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) 1996-2003, 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 1.164 2004/01/12 22:20:28 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_LIMIT  3
  51 #define EXPRKIND_ININFO 4
  52
  53
  54 static Node *preprocess_expression(Query *parse, Node *expr, int kind);
  55 static void preprocess_qual_conditions(Query *parse, Node *jtnode);
  56 static Plan *inheritance_planner(Query *parse, List *inheritlist);
  57 static Plan *grouping_planner(Query *parse, double tuple_fraction);
  58 static bool hash_safe_grouping(Query *parse);
  59 static List *make_subplanTargetList(Query *parse, List *tlist,
  60                                            AttrNumber **groupColIdx, bool *need_tlist_eval);
  61 static void locate_grouping_columns(Query *parse,
  62                                                 List *tlist,
  63                                                 List *sub_tlist,
  64                                                 AttrNumber *groupColIdx);
  65 static List *postprocess_setop_tlist(List *new_tlist, List *orig_tlist);
  66
  67
  68 /*****************************************************************************
  69  *
  70  *         Query optimizer entry point
  71  *
  72  *****************************************************************************/
  73 Plan *
  74 planner(Query *parse, bool isCursor, int cursorOptions)
  75 {
  76         double          tuple_fraction;
  77         Plan       *result_plan;
  78         Index           save_PlannerQueryLevel;
  79         List       *save_PlannerParamList;
  80
  81         /*
  82          * The planner can be called recursively (an example is when
  83          * eval_const_expressions tries to pre-evaluate an SQL function). So,
  84          * these global state variables must be saved and restored.
  85          *
  86          * These vars cannot be moved into the Query structure since their whole
  87          * purpose is communication across multiple sub-Queries.
  88          *
  89          * Note we do NOT save and restore PlannerPlanId: it exists to assign
  90          * unique IDs to SubPlan nodes, and we want those IDs to be unique for
  91          * the life of a backend.  Also, PlannerInitPlan is saved/restored in
  92          * subquery_planner, not here.
  93          */
  94         save_PlannerQueryLevel = PlannerQueryLevel;
  95         save_PlannerParamList = PlannerParamList;
  96
  97         /* Initialize state for handling outer-level references and params */
  98         PlannerQueryLevel = 0;          /* will be 1 in top-level subquery_planner */
  99         PlannerParamList = NIL;
 100
 101         /* Determine what fraction of the plan is likely to be scanned */
 102         if (isCursor)
 103         {
 104                 /*
 105                  * We have no real idea how many tuples the user will ultimately
 106                  * FETCH from a cursor, but it seems a good bet that he doesn't
 107                  * want 'em all.  Optimize for 10% retrieval (you gotta better
 108                  * number?      Should this be a SETtable parameter?)
 109                  */
 110                 tuple_fraction = 0.10;
 111         }
 112         else
 113         {
 114                 /* Default assumption is we need all the tuples */
 115                 tuple_fraction = 0.0;
 116         }
 117
 118         /* primary planning entry point (may recurse for subqueries) */
 119         result_plan = subquery_planner(parse, tuple_fraction);
 120
 121         Assert(PlannerQueryLevel == 0);
 122
 123         /*
 124          * If creating a plan for a scrollable cursor, make sure it can run
 125          * backwards on demand.  Add a Material node at the top at need.
 126          */
 127         if (isCursor && (cursorOptions & CURSOR_OPT_SCROLL))
 128         {
 129                 if (!ExecSupportsBackwardScan(result_plan))
 130                         result_plan = materialize_finished_plan(result_plan);
 131         }
 132
 133         /* executor wants to know total number of Params used overall */
 134         result_plan->nParamExec = length(PlannerParamList);
 135
 136         /* final cleanup of the plan */
 137         set_plan_references(result_plan, parse->rtable);
 138
 139         /* restore state for outer planner, if any */
 140         PlannerQueryLevel = save_PlannerQueryLevel;
 141         PlannerParamList = save_PlannerParamList;
 142
 143         return result_plan;
 144 }
 145
 146
 147 /*--------------------
 148  * subquery_planner
 149  *        Invokes the planner on a subquery.  We recurse to here for each
 150  *        sub-SELECT found in the query tree.
 151  *
 152  * parse is the querytree produced by the parser & rewriter.
 153  * tuple_fraction is the fraction of tuples we expect will be retrieved.
 154  * tuple_fraction is interpreted as explained for grouping_planner, below.
 155  *
 156  * Basically, this routine does the stuff that should only be done once
 157  * per Query object.  It then calls grouping_planner.  At one time,
 158  * grouping_planner could be invoked recursively on the same Query object;
 159  * that's not currently true, but we keep the separation between the two
 160  * routines anyway, in case we need it again someday.
 161  *
 162  * subquery_planner will be called recursively to handle sub-Query nodes
 163  * found within the query's expressions and rangetable.
 164  *
 165  * Returns a query plan.
 166  *--------------------
 167  */
 168 Plan *
 169 subquery_planner(Query *parse, double tuple_fraction)
 170 {
 171         List       *saved_initplan = PlannerInitPlan;
 172         int                     saved_planid = PlannerPlanId;
 173         bool            hasOuterJoins;
 174         Plan       *plan;
 175         List       *newHaving;
 176         List       *lst;
 177
 178         /* Set up for a new level of subquery */
 179         PlannerQueryLevel++;
 180         PlannerInitPlan = NIL;
 181
 182         /*
 183          * Look for IN clauses at the top level of WHERE, and transform them
 184          * into joins.  Note that this step only handles IN clauses originally
 185          * at top level of WHERE; if we pull up any subqueries in the next
 186          * step, their INs are processed just before pulling them up.
 187          */
 188         parse->in_info_list = NIL;
 189         if (parse->hasSubLinks)
 190                 parse->jointree->quals = pull_up_IN_clauses(parse,
 191                                                                                                  parse->jointree->quals);
 192
 193         /*
 194          * Check to see if any subqueries in the rangetable can be merged into
 195          * this query.
 196          */
 197         parse->jointree = (FromExpr *)
 198                 pull_up_subqueries(parse, (Node *) parse->jointree, false);
 199
 200         /*
 201          * Detect whether any rangetable entries are RTE_JOIN kind; if not, we
 202          * can avoid the expense of doing flatten_join_alias_vars().  Also
 203          * check for outer joins --- if none, we can skip
 204          * reduce_outer_joins(). This must be done after we have done
 205          * pull_up_subqueries, of course.
 206          */
 207         parse->hasJoinRTEs = false;
 208         hasOuterJoins = false;
 209         foreach(lst, parse->rtable)
 210         {
 211                 RangeTblEntry *rte = (RangeTblEntry *) lfirst(lst);
 212
 213                 if (rte->rtekind == RTE_JOIN)
 214                 {
 215                         parse->hasJoinRTEs = true;
 216                         if (IS_OUTER_JOIN(rte->jointype))
 217                         {
 218                                 hasOuterJoins = true;
 219                                 /* Can quit scanning once we find an outer join */
 220                                 break;
 221                         }
 222                 }
 223         }
 224
 225         /*
 226          * Do expression preprocessing on targetlist and quals.
 227          */
 228         parse->targetList = (List *)
 229                 preprocess_expression(parse, (Node *) parse->targetList,
 230                                                           EXPRKIND_TARGET);
 231
 232         preprocess_qual_conditions(parse, (Node *) parse->jointree);
 233
 234         parse->havingQual = preprocess_expression(parse, parse->havingQual,
 235                                                                                           EXPRKIND_QUAL);
 236
 237         parse->limitOffset = preprocess_expression(parse, parse->limitOffset,
 238                                                                                            EXPRKIND_LIMIT);
 239         parse->limitCount = preprocess_expression(parse, parse->limitCount,
 240                                                                                           EXPRKIND_LIMIT);
 241
 242         parse->in_info_list = (List *)
 243                 preprocess_expression(parse, (Node *) parse->in_info_list,
 244                                                           EXPRKIND_ININFO);
 245
 246         /* Also need to preprocess expressions for function RTEs */
 247         foreach(lst, parse->rtable)
 248         {
 249                 RangeTblEntry *rte = (RangeTblEntry *) lfirst(lst);
 250
 251                 if (rte->rtekind == RTE_FUNCTION)
 252                         rte->funcexpr = preprocess_expression(parse, rte->funcexpr,
 253                                                                                                   EXPRKIND_RTFUNC);
 254         }
 255
 256         /*
 257          * A HAVING clause without aggregates is equivalent to a WHERE clause
 258          * (except it can only refer to grouped fields).  Transfer any
 259          * agg-free clauses of the HAVING qual into WHERE.      This may seem like
 260          * wasting cycles to cater to stupidly-written queries, but there are
 261          * other reasons for doing it.  Firstly, if the query contains no aggs
 262          * at all, then we aren't going to generate an Agg plan node, and so
 263          * there'll be no place to execute HAVING conditions; without this
 264          * transfer, we'd lose the HAVING condition entirely, which is wrong.
 265          * Secondly, when we push down a qual condition into a sub-query, it's
 266          * easiest to push the qual into HAVING always, in case it contains
 267          * aggs, and then let this code sort it out.
 268          *
 269          * Note that both havingQual and parse->jointree->quals are in
 270          * implicitly-ANDed-list form at this point, even though they are
 271          * declared as Node *.
 272          */
 273         newHaving = NIL;
 274         foreach(lst, (List *) parse->havingQual)
 275         {
 276                 Node       *havingclause = (Node *) lfirst(lst);
 277
 278                 if (contain_agg_clause(havingclause))
 279                         newHaving = lappend(newHaving, havingclause);
 280                 else
 281                         parse->jointree->quals = (Node *)
 282                                 lappend((List *) parse->jointree->quals, havingclause);
 283         }
 284         parse->havingQual = (Node *) newHaving;
 285
 286         /*
 287          * If we have any outer joins, try to reduce them to plain inner
 288          * joins. This step is most easily done after we've done expression
 289          * preprocessing.
 290          */
 291         if (hasOuterJoins)
 292                 reduce_outer_joins(parse);
 293
 294         /*
 295          * See if we can simplify the jointree; opportunities for this may
 296          * come from having pulled up subqueries, or from flattening explicit
 297          * JOIN syntax.  We must do this after flattening JOIN alias
 298          * variables, since eliminating explicit JOIN nodes from the jointree
 299          * will cause get_relids_for_join() to fail.  But it should happen
 300          * after reduce_outer_joins, anyway.
 301          */
 302         parse->jointree = (FromExpr *)
 303                 simplify_jointree(parse, (Node *) parse->jointree);
 304
 305         /*
 306          * Do the main planning.  If we have an inherited target relation,
 307          * that needs special processing, else go straight to
 308          * grouping_planner.
 309          */
 310         if (parse->resultRelation &&
 311                 (lst = expand_inherited_rtentry(parse, parse->resultRelation,
 312                                                                                 false)) != NIL)
 313                 plan = inheritance_planner(parse, lst);
 314         else
 315                 plan = grouping_planner(parse, tuple_fraction);
 316
 317         /*
 318          * If any subplans were generated, or if we're inside a subplan, build
 319          * initPlan list and extParam/allParam sets for plan nodes.
 320          */
 321         if (PlannerPlanId != saved_planid || PlannerQueryLevel > 1)
 322         {
 323                 Cost            initplan_cost = 0;
 324
 325                 /* Prepare extParam/allParam sets for all nodes in tree */
 326                 SS_finalize_plan(plan, parse->rtable);
 327
 328                 /*
 329                  * SS_finalize_plan doesn't handle initPlans, so we have to
 330                  * manually attach them to the topmost plan node, and add their
 331                  * extParams to the topmost node's, too.
 332                  *
 333                  * We also add the total_cost of each initPlan to the startup cost of
 334                  * the top node.  This is a conservative overestimate, since in
 335                  * fact each initPlan might be executed later than plan startup,
 336                  * or even not at all.
 337                  */
 338                 plan->initPlan = PlannerInitPlan;
 339
 340                 foreach(lst, plan->initPlan)
 341                 {
 342                         SubPlan    *initplan = (SubPlan *) lfirst(lst);
 343
 344                         plan->extParam = bms_add_members(plan->extParam,
 345                                                                                          initplan->plan->extParam);
 346                         initplan_cost += initplan->plan->total_cost;
 347                 }
 348
 349                 plan->startup_cost += initplan_cost;
 350                 plan->total_cost += initplan_cost;
 351         }
 352
 353         /* Return to outer subquery context */
 354         PlannerQueryLevel--;
 355         PlannerInitPlan = saved_initplan;
 356         /* we do NOT restore PlannerPlanId; that's not an oversight! */
 357
 358         return plan;
 359 }
 360
 361 /*
 362  * preprocess_expression
 363  *              Do subquery_planner's preprocessing work for an expression,
 364  *              which can be a targetlist, a WHERE clause (including JOIN/ON
 365  *              conditions), or a HAVING clause.
 366  */
 367 static Node *
 368 preprocess_expression(Query *parse, Node *expr, int kind)
 369 {
 370         /*
 371          * If the query has any join RTEs, replace join alias variables with
 372          * base-relation variables. We must do this before sublink processing,
 373          * else sublinks expanded out from join aliases wouldn't get
 374          * processed.
 375          */
 376         if (parse->hasJoinRTEs)
 377                 expr = flatten_join_alias_vars(parse, expr);
 378
 379         /*
 380          * If it's a qual or havingQual, canonicalize it.  It seems most useful
 381          * to do this before applying eval_const_expressions, since the latter
 382          * can optimize flattened AND/ORs better than unflattened ones.
 383          *
 384          * Note: all processing of a qual expression after this point must be
 385          * careful to maintain AND/OR flatness --- that is, do not generate a
 386          * tree with AND directly under AND, nor OR directly under OR.
 387          */
 388         if (kind == EXPRKIND_QUAL)
 389         {
 390                 expr = (Node *) canonicalize_qual((Expr *) expr);
 391
 392 #ifdef OPTIMIZER_DEBUG
 393                 printf("After canonicalize_qual()\n");
 394                 pprint(expr);
 395 #endif
 396         }
 397
 398         /*
 399          * Simplify constant expressions.
 400          */
 401         expr = eval_const_expressions(expr);
 402
 403         /* Expand SubLinks to SubPlans */
 404         if (parse->hasSubLinks)
 405                 expr = SS_process_sublinks(expr, (kind == EXPRKIND_QUAL));
 406
 407         /*
 408          * XXX do not insert anything here unless you have grokked the
 409          * comments in SS_replace_correlation_vars ...
 410          */
 411
 412         /* Replace uplevel vars with Param nodes */
 413         if (PlannerQueryLevel > 1)
 414                 expr = SS_replace_correlation_vars(expr);
 415
 416         /*
 417          * If it's a qual or havingQual, convert it to implicit-AND format.
 418          * (We don't want to do this before eval_const_expressions, since the
 419          * latter would be unable to simplify a top-level AND correctly.  Also,
 420          * SS_process_sublinks expects explicit-AND format.)
 421          */
 422         if (kind == EXPRKIND_QUAL)
 423                 expr = (Node *) make_ands_implicit((Expr *) expr);
 424
 425         return expr;
 426 }
 427
 428 /*
 429  * preprocess_qual_conditions
 430  *              Recursively scan the query's jointree and do subquery_planner's
 431  *              preprocessing work on each qual condition found therein.
 432  */
 433 static void
 434 preprocess_qual_conditions(Query *parse, Node *jtnode)
 435 {
 436         if (jtnode == NULL)
 437                 return;
 438         if (IsA(jtnode, RangeTblRef))
 439         {
 440                 /* nothing to do here */
 441         }
 442         else if (IsA(jtnode, FromExpr))
 443         {
 444                 FromExpr   *f = (FromExpr *) jtnode;
 445                 List       *l;
 446
 447                 foreach(l, f->fromlist)
 448                         preprocess_qual_conditions(parse, lfirst(l));
 449
 450                 f->quals = preprocess_expression(parse, f->quals, EXPRKIND_QUAL);
 451         }
 452         else if (IsA(jtnode, JoinExpr))
 453         {
 454                 JoinExpr   *j = (JoinExpr *) jtnode;
 455
 456                 preprocess_qual_conditions(parse, j->larg);
 457                 preprocess_qual_conditions(parse, j->rarg);
 458
 459                 j->quals = preprocess_expression(parse, j->quals, EXPRKIND_QUAL);
 460         }
 461         else
 462                 elog(ERROR, "unrecognized node type: %d",
 463                          (int) nodeTag(jtnode));
 464 }
 465
 466 /*--------------------
 467  * inheritance_planner
 468  *        Generate a plan in the case where the result relation is an
 469  *        inheritance set.
 470  *
 471  * We have to handle this case differently from cases where a source
 472  * relation is an inheritance set.      Source inheritance is expanded at
 473  * the bottom of the plan tree (see allpaths.c), but target inheritance
 474  * has to be expanded at the top.  The reason is that for UPDATE, each
 475  * target relation needs a different targetlist matching its own column
 476  * set.  (This is not so critical for DELETE, but for simplicity we treat
 477  * inherited DELETE the same way.)      Fortunately, the UPDATE/DELETE target
 478  * can never be the nullable side of an outer join, so it's OK to generate
 479  * the plan this way.
 480  *
 481  * parse is the querytree produced by the parser & rewriter.
 482  * inheritlist is an integer list of RT indexes for the result relation set.
 483  *
 484  * Returns a query plan.
 485  *--------------------
 486  */
 487 static Plan *
 488 inheritance_planner(Query *parse, List *inheritlist)
 489 {
 490         int                     parentRTindex = parse->resultRelation;
 491         Oid                     parentOID = getrelid(parentRTindex, parse->rtable);
 492         int                     mainrtlength = length(parse->rtable);
 493         List       *subplans = NIL;
 494         List       *tlist = NIL;
 495         List       *l;
 496
 497         foreach(l, inheritlist)
 498         {
 499                 int                     childRTindex = lfirsti(l);
 500                 Oid                     childOID = getrelid(childRTindex, parse->rtable);
 501                 int                     subrtlength;
 502                 Query      *subquery;
 503                 Plan       *subplan;
 504
 505                 /* Generate modified query with this rel as target */
 506                 subquery = (Query *) adjust_inherited_attrs((Node *) parse,
 507                                                                                                 parentRTindex, parentOID,
 508                                                                                                  childRTindex, childOID);
 509                 /* Generate plan */
 510                 subplan = grouping_planner(subquery, 0.0 /* retrieve all tuples */ );
 511                 subplans = lappend(subplans, subplan);
 512
 513                 /*
 514                  * It's possible that additional RTEs got added to the rangetable
 515                  * due to expansion of inherited source tables (see allpaths.c).
 516                  * If so, we must copy 'em back to the main parse tree's rtable.
 517                  *
 518                  * XXX my goodness this is ugly.  Really need to think about ways to
 519                  * rein in planner's habit of scribbling on its input.
 520                  */
 521                 subrtlength = length(subquery->rtable);
 522                 if (subrtlength > mainrtlength)
 523                 {
 524                         List       *subrt = subquery->rtable;
 525
 526                         while (mainrtlength-- > 0)      /* wish we had nthcdr() */
 527                                 subrt = lnext(subrt);
 528                         parse->rtable = nconc(parse->rtable, subrt);
 529                         mainrtlength = subrtlength;
 530                 }
 531                 /* Save preprocessed tlist from first rel for use in Append */
 532                 if (tlist == NIL)
 533                         tlist = subplan->targetlist;
 534         }
 535
 536         /* Save the target-relations list for the executor, too */
 537         parse->resultRelations = inheritlist;
 538
 539         /* Mark result as unordered (probably unnecessary) */
 540         parse->query_pathkeys = NIL;
 541
 542         return (Plan *) make_append(subplans, true, tlist);
 543 }
 544
 545 /*--------------------
 546  * grouping_planner
 547  *        Perform planning steps related to grouping, aggregation, etc.
 548  *        This primarily means adding top-level processing to the basic
 549  *        query plan produced by query_planner.
 550  *
 551  * parse is the querytree produced by the parser & rewriter.
 552  * tuple_fraction is the fraction of tuples we expect will be retrieved
 553  *
 554  * tuple_fraction is interpreted as follows:
 555  *        0: expect all tuples to be retrieved (normal case)
 556  *        0 < tuple_fraction < 1: expect the given fraction of tuples available
 557  *              from the plan to be retrieved
 558  *        tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
 559  *              expected to be retrieved (ie, a LIMIT specification)
 560  *
 561  * Returns a query plan.  Also, parse->query_pathkeys is returned as the
 562  * actual output ordering of the plan (in pathkey format).
 563  *--------------------
 564  */
 565 static Plan *
 566 grouping_planner(Query *parse, double tuple_fraction)
 567 {
 568         List       *tlist = parse->targetList;
 569         Plan       *result_plan;
 570         List       *current_pathkeys;
 571         List       *sort_pathkeys;
 572
 573         if (parse->setOperations)
 574         {
 575                 /*
 576                  * Construct the plan for set operations.  The result will not
 577                  * need any work except perhaps a top-level sort and/or LIMIT.
 578                  */
 579                 result_plan = plan_set_operations(parse);
 580
 581                 /*
 582                  * We should not need to call preprocess_targetlist, since we must
 583                  * be in a SELECT query node.  Instead, use the targetlist
 584                  * returned by plan_set_operations (since this tells whether it
 585                  * returned any resjunk columns!), and transfer any sort key
 586                  * information from the original tlist.
 587                  */
 588                 Assert(parse->commandType == CMD_SELECT);
 589
 590                 tlist = postprocess_setop_tlist(result_plan->targetlist, tlist);
 591
 592                 /*
 593                  * Can't handle FOR UPDATE here (parser should have checked
 594                  * already, but let's make sure).
 595                  */
 596                 if (parse->rowMarks)
 597                         ereport(ERROR,
 598                                         (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
 599                                          errmsg("SELECT FOR UPDATE is not allowed with UNION/INTERSECT/EXCEPT")));
 600
 601                 /*
 602                  * We set current_pathkeys NIL indicating we do not know sort
 603                  * order.  This is correct when the top set operation is UNION
 604                  * ALL, since the appended-together results are unsorted even if
 605                  * the subplans were sorted.  For other set operations we could be
 606                  * smarter --- room for future improvement!
 607                  */
 608                 current_pathkeys = NIL;
 609
 610                 /*
 611                  * Calculate pathkeys that represent ordering requirements
 612                  */
 613                 sort_pathkeys = make_pathkeys_for_sortclauses(parse->sortClause,
 614                                                                                                           tlist);
 615                 sort_pathkeys = canonicalize_pathkeys(parse, sort_pathkeys);
 616         }
 617         else
 618         {
 619                 /* No set operations, do regular planning */
 620                 List       *sub_tlist;
 621                 List       *group_pathkeys;
 622                 AttrNumber *groupColIdx = NULL;
 623                 bool            need_tlist_eval = true;
 624                 QualCost        tlist_cost;
 625                 double          sub_tuple_fraction;
 626                 Path       *cheapest_path;
 627                 Path       *sorted_path;
 628                 double          dNumGroups = 0;
 629                 long            numGroups = 0;
 630                 int                     numAggs = 0;
 631                 int                     numGroupCols = length(parse->groupClause);
 632                 bool            use_hashed_grouping = false;
 633
 634                 /* Preprocess targetlist in case we are inside an INSERT/UPDATE. */
 635                 tlist = preprocess_targetlist(tlist,
 636                                                                           parse->commandType,
 637                                                                           parse->resultRelation,
 638                                                                           parse->rtable);
 639
 640                 /*
 641                  * Add TID targets for rels selected FOR UPDATE (should this be
 642                  * done in preprocess_targetlist?).  The executor uses the TID to
 643                  * know which rows to lock, much as for UPDATE or DELETE.
 644                  */
 645                 if (parse->rowMarks)
 646                 {
 647                         List       *l;
 648
 649                         /*
 650                          * We've got trouble if the FOR UPDATE appears inside
 651                          * grouping, since grouping renders a reference to individual
 652                          * tuple CTIDs invalid.  This is also checked at parse time,
 653                          * but that's insufficient because of rule substitution, query
 654                          * pullup, etc.
 655                          */
 656                         CheckSelectForUpdate(parse);
 657
 658                         /*
 659                          * Currently the executor only supports FOR UPDATE at top
 660                          * level
 661                          */
 662                         if (PlannerQueryLevel > 1)
 663                                 ereport(ERROR,
 664                                                 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
 665                                                  errmsg("SELECT FOR UPDATE is not allowed in subqueries")));
 666
 667                         foreach(l, parse->rowMarks)
 668                         {
 669                                 Index           rti = lfirsti(l);
 670                                 char       *resname;
 671                                 Resdom     *resdom;
 672                                 Var                *var;
 673                                 TargetEntry *ctid;
 674
 675                                 resname = (char *) palloc(32);
 676                                 snprintf(resname, 32, "ctid%u", rti);
 677                                 resdom = makeResdom(length(tlist) + 1,
 678                                                                         TIDOID,
 679                                                                         -1,
 680                                                                         resname,
 681                                                                         true);
 682
 683                                 var = makeVar(rti,
 684                                                           SelfItemPointerAttributeNumber,
 685                                                           TIDOID,
 686                                                           -1,
 687                                                           0);
 688
 689                                 ctid = makeTargetEntry(resdom, (Expr *) var);
 690                                 tlist = lappend(tlist, ctid);
 691                         }
 692                 }
 693
 694                 /*
 695                  * Generate appropriate target list for subplan; may be different
 696                  * from tlist if grouping or aggregation is needed.
 697                  */
 698                 sub_tlist = make_subplanTargetList(parse, tlist,
 699                                                                                  &groupColIdx, &need_tlist_eval);
 700
 701                 /*
 702                  * Calculate pathkeys that represent grouping/ordering
 703                  * requirements
 704                  */
 705                 group_pathkeys = make_pathkeys_for_sortclauses(parse->groupClause,
 706                                                                                                            tlist);
 707                 sort_pathkeys = make_pathkeys_for_sortclauses(parse->sortClause,
 708                                                                                                           tlist);
 709
 710                 /*
 711                  * Will need actual number of aggregates for estimating costs.
 712                  * Also, it's possible that optimization has eliminated all
 713                  * aggregates, and we may as well check for that here.
 714                  *
 715                  * Note: we do not attempt to detect duplicate aggregates here; a
 716                  * somewhat-overestimated count is okay for our present purposes.
 717                  */
 718                 if (parse->hasAggs)
 719                 {
 720                         numAggs = count_agg_clause((Node *) tlist) +
 721                                 count_agg_clause(parse->havingQual);
 722                         if (numAggs == 0)
 723                                 parse->hasAggs = false;
 724                 }
 725
 726                 /*
 727                  * Figure out whether we need a sorted result from query_planner.
 728                  *
 729                  * If we have a GROUP BY clause, then we want a result sorted
 730                  * properly for grouping.  Otherwise, if there is an ORDER BY
 731                  * clause, we want to sort by the ORDER BY clause.      (Note: if we
 732                  * have both, and ORDER BY is a superset of GROUP BY, it would be
 733                  * tempting to request sort by ORDER BY --- but that might just
 734                  * leave us failing to exploit an available sort order at all.
 735                  * Needs more thought...)
 736                  */
 737                 if (parse->groupClause)
 738                         parse->query_pathkeys = group_pathkeys;
 739                 else if (parse->sortClause)
 740                         parse->query_pathkeys = sort_pathkeys;
 741                 else
 742                         parse->query_pathkeys = NIL;
 743
 744                 /*
 745                  * Adjust tuple_fraction if we see that we are going to apply
 746                  * limiting/grouping/aggregation/etc.  This is not overridable by
 747                  * the caller, since it reflects plan actions that this routine
 748                  * will certainly take, not assumptions about context.
 749                  */
 750                 if (parse->limitCount != NULL)
 751                 {
 752                         /*
 753                          * A LIMIT clause limits the absolute number of tuples
 754                          * returned. However, if it's not a constant LIMIT then we
 755                          * have to punt; for lack of a better idea, assume 10% of the
 756                          * plan's result is wanted.
 757                          */
 758                         double          limit_fraction = 0.0;
 759
 760                         if (IsA(parse->limitCount, Const))
 761                         {
 762                                 Const      *limitc = (Const *) parse->limitCount;
 763                                 int32           count = DatumGetInt32(limitc->constvalue);
 764
 765                                 /*
 766                                  * A NULL-constant LIMIT represents "LIMIT ALL", which we
 767                                  * treat the same as no limit (ie, expect to retrieve all
 768                                  * the tuples).
 769                                  */
 770                                 if (!limitc->constisnull && count > 0)
 771                                 {
 772                                         limit_fraction = (double) count;
 773                                         /* We must also consider the OFFSET, if present */
 774                                         if (parse->limitOffset != NULL)
 775                                         {
 776                                                 if (IsA(parse->limitOffset, Const))
 777                                                 {
 778                                                         int32           offset;
 779
 780                                                         limitc = (Const *) parse->limitOffset;
 781                                                         offset = DatumGetInt32(limitc->constvalue);
 782                                                         if (!limitc->constisnull && offset > 0)
 783                                                                 limit_fraction += (double) offset;
 784                                                 }
 785                                                 else
 786                                                 {
 787                                                         /* OFFSET is an expression ... punt ... */
 788                                                         limit_fraction = 0.10;
 789                                                 }
 790                                         }
 791                                 }
 792                         }
 793                         else
 794                         {
 795                                 /* LIMIT is an expression ... punt ... */
 796                                 limit_fraction = 0.10;
 797                         }
 798
 799                         if (limit_fraction > 0.0)
 800                         {
 801                                 /*
 802                                  * If we have absolute limits from both caller and LIMIT,
 803                                  * use the smaller value; if one is fractional and the
 804                                  * other absolute, treat the fraction as a fraction of the
 805                                  * absolute value; else we can multiply the two fractions
 806                                  * together.
 807                                  */
 808                                 if (tuple_fraction >= 1.0)
 809                                 {
 810                                         if (limit_fraction >= 1.0)
 811                                         {
 812                                                 /* both absolute */
 813                                                 tuple_fraction = Min(tuple_fraction, limit_fraction);
 814                                         }
 815                                         else
 816                                         {
 817                                                 /* caller absolute, limit fractional */
 818                                                 tuple_fraction *= limit_fraction;
 819                                                 if (tuple_fraction < 1.0)
 820                                                         tuple_fraction = 1.0;
 821                                         }
 822                                 }
 823                                 else if (tuple_fraction > 0.0)
 824                                 {
 825                                         if (limit_fraction >= 1.0)
 826                                         {
 827                                                 /* caller fractional, limit absolute */
 828                                                 tuple_fraction *= limit_fraction;
 829                                                 if (tuple_fraction < 1.0)
 830                                                         tuple_fraction = 1.0;
 831                                         }
 832                                         else
 833                                         {
 834                                                 /* both fractional */
 835                                                 tuple_fraction *= limit_fraction;
 836                                         }
 837                                 }
 838                                 else
 839                                 {
 840                                         /* no info from caller, just use limit */
 841                                         tuple_fraction = limit_fraction;
 842                                 }
 843                         }
 844                 }
 845
 846                 /*
 847                  * With grouping or aggregation, the tuple fraction to pass to
 848                  * query_planner() may be different from what it is at top level.
 849                  */
 850                 sub_tuple_fraction = tuple_fraction;
 851
 852                 if (parse->groupClause)
 853                 {
 854                         /*
 855                          * In GROUP BY mode, we have the little problem that we don't
 856                          * really know how many input tuples will be needed to make a
 857                          * group, so we can't translate an output LIMIT count into an
 858                          * input count.  For lack of a better idea, assume 25% of the
 859                          * input data will be processed if there is any output limit.
 860                          * However, if the caller gave us a fraction rather than an
 861                          * absolute count, we can keep using that fraction (which
 862                          * amounts to assuming that all the groups are about the same
 863                          * size).
 864                          */
 865                         if (sub_tuple_fraction >= 1.0)
 866                                 sub_tuple_fraction = 0.25;
 867
 868                         /*
 869                          * If both GROUP BY and ORDER BY are specified, we will need
 870                          * two levels of sort --- and, therefore, certainly need to
 871                          * read all the input tuples --- unless ORDER BY is a subset
 872                          * of GROUP BY.  (We have not yet canonicalized the pathkeys,
 873                          * so must use the slower noncanonical comparison method.)
 874                          */
 875                         if (parse->groupClause && parse->sortClause &&
 876                                 !noncanonical_pathkeys_contained_in(sort_pathkeys,
 877                                                                                                         group_pathkeys))
 878                                 sub_tuple_fraction = 0.0;
 879                 }
 880                 else if (parse->hasAggs)
 881                 {
 882                         /*
 883                          * Ungrouped aggregate will certainly want all the input
 884                          * tuples.
 885                          */
 886                         sub_tuple_fraction = 0.0;
 887                 }
 888                 else if (parse->distinctClause)
 889                 {
 890                         /*
 891                          * SELECT DISTINCT, like GROUP, will absorb an unpredictable
 892                          * number of input tuples per output tuple.  Handle the same
 893                          * way.
 894                          */
 895                         if (sub_tuple_fraction >= 1.0)
 896                                 sub_tuple_fraction = 0.25;
 897                 }
 898
 899                 /*
 900                  * Generate the best unsorted and presorted paths for this Query
 901                  * (but note there may not be any presorted path).
 902                  */
 903                 query_planner(parse, sub_tlist, sub_tuple_fraction,
 904                                           &cheapest_path, &sorted_path);
 905
 906                 /*
 907                  * We couldn't canonicalize group_pathkeys and sort_pathkeys
 908                  * before running query_planner(), so do it now.
 909                  */
 910                 group_pathkeys = canonicalize_pathkeys(parse, group_pathkeys);
 911                 sort_pathkeys = canonicalize_pathkeys(parse, sort_pathkeys);
 912
 913                 /*
 914                  * Consider whether we might want to use hashed grouping.
 915                  */
 916                 if (parse->groupClause)
 917                 {
 918                         List       *groupExprs;
 919                         double          cheapest_path_rows;
 920                         int                     cheapest_path_width;
 921
 922                         /*
 923                          * Beware in this section of the possibility that
 924                          * cheapest_path->parent is NULL.  This could happen if user
 925                          * does something silly like SELECT 'foo' GROUP BY 1;
 926                          */
 927                         if (cheapest_path->parent)
 928                         {
 929                                 cheapest_path_rows = cheapest_path->parent->rows;
 930                                 cheapest_path_width = cheapest_path->parent->width;
 931                         }
 932                         else
 933                         {
 934                                 cheapest_path_rows = 1; /* assume non-set result */
 935                                 cheapest_path_width = 100;              /* arbitrary */
 936                         }
 937
 938                         /*
 939                          * Always estimate the number of groups.  We can't do this
 940                          * until after running query_planner(), either.
 941                          */
 942                         groupExprs = get_sortgrouplist_exprs(parse->groupClause,
 943                                                                                                  parse->targetList);
 944                         dNumGroups = estimate_num_groups(parse,
 945                                                                                          groupExprs,
 946                                                                                          cheapest_path_rows);
 947                         /* Also want it as a long int --- but 'ware overflow! */
 948                         numGroups = (long) Min(dNumGroups, (double) LONG_MAX);
 949
 950                         /*
 951                          * Check can't-do-it conditions, including whether the
 952                          * grouping operators are hashjoinable.
 953                          *
 954                          * Executor doesn't support hashed aggregation with DISTINCT
 955                          * aggregates.  (Doing so would imply storing *all* the input
 956                          * values in the hash table, which seems like a certain
 957                          * loser.)
 958                          */
 959                         if (!enable_hashagg || !hash_safe_grouping(parse))
 960                                 use_hashed_grouping = false;
 961                         else if (parse->hasAggs &&
 962                                          (contain_distinct_agg_clause((Node *) tlist) ||
 963                                           contain_distinct_agg_clause(parse->havingQual)))
 964                                 use_hashed_grouping = false;
 965                         else
 966                         {
 967                                 /*
 968                                  * Use hashed grouping if (a) we think we can fit the
 969                                  * hashtable into SortMem, *and* (b) the estimated cost is
 970                                  * no more than doing it the other way.  While avoiding
 971                                  * the need for sorted input is usually a win, the fact
 972                                  * that the output won't be sorted may be a loss; so we
 973                                  * need to do an actual cost comparison.
 974                                  *
 975                                  * In most cases we have no good way to estimate the size of
 976                                  * the transition value needed by an aggregate;
 977                                  * arbitrarily assume it is 100 bytes.  Also set the
 978                                  * overhead per hashtable entry at 64 bytes.
 979                                  */
 980                                 int                     hashentrysize = cheapest_path_width + 64 + numAggs * 100;
 981
 982                                 if (hashentrysize * dNumGroups <= SortMem * 1024L)
 983                                 {
 984                                         /*
 985                                          * Okay, do the cost comparison.  We need to consider
 986                                          * cheapest_path + hashagg [+ final sort] versus
 987                                          * either cheapest_path [+ sort] + group or agg [+
 988                                          * final sort] or presorted_path + group or agg [+
 989                                          * final sort] where brackets indicate a step that may
 990                                          * not be needed. We assume query_planner() will have
 991                                          * returned a presorted path only if it's a winner
 992                                          * compared to cheapest_path for this purpose.
 993                                          *
 994                                          * These path variables are dummies that just hold cost
 995                                          * fields; we don't make actual Paths for these steps.
 996                                          */
 997                                         Path            hashed_p;
 998                                         Path            sorted_p;
 999
1000                                         cost_agg(&hashed_p, parse,
1001                                                          AGG_HASHED, numAggs,
1002                                                          numGroupCols, dNumGroups,
1003                                                          cheapest_path->startup_cost,
1004                                                          cheapest_path->total_cost,
1005                                                          cheapest_path_rows);
1006                                         /* Result of hashed agg is always unsorted */
1007                                         if (sort_pathkeys)
1008                                                 cost_sort(&hashed_p, parse, sort_pathkeys,
1009                                                                   hashed_p.total_cost,
1010                                                                   dNumGroups,
1011                                                                   cheapest_path_width);
1012
1013                                         if (sorted_path)
1014                                         {
1015                                                 sorted_p.startup_cost = sorted_path->startup_cost;
1016                                                 sorted_p.total_cost = sorted_path->total_cost;
1017                                                 current_pathkeys = sorted_path->pathkeys;
1018                                         }
1019                                         else
1020                                         {
1021                                                 sorted_p.startup_cost = cheapest_path->startup_cost;
1022                                                 sorted_p.total_cost = cheapest_path->total_cost;
1023                                                 current_pathkeys = cheapest_path->pathkeys;
1024                                         }
1025                                         if (!pathkeys_contained_in(group_pathkeys,
1026                                                                                            current_pathkeys))
1027                                         {
1028                                                 cost_sort(&sorted_p, parse, group_pathkeys,
1029                                                                   sorted_p.total_cost,
1030                                                                   cheapest_path_rows,
1031                                                                   cheapest_path_width);
1032                                                 current_pathkeys = group_pathkeys;
1033                                         }
1034                                         if (parse->hasAggs)
1035                                                 cost_agg(&sorted_p, parse,
1036                                                                  AGG_SORTED, numAggs,
1037                                                                  numGroupCols, dNumGroups,
1038                                                                  sorted_p.startup_cost,
1039                                                                  sorted_p.total_cost,
1040                                                                  cheapest_path_rows);
1041                                         else
1042                                                 cost_group(&sorted_p, parse,
1043                                                                    numGroupCols, dNumGroups,
1044                                                                    sorted_p.startup_cost,
1045                                                                    sorted_p.total_cost,
1046                                                                    cheapest_path_rows);
1047                                         /* The Agg or Group node will preserve ordering */
1048                                         if (sort_pathkeys &&
1049                                                 !pathkeys_contained_in(sort_pathkeys,
1050                                                                                            current_pathkeys))
1051                                         {
1052                                                 cost_sort(&sorted_p, parse, sort_pathkeys,
1053                                                                   sorted_p.total_cost,
1054                                                                   dNumGroups,
1055                                                                   cheapest_path_width);
1056                                         }
1057
1058                                         /*
1059                                          * Now make the decision using the top-level tuple
1060                                          * fraction.  First we have to convert an absolute
1061                                          * count (LIMIT) into fractional form.
1062                                          */
1063                                         if (tuple_fraction >= 1.0)
1064                                                 tuple_fraction /= dNumGroups;
1065
1066                                         if (compare_fractional_path_costs(&hashed_p, &sorted_p,
1067                                                                                                           tuple_fraction) < 0)
1068                                         {
1069                                                 /* Hashed is cheaper, so use it */
1070                                                 use_hashed_grouping = true;
1071                                         }
1072                                 }
1073                         }
1074                 }
1075
1076                 /*
1077                  * Select the best path and create a plan to execute it.
1078                  *
1079                  * If we are doing hashed grouping, we will always read all the input
1080                  * tuples, so use the cheapest-total path.      Otherwise, trust
1081                  * query_planner's decision about which to use.
1082                  */
1083                 if (sorted_path && !use_hashed_grouping)
1084                 {
1085                         result_plan = create_plan(parse, sorted_path);
1086                         current_pathkeys = sorted_path->pathkeys;
1087                 }
1088                 else
1089                 {
1090                         result_plan = create_plan(parse, cheapest_path);
1091                         current_pathkeys = cheapest_path->pathkeys;
1092                 }
1093
1094                 /*
1095                  * create_plan() returns a plan with just a "flat" tlist of
1096                  * required Vars.  Usually we need to insert the sub_tlist as the
1097                  * tlist of the top plan node.  However, we can skip that if we
1098                  * determined that whatever query_planner chose to return will be
1099                  * good enough.
1100                  */
1101                 if (need_tlist_eval)
1102                 {
1103                         /*
1104                          * If the top-level plan node is one that cannot do expression
1105                          * evaluation, we must insert a Result node to project the
1106                          * desired tlist. Currently, the only plan node we might see
1107                          * here that falls into that category is Append.
1108                          */
1109                         if (IsA(result_plan, Append))
1110                         {
1111                                 result_plan = (Plan *) make_result(sub_tlist, NULL,
1112                                                                                                    result_plan);
1113                         }
1114                         else
1115                         {
1116                                 /*
1117                                  * Otherwise, just replace the subplan's flat tlist with
1118                                  * the desired tlist.
1119                                  */
1120                                 result_plan->targetlist = sub_tlist;
1121                         }
1122
1123                         /*
1124                          * Also, account for the cost of evaluation of the sub_tlist.
1125                          *
1126                          * Up to now, we have only been dealing with "flat" tlists,
1127                          * containing just Vars.  So their evaluation cost is zero
1128                          * according to the model used by cost_qual_eval() (or if you
1129                          * prefer, the cost is factored into cpu_tuple_cost).  Thus we
1130                          * can avoid accounting for tlist cost throughout
1131                          * query_planner() and subroutines.  But now we've inserted a
1132                          * tlist that might contain actual operators, sub-selects, etc
1133                          * --- so we'd better account for its cost.
1134                          *
1135                          * Below this point, any tlist eval cost for added-on nodes
1136                          * should be accounted for as we create those nodes.
1137                          * Presently, of the node types we can add on, only Agg and
1138                          * Group project new tlists (the rest just copy their input
1139                          * tuples) --- so make_agg() and make_group() are responsible
1140                          * for computing the added cost.
1141                          */
1142                         cost_qual_eval(&tlist_cost, sub_tlist);
1143                         result_plan->startup_cost += tlist_cost.startup;
1144                         result_plan->total_cost += tlist_cost.startup +
1145                                 tlist_cost.per_tuple * result_plan->plan_rows;
1146                 }
1147                 else
1148                 {
1149                         /*
1150                          * Since we're using query_planner's tlist and not the one
1151                          * make_subplanTargetList calculated, we have to refigure any
1152                          * grouping-column indexes make_subplanTargetList computed.
1153                          */
1154                         locate_grouping_columns(parse, tlist, result_plan->targetlist,
1155                                                                         groupColIdx);
1156                 }
1157
1158                 /*
1159                  * Insert AGG or GROUP node if needed, plus an explicit sort step
1160                  * if necessary.
1161                  *
1162                  * HAVING clause, if any, becomes qual of the Agg node
1163                  */
1164                 if (use_hashed_grouping)
1165                 {
1166                         /* Hashed aggregate plan --- no sort needed */
1167                         result_plan = (Plan *) make_agg(parse,
1168                                                                                         tlist,
1169                                                                                         (List *) parse->havingQual,
1170                                                                                         AGG_HASHED,
1171                                                                                         numGroupCols,
1172                                                                                         groupColIdx,
1173                                                                                         numGroups,
1174                                                                                         numAggs,
1175                                                                                         result_plan);
1176                         /* Hashed aggregation produces randomly-ordered results */
1177                         current_pathkeys = NIL;
1178                 }
1179                 else if (parse->hasAggs)
1180                 {
1181                         /* Plain aggregate plan --- sort if needed */
1182                         AggStrategy aggstrategy;
1183
1184                         if (parse->groupClause)
1185                         {
1186                                 if (!pathkeys_contained_in(group_pathkeys, current_pathkeys))
1187                                 {
1188                                         result_plan = (Plan *)
1189                                                 make_sort_from_groupcols(parse,
1190                                                                                                  parse->groupClause,
1191                                                                                                  groupColIdx,
1192                                                                                                  result_plan);
1193                                         current_pathkeys = group_pathkeys;
1194                                 }
1195                                 aggstrategy = AGG_SORTED;
1196
1197                                 /*
1198                                  * The AGG node will not change the sort ordering of its
1199                                  * groups, so current_pathkeys describes the result too.
1200                                  */
1201                         }
1202                         else
1203                         {
1204                                 aggstrategy = AGG_PLAIN;
1205                                 /* Result will be only one row anyway; no sort order */
1206                                 current_pathkeys = NIL;
1207                         }
1208
1209                         result_plan = (Plan *) make_agg(parse,
1210                                                                                         tlist,
1211                                                                                         (List *) parse->havingQual,
1212                                                                                         aggstrategy,
1213                                                                                         numGroupCols,
1214                                                                                         groupColIdx,
1215                                                                                         numGroups,
1216                                                                                         numAggs,
1217                                                                                         result_plan);
1218                 }
1219                 else
1220                 {
1221                         /*
1222                          * If there are no Aggs, we shouldn't have any HAVING qual
1223                          * anymore
1224                          */
1225                         Assert(parse->havingQual == NULL);
1226
1227                         /*
1228                          * If we have a GROUP BY clause, insert a group node (plus the
1229                          * appropriate sort node, if necessary).
1230                          */
1231                         if (parse->groupClause)
1232                         {
1233                                 /*
1234                                  * Add an explicit sort if we couldn't make the path come
1235                                  * out the way the GROUP node needs it.
1236                                  */
1237                                 if (!pathkeys_contained_in(group_pathkeys, current_pathkeys))
1238                                 {
1239                                         result_plan = (Plan *)
1240                                                 make_sort_from_groupcols(parse,
1241                                                                                                  parse->groupClause,
1242                                                                                                  groupColIdx,
1243                                                                                                  result_plan);
1244                                         current_pathkeys = group_pathkeys;
1245                                 }
1246
1247                                 result_plan = (Plan *) make_group(parse,
1248                                                                                                   tlist,
1249                                                                                                   numGroupCols,
1250                                                                                                   groupColIdx,
1251                                                                                                   dNumGroups,
1252                                                                                                   result_plan);
1253                                 /* The Group node won't change sort ordering */
1254                         }
1255                 }
1256         }                                                       /* end of if (setOperations) */
1257
1258         /*
1259          * If we were not able to make the plan come out in the right order,
1260          * add an explicit sort step.
1261          */
1262         if (parse->sortClause)
1263         {
1264                 if (!pathkeys_contained_in(sort_pathkeys, current_pathkeys))
1265                 {
1266                         result_plan = (Plan *)
1267                                 make_sort_from_sortclauses(parse,
1268                                                                                    tlist,
1269                                                                                    result_plan,
1270                                                                                    parse->sortClause);
1271                         current_pathkeys = sort_pathkeys;
1272                 }
1273         }
1274
1275         /*
1276          * If there is a DISTINCT clause, add the UNIQUE node.
1277          */
1278         if (parse->distinctClause)
1279         {
1280                 result_plan = (Plan *) make_unique(tlist, result_plan,
1281                                                                                    parse->distinctClause);
1282
1283                 /*
1284                  * If there was grouping or aggregation, leave plan_rows as-is
1285                  * (ie, assume the result was already mostly unique).  If not,
1286                  * it's reasonable to assume the UNIQUE filter has effects
1287                  * comparable to GROUP BY.
1288                  */
1289                 if (!parse->groupClause && !parse->hasAggs)
1290                 {
1291                         List       *distinctExprs;
1292
1293                         distinctExprs = get_sortgrouplist_exprs(parse->distinctClause,
1294                                                                                                         parse->targetList);
1295                         result_plan->plan_rows = estimate_num_groups(parse,
1296                                                                                                                  distinctExprs,
1297                                                                                                  result_plan->plan_rows);
1298                 }
1299         }
1300
1301         /*
1302          * Finally, if there is a LIMIT/OFFSET clause, add the LIMIT node.
1303          */
1304         if (parse->limitOffset || parse->limitCount)
1305         {
1306                 result_plan = (Plan *) make_limit(tlist, result_plan,
1307                                                                                   parse->limitOffset,
1308                                                                                   parse->limitCount);
1309         }
1310
1311         /*
1312          * Return the actual output ordering in query_pathkeys for possible
1313          * use by an outer query level.
1314          */
1315         parse->query_pathkeys = current_pathkeys;
1316
1317         return result_plan;
1318 }
1319
1320 /*
1321  * hash_safe_grouping - are grouping operators hashable?
1322  *
1323  * We assume hashed aggregation will work if the datatype's equality operator
1324  * is marked hashjoinable.
1325  */
1326 static bool
1327 hash_safe_grouping(Query *parse)
1328 {
1329         List       *gl;
1330
1331         foreach(gl, parse->groupClause)
1332         {
1333                 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1334                 TargetEntry *tle = get_sortgroupclause_tle(grpcl, parse->targetList);
1335                 Operator        optup;
1336                 bool            oprcanhash;
1337
1338                 optup = equality_oper(tle->resdom->restype, true);
1339                 if (!optup)
1340                         return false;
1341                 oprcanhash = ((Form_pg_operator) GETSTRUCT(optup))->oprcanhash;
1342                 ReleaseSysCache(optup);
1343                 if (!oprcanhash)
1344                         return false;
1345         }
1346         return true;
1347 }
1348
1349 /*---------------
1350  * make_subplanTargetList
1351  *        Generate appropriate target list when grouping is required.
1352  *
1353  * When grouping_planner inserts Aggregate or Group plan nodes above
1354  * the result of query_planner, we typically want to pass a different
1355  * target list to query_planner than the outer plan nodes should have.
1356  * This routine generates the correct target list for the subplan.
1357  *
1358  * The initial target list passed from the parser already contains entries
1359  * for all ORDER BY and GROUP BY expressions, but it will not have entries
1360  * for variables used only in HAVING clauses; so we need to add those
1361  * variables to the subplan target list.  Also, if we are doing either
1362  * grouping or aggregation, we flatten all expressions except GROUP BY items
1363  * into their component variables; the other expressions will be computed by
1364  * the inserted nodes rather than by the subplan.  For example,
1365  * given a query like
1366  *              SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
1367  * we want to pass this targetlist to the subplan:
1368  *              a,b,c,d,a+b
1369  * where the a+b target will be used by the Sort/Group steps, and the
1370  * other targets will be used for computing the final results.  (In the
1371  * above example we could theoretically suppress the a and b targets and
1372  * pass down only c,d,a+b, but it's not really worth the trouble to
1373  * eliminate simple var references from the subplan.  We will avoid doing
1374  * the extra computation to recompute a+b at the outer level; see
1375  * replace_vars_with_subplan_refs() in setrefs.c.)
1376  *
1377  * If we are grouping or aggregating, *and* there are no non-Var grouping
1378  * expressions, then the returned tlist is effectively dummy; we do not
1379  * need to force it to be evaluated, because all the Vars it contains
1380  * should be present in the output of query_planner anyway.
1381  *
1382  * 'parse' is the query being processed.
1383  * 'tlist' is the query's target list.
1384  * 'groupColIdx' receives an array of column numbers for the GROUP BY
1385  *                      expressions (if there are any) in the subplan's target list.
1386  * 'need_tlist_eval' is set true if we really need to evaluate the
1387  *                      result tlist.
1388  *
1389  * The result is the targetlist to be passed to the subplan.
1390  *---------------
1391  */
1392 static List *
1393 make_subplanTargetList(Query *parse,
1394                                            List *tlist,
1395                                            AttrNumber **groupColIdx,
1396                                            bool *need_tlist_eval)
1397 {
1398         List       *sub_tlist;
1399         List       *extravars;
1400         int                     numCols;
1401
1402         *groupColIdx = NULL;
1403
1404         /*
1405          * If we're not grouping or aggregating, nothing to do here;
1406          * query_planner should receive the unmodified target list.
1407          */
1408         if (!parse->hasAggs && !parse->groupClause)
1409         {
1410                 *need_tlist_eval = true;
1411                 return tlist;
1412         }
1413
1414         /*
1415          * Otherwise, start with a "flattened" tlist (having just the vars
1416          * mentioned in the targetlist and HAVING qual --- but not upper-
1417          * level Vars; they will be replaced by Params later on).
1418          */
1419         sub_tlist = flatten_tlist(tlist);
1420         extravars = pull_var_clause(parse->havingQual, false);
1421         sub_tlist = add_to_flat_tlist(sub_tlist, extravars);
1422         freeList(extravars);
1423         *need_tlist_eval = false;       /* only eval if not flat tlist */
1424
1425         /*
1426          * If grouping, create sub_tlist entries for all GROUP BY expressions
1427          * (GROUP BY items that are simple Vars should be in the list
1428          * already), and make an array showing where the group columns are in
1429          * the sub_tlist.
1430          */
1431         numCols = length(parse->groupClause);
1432         if (numCols > 0)
1433         {
1434                 int                     keyno = 0;
1435                 AttrNumber *grpColIdx;
1436                 List       *gl;
1437
1438                 grpColIdx = (AttrNumber *) palloc(sizeof(AttrNumber) * numCols);
1439                 *groupColIdx = grpColIdx;
1440
1441                 foreach(gl, parse->groupClause)
1442                 {
1443                         GroupClause *grpcl = (GroupClause *) lfirst(gl);
1444                         Node       *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1445                         TargetEntry *te = NULL;
1446                         List       *sl;
1447
1448                         /* Find or make a matching sub_tlist entry */
1449                         foreach(sl, sub_tlist)
1450                         {
1451                                 te = (TargetEntry *) lfirst(sl);
1452                                 if (equal(groupexpr, te->expr))
1453                                         break;
1454                         }
1455                         if (!sl)
1456                         {
1457                                 te = makeTargetEntry(makeResdom(length(sub_tlist) + 1,
1458                                                                                                 exprType(groupexpr),
1459                                                                                                 exprTypmod(groupexpr),
1460                                                                                                 NULL,
1461                                                                                                 false),
1462                                                                          (Expr *) groupexpr);
1463                                 sub_tlist = lappend(sub_tlist, te);
1464                                 *need_tlist_eval = true;                /* it's not flat anymore */
1465                         }
1466
1467                         /* and save its resno */
1468                         grpColIdx[keyno++] = te->resdom->resno;
1469                 }
1470         }
1471
1472         return sub_tlist;
1473 }
1474
1475 /*
1476  * locate_grouping_columns
1477  *              Locate grouping columns in the tlist chosen by query_planner.
1478  *
1479  * This is only needed if we don't use the sub_tlist chosen by
1480  * make_subplanTargetList.      We have to forget the column indexes found
1481  * by that routine and re-locate the grouping vars in the real sub_tlist.
1482  */
1483 static void
1484 locate_grouping_columns(Query *parse,
1485                                                 List *tlist,
1486                                                 List *sub_tlist,
1487                                                 AttrNumber *groupColIdx)
1488 {
1489         int                     keyno = 0;
1490         List       *gl;
1491
1492         /*
1493          * No work unless grouping.
1494          */
1495         if (!parse->groupClause)
1496         {
1497                 Assert(groupColIdx == NULL);
1498                 return;
1499         }
1500         Assert(groupColIdx != NULL);
1501
1502         foreach(gl, parse->groupClause)
1503         {
1504                 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1505                 Node       *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1506                 TargetEntry *te = NULL;
1507                 List       *sl;
1508
1509                 foreach(sl, sub_tlist)
1510                 {
1511                         te = (TargetEntry *) lfirst(sl);
1512                         if (equal(groupexpr, te->expr))
1513                                 break;
1514                 }
1515                 if (!sl)
1516                         elog(ERROR, "failed to locate grouping columns");
1517
1518                 groupColIdx[keyno++] = te->resdom->resno;
1519         }
1520 }
1521
1522 /*
1523  * postprocess_setop_tlist
1524  *        Fix up targetlist returned by plan_set_operations().
1525  *
1526  * We need to transpose sort key info from the orig_tlist into new_tlist.
1527  * NOTE: this would not be good enough if we supported resjunk sort keys
1528  * for results of set operations --- then, we'd need to project a whole
1529  * new tlist to evaluate the resjunk columns.  For now, just ereport if we
1530  * find any resjunk columns in orig_tlist.
1531  */
1532 static List *
1533 postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
1534 {
1535         List       *l;
1536
1537         foreach(l, new_tlist)
1538         {
1539                 TargetEntry *new_tle = (TargetEntry *) lfirst(l);
1540                 TargetEntry *orig_tle;
1541
1542                 /* ignore resjunk columns in setop result */
1543                 if (new_tle->resdom->resjunk)
1544                         continue;
1545
1546                 Assert(orig_tlist != NIL);
1547                 orig_tle = (TargetEntry *) lfirst(orig_tlist);
1548                 orig_tlist = lnext(orig_tlist);
1549                 if (orig_tle->resdom->resjunk)  /* should not happen */
1550                         elog(ERROR, "resjunk output columns are not implemented");
1551                 Assert(new_tle->resdom->resno == orig_tle->resdom->resno);
1552                 Assert(new_tle->resdom->restype == orig_tle->resdom->restype);
1553                 new_tle->resdom->ressortgroupref = orig_tle->resdom->ressortgroupref;
1554         }
1555         if (orig_tlist != NIL)
1556                 elog(ERROR, "resjunk output columns are not implemented");
1557         return new_tlist;
1558 }