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) 1996-2007, 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 1.214 2007/02/20 17:32:15 tgl Exp $
  12  *
  13  *-------------------------------------------------------------------------
  14  */
  15
  16 #include "postgres.h"
  17
  18 #include <limits.h>
  19
  20 #include "catalog/pg_operator.h"
  21 #include "executor/executor.h"
  22 #include "executor/nodeAgg.h"
  23 #include "miscadmin.h"
  24 #include "nodes/makefuncs.h"
  25 #include "optimizer/clauses.h"
  26 #include "optimizer/cost.h"
  27 #include "optimizer/pathnode.h"
  28 #include "optimizer/paths.h"
  29 #include "optimizer/planmain.h"
  30 #include "optimizer/planner.h"
  31 #include "optimizer/prep.h"
  32 #include "optimizer/subselect.h"
  33 #include "optimizer/tlist.h"
  34 #include "optimizer/var.h"
  35 #ifdef OPTIMIZER_DEBUG
  36 #include "nodes/print.h"
  37 #endif
  38 #include "parser/parse_expr.h"
  39 #include "parser/parse_oper.h"
  40 #include "parser/parsetree.h"
  41 #include "utils/lsyscache.h"
  42 #include "utils/syscache.h"
  43
  44
  45 /* Expression kind codes for preprocess_expression */
  46 #define EXPRKIND_QUAL           0
  47 #define EXPRKIND_TARGET         1
  48 #define EXPRKIND_RTFUNC         2
  49 #define EXPRKIND_VALUES         3
  50 #define EXPRKIND_LIMIT          4
  51 #define EXPRKIND_ININFO         5
  52 #define EXPRKIND_APPINFO        6
  53
  54
  55 static Node *preprocess_expression(PlannerInfo *root, Node *expr, int kind);
  56 static void preprocess_qual_conditions(PlannerInfo *root, Node *jtnode);
  57 static Plan *inheritance_planner(PlannerInfo *root);
  58 static Plan *grouping_planner(PlannerInfo *root, double tuple_fraction);
  59 static bool is_dummy_plan(Plan *plan);
  60 static double preprocess_limit(PlannerInfo *root,
  61                                  double tuple_fraction,
  62                                  int64 *offset_est, int64 *count_est);
  63 static Oid *extract_grouping_ops(List *groupClause);
  64 static bool choose_hashed_grouping(PlannerInfo *root, double tuple_fraction,
  65                                            Path *cheapest_path, Path *sorted_path,
  66                                            Oid *groupOperators, double dNumGroups,
  67                                            AggClauseCounts *agg_counts);
  68 static List *make_subplanTargetList(PlannerInfo *root, List *tlist,
  69                                            AttrNumber **groupColIdx, bool *need_tlist_eval);
  70 static void locate_grouping_columns(PlannerInfo *root,
  71                                                 List *tlist,
  72                                                 List *sub_tlist,
  73                                                 AttrNumber *groupColIdx);
  74 static List *postprocess_setop_tlist(List *new_tlist, List *orig_tlist);
  75
  76
  77 /*****************************************************************************
  78  *
  79  *         Query optimizer entry point
  80  *
  81  *****************************************************************************/
  82 PlannedStmt *
  83 planner(Query *parse, bool isCursor, int cursorOptions,
  84                 ParamListInfo boundParams)
  85 {
  86         PlannedStmt *result;
  87         PlannerGlobal *glob;
  88         double          tuple_fraction;
  89         PlannerInfo *root;
  90         Plan       *top_plan;
  91
  92         /*
  93          * Set up global state for this planner invocation.  This data is needed
  94          * across all levels of sub-Query that might exist in the given command,
  95          * so we keep it in a separate struct that's linked to by each per-Query
  96          * PlannerInfo.
  97          */
  98         glob = makeNode(PlannerGlobal);
  99
 100         glob->boundParams = boundParams;
 101         glob->paramlist = NIL;
 102         glob->next_plan_id = 0;
 103
 104         /* Determine what fraction of the plan is likely to be scanned */
 105         if (isCursor)
 106         {
 107                 /*
 108                  * We have no real idea how many tuples the user will ultimately FETCH
 109                  * from a cursor, but it seems a good bet that he doesn't want 'em
 110                  * all.  Optimize for 10% retrieval (you gotta better number?  Should
 111                  * this be a SETtable parameter?)
 112                  */
 113                 tuple_fraction = 0.10;
 114         }
 115         else
 116         {
 117                 /* Default assumption is we need all the tuples */
 118                 tuple_fraction = 0.0;
 119         }
 120
 121         /* primary planning entry point (may recurse for subqueries) */
 122         top_plan = subquery_planner(glob, parse, 1, tuple_fraction, &root);
 123
 124         /*
 125          * If creating a plan for a scrollable cursor, make sure it can run
 126          * backwards on demand.  Add a Material node at the top at need.
 127          */
 128         if (isCursor && (cursorOptions & CURSOR_OPT_SCROLL))
 129         {
 130                 if (!ExecSupportsBackwardScan(top_plan))
 131                         top_plan = materialize_finished_plan(top_plan);
 132         }
 133
 134         /* final cleanup of the plan */
 135         top_plan = set_plan_references(top_plan, parse->rtable);
 136
 137         /* build the PlannedStmt result */
 138         result = makeNode(PlannedStmt);
 139
 140         result->commandType = parse->commandType;
 141         result->canSetTag = parse->canSetTag;
 142         result->planTree = top_plan;
 143         result->rtable = parse->rtable;
 144         result->resultRelations = root->resultRelations;
 145         result->into = parse->into;
 146         result->returningLists = root->returningLists;
 147         result->rowMarks = parse->rowMarks;
 148         result->nParamExec = list_length(glob->paramlist);
 149
 150         return result;
 151 }
 152
 153
 154 /*--------------------
 155  * subquery_planner
 156  *        Invokes the planner on a subquery.  We recurse to here for each
 157  *        sub-SELECT found in the query tree.
 158  *
 159  * glob is the global state for the current planner run.
 160  * parse is the querytree produced by the parser & rewriter.
 161  * level is the current recursion depth (1 at the top-level Query).
 162  * tuple_fraction is the fraction of tuples we expect will be retrieved.
 163  * tuple_fraction is interpreted as explained for grouping_planner, below.
 164  *
 165  * If subroot isn't NULL, we pass back the query's final PlannerInfo struct;
 166  * among other things this tells the output sort ordering of the plan.
 167  *
 168  * Basically, this routine does the stuff that should only be done once
 169  * per Query object.  It then calls grouping_planner.  At one time,
 170  * grouping_planner could be invoked recursively on the same Query object;
 171  * that's not currently true, but we keep the separation between the two
 172  * routines anyway, in case we need it again someday.
 173  *
 174  * subquery_planner will be called recursively to handle sub-Query nodes
 175  * found within the query's expressions and rangetable.
 176  *
 177  * Returns a query plan.
 178  *--------------------
 179  */
 180 Plan *
 181 subquery_planner(PlannerGlobal *glob, Query *parse,
 182                                  Index level, double tuple_fraction,
 183                                  PlannerInfo **subroot)
 184 {
 185         int                     saved_plan_id = glob->next_plan_id;
 186         PlannerInfo *root;
 187         Plan       *plan;
 188         List       *newHaving;
 189         ListCell   *l;
 190
 191         /* Create a PlannerInfo data structure for this subquery */
 192         root = makeNode(PlannerInfo);
 193         root->parse = parse;
 194         root->glob = glob;
 195         root->query_level = level;
 196         root->planner_cxt = CurrentMemoryContext;
 197         root->init_plans = NIL;
 198         root->eq_classes = NIL;
 199         root->in_info_list = NIL;
 200         root->append_rel_list = NIL;
 201
 202         /*
 203          * Look for IN clauses at the top level of WHERE, and transform them into
 204          * joins.  Note that this step only handles IN clauses originally at top
 205          * level of WHERE; if we pull up any subqueries in the next step, their
 206          * INs are processed just before pulling them up.
 207          */
 208         if (parse->hasSubLinks)
 209                 parse->jointree->quals = pull_up_IN_clauses(root,
 210                                                                                                         parse->jointree->quals);
 211
 212         /*
 213          * Check to see if any subqueries in the rangetable can be merged into
 214          * this query.
 215          */
 216         parse->jointree = (FromExpr *)
 217                 pull_up_subqueries(root, (Node *) parse->jointree, false, false);
 218
 219         /*
 220          * Detect whether any rangetable entries are RTE_JOIN kind; if not, we can
 221          * avoid the expense of doing flatten_join_alias_vars().  Also check for
 222          * outer joins --- if none, we can skip reduce_outer_joins() and some
 223          * other processing.  This must be done after we have done
 224          * pull_up_subqueries, of course.
 225          *
 226          * Note: if reduce_outer_joins manages to eliminate all outer joins,
 227          * root->hasOuterJoins is not reset currently.  This is OK since its
 228          * purpose is merely to suppress unnecessary processing in simple cases.
 229          */
 230         root->hasJoinRTEs = false;
 231         root->hasOuterJoins = false;
 232         foreach(l, parse->rtable)
 233         {
 234                 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
 235
 236                 if (rte->rtekind == RTE_JOIN)
 237                 {
 238                         root->hasJoinRTEs = true;
 239                         if (IS_OUTER_JOIN(rte->jointype))
 240                         {
 241                                 root->hasOuterJoins = true;
 242                                 /* Can quit scanning once we find an outer join */
 243                                 break;
 244                         }
 245                 }
 246         }
 247
 248         /*
 249          * Expand any rangetable entries that are inheritance sets into "append
 250          * relations".  This can add entries to the rangetable, but they must be
 251          * plain base relations not joins, so it's OK (and marginally more
 252          * efficient) to do it after checking for join RTEs.  We must do it after
 253          * pulling up subqueries, else we'd fail to handle inherited tables in
 254          * subqueries.
 255          */
 256         expand_inherited_tables(root);
 257
 258         /*
 259          * Set hasHavingQual to remember if HAVING clause is present.  Needed
 260          * because preprocess_expression will reduce a constant-true condition to
 261          * an empty qual list ... but "HAVING TRUE" is not a semantic no-op.
 262          */
 263         root->hasHavingQual = (parse->havingQual != NULL);
 264
 265         /* Clear this flag; might get set in distribute_qual_to_rels */
 266         root->hasPseudoConstantQuals = false;
 267
 268         /*
 269          * Do expression preprocessing on targetlist and quals.
 270          */
 271         parse->targetList = (List *)
 272                 preprocess_expression(root, (Node *) parse->targetList,
 273                                                           EXPRKIND_TARGET);
 274
 275         parse->returningList = (List *)
 276                 preprocess_expression(root, (Node *) parse->returningList,
 277                                                           EXPRKIND_TARGET);
 278
 279         preprocess_qual_conditions(root, (Node *) parse->jointree);
 280
 281         parse->havingQual = preprocess_expression(root, parse->havingQual,
 282                                                                                           EXPRKIND_QUAL);
 283
 284         parse->limitOffset = preprocess_expression(root, parse->limitOffset,
 285                                                                                            EXPRKIND_LIMIT);
 286         parse->limitCount = preprocess_expression(root, parse->limitCount,
 287                                                                                           EXPRKIND_LIMIT);
 288
 289         root->in_info_list = (List *)
 290                 preprocess_expression(root, (Node *) root->in_info_list,
 291                                                           EXPRKIND_ININFO);
 292         root->append_rel_list = (List *)
 293                 preprocess_expression(root, (Node *) root->append_rel_list,
 294                                                           EXPRKIND_APPINFO);
 295
 296         /* Also need to preprocess expressions for function and values RTEs */
 297         foreach(l, parse->rtable)
 298         {
 299                 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
 300
 301                 if (rte->rtekind == RTE_FUNCTION)
 302                         rte->funcexpr = preprocess_expression(root, rte->funcexpr,
 303                                                                                                   EXPRKIND_RTFUNC);
 304                 else if (rte->rtekind == RTE_VALUES)
 305                         rte->values_lists = (List *)
 306                                 preprocess_expression(root, (Node *) rte->values_lists,
 307                                                                           EXPRKIND_VALUES);
 308         }
 309
 310         /*
 311          * In some cases we may want to transfer a HAVING clause into WHERE. We
 312          * cannot do so if the HAVING clause contains aggregates (obviously) or
 313          * volatile functions (since a HAVING clause is supposed to be executed
 314          * only once per group).  Also, it may be that the clause is so expensive
 315          * to execute that we're better off doing it only once per group, despite
 316          * the loss of selectivity.  This is hard to estimate short of doing the
 317          * entire planning process twice, so we use a heuristic: clauses
 318          * containing subplans are left in HAVING.      Otherwise, we move or copy the
 319          * HAVING clause into WHERE, in hopes of eliminating tuples before
 320          * aggregation instead of after.
 321          *
 322          * If the query has explicit grouping then we can simply move such a
 323          * clause into WHERE; any group that fails the clause will not be in the
 324          * output because none of its tuples will reach the grouping or
 325          * aggregation stage.  Otherwise we must have a degenerate (variable-free)
 326          * HAVING clause, which we put in WHERE so that query_planner() can use it
 327          * in a gating Result node, but also keep in HAVING to ensure that we
 328          * don't emit a bogus aggregated row. (This could be done better, but it
 329          * seems not worth optimizing.)
 330          *
 331          * Note that both havingQual and parse->jointree->quals are in
 332          * implicitly-ANDed-list form at this point, even though they are declared
 333          * as Node *.
 334          */
 335         newHaving = NIL;
 336         foreach(l, (List *) parse->havingQual)
 337         {
 338                 Node       *havingclause = (Node *) lfirst(l);
 339
 340                 if (contain_agg_clause(havingclause) ||
 341                         contain_volatile_functions(havingclause) ||
 342                         contain_subplans(havingclause))
 343                 {
 344                         /* keep it in HAVING */
 345                         newHaving = lappend(newHaving, havingclause);
 346                 }
 347                 else if (parse->groupClause)
 348                 {
 349                         /* move it to WHERE */
 350                         parse->jointree->quals = (Node *)
 351                                 lappend((List *) parse->jointree->quals, havingclause);
 352                 }
 353                 else
 354                 {
 355                         /* put a copy in WHERE, keep it in HAVING */
 356                         parse->jointree->quals = (Node *)
 357                                 lappend((List *) parse->jointree->quals,
 358                                                 copyObject(havingclause));
 359                         newHaving = lappend(newHaving, havingclause);
 360                 }
 361         }
 362         parse->havingQual = (Node *) newHaving;
 363
 364         /*
 365          * If we have any outer joins, try to reduce them to plain inner joins.
 366          * This step is most easily done after we've done expression
 367          * preprocessing.
 368          */
 369         if (root->hasOuterJoins)
 370                 reduce_outer_joins(root);
 371
 372         /*
 373          * Do the main planning.  If we have an inherited target relation, that
 374          * needs special processing, else go straight to grouping_planner.
 375          */
 376         if (parse->resultRelation &&
 377                 rt_fetch(parse->resultRelation, parse->rtable)->inh)
 378                 plan = inheritance_planner(root);
 379         else
 380                 plan = grouping_planner(root, tuple_fraction);
 381
 382         /*
 383          * If any subplans were generated, or if we're inside a subplan, build
 384          * initPlan list and extParam/allParam sets for plan nodes, and attach the
 385          * initPlans to the top plan node.
 386          */
 387         if (root->glob->next_plan_id != saved_plan_id || root->query_level > 1)
 388                 SS_finalize_plan(root, plan);
 389
 390         /* Return internal info if caller wants it */
 391         if (subroot)
 392                 *subroot = root;
 393
 394         return plan;
 395 }
 396
 397 /*
 398  * preprocess_expression
 399  *              Do subquery_planner's preprocessing work for an expression,
 400  *              which can be a targetlist, a WHERE clause (including JOIN/ON
 401  *              conditions), or a HAVING clause.
 402  */
 403 static Node *
 404 preprocess_expression(PlannerInfo *root, Node *expr, int kind)
 405 {
 406         /*
 407          * Fall out quickly if expression is empty.  This occurs often enough to
 408          * be worth checking.  Note that null->null is the correct conversion for
 409          * implicit-AND result format, too.
 410          */
 411         if (expr == NULL)
 412                 return NULL;
 413
 414         /*
 415          * If the query has any join RTEs, replace join alias variables with
 416          * base-relation variables. We must do this before sublink processing,
 417          * else sublinks expanded out from join aliases wouldn't get processed. We
 418          * can skip it in VALUES lists, however, since they can't contain any Vars
 419          * at all.
 420          */
 421         if (root->hasJoinRTEs && kind != EXPRKIND_VALUES)
 422                 expr = flatten_join_alias_vars(root, expr);
 423
 424         /*
 425          * Simplify constant expressions.
 426          *
 427          * Note: this also flattens nested AND and OR expressions into N-argument
 428          * form.  All processing of a qual expression after this point must be
 429          * careful to maintain AND/OR flatness --- that is, do not generate a tree
 430          * with AND directly under AND, nor OR directly under OR.
 431          *
 432          * Because this is a relatively expensive process, we skip it when the
 433          * query is trivial, such as "SELECT 2+2;" or "INSERT ... VALUES()". The
 434          * expression will only be evaluated once anyway, so no point in
 435          * pre-simplifying; we can't execute it any faster than the executor can,
 436          * and we will waste cycles copying the tree.  Notice however that we
 437          * still must do it for quals (to get AND/OR flatness); and if we are in a
 438          * subquery we should not assume it will be done only once.
 439          *
 440          * For VALUES lists we never do this at all, again on the grounds that we
 441          * should optimize for one-time evaluation.
 442          */
 443         if (kind != EXPRKIND_VALUES &&
 444                 (root->parse->jointree->fromlist != NIL ||
 445                  kind == EXPRKIND_QUAL ||
 446                  root->query_level > 1))
 447                 expr = eval_const_expressions(expr);
 448
 449         /*
 450          * If it's a qual or havingQual, canonicalize it.
 451          */
 452         if (kind == EXPRKIND_QUAL)
 453         {
 454                 expr = (Node *) canonicalize_qual((Expr *) expr);
 455
 456 #ifdef OPTIMIZER_DEBUG
 457                 printf("After canonicalize_qual()\n");
 458                 pprint(expr);
 459 #endif
 460         }
 461
 462         /* Expand SubLinks to SubPlans */
 463         if (root->parse->hasSubLinks)
 464                 expr = SS_process_sublinks(root, expr, (kind == EXPRKIND_QUAL));
 465
 466         /*
 467          * XXX do not insert anything here unless you have grokked the comments in
 468          * SS_replace_correlation_vars ...
 469          */
 470
 471         /* Replace uplevel vars with Param nodes (this IS possible in VALUES) */
 472         if (root->query_level > 1)
 473                 expr = SS_replace_correlation_vars(root, expr);
 474
 475         /*
 476          * If it's a qual or havingQual, convert it to implicit-AND format. (We
 477          * don't want to do this before eval_const_expressions, since the latter
 478          * would be unable to simplify a top-level AND correctly. Also,
 479          * SS_process_sublinks expects explicit-AND format.)
 480          */
 481         if (kind == EXPRKIND_QUAL)
 482                 expr = (Node *) make_ands_implicit((Expr *) expr);
 483
 484         return expr;
 485 }
 486
 487 /*
 488  * preprocess_qual_conditions
 489  *              Recursively scan the query's jointree and do subquery_planner's
 490  *              preprocessing work on each qual condition found therein.
 491  */
 492 static void
 493 preprocess_qual_conditions(PlannerInfo *root, Node *jtnode)
 494 {
 495         if (jtnode == NULL)
 496                 return;
 497         if (IsA(jtnode, RangeTblRef))
 498         {
 499                 /* nothing to do here */
 500         }
 501         else if (IsA(jtnode, FromExpr))
 502         {
 503                 FromExpr   *f = (FromExpr *) jtnode;
 504                 ListCell   *l;
 505
 506                 foreach(l, f->fromlist)
 507                         preprocess_qual_conditions(root, lfirst(l));
 508
 509                 f->quals = preprocess_expression(root, f->quals, EXPRKIND_QUAL);
 510         }
 511         else if (IsA(jtnode, JoinExpr))
 512         {
 513                 JoinExpr   *j = (JoinExpr *) jtnode;
 514
 515                 preprocess_qual_conditions(root, j->larg);
 516                 preprocess_qual_conditions(root, j->rarg);
 517
 518                 j->quals = preprocess_expression(root, j->quals, EXPRKIND_QUAL);
 519         }
 520         else
 521                 elog(ERROR, "unrecognized node type: %d",
 522                          (int) nodeTag(jtnode));
 523 }
 524
 525 /*
 526  * inheritance_planner
 527  *        Generate a plan in the case where the result relation is an
 528  *        inheritance set.
 529  *
 530  * We have to handle this case differently from cases where a source relation
 531  * is an inheritance set. Source inheritance is expanded at the bottom of the
 532  * plan tree (see allpaths.c), but target inheritance has to be expanded at
 533  * the top.  The reason is that for UPDATE, each target relation needs a
 534  * different targetlist matching its own column set.  Also, for both UPDATE
 535  * and DELETE, the executor needs the Append plan node at the top, else it
 536  * can't keep track of which table is the current target table.  Fortunately,
 537  * the UPDATE/DELETE target can never be the nullable side of an outer join,
 538  * so it's OK to generate the plan this way.
 539  *
 540  * Returns a query plan.
 541  */
 542 static Plan *
 543 inheritance_planner(PlannerInfo *root)
 544 {
 545         Query      *parse = root->parse;
 546         int                     parentRTindex = parse->resultRelation;
 547         List       *subplans = NIL;
 548         List       *resultRelations = NIL;
 549         List       *returningLists = NIL;
 550         List       *rtable = NIL;
 551         List       *tlist = NIL;
 552         PlannerInfo subroot;
 553         ListCell   *l;
 554
 555         foreach(l, root->append_rel_list)
 556         {
 557                 AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
 558                 Plan       *subplan;
 559
 560                 /* append_rel_list contains all append rels; ignore others */
 561                 if (appinfo->parent_relid != parentRTindex)
 562                         continue;
 563
 564                 /*
 565                  * Generate modified query with this rel as target.  We have to be
 566                  * prepared to translate varnos in in_info_list as well as in the
 567                  * Query proper.
 568                  */
 569                 memcpy(&subroot, root, sizeof(PlannerInfo));
 570                 subroot.parse = (Query *)
 571                         adjust_appendrel_attrs((Node *) parse,
 572                                                                    appinfo);
 573                 subroot.in_info_list = (List *)
 574                         adjust_appendrel_attrs((Node *) root->in_info_list,
 575                                                                    appinfo);
 576                 subroot.init_plans = NIL;
 577                 /* There shouldn't be any OJ info to translate, as yet */
 578                 Assert(subroot.oj_info_list == NIL);
 579
 580                 /* Generate plan */
 581                 subplan = grouping_planner(&subroot, 0.0 /* retrieve all tuples */ );
 582
 583                 /*
 584                  * If this child rel was excluded by constraint exclusion, exclude it
 585                  * from the plan.
 586                  */
 587                 if (is_dummy_plan(subplan))
 588                         continue;
 589
 590                 /* Save rtable and tlist from first rel for use below */
 591                 if (subplans == NIL)
 592                 {
 593                         rtable = subroot.parse->rtable;
 594                         tlist = subplan->targetlist;
 595                 }
 596
 597                 subplans = lappend(subplans, subplan);
 598
 599                 /* Make sure any initplans from this rel get into the outer list */
 600                 root->init_plans = list_concat(root->init_plans, subroot.init_plans);
 601
 602                 /* Build target-relations list for the executor */
 603                 resultRelations = lappend_int(resultRelations, appinfo->child_relid);
 604
 605                 /* Build list of per-relation RETURNING targetlists */
 606                 if (parse->returningList)
 607                 {
 608                         Assert(list_length(subroot.returningLists) == 1);
 609                         returningLists = list_concat(returningLists,
 610                                                                                  subroot.returningLists);
 611                 }
 612         }
 613
 614         root->resultRelations = resultRelations;
 615         root->returningLists = returningLists;
 616
 617         /* Mark result as unordered (probably unnecessary) */
 618         root->query_pathkeys = NIL;
 619
 620         /*
 621          * If we managed to exclude every child rel, return a dummy plan
 622          */
 623         if (subplans == NIL)
 624                 return (Plan *) make_result(tlist,
 625                                                                         (Node *) list_make1(makeBoolConst(false,
 626                                                                                                                                           false)),
 627                                                                         NULL);
 628
 629         /*
 630          * Planning might have modified the rangetable, due to changes of the
 631          * Query structures inside subquery RTEs.  We have to ensure that this
 632          * gets propagated back to the master copy.  But can't do this until we
 633          * are done planning, because all the calls to grouping_planner need
 634          * virgin sub-Queries to work from.  (We are effectively assuming that
 635          * sub-Queries will get planned identically each time, or at least that
 636          * the impacts on their rangetables will be the same each time.)
 637          *
 638          * XXX should clean this up someday
 639          */
 640         parse->rtable = rtable;
 641
 642         return (Plan *) make_append(subplans, true, tlist);
 643 }
 644
 645 /*--------------------
 646  * grouping_planner
 647  *        Perform planning steps related to grouping, aggregation, etc.
 648  *        This primarily means adding top-level processing to the basic
 649  *        query plan produced by query_planner.
 650  *
 651  * tuple_fraction is the fraction of tuples we expect will be retrieved
 652  *
 653  * tuple_fraction is interpreted as follows:
 654  *        0: expect all tuples to be retrieved (normal case)
 655  *        0 < tuple_fraction < 1: expect the given fraction of tuples available
 656  *              from the plan to be retrieved
 657  *        tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
 658  *              expected to be retrieved (ie, a LIMIT specification)
 659  *
 660  * Returns a query plan.  Also, root->query_pathkeys is returned as the
 661  * actual output ordering of the plan (in pathkey format).
 662  *--------------------
 663  */
 664 static Plan *
 665 grouping_planner(PlannerInfo *root, double tuple_fraction)
 666 {
 667         Query      *parse = root->parse;
 668         List       *tlist = parse->targetList;
 669         int64           offset_est = 0;
 670         int64           count_est = 0;
 671         Plan       *result_plan;
 672         List       *current_pathkeys;
 673         List       *sort_pathkeys;
 674         double          dNumGroups = 0;
 675
 676         /* Tweak caller-supplied tuple_fraction if have LIMIT/OFFSET */
 677         if (parse->limitCount || parse->limitOffset)
 678                 tuple_fraction = preprocess_limit(root, tuple_fraction,
 679                                                                                   &offset_est, &count_est);
 680
 681         if (parse->setOperations)
 682         {
 683                 List       *set_sortclauses;
 684
 685                 /*
 686                  * If there's a top-level ORDER BY, assume we have to fetch all the
 687                  * tuples.      This might seem too simplistic given all the hackery below
 688                  * to possibly avoid the sort ... but a nonzero tuple_fraction is only
 689                  * of use to plan_set_operations() when the setop is UNION ALL, and
 690                  * the result of UNION ALL is always unsorted.
 691                  */
 692                 if (parse->sortClause)
 693                         tuple_fraction = 0.0;
 694
 695                 /*
 696                  * Construct the plan for set operations.  The result will not need
 697                  * any work except perhaps a top-level sort and/or LIMIT.
 698                  */
 699                 result_plan = plan_set_operations(root, tuple_fraction,
 700                                                                                   &set_sortclauses);
 701
 702                 /*
 703                  * Calculate pathkeys representing the sort order (if any) of the set
 704                  * operation's result.  We have to do this before overwriting the sort
 705                  * key information...
 706                  */
 707                 current_pathkeys = make_pathkeys_for_sortclauses(root,
 708                                                                                                                  set_sortclauses,
 709                                                                                                         result_plan->targetlist,
 710                                                                                                                  true);
 711
 712                 /*
 713                  * We should not need to call preprocess_targetlist, since we must be
 714                  * in a SELECT query node.      Instead, use the targetlist returned by
 715                  * plan_set_operations (since this tells whether it returned any
 716                  * resjunk columns!), and transfer any sort key information from the
 717                  * original tlist.
 718                  */
 719                 Assert(parse->commandType == CMD_SELECT);
 720
 721                 tlist = postprocess_setop_tlist(result_plan->targetlist, tlist);
 722
 723                 /*
 724                  * Can't handle FOR UPDATE/SHARE here (parser should have checked
 725                  * already, but let's make sure).
 726                  */
 727                 if (parse->rowMarks)
 728                         ereport(ERROR,
 729                                         (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
 730                                          errmsg("SELECT FOR UPDATE/SHARE is not allowed with UNION/INTERSECT/EXCEPT")));
 731
 732                 /*
 733                  * Calculate pathkeys that represent result ordering requirements
 734                  */
 735                 sort_pathkeys = make_pathkeys_for_sortclauses(root,
 736                                                                                                           parse->sortClause,
 737                                                                                                           tlist,
 738                                                                                                           true);
 739         }
 740         else
 741         {
 742                 /* No set operations, do regular planning */
 743                 List       *sub_tlist;
 744                 List       *group_pathkeys;
 745                 AttrNumber *groupColIdx = NULL;
 746                 Oid                *groupOperators = NULL;
 747                 bool            need_tlist_eval = true;
 748                 QualCost        tlist_cost;
 749                 Path       *cheapest_path;
 750                 Path       *sorted_path;
 751                 Path       *best_path;
 752                 long            numGroups = 0;
 753                 AggClauseCounts agg_counts;
 754                 int                     numGroupCols = list_length(parse->groupClause);
 755                 bool            use_hashed_grouping = false;
 756
 757                 MemSet(&agg_counts, 0, sizeof(AggClauseCounts));
 758
 759                 /* Preprocess targetlist */
 760                 tlist = preprocess_targetlist(root, tlist);
 761
 762                 /*
 763                  * Generate appropriate target list for subplan; may be different from
 764                  * tlist if grouping or aggregation is needed.
 765                  */
 766                 sub_tlist = make_subplanTargetList(root, tlist,
 767                                                                                    &groupColIdx, &need_tlist_eval);
 768
 769                 /*
 770                  * Calculate pathkeys that represent grouping/ordering requirements.
 771                  * Stash them in PlannerInfo so that query_planner can canonicalize
 772                  * them after EquivalenceClasses have been formed.
 773                  */
 774                 root->group_pathkeys =
 775                         make_pathkeys_for_sortclauses(root,
 776                                                                                   parse->groupClause,
 777                                                                                   tlist,
 778                                                                                   false);
 779                 root->sort_pathkeys =
 780                         make_pathkeys_for_sortclauses(root,
 781                                                                                   parse->sortClause,
 782                                                                                   tlist,
 783                                                                                   false);
 784
 785                 /*
 786                  * Will need actual number of aggregates for estimating costs.
 787                  *
 788                  * Note: we do not attempt to detect duplicate aggregates here; a
 789                  * somewhat-overestimated count is okay for our present purposes.
 790                  *
 791                  * Note: think not that we can turn off hasAggs if we find no aggs. It
 792                  * is possible for constant-expression simplification to remove all
 793                  * explicit references to aggs, but we still have to follow the
 794                  * aggregate semantics (eg, producing only one output row).
 795                  */
 796                 if (parse->hasAggs)
 797                 {
 798                         count_agg_clauses((Node *) tlist, &agg_counts);
 799                         count_agg_clauses(parse->havingQual, &agg_counts);
 800                 }
 801
 802                 /*
 803                  * Figure out whether we need a sorted result from query_planner.
 804                  *
 805                  * If we have a GROUP BY clause, then we want a result sorted properly
 806                  * for grouping.  Otherwise, if there is an ORDER BY clause, we want
 807                  * to sort by the ORDER BY clause.      (Note: if we have both, and ORDER
 808                  * BY is a superset of GROUP BY, it would be tempting to request sort
 809                  * by ORDER BY --- but that might just leave us failing to exploit an
 810                  * available sort order at all. Needs more thought...)
 811                  */
 812                 if (parse->groupClause)
 813                         root->query_pathkeys = root->group_pathkeys;
 814                 else if (parse->sortClause)
 815                         root->query_pathkeys = root->sort_pathkeys;
 816                 else
 817                         root->query_pathkeys = NIL;
 818
 819                 /*
 820                  * Generate the best unsorted and presorted paths for this Query (but
 821                  * note there may not be any presorted path).  query_planner will also
 822                  * estimate the number of groups in the query, and canonicalize all
 823                  * the pathkeys.
 824                  */
 825                 query_planner(root, sub_tlist, tuple_fraction,
 826                                           &cheapest_path, &sorted_path, &dNumGroups);
 827
 828                 group_pathkeys = root->group_pathkeys;
 829                 sort_pathkeys = root->sort_pathkeys;
 830
 831                 /*
 832                  * If grouping, extract the grouping operators and decide whether we
 833                  * want to use hashed grouping.
 834                  */
 835                 if (parse->groupClause)
 836                 {
 837                         groupOperators = extract_grouping_ops(parse->groupClause);
 838                         use_hashed_grouping =
 839                                 choose_hashed_grouping(root, tuple_fraction,
 840                                                                            cheapest_path, sorted_path,
 841                                                                            groupOperators, dNumGroups,
 842                                                                            &agg_counts);
 843
 844                         /* Also convert # groups to long int --- but 'ware overflow! */
 845                         numGroups = (long) Min(dNumGroups, (double) LONG_MAX);
 846                 }
 847
 848                 /*
 849                  * Select the best path.  If we are doing hashed grouping, we will
 850                  * always read all the input tuples, so use the cheapest-total path.
 851                  * Otherwise, trust query_planner's decision about which to use.
 852                  */
 853                 if (use_hashed_grouping || !sorted_path)
 854                         best_path = cheapest_path;
 855                 else
 856                         best_path = sorted_path;
 857
 858                 /*
 859                  * Check to see if it's possible to optimize MIN/MAX aggregates. If
 860                  * so, we will forget all the work we did so far to choose a "regular"
 861                  * path ... but we had to do it anyway to be able to tell which way is
 862                  * cheaper.
 863                  */
 864                 result_plan = optimize_minmax_aggregates(root,
 865                                                                                                  tlist,
 866                                                                                                  best_path);
 867                 if (result_plan != NULL)
 868                 {
 869                         /*
 870                          * optimize_minmax_aggregates generated the full plan, with the
 871                          * right tlist, and it has no sort order.
 872                          */
 873                         current_pathkeys = NIL;
 874                 }
 875                 else
 876                 {
 877                         /*
 878                          * Normal case --- create a plan according to query_planner's
 879                          * results.
 880                          */
 881                         result_plan = create_plan(root, best_path);
 882                         current_pathkeys = best_path->pathkeys;
 883
 884                         /*
 885                          * create_plan() returns a plan with just a "flat" tlist of
 886                          * required Vars.  Usually we need to insert the sub_tlist as the
 887                          * tlist of the top plan node.  However, we can skip that if we
 888                          * determined that whatever query_planner chose to return will be
 889                          * good enough.
 890                          */
 891                         if (need_tlist_eval)
 892                         {
 893                                 /*
 894                                  * If the top-level plan node is one that cannot do expression
 895                                  * evaluation, we must insert a Result node to project the
 896                                  * desired tlist.
 897                                  */
 898                                 if (!is_projection_capable_plan(result_plan))
 899                                 {
 900                                         result_plan = (Plan *) make_result(sub_tlist, NULL,
 901                                                                                                            result_plan);
 902                                 }
 903                                 else
 904                                 {
 905                                         /*
 906                                          * Otherwise, just replace the subplan's flat tlist with
 907                                          * the desired tlist.
 908                                          */
 909                                         result_plan->targetlist = sub_tlist;
 910                                 }
 911
 912                                 /*
 913                                  * Also, account for the cost of evaluation of the sub_tlist.
 914                                  *
 915                                  * Up to now, we have only been dealing with "flat" tlists,
 916                                  * containing just Vars.  So their evaluation cost is zero
 917                                  * according to the model used by cost_qual_eval() (or if you
 918                                  * prefer, the cost is factored into cpu_tuple_cost).  Thus we
 919                                  * can avoid accounting for tlist cost throughout
 920                                  * query_planner() and subroutines.  But now we've inserted a
 921                                  * tlist that might contain actual operators, sub-selects, etc
 922                                  * --- so we'd better account for its cost.
 923                                  *
 924                                  * Below this point, any tlist eval cost for added-on nodes
 925                                  * should be accounted for as we create those nodes.
 926                                  * Presently, of the node types we can add on, only Agg and
 927                                  * Group project new tlists (the rest just copy their input
 928                                  * tuples) --- so make_agg() and make_group() are responsible
 929                                  * for computing the added cost.
 930                                  */
 931                                 cost_qual_eval(&tlist_cost, sub_tlist);
 932                                 result_plan->startup_cost += tlist_cost.startup;
 933                                 result_plan->total_cost += tlist_cost.startup +
 934                                         tlist_cost.per_tuple * result_plan->plan_rows;
 935                         }
 936                         else
 937                         {
 938                                 /*
 939                                  * Since we're using query_planner's tlist and not the one
 940                                  * make_subplanTargetList calculated, we have to refigure any
 941                                  * grouping-column indexes make_subplanTargetList computed.
 942                                  */
 943                                 locate_grouping_columns(root, tlist, result_plan->targetlist,
 944                                                                                 groupColIdx);
 945                         }
 946
 947                         /*
 948                          * Insert AGG or GROUP node if needed, plus an explicit sort step
 949                          * if necessary.
 950                          *
 951                          * HAVING clause, if any, becomes qual of the Agg or Group node.
 952                          */
 953                         if (use_hashed_grouping)
 954                         {
 955                                 /* Hashed aggregate plan --- no sort needed */
 956                                 result_plan = (Plan *) make_agg(root,
 957                                                                                                 tlist,
 958                                                                                                 (List *) parse->havingQual,
 959                                                                                                 AGG_HASHED,
 960                                                                                                 numGroupCols,
 961                                                                                                 groupColIdx,
 962                                                                                                 groupOperators,
 963                                                                                                 numGroups,
 964                                                                                                 agg_counts.numAggs,
 965                                                                                                 result_plan);
 966                                 /* Hashed aggregation produces randomly-ordered results */
 967                                 current_pathkeys = NIL;
 968                         }
 969                         else if (parse->hasAggs)
 970                         {
 971                                 /* Plain aggregate plan --- sort if needed */
 972                                 AggStrategy aggstrategy;
 973
 974                                 if (parse->groupClause)
 975                                 {
 976                                         if (!pathkeys_contained_in(group_pathkeys,
 977                                                                                            current_pathkeys))
 978                                         {
 979                                                 result_plan = (Plan *)
 980                                                         make_sort_from_groupcols(root,
 981                                                                                                          parse->groupClause,
 982                                                                                                          groupColIdx,
 983                                                                                                          result_plan);
 984                                                 current_pathkeys = group_pathkeys;
 985                                         }
 986                                         aggstrategy = AGG_SORTED;
 987
 988                                         /*
 989                                          * The AGG node will not change the sort ordering of its
 990                                          * groups, so current_pathkeys describes the result too.
 991                                          */
 992                                 }
 993                                 else
 994                                 {
 995                                         aggstrategy = AGG_PLAIN;
 996                                         /* Result will be only one row anyway; no sort order */
 997                                         current_pathkeys = NIL;
 998                                 }
 999
1000                                 result_plan = (Plan *) make_agg(root,
1001                                                                                                 tlist,
1002                                                                                                 (List *) parse->havingQual,
1003                                                                                                 aggstrategy,
1004                                                                                                 numGroupCols,
1005                                                                                                 groupColIdx,
1006                                                                                                 groupOperators,
1007                                                                                                 numGroups,
1008                                                                                                 agg_counts.numAggs,
1009                                                                                                 result_plan);
1010                         }
1011                         else if (parse->groupClause)
1012                         {
1013                                 /*
1014                                  * GROUP BY without aggregation, so insert a group node (plus
1015                                  * the appropriate sort node, if necessary).
1016                                  *
1017                                  * Add an explicit sort if we couldn't make the path come out
1018                                  * the way the GROUP node needs it.
1019                                  */
1020                                 if (!pathkeys_contained_in(group_pathkeys, current_pathkeys))
1021                                 {
1022                                         result_plan = (Plan *)
1023                                                 make_sort_from_groupcols(root,
1024                                                                                                  parse->groupClause,
1025                                                                                                  groupColIdx,
1026                                                                                                  result_plan);
1027                                         current_pathkeys = group_pathkeys;
1028                                 }
1029
1030                                 result_plan = (Plan *) make_group(root,
1031                                                                                                   tlist,
1032                                                                                                   (List *) parse->havingQual,
1033                                                                                                   numGroupCols,
1034                                                                                                   groupColIdx,
1035                                                                                                   groupOperators,
1036                                                                                                   dNumGroups,
1037                                                                                                   result_plan);
1038                                 /* The Group node won't change sort ordering */
1039                         }
1040                         else if (root->hasHavingQual)
1041                         {
1042                                 /*
1043                                  * No aggregates, and no GROUP BY, but we have a HAVING qual.
1044                                  * This is a degenerate case in which we are supposed to emit
1045                                  * either 0 or 1 row depending on whether HAVING succeeds.
1046                                  * Furthermore, there cannot be any variables in either HAVING
1047                                  * or the targetlist, so we actually do not need the FROM
1048                                  * table at all!  We can just throw away the plan-so-far and
1049                                  * generate a Result node.      This is a sufficiently unusual
1050                                  * corner case that it's not worth contorting the structure of
1051                                  * this routine to avoid having to generate the plan in the
1052                                  * first place.
1053                                  */
1054                                 result_plan = (Plan *) make_result(tlist,
1055                                                                                                    parse->havingQual,
1056                                                                                                    NULL);
1057                         }
1058                 }                                               /* end of non-minmax-aggregate case */
1059         }                                                       /* end of if (setOperations) */
1060
1061         /*
1062          * If we were not able to make the plan come out in the right order, add
1063          * an explicit sort step.
1064          */
1065         if (parse->sortClause)
1066         {
1067                 if (!pathkeys_contained_in(sort_pathkeys, current_pathkeys))
1068                 {
1069                         result_plan = (Plan *) make_sort_from_pathkeys(root,
1070                                                                                                                    result_plan,
1071                                                                                                                    sort_pathkeys);
1072                         current_pathkeys = sort_pathkeys;
1073                 }
1074         }
1075
1076         /*
1077          * If there is a DISTINCT clause, add the UNIQUE node.
1078          */
1079         if (parse->distinctClause)
1080         {
1081                 result_plan = (Plan *) make_unique(result_plan, parse->distinctClause);
1082
1083                 /*
1084                  * If there was grouping or aggregation, leave plan_rows as-is (ie,
1085                  * assume the result was already mostly unique).  If not, use the
1086                  * number of distinct-groups calculated by query_planner.
1087                  */
1088                 if (!parse->groupClause && !root->hasHavingQual && !parse->hasAggs)
1089                         result_plan->plan_rows = dNumGroups;
1090         }
1091
1092         /*
1093          * Finally, if there is a LIMIT/OFFSET clause, add the LIMIT node.
1094          */
1095         if (parse->limitCount || parse->limitOffset)
1096         {
1097                 result_plan = (Plan *) make_limit(result_plan,
1098                                                                                   parse->limitOffset,
1099                                                                                   parse->limitCount,
1100                                                                                   offset_est,
1101                                                                                   count_est);
1102         }
1103
1104         /*
1105          * Deal with the RETURNING clause if any.  It's convenient to pass the
1106          * returningList through setrefs.c now rather than at top level (if we
1107          * waited, handling inherited UPDATE/DELETE would be much harder).
1108          */
1109         if (parse->returningList)
1110         {
1111                 List       *rlist;
1112
1113                 rlist = set_returning_clause_references(parse->returningList,
1114                                                                                                 result_plan,
1115                                                                                                 parse->resultRelation);
1116                 root->returningLists = list_make1(rlist);
1117         }
1118         else
1119                 root->returningLists = NIL;
1120
1121         /* Compute result-relations list if needed */
1122         if (parse->resultRelation)
1123                 root->resultRelations = list_make1_int(parse->resultRelation);
1124         else
1125                 root->resultRelations = NIL;
1126
1127         /*
1128          * Return the actual output ordering in query_pathkeys for possible use by
1129          * an outer query level.
1130          */
1131         root->query_pathkeys = current_pathkeys;
1132
1133         return result_plan;
1134 }
1135
1136 /*
1137  * Detect whether a plan node is a "dummy" plan created when a relation
1138  * is deemed not to need scanning due to constraint exclusion.
1139  *
1140  * Currently, such dummy plans are Result nodes with constant FALSE
1141  * filter quals.
1142  */
1143 static bool
1144 is_dummy_plan(Plan *plan)
1145 {
1146         if (IsA(plan, Result))
1147         {
1148                 List       *rcqual = (List *) ((Result *) plan)->resconstantqual;
1149
1150                 if (list_length(rcqual) == 1)
1151                 {
1152                         Const      *constqual = (Const *) linitial(rcqual);
1153
1154                         if (constqual && IsA(constqual, Const))
1155                         {
1156                                 if (!constqual->constisnull &&
1157                                         !DatumGetBool(constqual->constvalue))
1158                                         return true;
1159                         }
1160                 }
1161         }
1162         return false;
1163 }
1164
1165 /*
1166  * preprocess_limit - do pre-estimation for LIMIT and/or OFFSET clauses
1167  *
1168  * We try to estimate the values of the LIMIT/OFFSET clauses, and pass the
1169  * results back in *count_est and *offset_est.  These variables are set to
1170  * 0 if the corresponding clause is not present, and -1 if it's present
1171  * but we couldn't estimate the value for it.  (The "0" convention is OK
1172  * for OFFSET but a little bit bogus for LIMIT: effectively we estimate
1173  * LIMIT 0 as though it were LIMIT 1.  But this is in line with the planner's
1174  * usual practice of never estimating less than one row.)  These values will
1175  * be passed to make_limit, which see if you change this code.
1176  *
1177  * The return value is the suitably adjusted tuple_fraction to use for
1178  * planning the query.  This adjustment is not overridable, since it reflects
1179  * plan actions that grouping_planner() will certainly take, not assumptions
1180  * about context.
1181  */
1182 static double
1183 preprocess_limit(PlannerInfo *root, double tuple_fraction,
1184                                  int64 *offset_est, int64 *count_est)
1185 {
1186         Query      *parse = root->parse;
1187         Node       *est;
1188         double          limit_fraction;
1189
1190         /* Should not be called unless LIMIT or OFFSET */
1191         Assert(parse->limitCount || parse->limitOffset);
1192
1193         /*
1194          * Try to obtain the clause values.  We use estimate_expression_value
1195          * primarily because it can sometimes do something useful with Params.
1196          */
1197         if (parse->limitCount)
1198         {
1199                 est = estimate_expression_value(root, parse->limitCount);
1200                 if (est && IsA(est, Const))
1201                 {
1202                         if (((Const *) est)->constisnull)
1203                         {
1204                                 /* NULL indicates LIMIT ALL, ie, no limit */
1205                                 *count_est = 0; /* treat as not present */
1206                         }
1207                         else
1208                         {
1209                                 *count_est = DatumGetInt64(((Const *) est)->constvalue);
1210                                 if (*count_est <= 0)
1211                                         *count_est = 1;         /* force to at least 1 */
1212                         }
1213                 }
1214                 else
1215                         *count_est = -1;        /* can't estimate */
1216         }
1217         else
1218                 *count_est = 0;                 /* not present */
1219
1220         if (parse->limitOffset)
1221         {
1222                 est = estimate_expression_value(root, parse->limitOffset);
1223                 if (est && IsA(est, Const))
1224                 {
1225                         if (((Const *) est)->constisnull)
1226                         {
1227                                 /* Treat NULL as no offset; the executor will too */
1228                                 *offset_est = 0;        /* treat as not present */
1229                         }
1230                         else
1231                         {
1232                                 *offset_est = DatumGetInt64(((Const *) est)->constvalue);
1233                                 if (*offset_est < 0)
1234                                         *offset_est = 0;        /* less than 0 is same as 0 */
1235                         }
1236                 }
1237                 else
1238                         *offset_est = -1;       /* can't estimate */
1239         }
1240         else
1241                 *offset_est = 0;                /* not present */
1242
1243         if (*count_est != 0)
1244         {
1245                 /*
1246                  * A LIMIT clause limits the absolute number of tuples returned.
1247                  * However, if it's not a constant LIMIT then we have to guess; for
1248                  * lack of a better idea, assume 10% of the plan's result is wanted.
1249                  */
1250                 if (*count_est < 0 || *offset_est < 0)
1251                 {
1252                         /* LIMIT or OFFSET is an expression ... punt ... */
1253                         limit_fraction = 0.10;
1254                 }
1255                 else
1256                 {
1257                         /* LIMIT (plus OFFSET, if any) is max number of tuples needed */
1258                         limit_fraction = (double) *count_est + (double) *offset_est;
1259                 }
1260
1261                 /*
1262                  * If we have absolute limits from both caller and LIMIT, use the
1263                  * smaller value; likewise if they are both fractional.  If one is
1264                  * fractional and the other absolute, we can't easily determine which
1265                  * is smaller, but we use the heuristic that the absolute will usually
1266                  * be smaller.
1267                  */
1268                 if (tuple_fraction >= 1.0)
1269                 {
1270                         if (limit_fraction >= 1.0)
1271                         {
1272                                 /* both absolute */
1273                                 tuple_fraction = Min(tuple_fraction, limit_fraction);
1274                         }
1275                         else
1276                         {
1277                                 /* caller absolute, limit fractional; use caller's value */
1278                         }
1279                 }
1280                 else if (tuple_fraction > 0.0)
1281                 {
1282                         if (limit_fraction >= 1.0)
1283                         {
1284                                 /* caller fractional, limit absolute; use limit */
1285                                 tuple_fraction = limit_fraction;
1286                         }
1287                         else
1288                         {
1289                                 /* both fractional */
1290                                 tuple_fraction = Min(tuple_fraction, limit_fraction);
1291                         }
1292                 }
1293                 else
1294                 {
1295                         /* no info from caller, just use limit */
1296                         tuple_fraction = limit_fraction;
1297                 }
1298         }
1299         else if (*offset_est != 0 && tuple_fraction > 0.0)
1300         {
1301                 /*
1302                  * We have an OFFSET but no LIMIT.      This acts entirely differently
1303                  * from the LIMIT case: here, we need to increase rather than decrease
1304                  * the caller's tuple_fraction, because the OFFSET acts to cause more
1305                  * tuples to be fetched instead of fewer.  This only matters if we got
1306                  * a tuple_fraction > 0, however.
1307                  *
1308                  * As above, use 10% if OFFSET is present but unestimatable.
1309                  */
1310                 if (*offset_est < 0)
1311                         limit_fraction = 0.10;
1312                 else
1313                         limit_fraction = (double) *offset_est;
1314
1315                 /*
1316                  * If we have absolute counts from both caller and OFFSET, add them
1317                  * together; likewise if they are both fractional.      If one is
1318                  * fractional and the other absolute, we want to take the larger, and
1319                  * we heuristically assume that's the fractional one.
1320                  */
1321                 if (tuple_fraction >= 1.0)
1322                 {
1323                         if (limit_fraction >= 1.0)
1324                         {
1325                                 /* both absolute, so add them together */
1326                                 tuple_fraction += limit_fraction;
1327                         }
1328                         else
1329                         {
1330                                 /* caller absolute, limit fractional; use limit */
1331                                 tuple_fraction = limit_fraction;
1332                         }
1333                 }
1334                 else
1335                 {
1336                         if (limit_fraction >= 1.0)
1337                         {
1338                                 /* caller fractional, limit absolute; use caller's value */
1339                         }
1340                         else
1341                         {
1342                                 /* both fractional, so add them together */
1343                                 tuple_fraction += limit_fraction;
1344                                 if (tuple_fraction >= 1.0)
1345                                         tuple_fraction = 0.0;           /* assume fetch all */
1346                         }
1347                 }
1348         }
1349
1350         return tuple_fraction;
1351 }
1352
1353 /*
1354  * extract_grouping_ops - make an array of the equality operator OIDs
1355  *              for the GROUP BY clause
1356  */
1357 static Oid *
1358 extract_grouping_ops(List *groupClause)
1359 {
1360         int                     numCols = list_length(groupClause);
1361         int                     colno = 0;
1362         Oid                *groupOperators;
1363         ListCell   *glitem;
1364
1365         groupOperators = (Oid *) palloc(sizeof(Oid) * numCols);
1366
1367         foreach(glitem, groupClause)
1368         {
1369                 GroupClause *groupcl = (GroupClause *) lfirst(glitem);
1370
1371                 groupOperators[colno] = get_equality_op_for_ordering_op(groupcl->sortop);
1372                 if (!OidIsValid(groupOperators[colno]))         /* shouldn't happen */
1373                         elog(ERROR, "could not find equality operator for ordering operator %u",
1374                                  groupcl->sortop);
1375                 colno++;
1376         }
1377
1378         return groupOperators;
1379 }
1380
1381 /*
1382  * choose_hashed_grouping - should we use hashed grouping?
1383  */
1384 static bool
1385 choose_hashed_grouping(PlannerInfo *root, double tuple_fraction,
1386                                            Path *cheapest_path, Path *sorted_path,
1387                                            Oid *groupOperators, double dNumGroups,
1388                                            AggClauseCounts *agg_counts)
1389 {
1390         int                     numGroupCols = list_length(root->parse->groupClause);
1391         double          cheapest_path_rows;
1392         int                     cheapest_path_width;
1393         Size            hashentrysize;
1394         List       *current_pathkeys;
1395         Path            hashed_p;
1396         Path            sorted_p;
1397         int                     i;
1398
1399         /*
1400          * Check can't-do-it conditions, including whether the grouping operators
1401          * are hashjoinable.  (We assume hashing is OK if they are marked
1402          * oprcanhash.  If there isn't actually a supporting hash function,
1403          * the executor will complain at runtime.)
1404          *
1405          * Executor doesn't support hashed aggregation with DISTINCT aggregates.
1406          * (Doing so would imply storing *all* the input values in the hash table,
1407          * which seems like a certain loser.)
1408          */
1409         if (!enable_hashagg)
1410                 return false;
1411         if (agg_counts->numDistinctAggs != 0)
1412                 return false;
1413         for (i = 0; i < numGroupCols; i++)
1414         {
1415                 if (!op_hashjoinable(groupOperators[i]))
1416                         return false;
1417         }
1418
1419         /*
1420          * Don't do it if it doesn't look like the hashtable will fit into
1421          * work_mem.
1422          *
1423          * Beware here of the possibility that cheapest_path->parent is NULL. This
1424          * could happen if user does something silly like SELECT 'foo' GROUP BY 1;
1425          */
1426         if (cheapest_path->parent)
1427         {
1428                 cheapest_path_rows = cheapest_path->parent->rows;
1429                 cheapest_path_width = cheapest_path->parent->width;
1430         }
1431         else
1432         {
1433                 cheapest_path_rows = 1; /* assume non-set result */
1434                 cheapest_path_width = 100;              /* arbitrary */
1435         }
1436
1437         /* Estimate per-hash-entry space at tuple width... */
1438         hashentrysize = MAXALIGN(cheapest_path_width) + MAXALIGN(sizeof(MinimalTupleData));
1439         /* plus space for pass-by-ref transition values... */
1440         hashentrysize += agg_counts->transitionSpace;
1441         /* plus the per-hash-entry overhead */
1442         hashentrysize += hash_agg_entry_size(agg_counts->numAggs);
1443
1444         if (hashentrysize * dNumGroups > work_mem * 1024L)
1445                 return false;
1446
1447         /*
1448          * See if the estimated cost is no more than doing it the other way. While
1449          * avoiding the need for sorted input is usually a win, the fact that the
1450          * output won't be sorted may be a loss; so we need to do an actual cost
1451          * comparison.
1452          *
1453          * We need to consider cheapest_path + hashagg [+ final sort] versus
1454          * either cheapest_path [+ sort] + group or agg [+ final sort] or
1455          * presorted_path + group or agg [+ final sort] where brackets indicate a
1456          * step that may not be needed. We assume query_planner() will have
1457          * returned a presorted path only if it's a winner compared to
1458          * cheapest_path for this purpose.
1459          *
1460          * These path variables are dummies that just hold cost fields; we don't
1461          * make actual Paths for these steps.
1462          */
1463         cost_agg(&hashed_p, root, AGG_HASHED, agg_counts->numAggs,
1464                          numGroupCols, dNumGroups,
1465                          cheapest_path->startup_cost, cheapest_path->total_cost,
1466                          cheapest_path_rows);
1467         /* Result of hashed agg is always unsorted */
1468         if (root->sort_pathkeys)
1469                 cost_sort(&hashed_p, root, root->sort_pathkeys, hashed_p.total_cost,
1470                                   dNumGroups, cheapest_path_width);
1471
1472         if (sorted_path)
1473         {
1474                 sorted_p.startup_cost = sorted_path->startup_cost;
1475                 sorted_p.total_cost = sorted_path->total_cost;
1476                 current_pathkeys = sorted_path->pathkeys;
1477         }
1478         else
1479         {
1480                 sorted_p.startup_cost = cheapest_path->startup_cost;
1481                 sorted_p.total_cost = cheapest_path->total_cost;
1482                 current_pathkeys = cheapest_path->pathkeys;
1483         }
1484         if (!pathkeys_contained_in(root->group_pathkeys, current_pathkeys))
1485         {
1486                 cost_sort(&sorted_p, root, root->group_pathkeys, sorted_p.total_cost,
1487                                   cheapest_path_rows, cheapest_path_width);
1488                 current_pathkeys = root->group_pathkeys;
1489         }
1490
1491         if (root->parse->hasAggs)
1492                 cost_agg(&sorted_p, root, AGG_SORTED, agg_counts->numAggs,
1493                                  numGroupCols, dNumGroups,
1494                                  sorted_p.startup_cost, sorted_p.total_cost,
1495                                  cheapest_path_rows);
1496         else
1497                 cost_group(&sorted_p, root, numGroupCols, dNumGroups,
1498                                    sorted_p.startup_cost, sorted_p.total_cost,
1499                                    cheapest_path_rows);
1500         /* The Agg or Group node will preserve ordering */
1501         if (root->sort_pathkeys &&
1502                 !pathkeys_contained_in(root->sort_pathkeys, current_pathkeys))
1503                 cost_sort(&sorted_p, root, root->sort_pathkeys, sorted_p.total_cost,
1504                                   dNumGroups, cheapest_path_width);
1505
1506         /*
1507          * Now make the decision using the top-level tuple fraction.  First we
1508          * have to convert an absolute count (LIMIT) into fractional form.
1509          */
1510         if (tuple_fraction >= 1.0)
1511                 tuple_fraction /= dNumGroups;
1512
1513         if (compare_fractional_path_costs(&hashed_p, &sorted_p,
1514                                                                           tuple_fraction) < 0)
1515         {
1516                 /* Hashed is cheaper, so use it */
1517                 return true;
1518         }
1519         return false;
1520 }
1521
1522 /*---------------
1523  * make_subplanTargetList
1524  *        Generate appropriate target list when grouping is required.
1525  *
1526  * When grouping_planner inserts Aggregate, Group, or Result plan nodes
1527  * above the result of query_planner, we typically want to pass a different
1528  * target list to query_planner than the outer plan nodes should have.
1529  * This routine generates the correct target list for the subplan.
1530  *
1531  * The initial target list passed from the parser already contains entries
1532  * for all ORDER BY and GROUP BY expressions, but it will not have entries
1533  * for variables used only in HAVING clauses; so we need to add those
1534  * variables to the subplan target list.  Also, we flatten all expressions
1535  * except GROUP BY items into their component variables; the other expressions
1536  * will be computed by the inserted nodes rather than by the subplan.
1537  * For example, given a query like
1538  *              SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
1539  * we want to pass this targetlist to the subplan:
1540  *              a,b,c,d,a+b
1541  * where the a+b target will be used by the Sort/Group steps, and the
1542  * other targets will be used for computing the final results.  (In the
1543  * above example we could theoretically suppress the a and b targets and
1544  * pass down only c,d,a+b, but it's not really worth the trouble to
1545  * eliminate simple var references from the subplan.  We will avoid doing
1546  * the extra computation to recompute a+b at the outer level; see
1547  * replace_vars_with_subplan_refs() in setrefs.c.)
1548  *
1549  * If we are grouping or aggregating, *and* there are no non-Var grouping
1550  * expressions, then the returned tlist is effectively dummy; we do not
1551  * need to force it to be evaluated, because all the Vars it contains
1552  * should be present in the output of query_planner anyway.
1553  *
1554  * 'tlist' is the query's target list.
1555  * 'groupColIdx' receives an array of column numbers for the GROUP BY
1556  *                      expressions (if there are any) in the subplan's target list.
1557  * 'need_tlist_eval' is set true if we really need to evaluate the
1558  *                      result tlist.
1559  *
1560  * The result is the targetlist to be passed to the subplan.
1561  *---------------
1562  */
1563 static List *
1564 make_subplanTargetList(PlannerInfo *root,
1565                                            List *tlist,
1566                                            AttrNumber **groupColIdx,
1567                                            bool *need_tlist_eval)
1568 {
1569         Query      *parse = root->parse;
1570         List       *sub_tlist;
1571         List       *extravars;
1572         int                     numCols;
1573
1574         *groupColIdx = NULL;
1575
1576         /*
1577          * If we're not grouping or aggregating, there's nothing to do here;
1578          * query_planner should receive the unmodified target list.
1579          */
1580         if (!parse->hasAggs && !parse->groupClause && !root->hasHavingQual)
1581         {
1582                 *need_tlist_eval = true;
1583                 return tlist;
1584         }
1585
1586         /*
1587          * Otherwise, start with a "flattened" tlist (having just the vars
1588          * mentioned in the targetlist and HAVING qual --- but not upper- level
1589          * Vars; they will be replaced by Params later on).
1590          */
1591         sub_tlist = flatten_tlist(tlist);
1592         extravars = pull_var_clause(parse->havingQual, false);
1593         sub_tlist = add_to_flat_tlist(sub_tlist, extravars);
1594         list_free(extravars);
1595         *need_tlist_eval = false;       /* only eval if not flat tlist */
1596
1597         /*
1598          * If grouping, create sub_tlist entries for all GROUP BY expressions
1599          * (GROUP BY items that are simple Vars should be in the list already),
1600          * and make an array showing where the group columns are in the sub_tlist.
1601          */
1602         numCols = list_length(parse->groupClause);
1603         if (numCols > 0)
1604         {
1605                 int                     keyno = 0;
1606                 AttrNumber *grpColIdx;
1607                 ListCell   *gl;
1608
1609                 grpColIdx = (AttrNumber *) palloc(sizeof(AttrNumber) * numCols);
1610                 *groupColIdx = grpColIdx;
1611
1612                 foreach(gl, parse->groupClause)
1613                 {
1614                         GroupClause *grpcl = (GroupClause *) lfirst(gl);
1615                         Node       *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1616                         TargetEntry *te = NULL;
1617                         ListCell   *sl;
1618
1619                         /* Find or make a matching sub_tlist entry */
1620                         foreach(sl, sub_tlist)
1621                         {
1622                                 te = (TargetEntry *) lfirst(sl);
1623                                 if (equal(groupexpr, te->expr))
1624                                         break;
1625                         }
1626                         if (!sl)
1627                         {
1628                                 te = makeTargetEntry((Expr *) groupexpr,
1629                                                                          list_length(sub_tlist) + 1,
1630                                                                          NULL,
1631                                                                          false);
1632                                 sub_tlist = lappend(sub_tlist, te);
1633                                 *need_tlist_eval = true;                /* it's not flat anymore */
1634                         }
1635
1636                         /* and save its resno */
1637                         grpColIdx[keyno++] = te->resno;
1638                 }
1639         }
1640
1641         return sub_tlist;
1642 }
1643
1644 /*
1645  * locate_grouping_columns
1646  *              Locate grouping columns in the tlist chosen by query_planner.
1647  *
1648  * This is only needed if we don't use the sub_tlist chosen by
1649  * make_subplanTargetList.      We have to forget the column indexes found
1650  * by that routine and re-locate the grouping vars in the real sub_tlist.
1651  */
1652 static void
1653 locate_grouping_columns(PlannerInfo *root,
1654                                                 List *tlist,
1655                                                 List *sub_tlist,
1656                                                 AttrNumber *groupColIdx)
1657 {
1658         int                     keyno = 0;
1659         ListCell   *gl;
1660
1661         /*
1662          * No work unless grouping.
1663          */
1664         if (!root->parse->groupClause)
1665         {
1666                 Assert(groupColIdx == NULL);
1667                 return;
1668         }
1669         Assert(groupColIdx != NULL);
1670
1671         foreach(gl, root->parse->groupClause)
1672         {
1673                 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1674                 Node       *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1675                 TargetEntry *te = NULL;
1676                 ListCell   *sl;
1677
1678                 foreach(sl, sub_tlist)
1679                 {
1680                         te = (TargetEntry *) lfirst(sl);
1681                         if (equal(groupexpr, te->expr))
1682                                 break;
1683                 }
1684                 if (!sl)
1685                         elog(ERROR, "failed to locate grouping columns");
1686
1687                 groupColIdx[keyno++] = te->resno;
1688         }
1689 }
1690
1691 /*
1692  * postprocess_setop_tlist
1693  *        Fix up targetlist returned by plan_set_operations().
1694  *
1695  * We need to transpose sort key info from the orig_tlist into new_tlist.
1696  * NOTE: this would not be good enough if we supported resjunk sort keys
1697  * for results of set operations --- then, we'd need to project a whole
1698  * new tlist to evaluate the resjunk columns.  For now, just ereport if we
1699  * find any resjunk columns in orig_tlist.
1700  */
1701 static List *
1702 postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
1703 {
1704         ListCell   *l;
1705         ListCell   *orig_tlist_item = list_head(orig_tlist);
1706
1707         foreach(l, new_tlist)
1708         {
1709                 TargetEntry *new_tle = (TargetEntry *) lfirst(l);
1710                 TargetEntry *orig_tle;
1711
1712                 /* ignore resjunk columns in setop result */
1713                 if (new_tle->resjunk)
1714                         continue;
1715
1716                 Assert(orig_tlist_item != NULL);
1717                 orig_tle = (TargetEntry *) lfirst(orig_tlist_item);
1718                 orig_tlist_item = lnext(orig_tlist_item);
1719                 if (orig_tle->resjunk)  /* should not happen */
1720                         elog(ERROR, "resjunk output columns are not implemented");
1721                 Assert(new_tle->resno == orig_tle->resno);
1722                 new_tle->ressortgroupref = orig_tle->ressortgroupref;
1723         }
1724         if (orig_tlist_item != NULL)
1725                 elog(ERROR, "resjunk output columns are not implemented");
1726         return new_tlist;
1727 }