Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * planagg.c
4 : * Special planning for aggregate queries.
5 : *
6 : * This module tries to replace MIN/MAX aggregate functions by subqueries
7 : * of the form
8 : * (SELECT col FROM tab
9 : * WHERE col IS NOT NULL AND existing-quals
10 : * ORDER BY col ASC/DESC
11 : * LIMIT 1)
12 : * Given a suitable index on tab.col, this can be much faster than the
13 : * generic scan-all-the-rows aggregation plan. We can handle multiple
14 : * MIN/MAX aggregates by generating multiple subqueries, and their
15 : * orderings can be different. However, if the query contains any
16 : * non-optimizable aggregates, there's no point since we'll have to
17 : * scan all the rows anyway.
18 : *
19 : *
20 : * Portions Copyright (c) 1996-2024, PostgreSQL Global Development Group
21 : * Portions Copyright (c) 1994, Regents of the University of California
22 : *
23 : *
24 : * IDENTIFICATION
25 : * src/backend/optimizer/plan/planagg.c
26 : *
27 : *-------------------------------------------------------------------------
28 : */
29 : #include "postgres.h"
30 :
31 : #include "access/htup_details.h"
32 : #include "catalog/pg_aggregate.h"
33 : #include "catalog/pg_type.h"
34 : #include "nodes/makefuncs.h"
35 : #include "nodes/nodeFuncs.h"
36 : #include "optimizer/cost.h"
37 : #include "optimizer/optimizer.h"
38 : #include "optimizer/pathnode.h"
39 : #include "optimizer/paths.h"
40 : #include "optimizer/planmain.h"
41 : #include "optimizer/subselect.h"
42 : #include "optimizer/tlist.h"
43 : #include "parser/parse_clause.h"
44 : #include "parser/parsetree.h"
45 : #include "rewrite/rewriteManip.h"
46 : #include "utils/lsyscache.h"
47 : #include "utils/syscache.h"
48 :
49 : static bool can_minmax_aggs(PlannerInfo *root, List **context);
50 : static bool build_minmax_path(PlannerInfo *root, MinMaxAggInfo *mminfo,
51 : Oid eqop, Oid sortop, bool nulls_first);
52 : static void minmax_qp_callback(PlannerInfo *root, void *extra);
53 : static Oid fetch_agg_sort_op(Oid aggfnoid);
54 :
55 :
56 : /*
57 : * preprocess_minmax_aggregates - preprocess MIN/MAX aggregates
58 : *
59 : * Check to see whether the query contains MIN/MAX aggregate functions that
60 : * might be optimizable via indexscans. If it does, and all the aggregates
61 : * are potentially optimizable, then create a MinMaxAggPath and add it to
62 : * the (UPPERREL_GROUP_AGG, NULL) upperrel.
63 : *
64 : * This should be called by grouping_planner() just before it's ready to call
65 : * query_planner(), because we generate indexscan paths by cloning the
66 : * planner's state and invoking query_planner() on a modified version of
67 : * the query parsetree. Thus, all preprocessing needed before query_planner()
68 : * must already be done. This relies on the list of aggregates in
69 : * root->agginfos, so preprocess_aggrefs() must have been called already, too.
70 : */
71 : void
72 33648 : preprocess_minmax_aggregates(PlannerInfo *root)
73 : {
74 33648 : Query *parse = root->parse;
75 : FromExpr *jtnode;
76 : RangeTblRef *rtr;
77 : RangeTblEntry *rte;
78 : List *aggs_list;
79 : RelOptInfo *grouped_rel;
80 : ListCell *lc;
81 :
82 : /* minmax_aggs list should be empty at this point */
83 : Assert(root->minmax_aggs == NIL);
84 :
85 : /* Nothing to do if query has no aggregates */
86 33648 : if (!parse->hasAggs)
87 33262 : return;
88 :
89 : Assert(!parse->setOperations); /* shouldn't get here if a setop */
90 : Assert(parse->rowMarks == NIL); /* nor if FOR UPDATE */
91 :
92 : /*
93 : * Reject unoptimizable cases.
94 : *
95 : * We don't handle GROUP BY or windowing, because our current
96 : * implementations of grouping require looking at all the rows anyway, and
97 : * so there's not much point in optimizing MIN/MAX.
98 : */
99 33648 : if (parse->groupClause || list_length(parse->groupingSets) > 1 ||
100 30038 : parse->hasWindowFuncs)
101 3616 : return;
102 :
103 : /*
104 : * Reject if query contains any CTEs; there's no way to build an indexscan
105 : * on one so we couldn't succeed here. (If the CTEs are unreferenced,
106 : * that's not true, but it doesn't seem worth expending cycles to check.)
107 : */
108 30032 : if (parse->cteList)
109 62 : return;
110 :
111 : /*
112 : * We also restrict the query to reference exactly one table, since join
113 : * conditions can't be handled reasonably. (We could perhaps handle a
114 : * query containing cartesian-product joins, but it hardly seems worth the
115 : * trouble.) However, the single table could be buried in several levels
116 : * of FromExpr due to subqueries. Note the "single" table could be an
117 : * inheritance parent, too, including the case of a UNION ALL subquery
118 : * that's been flattened to an appendrel.
119 : */
120 29970 : jtnode = parse->jointree;
121 56762 : while (IsA(jtnode, FromExpr))
122 : {
123 30004 : if (list_length(jtnode->fromlist) != 1)
124 3212 : return;
125 26792 : jtnode = linitial(jtnode->fromlist);
126 : }
127 26758 : if (!IsA(jtnode, RangeTblRef))
128 1106 : return;
129 25652 : rtr = (RangeTblRef *) jtnode;
130 25652 : rte = planner_rt_fetch(rtr->rtindex, root);
131 25652 : if (rte->rtekind == RTE_RELATION)
132 : /* ordinary relation, ok */ ;
133 3164 : else if (rte->rtekind == RTE_SUBQUERY && rte->inh)
134 : /* flattened UNION ALL subquery, ok */ ;
135 : else
136 3110 : return;
137 :
138 : /*
139 : * Examine all the aggregates and verify all are MIN/MAX aggregates. Stop
140 : * as soon as we find one that isn't.
141 : */
142 22542 : aggs_list = NIL;
143 22542 : if (!can_minmax_aggs(root, &aggs_list))
144 21868 : return;
145 :
146 : /*
147 : * OK, there is at least the possibility of performing the optimization.
148 : * Build an access path for each aggregate. If any of the aggregates
149 : * prove to be non-indexable, give up; there is no point in optimizing
150 : * just some of them.
151 : */
152 1096 : foreach(lc, aggs_list)
153 : {
154 710 : MinMaxAggInfo *mminfo = (MinMaxAggInfo *) lfirst(lc);
155 : Oid eqop;
156 : bool reverse;
157 :
158 : /*
159 : * We'll need the equality operator that goes with the aggregate's
160 : * ordering operator.
161 : */
162 710 : eqop = get_equality_op_for_ordering_op(mminfo->aggsortop, &reverse);
163 710 : if (!OidIsValid(eqop)) /* shouldn't happen */
164 0 : elog(ERROR, "could not find equality operator for ordering operator %u",
165 : mminfo->aggsortop);
166 :
167 : /*
168 : * We can use either an ordering that gives NULLS FIRST or one that
169 : * gives NULLS LAST; furthermore there's unlikely to be much
170 : * performance difference between them, so it doesn't seem worth
171 : * costing out both ways if we get a hit on the first one. NULLS
172 : * FIRST is more likely to be available if the operator is a
173 : * reverse-sort operator, so try that first if reverse.
174 : */
175 710 : if (build_minmax_path(root, mminfo, eqop, mminfo->aggsortop, reverse))
176 422 : continue;
177 288 : if (build_minmax_path(root, mminfo, eqop, mminfo->aggsortop, !reverse))
178 0 : continue;
179 :
180 : /* No indexable path for this aggregate, so fail */
181 288 : return;
182 : }
183 :
184 : /*
185 : * OK, we can do the query this way. Prepare to create a MinMaxAggPath
186 : * node.
187 : *
188 : * First, create an output Param node for each agg. (If we end up not
189 : * using the MinMaxAggPath, we'll waste a PARAM_EXEC slot for each agg,
190 : * which is not worth worrying about. We can't wait till create_plan time
191 : * to decide whether to make the Param, unfortunately.)
192 : */
193 808 : foreach(lc, aggs_list)
194 : {
195 422 : MinMaxAggInfo *mminfo = (MinMaxAggInfo *) lfirst(lc);
196 :
197 422 : mminfo->param =
198 422 : SS_make_initplan_output_param(root,
199 422 : exprType((Node *) mminfo->target),
200 : -1,
201 422 : exprCollation((Node *) mminfo->target));
202 : }
203 :
204 : /*
205 : * Create a MinMaxAggPath node with the appropriate estimated costs and
206 : * other needed data, and add it to the UPPERREL_GROUP_AGG upperrel, where
207 : * it will compete against the standard aggregate implementation. (It
208 : * will likely always win, but we need not assume that here.)
209 : *
210 : * Note: grouping_planner won't have created this upperrel yet, but it's
211 : * fine for us to create it first. We will not have inserted the correct
212 : * consider_parallel value in it, but MinMaxAggPath paths are currently
213 : * never parallel-safe anyway, so that doesn't matter. Likewise, it
214 : * doesn't matter that we haven't filled FDW-related fields in the rel.
215 : * Also, because there are no rowmarks, we know that the processed_tlist
216 : * doesn't need to change anymore, so making the pathtarget now is safe.
217 : */
218 386 : grouped_rel = fetch_upper_rel(root, UPPERREL_GROUP_AGG, NULL);
219 386 : add_path(grouped_rel, (Path *)
220 386 : create_minmaxagg_path(root, grouped_rel,
221 : create_pathtarget(root,
222 : root->processed_tlist),
223 : aggs_list,
224 386 : (List *) parse->havingQual));
225 : }
226 :
227 : /*
228 : * can_minmax_aggs
229 : * Examine all the aggregates in the query, and check if they are
230 : * all MIN/MAX aggregates. If so, build a list of MinMaxAggInfo
231 : * nodes for them.
232 : *
233 : * Returns false if a non-MIN/MAX aggregate is found, true otherwise.
234 : */
235 : static bool
236 22542 : can_minmax_aggs(PlannerInfo *root, List **context)
237 : {
238 : ListCell *lc;
239 :
240 : /*
241 : * This function used to have to scan the query for itself, but now we can
242 : * just thumb through the AggInfo list made by preprocess_aggrefs.
243 : */
244 23576 : foreach(lc, root->agginfos)
245 : {
246 22902 : AggInfo *agginfo = lfirst_node(AggInfo, lc);
247 22902 : Aggref *aggref = linitial_node(Aggref, agginfo->aggrefs);
248 : Oid aggsortop;
249 : TargetEntry *curTarget;
250 : MinMaxAggInfo *mminfo;
251 :
252 : Assert(aggref->agglevelsup == 0);
253 22902 : if (list_length(aggref->args) != 1)
254 21868 : return false; /* it couldn't be MIN/MAX */
255 :
256 : /*
257 : * ORDER BY is usually irrelevant for MIN/MAX, but it can change the
258 : * outcome if the aggsortop's operator class recognizes non-identical
259 : * values as equal. For example, 4.0 and 4.00 are equal according to
260 : * numeric_ops, yet distinguishable. If MIN() receives more than one
261 : * value equal to 4.0 and no value less than 4.0, it is unspecified
262 : * which of those equal values MIN() returns. An ORDER BY expression
263 : * that differs for each of those equal values of the argument
264 : * expression makes the result predictable once again. This is a
265 : * niche requirement, and we do not implement it with subquery paths.
266 : * In any case, this test lets us reject ordered-set aggregates
267 : * quickly.
268 : */
269 14236 : if (aggref->aggorder != NIL)
270 452 : return false;
271 : /* note: we do not care if DISTINCT is mentioned ... */
272 :
273 : /*
274 : * We might implement the optimization when a FILTER clause is present
275 : * by adding the filter to the quals of the generated subquery. For
276 : * now, just punt.
277 : */
278 13784 : if (aggref->aggfilter != NULL)
279 442 : return false;
280 :
281 13342 : aggsortop = fetch_agg_sort_op(aggref->aggfnoid);
282 13342 : if (!OidIsValid(aggsortop))
283 12304 : return false; /* not a MIN/MAX aggregate */
284 :
285 1038 : curTarget = (TargetEntry *) linitial(aggref->args);
286 :
287 1038 : if (contain_mutable_functions((Node *) curTarget->expr))
288 4 : return false; /* not potentially indexable */
289 :
290 1034 : if (type_is_rowtype(exprType((Node *) curTarget->expr)))
291 0 : return false; /* IS NOT NULL would have weird semantics */
292 :
293 1034 : mminfo = makeNode(MinMaxAggInfo);
294 1034 : mminfo->aggfnoid = aggref->aggfnoid;
295 1034 : mminfo->aggsortop = aggsortop;
296 1034 : mminfo->target = curTarget->expr;
297 1034 : mminfo->subroot = NULL; /* don't compute path yet */
298 1034 : mminfo->path = NULL;
299 1034 : mminfo->pathcost = 0;
300 1034 : mminfo->param = NULL;
301 :
302 1034 : *context = lappend(*context, mminfo);
303 : }
304 674 : return true;
305 : }
306 :
307 : /*
308 : * build_minmax_path
309 : * Given a MIN/MAX aggregate, try to build an indexscan Path it can be
310 : * optimized with.
311 : *
312 : * If successful, stash the best path in *mminfo and return true.
313 : * Otherwise, return false.
314 : */
315 : static bool
316 998 : build_minmax_path(PlannerInfo *root, MinMaxAggInfo *mminfo,
317 : Oid eqop, Oid sortop, bool nulls_first)
318 : {
319 : PlannerInfo *subroot;
320 : Query *parse;
321 : TargetEntry *tle;
322 : List *tlist;
323 : NullTest *ntest;
324 : SortGroupClause *sortcl;
325 : RelOptInfo *final_rel;
326 : Path *sorted_path;
327 : Cost path_cost;
328 : double path_fraction;
329 :
330 : /*
331 : * We are going to construct what is effectively a sub-SELECT query, so
332 : * clone the current query level's state and adjust it to make it look
333 : * like a subquery. Any outer references will now be one level higher
334 : * than before. (This means that when we are done, there will be no Vars
335 : * of level 1, which is why the subquery can become an initplan.)
336 : */
337 998 : subroot = (PlannerInfo *) palloc(sizeof(PlannerInfo));
338 998 : memcpy(subroot, root, sizeof(PlannerInfo));
339 998 : subroot->query_level++;
340 998 : subroot->parent_root = root;
341 : /* reset subplan-related stuff */
342 998 : subroot->plan_params = NIL;
343 998 : subroot->outer_params = NULL;
344 998 : subroot->init_plans = NIL;
345 998 : subroot->agginfos = NIL;
346 998 : subroot->aggtransinfos = NIL;
347 :
348 998 : subroot->parse = parse = copyObject(root->parse);
349 998 : IncrementVarSublevelsUp((Node *) parse, 1, 1);
350 :
351 : /* append_rel_list might contain outer Vars? */
352 998 : subroot->append_rel_list = copyObject(root->append_rel_list);
353 998 : IncrementVarSublevelsUp((Node *) subroot->append_rel_list, 1, 1);
354 : /* There shouldn't be any OJ info to translate, as yet */
355 : Assert(subroot->join_info_list == NIL);
356 : /* and we haven't made equivalence classes, either */
357 : Assert(subroot->eq_classes == NIL);
358 : /* and we haven't created PlaceHolderInfos, either */
359 : Assert(subroot->placeholder_list == NIL);
360 :
361 : /*----------
362 : * Generate modified query of the form
363 : * (SELECT col FROM tab
364 : * WHERE col IS NOT NULL AND existing-quals
365 : * ORDER BY col ASC/DESC
366 : * LIMIT 1)
367 : *----------
368 : */
369 : /* single tlist entry that is the aggregate target */
370 998 : tle = makeTargetEntry(copyObject(mminfo->target),
371 : (AttrNumber) 1,
372 : pstrdup("agg_target"),
373 : false);
374 998 : tlist = list_make1(tle);
375 998 : subroot->processed_tlist = parse->targetList = tlist;
376 :
377 : /* No HAVING, no DISTINCT, no aggregates anymore */
378 998 : parse->havingQual = NULL;
379 998 : subroot->hasHavingQual = false;
380 998 : parse->distinctClause = NIL;
381 998 : parse->hasDistinctOn = false;
382 998 : parse->hasAggs = false;
383 :
384 : /* Build "target IS NOT NULL" expression */
385 998 : ntest = makeNode(NullTest);
386 998 : ntest->nulltesttype = IS_NOT_NULL;
387 998 : ntest->arg = copyObject(mminfo->target);
388 : /* we checked it wasn't a rowtype in can_minmax_aggs */
389 998 : ntest->argisrow = false;
390 998 : ntest->location = -1;
391 :
392 : /* User might have had that in WHERE already */
393 998 : if (!list_member((List *) parse->jointree->quals, ntest))
394 998 : parse->jointree->quals = (Node *)
395 998 : lcons(ntest, (List *) parse->jointree->quals);
396 :
397 : /* Build suitable ORDER BY clause */
398 998 : sortcl = makeNode(SortGroupClause);
399 998 : sortcl->tleSortGroupRef = assignSortGroupRef(tle, subroot->processed_tlist);
400 998 : sortcl->eqop = eqop;
401 998 : sortcl->sortop = sortop;
402 998 : sortcl->nulls_first = nulls_first;
403 998 : sortcl->hashable = false; /* no need to make this accurate */
404 998 : parse->sortClause = list_make1(sortcl);
405 :
406 : /* set up expressions for LIMIT 1 */
407 998 : parse->limitOffset = NULL;
408 998 : parse->limitCount = (Node *) makeConst(INT8OID, -1, InvalidOid,
409 : sizeof(int64),
410 : Int64GetDatum(1), false,
411 : FLOAT8PASSBYVAL);
412 :
413 : /*
414 : * Generate the best paths for this query, telling query_planner that we
415 : * have LIMIT 1.
416 : */
417 998 : subroot->tuple_fraction = 1.0;
418 998 : subroot->limit_tuples = 1.0;
419 :
420 998 : final_rel = query_planner(subroot, minmax_qp_callback, NULL);
421 :
422 : /*
423 : * Since we didn't go through subquery_planner() to handle the subquery,
424 : * we have to do some of the same cleanup it would do, in particular cope
425 : * with params and initplans used within this subquery. (This won't
426 : * matter if we end up not using the subplan.)
427 : */
428 998 : SS_identify_outer_params(subroot);
429 998 : SS_charge_for_initplans(subroot, final_rel);
430 :
431 : /*
432 : * Get the best presorted path, that being the one that's cheapest for
433 : * fetching just one row. If there's no such path, fail.
434 : */
435 998 : if (final_rel->rows > 1.0)
436 986 : path_fraction = 1.0 / final_rel->rows;
437 : else
438 12 : path_fraction = 1.0;
439 :
440 : sorted_path =
441 998 : get_cheapest_fractional_path_for_pathkeys(final_rel->pathlist,
442 : subroot->query_pathkeys,
443 : NULL,
444 : path_fraction);
445 998 : if (!sorted_path)
446 576 : return false;
447 :
448 : /*
449 : * The path might not return exactly what we want, so fix that. (We
450 : * assume that this won't change any conclusions about which was the
451 : * cheapest path.)
452 : */
453 422 : sorted_path = apply_projection_to_path(subroot, final_rel, sorted_path,
454 : create_pathtarget(subroot,
455 : subroot->processed_tlist));
456 :
457 : /*
458 : * Determine cost to get just the first row of the presorted path.
459 : *
460 : * Note: cost calculation here should match
461 : * compare_fractional_path_costs().
462 : */
463 422 : path_cost = sorted_path->startup_cost +
464 422 : path_fraction * (sorted_path->total_cost - sorted_path->startup_cost);
465 :
466 : /* Save state for further processing */
467 422 : mminfo->subroot = subroot;
468 422 : mminfo->path = sorted_path;
469 422 : mminfo->pathcost = path_cost;
470 :
471 422 : return true;
472 : }
473 :
474 : /*
475 : * Compute query_pathkeys and other pathkeys during query_planner()
476 : */
477 : static void
478 998 : minmax_qp_callback(PlannerInfo *root, void *extra)
479 : {
480 998 : root->group_pathkeys = NIL;
481 998 : root->window_pathkeys = NIL;
482 998 : root->distinct_pathkeys = NIL;
483 :
484 998 : root->sort_pathkeys =
485 998 : make_pathkeys_for_sortclauses(root,
486 998 : root->parse->sortClause,
487 998 : root->parse->targetList);
488 :
489 998 : root->query_pathkeys = root->sort_pathkeys;
490 998 : }
491 :
492 : /*
493 : * Get the OID of the sort operator, if any, associated with an aggregate.
494 : * Returns InvalidOid if there is no such operator.
495 : */
496 : static Oid
497 13342 : fetch_agg_sort_op(Oid aggfnoid)
498 : {
499 : HeapTuple aggTuple;
500 : Form_pg_aggregate aggform;
501 : Oid aggsortop;
502 :
503 : /* fetch aggregate entry from pg_aggregate */
504 13342 : aggTuple = SearchSysCache1(AGGFNOID, ObjectIdGetDatum(aggfnoid));
505 13342 : if (!HeapTupleIsValid(aggTuple))
506 0 : return InvalidOid;
507 13342 : aggform = (Form_pg_aggregate) GETSTRUCT(aggTuple);
508 13342 : aggsortop = aggform->aggsortop;
509 13342 : ReleaseSysCache(aggTuple);
510 :
511 13342 : return aggsortop;
512 : }
|