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

   1 /*-------------------------------------------------------------------------
   2  *
   3  * planner.c
   4  *        The query optimizer external interface.
   5  *
   6  * Portions Copyright (c) 1996-2002, PostgreSQL Global Development Group
   7  * Portions Copyright (c) 1994, Regents of the University of California
   8  *
   9  *
  10  * IDENTIFICATION
  11  *        $Header: /cvsroot/pgsql/src/backend/optimizer/plan/planner.c,v 1.156 2003/07/03 19:07:20 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
 107                  * doesn't want 'em all.  Optimize for 10% retrieval (you
 108                  * gotta better 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
 125          * run 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 step,
 186          * 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,
 202          * we can avoid the expense of doing flatten_join_alias_vars().  Also
 203          * check for outer joins --- if none, we can skip reduce_outer_joins().
 204          * This must be done after we have done pull_up_subqueries, of course.
 205          */
 206         parse->hasJoinRTEs = false;
 207         hasOuterJoins = false;
 208         foreach(lst, parse->rtable)
 209         {
 210                 RangeTblEntry *rte = (RangeTblEntry *) lfirst(lst);
 211
 212                 if (rte->rtekind == RTE_JOIN)
 213                 {
 214                         parse->hasJoinRTEs = true;
 215                         if (IS_OUTER_JOIN(rte->jointype))
 216                         {
 217                                 hasOuterJoins = true;
 218                                 /* Can quit scanning once we find an outer join */
 219                                 break;
 220                         }
 221                 }
 222         }
 223
 224         /*
 225          * Do expression preprocessing on targetlist and quals.
 226          */
 227         parse->targetList = (List *)
 228                 preprocess_expression(parse, (Node *) parse->targetList,
 229                                                           EXPRKIND_TARGET);
 230
 231         preprocess_qual_conditions(parse, (Node *) parse->jointree);
 232
 233         parse->havingQual = preprocess_expression(parse, parse->havingQual,
 234                                                                                           EXPRKIND_QUAL);
 235
 236         parse->limitOffset = preprocess_expression(parse, parse->limitOffset,
 237                                                                                            EXPRKIND_LIMIT);
 238         parse->limitCount = preprocess_expression(parse, parse->limitCount,
 239                                                                                           EXPRKIND_LIMIT);
 240
 241         parse->in_info_list = (List *)
 242                 preprocess_expression(parse, (Node *) parse->in_info_list,
 243                                                           EXPRKIND_ININFO);
 244
 245         /* Also need to preprocess expressions for function RTEs */
 246         foreach(lst, parse->rtable)
 247         {
 248                 RangeTblEntry *rte = (RangeTblEntry *) lfirst(lst);
 249
 250                 if (rte->rtekind == RTE_FUNCTION)
 251                         rte->funcexpr = preprocess_expression(parse, rte->funcexpr,
 252                                                                                                   EXPRKIND_RTFUNC);
 253         }
 254
 255         /*
 256          * A HAVING clause without aggregates is equivalent to a WHERE clause
 257          * (except it can only refer to grouped fields).  Transfer any
 258          * agg-free clauses of the HAVING qual into WHERE.      This may seem like
 259          * wasting cycles to cater to stupidly-written queries, but there are
 260          * other reasons for doing it.  Firstly, if the query contains no aggs
 261          * at all, then we aren't going to generate an Agg plan node, and so
 262          * there'll be no place to execute HAVING conditions; without this
 263          * transfer, we'd lose the HAVING condition entirely, which is wrong.
 264          * Secondly, when we push down a qual condition into a sub-query, it's
 265          * easiest to push the qual into HAVING always, in case it contains
 266          * aggs, and then let this code sort it out.
 267          *
 268          * Note that both havingQual and parse->jointree->quals are in
 269          * implicitly-ANDed-list form at this point, even though they are
 270          * declared as Node *.
 271          */
 272         newHaving = NIL;
 273         foreach(lst, (List *) parse->havingQual)
 274         {
 275                 Node       *havingclause = (Node *) lfirst(lst);
 276
 277                 if (contain_agg_clause(havingclause))
 278                         newHaving = lappend(newHaving, havingclause);
 279                 else
 280                         parse->jointree->quals = (Node *)
 281                                 lappend((List *) parse->jointree->quals, havingclause);
 282         }
 283         parse->havingQual = (Node *) newHaving;
 284
 285         /*
 286          * If we have any outer joins, try to reduce them to plain inner joins.
 287          * This step is most easily done after we've done expression preprocessing.
 288          */
 289         if (hasOuterJoins)
 290                 reduce_outer_joins(parse);
 291
 292         /*
 293          * See if we can simplify the jointree; opportunities for this may come
 294          * from having pulled up subqueries, or from flattening explicit JOIN
 295          * syntax.  We must do this after flattening JOIN alias variables, since
 296          * eliminating explicit JOIN nodes from the jointree will cause
 297          * get_relids_for_join() to fail.  But it should happen after
 298          * reduce_outer_joins, anyway.
 299          */
 300         parse->jointree = (FromExpr *)
 301                 simplify_jointree(parse, (Node *) parse->jointree);
 302
 303         /*
 304          * Do the main planning.  If we have an inherited target relation,
 305          * that needs special processing, else go straight to
 306          * grouping_planner.
 307          */
 308         if (parse->resultRelation &&
 309                 (lst = expand_inherited_rtentry(parse, parse->resultRelation,
 310                                                                                 false)) != NIL)
 311                 plan = inheritance_planner(parse, lst);
 312         else
 313                 plan = grouping_planner(parse, tuple_fraction);
 314
 315         /*
 316          * If any subplans were generated, or if we're inside a subplan, build
 317          * initPlan list and extParam/allParam sets for plan nodes.
 318          */
 319         if (PlannerPlanId != saved_planid || PlannerQueryLevel > 1)
 320         {
 321                 Cost    initplan_cost = 0;
 322
 323                 /* Prepare extParam/allParam sets for all nodes in tree */
 324                 SS_finalize_plan(plan, parse->rtable);
 325
 326                 /*
 327                  * SS_finalize_plan doesn't handle initPlans, so we have to manually
 328                  * attach them to the topmost plan node, and add their extParams to
 329                  * the topmost node's, too.
 330                  *
 331                  * We also add the total_cost of each initPlan to the startup cost
 332                  * of the top node.  This is a conservative overestimate, since in
 333                  * fact each initPlan might be executed later than plan startup, or
 334                  * even not at all.
 335                  */
 336                 plan->initPlan = PlannerInitPlan;
 337
 338                 foreach(lst, plan->initPlan)
 339                 {
 340                         SubPlan    *initplan = (SubPlan *) lfirst(lst);
 341
 342                         plan->extParam = bms_add_members(plan->extParam,
 343                                                                                          initplan->plan->extParam);
 344                         initplan_cost += initplan->plan->total_cost;
 345                 }
 346
 347                 plan->startup_cost += initplan_cost;
 348                 plan->total_cost += initplan_cost;
 349         }
 350
 351         /* Return to outer subquery context */
 352         PlannerQueryLevel--;
 353         PlannerInitPlan = saved_initplan;
 354         /* we do NOT restore PlannerPlanId; that's not an oversight! */
 355
 356         return plan;
 357 }
 358
 359 /*
 360  * preprocess_expression
 361  *              Do subquery_planner's preprocessing work for an expression,
 362  *              which can be a targetlist, a WHERE clause (including JOIN/ON
 363  *              conditions), or a HAVING clause.
 364  */
 365 static Node *
 366 preprocess_expression(Query *parse, Node *expr, int kind)
 367 {
 368         /*
 369          * If the query has any join RTEs, replace join alias variables with
 370          * base-relation variables. We must do this before sublink processing,
 371          * else sublinks expanded out from join aliases wouldn't get processed.
 372          */
 373         if (parse->hasJoinRTEs)
 374                 expr = flatten_join_alias_vars(parse, expr);
 375
 376         /*
 377          * Simplify constant expressions.
 378          *
 379          * Note that at this point quals have not yet been converted to
 380          * implicit-AND form, so we can apply eval_const_expressions directly.
 381          */
 382         expr = eval_const_expressions(expr);
 383
 384         /*
 385          * If it's a qual or havingQual, canonicalize it, and convert it to
 386          * implicit-AND format.
 387          *
 388          * XXX Is there any value in re-applying eval_const_expressions after
 389          * canonicalize_qual?
 390          */
 391         if (kind == EXPRKIND_QUAL)
 392         {
 393                 expr = (Node *) canonicalize_qual((Expr *) expr, true);
 394
 395 #ifdef OPTIMIZER_DEBUG
 396                 printf("After canonicalize_qual()\n");
 397                 pprint(expr);
 398 #endif
 399         }
 400
 401         /* Expand SubLinks to SubPlans */
 402         if (parse->hasSubLinks)
 403                 expr = SS_process_sublinks(expr, (kind == EXPRKIND_QUAL));
 404
 405         /*
 406          * XXX do not insert anything here unless you have grokked the comments
 407          * in SS_replace_correlation_vars ...
 408          */
 409
 410         /* Replace uplevel vars with Param nodes */
 411         if (PlannerQueryLevel > 1)
 412                 expr = SS_replace_correlation_vars(expr);
 413
 414         return expr;
 415 }
 416
 417 /*
 418  * preprocess_qual_conditions
 419  *              Recursively scan the query's jointree and do subquery_planner's
 420  *              preprocessing work on each qual condition found therein.
 421  */
 422 static void
 423 preprocess_qual_conditions(Query *parse, Node *jtnode)
 424 {
 425         if (jtnode == NULL)
 426                 return;
 427         if (IsA(jtnode, RangeTblRef))
 428         {
 429                 /* nothing to do here */
 430         }
 431         else if (IsA(jtnode, FromExpr))
 432         {
 433                 FromExpr   *f = (FromExpr *) jtnode;
 434                 List       *l;
 435
 436                 foreach(l, f->fromlist)
 437                         preprocess_qual_conditions(parse, lfirst(l));
 438
 439                 f->quals = preprocess_expression(parse, f->quals, EXPRKIND_QUAL);
 440         }
 441         else if (IsA(jtnode, JoinExpr))
 442         {
 443                 JoinExpr   *j = (JoinExpr *) jtnode;
 444
 445                 preprocess_qual_conditions(parse, j->larg);
 446                 preprocess_qual_conditions(parse, j->rarg);
 447
 448                 j->quals = preprocess_expression(parse, j->quals, EXPRKIND_QUAL);
 449         }
 450         else
 451                 elog(ERROR, "preprocess_qual_conditions: unexpected node type %d",
 452                          nodeTag(jtnode));
 453 }
 454
 455 /*--------------------
 456  * inheritance_planner
 457  *        Generate a plan in the case where the result relation is an
 458  *        inheritance set.
 459  *
 460  * We have to handle this case differently from cases where a source
 461  * relation is an inheritance set.      Source inheritance is expanded at
 462  * the bottom of the plan tree (see allpaths.c), but target inheritance
 463  * has to be expanded at the top.  The reason is that for UPDATE, each
 464  * target relation needs a different targetlist matching its own column
 465  * set.  (This is not so critical for DELETE, but for simplicity we treat
 466  * inherited DELETE the same way.)      Fortunately, the UPDATE/DELETE target
 467  * can never be the nullable side of an outer join, so it's OK to generate
 468  * the plan this way.
 469  *
 470  * parse is the querytree produced by the parser & rewriter.
 471  * inheritlist is an integer list of RT indexes for the result relation set.
 472  *
 473  * Returns a query plan.
 474  *--------------------
 475  */
 476 static Plan *
 477 inheritance_planner(Query *parse, List *inheritlist)
 478 {
 479         int                     parentRTindex = parse->resultRelation;
 480         Oid                     parentOID = getrelid(parentRTindex, parse->rtable);
 481         int                     mainrtlength = length(parse->rtable);
 482         List       *subplans = NIL;
 483         List       *tlist = NIL;
 484         List       *l;
 485
 486         foreach(l, inheritlist)
 487         {
 488                 int                     childRTindex = lfirsti(l);
 489                 Oid                     childOID = getrelid(childRTindex, parse->rtable);
 490                 int                     subrtlength;
 491                 Query      *subquery;
 492                 Plan       *subplan;
 493
 494                 /* Generate modified query with this rel as target */
 495                 subquery = (Query *) adjust_inherited_attrs((Node *) parse,
 496                                                                                                 parentRTindex, parentOID,
 497                                                                                                  childRTindex, childOID);
 498                 /* Generate plan */
 499                 subplan = grouping_planner(subquery, 0.0 /* retrieve all tuples */ );
 500                 subplans = lappend(subplans, subplan);
 501                 /*
 502                  * It's possible that additional RTEs got added to the rangetable
 503                  * due to expansion of inherited source tables (see allpaths.c).
 504                  * If so, we must copy 'em back to the main parse tree's rtable.
 505                  *
 506                  * XXX my goodness this is ugly.  Really need to think about ways
 507                  * to rein in planner's habit of scribbling on its input.
 508                  */
 509                 subrtlength = length(subquery->rtable);
 510                 if (subrtlength > mainrtlength)
 511                 {
 512                         List   *subrt = subquery->rtable;
 513
 514                         while (mainrtlength-- > 0) /* wish we had nthcdr() */
 515                                 subrt = lnext(subrt);
 516                         parse->rtable = nconc(parse->rtable, subrt);
 517                         mainrtlength = subrtlength;
 518                 }
 519                 /* Save preprocessed tlist from first rel for use in Append */
 520                 if (tlist == NIL)
 521                         tlist = subplan->targetlist;
 522         }
 523
 524         /* Save the target-relations list for the executor, too */
 525         parse->resultRelations = inheritlist;
 526
 527         /* Mark result as unordered (probably unnecessary) */
 528         parse->query_pathkeys = NIL;
 529
 530         return (Plan *) make_append(subplans, true, tlist);
 531 }
 532
 533 /*--------------------
 534  * grouping_planner
 535  *        Perform planning steps related to grouping, aggregation, etc.
 536  *        This primarily means adding top-level processing to the basic
 537  *        query plan produced by query_planner.
 538  *
 539  * parse is the querytree produced by the parser & rewriter.
 540  * tuple_fraction is the fraction of tuples we expect will be retrieved
 541  *
 542  * tuple_fraction is interpreted as follows:
 543  *        0: expect all tuples to be retrieved (normal case)
 544  *        0 < tuple_fraction < 1: expect the given fraction of tuples available
 545  *              from the plan to be retrieved
 546  *        tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
 547  *              expected to be retrieved (ie, a LIMIT specification)
 548  *
 549  * Returns a query plan.  Also, parse->query_pathkeys is returned as the
 550  * actual output ordering of the plan (in pathkey format).
 551  *--------------------
 552  */
 553 static Plan *
 554 grouping_planner(Query *parse, double tuple_fraction)
 555 {
 556         List       *tlist = parse->targetList;
 557         Plan       *result_plan;
 558         List       *current_pathkeys;
 559         List       *sort_pathkeys;
 560
 561         if (parse->setOperations)
 562         {
 563                 /*
 564                  * Construct the plan for set operations.  The result will not
 565                  * need any work except perhaps a top-level sort and/or LIMIT.
 566                  */
 567                 result_plan = plan_set_operations(parse);
 568
 569                 /*
 570                  * We should not need to call preprocess_targetlist, since we must
 571                  * be in a SELECT query node.  Instead, use the targetlist
 572                  * returned by plan_set_operations (since this tells whether it
 573                  * returned any resjunk columns!), and transfer any sort key
 574                  * information from the original tlist.
 575                  */
 576                 Assert(parse->commandType == CMD_SELECT);
 577
 578                 tlist = postprocess_setop_tlist(result_plan->targetlist, tlist);
 579
 580                 /*
 581                  * Can't handle FOR UPDATE here (parser should have checked
 582                  * already, but let's make sure).
 583                  */
 584                 if (parse->rowMarks)
 585                         elog(ERROR, "SELECT FOR UPDATE is not allowed with UNION/INTERSECT/EXCEPT");
 586
 587                 /*
 588                  * We set current_pathkeys NIL indicating we do not know sort
 589                  * order.  This is correct when the top set operation is UNION
 590                  * ALL, since the appended-together results are unsorted even if
 591                  * the subplans were sorted.  For other set operations we could be
 592                  * smarter --- room for future improvement!
 593                  */
 594                 current_pathkeys = NIL;
 595
 596                 /*
 597                  * Calculate pathkeys that represent ordering requirements
 598                  */
 599                 sort_pathkeys = make_pathkeys_for_sortclauses(parse->sortClause,
 600                                                                                                           tlist);
 601                 sort_pathkeys = canonicalize_pathkeys(parse, sort_pathkeys);
 602         }
 603         else
 604         {
 605                 /* No set operations, do regular planning */
 606                 List       *sub_tlist;
 607                 List       *group_pathkeys;
 608                 AttrNumber *groupColIdx = NULL;
 609                 bool            need_tlist_eval = true;
 610                 QualCost        tlist_cost;
 611                 double          sub_tuple_fraction;
 612                 Path       *cheapest_path;
 613                 Path       *sorted_path;
 614                 double          dNumGroups = 0;
 615                 long            numGroups = 0;
 616                 int                     numAggs = 0;
 617                 int                     numGroupCols = length(parse->groupClause);
 618                 bool            use_hashed_grouping = false;
 619
 620                 /* Preprocess targetlist in case we are inside an INSERT/UPDATE. */
 621                 tlist = preprocess_targetlist(tlist,
 622                                                                           parse->commandType,
 623                                                                           parse->resultRelation,
 624                                                                           parse->rtable);
 625
 626                 /*
 627                  * Add TID targets for rels selected FOR UPDATE (should this be
 628                  * done in preprocess_targetlist?).  The executor uses the TID to
 629                  * know which rows to lock, much as for UPDATE or DELETE.
 630                  */
 631                 if (parse->rowMarks)
 632                 {
 633                         List       *l;
 634
 635                         /*
 636                          * We've got trouble if the FOR UPDATE appears inside
 637                          * grouping, since grouping renders a reference to individual
 638                          * tuple CTIDs invalid.  This is also checked at parse time,
 639                          * but that's insufficient because of rule substitution, query
 640                          * pullup, etc.
 641                          */
 642                         CheckSelectForUpdate(parse);
 643
 644                         /*
 645                          * Currently the executor only supports FOR UPDATE at top
 646                          * level
 647                          */
 648                         if (PlannerQueryLevel > 1)
 649                                 elog(ERROR, "SELECT FOR UPDATE is not allowed in subselects");
 650
 651                         foreach(l, parse->rowMarks)
 652                         {
 653                                 Index           rti = lfirsti(l);
 654                                 char       *resname;
 655                                 Resdom     *resdom;
 656                                 Var                *var;
 657                                 TargetEntry *ctid;
 658
 659                                 resname = (char *) palloc(32);
 660                                 snprintf(resname, 32, "ctid%u", rti);
 661                                 resdom = makeResdom(length(tlist) + 1,
 662                                                                         TIDOID,
 663                                                                         -1,
 664                                                                         resname,
 665                                                                         true);
 666
 667                                 var = makeVar(rti,
 668                                                           SelfItemPointerAttributeNumber,
 669                                                           TIDOID,
 670                                                           -1,
 671                                                           0);
 672
 673                                 ctid = makeTargetEntry(resdom, (Expr *) var);
 674                                 tlist = lappend(tlist, ctid);
 675                         }
 676                 }
 677
 678                 /*
 679                  * Generate appropriate target list for subplan; may be different
 680                  * from tlist if grouping or aggregation is needed.
 681                  */
 682                 sub_tlist = make_subplanTargetList(parse, tlist,
 683                                                                                    &groupColIdx, &need_tlist_eval);
 684
 685                 /*
 686                  * Calculate pathkeys that represent grouping/ordering
 687                  * requirements
 688                  */
 689                 group_pathkeys = make_pathkeys_for_sortclauses(parse->groupClause,
 690                                                                                                            tlist);
 691                 sort_pathkeys = make_pathkeys_for_sortclauses(parse->sortClause,
 692                                                                                                           tlist);
 693
 694                 /*
 695                  * Will need actual number of aggregates for estimating costs.
 696                  * Also, it's possible that optimization has eliminated all
 697                  * aggregates, and we may as well check for that here.
 698                  *
 699                  * Note: we do not attempt to detect duplicate aggregates here;
 700                  * a somewhat-overestimated count is okay for our present purposes.
 701                  */
 702                 if (parse->hasAggs)
 703                 {
 704                         numAggs = count_agg_clause((Node *) tlist) +
 705                                 count_agg_clause(parse->havingQual);
 706                         if (numAggs == 0)
 707                                 parse->hasAggs = false;
 708                 }
 709
 710                 /*
 711                  * Figure out whether we need a sorted result from query_planner.
 712                  *
 713                  * If we have a GROUP BY clause, then we want a result sorted
 714                  * properly for grouping.  Otherwise, if there is an ORDER BY
 715                  * clause, we want to sort by the ORDER BY clause.      (Note: if we
 716                  * have both, and ORDER BY is a superset of GROUP BY, it would be
 717                  * tempting to request sort by ORDER BY --- but that might just
 718                  * leave us failing to exploit an available sort order at all.
 719                  * Needs more thought...)
 720                  */
 721                 if (parse->groupClause)
 722                         parse->query_pathkeys = group_pathkeys;
 723                 else if (parse->sortClause)
 724                         parse->query_pathkeys = sort_pathkeys;
 725                 else
 726                         parse->query_pathkeys = NIL;
 727
 728                 /*
 729                  * Adjust tuple_fraction if we see that we are going to apply
 730                  * limiting/grouping/aggregation/etc.  This is not overridable by
 731                  * the caller, since it reflects plan actions that this routine
 732                  * will certainly take, not assumptions about context.
 733                  */
 734                 if (parse->limitCount != NULL)
 735                 {
 736                         /*
 737                          * A LIMIT clause limits the absolute number of tuples
 738                          * returned. However, if it's not a constant LIMIT then we
 739                          * have to punt; for lack of a better idea, assume 10% of the
 740                          * plan's result is wanted.
 741                          */
 742                         double          limit_fraction = 0.0;
 743
 744                         if (IsA(parse->limitCount, Const))
 745                         {
 746                                 Const      *limitc = (Const *) parse->limitCount;
 747                                 int32           count = DatumGetInt32(limitc->constvalue);
 748
 749                                 /*
 750                                  * A NULL-constant LIMIT represents "LIMIT ALL", which we
 751                                  * treat the same as no limit (ie, expect to retrieve all
 752                                  * the tuples).
 753                                  */
 754                                 if (!limitc->constisnull && count > 0)
 755                                 {
 756                                         limit_fraction = (double) count;
 757                                         /* We must also consider the OFFSET, if present */
 758                                         if (parse->limitOffset != NULL)
 759                                         {
 760                                                 if (IsA(parse->limitOffset, Const))
 761                                                 {
 762                                                         int32           offset;
 763
 764                                                         limitc = (Const *) parse->limitOffset;
 765                                                         offset = DatumGetInt32(limitc->constvalue);
 766                                                         if (!limitc->constisnull && offset > 0)
 767                                                                 limit_fraction += (double) offset;
 768                                                 }
 769                                                 else
 770                                                 {
 771                                                         /* OFFSET is an expression ... punt ... */
 772                                                         limit_fraction = 0.10;
 773                                                 }
 774                                         }
 775                                 }
 776                         }
 777                         else
 778                         {
 779                                 /* LIMIT is an expression ... punt ... */
 780                                 limit_fraction = 0.10;
 781                         }
 782
 783                         if (limit_fraction > 0.0)
 784                         {
 785                                 /*
 786                                  * If we have absolute limits from both caller and LIMIT,
 787                                  * use the smaller value; if one is fractional and the
 788                                  * other absolute, treat the fraction as a fraction of the
 789                                  * absolute value; else we can multiply the two fractions
 790                                  * together.
 791                                  */
 792                                 if (tuple_fraction >= 1.0)
 793                                 {
 794                                         if (limit_fraction >= 1.0)
 795                                         {
 796                                                 /* both absolute */
 797                                                 tuple_fraction = Min(tuple_fraction, limit_fraction);
 798                                         }
 799                                         else
 800                                         {
 801                                                 /* caller absolute, limit fractional */
 802                                                 tuple_fraction *= limit_fraction;
 803                                                 if (tuple_fraction < 1.0)
 804                                                         tuple_fraction = 1.0;
 805                                         }
 806                                 }
 807                                 else if (tuple_fraction > 0.0)
 808                                 {
 809                                         if (limit_fraction >= 1.0)
 810                                         {
 811                                                 /* caller fractional, limit absolute */
 812                                                 tuple_fraction *= limit_fraction;
 813                                                 if (tuple_fraction < 1.0)
 814                                                         tuple_fraction = 1.0;
 815                                         }
 816                                         else
 817                                         {
 818                                                 /* both fractional */
 819                                                 tuple_fraction *= limit_fraction;
 820                                         }
 821                                 }
 822                                 else
 823                                 {
 824                                         /* no info from caller, just use limit */
 825                                         tuple_fraction = limit_fraction;
 826                                 }
 827                         }
 828                 }
 829
 830                 /*
 831                  * With grouping or aggregation, the tuple fraction to pass to
 832                  * query_planner() may be different from what it is at top level.
 833                  */
 834                 sub_tuple_fraction = tuple_fraction;
 835
 836                 if (parse->groupClause)
 837                 {
 838                         /*
 839                          * In GROUP BY mode, we have the little problem that we don't
 840                          * really know how many input tuples will be needed to make a
 841                          * group, so we can't translate an output LIMIT count into an
 842                          * input count.  For lack of a better idea, assume 25% of the
 843                          * input data will be processed if there is any output limit.
 844                          * However, if the caller gave us a fraction rather than an
 845                          * absolute count, we can keep using that fraction (which
 846                          * amounts to assuming that all the groups are about the same
 847                          * size).
 848                          */
 849                         if (sub_tuple_fraction >= 1.0)
 850                                 sub_tuple_fraction = 0.25;
 851
 852                         /*
 853                          * If both GROUP BY and ORDER BY are specified, we will need
 854                          * two levels of sort --- and, therefore, certainly need to
 855                          * read all the input tuples --- unless ORDER BY is a subset
 856                          * of GROUP BY.  (We have not yet canonicalized the pathkeys,
 857                          * so must use the slower noncanonical comparison method.)
 858                          */
 859                         if (parse->groupClause && parse->sortClause &&
 860                                 !noncanonical_pathkeys_contained_in(sort_pathkeys,
 861                                                                                                         group_pathkeys))
 862                                 sub_tuple_fraction = 0.0;
 863                 }
 864                 else if (parse->hasAggs)
 865                 {
 866                         /*
 867                          * Ungrouped aggregate will certainly want all the input
 868                          * tuples.
 869                          */
 870                         sub_tuple_fraction = 0.0;
 871                 }
 872                 else if (parse->distinctClause)
 873                 {
 874                         /*
 875                          * SELECT DISTINCT, like GROUP, will absorb an unpredictable
 876                          * number of input tuples per output tuple.  Handle the same
 877                          * way.
 878                          */
 879                         if (sub_tuple_fraction >= 1.0)
 880                                 sub_tuple_fraction = 0.25;
 881                 }
 882
 883                 /*
 884                  * Generate the best unsorted and presorted paths for this Query
 885                  * (but note there may not be any presorted path).
 886                  */
 887                 query_planner(parse, sub_tlist, sub_tuple_fraction,
 888                                           &cheapest_path, &sorted_path);
 889
 890                 /*
 891                  * We couldn't canonicalize group_pathkeys and sort_pathkeys before
 892                  * running query_planner(), so do it now.
 893                  */
 894                 group_pathkeys = canonicalize_pathkeys(parse, group_pathkeys);
 895                 sort_pathkeys = canonicalize_pathkeys(parse, sort_pathkeys);
 896
 897                 /*
 898                  * Consider whether we might want to use hashed grouping.
 899                  */
 900                 if (parse->groupClause)
 901                 {
 902                         List   *groupExprs;
 903                         double  cheapest_path_rows;
 904                         int             cheapest_path_width;
 905
 906                         /*
 907                          * Beware in this section of the possibility that
 908                          * cheapest_path->parent is NULL.  This could happen if user
 909                          * does something silly like SELECT 'foo' GROUP BY 1;
 910                          */
 911                         if (cheapest_path->parent)
 912                         {
 913                                 cheapest_path_rows = cheapest_path->parent->rows;
 914                                 cheapest_path_width = cheapest_path->parent->width;
 915                         }
 916                         else
 917                         {
 918                                 cheapest_path_rows = 1; /* assume non-set result */
 919                                 cheapest_path_width = 100; /* arbitrary */
 920                         }
 921
 922                         /*
 923                          * Always estimate the number of groups.  We can't do this until
 924                          * after running query_planner(), either.
 925                          */
 926                         groupExprs = get_sortgrouplist_exprs(parse->groupClause,
 927                                                                                                  parse->targetList);
 928                         dNumGroups = estimate_num_groups(parse,
 929                                                                                          groupExprs,
 930                                                                                          cheapest_path_rows);
 931                         /* Also want it as a long int --- but 'ware overflow! */
 932                         numGroups = (long) Min(dNumGroups, (double) LONG_MAX);
 933
 934                         /*
 935                          * Check can't-do-it conditions, including whether the grouping
 936                          * operators are hashjoinable.
 937                          *
 938                          * Executor doesn't support hashed aggregation with DISTINCT
 939                          * aggregates.  (Doing so would imply storing *all* the input
 940                          * values in the hash table, which seems like a certain loser.)
 941                          */
 942                         if (!enable_hashagg || !hash_safe_grouping(parse))
 943                                 use_hashed_grouping = false;
 944                         else if (parse->hasAggs &&
 945                                          (contain_distinct_agg_clause((Node *) tlist) ||
 946                                           contain_distinct_agg_clause(parse->havingQual)))
 947                                 use_hashed_grouping = false;
 948                         else
 949                         {
 950                                 /*
 951                                  * Use hashed grouping if (a) we think we can fit the
 952                                  * hashtable into SortMem, *and* (b) the estimated cost
 953                                  * is no more than doing it the other way.  While avoiding
 954                                  * the need for sorted input is usually a win, the fact
 955                                  * that the output won't be sorted may be a loss; so we
 956                                  * need to do an actual cost comparison.
 957                                  *
 958                                  * In most cases we have no good way to estimate the size of
 959                                  * the transition value needed by an aggregate; arbitrarily
 960                                  * assume it is 100 bytes.  Also set the overhead per hashtable
 961                                  * entry at 64 bytes.
 962                                  */
 963                                 int             hashentrysize = cheapest_path_width + 64 + numAggs * 100;
 964
 965                                 if (hashentrysize * dNumGroups <= SortMem * 1024L)
 966                                 {
 967                                         /*
 968                                          * Okay, do the cost comparison.  We need to consider
 969                                          *      cheapest_path + hashagg [+ final sort]
 970                                          * versus either
 971                                          *      cheapest_path [+ sort] + group or agg [+ final sort]
 972                                          * or
 973                                          *      presorted_path + group or agg [+ final sort]
 974                                          * where brackets indicate a step that may not be needed.
 975                                          * We assume query_planner() will have returned a
 976                                          * presorted path only if it's a winner compared to
 977                                          * cheapest_path for this purpose.
 978                                          *
 979                                          * These path variables are dummies that just hold cost
 980                                          * fields; we don't make actual Paths for these steps.
 981                                          */
 982                                         Path            hashed_p;
 983                                         Path            sorted_p;
 984
 985                                         cost_agg(&hashed_p, parse,
 986                                                          AGG_HASHED, numAggs,
 987                                                          numGroupCols, dNumGroups,
 988                                                          cheapest_path->startup_cost,
 989                                                          cheapest_path->total_cost,
 990                                                          cheapest_path_rows);
 991                                         /* Result of hashed agg is always unsorted */
 992                                         if (sort_pathkeys)
 993                                                 cost_sort(&hashed_p, parse, sort_pathkeys,
 994                                                                   hashed_p.total_cost,
 995                                                                   dNumGroups,
 996                                                                   cheapest_path_width);
 997
 998                                         if (sorted_path)
 999                                         {
1000                                                 sorted_p.startup_cost = sorted_path->startup_cost;
1001                                                 sorted_p.total_cost = sorted_path->total_cost;
1002                                                 current_pathkeys = sorted_path->pathkeys;
1003                                         }
1004                                         else
1005                                         {
1006                                                 sorted_p.startup_cost = cheapest_path->startup_cost;
1007                                                 sorted_p.total_cost = cheapest_path->total_cost;
1008                                                 current_pathkeys = cheapest_path->pathkeys;
1009                                         }
1010                                         if (!pathkeys_contained_in(group_pathkeys,
1011                                                                                            current_pathkeys))
1012                                         {
1013                                                 cost_sort(&sorted_p, parse, group_pathkeys,
1014                                                                   sorted_p.total_cost,
1015                                                                   cheapest_path_rows,
1016                                                                   cheapest_path_width);
1017                                                 current_pathkeys = group_pathkeys;
1018                                         }
1019                                         if (parse->hasAggs)
1020                                                 cost_agg(&sorted_p, parse,
1021                                                                  AGG_SORTED, numAggs,
1022                                                                  numGroupCols, dNumGroups,
1023                                                                  sorted_p.startup_cost,
1024                                                                  sorted_p.total_cost,
1025                                                                  cheapest_path_rows);
1026                                         else
1027                                                 cost_group(&sorted_p, parse,
1028                                                                    numGroupCols, dNumGroups,
1029                                                                    sorted_p.startup_cost,
1030                                                                    sorted_p.total_cost,
1031                                                                    cheapest_path_rows);
1032                                         /* The Agg or Group node will preserve ordering */
1033                                         if (sort_pathkeys &&
1034                                                 !pathkeys_contained_in(sort_pathkeys,
1035                                                                                            current_pathkeys))
1036                                         {
1037                                                 cost_sort(&sorted_p, parse, sort_pathkeys,
1038                                                                   sorted_p.total_cost,
1039                                                                   dNumGroups,
1040                                                                   cheapest_path_width);
1041                                         }
1042
1043                                         /*
1044                                          * Now make the decision using the top-level tuple
1045                                          * fraction.  First we have to convert an absolute
1046                                          * count (LIMIT) into fractional form.
1047                                          */
1048                                         if (tuple_fraction >= 1.0)
1049                                                 tuple_fraction /= dNumGroups;
1050
1051                                         if (compare_fractional_path_costs(&hashed_p, &sorted_p,
1052                                                                                                           tuple_fraction) < 0)
1053                                         {
1054                                                 /* Hashed is cheaper, so use it */
1055                                                 use_hashed_grouping = true;
1056                                         }
1057                                 }
1058                         }
1059                 }
1060
1061                 /*
1062                  * Select the best path and create a plan to execute it.
1063                  *
1064                  * If we are doing hashed grouping, we will always read all the
1065                  * input tuples, so use the cheapest-total path.  Otherwise,
1066                  * trust query_planner's decision about which to use.
1067                  */
1068                 if (sorted_path && !use_hashed_grouping)
1069                 {
1070                         result_plan = create_plan(parse, sorted_path);
1071                         current_pathkeys = sorted_path->pathkeys;
1072                 }
1073                 else
1074                 {
1075                         result_plan = create_plan(parse, cheapest_path);
1076                         current_pathkeys = cheapest_path->pathkeys;
1077                 }
1078
1079                 /*
1080                  * create_plan() returns a plan with just a "flat" tlist of required
1081                  * Vars.  Usually we need to insert the sub_tlist as the tlist of the
1082                  * top plan node.  However, we can skip that if we determined that
1083                  * whatever query_planner chose to return will be good enough.
1084                  */
1085                 if (need_tlist_eval)
1086                 {
1087                         /*
1088                          * If the top-level plan node is one that cannot do expression
1089                          * evaluation, we must insert a Result node to project the desired
1090                          * tlist.
1091                          * Currently, the only plan node we might see here that falls into
1092                          * that category is Append.
1093                          */
1094                         if (IsA(result_plan, Append))
1095                         {
1096                                 result_plan = (Plan *) make_result(sub_tlist, NULL,
1097                                                                                                    result_plan);
1098                         }
1099                         else
1100                         {
1101                                 /*
1102                                  * Otherwise, just replace the subplan's flat tlist with
1103                                  * the desired tlist.
1104                                  */
1105                                 result_plan->targetlist = sub_tlist;
1106                         }
1107                         /*
1108                          * Also, account for the cost of evaluation of the sub_tlist.
1109                          *
1110                          * Up to now, we have only been dealing with "flat" tlists,
1111                          * containing just Vars.  So their evaluation cost is zero
1112                          * according to the model used by cost_qual_eval() (or if you
1113                          * prefer, the cost is factored into cpu_tuple_cost).  Thus we can
1114                          * avoid accounting for tlist cost throughout query_planner() and
1115                          * subroutines.  But now we've inserted a tlist that might contain
1116                          * actual operators, sub-selects, etc --- so we'd better account
1117                          * for its cost.
1118                          *
1119                          * Below this point, any tlist eval cost for added-on nodes should
1120                          * be accounted for as we create those nodes.  Presently, of the
1121                          * node types we can add on, only Agg and Group project new tlists
1122                          * (the rest just copy their input tuples) --- so make_agg() and
1123                          * make_group() are responsible for computing the added cost.
1124                          */
1125                         cost_qual_eval(&tlist_cost, sub_tlist);
1126                         result_plan->startup_cost += tlist_cost.startup;
1127                         result_plan->total_cost += tlist_cost.startup +
1128                                 tlist_cost.per_tuple * result_plan->plan_rows;
1129                 }
1130                 else
1131                 {
1132                         /*
1133                          * Since we're using query_planner's tlist and not the one
1134                          * make_subplanTargetList calculated, we have to refigure
1135                          * any grouping-column indexes make_subplanTargetList computed.
1136                          */
1137                         locate_grouping_columns(parse, tlist, result_plan->targetlist,
1138                                                                         groupColIdx);
1139                 }
1140
1141                 /*
1142                  * Insert AGG or GROUP node if needed, plus an explicit sort step
1143                  * if necessary.
1144                  *
1145                  * HAVING clause, if any, becomes qual of the Agg node
1146                  */
1147                 if (use_hashed_grouping)
1148                 {
1149                         /* Hashed aggregate plan --- no sort needed */
1150                         result_plan = (Plan *) make_agg(parse,
1151                                                                                         tlist,
1152                                                                                         (List *) parse->havingQual,
1153                                                                                         AGG_HASHED,
1154                                                                                         numGroupCols,
1155                                                                                         groupColIdx,
1156                                                                                         numGroups,
1157                                                                                         numAggs,
1158                                                                                         result_plan);
1159                         /* Hashed aggregation produces randomly-ordered results */
1160                         current_pathkeys = NIL;
1161                 }
1162                 else if (parse->hasAggs)
1163                 {
1164                         /* Plain aggregate plan --- sort if needed */
1165                         AggStrategy aggstrategy;
1166
1167                         if (parse->groupClause)
1168                         {
1169                                 if (!pathkeys_contained_in(group_pathkeys, current_pathkeys))
1170                                 {
1171                                         result_plan = (Plan *)
1172                                                 make_sort_from_groupcols(parse,
1173                                                                                                  parse->groupClause,
1174                                                                                                  groupColIdx,
1175                                                                                                  result_plan);
1176                                         current_pathkeys = group_pathkeys;
1177                                 }
1178                                 aggstrategy = AGG_SORTED;
1179                                 /*
1180                                  * The AGG node will not change the sort ordering of its
1181                                  * groups, so current_pathkeys describes the result too.
1182                                  */
1183                         }
1184                         else
1185                         {
1186                                 aggstrategy = AGG_PLAIN;
1187                                 /* Result will be only one row anyway; no sort order */
1188                                 current_pathkeys = NIL;
1189                         }
1190
1191                         result_plan = (Plan *) make_agg(parse,
1192                                                                                         tlist,
1193                                                                                         (List *) parse->havingQual,
1194                                                                                         aggstrategy,
1195                                                                                         numGroupCols,
1196                                                                                         groupColIdx,
1197                                                                                         numGroups,
1198                                                                                         numAggs,
1199                                                                                         result_plan);
1200                 }
1201                 else
1202                 {
1203                         /*
1204                          * If there are no Aggs, we shouldn't have any HAVING qual anymore
1205                          */
1206                         Assert(parse->havingQual == NULL);
1207
1208                         /*
1209                          * If we have a GROUP BY clause, insert a group node (plus the
1210                          * appropriate sort node, if necessary).
1211                          */
1212                         if (parse->groupClause)
1213                         {
1214                                 /*
1215                                  * Add an explicit sort if we couldn't make the path come out
1216                                  * the way the GROUP node needs it.
1217                                  */
1218                                 if (!pathkeys_contained_in(group_pathkeys, current_pathkeys))
1219                                 {
1220                                         result_plan = (Plan *)
1221                                                 make_sort_from_groupcols(parse,
1222                                                                                                  parse->groupClause,
1223                                                                                                  groupColIdx,
1224                                                                                                  result_plan);
1225                                         current_pathkeys = group_pathkeys;
1226                                 }
1227
1228                                 result_plan = (Plan *) make_group(parse,
1229                                                                                                   tlist,
1230                                                                                                   numGroupCols,
1231                                                                                                   groupColIdx,
1232                                                                                                   dNumGroups,
1233                                                                                                   result_plan);
1234                                 /* The Group node won't change sort ordering */
1235                         }
1236                 }
1237         } /* end of if (setOperations) */
1238
1239         /*
1240          * If we were not able to make the plan come out in the right order,
1241          * add an explicit sort step.
1242          */
1243         if (parse->sortClause)
1244         {
1245                 if (!pathkeys_contained_in(sort_pathkeys, current_pathkeys))
1246                 {
1247                         result_plan = (Plan *)
1248                                 make_sort_from_sortclauses(parse,
1249                                                                                    tlist,
1250                                                                                    result_plan,
1251                                                                                    parse->sortClause);
1252                         current_pathkeys = sort_pathkeys;
1253                 }
1254         }
1255
1256         /*
1257          * If there is a DISTINCT clause, add the UNIQUE node.
1258          */
1259         if (parse->distinctClause)
1260         {
1261                 result_plan = (Plan *) make_unique(tlist, result_plan,
1262                                                                                    parse->distinctClause);
1263                 /*
1264                  * If there was grouping or aggregation, leave plan_rows as-is
1265                  * (ie, assume the result was already mostly unique).  If not,
1266                  * it's reasonable to assume the UNIQUE filter has effects
1267                  * comparable to GROUP BY.
1268                  */
1269                 if (!parse->groupClause && !parse->hasAggs)
1270                 {
1271                         List   *distinctExprs;
1272
1273                         distinctExprs = get_sortgrouplist_exprs(parse->distinctClause,
1274                                                                                                         parse->targetList);
1275                         result_plan->plan_rows = estimate_num_groups(parse,
1276                                                                                                                  distinctExprs,
1277                                                                                                                  result_plan->plan_rows);
1278                 }
1279         }
1280
1281         /*
1282          * Finally, if there is a LIMIT/OFFSET clause, add the LIMIT node.
1283          */
1284         if (parse->limitOffset || parse->limitCount)
1285         {
1286                 result_plan = (Plan *) make_limit(tlist, result_plan,
1287                                                                                   parse->limitOffset,
1288                                                                                   parse->limitCount);
1289         }
1290
1291         /*
1292          * Return the actual output ordering in query_pathkeys for possible
1293          * use by an outer query level.
1294          */
1295         parse->query_pathkeys = current_pathkeys;
1296
1297         return result_plan;
1298 }
1299
1300 /*
1301  * hash_safe_grouping - are grouping operators hashable?
1302  *
1303  * We assume hashed aggregation will work if the datatype's equality operator
1304  * is marked hashjoinable.
1305  */
1306 static bool
1307 hash_safe_grouping(Query *parse)
1308 {
1309         List       *gl;
1310
1311         foreach(gl, parse->groupClause)
1312         {
1313                 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1314                 TargetEntry *tle = get_sortgroupclause_tle(grpcl, parse->targetList);
1315                 Operator        optup;
1316                 bool            oprcanhash;
1317
1318                 optup = equality_oper(tle->resdom->restype, false);
1319                 oprcanhash = ((Form_pg_operator) GETSTRUCT(optup))->oprcanhash;
1320                 ReleaseSysCache(optup);
1321                 if (!oprcanhash)
1322                         return false;
1323         }
1324         return true;
1325 }
1326
1327 /*---------------
1328  * make_subplanTargetList
1329  *        Generate appropriate target list when grouping is required.
1330  *
1331  * When grouping_planner inserts Aggregate or Group plan nodes above
1332  * the result of query_planner, we typically want to pass a different
1333  * target list to query_planner than the outer plan nodes should have.
1334  * This routine generates the correct target list for the subplan.
1335  *
1336  * The initial target list passed from the parser already contains entries
1337  * for all ORDER BY and GROUP BY expressions, but it will not have entries
1338  * for variables used only in HAVING clauses; so we need to add those
1339  * variables to the subplan target list.  Also, if we are doing either
1340  * grouping or aggregation, we flatten all expressions except GROUP BY items
1341  * into their component variables; the other expressions will be computed by
1342  * the inserted nodes rather than by the subplan.  For example,
1343  * given a query like
1344  *              SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
1345  * we want to pass this targetlist to the subplan:
1346  *              a,b,c,d,a+b
1347  * where the a+b target will be used by the Sort/Group steps, and the
1348  * other targets will be used for computing the final results.  (In the
1349  * above example we could theoretically suppress the a and b targets and
1350  * pass down only c,d,a+b, but it's not really worth the trouble to
1351  * eliminate simple var references from the subplan.  We will avoid doing
1352  * the extra computation to recompute a+b at the outer level; see
1353  * replace_vars_with_subplan_refs() in setrefs.c.)
1354  *
1355  * If we are grouping or aggregating, *and* there are no non-Var grouping
1356  * expressions, then the returned tlist is effectively dummy; we do not
1357  * need to force it to be evaluated, because all the Vars it contains
1358  * should be present in the output of query_planner anyway.
1359  *
1360  * 'parse' is the query being processed.
1361  * 'tlist' is the query's target list.
1362  * 'groupColIdx' receives an array of column numbers for the GROUP BY
1363  *                      expressions (if there are any) in the subplan's target list.
1364  * 'need_tlist_eval' is set true if we really need to evaluate the
1365  *                      result tlist.
1366  *
1367  * The result is the targetlist to be passed to the subplan.
1368  *---------------
1369  */
1370 static List *
1371 make_subplanTargetList(Query *parse,
1372                                            List *tlist,
1373                                            AttrNumber **groupColIdx,
1374                                            bool *need_tlist_eval)
1375 {
1376         List       *sub_tlist;
1377         List       *extravars;
1378         int                     numCols;
1379
1380         *groupColIdx = NULL;
1381
1382         /*
1383          * If we're not grouping or aggregating, nothing to do here;
1384          * query_planner should receive the unmodified target list.
1385          */
1386         if (!parse->hasAggs && !parse->groupClause)
1387         {
1388                 *need_tlist_eval = true;
1389                 return tlist;
1390         }
1391
1392         /*
1393          * Otherwise, start with a "flattened" tlist (having just the vars
1394          * mentioned in the targetlist and HAVING qual --- but not upper-
1395          * level Vars; they will be replaced by Params later on).
1396          */
1397         sub_tlist = flatten_tlist(tlist);
1398         extravars = pull_var_clause(parse->havingQual, false);
1399         sub_tlist = add_to_flat_tlist(sub_tlist, extravars);
1400         freeList(extravars);
1401         *need_tlist_eval = false;       /* only eval if not flat tlist */
1402
1403         /*
1404          * If grouping, create sub_tlist entries for all GROUP BY expressions
1405          * (GROUP BY items that are simple Vars should be in the list
1406          * already), and make an array showing where the group columns are in
1407          * the sub_tlist.
1408          */
1409         numCols = length(parse->groupClause);
1410         if (numCols > 0)
1411         {
1412                 int                     keyno = 0;
1413                 AttrNumber *grpColIdx;
1414                 List       *gl;
1415
1416                 grpColIdx = (AttrNumber *) palloc(sizeof(AttrNumber) * numCols);
1417                 *groupColIdx = grpColIdx;
1418
1419                 foreach(gl, parse->groupClause)
1420                 {
1421                         GroupClause *grpcl = (GroupClause *) lfirst(gl);
1422                         Node       *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1423                         TargetEntry *te = NULL;
1424                         List       *sl;
1425
1426                         /* Find or make a matching sub_tlist entry */
1427                         foreach(sl, sub_tlist)
1428                         {
1429                                 te = (TargetEntry *) lfirst(sl);
1430                                 if (equal(groupexpr, te->expr))
1431                                         break;
1432                         }
1433                         if (!sl)
1434                         {
1435                                 te = makeTargetEntry(makeResdom(length(sub_tlist) + 1,
1436                                                                                                 exprType(groupexpr),
1437                                                                                                 exprTypmod(groupexpr),
1438                                                                                                 NULL,
1439                                                                                                 false),
1440                                                                          (Expr *) groupexpr);
1441                                 sub_tlist = lappend(sub_tlist, te);
1442                                 *need_tlist_eval = true; /* it's not flat anymore */
1443                         }
1444
1445                         /* and save its resno */
1446                         grpColIdx[keyno++] = te->resdom->resno;
1447                 }
1448         }
1449
1450         return sub_tlist;
1451 }
1452
1453 /*
1454  * locate_grouping_columns
1455  *              Locate grouping columns in the tlist chosen by query_planner.
1456  *
1457  * This is only needed if we don't use the sub_tlist chosen by
1458  * make_subplanTargetList.  We have to forget the column indexes found
1459  * by that routine and re-locate the grouping vars in the real sub_tlist.
1460  */
1461 static void
1462 locate_grouping_columns(Query *parse,
1463                                                 List *tlist,
1464                                                 List *sub_tlist,
1465                                                 AttrNumber *groupColIdx)
1466 {
1467         int                     keyno = 0;
1468         List       *gl;
1469
1470         /*
1471          * No work unless grouping.
1472          */
1473         if (!parse->groupClause)
1474         {
1475                 Assert(groupColIdx == NULL);
1476                 return;
1477         }
1478         Assert(groupColIdx != NULL);
1479
1480         foreach(gl, parse->groupClause)
1481         {
1482                 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1483                 Node       *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1484                 TargetEntry *te = NULL;
1485                 List       *sl;
1486
1487                 foreach(sl, sub_tlist)
1488                 {
1489                         te = (TargetEntry *) lfirst(sl);
1490                         if (equal(groupexpr, te->expr))
1491                                 break;
1492                 }
1493                 if (!sl)
1494                         elog(ERROR, "locate_grouping_columns: failed");
1495
1496                 groupColIdx[keyno++] = te->resdom->resno;
1497         }
1498 }
1499
1500 /*
1501  * postprocess_setop_tlist
1502  *        Fix up targetlist returned by plan_set_operations().
1503  *
1504  * We need to transpose sort key info from the orig_tlist into new_tlist.
1505  * NOTE: this would not be good enough if we supported resjunk sort keys
1506  * for results of set operations --- then, we'd need to project a whole
1507  * new tlist to evaluate the resjunk columns.  For now, just elog if we
1508  * find any resjunk columns in orig_tlist.
1509  */
1510 static List *
1511 postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
1512 {
1513         List       *l;
1514
1515         foreach(l, new_tlist)
1516         {
1517                 TargetEntry *new_tle = (TargetEntry *) lfirst(l);
1518                 TargetEntry *orig_tle;
1519
1520                 /* ignore resjunk columns in setop result */
1521                 if (new_tle->resdom->resjunk)
1522                         continue;
1523
1524                 Assert(orig_tlist != NIL);
1525                 orig_tle = (TargetEntry *) lfirst(orig_tlist);
1526                 orig_tlist = lnext(orig_tlist);
1527                 if (orig_tle->resdom->resjunk)
1528                         elog(ERROR, "postprocess_setop_tlist: resjunk output columns not implemented");
1529                 Assert(new_tle->resdom->resno == orig_tle->resdom->resno);
1530                 Assert(new_tle->resdom->restype == orig_tle->resdom->restype);
1531                 new_tle->resdom->ressortgroupref = orig_tle->resdom->ressortgroupref;
1532         }
1533         if (orig_tlist != NIL)
1534                 elog(ERROR, "postprocess_setop_tlist: resjunk output columns not implemented");
1535         return new_tlist;
1536 }