Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * placeholder.c
4 : * PlaceHolderVar and PlaceHolderInfo manipulation routines
5 : *
6 : *
7 : * Portions Copyright (c) 1996-2026, PostgreSQL Global Development Group
8 : * Portions Copyright (c) 1994, Regents of the University of California
9 : *
10 : *
11 : * IDENTIFICATION
12 : * src/backend/optimizer/util/placeholder.c
13 : *
14 : *-------------------------------------------------------------------------
15 : */
16 : #include "postgres.h"
17 :
18 : #include "nodes/nodeFuncs.h"
19 : #include "optimizer/cost.h"
20 : #include "optimizer/optimizer.h"
21 : #include "optimizer/pathnode.h"
22 : #include "optimizer/placeholder.h"
23 : #include "optimizer/planmain.h"
24 : #include "utils/lsyscache.h"
25 :
26 :
27 : typedef struct contain_placeholder_references_context
28 : {
29 : int relid;
30 : int sublevels_up;
31 : } contain_placeholder_references_context;
32 :
33 : /* Local functions */
34 : static void find_placeholders_recurse(PlannerInfo *root, Node *jtnode);
35 : static void find_placeholders_in_expr(PlannerInfo *root, Node *expr);
36 : static bool contain_placeholder_references_walker(Node *node,
37 : contain_placeholder_references_context *context);
38 :
39 :
40 : /*
41 : * make_placeholder_expr
42 : * Make a PlaceHolderVar for the given expression.
43 : *
44 : * phrels is the syntactic location (as a set of relids) to attribute
45 : * to the expression.
46 : *
47 : * The caller is responsible for adjusting phlevelsup and phnullingrels
48 : * as needed. Because we do not know here which query level the PHV
49 : * will be associated with, it's important that this function touches
50 : * only root->glob; messing with other parts of PlannerInfo would be
51 : * likely to do the wrong thing.
52 : */
53 : PlaceHolderVar *
54 1382 : make_placeholder_expr(PlannerInfo *root, Expr *expr, Relids phrels)
55 : {
56 1382 : PlaceHolderVar *phv = makeNode(PlaceHolderVar);
57 :
58 1382 : phv->phexpr = expr;
59 1382 : phv->phrels = phrels;
60 1382 : phv->phnullingrels = NULL; /* caller may change this later */
61 1382 : phv->phid = ++(root->glob->lastPHId);
62 1382 : phv->phlevelsup = 0; /* caller may change this later */
63 :
64 1382 : return phv;
65 : }
66 :
67 : /*
68 : * find_placeholder_info
69 : * Fetch the PlaceHolderInfo for the given PHV
70 : *
71 : * If the PlaceHolderInfo doesn't exist yet, create it if we haven't yet
72 : * frozen the set of PlaceHolderInfos for the query; else throw an error.
73 : *
74 : * This is separate from make_placeholder_expr because subquery pullup has
75 : * to make PlaceHolderVars for expressions that might not be used at all in
76 : * the upper query, or might not remain after const-expression simplification.
77 : * We build PlaceHolderInfos only for PHVs that are still present in the
78 : * simplified query passed to query_planner().
79 : *
80 : * Note: this should only be called after query_planner() has started.
81 : */
82 : PlaceHolderInfo *
83 4990 : find_placeholder_info(PlannerInfo *root, PlaceHolderVar *phv)
84 : {
85 : PlaceHolderInfo *phinfo;
86 : Relids rels_used;
87 :
88 : /* if this ever isn't true, we'd need to be able to look in parent lists */
89 : Assert(phv->phlevelsup == 0);
90 :
91 : /* Use placeholder_array to look up existing PlaceHolderInfo quickly */
92 4990 : if (phv->phid < root->placeholder_array_size)
93 4269 : phinfo = root->placeholder_array[phv->phid];
94 : else
95 721 : phinfo = NULL;
96 4990 : if (phinfo != NULL)
97 : {
98 : Assert(phinfo->phid == phv->phid);
99 3853 : return phinfo;
100 : }
101 :
102 : /* Not found, so create it */
103 1137 : if (root->placeholdersFrozen)
104 0 : elog(ERROR, "too late to create a new PlaceHolderInfo");
105 :
106 1137 : phinfo = makeNode(PlaceHolderInfo);
107 :
108 1137 : phinfo->phid = phv->phid;
109 1137 : phinfo->ph_var = copyObject(phv);
110 :
111 : /*
112 : * By convention, phinfo->ph_var->phnullingrels is always empty, since the
113 : * PlaceHolderInfo represents the initially-calculated state of the
114 : * PlaceHolderVar. PlaceHolderVars appearing in the query tree might have
115 : * varying values of phnullingrels, reflecting outer joins applied above
116 : * the calculation level.
117 : */
118 1137 : phinfo->ph_var->phnullingrels = NULL;
119 :
120 : /*
121 : * Any referenced rels that are outside the PHV's syntactic scope are
122 : * LATERAL references, which should be included in ph_lateral but not in
123 : * ph_eval_at. If no referenced rels are within the syntactic scope,
124 : * force evaluation at the syntactic location.
125 : */
126 1137 : rels_used = pull_varnos(root, (Node *) phv->phexpr);
127 1137 : phinfo->ph_lateral = bms_difference(rels_used, phv->phrels);
128 1137 : phinfo->ph_eval_at = bms_int_members(rels_used, phv->phrels);
129 : /* If no contained vars, force evaluation at syntactic location */
130 1137 : if (bms_is_empty(phinfo->ph_eval_at))
131 : {
132 561 : phinfo->ph_eval_at = bms_copy(phv->phrels);
133 : Assert(!bms_is_empty(phinfo->ph_eval_at));
134 : }
135 1137 : phinfo->ph_needed = NULL; /* initially it's unused */
136 : /* for the moment, estimate width using just the datatype info */
137 1137 : phinfo->ph_width = get_typavgwidth(exprType((Node *) phv->phexpr),
138 1137 : exprTypmod((Node *) phv->phexpr));
139 :
140 : /*
141 : * Add to both placeholder_list and placeholder_array. Note: because we
142 : * store pointers to the PlaceHolderInfos in two data structures, it'd be
143 : * unsafe to pass the whole placeholder_list structure through
144 : * expression_tree_mutator or the like --- or at least, you'd have to
145 : * rebuild the placeholder_array afterwards.
146 : */
147 1137 : root->placeholder_list = lappend(root->placeholder_list, phinfo);
148 :
149 1137 : if (phinfo->phid >= root->placeholder_array_size)
150 : {
151 : /* Must allocate or enlarge placeholder_array */
152 : int new_size;
153 :
154 721 : new_size = root->placeholder_array_size ? root->placeholder_array_size * 2 : 8;
155 721 : while (phinfo->phid >= new_size)
156 0 : new_size *= 2;
157 721 : if (root->placeholder_array)
158 0 : root->placeholder_array =
159 0 : repalloc0_array(root->placeholder_array, PlaceHolderInfo *, root->placeholder_array_size, new_size);
160 : else
161 721 : root->placeholder_array =
162 721 : palloc0_array(PlaceHolderInfo *, new_size);
163 721 : root->placeholder_array_size = new_size;
164 : }
165 1137 : root->placeholder_array[phinfo->phid] = phinfo;
166 :
167 : /*
168 : * The PHV's contained expression may contain other, lower-level PHVs. We
169 : * now know we need to get those into the PlaceHolderInfo list, too, so we
170 : * may as well do that immediately.
171 : */
172 1137 : find_placeholders_in_expr(root, (Node *) phinfo->ph_var->phexpr);
173 :
174 1137 : return phinfo;
175 : }
176 :
177 : /*
178 : * find_placeholders_in_jointree
179 : * Search the jointree for PlaceHolderVars, and build PlaceHolderInfos
180 : *
181 : * We don't need to look at the targetlist because build_base_rel_tlists()
182 : * will already have made entries for any PHVs in the tlist.
183 : */
184 : void
185 172439 : find_placeholders_in_jointree(PlannerInfo *root)
186 : {
187 : /* This must be done before freezing the set of PHIs */
188 : Assert(!root->placeholdersFrozen);
189 :
190 : /* We need do nothing if the query contains no PlaceHolderVars */
191 172439 : if (root->glob->lastPHId != 0)
192 : {
193 : /* Start recursion at top of jointree */
194 : Assert(root->parse->jointree != NULL &&
195 : IsA(root->parse->jointree, FromExpr));
196 961 : find_placeholders_recurse(root, (Node *) root->parse->jointree);
197 : }
198 172439 : }
199 :
200 : /*
201 : * find_placeholders_recurse
202 : * One recursion level of find_placeholders_in_jointree.
203 : *
204 : * jtnode is the current jointree node to examine.
205 : */
206 : static void
207 4557 : find_placeholders_recurse(PlannerInfo *root, Node *jtnode)
208 : {
209 4557 : if (jtnode == NULL)
210 0 : return;
211 4557 : if (IsA(jtnode, RangeTblRef))
212 : {
213 : /* No quals to deal with here */
214 : }
215 2306 : else if (IsA(jtnode, FromExpr))
216 : {
217 1156 : FromExpr *f = (FromExpr *) jtnode;
218 : ListCell *l;
219 :
220 : /*
221 : * First, recurse to handle child joins.
222 : */
223 2452 : foreach(l, f->fromlist)
224 : {
225 1296 : find_placeholders_recurse(root, lfirst(l));
226 : }
227 :
228 : /*
229 : * Now process the top-level quals.
230 : */
231 1156 : find_placeholders_in_expr(root, f->quals);
232 : }
233 1150 : else if (IsA(jtnode, JoinExpr))
234 : {
235 1150 : JoinExpr *j = (JoinExpr *) jtnode;
236 :
237 : /*
238 : * First, recurse to handle child joins.
239 : */
240 1150 : find_placeholders_recurse(root, j->larg);
241 1150 : find_placeholders_recurse(root, j->rarg);
242 :
243 : /* Process the qual clauses */
244 1150 : find_placeholders_in_expr(root, j->quals);
245 : }
246 : else
247 0 : elog(ERROR, "unrecognized node type: %d",
248 : (int) nodeTag(jtnode));
249 : }
250 :
251 : /*
252 : * find_placeholders_in_expr
253 : * Find all PlaceHolderVars in the given expression, and create
254 : * PlaceHolderInfo entries for them.
255 : */
256 : static void
257 3443 : find_placeholders_in_expr(PlannerInfo *root, Node *expr)
258 : {
259 : List *vars;
260 : ListCell *vl;
261 :
262 : /*
263 : * pull_var_clause does more than we need here, but it'll do and it's
264 : * convenient to use.
265 : */
266 3443 : vars = pull_var_clause(expr,
267 : PVC_RECURSE_AGGREGATES |
268 : PVC_RECURSE_WINDOWFUNCS |
269 : PVC_INCLUDE_PLACEHOLDERS);
270 7529 : foreach(vl, vars)
271 : {
272 4086 : PlaceHolderVar *phv = (PlaceHolderVar *) lfirst(vl);
273 :
274 : /* Ignore any plain Vars */
275 4086 : if (!IsA(phv, PlaceHolderVar))
276 3635 : continue;
277 :
278 : /* Create a PlaceHolderInfo entry if there's not one already */
279 451 : (void) find_placeholder_info(root, phv);
280 : }
281 3443 : list_free(vars);
282 3443 : }
283 :
284 : /*
285 : * fix_placeholder_input_needed_levels
286 : * Adjust the "needed at" levels for placeholder inputs
287 : *
288 : * This is called after we've finished determining the eval_at levels for
289 : * all placeholders. We need to make sure that all vars and placeholders
290 : * needed to evaluate each placeholder will be available at the scan or join
291 : * level where the evaluation will be done. (It might seem that scan-level
292 : * evaluations aren't interesting, but that's not so: a LATERAL reference
293 : * within a placeholder's expression needs to cause the referenced var or
294 : * placeholder to be marked as needed in the scan where it's evaluated.)
295 : * Note that this loop can have side-effects on the ph_needed sets of other
296 : * PlaceHolderInfos; that's okay because we don't examine ph_needed here, so
297 : * there are no ordering issues to worry about.
298 : */
299 : void
300 172439 : fix_placeholder_input_needed_levels(PlannerInfo *root)
301 : {
302 : ListCell *lc;
303 :
304 173576 : foreach(lc, root->placeholder_list)
305 : {
306 1137 : PlaceHolderInfo *phinfo = (PlaceHolderInfo *) lfirst(lc);
307 1137 : List *vars = pull_var_clause((Node *) phinfo->ph_var->phexpr,
308 : PVC_RECURSE_AGGREGATES |
309 : PVC_RECURSE_WINDOWFUNCS |
310 : PVC_INCLUDE_PLACEHOLDERS);
311 :
312 1137 : add_vars_to_targetlist(root, vars, phinfo->ph_eval_at);
313 1137 : list_free(vars);
314 : }
315 172439 : }
316 :
317 : /*
318 : * rebuild_placeholder_attr_needed
319 : * Put back attr_needed bits for Vars/PHVs needed in PlaceHolderVars.
320 : *
321 : * This is used to rebuild attr_needed/ph_needed sets after removal of a
322 : * useless outer join. It should match what
323 : * fix_placeholder_input_needed_levels did, except that we call
324 : * add_vars_to_attr_needed not add_vars_to_targetlist.
325 : */
326 : void
327 6263 : rebuild_placeholder_attr_needed(PlannerInfo *root)
328 : {
329 : ListCell *lc;
330 :
331 6374 : foreach(lc, root->placeholder_list)
332 : {
333 111 : PlaceHolderInfo *phinfo = (PlaceHolderInfo *) lfirst(lc);
334 111 : List *vars = pull_var_clause((Node *) phinfo->ph_var->phexpr,
335 : PVC_RECURSE_AGGREGATES |
336 : PVC_RECURSE_WINDOWFUNCS |
337 : PVC_INCLUDE_PLACEHOLDERS);
338 :
339 111 : add_vars_to_attr_needed(root, vars, phinfo->ph_eval_at);
340 111 : list_free(vars);
341 : }
342 6263 : }
343 :
344 : /*
345 : * add_placeholders_to_base_rels
346 : * Add any required PlaceHolderVars to base rels' targetlists.
347 : *
348 : * If any placeholder can be computed at a base rel and is needed above it,
349 : * add it to that rel's targetlist. This might look like it could be merged
350 : * with fix_placeholder_input_needed_levels, but it must be separate because
351 : * join removal happens in between, and can change the ph_eval_at sets. There
352 : * is essentially the same logic in add_placeholders_to_joinrel, but we can't
353 : * do that part until joinrels are formed.
354 : */
355 : void
356 172439 : add_placeholders_to_base_rels(PlannerInfo *root)
357 : {
358 : ListCell *lc;
359 :
360 173573 : foreach(lc, root->placeholder_list)
361 : {
362 1134 : PlaceHolderInfo *phinfo = (PlaceHolderInfo *) lfirst(lc);
363 1134 : Relids eval_at = phinfo->ph_eval_at;
364 : int varno;
365 :
366 2056 : if (bms_get_singleton_member(eval_at, &varno) &&
367 922 : bms_nonempty_difference(phinfo->ph_needed, eval_at))
368 : {
369 883 : RelOptInfo *rel = find_base_rel(root, varno);
370 :
371 : /*
372 : * As in add_vars_to_targetlist(), a value computed at scan level
373 : * has not yet been nulled by any outer join, so its phnullingrels
374 : * should be empty.
375 : */
376 : Assert(phinfo->ph_var->phnullingrels == NULL);
377 :
378 : /* Copying the PHV might be unnecessary here, but be safe */
379 883 : rel->reltarget->exprs = lappend(rel->reltarget->exprs,
380 883 : copyObject(phinfo->ph_var));
381 : /* reltarget's cost and width fields will be updated later */
382 : }
383 : }
384 172439 : }
385 :
386 : /*
387 : * add_placeholders_to_joinrel
388 : * Add any newly-computable PlaceHolderVars to a join rel's targetlist;
389 : * and if computable PHVs contain lateral references, add those
390 : * references to the joinrel's direct_lateral_relids.
391 : *
392 : * A join rel should emit a PlaceHolderVar if (a) the PHV can be computed
393 : * at or below this join level and (b) the PHV is needed above this level.
394 : * Our caller build_join_rel() has already added any PHVs that were computed
395 : * in either join input rel, so we need add only newly-computable ones to
396 : * the targetlist. However, direct_lateral_relids must be updated for every
397 : * PHV computable at or below this join, as explained below.
398 : */
399 : void
400 113477 : add_placeholders_to_joinrel(PlannerInfo *root, RelOptInfo *joinrel,
401 : RelOptInfo *outer_rel, RelOptInfo *inner_rel,
402 : SpecialJoinInfo *sjinfo)
403 : {
404 113477 : Relids relids = joinrel->relids;
405 113477 : int64 tuple_width = joinrel->reltarget->width;
406 : ListCell *lc;
407 :
408 115708 : foreach(lc, root->placeholder_list)
409 : {
410 2231 : PlaceHolderInfo *phinfo = (PlaceHolderInfo *) lfirst(lc);
411 :
412 : /* Is it computable here? */
413 2231 : if (bms_is_subset(phinfo->ph_eval_at, relids))
414 : {
415 : /* Is it still needed above this joinrel? */
416 1543 : if (bms_nonempty_difference(phinfo->ph_needed, relids))
417 : {
418 : /*
419 : * Yes, but only add to tlist if it wasn't computed in either
420 : * input; otherwise it should be there already. Also, we
421 : * charge the cost of evaluating the contained expression if
422 : * the PHV can be computed here but not in either input. This
423 : * is a bit bogus because we make the decision based on the
424 : * first pair of possible input relations considered for the
425 : * joinrel. With other pairs, it might be possible to compute
426 : * the PHV in one input or the other, and then we'd be double
427 : * charging the PHV's cost for some join paths. For now, live
428 : * with that; but we might want to improve it later by
429 : * refiguring the reltarget costs for each pair of inputs.
430 : */
431 1044 : if (!bms_is_subset(phinfo->ph_eval_at, outer_rel->relids) &&
432 779 : !bms_is_subset(phinfo->ph_eval_at, inner_rel->relids))
433 : {
434 : /* Copying might be unnecessary here, but be safe */
435 247 : PlaceHolderVar *phv = copyObject(phinfo->ph_var);
436 : QualCost cost;
437 :
438 : /*
439 : * It'll start out not nulled by anything. Joins above
440 : * this one might add to its phnullingrels later, in much
441 : * the same way as for Vars.
442 : */
443 : Assert(phv->phnullingrels == NULL);
444 :
445 247 : joinrel->reltarget->exprs = lappend(joinrel->reltarget->exprs,
446 : phv);
447 247 : cost_qual_eval_node(&cost, (Node *) phv->phexpr, root);
448 247 : joinrel->reltarget->cost.startup += cost.startup;
449 247 : joinrel->reltarget->cost.per_tuple += cost.per_tuple;
450 247 : tuple_width += phinfo->ph_width;
451 : }
452 : }
453 :
454 : /*
455 : * Also adjust joinrel's direct_lateral_relids to include the
456 : * PHV's source rel(s). We must do this even if we're not
457 : * actually going to emit the PHV, otherwise join_is_legal() will
458 : * reject valid join orderings. (In principle maybe we could
459 : * instead remove the joinrel's lateral_relids dependency; but
460 : * that's complicated to get right, and cases where we're not
461 : * going to emit the PHV are too rare to justify the work.)
462 : *
463 : * In principle we should only do this if the join doesn't yet
464 : * include the PHV's source rel(s). But our caller
465 : * build_join_rel() will clean things up by removing the join's
466 : * own relids from its direct_lateral_relids, so we needn't
467 : * account for that here.
468 : */
469 1543 : joinrel->direct_lateral_relids =
470 1543 : bms_add_members(joinrel->direct_lateral_relids,
471 1543 : phinfo->ph_lateral);
472 : }
473 : }
474 :
475 113477 : joinrel->reltarget->width = clamp_width_est(tuple_width);
476 113477 : }
477 :
478 : /*
479 : * contain_placeholder_references_to
480 : * Detect whether any PlaceHolderVars in the given clause contain
481 : * references to the given relid (typically an OJ relid).
482 : *
483 : * "Contain" means that there's a use of the relid inside the PHV's
484 : * contained expression, so that changing the nullability status of
485 : * the rel might change what the PHV computes.
486 : *
487 : * The code here to cope with upper-level PHVs is likely dead, but keep it
488 : * anyway just in case.
489 : */
490 : bool
491 8932 : contain_placeholder_references_to(PlannerInfo *root, Node *clause,
492 : int relid)
493 : {
494 : contain_placeholder_references_context context;
495 :
496 : /* We can answer quickly in the common case that there's no PHVs at all */
497 8932 : if (root->glob->lastPHId == 0)
498 8616 : return false;
499 : /* Else run the recursive search */
500 316 : context.relid = relid;
501 316 : context.sublevels_up = 0;
502 316 : return contain_placeholder_references_walker(clause, &context);
503 : }
504 :
505 : static bool
506 1075 : contain_placeholder_references_walker(Node *node,
507 : contain_placeholder_references_context *context)
508 : {
509 1075 : if (node == NULL)
510 63 : return false;
511 1012 : if (IsA(node, PlaceHolderVar))
512 : {
513 27 : PlaceHolderVar *phv = (PlaceHolderVar *) node;
514 :
515 : /* We should just look through PHVs of other query levels */
516 27 : if (phv->phlevelsup == context->sublevels_up)
517 : {
518 : /* If phrels matches, we found what we came for */
519 27 : if (bms_is_member(context->relid, phv->phrels))
520 6 : return true;
521 :
522 : /*
523 : * We should not examine phnullingrels: what we are looking for is
524 : * references in the contained expression, not OJs that might null
525 : * the result afterwards. Also, we don't need to recurse into the
526 : * contained expression, because phrels should adequately
527 : * summarize what's in there. So we're done here.
528 : */
529 21 : return false;
530 : }
531 : }
532 985 : else if (IsA(node, Query))
533 : {
534 : /* Recurse into RTE subquery or not-yet-planned sublink subquery */
535 : bool result;
536 :
537 0 : context->sublevels_up++;
538 0 : result = query_tree_walker((Query *) node,
539 : contain_placeholder_references_walker,
540 : context,
541 : 0);
542 0 : context->sublevels_up--;
543 0 : return result;
544 : }
545 985 : return expression_tree_walker(node, contain_placeholder_references_walker,
546 : context);
547 : }
548 :
549 : /*
550 : * Compute the set of outer-join relids that can null a placeholder.
551 : *
552 : * This is analogous to RelOptInfo.nulling_relids for Vars, but we compute it
553 : * on-the-fly rather than saving it somewhere. Currently the value is needed
554 : * at most once per query, so there's little value in doing otherwise. If it
555 : * ever gains more widespread use, perhaps we should cache the result in
556 : * PlaceHolderInfo.
557 : */
558 : Relids
559 126 : get_placeholder_nulling_relids(PlannerInfo *root, PlaceHolderInfo *phinfo)
560 : {
561 126 : Relids result = NULL;
562 126 : int relid = -1;
563 :
564 : /*
565 : * Form the union of all potential nulling OJs for each baserel included
566 : * in ph_eval_at.
567 : */
568 288 : while ((relid = bms_next_member(phinfo->ph_eval_at, relid)) > 0)
569 : {
570 162 : RelOptInfo *rel = root->simple_rel_array[relid];
571 :
572 : /* ignore the RTE_GROUP RTE */
573 162 : if (relid == root->group_rtindex)
574 0 : continue;
575 :
576 162 : if (rel == NULL) /* must be an outer join */
577 : {
578 : Assert(bms_is_member(relid, root->outer_join_rels));
579 6 : continue;
580 : }
581 156 : result = bms_add_members(result, rel->nulling_relids);
582 : }
583 :
584 : /* Now remove any OJs already included in ph_eval_at, and we're done. */
585 126 : result = bms_del_members(result, phinfo->ph_eval_at);
586 126 : return result;
587 : }
|
