granicus.if.org Git - postgresql/blob - src/backend/optimizer/plan/planner.c

   1 /*-------------------------------------------------------------------------
   2  *
   3  * planner.c
   4  *        The query optimizer external interface.
   5  *
   6  * Portions Copyright (c) 1996-2002, PostgreSQL Global Development Group
   7  * Portions Copyright (c) 1994, Regents of the University of California
   8  *
   9  *
  10  * IDENTIFICATION
  11  *        $Header: /cvsroot/pgsql/src/backend/optimizer/plan/planner.c,v 1.151 2003/03/10 03:53:50 tgl Exp $
  12  *
  13  *-------------------------------------------------------------------------
  14  */
  15
  16 #include "postgres.h"
  17
  18 #include <limits.h>
  19
  20 #include "catalog/pg_operator.h"
  21 #include "catalog/pg_type.h"
  22 #include "executor/executor.h"
  23 #include "miscadmin.h"
  24 #include "nodes/makefuncs.h"
  25 #ifdef OPTIMIZER_DEBUG
  26 #include "nodes/print.h"
  27 #endif
  28 #include "optimizer/clauses.h"
  29 #include "optimizer/cost.h"
  30 #include "optimizer/pathnode.h"
  31 #include "optimizer/paths.h"
  32 #include "optimizer/planmain.h"
  33 #include "optimizer/planner.h"
  34 #include "optimizer/prep.h"
  35 #include "optimizer/subselect.h"
  36 #include "optimizer/tlist.h"
  37 #include "optimizer/var.h"
  38 #include "parser/analyze.h"
  39 #include "parser/parsetree.h"
  40 #include "parser/parse_expr.h"
  41 #include "parser/parse_oper.h"
  42 #include "utils/selfuncs.h"
  43 #include "utils/syscache.h"
  44
  45
  46 /* Expression kind codes for preprocess_expression */
  47 #define EXPRKIND_QUAL   0
  48 #define EXPRKIND_TARGET 1
  49 #define EXPRKIND_RTFUNC 2
  50 #define EXPRKIND_ININFO 3
  51
  52
  53 static Node *preprocess_expression(Query *parse, Node *expr, int kind);
  54 static void preprocess_qual_conditions(Query *parse, Node *jtnode);
  55 static Plan *inheritance_planner(Query *parse, List *inheritlist);
  56 static Plan *grouping_planner(Query *parse, double tuple_fraction);
  57 static bool hash_safe_grouping(Query *parse);
  58 static List *make_subplanTargetList(Query *parse, List *tlist,
  59                                            AttrNumber **groupColIdx, bool *need_tlist_eval);
  60 static void locate_grouping_columns(Query *parse,
  61                                                                         List *tlist,
  62                                                                         List *sub_tlist,
  63                                                                         AttrNumber *groupColIdx);
  64 static Plan *make_groupsortplan(Query *parse,
  65                                                                 List *groupClause,
  66                                                                 AttrNumber *grpColIdx,
  67                                                                 Plan *subplan);
  68 static List *postprocess_setop_tlist(List *new_tlist, List *orig_tlist);
  69
  70
  71 /*****************************************************************************
  72  *
  73  *         Query optimizer entry point
  74  *
  75  *****************************************************************************/
  76 Plan *
  77 planner(Query *parse, bool isCursor, int cursorOptions)
  78 {
  79         double          tuple_fraction;
  80         Plan       *result_plan;
  81         Index           save_PlannerQueryLevel;
  82         List       *save_PlannerParamVar;
  83
  84         /*
  85          * The planner can be called recursively (an example is when
  86          * eval_const_expressions tries to pre-evaluate an SQL function). So,
  87          * these global state variables must be saved and restored.
  88          *
  89          * These vars cannot be moved into the Query structure since their whole
  90          * purpose is communication across multiple sub-Queries.
  91          *
  92          * Note we do NOT save and restore PlannerPlanId: it exists to assign
  93          * unique IDs to SubPlan nodes, and we want those IDs to be unique for
  94          * the life of a backend.  Also, PlannerInitPlan is saved/restored in
  95          * subquery_planner, not here.
  96          */
  97         save_PlannerQueryLevel = PlannerQueryLevel;
  98         save_PlannerParamVar = PlannerParamVar;
  99
 100         /* Initialize state for handling outer-level references and params */
 101         PlannerQueryLevel = 0;          /* will be 1 in top-level subquery_planner */
 102         PlannerParamVar = NIL;
 103
 104         /* Determine what fraction of the plan is likely to be scanned */
 105         if (isCursor)
 106         {
 107                 /*
 108                  * We have no real idea how many tuples the user will ultimately
 109                  * FETCH from a cursor, but it seems a good bet that he
 110                  * doesn't want 'em all.  Optimize for 10% retrieval (you
 111                  * gotta better number?  Should this be a SETtable parameter?)
 112                  */
 113                 tuple_fraction = 0.10;
 114         }
 115         else
 116         {
 117                 /* Default assumption is we need all the tuples */
 118                 tuple_fraction = 0.0;
 119         }
 120
 121         /* primary planning entry point (may recurse for subqueries) */
 122         result_plan = subquery_planner(parse, tuple_fraction);
 123
 124         Assert(PlannerQueryLevel == 0);
 125
 126         /*
 127          * If creating a plan for a scrollable cursor, make sure it can
 128          * run backwards on demand.  Add a Material node at the top at need.
 129          */
 130         if (isCursor && (cursorOptions & CURSOR_OPT_SCROLL))
 131         {
 132                 if (!ExecSupportsBackwardScan(result_plan))
 133                         result_plan = materialize_finished_plan(result_plan);
 134         }
 135
 136         /* executor wants to know total number of Params used overall */
 137         result_plan->nParamExec = length(PlannerParamVar);
 138
 139         /* final cleanup of the plan */
 140         set_plan_references(result_plan, parse->rtable);
 141
 142         /* restore state for outer planner, if any */
 143         PlannerQueryLevel = save_PlannerQueryLevel;
 144         PlannerParamVar = save_PlannerParamVar;
 145
 146         return result_plan;
 147 }
 148
 149
 150 /*--------------------
 151  * subquery_planner
 152  *        Invokes the planner on a subquery.  We recurse to here for each
 153  *        sub-SELECT found in the query tree.
 154  *
 155  * parse is the querytree produced by the parser & rewriter.
 156  * tuple_fraction is the fraction of tuples we expect will be retrieved.
 157  * tuple_fraction is interpreted as explained for grouping_planner, below.
 158  *
 159  * Basically, this routine does the stuff that should only be done once
 160  * per Query object.  It then calls grouping_planner.  At one time,
 161  * grouping_planner could be invoked recursively on the same Query object;
 162  * that's not currently true, but we keep the separation between the two
 163  * routines anyway, in case we need it again someday.
 164  *
 165  * subquery_planner will be called recursively to handle sub-Query nodes
 166  * found within the query's expressions and rangetable.
 167  *
 168  * Returns a query plan.
 169  *--------------------
 170  */
 171 Plan *
 172 subquery_planner(Query *parse, double tuple_fraction)
 173 {
 174         List       *saved_initplan = PlannerInitPlan;
 175         int                     saved_planid = PlannerPlanId;
 176         bool            hasOuterJoins;
 177         Plan       *plan;
 178         List       *newHaving;
 179         List       *lst;
 180
 181         /* Set up for a new level of subquery */
 182         PlannerQueryLevel++;
 183         PlannerInitPlan = NIL;
 184
 185         /*
 186          * Look for IN clauses at the top level of WHERE, and transform them
 187          * into joins.  Note that this step only handles IN clauses originally
 188          * at top level of WHERE; if we pull up any subqueries in the next step,
 189          * their INs are processed just before pulling them up.
 190          */
 191         parse->in_info_list = NIL;
 192         if (parse->hasSubLinks)
 193                 parse->jointree->quals = pull_up_IN_clauses(parse,
 194                                                                                                         parse->jointree->quals);
 195
 196         /*
 197          * Check to see if any subqueries in the rangetable can be merged into
 198          * this query.
 199          */
 200         parse->jointree = (FromExpr *)
 201                 pull_up_subqueries(parse, (Node *) parse->jointree, false);
 202
 203         /*
 204          * Detect whether any rangetable entries are RTE_JOIN kind; if not,
 205          * we can avoid the expense of doing flatten_join_alias_vars().  Also
 206          * check for outer joins --- if none, we can skip reduce_outer_joins().
 207          * This must be done after we have done pull_up_subqueries, of course.
 208          */
 209         parse->hasJoinRTEs = false;
 210         hasOuterJoins = false;
 211         foreach(lst, parse->rtable)
 212         {
 213                 RangeTblEntry *rte = (RangeTblEntry *) lfirst(lst);
 214
 215                 if (rte->rtekind == RTE_JOIN)
 216                 {
 217                         parse->hasJoinRTEs = true;
 218                         if (IS_OUTER_JOIN(rte->jointype))
 219                         {
 220                                 hasOuterJoins = true;
 221                                 /* Can quit scanning once we find an outer join */
 222                                 break;
 223                         }
 224                 }
 225         }
 226
 227         /*
 228          * Do expression preprocessing on targetlist and quals.
 229          */
 230         parse->targetList = (List *)
 231                 preprocess_expression(parse, (Node *) parse->targetList,
 232                                                           EXPRKIND_TARGET);
 233
 234         preprocess_qual_conditions(parse, (Node *) parse->jointree);
 235
 236         parse->havingQual = preprocess_expression(parse, parse->havingQual,
 237                                                                                           EXPRKIND_QUAL);
 238
 239         parse->in_info_list = (List *)
 240                 preprocess_expression(parse, (Node *) parse->in_info_list,
 241                                                           EXPRKIND_ININFO);
 242
 243         /* Also need to preprocess expressions for function RTEs */
 244         foreach(lst, parse->rtable)
 245         {
 246                 RangeTblEntry *rte = (RangeTblEntry *) lfirst(lst);
 247
 248                 if (rte->rtekind == RTE_FUNCTION)
 249                         rte->funcexpr = preprocess_expression(parse, rte->funcexpr,
 250                                                                                                   EXPRKIND_RTFUNC);
 251         }
 252
 253         /*
 254          * A HAVING clause without aggregates is equivalent to a WHERE clause
 255          * (except it can only refer to grouped fields).  Transfer any
 256          * agg-free clauses of the HAVING qual into WHERE.      This may seem like
 257          * wasting cycles to cater to stupidly-written queries, but there are
 258          * other reasons for doing it.  Firstly, if the query contains no aggs
 259          * at all, then we aren't going to generate an Agg plan node, and so
 260          * there'll be no place to execute HAVING conditions; without this
 261          * transfer, we'd lose the HAVING condition entirely, which is wrong.
 262          * Secondly, when we push down a qual condition into a sub-query, it's
 263          * easiest to push the qual into HAVING always, in case it contains
 264          * aggs, and then let this code sort it out.
 265          *
 266          * Note that both havingQual and parse->jointree->quals are in
 267          * implicitly-ANDed-list form at this point, even though they are
 268          * declared as Node *.  Also note that contain_agg_clause does not
 269          * recurse into sub-selects, which is exactly what we need here.
 270          */
 271         newHaving = NIL;
 272         foreach(lst, (List *) parse->havingQual)
 273         {
 274                 Node       *havingclause = (Node *) lfirst(lst);
 275
 276                 if (contain_agg_clause(havingclause))
 277                         newHaving = lappend(newHaving, havingclause);
 278                 else
 279                         parse->jointree->quals = (Node *)
 280                                 lappend((List *) parse->jointree->quals, havingclause);
 281         }
 282         parse->havingQual = (Node *) newHaving;
 283
 284         /*
 285          * If we have any outer joins, try to reduce them to plain inner joins.
 286          * This step is most easily done after we've done expression preprocessing.
 287          */
 288         if (hasOuterJoins)
 289                 reduce_outer_joins(parse);
 290
 291         /*
 292          * See if we can simplify the jointree; opportunities for this may come
 293          * from having pulled up subqueries, or from flattening explicit JOIN
 294          * syntax.  We must do this after flattening JOIN alias variables, since
 295          * eliminating explicit JOIN nodes from the jointree will cause
 296          * get_relids_for_join() to fail.  But it should happen after
 297          * reduce_outer_joins, anyway.
 298          */
 299         parse->jointree = (FromExpr *)
 300                 simplify_jointree(parse, (Node *) parse->jointree);
 301
 302         /*
 303          * Do the main planning.  If we have an inherited target relation,
 304          * that needs special processing, else go straight to
 305          * grouping_planner.
 306          */
 307         if (parse->resultRelation &&
 308                 (lst = expand_inherited_rtentry(parse, parse->resultRelation,
 309                                                                                 false)) != NIL)
 310                 plan = inheritance_planner(parse, lst);
 311         else
 312                 plan = grouping_planner(parse, tuple_fraction);
 313
 314         /*
 315          * If any subplans were generated, or if we're inside a subplan, build
 316          * initPlan list and extParam/allParam sets for plan nodes.
 317          */
 318         if (PlannerPlanId != saved_planid || PlannerQueryLevel > 1)
 319         {
 320                 Cost    initplan_cost = 0;
 321
 322                 /* Prepare extParam/allParam sets for all nodes in tree */
 323                 SS_finalize_plan(plan, parse->rtable);
 324
 325                 /*
 326                  * SS_finalize_plan doesn't handle initPlans, so we have to manually
 327                  * attach them to the topmost plan node, and add their extParams to
 328                  * the topmost node's, too.
 329                  *
 330                  * We also add the total_cost of each initPlan to the startup cost
 331                  * of the top node.  This is a conservative overestimate, since in
 332                  * fact each initPlan might be executed later than plan startup, or
 333                  * even not at all.
 334                  */
 335                 plan->initPlan = PlannerInitPlan;
 336
 337                 foreach(lst, plan->initPlan)
 338                 {
 339                         SubPlan    *initplan = (SubPlan *) lfirst(lst);
 340
 341                         plan->extParam = bms_add_members(plan->extParam,
 342                                                                                          initplan->plan->extParam);
 343                         initplan_cost += initplan->plan->total_cost;
 344                 }
 345
 346                 plan->startup_cost += initplan_cost;
 347                 plan->total_cost += initplan_cost;
 348         }
 349
 350         /* Return to outer subquery context */
 351         PlannerQueryLevel--;
 352         PlannerInitPlan = saved_initplan;
 353         /* we do NOT restore PlannerPlanId; that's not an oversight! */
 354
 355         return plan;
 356 }
 357
 358 /*
 359  * preprocess_expression
 360  *              Do subquery_planner's preprocessing work for an expression,
 361  *              which can be a targetlist, a WHERE clause (including JOIN/ON
 362  *              conditions), or a HAVING clause.
 363  */
 364 static Node *
 365 preprocess_expression(Query *parse, Node *expr, int kind)
 366 {
 367         /*
 368          * If the query has any join RTEs, replace join alias variables with
 369          * base-relation variables. We must do this before sublink processing,
 370          * else sublinks expanded out from join aliases wouldn't get processed.
 371          */
 372         if (parse->hasJoinRTEs)
 373                 expr = flatten_join_alias_vars(parse, expr);
 374
 375         /*
 376          * Simplify constant expressions.
 377          *
 378          * Note that at this point quals have not yet been converted to
 379          * implicit-AND form, so we can apply eval_const_expressions directly.
 380          */
 381         expr = eval_const_expressions(expr);
 382
 383         /*
 384          * If it's a qual or havingQual, canonicalize it, and convert it to
 385          * implicit-AND format.
 386          *
 387          * XXX Is there any value in re-applying eval_const_expressions after
 388          * canonicalize_qual?
 389          */
 390         if (kind == EXPRKIND_QUAL)
 391         {
 392                 expr = (Node *) canonicalize_qual((Expr *) expr, true);
 393
 394 #ifdef OPTIMIZER_DEBUG
 395                 printf("After canonicalize_qual()\n");
 396                 pprint(expr);
 397 #endif
 398         }
 399
 400         /* Expand SubLinks to SubPlans */
 401         if (parse->hasSubLinks)
 402                 expr = SS_process_sublinks(expr, (kind == EXPRKIND_QUAL));
 403
 404         /* Replace uplevel vars with Param nodes */
 405         if (PlannerQueryLevel > 1)
 406                 expr = SS_replace_correlation_vars(expr);
 407
 408         return expr;
 409 }
 410
 411 /*
 412  * preprocess_qual_conditions
 413  *              Recursively scan the query's jointree and do subquery_planner's
 414  *              preprocessing work on each qual condition found therein.
 415  */
 416 static void
 417 preprocess_qual_conditions(Query *parse, Node *jtnode)
 418 {
 419         if (jtnode == NULL)
 420                 return;
 421         if (IsA(jtnode, RangeTblRef))
 422         {
 423                 /* nothing to do here */
 424         }
 425         else if (IsA(jtnode, FromExpr))
 426         {
 427                 FromExpr   *f = (FromExpr *) jtnode;
 428                 List       *l;
 429
 430                 foreach(l, f->fromlist)
 431                         preprocess_qual_conditions(parse, lfirst(l));
 432
 433                 f->quals = preprocess_expression(parse, f->quals, EXPRKIND_QUAL);
 434         }
 435         else if (IsA(jtnode, JoinExpr))
 436         {
 437                 JoinExpr   *j = (JoinExpr *) jtnode;
 438
 439                 preprocess_qual_conditions(parse, j->larg);
 440                 preprocess_qual_conditions(parse, j->rarg);
 441
 442                 j->quals = preprocess_expression(parse, j->quals, EXPRKIND_QUAL);
 443         }
 444         else
 445                 elog(ERROR, "preprocess_qual_conditions: unexpected node type %d",
 446                          nodeTag(jtnode));
 447 }
 448
 449 /*--------------------
 450  * inheritance_planner
 451  *        Generate a plan in the case where the result relation is an
 452  *        inheritance set.
 453  *
 454  * We have to handle this case differently from cases where a source
 455  * relation is an inheritance set.      Source inheritance is expanded at
 456  * the bottom of the plan tree (see allpaths.c), but target inheritance
 457  * has to be expanded at the top.  The reason is that for UPDATE, each
 458  * target relation needs a different targetlist matching its own column
 459  * set.  (This is not so critical for DELETE, but for simplicity we treat
 460  * inherited DELETE the same way.)      Fortunately, the UPDATE/DELETE target
 461  * can never be the nullable side of an outer join, so it's OK to generate
 462  * the plan this way.
 463  *
 464  * parse is the querytree produced by the parser & rewriter.
 465  * inheritlist is an integer list of RT indexes for the result relation set.
 466  *
 467  * Returns a query plan.
 468  *--------------------
 469  */
 470 static Plan *
 471 inheritance_planner(Query *parse, List *inheritlist)
 472 {
 473         int                     parentRTindex = parse->resultRelation;
 474         Oid                     parentOID = getrelid(parentRTindex, parse->rtable);
 475         int                     mainrtlength = length(parse->rtable);
 476         List       *subplans = NIL;
 477         List       *tlist = NIL;
 478         List       *l;
 479
 480         foreach(l, inheritlist)
 481         {
 482                 int                     childRTindex = lfirsti(l);
 483                 Oid                     childOID = getrelid(childRTindex, parse->rtable);
 484                 int                     subrtlength;
 485                 Query      *subquery;
 486                 Plan       *subplan;
 487
 488                 /* Generate modified query with this rel as target */
 489                 subquery = (Query *) adjust_inherited_attrs((Node *) parse,
 490                                                                                                 parentRTindex, parentOID,
 491                                                                                                  childRTindex, childOID);
 492                 /* Generate plan */
 493                 subplan = grouping_planner(subquery, 0.0 /* retrieve all tuples */ );
 494                 subplans = lappend(subplans, subplan);
 495                 /*
 496                  * It's possible that additional RTEs got added to the rangetable
 497                  * due to expansion of inherited source tables (see allpaths.c).
 498                  * If so, we must copy 'em back to the main parse tree's rtable.
 499                  *
 500                  * XXX my goodness this is ugly.  Really need to think about ways
 501                  * to rein in planner's habit of scribbling on its input.
 502                  */
 503                 subrtlength = length(subquery->rtable);
 504                 if (subrtlength > mainrtlength)
 505                 {
 506                         List   *subrt = subquery->rtable;
 507
 508                         while (mainrtlength-- > 0) /* wish we had nthcdr() */
 509                                 subrt = lnext(subrt);
 510                         parse->rtable = nconc(parse->rtable, subrt);
 511                         mainrtlength = subrtlength;
 512                 }
 513                 /* Save preprocessed tlist from first rel for use in Append */
 514                 if (tlist == NIL)
 515                         tlist = subplan->targetlist;
 516         }
 517
 518         /* Save the target-relations list for the executor, too */
 519         parse->resultRelations = inheritlist;
 520
 521         /* Mark result as unordered (probably unnecessary) */
 522         parse->query_pathkeys = NIL;
 523
 524         return (Plan *) make_append(subplans, true, tlist);
 525 }
 526
 527 /*--------------------
 528  * grouping_planner
 529  *        Perform planning steps related to grouping, aggregation, etc.
 530  *        This primarily means adding top-level processing to the basic
 531  *        query plan produced by query_planner.
 532  *
 533  * parse is the querytree produced by the parser & rewriter.
 534  * tuple_fraction is the fraction of tuples we expect will be retrieved
 535  *
 536  * tuple_fraction is interpreted as follows:
 537  *        0: expect all tuples to be retrieved (normal case)
 538  *        0 < tuple_fraction < 1: expect the given fraction of tuples available
 539  *              from the plan to be retrieved
 540  *        tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
 541  *              expected to be retrieved (ie, a LIMIT specification)
 542  *
 543  * Returns a query plan.  Also, parse->query_pathkeys is returned as the
 544  * actual output ordering of the plan (in pathkey format).
 545  *--------------------
 546  */
 547 static Plan *
 548 grouping_planner(Query *parse, double tuple_fraction)
 549 {
 550         List       *tlist = parse->targetList;
 551         Plan       *result_plan;
 552         List       *current_pathkeys;
 553         List       *sort_pathkeys;
 554
 555         if (parse->setOperations)
 556         {
 557                 /*
 558                  * Construct the plan for set operations.  The result will not
 559                  * need any work except perhaps a top-level sort and/or LIMIT.
 560                  */
 561                 result_plan = plan_set_operations(parse);
 562
 563                 /*
 564                  * We should not need to call preprocess_targetlist, since we must
 565                  * be in a SELECT query node.  Instead, use the targetlist
 566                  * returned by plan_set_operations (since this tells whether it
 567                  * returned any resjunk columns!), and transfer any sort key
 568                  * information from the original tlist.
 569                  */
 570                 Assert(parse->commandType == CMD_SELECT);
 571
 572                 tlist = postprocess_setop_tlist(result_plan->targetlist, tlist);
 573
 574                 /*
 575                  * Can't handle FOR UPDATE here (parser should have checked
 576                  * already, but let's make sure).
 577                  */
 578                 if (parse->rowMarks)
 579                         elog(ERROR, "SELECT FOR UPDATE is not allowed with UNION/INTERSECT/EXCEPT");
 580
 581                 /*
 582                  * We set current_pathkeys NIL indicating we do not know sort
 583                  * order.  This is correct when the top set operation is UNION
 584                  * ALL, since the appended-together results are unsorted even if
 585                  * the subplans were sorted.  For other set operations we could be
 586                  * smarter --- room for future improvement!
 587                  */
 588                 current_pathkeys = NIL;
 589
 590                 /*
 591                  * Calculate pathkeys that represent ordering requirements
 592                  */
 593                 sort_pathkeys = make_pathkeys_for_sortclauses(parse->sortClause,
 594                                                                                                           tlist);
 595                 sort_pathkeys = canonicalize_pathkeys(parse, sort_pathkeys);
 596         }
 597         else
 598         {
 599                 /* No set operations, do regular planning */
 600                 List       *sub_tlist;
 601                 List       *group_pathkeys;
 602                 AttrNumber *groupColIdx = NULL;
 603                 bool            need_tlist_eval = true;
 604                 QualCost        tlist_cost;
 605                 double          sub_tuple_fraction;
 606                 Path       *cheapest_path;
 607                 Path       *sorted_path;
 608                 double          dNumGroups = 0;
 609                 long            numGroups = 0;
 610                 int                     numAggs = 0;
 611                 int                     numGroupCols = length(parse->groupClause);
 612                 bool            use_hashed_grouping = false;
 613
 614                 /* Preprocess targetlist in case we are inside an INSERT/UPDATE. */
 615                 tlist = preprocess_targetlist(tlist,
 616                                                                           parse->commandType,
 617                                                                           parse->resultRelation,
 618                                                                           parse->rtable);
 619
 620                 /*
 621                  * Add TID targets for rels selected FOR UPDATE (should this be
 622                  * done in preprocess_targetlist?).  The executor uses the TID to
 623                  * know which rows to lock, much as for UPDATE or DELETE.
 624                  */
 625                 if (parse->rowMarks)
 626                 {
 627                         List       *l;
 628
 629                         /*
 630                          * We've got trouble if the FOR UPDATE appears inside
 631                          * grouping, since grouping renders a reference to individual
 632                          * tuple CTIDs invalid.  This is also checked at parse time,
 633                          * but that's insufficient because of rule substitution, query
 634                          * pullup, etc.
 635                          */
 636                         CheckSelectForUpdate(parse);
 637
 638                         /*
 639                          * Currently the executor only supports FOR UPDATE at top
 640                          * level
 641                          */
 642                         if (PlannerQueryLevel > 1)
 643                                 elog(ERROR, "SELECT FOR UPDATE is not allowed in subselects");
 644
 645                         foreach(l, parse->rowMarks)
 646                         {
 647                                 Index           rti = lfirsti(l);
 648                                 char       *resname;
 649                                 Resdom     *resdom;
 650                                 Var                *var;
 651                                 TargetEntry *ctid;
 652
 653                                 resname = (char *) palloc(32);
 654                                 snprintf(resname, 32, "ctid%u", rti);
 655                                 resdom = makeResdom(length(tlist) + 1,
 656                                                                         TIDOID,
 657                                                                         -1,
 658                                                                         resname,
 659                                                                         true);
 660
 661                                 var = makeVar(rti,
 662                                                           SelfItemPointerAttributeNumber,
 663                                                           TIDOID,
 664                                                           -1,
 665                                                           0);
 666
 667                                 ctid = makeTargetEntry(resdom, (Expr *) var);
 668                                 tlist = lappend(tlist, ctid);
 669                         }
 670                 }
 671
 672                 /*
 673                  * Generate appropriate target list for subplan; may be different
 674                  * from tlist if grouping or aggregation is needed.
 675                  */
 676                 sub_tlist = make_subplanTargetList(parse, tlist,
 677                                                                                    &groupColIdx, &need_tlist_eval);
 678
 679                 /*
 680                  * Calculate pathkeys that represent grouping/ordering
 681                  * requirements
 682                  */
 683                 group_pathkeys = make_pathkeys_for_sortclauses(parse->groupClause,
 684                                                                                                            tlist);
 685                 sort_pathkeys = make_pathkeys_for_sortclauses(parse->sortClause,
 686                                                                                                           tlist);
 687
 688                 /*
 689                  * Will need actual number of aggregates for estimating costs.
 690                  * Also, it's possible that optimization has eliminated all
 691                  * aggregates, and we may as well check for that here.
 692                  *
 693                  * Note: we do not attempt to detect duplicate aggregates here;
 694                  * a somewhat-overestimated count is okay for our present purposes.
 695                  */
 696                 if (parse->hasAggs)
 697                 {
 698                         numAggs = count_agg_clause((Node *) tlist) +
 699                                 count_agg_clause(parse->havingQual);
 700                         if (numAggs == 0)
 701                                 parse->hasAggs = false;
 702                 }
 703
 704                 /*
 705                  * Figure out whether we need a sorted result from query_planner.
 706                  *
 707                  * If we have a GROUP BY clause, then we want a result sorted
 708                  * properly for grouping.  Otherwise, if there is an ORDER BY
 709                  * clause, we want to sort by the ORDER BY clause.      (Note: if we
 710                  * have both, and ORDER BY is a superset of GROUP BY, it would be
 711                  * tempting to request sort by ORDER BY --- but that might just
 712                  * leave us failing to exploit an available sort order at all.
 713                  * Needs more thought...)
 714                  */
 715                 if (parse->groupClause)
 716                         parse->query_pathkeys = group_pathkeys;
 717                 else if (parse->sortClause)
 718                         parse->query_pathkeys = sort_pathkeys;
 719                 else
 720                         parse->query_pathkeys = NIL;
 721
 722                 /*
 723                  * Adjust tuple_fraction if we see that we are going to apply
 724                  * limiting/grouping/aggregation/etc.  This is not overridable by
 725                  * the caller, since it reflects plan actions that this routine
 726                  * will certainly take, not assumptions about context.
 727                  */
 728                 if (parse->limitCount != NULL)
 729                 {
 730                         /*
 731                          * A LIMIT clause limits the absolute number of tuples
 732                          * returned. However, if it's not a constant LIMIT then we
 733                          * have to punt; for lack of a better idea, assume 10% of the
 734                          * plan's result is wanted.
 735                          */
 736                         double          limit_fraction = 0.0;
 737
 738                         if (IsA(parse->limitCount, Const))
 739                         {
 740                                 Const      *limitc = (Const *) parse->limitCount;
 741                                 int32           count = DatumGetInt32(limitc->constvalue);
 742
 743                                 /*
 744                                  * A NULL-constant LIMIT represents "LIMIT ALL", which we
 745                                  * treat the same as no limit (ie, expect to retrieve all
 746                                  * the tuples).
 747                                  */
 748                                 if (!limitc->constisnull && count > 0)
 749                                 {
 750                                         limit_fraction = (double) count;
 751                                         /* We must also consider the OFFSET, if present */
 752                                         if (parse->limitOffset != NULL)
 753                                         {
 754                                                 if (IsA(parse->limitOffset, Const))
 755                                                 {
 756                                                         int32           offset;
 757
 758                                                         limitc = (Const *) parse->limitOffset;
 759                                                         offset = DatumGetInt32(limitc->constvalue);
 760                                                         if (!limitc->constisnull && offset > 0)
 761                                                                 limit_fraction += (double) offset;
 762                                                 }
 763                                                 else
 764                                                 {
 765                                                         /* OFFSET is an expression ... punt ... */
 766                                                         limit_fraction = 0.10;
 767                                                 }
 768                                         }
 769                                 }
 770                         }
 771                         else
 772                         {
 773                                 /* LIMIT is an expression ... punt ... */
 774                                 limit_fraction = 0.10;
 775                         }
 776
 777                         if (limit_fraction > 0.0)
 778                         {
 779                                 /*
 780                                  * If we have absolute limits from both caller and LIMIT,
 781                                  * use the smaller value; if one is fractional and the
 782                                  * other absolute, treat the fraction as a fraction of the
 783                                  * absolute value; else we can multiply the two fractions
 784                                  * together.
 785                                  */
 786                                 if (tuple_fraction >= 1.0)
 787                                 {
 788                                         if (limit_fraction >= 1.0)
 789                                         {
 790                                                 /* both absolute */
 791                                                 tuple_fraction = Min(tuple_fraction, limit_fraction);
 792                                         }
 793                                         else
 794                                         {
 795                                                 /* caller absolute, limit fractional */
 796                                                 tuple_fraction *= limit_fraction;
 797                                                 if (tuple_fraction < 1.0)
 798                                                         tuple_fraction = 1.0;
 799                                         }
 800                                 }
 801                                 else if (tuple_fraction > 0.0)
 802                                 {
 803                                         if (limit_fraction >= 1.0)
 804                                         {
 805                                                 /* caller fractional, limit absolute */
 806                                                 tuple_fraction *= limit_fraction;
 807                                                 if (tuple_fraction < 1.0)
 808                                                         tuple_fraction = 1.0;
 809                                         }
 810                                         else
 811                                         {
 812                                                 /* both fractional */
 813                                                 tuple_fraction *= limit_fraction;
 814                                         }
 815                                 }
 816                                 else
 817                                 {
 818                                         /* no info from caller, just use limit */
 819                                         tuple_fraction = limit_fraction;
 820                                 }
 821                         }
 822                 }
 823
 824                 /*
 825                  * With grouping or aggregation, the tuple fraction to pass to
 826                  * query_planner() may be different from what it is at top level.
 827                  */
 828                 sub_tuple_fraction = tuple_fraction;
 829
 830                 if (parse->groupClause)
 831                 {
 832                         /*
 833                          * In GROUP BY mode, we have the little problem that we don't
 834                          * really know how many input tuples will be needed to make a
 835                          * group, so we can't translate an output LIMIT count into an
 836                          * input count.  For lack of a better idea, assume 25% of the
 837                          * input data will be processed if there is any output limit.
 838                          * However, if the caller gave us a fraction rather than an
 839                          * absolute count, we can keep using that fraction (which
 840                          * amounts to assuming that all the groups are about the same
 841                          * size).
 842                          */
 843                         if (sub_tuple_fraction >= 1.0)
 844                                 sub_tuple_fraction = 0.25;
 845
 846                         /*
 847                          * If both GROUP BY and ORDER BY are specified, we will need
 848                          * two levels of sort --- and, therefore, certainly need to
 849                          * read all the input tuples --- unless ORDER BY is a subset
 850                          * of GROUP BY.  (We have not yet canonicalized the pathkeys,
 851                          * so must use the slower noncanonical comparison method.)
 852                          */
 853                         if (parse->groupClause && parse->sortClause &&
 854                                 !noncanonical_pathkeys_contained_in(sort_pathkeys,
 855                                                                                                         group_pathkeys))
 856                                 sub_tuple_fraction = 0.0;
 857                 }
 858                 else if (parse->hasAggs)
 859                 {
 860                         /*
 861                          * Ungrouped aggregate will certainly want all the input
 862                          * tuples.
 863                          */
 864                         sub_tuple_fraction = 0.0;
 865                 }
 866                 else if (parse->distinctClause)
 867                 {
 868                         /*
 869                          * SELECT DISTINCT, like GROUP, will absorb an unpredictable
 870                          * number of input tuples per output tuple.  Handle the same
 871                          * way.
 872                          */
 873                         if (sub_tuple_fraction >= 1.0)
 874                                 sub_tuple_fraction = 0.25;
 875                 }
 876
 877                 /*
 878                  * Generate the best unsorted and presorted paths for this Query
 879                  * (but note there may not be any presorted path).
 880                  */
 881                 query_planner(parse, sub_tlist, sub_tuple_fraction,
 882                                           &cheapest_path, &sorted_path);
 883
 884                 /*
 885                  * We couldn't canonicalize group_pathkeys and sort_pathkeys before
 886                  * running query_planner(), so do it now.
 887                  */
 888                 group_pathkeys = canonicalize_pathkeys(parse, group_pathkeys);
 889                 sort_pathkeys = canonicalize_pathkeys(parse, sort_pathkeys);
 890
 891                 /*
 892                  * Consider whether we might want to use hashed grouping.
 893                  */
 894                 if (parse->groupClause)
 895                 {
 896                         List   *groupExprs;
 897
 898                         /*
 899                          * Always estimate the number of groups.  We can't do this until
 900                          * after running query_planner(), either.
 901                          */
 902                         groupExprs = get_sortgrouplist_exprs(parse->groupClause,
 903                                                                                                  parse->targetList);
 904                         dNumGroups = estimate_num_groups(parse,
 905                                                                                          groupExprs,
 906                                                                                          cheapest_path->parent->rows);
 907                         /* Also want it as a long int --- but 'ware overflow! */
 908                         numGroups = (long) Min(dNumGroups, (double) LONG_MAX);
 909
 910                         /*
 911                          * Check can't-do-it conditions, including whether the grouping
 912                          * operators are hashjoinable.
 913                          *
 914                          * Executor doesn't support hashed aggregation with DISTINCT
 915                          * aggregates.  (Doing so would imply storing *all* the input
 916                          * values in the hash table, which seems like a certain loser.)
 917                          */
 918                         if (!enable_hashagg || !hash_safe_grouping(parse))
 919                                 use_hashed_grouping = false;
 920                         else if (parse->hasAggs &&
 921                                          (contain_distinct_agg_clause((Node *) tlist) ||
 922                                           contain_distinct_agg_clause(parse->havingQual)))
 923                                 use_hashed_grouping = false;
 924                         else
 925                         {
 926                                 /*
 927                                  * Use hashed grouping if (a) we think we can fit the
 928                                  * hashtable into SortMem, *and* (b) the estimated cost
 929                                  * is no more than doing it the other way.  While avoiding
 930                                  * the need for sorted input is usually a win, the fact
 931                                  * that the output won't be sorted may be a loss; so we
 932                                  * need to do an actual cost comparison.
 933                                  *
 934                                  * In most cases we have no good way to estimate the size of
 935                                  * the transition value needed by an aggregate; arbitrarily
 936                                  * assume it is 100 bytes.  Also set the overhead per hashtable
 937                                  * entry at 64 bytes.
 938                                  */
 939                                 int             hashentrysize = cheapest_path->parent->width + 64 +
 940                                         numAggs * 100;
 941
 942                                 if (hashentrysize * dNumGroups <= SortMem * 1024L)
 943                                 {
 944                                         /*
 945                                          * Okay, do the cost comparison.  We need to consider
 946                                          *      cheapest_path + hashagg [+ final sort]
 947                                          * versus either
 948                                          *      cheapest_path [+ sort] + group or agg [+ final sort]
 949                                          * or
 950                                          *      presorted_path + group or agg [+ final sort]
 951                                          * where brackets indicate a step that may not be needed.
 952                                          * We assume query_planner() will have returned a
 953                                          * presorted path only if it's a winner compared to
 954                                          * cheapest_path for this purpose.
 955                                          *
 956                                          * These path variables are dummies that just hold cost
 957                                          * fields; we don't make actual Paths for these steps.
 958                                          */
 959                                         Path            hashed_p;
 960                                         Path            sorted_p;
 961
 962                                         cost_agg(&hashed_p, parse,
 963                                                          AGG_HASHED, numAggs,
 964                                                          numGroupCols, dNumGroups,
 965                                                          cheapest_path->startup_cost,
 966                                                          cheapest_path->total_cost,
 967                                                          cheapest_path->parent->rows);
 968                                         /* Result of hashed agg is always unsorted */
 969                                         if (sort_pathkeys)
 970                                                 cost_sort(&hashed_p, parse, sort_pathkeys,
 971                                                                   hashed_p.total_cost,
 972                                                                   dNumGroups,
 973                                                                   cheapest_path->parent->width);
 974
 975                                         if (sorted_path)
 976                                         {
 977                                                 sorted_p.startup_cost = sorted_path->startup_cost;
 978                                                 sorted_p.total_cost = sorted_path->total_cost;
 979                                                 current_pathkeys = sorted_path->pathkeys;
 980                                         }
 981                                         else
 982                                         {
 983                                                 sorted_p.startup_cost = cheapest_path->startup_cost;
 984                                                 sorted_p.total_cost = cheapest_path->total_cost;
 985                                                 current_pathkeys = cheapest_path->pathkeys;
 986                                         }
 987                                         if (!pathkeys_contained_in(group_pathkeys,
 988                                                                                            current_pathkeys))
 989                                         {
 990                                                 cost_sort(&sorted_p, parse, group_pathkeys,
 991                                                                   sorted_p.total_cost,
 992                                                                   cheapest_path->parent->rows,
 993                                                                   cheapest_path->parent->width);
 994                                                 current_pathkeys = group_pathkeys;
 995                                         }
 996                                         if (parse->hasAggs)
 997                                                 cost_agg(&sorted_p, parse,
 998                                                                  AGG_SORTED, numAggs,
 999                                                                  numGroupCols, dNumGroups,
1000                                                                  sorted_p.startup_cost,
1001                                                                  sorted_p.total_cost,
1002                                                                  cheapest_path->parent->rows);
1003                                         else
1004                                                 cost_group(&sorted_p, parse,
1005                                                                    numGroupCols, dNumGroups,
1006                                                                    sorted_p.startup_cost,
1007                                                                    sorted_p.total_cost,
1008                                                                    cheapest_path->parent->rows);
1009                                         /* The Agg or Group node will preserve ordering */
1010                                         if (sort_pathkeys &&
1011                                                 !pathkeys_contained_in(sort_pathkeys,
1012                                                                                            current_pathkeys))
1013                                         {
1014                                                 cost_sort(&sorted_p, parse, sort_pathkeys,
1015                                                                   sorted_p.total_cost,
1016                                                                   dNumGroups,
1017                                                                   cheapest_path->parent->width);
1018                                         }
1019
1020                                         /*
1021                                          * Now make the decision using the top-level tuple
1022                                          * fraction.  First we have to convert an absolute
1023                                          * count (LIMIT) into fractional form.
1024                                          */
1025                                         if (tuple_fraction >= 1.0)
1026                                                 tuple_fraction /= dNumGroups;
1027
1028                                         if (compare_fractional_path_costs(&hashed_p, &sorted_p,
1029                                                                                                           tuple_fraction) < 0)
1030                                         {
1031                                                 /* Hashed is cheaper, so use it */
1032                                                 use_hashed_grouping = true;
1033                                         }
1034                                 }
1035                         }
1036                 }
1037
1038                 /*
1039                  * Select the best path and create a plan to execute it.
1040                  *
1041                  * If we are doing hashed grouping, we will always read all the
1042                  * input tuples, so use the cheapest-total path.  Otherwise,
1043                  * trust query_planner's decision about which to use.
1044                  */
1045                 if (sorted_path && !use_hashed_grouping)
1046                 {
1047                         result_plan = create_plan(parse, sorted_path);
1048                         current_pathkeys = sorted_path->pathkeys;
1049                 }
1050                 else
1051                 {
1052                         result_plan = create_plan(parse, cheapest_path);
1053                         current_pathkeys = cheapest_path->pathkeys;
1054                 }
1055
1056                 /*
1057                  * create_plan() returns a plan with just a "flat" tlist of required
1058                  * Vars.  Usually we need to insert the sub_tlist as the tlist of the
1059                  * top plan node.  However, we can skip that if we determined that
1060                  * whatever query_planner chose to return will be good enough.
1061                  */
1062                 if (need_tlist_eval)
1063                 {
1064                         /*
1065                          * If the top-level plan node is one that cannot do expression
1066                          * evaluation, we must insert a Result node to project the desired
1067                          * tlist.
1068                          * Currently, the only plan node we might see here that falls into
1069                          * that category is Append.
1070                          */
1071                         if (IsA(result_plan, Append))
1072                         {
1073                                 result_plan = (Plan *) make_result(sub_tlist, NULL,
1074                                                                                                    result_plan);
1075                         }
1076                         else
1077                         {
1078                                 /*
1079                                  * Otherwise, just replace the subplan's flat tlist with
1080                                  * the desired tlist.
1081                                  */
1082                                 result_plan->targetlist = sub_tlist;
1083                         }
1084                         /*
1085                          * Also, account for the cost of evaluation of the sub_tlist.
1086                          *
1087                          * Up to now, we have only been dealing with "flat" tlists,
1088                          * containing just Vars.  So their evaluation cost is zero
1089                          * according to the model used by cost_qual_eval() (or if you
1090                          * prefer, the cost is factored into cpu_tuple_cost).  Thus we can
1091                          * avoid accounting for tlist cost throughout query_planner() and
1092                          * subroutines.  But now we've inserted a tlist that might contain
1093                          * actual operators, sub-selects, etc --- so we'd better account
1094                          * for its cost.
1095                          *
1096                          * Below this point, any tlist eval cost for added-on nodes should
1097                          * be accounted for as we create those nodes.  Presently, of the
1098                          * node types we can add on, only Agg and Group project new tlists
1099                          * (the rest just copy their input tuples) --- so make_agg() and
1100                          * make_group() are responsible for computing the added cost.
1101                          */
1102                         cost_qual_eval(&tlist_cost, sub_tlist);
1103                         result_plan->startup_cost += tlist_cost.startup;
1104                         result_plan->total_cost += tlist_cost.startup +
1105                                 tlist_cost.per_tuple * result_plan->plan_rows;
1106                 }
1107                 else
1108                 {
1109                         /*
1110                          * Since we're using query_planner's tlist and not the one
1111                          * make_subplanTargetList calculated, we have to refigure
1112                          * any grouping-column indexes make_subplanTargetList computed.
1113                          */
1114                         locate_grouping_columns(parse, tlist, result_plan->targetlist,
1115                                                                         groupColIdx);
1116                 }
1117
1118                 /*
1119                  * Insert AGG or GROUP node if needed, plus an explicit sort step
1120                  * if necessary.
1121                  *
1122                  * HAVING clause, if any, becomes qual of the Agg node
1123                  */
1124                 if (use_hashed_grouping)
1125                 {
1126                         /* Hashed aggregate plan --- no sort needed */
1127                         result_plan = (Plan *) make_agg(parse,
1128                                                                                         tlist,
1129                                                                                         (List *) parse->havingQual,
1130                                                                                         AGG_HASHED,
1131                                                                                         numGroupCols,
1132                                                                                         groupColIdx,
1133                                                                                         numGroups,
1134                                                                                         numAggs,
1135                                                                                         result_plan);
1136                         /* Hashed aggregation produces randomly-ordered results */
1137                         current_pathkeys = NIL;
1138                 }
1139                 else if (parse->hasAggs)
1140                 {
1141                         /* Plain aggregate plan --- sort if needed */
1142                         AggStrategy aggstrategy;
1143
1144                         if (parse->groupClause)
1145                         {
1146                                 if (!pathkeys_contained_in(group_pathkeys, current_pathkeys))
1147                                 {
1148                                         result_plan = make_groupsortplan(parse,
1149                                                                                                          parse->groupClause,
1150                                                                                                          groupColIdx,
1151                                                                                                          result_plan);
1152                                         current_pathkeys = group_pathkeys;
1153                                 }
1154                                 aggstrategy = AGG_SORTED;
1155                                 /*
1156                                  * The AGG node will not change the sort ordering of its
1157                                  * groups, so current_pathkeys describes the result too.
1158                                  */
1159                         }
1160                         else
1161                         {
1162                                 aggstrategy = AGG_PLAIN;
1163                                 /* Result will be only one row anyway; no sort order */
1164                                 current_pathkeys = NIL;
1165                         }
1166
1167                         result_plan = (Plan *) make_agg(parse,
1168                                                                                         tlist,
1169                                                                                         (List *) parse->havingQual,
1170                                                                                         aggstrategy,
1171                                                                                         numGroupCols,
1172                                                                                         groupColIdx,
1173                                                                                         numGroups,
1174                                                                                         numAggs,
1175                                                                                         result_plan);
1176                 }
1177                 else
1178                 {
1179                         /*
1180                          * If there are no Aggs, we shouldn't have any HAVING qual anymore
1181                          */
1182                         Assert(parse->havingQual == NULL);
1183
1184                         /*
1185                          * If we have a GROUP BY clause, insert a group node (plus the
1186                          * appropriate sort node, if necessary).
1187                          */
1188                         if (parse->groupClause)
1189                         {
1190                                 /*
1191                                  * Add an explicit sort if we couldn't make the path come out
1192                                  * the way the GROUP node needs it.
1193                                  */
1194                                 if (!pathkeys_contained_in(group_pathkeys, current_pathkeys))
1195                                 {
1196                                         result_plan = make_groupsortplan(parse,
1197                                                                                                          parse->groupClause,
1198                                                                                                          groupColIdx,
1199                                                                                                          result_plan);
1200                                         current_pathkeys = group_pathkeys;
1201                                 }
1202
1203                                 result_plan = (Plan *) make_group(parse,
1204                                                                                                   tlist,
1205                                                                                                   numGroupCols,
1206                                                                                                   groupColIdx,
1207                                                                                                   dNumGroups,
1208                                                                                                   result_plan);
1209                                 /* The Group node won't change sort ordering */
1210                         }
1211                 }
1212         } /* end of if (setOperations) */
1213
1214         /*
1215          * If we were not able to make the plan come out in the right order,
1216          * add an explicit sort step.
1217          */
1218         if (parse->sortClause)
1219         {
1220                 if (!pathkeys_contained_in(sort_pathkeys, current_pathkeys))
1221                 {
1222                         result_plan = (Plan *) make_sort_from_sortclauses(parse,
1223                                                                                                                           tlist,
1224                                                                                                                           result_plan,
1225                                                                                                                           parse->sortClause);
1226                         current_pathkeys = sort_pathkeys;
1227                 }
1228         }
1229
1230         /*
1231          * If there is a DISTINCT clause, add the UNIQUE node.
1232          */
1233         if (parse->distinctClause)
1234         {
1235                 result_plan = (Plan *) make_unique(tlist, result_plan,
1236                                                                                    parse->distinctClause);
1237                 /*
1238                  * If there was grouping or aggregation, leave plan_rows as-is
1239                  * (ie, assume the result was already mostly unique).  If not,
1240                  * it's reasonable to assume the UNIQUE filter has effects
1241                  * comparable to GROUP BY.
1242                  */
1243                 if (!parse->groupClause && !parse->hasAggs)
1244                 {
1245                         List   *distinctExprs;
1246
1247                         distinctExprs = get_sortgrouplist_exprs(parse->distinctClause,
1248                                                                                                         parse->targetList);
1249                         result_plan->plan_rows = estimate_num_groups(parse,
1250                                                                                                                  distinctExprs,
1251                                                                                                                  result_plan->plan_rows);
1252                 }
1253         }
1254
1255         /*
1256          * Finally, if there is a LIMIT/OFFSET clause, add the LIMIT node.
1257          */
1258         if (parse->limitOffset || parse->limitCount)
1259         {
1260                 result_plan = (Plan *) make_limit(tlist, result_plan,
1261                                                                                   parse->limitOffset,
1262                                                                                   parse->limitCount);
1263         }
1264
1265         /*
1266          * Return the actual output ordering in query_pathkeys for possible
1267          * use by an outer query level.
1268          */
1269         parse->query_pathkeys = current_pathkeys;
1270
1271         return result_plan;
1272 }
1273
1274 /*
1275  * hash_safe_grouping - are grouping operators hashable?
1276  *
1277  * We assume hashed aggregation will work if the datatype's equality operator
1278  * is marked hashjoinable.
1279  */
1280 static bool
1281 hash_safe_grouping(Query *parse)
1282 {
1283         List       *gl;
1284
1285         foreach(gl, parse->groupClause)
1286         {
1287                 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1288                 TargetEntry *tle = get_sortgroupclause_tle(grpcl, parse->targetList);
1289                 Operator        optup;
1290                 bool            oprcanhash;
1291
1292                 optup = equality_oper(tle->resdom->restype, false);
1293                 oprcanhash = ((Form_pg_operator) GETSTRUCT(optup))->oprcanhash;
1294                 ReleaseSysCache(optup);
1295                 if (!oprcanhash)
1296                         return false;
1297         }
1298         return true;
1299 }
1300
1301 /*---------------
1302  * make_subplanTargetList
1303  *        Generate appropriate target list when grouping is required.
1304  *
1305  * When grouping_planner inserts Aggregate or Group plan nodes above
1306  * the result of query_planner, we typically want to pass a different
1307  * target list to query_planner than the outer plan nodes should have.
1308  * This routine generates the correct target list for the subplan.
1309  *
1310  * The initial target list passed from the parser already contains entries
1311  * for all ORDER BY and GROUP BY expressions, but it will not have entries
1312  * for variables used only in HAVING clauses; so we need to add those
1313  * variables to the subplan target list.  Also, if we are doing either
1314  * grouping or aggregation, we flatten all expressions except GROUP BY items
1315  * into their component variables; the other expressions will be computed by
1316  * the inserted nodes rather than by the subplan.  For example,
1317  * given a query like
1318  *              SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
1319  * we want to pass this targetlist to the subplan:
1320  *              a,b,c,d,a+b
1321  * where the a+b target will be used by the Sort/Group steps, and the
1322  * other targets will be used for computing the final results.  (In the
1323  * above example we could theoretically suppress the a and b targets and
1324  * pass down only c,d,a+b, but it's not really worth the trouble to
1325  * eliminate simple var references from the subplan.  We will avoid doing
1326  * the extra computation to recompute a+b at the outer level; see
1327  * replace_vars_with_subplan_refs() in setrefs.c.)
1328  *
1329  * If we are grouping or aggregating, *and* there are no non-Var grouping
1330  * expressions, then the returned tlist is effectively dummy; we do not
1331  * need to force it to be evaluated, because all the Vars it contains
1332  * should be present in the output of query_planner anyway.
1333  *
1334  * 'parse' is the query being processed.
1335  * 'tlist' is the query's target list.
1336  * 'groupColIdx' receives an array of column numbers for the GROUP BY
1337  *                      expressions (if there are any) in the subplan's target list.
1338  * 'need_tlist_eval' is set true if we really need to evaluate the
1339  *                      result tlist.
1340  *
1341  * The result is the targetlist to be passed to the subplan.
1342  *---------------
1343  */
1344 static List *
1345 make_subplanTargetList(Query *parse,
1346                                            List *tlist,
1347                                            AttrNumber **groupColIdx,
1348                                            bool *need_tlist_eval)
1349 {
1350         List       *sub_tlist;
1351         List       *extravars;
1352         int                     numCols;
1353
1354         *groupColIdx = NULL;
1355
1356         /*
1357          * If we're not grouping or aggregating, nothing to do here;
1358          * query_planner should receive the unmodified target list.
1359          */
1360         if (!parse->hasAggs && !parse->groupClause && !parse->havingQual)
1361         {
1362                 *need_tlist_eval = true;
1363                 return tlist;
1364         }
1365
1366         /*
1367          * Otherwise, start with a "flattened" tlist (having just the vars
1368          * mentioned in the targetlist and HAVING qual --- but not upper-
1369          * level Vars; they will be replaced by Params later on).
1370          */
1371         sub_tlist = flatten_tlist(tlist);
1372         extravars = pull_var_clause(parse->havingQual, false);
1373         sub_tlist = add_to_flat_tlist(sub_tlist, extravars);
1374         freeList(extravars);
1375         *need_tlist_eval = false;       /* only eval if not flat tlist */
1376
1377         /*
1378          * If grouping, create sub_tlist entries for all GROUP BY expressions
1379          * (GROUP BY items that are simple Vars should be in the list
1380          * already), and make an array showing where the group columns are in
1381          * the sub_tlist.
1382          */
1383         numCols = length(parse->groupClause);
1384         if (numCols > 0)
1385         {
1386                 int                     keyno = 0;
1387                 AttrNumber *grpColIdx;
1388                 List       *gl;
1389
1390                 grpColIdx = (AttrNumber *) palloc(sizeof(AttrNumber) * numCols);
1391                 *groupColIdx = grpColIdx;
1392
1393                 foreach(gl, parse->groupClause)
1394                 {
1395                         GroupClause *grpcl = (GroupClause *) lfirst(gl);
1396                         Node       *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1397                         TargetEntry *te = NULL;
1398                         List       *sl;
1399
1400                         /* Find or make a matching sub_tlist entry */
1401                         foreach(sl, sub_tlist)
1402                         {
1403                                 te = (TargetEntry *) lfirst(sl);
1404                                 if (equal(groupexpr, te->expr))
1405                                         break;
1406                         }
1407                         if (!sl)
1408                         {
1409                                 te = makeTargetEntry(makeResdom(length(sub_tlist) + 1,
1410                                                                                                 exprType(groupexpr),
1411                                                                                                 exprTypmod(groupexpr),
1412                                                                                                 NULL,
1413                                                                                                 false),
1414                                                                          (Expr *) groupexpr);
1415                                 sub_tlist = lappend(sub_tlist, te);
1416                                 *need_tlist_eval = true; /* it's not flat anymore */
1417                         }
1418
1419                         /* and save its resno */
1420                         grpColIdx[keyno++] = te->resdom->resno;
1421                 }
1422         }
1423
1424         return sub_tlist;
1425 }
1426
1427 /*
1428  * locate_grouping_columns
1429  *              Locate grouping columns in the tlist chosen by query_planner.
1430  *
1431  * This is only needed if we don't use the sub_tlist chosen by
1432  * make_subplanTargetList.  We have to forget the column indexes found
1433  * by that routine and re-locate the grouping vars in the real sub_tlist.
1434  */
1435 static void
1436 locate_grouping_columns(Query *parse,
1437                                                 List *tlist,
1438                                                 List *sub_tlist,
1439                                                 AttrNumber *groupColIdx)
1440 {
1441         int                     keyno = 0;
1442         List       *gl;
1443
1444         /*
1445          * No work unless grouping.
1446          */
1447         if (!parse->groupClause)
1448         {
1449                 Assert(groupColIdx == NULL);
1450                 return;
1451         }
1452         Assert(groupColIdx != NULL);
1453
1454         foreach(gl, parse->groupClause)
1455         {
1456                 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1457                 Node       *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1458                 TargetEntry *te = NULL;
1459                 List       *sl;
1460
1461                 foreach(sl, sub_tlist)
1462                 {
1463                         te = (TargetEntry *) lfirst(sl);
1464                         if (equal(groupexpr, te->expr))
1465                                 break;
1466                 }
1467                 if (!sl)
1468                         elog(ERROR, "locate_grouping_columns: failed");
1469
1470                 groupColIdx[keyno++] = te->resdom->resno;
1471         }
1472 }
1473
1474 /*
1475  * make_groupsortplan
1476  *              Add a Sort node to explicitly sort according to the GROUP BY clause.
1477  *
1478  * Note: the Sort node always just takes a copy of the subplan's tlist
1479  * plus ordering information.  (This might seem inefficient if the
1480  * subplan contains complex GROUP BY expressions, but in fact Sort
1481  * does not evaluate its targetlist --- it only outputs the same
1482  * tuples in a new order.  So the expressions we might be copying
1483  * are just dummies with no extra execution cost.)
1484  */
1485 static Plan *
1486 make_groupsortplan(Query *parse,
1487                                    List *groupClause,
1488                                    AttrNumber *grpColIdx,
1489                                    Plan *subplan)
1490 {
1491         List       *sort_tlist = new_unsorted_tlist(subplan->targetlist);
1492         int                     keyno = 0;
1493         List       *gl;
1494
1495         foreach(gl, groupClause)
1496         {
1497                 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1498                 TargetEntry *te = nth(grpColIdx[keyno] - 1, sort_tlist);
1499                 Resdom     *resdom = te->resdom;
1500
1501                 /*
1502                  * Check for the possibility of duplicate group-by clauses ---
1503                  * the parser should have removed 'em, but the Sort executor
1504                  * will get terribly confused if any get through!
1505                  */
1506                 if (resdom->reskey == 0)
1507                 {
1508                         /* OK, insert the ordering info needed by the executor. */
1509                         resdom->reskey = ++keyno;
1510                         resdom->reskeyop = grpcl->sortop;
1511                 }
1512         }
1513
1514         Assert(keyno > 0);
1515
1516         return (Plan *) make_sort(parse, sort_tlist, subplan, keyno);
1517 }
1518
1519 /*
1520  * postprocess_setop_tlist
1521  *        Fix up targetlist returned by plan_set_operations().
1522  *
1523  * We need to transpose sort key info from the orig_tlist into new_tlist.
1524  * NOTE: this would not be good enough if we supported resjunk sort keys
1525  * for results of set operations --- then, we'd need to project a whole
1526  * new tlist to evaluate the resjunk columns.  For now, just elog if we
1527  * find any resjunk columns in orig_tlist.
1528  */
1529 static List *
1530 postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
1531 {
1532         List       *l;
1533
1534         foreach(l, new_tlist)
1535         {
1536                 TargetEntry *new_tle = (TargetEntry *) lfirst(l);
1537                 TargetEntry *orig_tle;
1538
1539                 /* ignore resjunk columns in setop result */
1540                 if (new_tle->resdom->resjunk)
1541                         continue;
1542
1543                 Assert(orig_tlist != NIL);
1544                 orig_tle = (TargetEntry *) lfirst(orig_tlist);
1545                 orig_tlist = lnext(orig_tlist);
1546                 if (orig_tle->resdom->resjunk)
1547                         elog(ERROR, "postprocess_setop_tlist: resjunk output columns not implemented");
1548                 Assert(new_tle->resdom->resno == orig_tle->resdom->resno);
1549                 Assert(new_tle->resdom->restype == orig_tle->resdom->restype);
1550                 new_tle->resdom->ressortgroupref = orig_tle->resdom->ressortgroupref;
1551         }
1552         if (orig_tlist != NIL)
1553                 elog(ERROR, "postprocess_setop_tlist: resjunk output columns not implemented");
1554         return new_tlist;
1555 }