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) 1996-2005, PostgreSQL Global Development Group
   7  * Portions Copyright (c) 1994, Regents of the University of California
   8  *
   9  *
  10  * IDENTIFICATION
  11  *        $PostgreSQL: pgsql/src/backend/optimizer/plan/planner.c,v 1.178 2005/01/28 19:34:05 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 "executor/nodeAgg.h"
  24 #include "miscadmin.h"
  25 #include "nodes/makefuncs.h"
  26 #ifdef OPTIMIZER_DEBUG
  27 #include "nodes/print.h"
  28 #endif
  29 #include "optimizer/clauses.h"
  30 #include "optimizer/cost.h"
  31 #include "optimizer/pathnode.h"
  32 #include "optimizer/paths.h"
  33 #include "optimizer/planmain.h"
  34 #include "optimizer/planner.h"
  35 #include "optimizer/prep.h"
  36 #include "optimizer/subselect.h"
  37 #include "optimizer/tlist.h"
  38 #include "optimizer/var.h"
  39 #include "parser/analyze.h"
  40 #include "parser/parsetree.h"
  41 #include "parser/parse_expr.h"
  42 #include "parser/parse_oper.h"
  43 #include "utils/selfuncs.h"
  44 #include "utils/syscache.h"
  45
  46
  47 ParamListInfo PlannerBoundParamList = NULL;             /* current boundParams */
  48
  49
  50 /* Expression kind codes for preprocess_expression */
  51 #define EXPRKIND_QUAL   0
  52 #define EXPRKIND_TARGET 1
  53 #define EXPRKIND_RTFUNC 2
  54 #define EXPRKIND_LIMIT  3
  55 #define EXPRKIND_ININFO 4
  56
  57
  58 static Node *preprocess_expression(Query *parse, Node *expr, int kind);
  59 static void preprocess_qual_conditions(Query *parse, Node *jtnode);
  60 static Plan *inheritance_planner(Query *parse, List *inheritlist);
  61 static Plan *grouping_planner(Query *parse, double tuple_fraction);
  62 static bool hash_safe_grouping(Query *parse);
  63 static List *make_subplanTargetList(Query *parse, List *tlist,
  64                                            AttrNumber **groupColIdx, bool *need_tlist_eval);
  65 static void locate_grouping_columns(Query *parse,
  66                                                 List *tlist,
  67                                                 List *sub_tlist,
  68                                                 AttrNumber *groupColIdx);
  69 static List *postprocess_setop_tlist(List *new_tlist, List *orig_tlist);
  70
  71
  72 /*****************************************************************************
  73  *
  74  *         Query optimizer entry point
  75  *
  76  *****************************************************************************/
  77 Plan *
  78 planner(Query *parse, bool isCursor, int cursorOptions,
  79                 ParamListInfo boundParams)
  80 {
  81         double          tuple_fraction;
  82         Plan       *result_plan;
  83         Index           save_PlannerQueryLevel;
  84         List       *save_PlannerParamList;
  85         ParamListInfo save_PlannerBoundParamList;
  86
  87         /*
  88          * The planner can be called recursively (an example is when
  89          * eval_const_expressions tries to pre-evaluate an SQL function). So,
  90          * these global state variables must be saved and restored.
  91          *
  92          * Query level and the param list cannot be moved into the Query
  93          * structure since their whole purpose is communication across
  94          * multiple sub-Queries. Also, boundParams is explicitly info from
  95          * outside the Query, and so is likewise better handled as a global
  96          * variable.
  97          *
  98          * Note we do NOT save and restore PlannerPlanId: it exists to assign
  99          * unique IDs to SubPlan nodes, and we want those IDs to be unique for
 100          * the life of a backend.  Also, PlannerInitPlan is saved/restored in
 101          * subquery_planner, not here.
 102          */
 103         save_PlannerQueryLevel = PlannerQueryLevel;
 104         save_PlannerParamList = PlannerParamList;
 105         save_PlannerBoundParamList = PlannerBoundParamList;
 106
 107         /* Initialize state for handling outer-level references and params */
 108         PlannerQueryLevel = 0;          /* will be 1 in top-level subquery_planner */
 109         PlannerParamList = NIL;
 110         PlannerBoundParamList = boundParams;
 111
 112         /* Determine what fraction of the plan is likely to be scanned */
 113         if (isCursor)
 114         {
 115                 /*
 116                  * We have no real idea how many tuples the user will ultimately
 117                  * FETCH from a cursor, but it seems a good bet that he doesn't
 118                  * want 'em all.  Optimize for 10% retrieval (you gotta better
 119                  * number?      Should this be a SETtable parameter?)
 120                  */
 121                 tuple_fraction = 0.10;
 122         }
 123         else
 124         {
 125                 /* Default assumption is we need all the tuples */
 126                 tuple_fraction = 0.0;
 127         }
 128
 129         /* primary planning entry point (may recurse for subqueries) */
 130         result_plan = subquery_planner(parse, tuple_fraction);
 131
 132         Assert(PlannerQueryLevel == 0);
 133
 134         /*
 135          * If creating a plan for a scrollable cursor, make sure it can run
 136          * backwards on demand.  Add a Material node at the top at need.
 137          */
 138         if (isCursor && (cursorOptions & CURSOR_OPT_SCROLL))
 139         {
 140                 if (!ExecSupportsBackwardScan(result_plan))
 141                         result_plan = materialize_finished_plan(result_plan);
 142         }
 143
 144         /* executor wants to know total number of Params used overall */
 145         result_plan->nParamExec = list_length(PlannerParamList);
 146
 147         /* final cleanup of the plan */
 148         set_plan_references(result_plan, parse->rtable);
 149
 150         /* restore state for outer planner, if any */
 151         PlannerQueryLevel = save_PlannerQueryLevel;
 152         PlannerParamList = save_PlannerParamList;
 153         PlannerBoundParamList = save_PlannerBoundParamList;
 154
 155         return result_plan;
 156 }
 157
 158
 159 /*--------------------
 160  * subquery_planner
 161  *        Invokes the planner on a subquery.  We recurse to here for each
 162  *        sub-SELECT found in the query tree.
 163  *
 164  * parse is the querytree produced by the parser & rewriter.
 165  * tuple_fraction is the fraction of tuples we expect will be retrieved.
 166  * tuple_fraction is interpreted as explained for grouping_planner, below.
 167  *
 168  * Basically, this routine does the stuff that should only be done once
 169  * per Query object.  It then calls grouping_planner.  At one time,
 170  * grouping_planner could be invoked recursively on the same Query object;
 171  * that's not currently true, but we keep the separation between the two
 172  * routines anyway, in case we need it again someday.
 173  *
 174  * subquery_planner will be called recursively to handle sub-Query nodes
 175  * found within the query's expressions and rangetable.
 176  *
 177  * Returns a query plan.
 178  *--------------------
 179  */
 180 Plan *
 181 subquery_planner(Query *parse, double tuple_fraction)
 182 {
 183         List       *saved_initplan = PlannerInitPlan;
 184         int                     saved_planid = PlannerPlanId;
 185         bool            hasOuterJoins;
 186         Plan       *plan;
 187         List       *newHaving;
 188         List       *lst;
 189         ListCell   *l;
 190
 191         /* Set up for a new level of subquery */
 192         PlannerQueryLevel++;
 193         PlannerInitPlan = NIL;
 194
 195         /*
 196          * Look for IN clauses at the top level of WHERE, and transform them
 197          * into joins.  Note that this step only handles IN clauses originally
 198          * at top level of WHERE; if we pull up any subqueries in the next
 199          * step, their INs are processed just before pulling them up.
 200          */
 201         parse->in_info_list = NIL;
 202         if (parse->hasSubLinks)
 203                 parse->jointree->quals = pull_up_IN_clauses(parse,
 204                                                                                                  parse->jointree->quals);
 205
 206         /*
 207          * Check to see if any subqueries in the rangetable can be merged into
 208          * this query.
 209          */
 210         parse->jointree = (FromExpr *)
 211                 pull_up_subqueries(parse, (Node *) parse->jointree, false);
 212
 213         /*
 214          * Detect whether any rangetable entries are RTE_JOIN kind; if not, we
 215          * can avoid the expense of doing flatten_join_alias_vars().  Also
 216          * check for outer joins --- if none, we can skip
 217          * reduce_outer_joins(). This must be done after we have done
 218          * pull_up_subqueries, of course.
 219          */
 220         parse->hasJoinRTEs = false;
 221         hasOuterJoins = false;
 222         foreach(l, parse->rtable)
 223         {
 224                 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
 225
 226                 if (rte->rtekind == RTE_JOIN)
 227                 {
 228                         parse->hasJoinRTEs = true;
 229                         if (IS_OUTER_JOIN(rte->jointype))
 230                         {
 231                                 hasOuterJoins = true;
 232                                 /* Can quit scanning once we find an outer join */
 233                                 break;
 234                         }
 235                 }
 236         }
 237
 238         /*
 239          * Do expression preprocessing on targetlist and quals.
 240          */
 241         parse->targetList = (List *)
 242                 preprocess_expression(parse, (Node *) parse->targetList,
 243                                                           EXPRKIND_TARGET);
 244
 245         preprocess_qual_conditions(parse, (Node *) parse->jointree);
 246
 247         parse->havingQual = preprocess_expression(parse, parse->havingQual,
 248                                                                                           EXPRKIND_QUAL);
 249
 250         parse->limitOffset = preprocess_expression(parse, parse->limitOffset,
 251                                                                                            EXPRKIND_LIMIT);
 252         parse->limitCount = preprocess_expression(parse, parse->limitCount,
 253                                                                                           EXPRKIND_LIMIT);
 254
 255         parse->in_info_list = (List *)
 256                 preprocess_expression(parse, (Node *) parse->in_info_list,
 257                                                           EXPRKIND_ININFO);
 258
 259         /* Also need to preprocess expressions for function RTEs */
 260         foreach(l, parse->rtable)
 261         {
 262                 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
 263
 264                 if (rte->rtekind == RTE_FUNCTION)
 265                         rte->funcexpr = preprocess_expression(parse, rte->funcexpr,
 266                                                                                                   EXPRKIND_RTFUNC);
 267         }
 268
 269         /*
 270          * A HAVING clause without aggregates is equivalent to a WHERE clause
 271          * (except it can only refer to grouped fields).  Transfer any
 272          * agg-free clauses of the HAVING qual into WHERE.      This may seem like
 273          * wasting cycles to cater to stupidly-written queries, but there are
 274          * other reasons for doing it.  Firstly, if the query contains no aggs
 275          * at all, then we aren't going to generate an Agg plan node, and so
 276          * there'll be no place to execute HAVING conditions; without this
 277          * transfer, we'd lose the HAVING condition entirely, which is wrong.
 278          * Secondly, when we push down a qual condition into a sub-query, it's
 279          * easiest to push the qual into HAVING always, in case it contains
 280          * aggs, and then let this code sort it out.
 281          *
 282          * Note that both havingQual and parse->jointree->quals are in
 283          * implicitly-ANDed-list form at this point, even though they are
 284          * declared as Node *.
 285          */
 286         newHaving = NIL;
 287         foreach(l, (List *) parse->havingQual)
 288         {
 289                 Node       *havingclause = (Node *) lfirst(l);
 290
 291                 if (contain_agg_clause(havingclause))
 292                         newHaving = lappend(newHaving, havingclause);
 293                 else
 294                         parse->jointree->quals = (Node *)
 295                                 lappend((List *) parse->jointree->quals, havingclause);
 296         }
 297         parse->havingQual = (Node *) newHaving;
 298
 299         /*
 300          * If we have any outer joins, try to reduce them to plain inner
 301          * joins. This step is most easily done after we've done expression
 302          * preprocessing.
 303          */
 304         if (hasOuterJoins)
 305                 reduce_outer_joins(parse);
 306
 307         /*
 308          * See if we can simplify the jointree; opportunities for this may
 309          * come from having pulled up subqueries, or from flattening explicit
 310          * JOIN syntax.  We must do this after flattening JOIN alias
 311          * variables, since eliminating explicit JOIN nodes from the jointree
 312          * will cause get_relids_for_join() to fail.  But it should happen
 313          * after reduce_outer_joins, anyway.
 314          */
 315         parse->jointree = (FromExpr *)
 316                 simplify_jointree(parse, (Node *) parse->jointree);
 317
 318         /*
 319          * Do the main planning.  If we have an inherited target relation,
 320          * that needs special processing, else go straight to
 321          * grouping_planner.
 322          */
 323         if (parse->resultRelation &&
 324                 (lst = expand_inherited_rtentry(parse, parse->resultRelation)) != NIL)
 325                 plan = inheritance_planner(parse, lst);
 326         else
 327                 plan = grouping_planner(parse, tuple_fraction);
 328
 329         /*
 330          * If any subplans were generated, or if we're inside a subplan, build
 331          * initPlan list and extParam/allParam sets for plan nodes.
 332          */
 333         if (PlannerPlanId != saved_planid || PlannerQueryLevel > 1)
 334         {
 335                 Cost            initplan_cost = 0;
 336
 337                 /* Prepare extParam/allParam sets for all nodes in tree */
 338                 SS_finalize_plan(plan, parse->rtable);
 339
 340                 /*
 341                  * SS_finalize_plan doesn't handle initPlans, so we have to
 342                  * manually attach them to the topmost plan node, and add their
 343                  * extParams to the topmost node's, too.
 344                  *
 345                  * We also add the total_cost of each initPlan to the startup cost of
 346                  * the top node.  This is a conservative overestimate, since in
 347                  * fact each initPlan might be executed later than plan startup,
 348                  * or even not at all.
 349                  */
 350                 plan->initPlan = PlannerInitPlan;
 351
 352                 foreach(l, plan->initPlan)
 353                 {
 354                         SubPlan    *initplan = (SubPlan *) lfirst(l);
 355
 356                         plan->extParam = bms_add_members(plan->extParam,
 357                                                                                          initplan->plan->extParam);
 358                         /* allParam must include all members of extParam */
 359                         plan->allParam = bms_add_members(plan->allParam,
 360                                                                                          plan->extParam);
 361                         initplan_cost += initplan->plan->total_cost;
 362                 }
 363
 364                 plan->startup_cost += initplan_cost;
 365                 plan->total_cost += initplan_cost;
 366         }
 367
 368         /* Return to outer subquery context */
 369         PlannerQueryLevel--;
 370         PlannerInitPlan = saved_initplan;
 371         /* we do NOT restore PlannerPlanId; that's not an oversight! */
 372
 373         return plan;
 374 }
 375
 376 /*
 377  * preprocess_expression
 378  *              Do subquery_planner's preprocessing work for an expression,
 379  *              which can be a targetlist, a WHERE clause (including JOIN/ON
 380  *              conditions), or a HAVING clause.
 381  */
 382 static Node *
 383 preprocess_expression(Query *parse, Node *expr, int kind)
 384 {
 385         /*
 386          * If the query has any join RTEs, replace join alias variables with
 387          * base-relation variables. We must do this before sublink processing,
 388          * else sublinks expanded out from join aliases wouldn't get
 389          * processed.
 390          */
 391         if (parse->hasJoinRTEs)
 392                 expr = flatten_join_alias_vars(parse, expr);
 393
 394         /*
 395          * If it's a qual or havingQual, canonicalize it.  It seems most
 396          * useful to do this before applying eval_const_expressions, since the
 397          * latter can optimize flattened AND/ORs better than unflattened ones.
 398          *
 399          * Note: all processing of a qual expression after this point must be
 400          * careful to maintain AND/OR flatness --- that is, do not generate a
 401          * tree with AND directly under AND, nor OR directly under OR.
 402          */
 403         if (kind == EXPRKIND_QUAL)
 404         {
 405                 expr = (Node *) canonicalize_qual((Expr *) expr);
 406
 407 #ifdef OPTIMIZER_DEBUG
 408                 printf("After canonicalize_qual()\n");
 409                 pprint(expr);
 410 #endif
 411         }
 412
 413         /*
 414          * Simplify constant expressions.
 415          */
 416         expr = eval_const_expressions(expr);
 417
 418         /* Expand SubLinks to SubPlans */
 419         if (parse->hasSubLinks)
 420                 expr = SS_process_sublinks(expr, (kind == EXPRKIND_QUAL));
 421
 422         /*
 423          * XXX do not insert anything here unless you have grokked the
 424          * comments in SS_replace_correlation_vars ...
 425          */
 426
 427         /* Replace uplevel vars with Param nodes */
 428         if (PlannerQueryLevel > 1)
 429                 expr = SS_replace_correlation_vars(expr);
 430
 431         /*
 432          * If it's a qual or havingQual, convert it to implicit-AND format.
 433          * (We don't want to do this before eval_const_expressions, since the
 434          * latter would be unable to simplify a top-level AND correctly. Also,
 435          * SS_process_sublinks expects explicit-AND format.)
 436          */
 437         if (kind == EXPRKIND_QUAL)
 438                 expr = (Node *) make_ands_implicit((Expr *) expr);
 439
 440         return expr;
 441 }
 442
 443 /*
 444  * preprocess_qual_conditions
 445  *              Recursively scan the query's jointree and do subquery_planner's
 446  *              preprocessing work on each qual condition found therein.
 447  */
 448 static void
 449 preprocess_qual_conditions(Query *parse, Node *jtnode)
 450 {
 451         if (jtnode == NULL)
 452                 return;
 453         if (IsA(jtnode, RangeTblRef))
 454         {
 455                 /* nothing to do here */
 456         }
 457         else if (IsA(jtnode, FromExpr))
 458         {
 459                 FromExpr   *f = (FromExpr *) jtnode;
 460                 ListCell   *l;
 461
 462                 foreach(l, f->fromlist)
 463                         preprocess_qual_conditions(parse, lfirst(l));
 464
 465                 f->quals = preprocess_expression(parse, f->quals, EXPRKIND_QUAL);
 466         }
 467         else if (IsA(jtnode, JoinExpr))
 468         {
 469                 JoinExpr   *j = (JoinExpr *) jtnode;
 470
 471                 preprocess_qual_conditions(parse, j->larg);
 472                 preprocess_qual_conditions(parse, j->rarg);
 473
 474                 j->quals = preprocess_expression(parse, j->quals, EXPRKIND_QUAL);
 475         }
 476         else
 477                 elog(ERROR, "unrecognized node type: %d",
 478                          (int) nodeTag(jtnode));
 479 }
 480
 481 /*--------------------
 482  * inheritance_planner
 483  *        Generate a plan in the case where the result relation is an
 484  *        inheritance set.
 485  *
 486  * We have to handle this case differently from cases where a source
 487  * relation is an inheritance set.      Source inheritance is expanded at
 488  * the bottom of the plan tree (see allpaths.c), but target inheritance
 489  * has to be expanded at the top.  The reason is that for UPDATE, each
 490  * target relation needs a different targetlist matching its own column
 491  * set.  (This is not so critical for DELETE, but for simplicity we treat
 492  * inherited DELETE the same way.)      Fortunately, the UPDATE/DELETE target
 493  * can never be the nullable side of an outer join, so it's OK to generate
 494  * the plan this way.
 495  *
 496  * parse is the querytree produced by the parser & rewriter.
 497  * inheritlist is an integer list of RT indexes for the result relation set.
 498  *
 499  * Returns a query plan.
 500  *--------------------
 501  */
 502 static Plan *
 503 inheritance_planner(Query *parse, List *inheritlist)
 504 {
 505         int                     parentRTindex = parse->resultRelation;
 506         Oid                     parentOID = getrelid(parentRTindex, parse->rtable);
 507         int                     mainrtlength = list_length(parse->rtable);
 508         List       *subplans = NIL;
 509         List       *tlist = NIL;
 510         ListCell   *l;
 511
 512         foreach(l, inheritlist)
 513         {
 514                 int                     childRTindex = lfirst_int(l);
 515                 Oid                     childOID = getrelid(childRTindex, parse->rtable);
 516                 Query      *subquery;
 517                 Plan       *subplan;
 518
 519                 /* Generate modified query with this rel as target */
 520                 subquery = (Query *) adjust_inherited_attrs((Node *) parse,
 521                                                                                                 parentRTindex, parentOID,
 522                                                                                                  childRTindex, childOID);
 523                 /* Generate plan */
 524                 subplan = grouping_planner(subquery, 0.0 /* retrieve all tuples */ );
 525                 subplans = lappend(subplans, subplan);
 526
 527                 /*
 528                  * XXX my goodness this next bit is ugly.  Really need to think about
 529                  * ways to rein in planner's habit of scribbling on its input.
 530                  *
 531                  * Planning of the subquery might have modified the rangetable,
 532                  * either by addition of RTEs due to expansion of inherited source
 533                  * tables, or by changes of the Query structures inside subquery
 534                  * RTEs.  We have to ensure that this gets propagated back to the
 535                  * master copy.  However, if we aren't done planning yet, we also
 536                  * need to ensure that subsequent calls to grouping_planner have
 537                  * virgin sub-Queries to work from.  So, if we are at the last
 538                  * list entry, just copy the subquery rangetable back to the master
 539                  * copy; if we are not, then extend the master copy by adding
 540                  * whatever the subquery added.  (We assume these added entries
 541                  * will go untouched by the future grouping_planner calls.  We are
 542                  * also effectively assuming that sub-Queries will get planned
 543                  * identically each time, or at least that the impacts on their
 544                  * rangetables will be the same each time.  Did I say this is ugly?)
 545                  */
 546                 if (lnext(l) == NULL)
 547                         parse->rtable = subquery->rtable;
 548                 else
 549                 {
 550                         int             subrtlength = list_length(subquery->rtable);
 551
 552                         if (subrtlength > mainrtlength)
 553                         {
 554                                 List       *subrt;
 555
 556                                 subrt = list_copy_tail(subquery->rtable, mainrtlength);
 557                                 parse->rtable = list_concat(parse->rtable, subrt);
 558                                 mainrtlength = subrtlength;
 559                         }
 560                 }
 561
 562                 /* Save preprocessed tlist from first rel for use in Append */
 563                 if (tlist == NIL)
 564                         tlist = subplan->targetlist;
 565         }
 566
 567         /* Save the target-relations list for the executor, too */
 568         parse->resultRelations = inheritlist;
 569
 570         /* Mark result as unordered (probably unnecessary) */
 571         parse->query_pathkeys = NIL;
 572
 573         return (Plan *) make_append(subplans, true, tlist);
 574 }
 575
 576 /*--------------------
 577  * grouping_planner
 578  *        Perform planning steps related to grouping, aggregation, etc.
 579  *        This primarily means adding top-level processing to the basic
 580  *        query plan produced by query_planner.
 581  *
 582  * parse is the querytree produced by the parser & rewriter.
 583  * tuple_fraction is the fraction of tuples we expect will be retrieved
 584  *
 585  * tuple_fraction is interpreted as follows:
 586  *        0: expect all tuples to be retrieved (normal case)
 587  *        0 < tuple_fraction < 1: expect the given fraction of tuples available
 588  *              from the plan to be retrieved
 589  *        tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
 590  *              expected to be retrieved (ie, a LIMIT specification)
 591  *
 592  * Returns a query plan.  Also, parse->query_pathkeys is returned as the
 593  * actual output ordering of the plan (in pathkey format).
 594  *--------------------
 595  */
 596 static Plan *
 597 grouping_planner(Query *parse, double tuple_fraction)
 598 {
 599         List       *tlist = parse->targetList;
 600         Plan       *result_plan;
 601         List       *current_pathkeys;
 602         List       *sort_pathkeys;
 603
 604         if (parse->setOperations)
 605         {
 606                 List       *set_sortclauses;
 607
 608                 /*
 609                  * Construct the plan for set operations.  The result will not
 610                  * need any work except perhaps a top-level sort and/or LIMIT.
 611                  */
 612                 result_plan = plan_set_operations(parse,
 613                                                                                   &set_sortclauses);
 614
 615                 /*
 616                  * Calculate pathkeys representing the sort order (if any) of the
 617                  * set operation's result.  We have to do this before overwriting
 618                  * the sort key information...
 619                  */
 620                 current_pathkeys = make_pathkeys_for_sortclauses(set_sortclauses,
 621                                                                                                 result_plan->targetlist);
 622                 current_pathkeys = canonicalize_pathkeys(parse, current_pathkeys);
 623
 624                 /*
 625                  * We should not need to call preprocess_targetlist, since we must
 626                  * be in a SELECT query node.  Instead, use the targetlist
 627                  * returned by plan_set_operations (since this tells whether it
 628                  * returned any resjunk columns!), and transfer any sort key
 629                  * information from the original tlist.
 630                  */
 631                 Assert(parse->commandType == CMD_SELECT);
 632
 633                 tlist = postprocess_setop_tlist(result_plan->targetlist, tlist);
 634
 635                 /*
 636                  * Can't handle FOR UPDATE here (parser should have checked
 637                  * already, but let's make sure).
 638                  */
 639                 if (parse->rowMarks)
 640                         ereport(ERROR,
 641                                         (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
 642                                          errmsg("SELECT FOR UPDATE is not allowed with UNION/INTERSECT/EXCEPT")));
 643
 644                 /*
 645                  * Calculate pathkeys that represent result ordering requirements
 646                  */
 647                 sort_pathkeys = make_pathkeys_for_sortclauses(parse->sortClause,
 648                                                                                                           tlist);
 649                 sort_pathkeys = canonicalize_pathkeys(parse, sort_pathkeys);
 650         }
 651         else
 652         {
 653                 /* No set operations, do regular planning */
 654                 List       *sub_tlist;
 655                 List       *group_pathkeys;
 656                 AttrNumber *groupColIdx = NULL;
 657                 bool            need_tlist_eval = true;
 658                 QualCost        tlist_cost;
 659                 double          sub_tuple_fraction;
 660                 Path       *cheapest_path;
 661                 Path       *sorted_path;
 662                 double          dNumGroups = 0;
 663                 long            numGroups = 0;
 664                 AggClauseCounts agg_counts;
 665                 int                     numGroupCols = list_length(parse->groupClause);
 666                 bool            use_hashed_grouping = false;
 667
 668                 MemSet(&agg_counts, 0, sizeof(AggClauseCounts));
 669
 670                 /* Preprocess targetlist in case we are inside an INSERT/UPDATE. */
 671                 tlist = preprocess_targetlist(tlist,
 672                                                                           parse->commandType,
 673                                                                           parse->resultRelation,
 674                                                                           parse->rtable);
 675
 676                 /*
 677                  * Add TID targets for rels selected FOR UPDATE (should this be
 678                  * done in preprocess_targetlist?).  The executor uses the TID to
 679                  * know which rows to lock, much as for UPDATE or DELETE.
 680                  */
 681                 if (parse->rowMarks)
 682                 {
 683                         ListCell   *l;
 684
 685                         /*
 686                          * We've got trouble if the FOR UPDATE appears inside
 687                          * grouping, since grouping renders a reference to individual
 688                          * tuple CTIDs invalid.  This is also checked at parse time,
 689                          * but that's insufficient because of rule substitution, query
 690                          * pullup, etc.
 691                          */
 692                         CheckSelectForUpdate(parse);
 693
 694                         /*
 695                          * Currently the executor only supports FOR UPDATE at top
 696                          * level
 697                          */
 698                         if (PlannerQueryLevel > 1)
 699                                 ereport(ERROR,
 700                                                 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
 701                                                  errmsg("SELECT FOR UPDATE is not allowed in subqueries")));
 702
 703                         foreach(l, parse->rowMarks)
 704                         {
 705                                 Index           rti = lfirst_int(l);
 706                                 char       *resname;
 707                                 Resdom     *resdom;
 708                                 Var                *var;
 709                                 TargetEntry *ctid;
 710
 711                                 resname = (char *) palloc(32);
 712                                 snprintf(resname, 32, "ctid%u", rti);
 713                                 resdom = makeResdom(list_length(tlist) + 1,
 714                                                                         TIDOID,
 715                                                                         -1,
 716                                                                         resname,
 717                                                                         true);
 718
 719                                 var = makeVar(rti,
 720                                                           SelfItemPointerAttributeNumber,
 721                                                           TIDOID,
 722                                                           -1,
 723                                                           0);
 724
 725                                 ctid = makeTargetEntry(resdom, (Expr *) var);
 726                                 tlist = lappend(tlist, ctid);
 727                         }
 728                 }
 729
 730                 /*
 731                  * Generate appropriate target list for subplan; may be different
 732                  * from tlist if grouping or aggregation is needed.
 733                  */
 734                 sub_tlist = make_subplanTargetList(parse, tlist,
 735                                                                                  &groupColIdx, &need_tlist_eval);
 736
 737                 /*
 738                  * Calculate pathkeys that represent grouping/ordering
 739                  * requirements
 740                  */
 741                 group_pathkeys = make_pathkeys_for_sortclauses(parse->groupClause,
 742                                                                                                            tlist);
 743                 sort_pathkeys = make_pathkeys_for_sortclauses(parse->sortClause,
 744                                                                                                           tlist);
 745
 746                 /*
 747                  * Will need actual number of aggregates for estimating costs.
 748                  *
 749                  * Note: we do not attempt to detect duplicate aggregates here; a
 750                  * somewhat-overestimated count is okay for our present purposes.
 751                  *
 752                  * Note: think not that we can turn off hasAggs if we find no aggs.
 753                  * It is possible for constant-expression simplification to remove
 754                  * all explicit references to aggs, but we still have to follow
 755                  * the aggregate semantics (eg, producing only one output row).
 756                  */
 757                 if (parse->hasAggs)
 758                 {
 759                         count_agg_clauses((Node *) tlist, &agg_counts);
 760                         count_agg_clauses(parse->havingQual, &agg_counts);
 761                 }
 762
 763                 /*
 764                  * Figure out whether we need a sorted result from query_planner.
 765                  *
 766                  * If we have a GROUP BY clause, then we want a result sorted
 767                  * properly for grouping.  Otherwise, if there is an ORDER BY
 768                  * clause, we want to sort by the ORDER BY clause.      (Note: if we
 769                  * have both, and ORDER BY is a superset of GROUP BY, it would be
 770                  * tempting to request sort by ORDER BY --- but that might just
 771                  * leave us failing to exploit an available sort order at all.
 772                  * Needs more thought...)
 773                  */
 774                 if (parse->groupClause)
 775                         parse->query_pathkeys = group_pathkeys;
 776                 else if (parse->sortClause)
 777                         parse->query_pathkeys = sort_pathkeys;
 778                 else
 779                         parse->query_pathkeys = NIL;
 780
 781                 /*
 782                  * Adjust tuple_fraction if we see that we are going to apply
 783                  * limiting/grouping/aggregation/etc.  This is not overridable by
 784                  * the caller, since it reflects plan actions that this routine
 785                  * will certainly take, not assumptions about context.
 786                  */
 787                 if (parse->limitCount != NULL)
 788                 {
 789                         /*
 790                          * A LIMIT clause limits the absolute number of tuples
 791                          * returned. However, if it's not a constant LIMIT then we
 792                          * have to punt; for lack of a better idea, assume 10% of the
 793                          * plan's result is wanted.
 794                          */
 795                         double          limit_fraction = 0.0;
 796
 797                         if (IsA(parse->limitCount, Const))
 798                         {
 799                                 Const      *limitc = (Const *) parse->limitCount;
 800                                 int32           count = DatumGetInt32(limitc->constvalue);
 801
 802                                 /*
 803                                  * A NULL-constant LIMIT represents "LIMIT ALL", which we
 804                                  * treat the same as no limit (ie, expect to retrieve all
 805                                  * the tuples).
 806                                  */
 807                                 if (!limitc->constisnull && count > 0)
 808                                 {
 809                                         limit_fraction = (double) count;
 810                                         /* We must also consider the OFFSET, if present */
 811                                         if (parse->limitOffset != NULL)
 812                                         {
 813                                                 if (IsA(parse->limitOffset, Const))
 814                                                 {
 815                                                         int32           offset;
 816
 817                                                         limitc = (Const *) parse->limitOffset;
 818                                                         offset = DatumGetInt32(limitc->constvalue);
 819                                                         if (!limitc->constisnull && offset > 0)
 820                                                                 limit_fraction += (double) offset;
 821                                                 }
 822                                                 else
 823                                                 {
 824                                                         /* OFFSET is an expression ... punt ... */
 825                                                         limit_fraction = 0.10;
 826                                                 }
 827                                         }
 828                                 }
 829                         }
 830                         else
 831                         {
 832                                 /* LIMIT is an expression ... punt ... */
 833                                 limit_fraction = 0.10;
 834                         }
 835
 836                         if (limit_fraction > 0.0)
 837                         {
 838                                 /*
 839                                  * If we have absolute limits from both caller and LIMIT,
 840                                  * use the smaller value; if one is fractional and the
 841                                  * other absolute, treat the fraction as a fraction of the
 842                                  * absolute value; else we can multiply the two fractions
 843                                  * together.
 844                                  */
 845                                 if (tuple_fraction >= 1.0)
 846                                 {
 847                                         if (limit_fraction >= 1.0)
 848                                         {
 849                                                 /* both absolute */
 850                                                 tuple_fraction = Min(tuple_fraction, limit_fraction);
 851                                         }
 852                                         else
 853                                         {
 854                                                 /* caller absolute, limit fractional */
 855                                                 tuple_fraction *= limit_fraction;
 856                                                 if (tuple_fraction < 1.0)
 857                                                         tuple_fraction = 1.0;
 858                                         }
 859                                 }
 860                                 else if (tuple_fraction > 0.0)
 861                                 {
 862                                         if (limit_fraction >= 1.0)
 863                                         {
 864                                                 /* caller fractional, limit absolute */
 865                                                 tuple_fraction *= limit_fraction;
 866                                                 if (tuple_fraction < 1.0)
 867                                                         tuple_fraction = 1.0;
 868                                         }
 869                                         else
 870                                         {
 871                                                 /* both fractional */
 872                                                 tuple_fraction *= limit_fraction;
 873                                         }
 874                                 }
 875                                 else
 876                                 {
 877                                         /* no info from caller, just use limit */
 878                                         tuple_fraction = limit_fraction;
 879                                 }
 880                         }
 881                 }
 882
 883                 /*
 884                  * With grouping or aggregation, the tuple fraction to pass to
 885                  * query_planner() may be different from what it is at top level.
 886                  */
 887                 sub_tuple_fraction = tuple_fraction;
 888
 889                 if (parse->groupClause)
 890                 {
 891                         /*
 892                          * In GROUP BY mode, we have the little problem that we don't
 893                          * really know how many input tuples will be needed to make a
 894                          * group, so we can't translate an output LIMIT count into an
 895                          * input count.  For lack of a better idea, assume 25% of the
 896                          * input data will be processed if there is any output limit.
 897                          * However, if the caller gave us a fraction rather than an
 898                          * absolute count, we can keep using that fraction (which
 899                          * amounts to assuming that all the groups are about the same
 900                          * size).
 901                          */
 902                         if (sub_tuple_fraction >= 1.0)
 903                                 sub_tuple_fraction = 0.25;
 904
 905                         /*
 906                          * If both GROUP BY and ORDER BY are specified, we will need
 907                          * two levels of sort --- and, therefore, certainly need to
 908                          * read all the input tuples --- unless ORDER BY is a subset
 909                          * of GROUP BY.  (We have not yet canonicalized the pathkeys,
 910                          * so must use the slower noncanonical comparison method.)
 911                          */
 912                         if (parse->groupClause && parse->sortClause &&
 913                                 !noncanonical_pathkeys_contained_in(sort_pathkeys,
 914                                                                                                         group_pathkeys))
 915                                 sub_tuple_fraction = 0.0;
 916                 }
 917                 else if (parse->hasAggs)
 918                 {
 919                         /*
 920                          * Ungrouped aggregate will certainly want all the input
 921                          * tuples.
 922                          */
 923                         sub_tuple_fraction = 0.0;
 924                 }
 925                 else if (parse->distinctClause)
 926                 {
 927                         /*
 928                          * SELECT DISTINCT, like GROUP, will absorb an unpredictable
 929                          * number of input tuples per output tuple.  Handle the same
 930                          * way.
 931                          */
 932                         if (sub_tuple_fraction >= 1.0)
 933                                 sub_tuple_fraction = 0.25;
 934                 }
 935
 936                 /*
 937                  * Generate the best unsorted and presorted paths for this Query
 938                  * (but note there may not be any presorted path).
 939                  */
 940                 query_planner(parse, sub_tlist, sub_tuple_fraction,
 941                                           &cheapest_path, &sorted_path);
 942
 943                 /*
 944                  * We couldn't canonicalize group_pathkeys and sort_pathkeys
 945                  * before running query_planner(), so do it now.
 946                  */
 947                 group_pathkeys = canonicalize_pathkeys(parse, group_pathkeys);
 948                 sort_pathkeys = canonicalize_pathkeys(parse, sort_pathkeys);
 949
 950                 /*
 951                  * Consider whether we might want to use hashed grouping.
 952                  */
 953                 if (parse->groupClause)
 954                 {
 955                         List       *groupExprs;
 956                         double          cheapest_path_rows;
 957                         int                     cheapest_path_width;
 958
 959                         /*
 960                          * Beware in this section of the possibility that
 961                          * cheapest_path->parent is NULL.  This could happen if user
 962                          * does something silly like SELECT 'foo' GROUP BY 1;
 963                          */
 964                         if (cheapest_path->parent)
 965                         {
 966                                 cheapest_path_rows = cheapest_path->parent->rows;
 967                                 cheapest_path_width = cheapest_path->parent->width;
 968                         }
 969                         else
 970                         {
 971                                 cheapest_path_rows = 1; /* assume non-set result */
 972                                 cheapest_path_width = 100;              /* arbitrary */
 973                         }
 974
 975                         /*
 976                          * Always estimate the number of groups.  We can't do this
 977                          * until after running query_planner(), either.
 978                          */
 979                         groupExprs = get_sortgrouplist_exprs(parse->groupClause,
 980                                                                                                  parse->targetList);
 981                         dNumGroups = estimate_num_groups(parse,
 982                                                                                          groupExprs,
 983                                                                                          cheapest_path_rows);
 984                         /* Also want it as a long int --- but 'ware overflow! */
 985                         numGroups = (long) Min(dNumGroups, (double) LONG_MAX);
 986
 987                         /*
 988                          * Check can't-do-it conditions, including whether the
 989                          * grouping operators are hashjoinable.
 990                          *
 991                          * Executor doesn't support hashed aggregation with DISTINCT
 992                          * aggregates.  (Doing so would imply storing *all* the input
 993                          * values in the hash table, which seems like a certain
 994                          * loser.)
 995                          */
 996                         if (!enable_hashagg || !hash_safe_grouping(parse))
 997                                 use_hashed_grouping = false;
 998                         else if (agg_counts.numDistinctAggs != 0)
 999                                 use_hashed_grouping = false;
1000                         else
1001                         {
1002                                 /*
1003                                  * Use hashed grouping if (a) we think we can fit the
1004                                  * hashtable into work_mem, *and* (b) the estimated cost
1005                                  * is no more than doing it the other way.      While avoiding
1006                                  * the need for sorted input is usually a win, the fact
1007                                  * that the output won't be sorted may be a loss; so we
1008                                  * need to do an actual cost comparison.
1009                                  */
1010                                 Size            hashentrysize;
1011
1012                                 /* Estimate per-hash-entry space at tuple width... */
1013                                 hashentrysize = cheapest_path_width;
1014                                 /* plus space for pass-by-ref transition values... */
1015                                 hashentrysize += agg_counts.transitionSpace;
1016                                 /* plus the per-hash-entry overhead */
1017                                 hashentrysize += hash_agg_entry_size(agg_counts.numAggs);
1018
1019                                 if (hashentrysize * dNumGroups <= work_mem * 1024L)
1020                                 {
1021                                         /*
1022                                          * Okay, do the cost comparison.  We need to consider
1023                                          * cheapest_path + hashagg [+ final sort] versus
1024                                          * either cheapest_path [+ sort] + group or agg [+
1025                                          * final sort] or presorted_path + group or agg [+
1026                                          * final sort] where brackets indicate a step that may
1027                                          * not be needed. We assume query_planner() will have
1028                                          * returned a presorted path only if it's a winner
1029                                          * compared to cheapest_path for this purpose.
1030                                          *
1031                                          * These path variables are dummies that just hold cost
1032                                          * fields; we don't make actual Paths for these steps.
1033                                          */
1034                                         Path            hashed_p;
1035                                         Path            sorted_p;
1036
1037                                         cost_agg(&hashed_p, parse,
1038                                                          AGG_HASHED, agg_counts.numAggs,
1039                                                          numGroupCols, dNumGroups,
1040                                                          cheapest_path->startup_cost,
1041                                                          cheapest_path->total_cost,
1042                                                          cheapest_path_rows);
1043                                         /* Result of hashed agg is always unsorted */
1044                                         if (sort_pathkeys)
1045                                                 cost_sort(&hashed_p, parse, sort_pathkeys,
1046                                                                   hashed_p.total_cost,
1047                                                                   dNumGroups,
1048                                                                   cheapest_path_width);
1049
1050                                         if (sorted_path)
1051                                         {
1052                                                 sorted_p.startup_cost = sorted_path->startup_cost;
1053                                                 sorted_p.total_cost = sorted_path->total_cost;
1054                                                 current_pathkeys = sorted_path->pathkeys;
1055                                         }
1056                                         else
1057                                         {
1058                                                 sorted_p.startup_cost = cheapest_path->startup_cost;
1059                                                 sorted_p.total_cost = cheapest_path->total_cost;
1060                                                 current_pathkeys = cheapest_path->pathkeys;
1061                                         }
1062                                         if (!pathkeys_contained_in(group_pathkeys,
1063                                                                                            current_pathkeys))
1064                                         {
1065                                                 cost_sort(&sorted_p, parse, group_pathkeys,
1066                                                                   sorted_p.total_cost,
1067                                                                   cheapest_path_rows,
1068                                                                   cheapest_path_width);
1069                                                 current_pathkeys = group_pathkeys;
1070                                         }
1071                                         if (parse->hasAggs)
1072                                                 cost_agg(&sorted_p, parse,
1073                                                                  AGG_SORTED, agg_counts.numAggs,
1074                                                                  numGroupCols, dNumGroups,
1075                                                                  sorted_p.startup_cost,
1076                                                                  sorted_p.total_cost,
1077                                                                  cheapest_path_rows);
1078                                         else
1079                                                 cost_group(&sorted_p, parse,
1080                                                                    numGroupCols, dNumGroups,
1081                                                                    sorted_p.startup_cost,
1082                                                                    sorted_p.total_cost,
1083                                                                    cheapest_path_rows);
1084                                         /* The Agg or Group node will preserve ordering */
1085                                         if (sort_pathkeys &&
1086                                                 !pathkeys_contained_in(sort_pathkeys,
1087                                                                                            current_pathkeys))
1088                                         {
1089                                                 cost_sort(&sorted_p, parse, sort_pathkeys,
1090                                                                   sorted_p.total_cost,
1091                                                                   dNumGroups,
1092                                                                   cheapest_path_width);
1093                                         }
1094
1095                                         /*
1096                                          * Now make the decision using the top-level tuple
1097                                          * fraction.  First we have to convert an absolute
1098                                          * count (LIMIT) into fractional form.
1099                                          */
1100                                         if (tuple_fraction >= 1.0)
1101                                                 tuple_fraction /= dNumGroups;
1102
1103                                         if (compare_fractional_path_costs(&hashed_p, &sorted_p,
1104                                                                                                           tuple_fraction) < 0)
1105                                         {
1106                                                 /* Hashed is cheaper, so use it */
1107                                                 use_hashed_grouping = true;
1108                                         }
1109                                 }
1110                         }
1111                 }
1112
1113                 /*
1114                  * Select the best path and create a plan to execute it.
1115                  *
1116                  * If we are doing hashed grouping, we will always read all the input
1117                  * tuples, so use the cheapest-total path.      Otherwise, trust
1118                  * query_planner's decision about which to use.
1119                  */
1120                 if (sorted_path && !use_hashed_grouping)
1121                 {
1122                         result_plan = create_plan(parse, sorted_path);
1123                         current_pathkeys = sorted_path->pathkeys;
1124                 }
1125                 else
1126                 {
1127                         result_plan = create_plan(parse, cheapest_path);
1128                         current_pathkeys = cheapest_path->pathkeys;
1129                 }
1130
1131                 /*
1132                  * create_plan() returns a plan with just a "flat" tlist of
1133                  * required Vars.  Usually we need to insert the sub_tlist as the
1134                  * tlist of the top plan node.  However, we can skip that if we
1135                  * determined that whatever query_planner chose to return will be
1136                  * good enough.
1137                  */
1138                 if (need_tlist_eval)
1139                 {
1140                         /*
1141                          * If the top-level plan node is one that cannot do expression
1142                          * evaluation, we must insert a Result node to project the
1143                          * desired tlist.
1144                          */
1145                         if (!is_projection_capable_plan(result_plan))
1146                         {
1147                                 result_plan = (Plan *) make_result(sub_tlist, NULL,
1148                                                                                                    result_plan);
1149                         }
1150                         else
1151                         {
1152                                 /*
1153                                  * Otherwise, just replace the subplan's flat tlist with
1154                                  * the desired tlist.
1155                                  */
1156                                 result_plan->targetlist = sub_tlist;
1157                         }
1158
1159                         /*
1160                          * Also, account for the cost of evaluation of the sub_tlist.
1161                          *
1162                          * Up to now, we have only been dealing with "flat" tlists,
1163                          * containing just Vars.  So their evaluation cost is zero
1164                          * according to the model used by cost_qual_eval() (or if you
1165                          * prefer, the cost is factored into cpu_tuple_cost).  Thus we
1166                          * can avoid accounting for tlist cost throughout
1167                          * query_planner() and subroutines.  But now we've inserted a
1168                          * tlist that might contain actual operators, sub-selects, etc
1169                          * --- so we'd better account for its cost.
1170                          *
1171                          * Below this point, any tlist eval cost for added-on nodes
1172                          * should be accounted for as we create those nodes.
1173                          * Presently, of the node types we can add on, only Agg and
1174                          * Group project new tlists (the rest just copy their input
1175                          * tuples) --- so make_agg() and make_group() are responsible
1176                          * for computing the added cost.
1177                          */
1178                         cost_qual_eval(&tlist_cost, sub_tlist);
1179                         result_plan->startup_cost += tlist_cost.startup;
1180                         result_plan->total_cost += tlist_cost.startup +
1181                                 tlist_cost.per_tuple * result_plan->plan_rows;
1182                 }
1183                 else
1184                 {
1185                         /*
1186                          * Since we're using query_planner's tlist and not the one
1187                          * make_subplanTargetList calculated, we have to refigure any
1188                          * grouping-column indexes make_subplanTargetList computed.
1189                          */
1190                         locate_grouping_columns(parse, tlist, result_plan->targetlist,
1191                                                                         groupColIdx);
1192                 }
1193
1194                 /*
1195                  * Insert AGG or GROUP node if needed, plus an explicit sort step
1196                  * if necessary.
1197                  *
1198                  * HAVING clause, if any, becomes qual of the Agg node
1199                  */
1200                 if (use_hashed_grouping)
1201                 {
1202                         /* Hashed aggregate plan --- no sort needed */
1203                         result_plan = (Plan *) make_agg(parse,
1204                                                                                         tlist,
1205                                                                                         (List *) parse->havingQual,
1206                                                                                         AGG_HASHED,
1207                                                                                         numGroupCols,
1208                                                                                         groupColIdx,
1209                                                                                         numGroups,
1210                                                                                         agg_counts.numAggs,
1211                                                                                         result_plan);
1212                         /* Hashed aggregation produces randomly-ordered results */
1213                         current_pathkeys = NIL;
1214                 }
1215                 else if (parse->hasAggs)
1216                 {
1217                         /* Plain aggregate plan --- sort if needed */
1218                         AggStrategy aggstrategy;
1219
1220                         if (parse->groupClause)
1221                         {
1222                                 if (!pathkeys_contained_in(group_pathkeys, current_pathkeys))
1223                                 {
1224                                         result_plan = (Plan *)
1225                                                 make_sort_from_groupcols(parse,
1226                                                                                                  parse->groupClause,
1227                                                                                                  groupColIdx,
1228                                                                                                  result_plan);
1229                                         current_pathkeys = group_pathkeys;
1230                                 }
1231                                 aggstrategy = AGG_SORTED;
1232
1233                                 /*
1234                                  * The AGG node will not change the sort ordering of its
1235                                  * groups, so current_pathkeys describes the result too.
1236                                  */
1237                         }
1238                         else
1239                         {
1240                                 aggstrategy = AGG_PLAIN;
1241                                 /* Result will be only one row anyway; no sort order */
1242                                 current_pathkeys = NIL;
1243                         }
1244
1245                         result_plan = (Plan *) make_agg(parse,
1246                                                                                         tlist,
1247                                                                                         (List *) parse->havingQual,
1248                                                                                         aggstrategy,
1249                                                                                         numGroupCols,
1250                                                                                         groupColIdx,
1251                                                                                         numGroups,
1252                                                                                         agg_counts.numAggs,
1253                                                                                         result_plan);
1254                 }
1255                 else
1256                 {
1257                         /*
1258                          * If there are no Aggs, we shouldn't have any HAVING qual
1259                          * anymore
1260                          */
1261                         Assert(parse->havingQual == NULL);
1262
1263                         /*
1264                          * If we have a GROUP BY clause, insert a group node (plus the
1265                          * appropriate sort node, if necessary).
1266                          */
1267                         if (parse->groupClause)
1268                         {
1269                                 /*
1270                                  * Add an explicit sort if we couldn't make the path come
1271                                  * out the way the GROUP node needs it.
1272                                  */
1273                                 if (!pathkeys_contained_in(group_pathkeys, current_pathkeys))
1274                                 {
1275                                         result_plan = (Plan *)
1276                                                 make_sort_from_groupcols(parse,
1277                                                                                                  parse->groupClause,
1278                                                                                                  groupColIdx,
1279                                                                                                  result_plan);
1280                                         current_pathkeys = group_pathkeys;
1281                                 }
1282
1283                                 result_plan = (Plan *) make_group(parse,
1284                                                                                                   tlist,
1285                                                                                                   numGroupCols,
1286                                                                                                   groupColIdx,
1287                                                                                                   dNumGroups,
1288                                                                                                   result_plan);
1289                                 /* The Group node won't change sort ordering */
1290                         }
1291                 }
1292         }                                                       /* end of if (setOperations) */
1293
1294         /*
1295          * If we were not able to make the plan come out in the right order,
1296          * add an explicit sort step.
1297          */
1298         if (parse->sortClause)
1299         {
1300                 if (!pathkeys_contained_in(sort_pathkeys, current_pathkeys))
1301                 {
1302                         result_plan = (Plan *)
1303                                 make_sort_from_sortclauses(parse,
1304                                                                                    parse->sortClause,
1305                                                                                    result_plan);
1306                         current_pathkeys = sort_pathkeys;
1307                 }
1308         }
1309
1310         /*
1311          * If there is a DISTINCT clause, add the UNIQUE node.
1312          */
1313         if (parse->distinctClause)
1314         {
1315                 result_plan = (Plan *) make_unique(result_plan, parse->distinctClause);
1316
1317                 /*
1318                  * If there was grouping or aggregation, leave plan_rows as-is
1319                  * (ie, assume the result was already mostly unique).  If not,
1320                  * it's reasonable to assume the UNIQUE filter has effects
1321                  * comparable to GROUP BY.
1322                  */
1323                 if (!parse->groupClause && !parse->hasAggs)
1324                 {
1325                         List       *distinctExprs;
1326
1327                         distinctExprs = get_sortgrouplist_exprs(parse->distinctClause,
1328                                                                                                         parse->targetList);
1329                         result_plan->plan_rows = estimate_num_groups(parse,
1330                                                                                                                  distinctExprs,
1331                                                                                                  result_plan->plan_rows);
1332                 }
1333         }
1334
1335         /*
1336          * Finally, if there is a LIMIT/OFFSET clause, add the LIMIT node.
1337          */
1338         if (parse->limitOffset || parse->limitCount)
1339         {
1340                 result_plan = (Plan *) make_limit(result_plan,
1341                                                                                   parse->limitOffset,
1342                                                                                   parse->limitCount);
1343         }
1344
1345         /*
1346          * Return the actual output ordering in query_pathkeys for possible
1347          * use by an outer query level.
1348          */
1349         parse->query_pathkeys = current_pathkeys;
1350
1351         return result_plan;
1352 }
1353
1354 /*
1355  * hash_safe_grouping - are grouping operators hashable?
1356  *
1357  * We assume hashed aggregation will work if the datatype's equality operator
1358  * is marked hashjoinable.
1359  */
1360 static bool
1361 hash_safe_grouping(Query *parse)
1362 {
1363         ListCell   *gl;
1364
1365         foreach(gl, parse->groupClause)
1366         {
1367                 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1368                 TargetEntry *tle = get_sortgroupclause_tle(grpcl, parse->targetList);
1369                 Operator        optup;
1370                 bool            oprcanhash;
1371
1372                 optup = equality_oper(tle->resdom->restype, true);
1373                 if (!optup)
1374                         return false;
1375                 oprcanhash = ((Form_pg_operator) GETSTRUCT(optup))->oprcanhash;
1376                 ReleaseSysCache(optup);
1377                 if (!oprcanhash)
1378                         return false;
1379         }
1380         return true;
1381 }
1382
1383 /*---------------
1384  * make_subplanTargetList
1385  *        Generate appropriate target list when grouping is required.
1386  *
1387  * When grouping_planner inserts Aggregate or Group plan nodes above
1388  * the result of query_planner, we typically want to pass a different
1389  * target list to query_planner than the outer plan nodes should have.
1390  * This routine generates the correct target list for the subplan.
1391  *
1392  * The initial target list passed from the parser already contains entries
1393  * for all ORDER BY and GROUP BY expressions, but it will not have entries
1394  * for variables used only in HAVING clauses; so we need to add those
1395  * variables to the subplan target list.  Also, if we are doing either
1396  * grouping or aggregation, we flatten all expressions except GROUP BY items
1397  * into their component variables; the other expressions will be computed by
1398  * the inserted nodes rather than by the subplan.  For example,
1399  * given a query like
1400  *              SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
1401  * we want to pass this targetlist to the subplan:
1402  *              a,b,c,d,a+b
1403  * where the a+b target will be used by the Sort/Group steps, and the
1404  * other targets will be used for computing the final results.  (In the
1405  * above example we could theoretically suppress the a and b targets and
1406  * pass down only c,d,a+b, but it's not really worth the trouble to
1407  * eliminate simple var references from the subplan.  We will avoid doing
1408  * the extra computation to recompute a+b at the outer level; see
1409  * replace_vars_with_subplan_refs() in setrefs.c.)
1410  *
1411  * If we are grouping or aggregating, *and* there are no non-Var grouping
1412  * expressions, then the returned tlist is effectively dummy; we do not
1413  * need to force it to be evaluated, because all the Vars it contains
1414  * should be present in the output of query_planner anyway.
1415  *
1416  * 'parse' is the query being processed.
1417  * 'tlist' is the query's target list.
1418  * 'groupColIdx' receives an array of column numbers for the GROUP BY
1419  *                      expressions (if there are any) in the subplan's target list.
1420  * 'need_tlist_eval' is set true if we really need to evaluate the
1421  *                      result tlist.
1422  *
1423  * The result is the targetlist to be passed to the subplan.
1424  *---------------
1425  */
1426 static List *
1427 make_subplanTargetList(Query *parse,
1428                                            List *tlist,
1429                                            AttrNumber **groupColIdx,
1430                                            bool *need_tlist_eval)
1431 {
1432         List       *sub_tlist;
1433         List       *extravars;
1434         int                     numCols;
1435
1436         *groupColIdx = NULL;
1437
1438         /*
1439          * If we're not grouping or aggregating, nothing to do here;
1440          * query_planner should receive the unmodified target list.
1441          */
1442         if (!parse->hasAggs && !parse->groupClause)
1443         {
1444                 *need_tlist_eval = true;
1445                 return tlist;
1446         }
1447
1448         /*
1449          * Otherwise, start with a "flattened" tlist (having just the vars
1450          * mentioned in the targetlist and HAVING qual --- but not upper-
1451          * level Vars; they will be replaced by Params later on).
1452          */
1453         sub_tlist = flatten_tlist(tlist);
1454         extravars = pull_var_clause(parse->havingQual, false);
1455         sub_tlist = add_to_flat_tlist(sub_tlist, extravars);
1456         list_free(extravars);
1457         *need_tlist_eval = false;       /* only eval if not flat tlist */
1458
1459         /*
1460          * If grouping, create sub_tlist entries for all GROUP BY expressions
1461          * (GROUP BY items that are simple Vars should be in the list
1462          * already), and make an array showing where the group columns are in
1463          * the sub_tlist.
1464          */
1465         numCols = list_length(parse->groupClause);
1466         if (numCols > 0)
1467         {
1468                 int                     keyno = 0;
1469                 AttrNumber *grpColIdx;
1470                 ListCell   *gl;
1471
1472                 grpColIdx = (AttrNumber *) palloc(sizeof(AttrNumber) * numCols);
1473                 *groupColIdx = grpColIdx;
1474
1475                 foreach(gl, parse->groupClause)
1476                 {
1477                         GroupClause *grpcl = (GroupClause *) lfirst(gl);
1478                         Node       *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1479                         TargetEntry *te = NULL;
1480                         ListCell   *sl;
1481
1482                         /* Find or make a matching sub_tlist entry */
1483                         foreach(sl, sub_tlist)
1484                         {
1485                                 te = (TargetEntry *) lfirst(sl);
1486                                 if (equal(groupexpr, te->expr))
1487                                         break;
1488                         }
1489                         if (!sl)
1490                         {
1491                                 te = makeTargetEntry(makeResdom(list_length(sub_tlist) + 1,
1492                                                                                                 exprType(groupexpr),
1493                                                                                                 exprTypmod(groupexpr),
1494                                                                                                 NULL,
1495                                                                                                 false),
1496                                                                          (Expr *) groupexpr);
1497                                 sub_tlist = lappend(sub_tlist, te);
1498                                 *need_tlist_eval = true;                /* it's not flat anymore */
1499                         }
1500
1501                         /* and save its resno */
1502                         grpColIdx[keyno++] = te->resdom->resno;
1503                 }
1504         }
1505
1506         return sub_tlist;
1507 }
1508
1509 /*
1510  * locate_grouping_columns
1511  *              Locate grouping columns in the tlist chosen by query_planner.
1512  *
1513  * This is only needed if we don't use the sub_tlist chosen by
1514  * make_subplanTargetList.      We have to forget the column indexes found
1515  * by that routine and re-locate the grouping vars in the real sub_tlist.
1516  */
1517 static void
1518 locate_grouping_columns(Query *parse,
1519                                                 List *tlist,
1520                                                 List *sub_tlist,
1521                                                 AttrNumber *groupColIdx)
1522 {
1523         int                     keyno = 0;
1524         ListCell   *gl;
1525
1526         /*
1527          * No work unless grouping.
1528          */
1529         if (!parse->groupClause)
1530         {
1531                 Assert(groupColIdx == NULL);
1532                 return;
1533         }
1534         Assert(groupColIdx != NULL);
1535
1536         foreach(gl, parse->groupClause)
1537         {
1538                 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1539                 Node       *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1540                 TargetEntry *te = NULL;
1541                 ListCell   *sl;
1542
1543                 foreach(sl, sub_tlist)
1544                 {
1545                         te = (TargetEntry *) lfirst(sl);
1546                         if (equal(groupexpr, te->expr))
1547                                 break;
1548                 }
1549                 if (!sl)
1550                         elog(ERROR, "failed to locate grouping columns");
1551
1552                 groupColIdx[keyno++] = te->resdom->resno;
1553         }
1554 }
1555
1556 /*
1557  * postprocess_setop_tlist
1558  *        Fix up targetlist returned by plan_set_operations().
1559  *
1560  * We need to transpose sort key info from the orig_tlist into new_tlist.
1561  * NOTE: this would not be good enough if we supported resjunk sort keys
1562  * for results of set operations --- then, we'd need to project a whole
1563  * new tlist to evaluate the resjunk columns.  For now, just ereport if we
1564  * find any resjunk columns in orig_tlist.
1565  */
1566 static List *
1567 postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
1568 {
1569         ListCell   *l;
1570         ListCell   *orig_tlist_item = list_head(orig_tlist);
1571
1572         foreach(l, new_tlist)
1573         {
1574                 TargetEntry *new_tle = (TargetEntry *) lfirst(l);
1575                 TargetEntry *orig_tle;
1576
1577                 /* ignore resjunk columns in setop result */
1578                 if (new_tle->resdom->resjunk)
1579                         continue;
1580
1581                 Assert(orig_tlist_item != NULL);
1582                 orig_tle = (TargetEntry *) lfirst(orig_tlist_item);
1583                 orig_tlist_item = lnext(orig_tlist_item);
1584                 if (orig_tle->resdom->resjunk)  /* should not happen */
1585                         elog(ERROR, "resjunk output columns are not implemented");
1586                 Assert(new_tle->resdom->resno == orig_tle->resdom->resno);
1587                 Assert(new_tle->resdom->restype == orig_tle->resdom->restype);
1588                 new_tle->resdom->ressortgroupref = orig_tle->resdom->ressortgroupref;
1589         }
1590         if (orig_tlist_item != NULL)
1591                 elog(ERROR, "resjunk output columns are not implemented");
1592         return new_tlist;
1593 }