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 1.213 2007/02/19 07:03:29 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 Plan *
  83 planner(Query *parse, bool isCursor, int cursorOptions,
  84                 ParamListInfo boundParams)
  85 {
  86         PlannerGlobal *glob;
  87         double          tuple_fraction;
  88         Plan       *result_plan;
  89
  90         /*
  91          * Set up global state for this planner invocation.  This data is needed
  92          * across all levels of sub-Query that might exist in the given command,
  93          * so we keep it in a separate struct that's linked to by each per-Query
  94          * PlannerInfo.
  95          */
  96         glob = makeNode(PlannerGlobal);
  97
  98         glob->boundParams = boundParams;
  99         glob->paramlist = NIL;
 100         glob->next_plan_id = 0;
 101
 102         /* Determine what fraction of the plan is likely to be scanned */
 103         if (isCursor)
 104         {
 105                 /*
 106                  * We have no real idea how many tuples the user will ultimately FETCH
 107                  * from a cursor, but it seems a good bet that he doesn't want 'em
 108                  * all.  Optimize for 10% retrieval (you gotta better number?  Should
 109                  * this be a SETtable parameter?)
 110                  */
 111                 tuple_fraction = 0.10;
 112         }
 113         else
 114         {
 115                 /* Default assumption is we need all the tuples */
 116                 tuple_fraction = 0.0;
 117         }
 118
 119         /* primary planning entry point (may recurse for subqueries) */
 120         result_plan = subquery_planner(glob, parse, 1, tuple_fraction, NULL);
 121
 122         /*
 123          * If creating a plan for a scrollable cursor, make sure it can run
 124          * backwards on demand.  Add a Material node at the top at need.
 125          */
 126         if (isCursor && (cursorOptions & CURSOR_OPT_SCROLL))
 127         {
 128                 if (!ExecSupportsBackwardScan(result_plan))
 129                         result_plan = materialize_finished_plan(result_plan);
 130         }
 131
 132         /* final cleanup of the plan */
 133         result_plan = set_plan_references(result_plan, parse->rtable);
 134
 135         /* executor wants to know total number of Params used overall */
 136         result_plan->nParamExec = list_length(glob->paramlist);
 137
 138         return result_plan;
 139 }
 140
 141
 142 /*--------------------
 143  * subquery_planner
 144  *        Invokes the planner on a subquery.  We recurse to here for each
 145  *        sub-SELECT found in the query tree.
 146  *
 147  * glob is the global state for the current planner run.
 148  * parse is the querytree produced by the parser & rewriter.
 149  * level is the current recursion depth (1 at the top-level Query).
 150  * tuple_fraction is the fraction of tuples we expect will be retrieved.
 151  * tuple_fraction is interpreted as explained for grouping_planner, below.
 152  *
 153  * If subquery_pathkeys isn't NULL, it receives a list of pathkeys indicating
 154  * the output sort ordering of the completed plan.
 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(PlannerGlobal *glob, Query *parse,
 170                                  Index level, double tuple_fraction,
 171                                  List **subquery_pathkeys)
 172 {
 173         int                     saved_plan_id = glob->next_plan_id;
 174         PlannerInfo *root;
 175         Plan       *plan;
 176         List       *newHaving;
 177         ListCell   *l;
 178
 179         /* Create a PlannerInfo data structure for this subquery */
 180         root = makeNode(PlannerInfo);
 181         root->parse = parse;
 182         root->glob = glob;
 183         root->query_level = level;
 184         root->planner_cxt = CurrentMemoryContext;
 185         root->init_plans = NIL;
 186         root->eq_classes = NIL;
 187         root->in_info_list = NIL;
 188         root->append_rel_list = NIL;
 189
 190         /*
 191          * Look for IN clauses at the top level of WHERE, and transform them into
 192          * joins.  Note that this step only handles IN clauses originally at top
 193          * level of WHERE; if we pull up any subqueries in the next step, their
 194          * INs are processed just before pulling them up.
 195          */
 196         if (parse->hasSubLinks)
 197                 parse->jointree->quals = pull_up_IN_clauses(root,
 198                                                                                                         parse->jointree->quals);
 199
 200         /*
 201          * Check to see if any subqueries in the rangetable can be merged into
 202          * this query.
 203          */
 204         parse->jointree = (FromExpr *)
 205                 pull_up_subqueries(root, (Node *) parse->jointree, false, false);
 206
 207         /*
 208          * Detect whether any rangetable entries are RTE_JOIN kind; if not, we can
 209          * avoid the expense of doing flatten_join_alias_vars().  Also check for
 210          * outer joins --- if none, we can skip reduce_outer_joins() and some
 211          * other processing.  This must be done after we have done
 212          * pull_up_subqueries, of course.
 213          *
 214          * Note: if reduce_outer_joins manages to eliminate all outer joins,
 215          * root->hasOuterJoins is not reset currently.  This is OK since its
 216          * purpose is merely to suppress unnecessary processing in simple cases.
 217          */
 218         root->hasJoinRTEs = false;
 219         root->hasOuterJoins = false;
 220         foreach(l, parse->rtable)
 221         {
 222                 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
 223
 224                 if (rte->rtekind == RTE_JOIN)
 225                 {
 226                         root->hasJoinRTEs = true;
 227                         if (IS_OUTER_JOIN(rte->jointype))
 228                         {
 229                                 root->hasOuterJoins = true;
 230                                 /* Can quit scanning once we find an outer join */
 231                                 break;
 232                         }
 233                 }
 234         }
 235
 236         /*
 237          * Expand any rangetable entries that are inheritance sets into "append
 238          * relations".  This can add entries to the rangetable, but they must be
 239          * plain base relations not joins, so it's OK (and marginally more
 240          * efficient) to do it after checking for join RTEs.  We must do it after
 241          * pulling up subqueries, else we'd fail to handle inherited tables in
 242          * subqueries.
 243          */
 244         expand_inherited_tables(root);
 245
 246         /*
 247          * Set hasHavingQual to remember if HAVING clause is present.  Needed
 248          * because preprocess_expression will reduce a constant-true condition to
 249          * an empty qual list ... but "HAVING TRUE" is not a semantic no-op.
 250          */
 251         root->hasHavingQual = (parse->havingQual != NULL);
 252
 253         /* Clear this flag; might get set in distribute_qual_to_rels */
 254         root->hasPseudoConstantQuals = false;
 255
 256         /*
 257          * Do expression preprocessing on targetlist and quals.
 258          */
 259         parse->targetList = (List *)
 260                 preprocess_expression(root, (Node *) parse->targetList,
 261                                                           EXPRKIND_TARGET);
 262
 263         parse->returningList = (List *)
 264                 preprocess_expression(root, (Node *) parse->returningList,
 265                                                           EXPRKIND_TARGET);
 266
 267         preprocess_qual_conditions(root, (Node *) parse->jointree);
 268
 269         parse->havingQual = preprocess_expression(root, parse->havingQual,
 270                                                                                           EXPRKIND_QUAL);
 271
 272         parse->limitOffset = preprocess_expression(root, parse->limitOffset,
 273                                                                                            EXPRKIND_LIMIT);
 274         parse->limitCount = preprocess_expression(root, parse->limitCount,
 275                                                                                           EXPRKIND_LIMIT);
 276
 277         root->in_info_list = (List *)
 278                 preprocess_expression(root, (Node *) root->in_info_list,
 279                                                           EXPRKIND_ININFO);
 280         root->append_rel_list = (List *)
 281                 preprocess_expression(root, (Node *) root->append_rel_list,
 282                                                           EXPRKIND_APPINFO);
 283
 284         /* Also need to preprocess expressions for function and values RTEs */
 285         foreach(l, parse->rtable)
 286         {
 287                 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
 288
 289                 if (rte->rtekind == RTE_FUNCTION)
 290                         rte->funcexpr = preprocess_expression(root, rte->funcexpr,
 291                                                                                                   EXPRKIND_RTFUNC);
 292                 else if (rte->rtekind == RTE_VALUES)
 293                         rte->values_lists = (List *)
 294                                 preprocess_expression(root, (Node *) rte->values_lists,
 295                                                                           EXPRKIND_VALUES);
 296         }
 297
 298         /*
 299          * In some cases we may want to transfer a HAVING clause into WHERE. We
 300          * cannot do so if the HAVING clause contains aggregates (obviously) or
 301          * volatile functions (since a HAVING clause is supposed to be executed
 302          * only once per group).  Also, it may be that the clause is so expensive
 303          * to execute that we're better off doing it only once per group, despite
 304          * the loss of selectivity.  This is hard to estimate short of doing the
 305          * entire planning process twice, so we use a heuristic: clauses
 306          * containing subplans are left in HAVING.      Otherwise, we move or copy the
 307          * HAVING clause into WHERE, in hopes of eliminating tuples before
 308          * aggregation instead of after.
 309          *
 310          * If the query has explicit grouping then we can simply move such a
 311          * clause into WHERE; any group that fails the clause will not be in the
 312          * output because none of its tuples will reach the grouping or
 313          * aggregation stage.  Otherwise we must have a degenerate (variable-free)
 314          * HAVING clause, which we put in WHERE so that query_planner() can use it
 315          * in a gating Result node, but also keep in HAVING to ensure that we
 316          * don't emit a bogus aggregated row. (This could be done better, but it
 317          * seems not worth optimizing.)
 318          *
 319          * Note that both havingQual and parse->jointree->quals are in
 320          * implicitly-ANDed-list form at this point, even though they are declared
 321          * as Node *.
 322          */
 323         newHaving = NIL;
 324         foreach(l, (List *) parse->havingQual)
 325         {
 326                 Node       *havingclause = (Node *) lfirst(l);
 327
 328                 if (contain_agg_clause(havingclause) ||
 329                         contain_volatile_functions(havingclause) ||
 330                         contain_subplans(havingclause))
 331                 {
 332                         /* keep it in HAVING */
 333                         newHaving = lappend(newHaving, havingclause);
 334                 }
 335                 else if (parse->groupClause)
 336                 {
 337                         /* move it to WHERE */
 338                         parse->jointree->quals = (Node *)
 339                                 lappend((List *) parse->jointree->quals, havingclause);
 340                 }
 341                 else
 342                 {
 343                         /* put a copy in WHERE, keep it in HAVING */
 344                         parse->jointree->quals = (Node *)
 345                                 lappend((List *) parse->jointree->quals,
 346                                                 copyObject(havingclause));
 347                         newHaving = lappend(newHaving, havingclause);
 348                 }
 349         }
 350         parse->havingQual = (Node *) newHaving;
 351
 352         /*
 353          * If we have any outer joins, try to reduce them to plain inner joins.
 354          * This step is most easily done after we've done expression
 355          * preprocessing.
 356          */
 357         if (root->hasOuterJoins)
 358                 reduce_outer_joins(root);
 359
 360         /*
 361          * Do the main planning.  If we have an inherited target relation, that
 362          * needs special processing, else go straight to grouping_planner.
 363          */
 364         if (parse->resultRelation &&
 365                 rt_fetch(parse->resultRelation, parse->rtable)->inh)
 366                 plan = inheritance_planner(root);
 367         else
 368                 plan = grouping_planner(root, tuple_fraction);
 369
 370         /*
 371          * If any subplans were generated, or if we're inside a subplan, build
 372          * initPlan list and extParam/allParam sets for plan nodes, and attach the
 373          * initPlans to the top plan node.
 374          */
 375         if (root->glob->next_plan_id != saved_plan_id || root->query_level > 1)
 376                 SS_finalize_plan(root, plan);
 377
 378         /* Return sort ordering info if caller wants it */
 379         if (subquery_pathkeys)
 380                 *subquery_pathkeys = root->query_pathkeys;
 381
 382         return plan;
 383 }
 384
 385 /*
 386  * preprocess_expression
 387  *              Do subquery_planner's preprocessing work for an expression,
 388  *              which can be a targetlist, a WHERE clause (including JOIN/ON
 389  *              conditions), or a HAVING clause.
 390  */
 391 static Node *
 392 preprocess_expression(PlannerInfo *root, Node *expr, int kind)
 393 {
 394         /*
 395          * Fall out quickly if expression is empty.  This occurs often enough to
 396          * be worth checking.  Note that null->null is the correct conversion for
 397          * implicit-AND result format, too.
 398          */
 399         if (expr == NULL)
 400                 return NULL;
 401
 402         /*
 403          * If the query has any join RTEs, replace join alias variables with
 404          * base-relation variables. We must do this before sublink processing,
 405          * else sublinks expanded out from join aliases wouldn't get processed. We
 406          * can skip it in VALUES lists, however, since they can't contain any Vars
 407          * at all.
 408          */
 409         if (root->hasJoinRTEs && kind != EXPRKIND_VALUES)
 410                 expr = flatten_join_alias_vars(root, expr);
 411
 412         /*
 413          * Simplify constant expressions.
 414          *
 415          * Note: this also flattens nested AND and OR expressions into N-argument
 416          * form.  All processing of a qual expression after this point must be
 417          * careful to maintain AND/OR flatness --- that is, do not generate a tree
 418          * with AND directly under AND, nor OR directly under OR.
 419          *
 420          * Because this is a relatively expensive process, we skip it when the
 421          * query is trivial, such as "SELECT 2+2;" or "INSERT ... VALUES()". The
 422          * expression will only be evaluated once anyway, so no point in
 423          * pre-simplifying; we can't execute it any faster than the executor can,
 424          * and we will waste cycles copying the tree.  Notice however that we
 425          * still must do it for quals (to get AND/OR flatness); and if we are in a
 426          * subquery we should not assume it will be done only once.
 427          *
 428          * For VALUES lists we never do this at all, again on the grounds that we
 429          * should optimize for one-time evaluation.
 430          */
 431         if (kind != EXPRKIND_VALUES &&
 432                 (root->parse->jointree->fromlist != NIL ||
 433                  kind == EXPRKIND_QUAL ||
 434                  root->query_level > 1))
 435                 expr = eval_const_expressions(expr);
 436
 437         /*
 438          * If it's a qual or havingQual, canonicalize it.
 439          */
 440         if (kind == EXPRKIND_QUAL)
 441         {
 442                 expr = (Node *) canonicalize_qual((Expr *) expr);
 443
 444 #ifdef OPTIMIZER_DEBUG
 445                 printf("After canonicalize_qual()\n");
 446                 pprint(expr);
 447 #endif
 448         }
 449
 450         /* Expand SubLinks to SubPlans */
 451         if (root->parse->hasSubLinks)
 452                 expr = SS_process_sublinks(root, expr, (kind == EXPRKIND_QUAL));
 453
 454         /*
 455          * XXX do not insert anything here unless you have grokked the comments in
 456          * SS_replace_correlation_vars ...
 457          */
 458
 459         /* Replace uplevel vars with Param nodes (this IS possible in VALUES) */
 460         if (root->query_level > 1)
 461                 expr = SS_replace_correlation_vars(root, expr);
 462
 463         /*
 464          * If it's a qual or havingQual, convert it to implicit-AND format. (We
 465          * don't want to do this before eval_const_expressions, since the latter
 466          * would be unable to simplify a top-level AND correctly. Also,
 467          * SS_process_sublinks expects explicit-AND format.)
 468          */
 469         if (kind == EXPRKIND_QUAL)
 470                 expr = (Node *) make_ands_implicit((Expr *) expr);
 471
 472         return expr;
 473 }
 474
 475 /*
 476  * preprocess_qual_conditions
 477  *              Recursively scan the query's jointree and do subquery_planner's
 478  *              preprocessing work on each qual condition found therein.
 479  */
 480 static void
 481 preprocess_qual_conditions(PlannerInfo *root, Node *jtnode)
 482 {
 483         if (jtnode == NULL)
 484                 return;
 485         if (IsA(jtnode, RangeTblRef))
 486         {
 487                 /* nothing to do here */
 488         }
 489         else if (IsA(jtnode, FromExpr))
 490         {
 491                 FromExpr   *f = (FromExpr *) jtnode;
 492                 ListCell   *l;
 493
 494                 foreach(l, f->fromlist)
 495                         preprocess_qual_conditions(root, lfirst(l));
 496
 497                 f->quals = preprocess_expression(root, f->quals, EXPRKIND_QUAL);
 498         }
 499         else if (IsA(jtnode, JoinExpr))
 500         {
 501                 JoinExpr   *j = (JoinExpr *) jtnode;
 502
 503                 preprocess_qual_conditions(root, j->larg);
 504                 preprocess_qual_conditions(root, j->rarg);
 505
 506                 j->quals = preprocess_expression(root, j->quals, EXPRKIND_QUAL);
 507         }
 508         else
 509                 elog(ERROR, "unrecognized node type: %d",
 510                          (int) nodeTag(jtnode));
 511 }
 512
 513 /*
 514  * inheritance_planner
 515  *        Generate a plan in the case where the result relation is an
 516  *        inheritance set.
 517  *
 518  * We have to handle this case differently from cases where a source relation
 519  * is an inheritance set. Source inheritance is expanded at the bottom of the
 520  * plan tree (see allpaths.c), but target inheritance has to be expanded at
 521  * the top.  The reason is that for UPDATE, each target relation needs a
 522  * different targetlist matching its own column set.  Also, for both UPDATE
 523  * and DELETE, the executor needs the Append plan node at the top, else it
 524  * can't keep track of which table is the current target table.  Fortunately,
 525  * the UPDATE/DELETE target can never be the nullable side of an outer join,
 526  * so it's OK to generate the plan this way.
 527  *
 528  * Returns a query plan.
 529  */
 530 static Plan *
 531 inheritance_planner(PlannerInfo *root)
 532 {
 533         Query      *parse = root->parse;
 534         int                     parentRTindex = parse->resultRelation;
 535         List       *subplans = NIL;
 536         List       *resultRelations = NIL;
 537         List       *returningLists = NIL;
 538         List       *rtable = NIL;
 539         List       *tlist = NIL;
 540         PlannerInfo subroot;
 541         ListCell   *l;
 542
 543         foreach(l, root->append_rel_list)
 544         {
 545                 AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
 546                 Plan       *subplan;
 547
 548                 /* append_rel_list contains all append rels; ignore others */
 549                 if (appinfo->parent_relid != parentRTindex)
 550                         continue;
 551
 552                 /*
 553                  * Generate modified query with this rel as target.  We have to be
 554                  * prepared to translate varnos in in_info_list as well as in the
 555                  * Query proper.
 556                  */
 557                 memcpy(&subroot, root, sizeof(PlannerInfo));
 558                 subroot.parse = (Query *)
 559                         adjust_appendrel_attrs((Node *) parse,
 560                                                                    appinfo);
 561                 subroot.in_info_list = (List *)
 562                         adjust_appendrel_attrs((Node *) root->in_info_list,
 563                                                                    appinfo);
 564                 subroot.init_plans = NIL;
 565                 /* There shouldn't be any OJ info to translate, as yet */
 566                 Assert(subroot.oj_info_list == NIL);
 567
 568                 /* Generate plan */
 569                 subplan = grouping_planner(&subroot, 0.0 /* retrieve all tuples */ );
 570
 571                 /*
 572                  * If this child rel was excluded by constraint exclusion, exclude it
 573                  * from the plan.
 574                  */
 575                 if (is_dummy_plan(subplan))
 576                         continue;
 577
 578                 /* Save rtable and tlist from first rel for use below */
 579                 if (subplans == NIL)
 580                 {
 581                         rtable = subroot.parse->rtable;
 582                         tlist = subplan->targetlist;
 583                 }
 584
 585                 subplans = lappend(subplans, subplan);
 586
 587                 /* Make sure any initplans from this rel get into the outer list */
 588                 root->init_plans = list_concat(root->init_plans, subroot.init_plans);
 589
 590                 /* Build target-relations list for the executor */
 591                 resultRelations = lappend_int(resultRelations, appinfo->child_relid);
 592
 593                 /* Build list of per-relation RETURNING targetlists */
 594                 if (parse->returningList)
 595                 {
 596                         Assert(list_length(subroot.parse->returningLists) == 1);
 597                         returningLists = list_concat(returningLists,
 598                                                                                  subroot.parse->returningLists);
 599                 }
 600         }
 601
 602         parse->resultRelations = resultRelations;
 603         parse->returningLists = returningLists;
 604
 605         /* Mark result as unordered (probably unnecessary) */
 606         root->query_pathkeys = NIL;
 607
 608         /*
 609          * If we managed to exclude every child rel, return a dummy plan
 610          */
 611         if (subplans == NIL)
 612                 return (Plan *) make_result(tlist,
 613                                                                         (Node *) list_make1(makeBoolConst(false,
 614                                                                                                                                           false)),
 615                                                                         NULL);
 616
 617         /*
 618          * Planning might have modified the rangetable, due to changes of the
 619          * Query structures inside subquery RTEs.  We have to ensure that this
 620          * gets propagated back to the master copy.  But can't do this until we
 621          * are done planning, because all the calls to grouping_planner need
 622          * virgin sub-Queries to work from.  (We are effectively assuming that
 623          * sub-Queries will get planned identically each time, or at least that
 624          * the impacts on their rangetables will be the same each time.)
 625          *
 626          * XXX should clean this up someday
 627          */
 628         parse->rtable = rtable;
 629
 630         return (Plan *) make_append(subplans, true, tlist);
 631 }
 632
 633 /*--------------------
 634  * grouping_planner
 635  *        Perform planning steps related to grouping, aggregation, etc.
 636  *        This primarily means adding top-level processing to the basic
 637  *        query plan produced by query_planner.
 638  *
 639  * tuple_fraction is the fraction of tuples we expect will be retrieved
 640  *
 641  * tuple_fraction is interpreted as follows:
 642  *        0: expect all tuples to be retrieved (normal case)
 643  *        0 < tuple_fraction < 1: expect the given fraction of tuples available
 644  *              from the plan to be retrieved
 645  *        tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
 646  *              expected to be retrieved (ie, a LIMIT specification)
 647  *
 648  * Returns a query plan.  Also, root->query_pathkeys is returned as the
 649  * actual output ordering of the plan (in pathkey format).
 650  *--------------------
 651  */
 652 static Plan *
 653 grouping_planner(PlannerInfo *root, double tuple_fraction)
 654 {
 655         Query      *parse = root->parse;
 656         List       *tlist = parse->targetList;
 657         int64           offset_est = 0;
 658         int64           count_est = 0;
 659         Plan       *result_plan;
 660         List       *current_pathkeys;
 661         List       *sort_pathkeys;
 662         double          dNumGroups = 0;
 663
 664         /* Tweak caller-supplied tuple_fraction if have LIMIT/OFFSET */
 665         if (parse->limitCount || parse->limitOffset)
 666                 tuple_fraction = preprocess_limit(root, tuple_fraction,
 667                                                                                   &offset_est, &count_est);
 668
 669         if (parse->setOperations)
 670         {
 671                 List       *set_sortclauses;
 672
 673                 /*
 674                  * If there's a top-level ORDER BY, assume we have to fetch all the
 675                  * tuples.      This might seem too simplistic given all the hackery below
 676                  * to possibly avoid the sort ... but a nonzero tuple_fraction is only
 677                  * of use to plan_set_operations() when the setop is UNION ALL, and
 678                  * the result of UNION ALL is always unsorted.
 679                  */
 680                 if (parse->sortClause)
 681                         tuple_fraction = 0.0;
 682
 683                 /*
 684                  * Construct the plan for set operations.  The result will not need
 685                  * any work except perhaps a top-level sort and/or LIMIT.
 686                  */
 687                 result_plan = plan_set_operations(root, tuple_fraction,
 688                                                                                   &set_sortclauses);
 689
 690                 /*
 691                  * Calculate pathkeys representing the sort order (if any) of the set
 692                  * operation's result.  We have to do this before overwriting the sort
 693                  * key information...
 694                  */
 695                 current_pathkeys = make_pathkeys_for_sortclauses(root,
 696                                                                                                                  set_sortclauses,
 697                                                                                                         result_plan->targetlist,
 698                                                                                                                  true);
 699
 700                 /*
 701                  * We should not need to call preprocess_targetlist, since we must be
 702                  * in a SELECT query node.      Instead, use the targetlist returned by
 703                  * plan_set_operations (since this tells whether it returned any
 704                  * resjunk columns!), and transfer any sort key information from the
 705                  * original tlist.
 706                  */
 707                 Assert(parse->commandType == CMD_SELECT);
 708
 709                 tlist = postprocess_setop_tlist(result_plan->targetlist, tlist);
 710
 711                 /*
 712                  * Can't handle FOR UPDATE/SHARE here (parser should have checked
 713                  * already, but let's make sure).
 714                  */
 715                 if (parse->rowMarks)
 716                         ereport(ERROR,
 717                                         (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
 718                                          errmsg("SELECT FOR UPDATE/SHARE is not allowed with UNION/INTERSECT/EXCEPT")));
 719
 720                 /*
 721                  * Calculate pathkeys that represent result ordering requirements
 722                  */
 723                 sort_pathkeys = make_pathkeys_for_sortclauses(root,
 724                                                                                                           parse->sortClause,
 725                                                                                                           tlist,
 726                                                                                                           true);
 727         }
 728         else
 729         {
 730                 /* No set operations, do regular planning */
 731                 List       *sub_tlist;
 732                 List       *group_pathkeys;
 733                 AttrNumber *groupColIdx = NULL;
 734                 Oid                *groupOperators = NULL;
 735                 bool            need_tlist_eval = true;
 736                 QualCost        tlist_cost;
 737                 Path       *cheapest_path;
 738                 Path       *sorted_path;
 739                 Path       *best_path;
 740                 long            numGroups = 0;
 741                 AggClauseCounts agg_counts;
 742                 int                     numGroupCols = list_length(parse->groupClause);
 743                 bool            use_hashed_grouping = false;
 744
 745                 MemSet(&agg_counts, 0, sizeof(AggClauseCounts));
 746
 747                 /* Preprocess targetlist */
 748                 tlist = preprocess_targetlist(root, tlist);
 749
 750                 /*
 751                  * Generate appropriate target list for subplan; may be different from
 752                  * tlist if grouping or aggregation is needed.
 753                  */
 754                 sub_tlist = make_subplanTargetList(root, tlist,
 755                                                                                    &groupColIdx, &need_tlist_eval);
 756
 757                 /*
 758                  * Calculate pathkeys that represent grouping/ordering requirements.
 759                  * Stash them in PlannerInfo so that query_planner can canonicalize
 760                  * them after EquivalenceClasses have been formed.
 761                  */
 762                 root->group_pathkeys =
 763                         make_pathkeys_for_sortclauses(root,
 764                                                                                   parse->groupClause,
 765                                                                                   tlist,
 766                                                                                   false);
 767                 root->sort_pathkeys =
 768                         make_pathkeys_for_sortclauses(root,
 769                                                                                   parse->sortClause,
 770                                                                                   tlist,
 771                                                                                   false);
 772
 773                 /*
 774                  * Will need actual number of aggregates for estimating costs.
 775                  *
 776                  * Note: we do not attempt to detect duplicate aggregates here; a
 777                  * somewhat-overestimated count is okay for our present purposes.
 778                  *
 779                  * Note: think not that we can turn off hasAggs if we find no aggs. It
 780                  * is possible for constant-expression simplification to remove all
 781                  * explicit references to aggs, but we still have to follow the
 782                  * aggregate semantics (eg, producing only one output row).
 783                  */
 784                 if (parse->hasAggs)
 785                 {
 786                         count_agg_clauses((Node *) tlist, &agg_counts);
 787                         count_agg_clauses(parse->havingQual, &agg_counts);
 788                 }
 789
 790                 /*
 791                  * Figure out whether we need a sorted result from query_planner.
 792                  *
 793                  * If we have a GROUP BY clause, then we want a result sorted properly
 794                  * for grouping.  Otherwise, if there is an ORDER BY clause, we want
 795                  * to sort by the ORDER BY clause.      (Note: if we have both, and ORDER
 796                  * BY is a superset of GROUP BY, it would be tempting to request sort
 797                  * by ORDER BY --- but that might just leave us failing to exploit an
 798                  * available sort order at all. Needs more thought...)
 799                  */
 800                 if (parse->groupClause)
 801                         root->query_pathkeys = root->group_pathkeys;
 802                 else if (parse->sortClause)
 803                         root->query_pathkeys = root->sort_pathkeys;
 804                 else
 805                         root->query_pathkeys = NIL;
 806
 807                 /*
 808                  * Generate the best unsorted and presorted paths for this Query (but
 809                  * note there may not be any presorted path).  query_planner will also
 810                  * estimate the number of groups in the query, and canonicalize all
 811                  * the pathkeys.
 812                  */
 813                 query_planner(root, sub_tlist, tuple_fraction,
 814                                           &cheapest_path, &sorted_path, &dNumGroups);
 815
 816                 group_pathkeys = root->group_pathkeys;
 817                 sort_pathkeys = root->sort_pathkeys;
 818
 819                 /*
 820                  * If grouping, extract the grouping operators and decide whether we
 821                  * want to use hashed grouping.
 822                  */
 823                 if (parse->groupClause)
 824                 {
 825                         groupOperators = extract_grouping_ops(parse->groupClause);
 826                         use_hashed_grouping =
 827                                 choose_hashed_grouping(root, tuple_fraction,
 828                                                                            cheapest_path, sorted_path,
 829                                                                            groupOperators, dNumGroups,
 830                                                                            &agg_counts);
 831
 832                         /* Also convert # groups to long int --- but 'ware overflow! */
 833                         numGroups = (long) Min(dNumGroups, (double) LONG_MAX);
 834                 }
 835
 836                 /*
 837                  * Select the best path.  If we are doing hashed grouping, we will
 838                  * always read all the input tuples, so use the cheapest-total path.
 839                  * Otherwise, trust query_planner's decision about which to use.
 840                  */
 841                 if (use_hashed_grouping || !sorted_path)
 842                         best_path = cheapest_path;
 843                 else
 844                         best_path = sorted_path;
 845
 846                 /*
 847                  * Check to see if it's possible to optimize MIN/MAX aggregates. If
 848                  * so, we will forget all the work we did so far to choose a "regular"
 849                  * path ... but we had to do it anyway to be able to tell which way is
 850                  * cheaper.
 851                  */
 852                 result_plan = optimize_minmax_aggregates(root,
 853                                                                                                  tlist,
 854                                                                                                  best_path);
 855                 if (result_plan != NULL)
 856                 {
 857                         /*
 858                          * optimize_minmax_aggregates generated the full plan, with the
 859                          * right tlist, and it has no sort order.
 860                          */
 861                         current_pathkeys = NIL;
 862                 }
 863                 else
 864                 {
 865                         /*
 866                          * Normal case --- create a plan according to query_planner's
 867                          * results.
 868                          */
 869                         result_plan = create_plan(root, best_path);
 870                         current_pathkeys = best_path->pathkeys;
 871
 872                         /*
 873                          * create_plan() returns a plan with just a "flat" tlist of
 874                          * required Vars.  Usually we need to insert the sub_tlist as the
 875                          * tlist of the top plan node.  However, we can skip that if we
 876                          * determined that whatever query_planner chose to return will be
 877                          * good enough.
 878                          */
 879                         if (need_tlist_eval)
 880                         {
 881                                 /*
 882                                  * If the top-level plan node is one that cannot do expression
 883                                  * evaluation, we must insert a Result node to project the
 884                                  * desired tlist.
 885                                  */
 886                                 if (!is_projection_capable_plan(result_plan))
 887                                 {
 888                                         result_plan = (Plan *) make_result(sub_tlist, NULL,
 889                                                                                                            result_plan);
 890                                 }
 891                                 else
 892                                 {
 893                                         /*
 894                                          * Otherwise, just replace the subplan's flat tlist with
 895                                          * the desired tlist.
 896                                          */
 897                                         result_plan->targetlist = sub_tlist;
 898                                 }
 899
 900                                 /*
 901                                  * Also, account for the cost of evaluation of the sub_tlist.
 902                                  *
 903                                  * Up to now, we have only been dealing with "flat" tlists,
 904                                  * containing just Vars.  So their evaluation cost is zero
 905                                  * according to the model used by cost_qual_eval() (or if you
 906                                  * prefer, the cost is factored into cpu_tuple_cost).  Thus we
 907                                  * can avoid accounting for tlist cost throughout
 908                                  * query_planner() and subroutines.  But now we've inserted a
 909                                  * tlist that might contain actual operators, sub-selects, etc
 910                                  * --- so we'd better account for its cost.
 911                                  *
 912                                  * Below this point, any tlist eval cost for added-on nodes
 913                                  * should be accounted for as we create those nodes.
 914                                  * Presently, of the node types we can add on, only Agg and
 915                                  * Group project new tlists (the rest just copy their input
 916                                  * tuples) --- so make_agg() and make_group() are responsible
 917                                  * for computing the added cost.
 918                                  */
 919                                 cost_qual_eval(&tlist_cost, sub_tlist);
 920                                 result_plan->startup_cost += tlist_cost.startup;
 921                                 result_plan->total_cost += tlist_cost.startup +
 922                                         tlist_cost.per_tuple * result_plan->plan_rows;
 923                         }
 924                         else
 925                         {
 926                                 /*
 927                                  * Since we're using query_planner's tlist and not the one
 928                                  * make_subplanTargetList calculated, we have to refigure any
 929                                  * grouping-column indexes make_subplanTargetList computed.
 930                                  */
 931                                 locate_grouping_columns(root, tlist, result_plan->targetlist,
 932                                                                                 groupColIdx);
 933                         }
 934
 935                         /*
 936                          * Insert AGG or GROUP node if needed, plus an explicit sort step
 937                          * if necessary.
 938                          *
 939                          * HAVING clause, if any, becomes qual of the Agg or Group node.
 940                          */
 941                         if (use_hashed_grouping)
 942                         {
 943                                 /* Hashed aggregate plan --- no sort needed */
 944                                 result_plan = (Plan *) make_agg(root,
 945                                                                                                 tlist,
 946                                                                                                 (List *) parse->havingQual,
 947                                                                                                 AGG_HASHED,
 948                                                                                                 numGroupCols,
 949                                                                                                 groupColIdx,
 950                                                                                                 groupOperators,
 951                                                                                                 numGroups,
 952                                                                                                 agg_counts.numAggs,
 953                                                                                                 result_plan);
 954                                 /* Hashed aggregation produces randomly-ordered results */
 955                                 current_pathkeys = NIL;
 956                         }
 957                         else if (parse->hasAggs)
 958                         {
 959                                 /* Plain aggregate plan --- sort if needed */
 960                                 AggStrategy aggstrategy;
 961
 962                                 if (parse->groupClause)
 963                                 {
 964                                         if (!pathkeys_contained_in(group_pathkeys,
 965                                                                                            current_pathkeys))
 966                                         {
 967                                                 result_plan = (Plan *)
 968                                                         make_sort_from_groupcols(root,
 969                                                                                                          parse->groupClause,
 970                                                                                                          groupColIdx,
 971                                                                                                          result_plan);
 972                                                 current_pathkeys = group_pathkeys;
 973                                         }
 974                                         aggstrategy = AGG_SORTED;
 975
 976                                         /*
 977                                          * The AGG node will not change the sort ordering of its
 978                                          * groups, so current_pathkeys describes the result too.
 979                                          */
 980                                 }
 981                                 else
 982                                 {
 983                                         aggstrategy = AGG_PLAIN;
 984                                         /* Result will be only one row anyway; no sort order */
 985                                         current_pathkeys = NIL;
 986                                 }
 987
 988                                 result_plan = (Plan *) make_agg(root,
 989                                                                                                 tlist,
 990                                                                                                 (List *) parse->havingQual,
 991                                                                                                 aggstrategy,
 992                                                                                                 numGroupCols,
 993                                                                                                 groupColIdx,
 994                                                                                                 groupOperators,
 995                                                                                                 numGroups,
 996                                                                                                 agg_counts.numAggs,
 997                                                                                                 result_plan);
 998                         }
 999                         else if (parse->groupClause)
1000                         {
1001                                 /*
1002                                  * GROUP BY without aggregation, so insert a group node (plus
1003                                  * the appropriate sort node, if necessary).
1004                                  *
1005                                  * Add an explicit sort if we couldn't make the path come out
1006                                  * the way the GROUP node needs it.
1007                                  */
1008                                 if (!pathkeys_contained_in(group_pathkeys, current_pathkeys))
1009                                 {
1010                                         result_plan = (Plan *)
1011                                                 make_sort_from_groupcols(root,
1012                                                                                                  parse->groupClause,
1013                                                                                                  groupColIdx,
1014                                                                                                  result_plan);
1015                                         current_pathkeys = group_pathkeys;
1016                                 }
1017
1018                                 result_plan = (Plan *) make_group(root,
1019                                                                                                   tlist,
1020                                                                                                   (List *) parse->havingQual,
1021                                                                                                   numGroupCols,
1022                                                                                                   groupColIdx,
1023                                                                                                   groupOperators,
1024                                                                                                   dNumGroups,
1025                                                                                                   result_plan);
1026                                 /* The Group node won't change sort ordering */
1027                         }
1028                         else if (root->hasHavingQual)
1029                         {
1030                                 /*
1031                                  * No aggregates, and no GROUP BY, but we have a HAVING qual.
1032                                  * This is a degenerate case in which we are supposed to emit
1033                                  * either 0 or 1 row depending on whether HAVING succeeds.
1034                                  * Furthermore, there cannot be any variables in either HAVING
1035                                  * or the targetlist, so we actually do not need the FROM
1036                                  * table at all!  We can just throw away the plan-so-far and
1037                                  * generate a Result node.      This is a sufficiently unusual
1038                                  * corner case that it's not worth contorting the structure of
1039                                  * this routine to avoid having to generate the plan in the
1040                                  * first place.
1041                                  */
1042                                 result_plan = (Plan *) make_result(tlist,
1043                                                                                                    parse->havingQual,
1044                                                                                                    NULL);
1045                         }
1046                 }                                               /* end of non-minmax-aggregate case */
1047         }                                                       /* end of if (setOperations) */
1048
1049         /*
1050          * If we were not able to make the plan come out in the right order, add
1051          * an explicit sort step.
1052          */
1053         if (parse->sortClause)
1054         {
1055                 if (!pathkeys_contained_in(sort_pathkeys, current_pathkeys))
1056                 {
1057                         result_plan = (Plan *) make_sort_from_pathkeys(root,
1058                                                                                                                    result_plan,
1059                                                                                                                    sort_pathkeys);
1060                         current_pathkeys = sort_pathkeys;
1061                 }
1062         }
1063
1064         /*
1065          * If there is a DISTINCT clause, add the UNIQUE node.
1066          */
1067         if (parse->distinctClause)
1068         {
1069                 result_plan = (Plan *) make_unique(result_plan, parse->distinctClause);
1070
1071                 /*
1072                  * If there was grouping or aggregation, leave plan_rows as-is (ie,
1073                  * assume the result was already mostly unique).  If not, use the
1074                  * number of distinct-groups calculated by query_planner.
1075                  */
1076                 if (!parse->groupClause && !root->hasHavingQual && !parse->hasAggs)
1077                         result_plan->plan_rows = dNumGroups;
1078         }
1079
1080         /*
1081          * Finally, if there is a LIMIT/OFFSET clause, add the LIMIT node.
1082          */
1083         if (parse->limitCount || parse->limitOffset)
1084         {
1085                 result_plan = (Plan *) make_limit(result_plan,
1086                                                                                   parse->limitOffset,
1087                                                                                   parse->limitCount,
1088                                                                                   offset_est,
1089                                                                                   count_est);
1090         }
1091
1092         /*
1093          * Deal with the RETURNING clause if any.  It's convenient to pass the
1094          * returningList through setrefs.c now rather than at top level (if we
1095          * waited, handling inherited UPDATE/DELETE would be much harder).
1096          */
1097         if (parse->returningList)
1098         {
1099                 List       *rlist;
1100
1101                 rlist = set_returning_clause_references(parse->returningList,
1102                                                                                                 result_plan,
1103                                                                                                 parse->resultRelation);
1104                 parse->returningLists = list_make1(rlist);
1105         }
1106
1107         /*
1108          * Return the actual output ordering in query_pathkeys for possible use by
1109          * an outer query level.
1110          */
1111         root->query_pathkeys = current_pathkeys;
1112
1113         return result_plan;
1114 }
1115
1116 /*
1117  * Detect whether a plan node is a "dummy" plan created when a relation
1118  * is deemed not to need scanning due to constraint exclusion.
1119  *
1120  * Currently, such dummy plans are Result nodes with constant FALSE
1121  * filter quals.
1122  */
1123 static bool
1124 is_dummy_plan(Plan *plan)
1125 {
1126         if (IsA(plan, Result))
1127         {
1128                 List       *rcqual = (List *) ((Result *) plan)->resconstantqual;
1129
1130                 if (list_length(rcqual) == 1)
1131                 {
1132                         Const      *constqual = (Const *) linitial(rcqual);
1133
1134                         if (constqual && IsA(constqual, Const))
1135                         {
1136                                 if (!constqual->constisnull &&
1137                                         !DatumGetBool(constqual->constvalue))
1138                                         return true;
1139                         }
1140                 }
1141         }
1142         return false;
1143 }
1144
1145 /*
1146  * preprocess_limit - do pre-estimation for LIMIT and/or OFFSET clauses
1147  *
1148  * We try to estimate the values of the LIMIT/OFFSET clauses, and pass the
1149  * results back in *count_est and *offset_est.  These variables are set to
1150  * 0 if the corresponding clause is not present, and -1 if it's present
1151  * but we couldn't estimate the value for it.  (The "0" convention is OK
1152  * for OFFSET but a little bit bogus for LIMIT: effectively we estimate
1153  * LIMIT 0 as though it were LIMIT 1.  But this is in line with the planner's
1154  * usual practice of never estimating less than one row.)  These values will
1155  * be passed to make_limit, which see if you change this code.
1156  *
1157  * The return value is the suitably adjusted tuple_fraction to use for
1158  * planning the query.  This adjustment is not overridable, since it reflects
1159  * plan actions that grouping_planner() will certainly take, not assumptions
1160  * about context.
1161  */
1162 static double
1163 preprocess_limit(PlannerInfo *root, double tuple_fraction,
1164                                  int64 *offset_est, int64 *count_est)
1165 {
1166         Query      *parse = root->parse;
1167         Node       *est;
1168         double          limit_fraction;
1169
1170         /* Should not be called unless LIMIT or OFFSET */
1171         Assert(parse->limitCount || parse->limitOffset);
1172
1173         /*
1174          * Try to obtain the clause values.  We use estimate_expression_value
1175          * primarily because it can sometimes do something useful with Params.
1176          */
1177         if (parse->limitCount)
1178         {
1179                 est = estimate_expression_value(root, parse->limitCount);
1180                 if (est && IsA(est, Const))
1181                 {
1182                         if (((Const *) est)->constisnull)
1183                         {
1184                                 /* NULL indicates LIMIT ALL, ie, no limit */
1185                                 *count_est = 0; /* treat as not present */
1186                         }
1187                         else
1188                         {
1189                                 *count_est = DatumGetInt64(((Const *) est)->constvalue);
1190                                 if (*count_est <= 0)
1191                                         *count_est = 1;         /* force to at least 1 */
1192                         }
1193                 }
1194                 else
1195                         *count_est = -1;        /* can't estimate */
1196         }
1197         else
1198                 *count_est = 0;                 /* not present */
1199
1200         if (parse->limitOffset)
1201         {
1202                 est = estimate_expression_value(root, parse->limitOffset);
1203                 if (est && IsA(est, Const))
1204                 {
1205                         if (((Const *) est)->constisnull)
1206                         {
1207                                 /* Treat NULL as no offset; the executor will too */
1208                                 *offset_est = 0;        /* treat as not present */
1209                         }
1210                         else
1211                         {
1212                                 *offset_est = DatumGetInt64(((Const *) est)->constvalue);
1213                                 if (*offset_est < 0)
1214                                         *offset_est = 0;        /* less than 0 is same as 0 */
1215                         }
1216                 }
1217                 else
1218                         *offset_est = -1;       /* can't estimate */
1219         }
1220         else
1221                 *offset_est = 0;                /* not present */
1222
1223         if (*count_est != 0)
1224         {
1225                 /*
1226                  * A LIMIT clause limits the absolute number of tuples returned.
1227                  * However, if it's not a constant LIMIT then we have to guess; for
1228                  * lack of a better idea, assume 10% of the plan's result is wanted.
1229                  */
1230                 if (*count_est < 0 || *offset_est < 0)
1231                 {
1232                         /* LIMIT or OFFSET is an expression ... punt ... */
1233                         limit_fraction = 0.10;
1234                 }
1235                 else
1236                 {
1237                         /* LIMIT (plus OFFSET, if any) is max number of tuples needed */
1238                         limit_fraction = (double) *count_est + (double) *offset_est;
1239                 }
1240
1241                 /*
1242                  * If we have absolute limits from both caller and LIMIT, use the
1243                  * smaller value; likewise if they are both fractional.  If one is
1244                  * fractional and the other absolute, we can't easily determine which
1245                  * is smaller, but we use the heuristic that the absolute will usually
1246                  * be smaller.
1247                  */
1248                 if (tuple_fraction >= 1.0)
1249                 {
1250                         if (limit_fraction >= 1.0)
1251                         {
1252                                 /* both absolute */
1253                                 tuple_fraction = Min(tuple_fraction, limit_fraction);
1254                         }
1255                         else
1256                         {
1257                                 /* caller absolute, limit fractional; use caller's value */
1258                         }
1259                 }
1260                 else if (tuple_fraction > 0.0)
1261                 {
1262                         if (limit_fraction >= 1.0)
1263                         {
1264                                 /* caller fractional, limit absolute; use limit */
1265                                 tuple_fraction = limit_fraction;
1266                         }
1267                         else
1268                         {
1269                                 /* both fractional */
1270                                 tuple_fraction = Min(tuple_fraction, limit_fraction);
1271                         }
1272                 }
1273                 else
1274                 {
1275                         /* no info from caller, just use limit */
1276                         tuple_fraction = limit_fraction;
1277                 }
1278         }
1279         else if (*offset_est != 0 && tuple_fraction > 0.0)
1280         {
1281                 /*
1282                  * We have an OFFSET but no LIMIT.      This acts entirely differently
1283                  * from the LIMIT case: here, we need to increase rather than decrease
1284                  * the caller's tuple_fraction, because the OFFSET acts to cause more
1285                  * tuples to be fetched instead of fewer.  This only matters if we got
1286                  * a tuple_fraction > 0, however.
1287                  *
1288                  * As above, use 10% if OFFSET is present but unestimatable.
1289                  */
1290                 if (*offset_est < 0)
1291                         limit_fraction = 0.10;
1292                 else
1293                         limit_fraction = (double) *offset_est;
1294
1295                 /*
1296                  * If we have absolute counts from both caller and OFFSET, add them
1297                  * together; likewise if they are both fractional.      If one is
1298                  * fractional and the other absolute, we want to take the larger, and
1299                  * we heuristically assume that's the fractional one.
1300                  */
1301                 if (tuple_fraction >= 1.0)
1302                 {
1303                         if (limit_fraction >= 1.0)
1304                         {
1305                                 /* both absolute, so add them together */
1306                                 tuple_fraction += limit_fraction;
1307                         }
1308                         else
1309                         {
1310                                 /* caller absolute, limit fractional; use limit */
1311                                 tuple_fraction = limit_fraction;
1312                         }
1313                 }
1314                 else
1315                 {
1316                         if (limit_fraction >= 1.0)
1317                         {
1318                                 /* caller fractional, limit absolute; use caller's value */
1319                         }
1320                         else
1321                         {
1322                                 /* both fractional, so add them together */
1323                                 tuple_fraction += limit_fraction;
1324                                 if (tuple_fraction >= 1.0)
1325                                         tuple_fraction = 0.0;           /* assume fetch all */
1326                         }
1327                 }
1328         }
1329
1330         return tuple_fraction;
1331 }
1332
1333 /*
1334  * extract_grouping_ops - make an array of the equality operator OIDs
1335  *              for the GROUP BY clause
1336  */
1337 static Oid *
1338 extract_grouping_ops(List *groupClause)
1339 {
1340         int                     numCols = list_length(groupClause);
1341         int                     colno = 0;
1342         Oid                *groupOperators;
1343         ListCell   *glitem;
1344
1345         groupOperators = (Oid *) palloc(sizeof(Oid) * numCols);
1346
1347         foreach(glitem, groupClause)
1348         {
1349                 GroupClause *groupcl = (GroupClause *) lfirst(glitem);
1350
1351                 groupOperators[colno] = get_equality_op_for_ordering_op(groupcl->sortop);
1352                 if (!OidIsValid(groupOperators[colno]))         /* shouldn't happen */
1353                         elog(ERROR, "could not find equality operator for ordering operator %u",
1354                                  groupcl->sortop);
1355                 colno++;
1356         }
1357
1358         return groupOperators;
1359 }
1360
1361 /*
1362  * choose_hashed_grouping - should we use hashed grouping?
1363  */
1364 static bool
1365 choose_hashed_grouping(PlannerInfo *root, double tuple_fraction,
1366                                            Path *cheapest_path, Path *sorted_path,
1367                                            Oid *groupOperators, double dNumGroups,
1368                                            AggClauseCounts *agg_counts)
1369 {
1370         int                     numGroupCols = list_length(root->parse->groupClause);
1371         double          cheapest_path_rows;
1372         int                     cheapest_path_width;
1373         Size            hashentrysize;
1374         List       *current_pathkeys;
1375         Path            hashed_p;
1376         Path            sorted_p;
1377         int                     i;
1378
1379         /*
1380          * Check can't-do-it conditions, including whether the grouping operators
1381          * are hashjoinable.  (We assume hashing is OK if they are marked
1382          * oprcanhash.  If there isn't actually a supporting hash function,
1383          * the executor will complain at runtime.)
1384          *
1385          * Executor doesn't support hashed aggregation with DISTINCT aggregates.
1386          * (Doing so would imply storing *all* the input values in the hash table,
1387          * which seems like a certain loser.)
1388          */
1389         if (!enable_hashagg)
1390                 return false;
1391         if (agg_counts->numDistinctAggs != 0)
1392                 return false;
1393         for (i = 0; i < numGroupCols; i++)
1394         {
1395                 if (!op_hashjoinable(groupOperators[i]))
1396                         return false;
1397         }
1398
1399         /*
1400          * Don't do it if it doesn't look like the hashtable will fit into
1401          * work_mem.
1402          *
1403          * Beware here of the possibility that cheapest_path->parent is NULL. This
1404          * could happen if user does something silly like SELECT 'foo' GROUP BY 1;
1405          */
1406         if (cheapest_path->parent)
1407         {
1408                 cheapest_path_rows = cheapest_path->parent->rows;
1409                 cheapest_path_width = cheapest_path->parent->width;
1410         }
1411         else
1412         {
1413                 cheapest_path_rows = 1; /* assume non-set result */
1414                 cheapest_path_width = 100;              /* arbitrary */
1415         }
1416
1417         /* Estimate per-hash-entry space at tuple width... */
1418         hashentrysize = MAXALIGN(cheapest_path_width) + MAXALIGN(sizeof(MinimalTupleData));
1419         /* plus space for pass-by-ref transition values... */
1420         hashentrysize += agg_counts->transitionSpace;
1421         /* plus the per-hash-entry overhead */
1422         hashentrysize += hash_agg_entry_size(agg_counts->numAggs);
1423
1424         if (hashentrysize * dNumGroups > work_mem * 1024L)
1425                 return false;
1426
1427         /*
1428          * See if the estimated cost is no more than doing it the other way. While
1429          * avoiding the need for sorted input is usually a win, the fact that the
1430          * output won't be sorted may be a loss; so we need to do an actual cost
1431          * comparison.
1432          *
1433          * We need to consider cheapest_path + hashagg [+ final sort] versus
1434          * either cheapest_path [+ sort] + group or agg [+ final sort] or
1435          * presorted_path + group or agg [+ final sort] where brackets indicate a
1436          * step that may not be needed. We assume query_planner() will have
1437          * returned a presorted path only if it's a winner compared to
1438          * cheapest_path for this purpose.
1439          *
1440          * These path variables are dummies that just hold cost fields; we don't
1441          * make actual Paths for these steps.
1442          */
1443         cost_agg(&hashed_p, root, AGG_HASHED, agg_counts->numAggs,
1444                          numGroupCols, dNumGroups,
1445                          cheapest_path->startup_cost, cheapest_path->total_cost,
1446                          cheapest_path_rows);
1447         /* Result of hashed agg is always unsorted */
1448         if (root->sort_pathkeys)
1449                 cost_sort(&hashed_p, root, root->sort_pathkeys, hashed_p.total_cost,
1450                                   dNumGroups, cheapest_path_width);
1451
1452         if (sorted_path)
1453         {
1454                 sorted_p.startup_cost = sorted_path->startup_cost;
1455                 sorted_p.total_cost = sorted_path->total_cost;
1456                 current_pathkeys = sorted_path->pathkeys;
1457         }
1458         else
1459         {
1460                 sorted_p.startup_cost = cheapest_path->startup_cost;
1461                 sorted_p.total_cost = cheapest_path->total_cost;
1462                 current_pathkeys = cheapest_path->pathkeys;
1463         }
1464         if (!pathkeys_contained_in(root->group_pathkeys, current_pathkeys))
1465         {
1466                 cost_sort(&sorted_p, root, root->group_pathkeys, sorted_p.total_cost,
1467                                   cheapest_path_rows, cheapest_path_width);
1468                 current_pathkeys = root->group_pathkeys;
1469         }
1470
1471         if (root->parse->hasAggs)
1472                 cost_agg(&sorted_p, root, AGG_SORTED, agg_counts->numAggs,
1473                                  numGroupCols, dNumGroups,
1474                                  sorted_p.startup_cost, sorted_p.total_cost,
1475                                  cheapest_path_rows);
1476         else
1477                 cost_group(&sorted_p, root, numGroupCols, dNumGroups,
1478                                    sorted_p.startup_cost, sorted_p.total_cost,
1479                                    cheapest_path_rows);
1480         /* The Agg or Group node will preserve ordering */
1481         if (root->sort_pathkeys &&
1482                 !pathkeys_contained_in(root->sort_pathkeys, current_pathkeys))
1483                 cost_sort(&sorted_p, root, root->sort_pathkeys, sorted_p.total_cost,
1484                                   dNumGroups, cheapest_path_width);
1485
1486         /*
1487          * Now make the decision using the top-level tuple fraction.  First we
1488          * have to convert an absolute count (LIMIT) into fractional form.
1489          */
1490         if (tuple_fraction >= 1.0)
1491                 tuple_fraction /= dNumGroups;
1492
1493         if (compare_fractional_path_costs(&hashed_p, &sorted_p,
1494                                                                           tuple_fraction) < 0)
1495         {
1496                 /* Hashed is cheaper, so use it */
1497                 return true;
1498         }
1499         return false;
1500 }
1501
1502 /*---------------
1503  * make_subplanTargetList
1504  *        Generate appropriate target list when grouping is required.
1505  *
1506  * When grouping_planner inserts Aggregate, Group, or Result plan nodes
1507  * above the result of query_planner, we typically want to pass a different
1508  * target list to query_planner than the outer plan nodes should have.
1509  * This routine generates the correct target list for the subplan.
1510  *
1511  * The initial target list passed from the parser already contains entries
1512  * for all ORDER BY and GROUP BY expressions, but it will not have entries
1513  * for variables used only in HAVING clauses; so we need to add those
1514  * variables to the subplan target list.  Also, we flatten all expressions
1515  * except GROUP BY items into their component variables; the other expressions
1516  * will be computed by the inserted nodes rather than by the subplan.
1517  * For example, given a query like
1518  *              SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
1519  * we want to pass this targetlist to the subplan:
1520  *              a,b,c,d,a+b
1521  * where the a+b target will be used by the Sort/Group steps, and the
1522  * other targets will be used for computing the final results.  (In the
1523  * above example we could theoretically suppress the a and b targets and
1524  * pass down only c,d,a+b, but it's not really worth the trouble to
1525  * eliminate simple var references from the subplan.  We will avoid doing
1526  * the extra computation to recompute a+b at the outer level; see
1527  * replace_vars_with_subplan_refs() in setrefs.c.)
1528  *
1529  * If we are grouping or aggregating, *and* there are no non-Var grouping
1530  * expressions, then the returned tlist is effectively dummy; we do not
1531  * need to force it to be evaluated, because all the Vars it contains
1532  * should be present in the output of query_planner anyway.
1533  *
1534  * 'tlist' is the query's target list.
1535  * 'groupColIdx' receives an array of column numbers for the GROUP BY
1536  *                      expressions (if there are any) in the subplan's target list.
1537  * 'need_tlist_eval' is set true if we really need to evaluate the
1538  *                      result tlist.
1539  *
1540  * The result is the targetlist to be passed to the subplan.
1541  *---------------
1542  */
1543 static List *
1544 make_subplanTargetList(PlannerInfo *root,
1545                                            List *tlist,
1546                                            AttrNumber **groupColIdx,
1547                                            bool *need_tlist_eval)
1548 {
1549         Query      *parse = root->parse;
1550         List       *sub_tlist;
1551         List       *extravars;
1552         int                     numCols;
1553
1554         *groupColIdx = NULL;
1555
1556         /*
1557          * If we're not grouping or aggregating, there's nothing to do here;
1558          * query_planner should receive the unmodified target list.
1559          */
1560         if (!parse->hasAggs && !parse->groupClause && !root->hasHavingQual)
1561         {
1562                 *need_tlist_eval = true;
1563                 return tlist;
1564         }
1565
1566         /*
1567          * Otherwise, start with a "flattened" tlist (having just the vars
1568          * mentioned in the targetlist and HAVING qual --- but not upper- level
1569          * Vars; they will be replaced by Params later on).
1570          */
1571         sub_tlist = flatten_tlist(tlist);
1572         extravars = pull_var_clause(parse->havingQual, false);
1573         sub_tlist = add_to_flat_tlist(sub_tlist, extravars);
1574         list_free(extravars);
1575         *need_tlist_eval = false;       /* only eval if not flat tlist */
1576
1577         /*
1578          * If grouping, create sub_tlist entries for all GROUP BY expressions
1579          * (GROUP BY items that are simple Vars should be in the list already),
1580          * and make an array showing where the group columns are in the sub_tlist.
1581          */
1582         numCols = list_length(parse->groupClause);
1583         if (numCols > 0)
1584         {
1585                 int                     keyno = 0;
1586                 AttrNumber *grpColIdx;
1587                 ListCell   *gl;
1588
1589                 grpColIdx = (AttrNumber *) palloc(sizeof(AttrNumber) * numCols);
1590                 *groupColIdx = grpColIdx;
1591
1592                 foreach(gl, parse->groupClause)
1593                 {
1594                         GroupClause *grpcl = (GroupClause *) lfirst(gl);
1595                         Node       *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1596                         TargetEntry *te = NULL;
1597                         ListCell   *sl;
1598
1599                         /* Find or make a matching sub_tlist entry */
1600                         foreach(sl, sub_tlist)
1601                         {
1602                                 te = (TargetEntry *) lfirst(sl);
1603                                 if (equal(groupexpr, te->expr))
1604                                         break;
1605                         }
1606                         if (!sl)
1607                         {
1608                                 te = makeTargetEntry((Expr *) groupexpr,
1609                                                                          list_length(sub_tlist) + 1,
1610                                                                          NULL,
1611                                                                          false);
1612                                 sub_tlist = lappend(sub_tlist, te);
1613                                 *need_tlist_eval = true;                /* it's not flat anymore */
1614                         }
1615
1616                         /* and save its resno */
1617                         grpColIdx[keyno++] = te->resno;
1618                 }
1619         }
1620
1621         return sub_tlist;
1622 }
1623
1624 /*
1625  * locate_grouping_columns
1626  *              Locate grouping columns in the tlist chosen by query_planner.
1627  *
1628  * This is only needed if we don't use the sub_tlist chosen by
1629  * make_subplanTargetList.      We have to forget the column indexes found
1630  * by that routine and re-locate the grouping vars in the real sub_tlist.
1631  */
1632 static void
1633 locate_grouping_columns(PlannerInfo *root,
1634                                                 List *tlist,
1635                                                 List *sub_tlist,
1636                                                 AttrNumber *groupColIdx)
1637 {
1638         int                     keyno = 0;
1639         ListCell   *gl;
1640
1641         /*
1642          * No work unless grouping.
1643          */
1644         if (!root->parse->groupClause)
1645         {
1646                 Assert(groupColIdx == NULL);
1647                 return;
1648         }
1649         Assert(groupColIdx != NULL);
1650
1651         foreach(gl, root->parse->groupClause)
1652         {
1653                 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1654                 Node       *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1655                 TargetEntry *te = NULL;
1656                 ListCell   *sl;
1657
1658                 foreach(sl, sub_tlist)
1659                 {
1660                         te = (TargetEntry *) lfirst(sl);
1661                         if (equal(groupexpr, te->expr))
1662                                 break;
1663                 }
1664                 if (!sl)
1665                         elog(ERROR, "failed to locate grouping columns");
1666
1667                 groupColIdx[keyno++] = te->resno;
1668         }
1669 }
1670
1671 /*
1672  * postprocess_setop_tlist
1673  *        Fix up targetlist returned by plan_set_operations().
1674  *
1675  * We need to transpose sort key info from the orig_tlist into new_tlist.
1676  * NOTE: this would not be good enough if we supported resjunk sort keys
1677  * for results of set operations --- then, we'd need to project a whole
1678  * new tlist to evaluate the resjunk columns.  For now, just ereport if we
1679  * find any resjunk columns in orig_tlist.
1680  */
1681 static List *
1682 postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
1683 {
1684         ListCell   *l;
1685         ListCell   *orig_tlist_item = list_head(orig_tlist);
1686
1687         foreach(l, new_tlist)
1688         {
1689                 TargetEntry *new_tle = (TargetEntry *) lfirst(l);
1690                 TargetEntry *orig_tle;
1691
1692                 /* ignore resjunk columns in setop result */
1693                 if (new_tle->resjunk)
1694                         continue;
1695
1696                 Assert(orig_tlist_item != NULL);
1697                 orig_tle = (TargetEntry *) lfirst(orig_tlist_item);
1698                 orig_tlist_item = lnext(orig_tlist_item);
1699                 if (orig_tle->resjunk)  /* should not happen */
1700                         elog(ERROR, "resjunk output columns are not implemented");
1701                 Assert(new_tle->resno == orig_tle->resno);
1702                 new_tle->ressortgroupref = orig_tle->ressortgroupref;
1703         }
1704         if (orig_tlist_item != NULL)
1705                 elog(ERROR, "resjunk output columns are not implemented");
1706         return new_tlist;
1707 }