if (aggparam != NULL)
3067 556 : context->paramids = bms_add_member(context->paramids,
3068 : aggparam->paramid);
3069 : /* Fall through to examine the agg's arguments */
3070 : }
3071 9298178 : else if (IsA(node, SubPlan))
3072 : {
3073 39260 : SubPlan *subplan = (SubPlan *) node;
3074 39260 : Plan *plan = planner_subplan_get_plan(context->root, subplan);
3075 : ListCell *lc;
3076 : Bitmapset *subparamids;
3077 :
3078 : /* Recurse into the testexpr, but not into the Plan */
3079 39260 : finalize_primnode(subplan->testexpr, context);
3080 :
3081 : /*
3082 : * Remove any param IDs of output parameters of the subplan that were
3083 : * referenced in the testexpr. These are not interesting for
3084 : * parameter change signaling since we always re-evaluate the subplan.
3085 : * Note that this wouldn't work too well if there might be uses of the
3086 : * same param IDs elsewhere in the plan, but that can't happen because
3087 : * generate_new_exec_param never tries to merge params.
3088 : */
3089 42552 : foreach(lc, subplan->paramIds)
3090 : {
3091 3292 : context->paramids = bms_del_member(context->paramids,
3092 : lfirst_int(lc));
3093 : }
3094 :
3095 : /* Also examine args list */
3096 39260 : finalize_primnode((Node *) subplan->args, context);
3097 :
3098 : /*
3099 : * Add params needed by the subplan to paramids, but excluding those
3100 : * we will pass down to it. (We assume SS_finalize_plan was run on
3101 : * the subplan already.)
3102 : */
3103 39260 : subparamids = bms_copy(plan->extParam);
3104 93992 : foreach(lc, subplan->parParam)
3105 : {
3106 54732 : subparamids = bms_del_member(subparamids, lfirst_int(lc));
3107 : }
3108 39260 : context->paramids = bms_join(context->paramids, subparamids);
3109 :
3110 39260 : return false; /* no more to do here */
3111 : }
3112 9272828 : return expression_tree_walker(node, finalize_primnode, context);
3113 : }
3114 :
3115 : /*
3116 : * finalize_agg_primnode: find all Aggref nodes in the given expression tree,
3117 : * and add IDs of all PARAM_EXEC params appearing within their aggregated
3118 : * arguments to the result set.
3119 : */
3120 : static bool
3121 10706 : finalize_agg_primnode(Node *node, finalize_primnode_context *context)
3122 : {
3123 10706 : if (node == NULL)
3124 1440 : return false;
3125 9266 : if (IsA(node, Aggref))
3126 : {
3127 1144 : Aggref *agg = (Aggref *) node;
3128 :
3129 : /* we should not consider the direct arguments, if any */
3130 1144 : finalize_primnode((Node *) agg->args, context);
3131 1144 : finalize_primnode((Node *) agg->aggfilter, context);
3132 1144 : return false; /* there can't be any Aggrefs below here */
3133 : }
3134 8122 : return expression_tree_walker(node, finalize_agg_primnode, context);
3135 : }
3136 :
3137 : /*
3138 : * SS_make_initplan_output_param - make a Param for an initPlan's output
3139 : *
3140 : * The plan is expected to return a scalar value of the given type/collation.
3141 : *
3142 : * Note that in some cases the initplan may not ever appear in the finished
3143 : * plan tree. If that happens, we'll have wasted a PARAM_EXEC slot, which
3144 : * is no big deal.
3145 : */
3146 : Param *
3147 446 : SS_make_initplan_output_param(PlannerInfo *root,
3148 : Oid resulttype, int32 resulttypmod,
3149 : Oid resultcollation)
3150 : {
3151 446 : return generate_new_exec_param(root, resulttype,
3152 : resulttypmod, resultcollation);
3153 : }
3154 :
3155 : /*
3156 : * SS_make_initplan_from_plan - given a plan tree, make it an InitPlan
3157 : *
3158 : * We build an EXPR_SUBLINK SubPlan node and put it into the initplan
3159 : * list for the outer query level. A Param that represents the initplan's
3160 : * output has already been assigned using SS_make_initplan_output_param.
3161 : */
3162 : void
3163 400 : SS_make_initplan_from_plan(PlannerInfo *root,
3164 : PlannerInfo *subroot, Plan *plan,
3165 : Param *prm)
3166 : {
3167 : SubPlan *node;
3168 :
3169 : /*
3170 : * Add the subplan and its PlannerInfo, as well as a dummy path entry, to
3171 : * the global lists. Ideally we'd save a real path, but right now our
3172 : * sole caller doesn't build a path that exactly matches the plan. Since
3173 : * we're not currently going to need the path for an initplan, it's not
3174 : * worth requiring construction of such a path.
3175 : */
3176 400 : root->glob->subplans = lappend(root->glob->subplans, plan);
3177 400 : root->glob->subpaths = lappend(root->glob->subpaths, NULL);
3178 400 : root->glob->subroots = lappend(root->glob->subroots, subroot);
3179 :
3180 : /*
3181 : * Create a SubPlan node and add it to the outer list of InitPlans. Note
3182 : * it has to appear after any other InitPlans it might depend on (see
3183 : * comments in ExecReScan).
3184 : */
3185 400 : node = makeNode(SubPlan);
3186 400 : node->subLinkType = EXPR_SUBLINK;
3187 400 : node->plan_id = list_length(root->glob->subplans);
3188 400 : node->plan_name = psprintf("InitPlan %d", node->plan_id);
3189 400 : get_first_col_type(plan, &node->firstColType, &node->firstColTypmod,
3190 : &node->firstColCollation);
3191 400 : node->parallel_safe = plan->parallel_safe;
3192 400 : node->setParam = list_make1_int(prm->paramid);
3193 :
3194 400 : root->init_plans = lappend(root->init_plans, node);
3195 :
3196 : /*
3197 : * The node can't have any inputs (since it's an initplan), so the
3198 : * parParam and args lists remain empty.
3199 : */
3200 :
3201 : /* Set costs of SubPlan using info from the plan tree */
3202 400 : cost_subplan(subroot, node, plan);
3203 400 : }
|