Age Owner Branch data TLA Line data Source code
1 : : /*-------------------------------------------------------------------------
2 : : *
3 : : * prepunion.c
4 : : * Routines to plan set-operation queries. The filename is a leftover
5 : : * from a time when only UNIONs were implemented.
6 : : *
7 : : * There are two code paths in the planner for set-operation queries.
8 : : * If a subquery consists entirely of simple UNION ALL operations, it
9 : : * is converted into an "append relation". Otherwise, it is handled
10 : : * by the general code in this module (plan_set_operations and its
11 : : * subroutines). There is some support code here for the append-relation
12 : : * case, but most of the heavy lifting for that is done elsewhere,
13 : : * notably in prepjointree.c and allpaths.c.
14 : : *
15 : : * Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group
16 : : * Portions Copyright (c) 1994, Regents of the University of California
17 : : *
18 : : *
19 : : * IDENTIFICATION
20 : : * src/backend/optimizer/prep/prepunion.c
21 : : *
22 : : *-------------------------------------------------------------------------
23 : : */
24 : : #include "postgres.h"
25 : :
26 : : #include <math.h>
27 : :
28 : : #include "access/htup_details.h"
29 : : #include "catalog/pg_type.h"
30 : : #include "miscadmin.h"
31 : : #include "nodes/makefuncs.h"
32 : : #include "nodes/nodeFuncs.h"
33 : : #include "optimizer/cost.h"
34 : : #include "optimizer/pathnode.h"
35 : : #include "optimizer/paths.h"
36 : : #include "optimizer/planner.h"
37 : : #include "optimizer/prep.h"
38 : : #include "optimizer/tlist.h"
39 : : #include "parser/parse_coerce.h"
40 : : #include "utils/selfuncs.h"
41 : :
42 : :
43 : : static RelOptInfo *recurse_set_operations(Node *setOp, PlannerInfo *root,
44 : : SetOperationStmt *parentOp,
45 : : List *colTypes, List *colCollations,
46 : : List *refnames_tlist,
47 : : List **pTargetList,
48 : : bool *istrivial_tlist);
49 : : static RelOptInfo *generate_recursion_path(SetOperationStmt *setOp,
50 : : PlannerInfo *root,
51 : : List *refnames_tlist,
52 : : List **pTargetList);
53 : : static void build_setop_child_paths(PlannerInfo *root, RelOptInfo *rel,
54 : : bool trivial_tlist, List *child_tlist,
55 : : List *interesting_pathkeys,
56 : : double *pNumGroups);
57 : : static RelOptInfo *generate_union_paths(SetOperationStmt *op, PlannerInfo *root,
58 : : List *refnames_tlist,
59 : : List **pTargetList);
60 : : static RelOptInfo *generate_nonunion_paths(SetOperationStmt *op, PlannerInfo *root,
61 : : List *refnames_tlist,
62 : : List **pTargetList);
63 : : static List *plan_union_children(PlannerInfo *root,
64 : : SetOperationStmt *top_union,
65 : : List *refnames_tlist,
66 : : List **tlist_list,
67 : : List **istrivial_tlist);
68 : : static void postprocess_setop_rel(PlannerInfo *root, RelOptInfo *rel);
69 : : static List *generate_setop_tlist(List *colTypes, List *colCollations,
70 : : Index varno,
71 : : bool hack_constants,
72 : : List *input_tlist,
73 : : List *refnames_tlist,
74 : : bool *trivial_tlist);
75 : : static List *generate_append_tlist(List *colTypes, List *colCollations,
76 : : List *input_tlists,
77 : : List *refnames_tlist);
78 : : static List *generate_setop_grouplist(SetOperationStmt *op, List *targetlist);
79 : : static PathTarget *create_setop_pathtarget(PlannerInfo *root, List *tlist,
80 : : List *child_pathlist);
81 : :
82 : :
83 : : /*
84 : : * plan_set_operations
85 : : *
86 : : * Plans the queries for a tree of set operations (UNION/INTERSECT/EXCEPT)
87 : : *
88 : : * This routine only deals with the setOperations tree of the given query.
89 : : * Any top-level ORDER BY requested in root->parse->sortClause will be handled
90 : : * when we return to grouping_planner; likewise for LIMIT.
91 : : *
92 : : * What we return is an "upperrel" RelOptInfo containing at least one Path
93 : : * that implements the set-operation tree. In addition, root->processed_tlist
94 : : * receives a targetlist representing the output of the topmost setop node.
95 : : */
96 : : RelOptInfo *
3522 tgl@sss.pgh.pa.us 97 :CBC 3076 : plan_set_operations(PlannerInfo *root)
98 : : {
7450 99 : 3076 : Query *parse = root->parse;
3171 peter_e@gmx.net 100 : 3076 : SetOperationStmt *topop = castNode(SetOperationStmt, parse->setOperations);
101 : : Node *node;
102 : : RangeTblEntry *leftmostRTE;
103 : : Query *leftmostQuery;
104 : : RelOptInfo *setop_rel;
105 : : List *top_tlist;
106 : :
107 [ - + ]: 3076 : Assert(topop);
108 : :
109 : : /* check for unsupported stuff */
8140 tgl@sss.pgh.pa.us 110 [ - + ]: 3076 : Assert(parse->jointree->fromlist == NIL);
111 [ - + ]: 3076 : Assert(parse->jointree->quals == NULL);
112 [ - + ]: 3076 : Assert(parse->groupClause == NIL);
113 [ - + ]: 3076 : Assert(parse->havingQual == NULL);
6148 114 [ - + ]: 3076 : Assert(parse->windowClause == NIL);
8140 115 [ - + ]: 3076 : Assert(parse->distinctClause == NIL);
116 : :
117 : : /*
118 : : * In the outer query level, equivalence classes are limited to classes
119 : : * which define that the top-level target entry is equivalent to the
120 : : * corresponding child target entry. There won't be any equivalence class
121 : : * merging. Mark that merging is complete to allow us to make pathkeys.
122 : : */
2291 drowley@postgresql.o 123 [ - + ]: 3076 : Assert(root->eq_classes == NIL);
124 : 3076 : root->ec_merging_done = true;
125 : :
126 : : /*
127 : : * We'll need to build RelOptInfos for each of the leaf subqueries, which
128 : : * are RTE_SUBQUERY rangetable entries in this Query. Prepare the index
129 : : * arrays for those, and for AppendRelInfos in case they're needed.
130 : : */
5169 tgl@sss.pgh.pa.us 131 : 3076 : setup_simple_rel_arrays(root);
132 : :
133 : : /*
134 : : * Find the leftmost component Query. We need to use its column names for
135 : : * all generated tlists (else SELECT INTO won't work right).
136 : : */
9154 137 : 3076 : node = topop->larg;
138 [ + - + + ]: 4970 : while (node && IsA(node, SetOperationStmt))
139 : 1894 : node = ((SetOperationStmt *) node)->larg;
140 [ + - - + ]: 3076 : Assert(node && IsA(node, RangeTblRef));
5169 141 : 3076 : leftmostRTE = root->simple_rte_array[((RangeTblRef *) node)->rtindex];
142 : 3076 : leftmostQuery = leftmostRTE->subquery;
9154 143 [ - + ]: 3076 : Assert(leftmostQuery != NULL);
144 : :
145 : : /*
146 : : * If the topmost node is a recursive union, it needs special processing.
147 : : */
6233 148 [ + + ]: 3076 : if (root->hasRecursion)
149 : : {
2780 rhaas@postgresql.org 150 : 470 : setop_rel = generate_recursion_path(topop, root,
151 : : leftmostQuery->targetList,
152 : : &top_tlist);
153 : : }
154 : : else
155 : : {
156 : : bool trivial_tlist;
157 : :
158 : : /*
159 : : * Recurse on setOperations tree to generate paths for set ops. The
160 : : * final output paths should have just the column types shown as the
161 : : * output from the top-level node.
162 : : */
163 : 2606 : setop_rel = recurse_set_operations((Node *) topop, root,
164 : : NULL, /* no parent */
165 : : topop->colTypes, topop->colCollations,
166 : : leftmostQuery->targetList,
167 : : &top_tlist,
168 : : &trivial_tlist);
169 : : }
170 : :
171 : : /* Must return the built tlist into root->processed_tlist. */
3522 tgl@sss.pgh.pa.us 172 : 3073 : root->processed_tlist = top_tlist;
173 : :
174 : 3073 : return setop_rel;
175 : : }
176 : :
177 : : /*
178 : : * recurse_set_operations
179 : : * Recursively handle one step in a tree of set operations
180 : : *
181 : : * setOp: current step (could be a SetOperationStmt or a leaf RangeTblRef)
182 : : * parentOp: parent step, or NULL if none (but see below)
183 : : * colTypes: OID list of set-op's result column datatypes
184 : : * colCollations: OID list of set-op's result column collations
185 : : * refnames_tlist: targetlist to take column names from
186 : : *
187 : : * parentOp should be passed as NULL unless that step is interested in
188 : : * getting sorted output from this step. ("Sorted" means "sorted according
189 : : * to the default btree opclasses of the result column datatypes".)
190 : : *
191 : : * Returns a RelOptInfo for the subtree, as well as these output parameters:
192 : : * *pTargetList: receives the fully-fledged tlist for the subtree's top plan
193 : : * *istrivial_tlist: true if, and only if, datatypes between parent and child
194 : : * match.
195 : : *
196 : : * If setOp is a leaf node, this function plans the sub-query but does
197 : : * not populate the pathlist of the returned RelOptInfo. The caller will
198 : : * generate SubqueryScan paths using useful path(s) of the subquery (see
199 : : * build_setop_child_paths). But this function does build the paths for
200 : : * set-operation nodes.
201 : : *
202 : : * The pTargetList output parameter is mostly redundant with the pathtarget
203 : : * of the returned RelOptInfo, but for the moment we need it because much of
204 : : * the logic in this file depends on flag columns being marked resjunk.
205 : : * XXX Now that there are no flag columns and hence no resjunk columns, we
206 : : * could probably refactor this file to deal only in pathtargets.
207 : : *
208 : : * We don't have to care about typmods here: the only allowed difference
209 : : * between set-op input and output typmods is input is a specific typmod
210 : : * and output is -1, and that does not require a coercion.
211 : : */
212 : : static RelOptInfo *
7450 213 : 10763 : recurse_set_operations(Node *setOp, PlannerInfo *root,
214 : : SetOperationStmt *parentOp,
215 : : List *colTypes, List *colCollations,
216 : : List *refnames_tlist,
217 : : List **pTargetList,
218 : : bool *istrivial_tlist)
219 : : {
220 : : RelOptInfo *rel;
221 : :
525 rhaas@postgresql.org 222 : 10763 : *istrivial_tlist = true; /* for now */
223 : :
224 : : /* Guard against stack overflow due to overly complex setop nests */
2830 tgl@sss.pgh.pa.us 225 : 10763 : check_stack_depth();
226 : :
9154 227 [ + + ]: 10763 : if (IsA(setOp, RangeTblRef))
228 : : {
229 : 8079 : RangeTblRef *rtr = (RangeTblRef *) setOp;
5169 230 : 8079 : RangeTblEntry *rte = root->simple_rte_array[rtr->rtindex];
8986 bruce@momjian.us 231 : 8079 : Query *subquery = rte->subquery;
232 : : PlannerInfo *subroot;
233 : : List *tlist;
234 : : bool trivial_tlist;
235 : : char *plan_name;
236 : :
9154 tgl@sss.pgh.pa.us 237 [ - + ]: 8079 : Assert(subquery != NULL);
238 : :
239 : : /* Build a RelOptInfo for this leaf subquery. */
3130 rhaas@postgresql.org 240 : 8079 : rel = build_simple_rel(root, rtr->rtindex, NULL);
241 : :
242 : : /* plan_params should not be in use in current query level */
4801 tgl@sss.pgh.pa.us 243 [ - + ]: 8079 : Assert(root->plan_params == NIL);
244 : :
245 : : /*
246 : : * Generate a subroot and Paths for the subquery. If we have a
247 : : * parentOp, pass that down to encourage subquery_planner to consider
248 : : * suitably-sorted Paths.
249 : : */
21 rhaas@postgresql.org 250 :GNC 8079 : plan_name = choose_plan_name(root->glob, "setop", true);
251 : 8079 : subroot = rel->subroot = subquery_planner(root->glob, subquery,
252 : : plan_name, root,
253 : : false, root->tuple_fraction,
254 : : parentOp);
255 : :
256 : : /*
257 : : * It should not be possible for the primitive query to contain any
258 : : * cross-references to other primitive queries in the setop tree.
259 : : */
4801 tgl@sss.pgh.pa.us 260 [ - + ]:CBC 8079 : if (root->plan_params)
4801 tgl@sss.pgh.pa.us 261 [ # # ]:UBC 0 : elog(ERROR, "unexpected outer reference in set operation subquery");
262 : :
263 : : /* Figure out the appropriate target list for this subquery. */
2780 rhaas@postgresql.org 264 :CBC 8079 : tlist = generate_setop_tlist(colTypes, colCollations,
265 : 8079 : rtr->rtindex,
266 : : true,
267 : : subroot->processed_tlist,
268 : : refnames_tlist,
269 : : &trivial_tlist);
270 : 8079 : rel->reltarget = create_pathtarget(root, tlist);
271 : :
272 : : /* Return the fully-fledged tlist to caller, too */
273 : 8079 : *pTargetList = tlist;
525 274 : 8079 : *istrivial_tlist = trivial_tlist;
275 : : }
9154 tgl@sss.pgh.pa.us 276 [ + - ]: 2684 : else if (IsA(setOp, SetOperationStmt))
277 : : {
278 : 2684 : SetOperationStmt *op = (SetOperationStmt *) setOp;
279 : :
280 : : /* UNIONs are much different from INTERSECT/EXCEPT */
281 [ + + ]: 2684 : if (op->op == SETOP_UNION)
2780 rhaas@postgresql.org 282 : 2338 : rel = generate_union_paths(op, root,
283 : : refnames_tlist,
284 : : pTargetList);
285 : : else
286 : 346 : rel = generate_nonunion_paths(op, root,
287 : : refnames_tlist,
288 : : pTargetList);
289 : :
290 : : /*
291 : : * If necessary, add a Result node to project the caller-requested
292 : : * output columns.
293 : : *
294 : : * XXX you don't really want to know about this: setrefs.c will apply
295 : : * fix_upper_expr() to the Result node's tlist. This would fail if the
296 : : * Vars generated by generate_setop_tlist() were not exactly equal()
297 : : * to the corresponding tlist entries of the subplan. However, since
298 : : * the subplan was generated by generate_union_paths() or
299 : : * generate_nonunion_paths(), and hence its tlist was generated by
300 : : * generate_append_tlist() or generate_setop_tlist(), this will work.
301 : : * We just tell generate_setop_tlist() to use varno 0.
302 : : */
313 tgl@sss.pgh.pa.us 303 [ + + ]: 2684 : if (!tlist_same_datatypes(*pTargetList, colTypes, false) ||
304 [ - + ]: 2678 : !tlist_same_collations(*pTargetList, colCollations, false))
305 : : {
306 : : PathTarget *target;
307 : : bool trivial_tlist;
308 : : ListCell *lc;
309 : :
3522 310 : 6 : *pTargetList = generate_setop_tlist(colTypes, colCollations,
311 : : 0,
312 : : false,
313 : : *pTargetList,
314 : : refnames_tlist,
315 : : &trivial_tlist);
525 rhaas@postgresql.org 316 : 6 : *istrivial_tlist = trivial_tlist;
2780 317 : 6 : target = create_pathtarget(root, *pTargetList);
318 : :
319 : : /* Apply projection to each path */
320 [ + - + + : 12 : foreach(lc, rel->pathlist)
+ + ]
321 : : {
322 : 6 : Path *subpath = (Path *) lfirst(lc);
323 : : Path *path;
324 : :
325 [ - + ]: 6 : Assert(subpath->param_info == NULL);
326 : 6 : path = apply_projection_to_path(root, subpath->parent,
327 : : subpath, target);
328 : : /* If we had to add a Result, path is different from subpath */
329 [ + - ]: 6 : if (path != subpath)
330 : 6 : lfirst(lc) = path;
331 : : }
332 : :
333 : : /* Apply projection to each partial path */
334 [ - + - - : 6 : foreach(lc, rel->partial_pathlist)
- + ]
335 : : {
2780 rhaas@postgresql.org 336 :UBC 0 : Path *subpath = (Path *) lfirst(lc);
337 : : Path *path;
338 : :
339 [ # # ]: 0 : Assert(subpath->param_info == NULL);
340 : :
341 : : /* avoid apply_projection_to_path, in case of multiple refs */
342 : 0 : path = (Path *) create_projection_path(root, subpath->parent,
343 : : subpath, target);
344 : 0 : lfirst(lc) = path;
345 : : }
346 : : }
525 rhaas@postgresql.org 347 :CBC 2684 : postprocess_setop_rel(root, rel);
348 : : }
349 : : else
350 : : {
8131 tgl@sss.pgh.pa.us 351 [ # # ]:UBC 0 : elog(ERROR, "unrecognized node type: %d",
352 : : (int) nodeTag(setOp));
353 : : *pTargetList = NIL;
354 : : rel = NULL; /* keep compiler quiet */
355 : : }
356 : :
2780 rhaas@postgresql.org 357 :CBC 10763 : return rel;
358 : : }
359 : :
360 : : /*
361 : : * Generate paths for a recursive UNION node
362 : : */
363 : : static RelOptInfo *
3522 tgl@sss.pgh.pa.us 364 : 470 : generate_recursion_path(SetOperationStmt *setOp, PlannerInfo *root,
365 : : List *refnames_tlist,
366 : : List **pTargetList)
367 : : {
368 : : RelOptInfo *result_rel;
369 : : Path *path;
370 : : RelOptInfo *lrel,
371 : : *rrel;
372 : : Path *lpath;
373 : : Path *rpath;
374 : : List *lpath_tlist;
375 : : bool lpath_trivial_tlist;
376 : : List *rpath_tlist;
377 : : bool rpath_trivial_tlist;
378 : : List *tlist;
379 : : List *groupList;
380 : : double dNumGroups;
381 : :
382 : : /* Parser should have rejected other cases */
6230 383 [ - + ]: 470 : if (setOp->op != SETOP_UNION)
6230 tgl@sss.pgh.pa.us 384 [ # # ]:UBC 0 : elog(ERROR, "only UNION queries can be recursive");
385 : : /* Worktable ID should be assigned */
6233 tgl@sss.pgh.pa.us 386 [ - + ]:CBC 470 : Assert(root->wt_param_id >= 0);
387 : :
388 : : /*
389 : : * Unlike a regular UNION node, process the left and right inputs
390 : : * separately without any intention of combining them into one Append.
391 : : */
2780 rhaas@postgresql.org 392 : 470 : lrel = recurse_set_operations(setOp->larg, root,
393 : : NULL, /* no value in sorted results */
394 : : setOp->colTypes, setOp->colCollations,
395 : : refnames_tlist,
396 : : &lpath_tlist,
397 : : &lpath_trivial_tlist);
525 398 [ + - ]: 470 : if (lrel->rtekind == RTE_SUBQUERY)
399 : 470 : build_setop_child_paths(root, lrel, lpath_trivial_tlist, lpath_tlist,
400 : : NIL, NULL);
2780 401 : 470 : lpath = lrel->cheapest_total_path;
402 : : /* The right path will want to look at the left one ... */
3522 tgl@sss.pgh.pa.us 403 : 470 : root->non_recursive_path = lpath;
2780 rhaas@postgresql.org 404 : 470 : rrel = recurse_set_operations(setOp->rarg, root,
405 : : NULL, /* no value in sorted results */
406 : : setOp->colTypes, setOp->colCollations,
407 : : refnames_tlist,
408 : : &rpath_tlist,
409 : : &rpath_trivial_tlist);
525 410 [ + + ]: 470 : if (rrel->rtekind == RTE_SUBQUERY)
411 : 467 : build_setop_child_paths(root, rrel, rpath_trivial_tlist, rpath_tlist,
412 : : NIL, NULL);
2780 413 : 470 : rpath = rrel->cheapest_total_path;
3522 tgl@sss.pgh.pa.us 414 : 470 : root->non_recursive_path = NULL;
415 : :
416 : : /*
417 : : * Generate tlist for RecursiveUnion path node --- same as in Append cases
418 : : */
313 419 : 470 : tlist = generate_append_tlist(setOp->colTypes, setOp->colCollations,
3522 420 : 470 : list_make2(lpath_tlist, rpath_tlist),
421 : : refnames_tlist);
422 : :
423 : 470 : *pTargetList = tlist;
424 : :
425 : : /* Build result relation. */
2780 rhaas@postgresql.org 426 : 470 : result_rel = fetch_upper_rel(root, UPPERREL_SETOP,
427 : 470 : bms_union(lrel->relids, rrel->relids));
428 : 470 : result_rel->reltarget = create_pathtarget(root, tlist);
429 : :
430 : : /*
431 : : * If UNION, identify the grouping operators
432 : : */
6230 tgl@sss.pgh.pa.us 433 [ + + ]: 470 : if (setOp->all)
434 : : {
435 : 274 : groupList = NIL;
3522 436 : 274 : dNumGroups = 0;
437 : : }
438 : : else
439 : : {
440 : : /* Identify the grouping semantics */
6230 441 : 196 : groupList = generate_setop_grouplist(setOp, tlist);
442 : :
443 : : /* We only support hashing here */
444 [ + + ]: 196 : if (!grouping_is_hashable(groupList))
445 [ + - ]: 3 : ereport(ERROR,
446 : : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
447 : : errmsg("could not implement recursive UNION"),
448 : : errdetail("All column datatypes must be hashable.")));
449 : :
450 : : /*
451 : : * For the moment, take the number of distinct groups as equal to the
452 : : * total input size, ie, the worst case.
453 : : */
3522 454 : 193 : dNumGroups = lpath->rows + rpath->rows * 10;
455 : : }
456 : :
457 : : /*
458 : : * And make the path node.
459 : : */
460 : 467 : path = (Path *) create_recursiveunion_path(root,
461 : : result_rel,
462 : : lpath,
463 : : rpath,
2780 rhaas@postgresql.org 464 : 467 : result_rel->reltarget,
465 : : groupList,
466 : : root->wt_param_id,
467 : : dNumGroups);
468 : :
469 : 467 : add_path(result_rel, path);
470 : 467 : postprocess_setop_rel(root, result_rel);
471 : 467 : return result_rel;
472 : : }
473 : :
474 : : /*
475 : : * build_setop_child_paths
476 : : * Build paths for the set op child relation denoted by 'rel'.
477 : : *
478 : : * 'rel' is an RTE_SUBQUERY relation. We have already generated paths within
479 : : * the subquery's subroot; the task here is to create SubqueryScan paths for
480 : : * 'rel', representing scans of the useful subquery paths.
481 : : *
482 : : * interesting_pathkeys: if not NIL, also include paths that suit these
483 : : * pathkeys, sorting any unsorted paths as required.
484 : : * *pNumGroups: if not NULL, we estimate the number of distinct groups
485 : : * in the result, and store it there.
486 : : */
487 : : static void
525 488 : 8079 : build_setop_child_paths(PlannerInfo *root, RelOptInfo *rel,
489 : : bool trivial_tlist, List *child_tlist,
490 : : List *interesting_pathkeys, double *pNumGroups)
491 : : {
492 : : RelOptInfo *final_rel;
493 : 8079 : List *setop_pathkeys = rel->subroot->setop_pathkeys;
494 : : ListCell *lc;
495 : :
496 : : /* it can't be a set op child rel if it's not a subquery */
497 [ - + ]: 8079 : Assert(rel->rtekind == RTE_SUBQUERY);
498 : :
499 : : /* when sorting is needed, add child rel equivalences */
500 [ + + ]: 8079 : if (interesting_pathkeys != NIL)
501 : 6049 : add_setop_child_rel_equivalences(root,
502 : : rel,
503 : : child_tlist,
504 : : interesting_pathkeys);
505 : :
506 : : /*
507 : : * Mark rel with estimated output rows, width, etc. Note that we have to
508 : : * do this before generating outer-query paths, else cost_subqueryscan is
509 : : * not happy.
510 : : */
511 : 8079 : set_subquery_size_estimates(root, rel);
512 : :
513 : : /*
514 : : * Since we may want to add a partial path to this relation, we must set
515 : : * its consider_parallel flag correctly.
516 : : */
517 : 8079 : final_rel = fetch_upper_rel(rel->subroot, UPPERREL_FINAL, NULL);
518 : 8079 : rel->consider_parallel = final_rel->consider_parallel;
519 : :
520 : : /* Generate subquery scan paths for any interesting path in final_rel */
521 [ + - + + : 20873 : foreach(lc, final_rel->pathlist)
+ + ]
522 : : {
523 : 12794 : Path *subpath = (Path *) lfirst(lc);
524 : : List *pathkeys;
525 : 12794 : Path *cheapest_input_path = final_rel->cheapest_total_path;
526 : : bool is_sorted;
527 : : int presorted_keys;
528 : :
529 : : /* If the input rel is dummy, propagate that to this query level */
24 drowley@postgresql.o 530 [ + + ]:GNC 12794 : if (is_dummy_rel(final_rel))
531 : : {
532 : 45 : mark_dummy_rel(rel);
533 : 2091 : continue;
534 : : }
535 : :
536 : : /*
537 : : * Include the cheapest path as-is so that the set operation can be
538 : : * cheaply implemented using a method which does not require the input
539 : : * to be sorted.
540 : : */
525 rhaas@postgresql.org 541 [ + + ]:CBC 12749 : if (subpath == cheapest_input_path)
542 : : {
543 : : /* Convert subpath's pathkeys to outer representation */
544 : 8034 : pathkeys = convert_subquery_pathkeys(root, rel, subpath->pathkeys,
545 : : make_tlist_from_pathtarget(subpath->pathtarget));
546 : :
547 : : /* Generate outer path using this subpath */
548 : 8034 : add_path(rel, (Path *) create_subqueryscan_path(root,
549 : : rel,
550 : : subpath,
551 : : trivial_tlist,
552 : : pathkeys,
553 : : NULL));
554 : : }
555 : :
556 : : /* skip dealing with sorted paths if the setop doesn't need them */
557 [ + + ]: 12749 : if (interesting_pathkeys == NIL)
558 : 2040 : continue;
559 : :
560 : : /*
561 : : * Create paths to suit final sort order required for setop_pathkeys.
562 : : * Here we'll sort the cheapest input path (if not sorted already) and
563 : : * incremental sort any paths which are partially sorted.
564 : : */
565 : 10709 : is_sorted = pathkeys_count_contained_in(setop_pathkeys,
566 : : subpath->pathkeys,
567 : : &presorted_keys);
568 : :
569 [ + + ]: 10709 : if (!is_sorted)
570 : : {
571 : 7039 : double limittuples = rel->subroot->limit_tuples;
572 : :
573 : : /*
574 : : * Try at least sorting the cheapest path and also try
575 : : * incrementally sorting any path which is partially sorted
576 : : * already (no need to deal with paths which have presorted keys
577 : : * when incremental sort is disabled unless it's the cheapest
578 : : * input path).
579 : : */
580 [ + + ]: 7039 : if (subpath != cheapest_input_path &&
581 [ + + - + ]: 1563 : (presorted_keys == 0 || !enable_incremental_sort))
582 : 6 : continue;
583 : :
584 : : /*
585 : : * We've no need to consider both a sort and incremental sort.
586 : : * We'll just do a sort if there are no presorted keys and an
587 : : * incremental sort when there are presorted keys.
588 : : */
589 [ + + + + ]: 7033 : if (presorted_keys == 0 || !enable_incremental_sort)
590 : 5452 : subpath = (Path *) create_sort_path(rel->subroot,
591 : : final_rel,
592 : : subpath,
593 : : setop_pathkeys,
594 : : limittuples);
595 : : else
596 : 1581 : subpath = (Path *) create_incremental_sort_path(rel->subroot,
597 : : final_rel,
598 : : subpath,
599 : : setop_pathkeys,
600 : : presorted_keys,
601 : : limittuples);
602 : : }
603 : :
604 : : /*
605 : : * subpath is now sorted, so add it to the pathlist. We already added
606 : : * the cheapest_input_path above, so don't add it again unless we just
607 : : * sorted it.
608 : : */
609 [ + + ]: 10703 : if (subpath != cheapest_input_path)
610 : : {
611 : : /* Convert subpath's pathkeys to outer representation */
612 : 10163 : pathkeys = convert_subquery_pathkeys(root, rel, subpath->pathkeys,
613 : : make_tlist_from_pathtarget(subpath->pathtarget));
614 : :
615 : : /* Generate outer path using this subpath */
616 : 10163 : add_path(rel, (Path *) create_subqueryscan_path(root,
617 : : rel,
618 : : subpath,
619 : : trivial_tlist,
620 : : pathkeys,
621 : : NULL));
622 : : }
623 : : }
624 : :
625 : : /* if consider_parallel is false, there should be no partial paths */
626 [ + + - + ]: 8079 : Assert(final_rel->consider_parallel ||
627 : : final_rel->partial_pathlist == NIL);
628 : :
629 : : /*
630 : : * If we have a partial path for the child relation, we can use that to
631 : : * build a partial path for this relation. But there's no point in
632 : : * considering any path but the cheapest.
633 : : */
634 [ + + + - ]: 8079 : if (rel->consider_parallel && bms_is_empty(rel->lateral_relids) &&
635 [ + + ]: 5611 : final_rel->partial_pathlist != NIL)
636 : : {
637 : : Path *partial_subpath;
638 : : Path *partial_path;
639 : :
640 : 6 : partial_subpath = linitial(final_rel->partial_pathlist);
641 : : partial_path = (Path *)
642 : 6 : create_subqueryscan_path(root, rel, partial_subpath,
643 : : trivial_tlist,
644 : : NIL, NULL);
645 : 6 : add_partial_path(rel, partial_path);
646 : : }
647 : :
648 : 8079 : postprocess_setop_rel(root, rel);
649 : :
650 : : /*
651 : : * Estimate number of groups if caller wants it. If the subquery used
652 : : * grouping or aggregation, its output is probably mostly unique anyway;
653 : : * otherwise do statistical estimation.
654 : : *
655 : : * XXX you don't really want to know about this: we do the estimation
656 : : * using the subroot->parse's original targetlist expressions, not the
657 : : * subroot->processed_tlist which might seem more appropriate. The reason
658 : : * is that if the subquery is itself a setop, it may return a
659 : : * processed_tlist containing "varno 0" Vars generated by
660 : : * generate_append_tlist, and those would confuse estimate_num_groups
661 : : * mightily. We ought to get rid of the "varno 0" hack, but that requires
662 : : * a redesign of the parsetree representation of setops, so that there can
663 : : * be an RTE corresponding to each setop's output. Note, we use this not
664 : : * subquery's targetlist but subroot->parse's targetlist, because it was
665 : : * revised by self-join removal. subquery's targetlist might contain the
666 : : * references to the removed relids.
667 : : */
668 [ + + ]: 8079 : if (pNumGroups)
669 : : {
670 : 677 : PlannerInfo *subroot = rel->subroot;
671 : 677 : Query *subquery = subroot->parse;
672 : :
673 [ + - + - ]: 677 : if (subquery->groupClause || subquery->groupingSets ||
674 [ + + + - ]: 677 : subquery->distinctClause || subroot->hasHavingQual ||
675 [ - + ]: 671 : subquery->hasAggs)
676 : 6 : *pNumGroups = rel->cheapest_total_path->rows;
677 : : else
678 : 671 : *pNumGroups = estimate_num_groups(subroot,
257 akorotkov@postgresql 679 : 671 : get_tlist_exprs(subroot->parse->targetList, false),
525 rhaas@postgresql.org 680 : 671 : rel->cheapest_total_path->rows,
681 : : NULL,
682 : : NULL);
683 : : }
684 : 8079 : }
685 : :
686 : : /*
687 : : * Generate paths for a UNION or UNION ALL node
688 : : */
689 : : static RelOptInfo *
2780 690 : 2338 : generate_union_paths(SetOperationStmt *op, PlannerInfo *root,
691 : : List *refnames_tlist,
692 : : List **pTargetList)
693 : : {
694 : 2338 : Relids relids = NULL;
695 : : RelOptInfo *result_rel;
696 : : ListCell *lc;
697 : : ListCell *lc2;
698 : : ListCell *lc3;
525 699 : 2338 : List *cheapest_pathlist = NIL;
700 : 2338 : List *ordered_pathlist = NIL;
2777 701 : 2338 : List *partial_pathlist = NIL;
702 : 2338 : bool partial_paths_valid = true;
703 : 2338 : bool consider_parallel = true;
704 : : List *rellist;
705 : : List *tlist_list;
706 : : List *trivial_tlist_list;
707 : : List *tlist;
525 708 : 2338 : List *groupList = NIL;
709 : : Path *apath;
710 : 2338 : Path *gpath = NULL;
711 : 2338 : bool try_sorted = false;
712 : 2338 : List *union_pathkeys = NIL;
713 : :
714 : : /*
715 : : * If any of my children are identical UNION nodes (same op, all-flag, and
716 : : * colTypes/colCollations) then they can be merged into this node so that
717 : : * we generate only one Append/MergeAppend and unique-ification for the
718 : : * lot. Recurse to find such nodes.
719 : : */
720 : 2338 : rellist = plan_union_children(root,
721 : : op,
722 : : refnames_tlist,
723 : : &tlist_list,
724 : : &trivial_tlist_list);
725 : :
726 : : /*
727 : : * Generate tlist for Append/MergeAppend plan node.
728 : : *
729 : : * The tlist for an Append plan isn't important as far as the Append is
730 : : * concerned, but we must make it look real anyway for the benefit of the
731 : : * next plan level up.
732 : : */
313 tgl@sss.pgh.pa.us 733 : 2338 : tlist = generate_append_tlist(op->colTypes, op->colCollations,
734 : : tlist_list, refnames_tlist);
526 735 : 2338 : *pTargetList = tlist;
736 : :
737 : : /* For UNIONs (not UNION ALL), try sorting, if sorting is possible */
525 rhaas@postgresql.org 738 [ + + ]: 2338 : if (!op->all)
739 : : {
740 : : /* Identify the grouping semantics */
741 : 1965 : groupList = generate_setop_grouplist(op, tlist);
742 : :
743 [ + + ]: 1965 : if (grouping_is_sortable(op->groupClauses))
744 : : {
745 : 1914 : try_sorted = true;
746 : : /* Determine the pathkeys for sorting by the whole target list */
747 : 1914 : union_pathkeys = make_pathkeys_for_sortclauses(root, groupList,
748 : : tlist);
749 : :
750 : 1914 : root->query_pathkeys = union_pathkeys;
751 : : }
752 : : }
753 : :
754 : : /*
755 : : * Now that we've got the append target list, we can build the union child
756 : : * paths.
757 : : */
758 [ + - + + : 8863 : forthree(lc, rellist, lc2, trivial_tlist_list, lc3, tlist_list)
+ - + + +
- + + + +
+ - + - +
+ ]
759 : : {
760 : 6525 : RelOptInfo *rel = lfirst(lc);
761 : 6525 : bool trivial_tlist = lfirst_int(lc2);
762 : 6525 : List *child_tlist = lfirst_node(List, lc3);
763 : :
764 : : /* only build paths for the union children */
765 [ + + ]: 6525 : if (rel->rtekind == RTE_SUBQUERY)
766 : 6465 : build_setop_child_paths(root, rel, trivial_tlist, child_tlist,
767 : : union_pathkeys, NULL);
768 : : }
769 : :
770 : : /* Build path lists and relid set. */
2780 771 [ + - + + : 8863 : foreach(lc, rellist)
+ + ]
772 : : {
773 : 6525 : RelOptInfo *rel = lfirst(lc);
774 : : Path *ordered_path;
775 : :
776 : : /* Skip any UNION children that are proven not to yield any rows */
24 drowley@postgresql.o 777 [ + + ]:GNC 6525 : if (is_dummy_rel(rel))
778 : 27 : continue;
779 : :
525 rhaas@postgresql.org 780 :CBC 6498 : cheapest_pathlist = lappend(cheapest_pathlist,
781 : 6498 : rel->cheapest_total_path);
782 : :
783 [ + + ]: 6498 : if (try_sorted)
784 : : {
785 : 2075 : ordered_path = get_cheapest_path_for_pathkeys(rel->pathlist,
786 : : union_pathkeys,
787 : : NULL,
788 : : TOTAL_COST,
789 : : false);
790 : :
791 [ + + ]: 2075 : if (ordered_path != NULL)
792 : 316 : ordered_pathlist = lappend(ordered_pathlist, ordered_path);
793 : : else
794 : : {
795 : : /*
796 : : * If we can't find a sorted path, just give up trying to
797 : : * generate a list of correctly sorted child paths. This can
798 : : * happen when type coercion was added to the targetlist due
799 : : * to mismatching types from the union children.
800 : : */
801 : 1759 : try_sorted = false;
802 : : }
803 : : }
804 : :
2777 805 [ + + ]: 6498 : if (consider_parallel)
806 : : {
807 [ + + ]: 4691 : if (!rel->consider_parallel)
808 : : {
809 : 1687 : consider_parallel = false;
810 : 1687 : partial_paths_valid = false;
811 : : }
812 [ + + ]: 3004 : else if (rel->partial_pathlist == NIL)
813 : 2998 : partial_paths_valid = false;
814 : : else
815 : 6 : partial_pathlist = lappend(partial_pathlist,
816 : 6 : linitial(rel->partial_pathlist));
817 : : }
818 : :
24 drowley@postgresql.o 819 :GNC 6498 : relids = bms_add_members(relids, rel->relids);
820 : : }
821 : :
822 : : /* Build result relation. */
2780 rhaas@postgresql.org 823 :CBC 2338 : result_rel = fetch_upper_rel(root, UPPERREL_SETOP, relids);
29 drowley@postgresql.o 824 :GNC 2338 : result_rel->reltarget = create_setop_pathtarget(root, tlist,
825 : : cheapest_pathlist);
2777 rhaas@postgresql.org 826 :CBC 2338 : result_rel->consider_parallel = consider_parallel;
525 827 : 2338 : result_rel->consider_startup = (root->tuple_fraction > 0);
828 : :
829 : : /* If all UNION children were dummy rels, make the resulting rel dummy */
24 drowley@postgresql.o 830 [ + + ]:GNC 2338 : if (cheapest_pathlist == NIL)
831 : : {
832 : 3 : mark_dummy_rel(result_rel);
833 : :
834 : 3 : return result_rel;
835 : : }
836 : :
837 : : /*
838 : : * Append the child results together using the cheapest paths from each
839 : : * union child.
840 : : */
525 rhaas@postgresql.org 841 :CBC 2335 : apath = (Path *) create_append_path(root, result_rel, cheapest_pathlist,
842 : : NIL, NIL, NULL, 0, false, -1);
843 : :
844 : : /*
845 : : * Estimate number of groups. For now we just assume the output is unique
846 : : * --- this is certainly true for the UNION case, and we want worst-case
847 : : * estimates anyway.
848 : : */
849 : 2335 : result_rel->rows = apath->rows;
850 : :
851 : : /*
852 : : * Now consider doing the same thing using the partial paths plus Append
853 : : * plus Gather.
854 : : */
2777 855 [ + + ]: 2335 : if (partial_paths_valid)
856 : : {
857 : : Path *papath;
858 : 3 : int parallel_workers = 0;
859 : :
860 : : /* Find the highest number of workers requested for any subpath. */
861 [ + - + + : 9 : foreach(lc, partial_pathlist)
+ + ]
862 : : {
1119 drowley@postgresql.o 863 : 6 : Path *subpath = lfirst(lc);
864 : :
865 : 6 : parallel_workers = Max(parallel_workers,
866 : : subpath->parallel_workers);
867 : : }
2777 rhaas@postgresql.org 868 [ - + ]: 3 : Assert(parallel_workers > 0);
869 : :
870 : : /*
871 : : * If the use of parallel append is permitted, always request at least
872 : : * log2(# of children) paths. We assume it can be useful to have
873 : : * extra workers in this case because they will be spread out across
874 : : * the children. The precise formula is just a guess; see
875 : : * add_paths_to_append_rel.
876 : : */
877 [ + - ]: 3 : if (enable_parallel_append)
878 : : {
879 [ + - ]: 3 : parallel_workers = Max(parallel_workers,
880 : : pg_leftmost_one_pos32(list_length(partial_pathlist)) + 1);
881 : 3 : parallel_workers = Min(parallel_workers,
882 : : max_parallel_workers_per_gather);
883 : : }
884 [ - + ]: 3 : Assert(parallel_workers > 0);
885 : :
886 : : papath = (Path *)
2761 alvherre@alvh.no-ip. 887 : 3 : create_append_path(root, result_rel, NIL, partial_pathlist,
888 : : NIL, NULL, parallel_workers,
889 : : enable_parallel_append, -1);
890 : : gpath = (Path *)
525 rhaas@postgresql.org 891 : 3 : create_gather_path(root, result_rel, papath,
2777 892 : 3 : result_rel->reltarget, NULL, NULL);
893 : : }
894 : :
525 895 [ + + ]: 2335 : if (!op->all)
896 : : {
897 : : double dNumGroups;
898 : 1962 : bool can_sort = grouping_is_sortable(groupList);
899 : 1962 : bool can_hash = grouping_is_hashable(groupList);
900 : :
24 drowley@postgresql.o 901 [ + + ]:GNC 1962 : if (list_length(cheapest_pathlist) == 1)
902 : : {
903 : 6 : Path *path = linitial(cheapest_pathlist);
904 : :
905 : : /*
906 : : * In the case where only one union child remains due to the
907 : : * detection of one or more dummy union children, obtain an
908 : : * estimate on the surviving child directly.
909 : : */
910 : 6 : dNumGroups = estimate_num_groups(root,
911 : 6 : path->pathtarget->exprs,
912 : : path->rows,
913 : : NULL,
914 : : NULL);
915 : : }
916 : : else
917 : : {
918 : : /*
919 : : * Otherwise, for the moment, take the number of distinct groups
920 : : * as equal to the total input size, i.e., the worst case. This
921 : : * is too conservative, but it's not clear how to get a decent
922 : : * estimate of the true size. One should note as well the
923 : : * propensity of novices to write UNION rather than UNION ALL even
924 : : * when they don't expect any duplicates...
925 : : */
926 : 1956 : dNumGroups = apath->rows;
927 : : }
928 : :
525 rhaas@postgresql.org 929 [ + + ]:CBC 1962 : if (can_hash)
930 : : {
931 : : Path *path;
932 : :
933 : : /*
934 : : * Try a hash aggregate plan on 'apath'. This is the cheapest
935 : : * available path containing each append child.
936 : : */
937 : 1926 : path = (Path *) create_agg_path(root,
938 : : result_rel,
939 : : apath,
29 drowley@postgresql.o 940 :GNC 1926 : result_rel->reltarget,
941 : : AGG_HASHED,
942 : : AGGSPLIT_SIMPLE,
943 : : groupList,
944 : : NIL,
945 : : NULL,
946 : : dNumGroups);
525 rhaas@postgresql.org 947 :CBC 1926 : add_path(result_rel, path);
948 : :
949 : : /* Try hash aggregate on the Gather path, if valid */
950 [ + + ]: 1926 : if (gpath != NULL)
951 : : {
952 : : /* Hashed aggregate plan --- no sort needed */
953 : 3 : path = (Path *) create_agg_path(root,
954 : : result_rel,
955 : : gpath,
29 drowley@postgresql.o 956 :GNC 3 : result_rel->reltarget,
957 : : AGG_HASHED,
958 : : AGGSPLIT_SIMPLE,
959 : : groupList,
960 : : NIL,
961 : : NULL,
962 : : dNumGroups);
525 rhaas@postgresql.org 963 :CBC 3 : add_path(result_rel, path);
964 : : }
965 : : }
966 : :
967 [ + + ]: 1962 : if (can_sort)
968 : : {
969 : 1911 : Path *path = apath;
970 : :
971 : : /* Try Sort -> Unique on the Append path */
972 [ + + ]: 1911 : if (groupList != NIL)
973 : 1896 : path = (Path *) create_sort_path(root, result_rel, path,
974 : : make_pathkeys_for_sortclauses(root, groupList, tlist),
975 : : -1.0);
976 : :
70 rguo@postgresql.org 977 :GNC 1911 : path = (Path *) create_unique_path(root,
978 : : result_rel,
979 : : path,
980 : 1911 : list_length(path->pathkeys),
981 : : dNumGroups);
982 : :
525 rhaas@postgresql.org 983 :CBC 1911 : add_path(result_rel, path);
984 : :
985 : : /* Try Sort -> Unique on the Gather path, if set */
986 [ + + ]: 1911 : if (gpath != NULL)
987 : : {
988 : 3 : path = gpath;
989 : :
990 : 3 : path = (Path *) create_sort_path(root, result_rel, path,
991 : : make_pathkeys_for_sortclauses(root, groupList, tlist),
992 : : -1.0);
993 : :
70 rguo@postgresql.org 994 :GNC 3 : path = (Path *) create_unique_path(root,
995 : : result_rel,
996 : : path,
997 : 3 : list_length(path->pathkeys),
998 : : dNumGroups);
525 rhaas@postgresql.org 999 :CBC 3 : add_path(result_rel, path);
1000 : : }
1001 : : }
1002 : :
1003 : : /*
1004 : : * Try making a MergeAppend path if we managed to find a path with the
1005 : : * correct pathkeys in each union child query.
1006 : : */
1007 [ + + + + ]: 1962 : if (try_sorted && groupList != NIL)
1008 : : {
1009 : : Path *path;
1010 : :
1011 : 137 : path = (Path *) create_merge_append_path(root,
1012 : : result_rel,
1013 : : ordered_pathlist,
1014 : : union_pathkeys,
1015 : : NULL);
1016 : :
1017 : : /* and make the MergeAppend unique */
70 rguo@postgresql.org 1018 :GNC 137 : path = (Path *) create_unique_path(root,
1019 : : result_rel,
1020 : : path,
1021 : : list_length(tlist),
1022 : : dNumGroups);
1023 : :
525 rhaas@postgresql.org 1024 :CBC 137 : add_path(result_rel, path);
1025 : : }
1026 : : }
1027 : : else
1028 : : {
1029 : : /* UNION ALL */
1030 : 373 : add_path(result_rel, apath);
1031 : :
1032 [ - + ]: 373 : if (gpath != NULL)
525 rhaas@postgresql.org 1033 :UBC 0 : add_path(result_rel, gpath);
1034 : : }
1035 : :
2780 rhaas@postgresql.org 1036 :CBC 2335 : return result_rel;
1037 : : }
1038 : :
1039 : : /*
1040 : : * Generate paths for an INTERSECT, INTERSECT ALL, EXCEPT, or EXCEPT ALL node
1041 : : */
1042 : : static RelOptInfo *
1043 : 346 : generate_nonunion_paths(SetOperationStmt *op, PlannerInfo *root,
1044 : : List *refnames_tlist,
1045 : : List **pTargetList)
1046 : : {
1047 : : RelOptInfo *result_rel;
1048 : : RelOptInfo *lrel,
1049 : : *rrel;
3522 tgl@sss.pgh.pa.us 1050 : 346 : double save_fraction = root->tuple_fraction;
1051 : : Path *lpath,
1052 : : *rpath,
1053 : : *path;
1054 : : List *lpath_tlist,
1055 : : *rpath_tlist,
1056 : : *tlist,
1057 : : *groupList;
1058 : : bool lpath_trivial_tlist,
1059 : : rpath_trivial_tlist,
1060 : : result_trivial_tlist;
313 1061 : 346 : List *nonunion_pathkeys = NIL;
1062 : : double dLeftGroups,
1063 : : dRightGroups,
1064 : : dNumGroups,
1065 : : dNumOutputRows;
1066 : : bool can_sort;
1067 : : bool can_hash;
1068 : : SetOpCmd cmd;
1069 : :
1070 : : /*
1071 : : * Tell children to fetch all tuples.
1072 : : */
3522 1073 : 346 : root->tuple_fraction = 0.0;
1074 : :
1075 : : /* Recurse on children */
2780 rhaas@postgresql.org 1076 : 346 : lrel = recurse_set_operations(op->larg, root,
1077 : : op,
1078 : : op->colTypes, op->colCollations,
1079 : : refnames_tlist,
1080 : : &lpath_tlist,
1081 : : &lpath_trivial_tlist);
1082 : :
1083 : 346 : rrel = recurse_set_operations(op->rarg, root,
1084 : : op,
1085 : : op->colTypes, op->colCollations,
1086 : : refnames_tlist,
1087 : : &rpath_tlist,
1088 : : &rpath_trivial_tlist);
1089 : :
1090 : : /*
1091 : : * Generate tlist for SetOp plan node.
1092 : : *
1093 : : * The tlist for a SetOp plan isn't important so far as the SetOp is
1094 : : * concerned, but we must make it look real anyway for the benefit of the
1095 : : * next plan level up.
1096 : : */
313 tgl@sss.pgh.pa.us 1097 : 346 : tlist = generate_setop_tlist(op->colTypes, op->colCollations,
1098 : : 0, false, lpath_tlist, refnames_tlist,
1099 : : &result_trivial_tlist);
1100 : :
1101 : : /* We should not have needed any type coercions in the tlist */
1102 [ - + ]: 346 : Assert(result_trivial_tlist);
1103 : :
1104 : 346 : *pTargetList = tlist;
1105 : :
1106 : : /* Identify the grouping semantics */
1107 : 346 : groupList = generate_setop_grouplist(op, tlist);
1108 : :
1109 : : /* Check whether the operators support sorting or hashing */
1110 : 346 : can_sort = grouping_is_sortable(groupList);
1111 : 346 : can_hash = grouping_is_hashable(groupList);
1112 [ - + - - ]: 346 : if (!can_sort && !can_hash)
313 tgl@sss.pgh.pa.us 1113 [ # # # # ]:UBC 0 : ereport(ERROR,
1114 : : (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1115 : : /* translator: %s is INTERSECT or EXCEPT */
1116 : : errmsg("could not implement %s",
1117 : : (op->op == SETOP_INTERSECT) ? "INTERSECT" : "EXCEPT"),
1118 : : errdetail("Some of the datatypes only support hashing, while others only support sorting.")));
1119 : :
313 tgl@sss.pgh.pa.us 1120 [ + - ]:CBC 346 : if (can_sort)
1121 : : {
1122 : : /* Determine the pathkeys for sorting by the whole target list */
1123 : 346 : nonunion_pathkeys = make_pathkeys_for_sortclauses(root, groupList,
1124 : : tlist);
1125 : :
1126 : 346 : root->query_pathkeys = nonunion_pathkeys;
1127 : : }
1128 : :
1129 : : /*
1130 : : * Now that we've got all that info, we can build the child paths.
1131 : : */
1132 [ + + ]: 346 : if (lrel->rtekind == RTE_SUBQUERY)
1133 : 334 : build_setop_child_paths(root, lrel, lpath_trivial_tlist, lpath_tlist,
1134 : : nonunion_pathkeys, &dLeftGroups);
1135 : : else
1136 : 12 : dLeftGroups = lrel->rows;
525 rhaas@postgresql.org 1137 [ + + ]: 346 : if (rrel->rtekind == RTE_SUBQUERY)
1138 : 343 : build_setop_child_paths(root, rrel, rpath_trivial_tlist, rpath_tlist,
1139 : : nonunion_pathkeys, &dRightGroups);
1140 : : else
1141 : 3 : dRightGroups = rrel->rows;
1142 : :
1143 : : /* Undo effects of forcing tuple_fraction to 0 */
3522 tgl@sss.pgh.pa.us 1144 : 346 : root->tuple_fraction = save_fraction;
1145 : :
1146 : : /*
1147 : : * For EXCEPT, we must put the left input first. For INTERSECT, either
1148 : : * order should give the same results, and we prefer to put the smaller
1149 : : * input first in order to (a) minimize the size of the hash table in the
1150 : : * hashing case, and (b) improve our chances of exploiting the executor's
1151 : : * fast path for empty left-hand input. "Smaller" means the one with the
1152 : : * fewer groups.
1153 : : */
313 1154 [ + + + + ]: 346 : if (op->op != SETOP_EXCEPT && dLeftGroups > dRightGroups)
1155 : : {
1156 : : /* need to swap the two inputs */
1157 : : RelOptInfo *tmprel;
1158 : : List *tmplist;
1159 : : double tmpd;
1160 : :
1161 : 18 : tmprel = lrel;
1162 : 18 : lrel = rrel;
1163 : 18 : rrel = tmprel;
1164 : 18 : tmplist = lpath_tlist;
1165 : 18 : lpath_tlist = rpath_tlist;
1166 : 18 : rpath_tlist = tmplist;
1167 : 18 : tmpd = dLeftGroups;
1168 : 18 : dLeftGroups = dRightGroups;
1169 : 18 : dRightGroups = tmpd;
1170 : : }
1171 : :
1172 : 346 : lpath = lrel->cheapest_total_path;
1173 : 346 : rpath = rrel->cheapest_total_path;
1174 : :
1175 : : /* Build result relation. */
2780 rhaas@postgresql.org 1176 : 346 : result_rel = fetch_upper_rel(root, UPPERREL_SETOP,
1177 : 346 : bms_union(lrel->relids, rrel->relids));
1178 : :
1179 : : /*
1180 : : * Create the PathTarget and set the width accordingly. For EXCEPT, since
1181 : : * the set op result won't contain rows from the rpath, we only account
1182 : : * for the width of the lpath. For INTERSECT, use both input paths.
1183 : : */
29 drowley@postgresql.o 1184 [ + + ]:GNC 346 : if (op->op == SETOP_EXCEPT)
1185 : 223 : result_rel->reltarget = create_setop_pathtarget(root, tlist,
1186 : 223 : list_make1(lpath));
1187 : : else
1188 : 123 : result_rel->reltarget = create_setop_pathtarget(root, tlist,
1189 : 123 : list_make2(lpath, rpath));
1190 : :
1191 : : /* Check for provably empty setop inputs and add short-circuit paths. */
21 1192 [ + + ]: 346 : if (op->op == SETOP_EXCEPT)
1193 : : {
1194 : : /*
1195 : : * For EXCEPTs, if the left side is dummy then there's no need to
1196 : : * inspect the right-hand side as scanning the right to find tuples to
1197 : : * remove won't make the left-hand input any more empty.
1198 : : */
1199 [ + + ]: 223 : if (is_dummy_rel(lrel))
1200 : : {
1201 : 3 : mark_dummy_rel(result_rel);
1202 : :
1203 : 3 : return result_rel;
1204 : : }
1205 : :
1206 : : /* Handle EXCEPTs with dummy right input */
1207 [ + + ]: 220 : if (is_dummy_rel(rrel))
1208 : : {
1209 [ + - ]: 3 : if (op->all)
1210 : : {
1211 : : Path *apath;
1212 : :
1213 : : /*
1214 : : * EXCEPT ALL: If the right-hand input is dummy then we can
1215 : : * simply scan the left-hand input. To keep createplan.c
1216 : : * happy, use a single child Append to handle the translation
1217 : : * between the set op targetlist and the targetlist of the
1218 : : * left input. The Append will be removed in setrefs.c.
1219 : : */
1220 : 3 : apath = (Path *) create_append_path(root, result_rel, list_make1(lpath),
1221 : : NIL, NIL, NULL, 0, false, -1);
1222 : :
1223 : 3 : add_path(result_rel, apath);
1224 : :
1225 : 3 : return result_rel;
1226 : : }
1227 : : else
1228 : : {
1229 : : /*
1230 : : * To make EXCEPT with a dummy RHS work means having to
1231 : : * deduplicate the left input. That could be done with
1232 : : * AggPaths, but it doesn't seem worth the effort. Let the
1233 : : * normal path generation code below handle this one.
1234 : : */
1235 : : }
1236 : : }
1237 : : }
1238 : : else
1239 : : {
1240 : : /*
1241 : : * For INTERSECT, if either input is a dummy rel then we can mark the
1242 : : * result_rel as dummy since intersecting with an empty relation can
1243 : : * never yield any results. This is true regardless of INTERSECT or
1244 : : * INTERSECT ALL.
1245 : : */
1246 [ + + - + ]: 123 : if (is_dummy_rel(lrel) || is_dummy_rel(rrel))
1247 : : {
1248 : 9 : mark_dummy_rel(result_rel);
1249 : :
1250 : 9 : return result_rel;
1251 : : }
1252 : : }
1253 : :
1254 : : /*
1255 : : * Estimate number of distinct groups that we'll need hashtable entries
1256 : : * for; this is the size of the left-hand input for EXCEPT, or the smaller
1257 : : * input for INTERSECT. Also estimate the number of eventual output rows.
1258 : : * In non-ALL cases, we estimate each group produces one output row; in
1259 : : * ALL cases use the relevant relation size. These are worst-case
1260 : : * estimates, of course, but we need to be conservative.
1261 : : */
6291 tgl@sss.pgh.pa.us 1262 [ + + ]:CBC 331 : if (op->op == SETOP_EXCEPT)
1263 : : {
1264 : 217 : dNumGroups = dLeftGroups;
3522 1265 [ + + ]: 217 : dNumOutputRows = op->all ? lpath->rows : dNumGroups;
1266 : : }
1267 : : else
1268 : : {
313 1269 : 114 : dNumGroups = dLeftGroups;
3522 1270 [ + + - + ]: 114 : dNumOutputRows = op->all ? Min(lpath->rows, rpath->rows) : dNumGroups;
1271 : : }
313 1272 : 331 : result_rel->rows = dNumOutputRows;
1273 : :
1274 : : /* Select the SetOpCmd type */
1275 [ + + - ]: 331 : switch (op->op)
1276 : : {
1277 : 114 : case SETOP_INTERSECT:
1278 : 114 : cmd = op->all ? SETOPCMD_INTERSECT_ALL : SETOPCMD_INTERSECT;
1279 : 114 : break;
1280 : 217 : case SETOP_EXCEPT:
1281 [ + + ]: 217 : cmd = op->all ? SETOPCMD_EXCEPT_ALL : SETOPCMD_EXCEPT;
1282 : 217 : break;
313 tgl@sss.pgh.pa.us 1283 :UBC 0 : default:
1284 [ # # ]: 0 : elog(ERROR, "unrecognized set op: %d", (int) op->op);
1285 : : cmd = SETOPCMD_INTERSECT; /* keep compiler quiet */
1286 : : break;
1287 : : }
1288 : :
1289 : : /*
1290 : : * If we can hash, that just requires a SetOp atop the cheapest inputs.
1291 : : */
313 tgl@sss.pgh.pa.us 1292 [ + + ]:CBC 331 : if (can_hash)
1293 : : {
1294 : 301 : path = (Path *) create_setop_path(root,
1295 : : result_rel,
1296 : : lpath,
1297 : : rpath,
1298 : : cmd,
1299 : : SETOP_HASHED,
1300 : : groupList,
1301 : : dNumGroups,
1302 : : dNumOutputRows);
1303 : 301 : add_path(result_rel, path);
1304 : : }
1305 : :
1306 : : /*
1307 : : * If we can sort, generate the cheapest sorted input paths, and add a
1308 : : * SetOp atop those.
1309 : : */
1310 [ + - ]: 331 : if (can_sort)
1311 : : {
1312 : : List *pathkeys;
1313 : : Path *slpath,
1314 : : *srpath;
1315 : :
1316 : : /* First the left input ... */
1317 : 331 : pathkeys = make_pathkeys_for_sortclauses(root,
1318 : : groupList,
1319 : : lpath_tlist);
1320 [ + + ]: 331 : if (pathkeys_contained_in(pathkeys, lpath->pathkeys))
1321 : 48 : slpath = lpath; /* cheapest path is already sorted */
1322 : : else
1323 : : {
1324 : 283 : slpath = get_cheapest_path_for_pathkeys(lrel->pathlist,
1325 : : nonunion_pathkeys,
1326 : : NULL,
1327 : : TOTAL_COST,
1328 : : false);
1329 : : /* Subquery failed to produce any presorted paths? */
1330 [ + + ]: 283 : if (slpath == NULL)
1331 : 72 : slpath = (Path *) create_sort_path(root,
1332 : : lpath->parent,
1333 : : lpath,
1334 : : pathkeys,
1335 : : -1.0);
1336 : : }
1337 : :
1338 : : /* and now the same for the right. */
1339 : 331 : pathkeys = make_pathkeys_for_sortclauses(root,
1340 : : groupList,
1341 : : rpath_tlist);
1342 [ + + ]: 331 : if (pathkeys_contained_in(pathkeys, rpath->pathkeys))
1343 : 54 : srpath = rpath; /* cheapest path is already sorted */
1344 : : else
1345 : : {
1346 : 277 : srpath = get_cheapest_path_for_pathkeys(rrel->pathlist,
1347 : : nonunion_pathkeys,
1348 : : NULL,
1349 : : TOTAL_COST,
1350 : : false);
1351 : : /* Subquery failed to produce any presorted paths? */
1352 [ + + ]: 277 : if (srpath == NULL)
1353 : 75 : srpath = (Path *) create_sort_path(root,
1354 : : rpath->parent,
1355 : : rpath,
1356 : : pathkeys,
1357 : : -1.0);
1358 : : }
1359 : :
1360 : 331 : path = (Path *) create_setop_path(root,
1361 : : result_rel,
1362 : : slpath,
1363 : : srpath,
1364 : : cmd,
1365 : : SETOP_SORTED,
1366 : : groupList,
1367 : : dNumGroups,
1368 : : dNumOutputRows);
1369 : 331 : add_path(result_rel, path);
1370 : : }
1371 : :
2780 rhaas@postgresql.org 1372 : 331 : return result_rel;
1373 : : }
1374 : :
1375 : : /*
1376 : : * Pull up children of a UNION node that are identically-propertied UNIONs,
1377 : : * and perform planning of the queries underneath the N-way UNION.
1378 : : *
1379 : : * The result is a list of RelOptInfos containing Paths for sub-nodes, with
1380 : : * one entry for each descendant that is a leaf query or non-identical setop.
1381 : : * We also return parallel lists of the childrens' targetlists and
1382 : : * is-trivial-tlist flags.
1383 : : *
1384 : : * NOTE: we can also pull a UNION ALL up into a UNION, since the distinct
1385 : : * output rows will be lost anyway.
1386 : : */
1387 : : static List *
1388 : 2338 : plan_union_children(PlannerInfo *root,
1389 : : SetOperationStmt *top_union,
1390 : : List *refnames_tlist,
1391 : : List **tlist_list,
1392 : : List **istrivial_tlist)
1393 : : {
1394 : 2338 : List *pending_rels = list_make1(top_union);
1395 : 2338 : List *result = NIL;
1396 : : List *child_tlist;
1397 : : bool trivial_tlist;
1398 : :
1399 : 2338 : *tlist_list = NIL;
525 1400 : 2338 : *istrivial_tlist = NIL;
1401 : :
2780 1402 [ + + ]: 13050 : while (pending_rels != NIL)
1403 : : {
1404 : 10712 : Node *setOp = linitial(pending_rels);
1405 : :
1406 : 10712 : pending_rels = list_delete_first(pending_rels);
1407 : :
1408 [ + + ]: 10712 : if (IsA(setOp, SetOperationStmt))
1409 : : {
1410 : 4247 : SetOperationStmt *op = (SetOperationStmt *) setOp;
1411 : :
1412 [ + + ]: 4247 : if (op->op == top_union->op &&
1413 [ + + + + : 8395 : (op->all == top_union->all || op->all) &&
+ + ]
343 tgl@sss.pgh.pa.us 1414 [ + - ]: 8380 : equal(op->colTypes, top_union->colTypes) &&
1415 : 4187 : equal(op->colCollations, top_union->colCollations))
1416 : : {
1417 : : /* Same UNION, so fold children into parent */
2780 rhaas@postgresql.org 1418 : 4187 : pending_rels = lcons(op->rarg, pending_rels);
1419 : 4187 : pending_rels = lcons(op->larg, pending_rels);
1420 : 4187 : continue;
1421 : : }
1422 : : }
1423 : :
1424 : : /*
1425 : : * Not same, so plan this child separately.
1426 : : *
1427 : : * If top_union isn't a UNION ALL, then we are interested in sorted
1428 : : * output from the child, so pass top_union as parentOp. Note that
1429 : : * this isn't necessarily the child node's immediate SetOperationStmt
1430 : : * parent, but that's fine: it's the effective parent.
1431 : : */
1432 : 6525 : result = lappend(result, recurse_set_operations(setOp, root,
313 tgl@sss.pgh.pa.us 1433 [ + + ]: 6525 : top_union->all ? NULL : top_union,
1434 : : top_union->colTypes,
1435 : : top_union->colCollations,
1436 : : refnames_tlist,
1437 : : &child_tlist,
1438 : : &trivial_tlist));
2780 rhaas@postgresql.org 1439 : 6525 : *tlist_list = lappend(*tlist_list, child_tlist);
525 1440 : 6525 : *istrivial_tlist = lappend_int(*istrivial_tlist, trivial_tlist);
1441 : : }
1442 : :
3522 tgl@sss.pgh.pa.us 1443 : 2338 : return result;
1444 : : }
1445 : :
1446 : : /*
1447 : : * postprocess_setop_rel - perform steps required after adding paths
1448 : : */
1449 : : static void
2780 rhaas@postgresql.org 1450 : 11230 : postprocess_setop_rel(PlannerInfo *root, RelOptInfo *rel)
1451 : : {
1452 : : /*
1453 : : * We don't currently worry about allowing FDWs to contribute paths to
1454 : : * this relation, but give extensions a chance.
1455 : : */
1456 [ - + ]: 11230 : if (create_upper_paths_hook)
2780 rhaas@postgresql.org 1457 :UBC 0 : (*create_upper_paths_hook) (root, UPPERREL_SETOP,
1458 : : NULL, rel, NULL);
1459 : :
1460 : : /* Select cheapest path */
2780 rhaas@postgresql.org 1461 :CBC 11230 : set_cheapest(rel);
1462 : 11230 : }
1463 : :
1464 : : /*
1465 : : * Generate targetlist for a set-operation plan node
1466 : : *
1467 : : * colTypes: OID list of set-op's result column datatypes
1468 : : * colCollations: OID list of set-op's result column collations
1469 : : * varno: varno to use in generated Vars
1470 : : * hack_constants: true to copy up constants (see comments in code)
1471 : : * input_tlist: targetlist of this node's input node
1472 : : * refnames_tlist: targetlist to take column names from
1473 : : * trivial_tlist: output parameter, set to true if targetlist is trivial
1474 : : */
1475 : : static List *
5309 tgl@sss.pgh.pa.us 1476 : 8431 : generate_setop_tlist(List *colTypes, List *colCollations,
1477 : : Index varno,
1478 : : bool hack_constants,
1479 : : List *input_tlist,
1480 : : List *refnames_tlist,
1481 : : bool *trivial_tlist)
1482 : : {
9154 1483 : 8431 : List *tlist = NIL;
1484 : 8431 : int resno = 1;
1485 : : ListCell *ctlc,
1486 : : *cclc,
1487 : : *itlc,
1488 : : *rtlc;
1489 : : TargetEntry *tle;
1490 : : Node *expr;
1491 : :
1197 1492 : 8431 : *trivial_tlist = true; /* until proven differently */
1493 : :
2434 1494 [ + + + + : 33523 : forfour(ctlc, colTypes, cclc, colCollations,
+ + + + +
+ + + + +
+ + + + +
- + - + -
+ + ]
1495 : : itlc, input_tlist, rtlc, refnames_tlist)
1496 : : {
5309 1497 : 25092 : Oid colType = lfirst_oid(ctlc);
1498 : 25092 : Oid colColl = lfirst_oid(cclc);
1499 : 25092 : TargetEntry *inputtle = (TargetEntry *) lfirst(itlc);
1500 : 25092 : TargetEntry *reftle = (TargetEntry *) lfirst(rtlc);
1501 : :
7510 1502 [ - + ]: 25092 : Assert(inputtle->resno == resno);
1503 [ - + ]: 25092 : Assert(reftle->resno == resno);
1504 [ - + ]: 25092 : Assert(!inputtle->resjunk);
1505 [ - + ]: 25092 : Assert(!reftle->resjunk);
1506 : :
1507 : : /*
1508 : : * Generate columns referencing input columns and having appropriate
1509 : : * data types and column names. Insert datatype coercions where
1510 : : * necessary.
1511 : : *
1512 : : * HACK: constants in the input's targetlist are copied up as-is
1513 : : * rather than being referenced as subquery outputs. This is mainly
1514 : : * to ensure that when we try to coerce them to the output column's
1515 : : * datatype, the right things happen for UNKNOWN constants. But do
1516 : : * this only at the first level of subquery-scan plans; we don't want
1517 : : * phony constants appearing in the output tlists of upper-level
1518 : : * nodes!
1519 : : *
1520 : : * Note that copying a constant doesn't in itself require us to mark
1521 : : * the tlist nontrivial; see trivial_subqueryscan() in setrefs.c.
1522 : : */
9119 1523 [ + + + - : 25092 : if (hack_constants && inputtle->expr && IsA(inputtle->expr, Const))
+ + ]
8356 1524 : 8010 : expr = (Node *) inputtle->expr;
1525 : : else
7464 1526 : 68328 : expr = (Node *) makeVar(varno,
7510 1527 : 17082 : inputtle->resno,
1528 : 17082 : exprType((Node *) inputtle->expr),
1529 : 17082 : exprTypmod((Node *) inputtle->expr),
5376 peter_e@gmx.net 1530 : 17082 : exprCollation((Node *) inputtle->expr),
1531 : : 0);
1532 : :
7510 tgl@sss.pgh.pa.us 1533 [ + + ]: 25092 : if (exprType(expr) != colType)
1534 : : {
1535 : : /*
1536 : : * Note: it's not really cool to be applying coerce_to_common_type
1537 : : * here; one notable point is that assign_expr_collations never
1538 : : * gets run on any generated nodes. For the moment that's not a
1539 : : * problem because we force the correct exposed collation below.
1540 : : * It would likely be best to make the parser generate the correct
1541 : : * output tlist for every set-op to begin with, though.
1542 : : */
8121 bruce@momjian.us 1543 : 847 : expr = coerce_to_common_type(NULL, /* no UNKNOWNs here */
1544 : : expr,
1545 : : colType,
1546 : : "UNION/INTERSECT/EXCEPT");
1197 tgl@sss.pgh.pa.us 1547 : 847 : *trivial_tlist = false; /* the coercion makes it not trivial */
1548 : : }
1549 : :
1550 : : /*
1551 : : * Ensure the tlist entry's exposed collation matches the set-op. This
1552 : : * is necessary because plan_set_operations() reports the result
1553 : : * ordering as a list of SortGroupClauses, which don't carry collation
1554 : : * themselves but just refer to tlist entries. If we don't show the
1555 : : * right collation then planner.c might do the wrong thing in
1556 : : * higher-level queries.
1557 : : *
1558 : : * Note we use RelabelType, not CollateExpr, since this expression
1559 : : * will reach the executor without any further processing.
1560 : : */
5309 1561 [ + + ]: 25092 : if (exprCollation(expr) != colColl)
1562 : : {
1896 1563 : 6571 : expr = applyRelabelType(expr,
1564 : : exprType(expr), exprTypmod(expr), colColl,
1565 : : COERCE_IMPLICIT_CAST, -1, false);
1197 1566 : 6571 : *trivial_tlist = false; /* the relabel makes it not trivial */
1567 : : }
1568 : :
7510 1569 : 50184 : tle = makeTargetEntry((Expr *) expr,
1570 : 25092 : (AttrNumber) resno++,
1571 : 25092 : pstrdup(reftle->resname),
1572 : : false);
1573 : :
1574 : : /*
1575 : : * By convention, all output columns in a setop tree have
1576 : : * ressortgroupref equal to their resno. In some cases the ref isn't
1577 : : * needed, but this is a cleaner way than modifying the tlist later.
1578 : : */
3522 1579 : 25092 : tle->ressortgroupref = tle->resno;
1580 : :
7510 1581 : 25092 : tlist = lappend(tlist, tle);
1582 : : }
1583 : :
9154 1584 : 8431 : return tlist;
1585 : : }
1586 : :
1587 : : /*
1588 : : * Generate targetlist for a set-operation Append node
1589 : : *
1590 : : * colTypes: OID list of set-op's result column datatypes
1591 : : * colCollations: OID list of set-op's result column collations
1592 : : * input_tlists: list of tlists for sub-plans of the Append
1593 : : * refnames_tlist: targetlist to take column names from
1594 : : *
1595 : : * The entries in the Append's targetlist should always be simple Vars;
1596 : : * we just have to make sure they have the right datatypes/typmods/collations.
1597 : : * The Vars are always generated with varno 0.
1598 : : *
1599 : : * XXX a problem with the varno-zero approach is that set_pathtarget_cost_width
1600 : : * cannot figure out a realistic width for the tlist we make here. But we
1601 : : * ought to refactor this code to produce a PathTarget directly, anyway.
1602 : : */
1603 : : static List *
5309 1604 : 2808 : generate_append_tlist(List *colTypes, List *colCollations,
1605 : : List *input_tlists,
1606 : : List *refnames_tlist)
1607 : : {
8638 1608 : 2808 : List *tlist = NIL;
1609 : 2808 : int resno = 1;
1610 : : ListCell *curColType;
1611 : : ListCell *curColCollation;
1612 : : ListCell *ref_tl_item;
1613 : : int colindex;
1614 : : TargetEntry *tle;
1615 : : Node *expr;
1616 : : ListCell *tlistl;
1617 : : int32 *colTypmods;
1618 : :
1619 : : /*
1620 : : * First extract typmods to use.
1621 : : *
1622 : : * If the inputs all agree on type and typmod of a particular column, use
1623 : : * that typmod; else use -1.
1624 : : */
7815 1625 : 2808 : colTypmods = (int32 *) palloc(list_length(colTypes) * sizeof(int32));
1626 : :
3522 1627 [ + - + + : 10273 : foreach(tlistl, input_tlists)
+ + ]
1628 : : {
1629 : 7465 : List *subtlist = (List *) lfirst(tlistl);
1630 : : ListCell *subtlistl;
1631 : :
7825 neilc@samurai.com 1632 : 7465 : curColType = list_head(colTypes);
8638 tgl@sss.pgh.pa.us 1633 : 7465 : colindex = 0;
3522 1634 [ + + + + : 29308 : foreach(subtlistl, subtlist)
+ + ]
1635 : : {
1636 : 21843 : TargetEntry *subtle = (TargetEntry *) lfirst(subtlistl);
1637 : :
313 1638 [ - + ]: 21843 : Assert(!subtle->resjunk);
7825 neilc@samurai.com 1639 [ - + ]: 21843 : Assert(curColType != NULL);
7510 tgl@sss.pgh.pa.us 1640 [ + - ]: 21843 : if (exprType((Node *) subtle->expr) == lfirst_oid(curColType))
1641 : : {
1642 : : /* If first subplan, copy the typmod; else compare */
1643 : 21843 : int32 subtypmod = exprTypmod((Node *) subtle->expr);
1644 : :
3522 1645 [ + + ]: 21843 : if (tlistl == list_head(input_tlists))
7510 1646 : 7743 : colTypmods[colindex] = subtypmod;
1647 [ + + ]: 14100 : else if (subtypmod != colTypmods[colindex])
8638 1648 : 6 : colTypmods[colindex] = -1;
1649 : : }
1650 : : else
1651 : : {
1652 : : /* types disagree, so force typmod to -1 */
8638 tgl@sss.pgh.pa.us 1653 :UBC 0 : colTypmods[colindex] = -1;
1654 : : }
2297 tgl@sss.pgh.pa.us 1655 :CBC 21843 : curColType = lnext(colTypes, curColType);
8638 1656 : 21843 : colindex++;
1657 : : }
7825 neilc@samurai.com 1658 [ - + ]: 7465 : Assert(curColType == NULL);
1659 : : }
1660 : :
1661 : : /*
1662 : : * Now we can build the tlist for the Append.
1663 : : */
8638 tgl@sss.pgh.pa.us 1664 : 2808 : colindex = 0;
5309 1665 [ + + + + : 10551 : forthree(curColType, colTypes, curColCollation, colCollations,
+ + + + +
+ + + + +
+ - + - +
+ ]
1666 : : ref_tl_item, refnames_tlist)
1667 : : {
7821 neilc@samurai.com 1668 : 7743 : Oid colType = lfirst_oid(curColType);
8638 tgl@sss.pgh.pa.us 1669 : 7743 : int32 colTypmod = colTypmods[colindex++];
5376 peter_e@gmx.net 1670 : 7743 : Oid colColl = lfirst_oid(curColCollation);
7825 neilc@samurai.com 1671 : 7743 : TargetEntry *reftle = (TargetEntry *) lfirst(ref_tl_item);
1672 : :
7510 tgl@sss.pgh.pa.us 1673 [ - + ]: 7743 : Assert(reftle->resno == resno);
1674 [ - + ]: 7743 : Assert(!reftle->resjunk);
8638 1675 : 7743 : expr = (Node *) makeVar(0,
1676 : : resno,
1677 : : colType,
1678 : : colTypmod,
1679 : : colColl,
1680 : : 0);
7510 1681 : 15486 : tle = makeTargetEntry((Expr *) expr,
1682 : 7743 : (AttrNumber) resno++,
1683 : 7743 : pstrdup(reftle->resname),
1684 : : false);
1685 : :
1686 : : /*
1687 : : * By convention, all output columns in a setop tree have
1688 : : * ressortgroupref equal to their resno. In some cases the ref isn't
1689 : : * needed, but this is a cleaner way than modifying the tlist later.
1690 : : */
3522 1691 : 7743 : tle->ressortgroupref = tle->resno;
1692 : :
7510 1693 : 7743 : tlist = lappend(tlist, tle);
1694 : : }
1695 : :
8638 1696 : 2808 : pfree(colTypmods);
1697 : :
1698 : 2808 : return tlist;
1699 : : }
1700 : :
1701 : : /*
1702 : : * generate_setop_grouplist
1703 : : * Build a SortGroupClause list defining the sort/grouping properties
1704 : : * of the setop's output columns.
1705 : : *
1706 : : * Parse analysis already determined the properties and built a suitable
1707 : : * list, except that the entries do not have sortgrouprefs set because
1708 : : * the parser output representation doesn't include a tlist for each
1709 : : * setop. So what we need to do here is copy that list and install
1710 : : * proper sortgrouprefs into it (copying those from the targetlist).
1711 : : */
1712 : : static List *
6291 1713 : 2507 : generate_setop_grouplist(SetOperationStmt *op, List *targetlist)
1714 : : {
3155 peter_e@gmx.net 1715 : 2507 : List *grouplist = copyObject(op->groupClauses);
1716 : : ListCell *lg;
1717 : : ListCell *lt;
1718 : :
6291 tgl@sss.pgh.pa.us 1719 : 2507 : lg = list_head(grouplist);
1720 [ + + + + : 9523 : foreach(lt, targetlist)
+ + ]
1721 : : {
1722 : 7016 : TargetEntry *tle = (TargetEntry *) lfirst(lt);
1723 : : SortGroupClause *sgc;
1724 : :
313 1725 [ - + ]: 7016 : Assert(!tle->resjunk);
1726 : :
1727 : : /* non-resjunk columns should have sortgroupref = resno */
3522 1728 [ - + ]: 7016 : Assert(tle->ressortgroupref == tle->resno);
1729 : :
1730 : : /* non-resjunk columns should have grouping clauses */
6291 1731 [ - + ]: 7016 : Assert(lg != NULL);
1732 : 7016 : sgc = (SortGroupClause *) lfirst(lg);
2297 1733 : 7016 : lg = lnext(grouplist, lg);
6291 1734 [ - + ]: 7016 : Assert(sgc->tleSortGroupRef == 0);
1735 : :
3522 1736 : 7016 : sgc->tleSortGroupRef = tle->ressortgroupref;
1737 : : }
6291 1738 [ - + ]: 2507 : Assert(lg == NULL);
1739 : 2507 : return grouplist;
1740 : : }
1741 : :
1742 : : /*
1743 : : * create_setop_pathtarget
1744 : : * Do the normal create_pathtarget() work, plus set the resulting
1745 : : * PathTarget's width to the average width of the Paths in child_pathlist
1746 : : * weighted using the estimated row count of each path.
1747 : : *
1748 : : * Note: This is required because set op target lists use varno==0, which
1749 : : * results in a type default width estimate rather than one that's based on
1750 : : * statistics of the columns from the set op children.
1751 : : */
1752 : : static PathTarget *
29 drowley@postgresql.o 1753 :GNC 2684 : create_setop_pathtarget(PlannerInfo *root, List *tlist, List *child_pathlist)
1754 : : {
1755 : : PathTarget *reltarget;
1756 : : ListCell *lc;
1757 : 2684 : double parent_rows = 0;
1758 : 2684 : double parent_size = 0;
1759 : :
1760 : 2684 : reltarget = create_pathtarget(root, tlist);
1761 : :
1762 : : /* Calculate the total rows and total size. */
1763 [ + + + + : 9651 : foreach(lc, child_pathlist)
+ + ]
1764 : : {
1765 : 6967 : Path *path = (Path *) lfirst(lc);
1766 : :
1767 : 6967 : parent_rows += path->rows;
1768 : 6967 : parent_size += path->parent->reltarget->width * path->rows;
1769 : : }
1770 : :
1771 [ + + ]: 2684 : if (parent_rows > 0)
1772 : 2675 : reltarget->width = rint(parent_size / parent_rows);
1773 : :
1774 : 2684 : return reltarget;
1775 : : }
|
