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 1.211 2007/01/10 18:06:03 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 ParamListInfo PlannerBoundParamList = NULL;             /* current boundParams */
  46
  47
  48 /* Expression kind codes for preprocess_expression */
  49 #define EXPRKIND_QUAL           0
  50 #define EXPRKIND_TARGET         1
  51 #define EXPRKIND_RTFUNC         2
  52 #define EXPRKIND_VALUES         3
  53 #define EXPRKIND_LIMIT          4
  54 #define EXPRKIND_ININFO         5
  55 #define EXPRKIND_APPINFO        6
  56
  57
  58 static Node *preprocess_expression(PlannerInfo *root, Node *expr, int kind);
  59 static void preprocess_qual_conditions(PlannerInfo *root, Node *jtnode);
  60 static Plan *inheritance_planner(PlannerInfo *root);
  61 static Plan *grouping_planner(PlannerInfo *root, double tuple_fraction);
  62 static bool is_dummy_plan(Plan *plan);
  63 static double preprocess_limit(PlannerInfo *root,
  64                                  double tuple_fraction,
  65                                  int64 *offset_est, int64 *count_est);
  66 static Oid *extract_grouping_ops(List *groupClause);
  67 static bool choose_hashed_grouping(PlannerInfo *root, double tuple_fraction,
  68                                            Path *cheapest_path, Path *sorted_path,
  69                                            Oid *groupOperators, double dNumGroups,
  70                                            AggClauseCounts *agg_counts);
  71 static List *make_subplanTargetList(PlannerInfo *root, List *tlist,
  72                                            AttrNumber **groupColIdx, bool *need_tlist_eval);
  73 static void locate_grouping_columns(PlannerInfo *root,
  74                                                 List *tlist,
  75                                                 List *sub_tlist,
  76                                                 AttrNumber *groupColIdx);
  77 static List *postprocess_setop_tlist(List *new_tlist, List *orig_tlist);
  78
  79
  80 /*****************************************************************************
  81  *
  82  *         Query optimizer entry point
  83  *
  84  *****************************************************************************/
  85 Plan *
  86 planner(Query *parse, bool isCursor, int cursorOptions,
  87                 ParamListInfo boundParams)
  88 {
  89         double          tuple_fraction;
  90         Plan       *result_plan;
  91         Index           save_PlannerQueryLevel;
  92         List       *save_PlannerParamList;
  93         ParamListInfo save_PlannerBoundParamList;
  94
  95         /*
  96          * The planner can be called recursively (an example is when
  97          * eval_const_expressions tries to pre-evaluate an SQL function). So,
  98          * these global state variables must be saved and restored.
  99          *
 100          * Query level and the param list cannot be moved into the per-query
 101          * PlannerInfo structure since their whole purpose is communication across
 102          * multiple sub-queries. Also, boundParams is explicitly info from outside
 103          * the query, and so is likewise better handled as a global variable.
 104          *
 105          * Note we do NOT save and restore PlannerPlanId: it exists to assign
 106          * unique IDs to SubPlan nodes, and we want those IDs to be unique for the
 107          * life of a backend.  Also, PlannerInitPlan is saved/restored in
 108          * subquery_planner, not here.
 109          */
 110         save_PlannerQueryLevel = PlannerQueryLevel;
 111         save_PlannerParamList = PlannerParamList;
 112         save_PlannerBoundParamList = PlannerBoundParamList;
 113
 114         /* Initialize state for handling outer-level references and params */
 115         PlannerQueryLevel = 0;          /* will be 1 in top-level subquery_planner */
 116         PlannerParamList = NIL;
 117         PlannerBoundParamList = boundParams;
 118
 119         /* Determine what fraction of the plan is likely to be scanned */
 120         if (isCursor)
 121         {
 122                 /*
 123                  * We have no real idea how many tuples the user will ultimately FETCH
 124                  * from a cursor, but it seems a good bet that he doesn't want 'em
 125                  * all.  Optimize for 10% retrieval (you gotta better number?  Should
 126                  * this be a SETtable parameter?)
 127                  */
 128                 tuple_fraction = 0.10;
 129         }
 130         else
 131         {
 132                 /* Default assumption is we need all the tuples */
 133                 tuple_fraction = 0.0;
 134         }
 135
 136         /* primary planning entry point (may recurse for subqueries) */
 137         result_plan = subquery_planner(parse, tuple_fraction, NULL);
 138
 139         /* check we popped out the right number of levels */
 140         Assert(PlannerQueryLevel == 0);
 141
 142         /*
 143          * If creating a plan for a scrollable cursor, make sure it can run
 144          * backwards on demand.  Add a Material node at the top at need.
 145          */
 146         if (isCursor && (cursorOptions & CURSOR_OPT_SCROLL))
 147         {
 148                 if (!ExecSupportsBackwardScan(result_plan))
 149                         result_plan = materialize_finished_plan(result_plan);
 150         }
 151
 152         /* final cleanup of the plan */
 153         result_plan = set_plan_references(result_plan, parse->rtable);
 154
 155         /* executor wants to know total number of Params used overall */
 156         result_plan->nParamExec = list_length(PlannerParamList);
 157
 158         /* restore state for outer planner, if any */
 159         PlannerQueryLevel = save_PlannerQueryLevel;
 160         PlannerParamList = save_PlannerParamList;
 161         PlannerBoundParamList = save_PlannerBoundParamList;
 162
 163         return result_plan;
 164 }
 165
 166
 167 /*--------------------
 168  * subquery_planner
 169  *        Invokes the planner on a subquery.  We recurse to here for each
 170  *        sub-SELECT found in the query tree.
 171  *
 172  * parse is the querytree produced by the parser & rewriter.
 173  * tuple_fraction is the fraction of tuples we expect will be retrieved.
 174  * tuple_fraction is interpreted as explained for grouping_planner, below.
 175  *
 176  * If subquery_pathkeys isn't NULL, it receives a list of pathkeys indicating
 177  * the output sort ordering of the completed plan.
 178  *
 179  * Basically, this routine does the stuff that should only be done once
 180  * per Query object.  It then calls grouping_planner.  At one time,
 181  * grouping_planner could be invoked recursively on the same Query object;
 182  * that's not currently true, but we keep the separation between the two
 183  * routines anyway, in case we need it again someday.
 184  *
 185  * subquery_planner will be called recursively to handle sub-Query nodes
 186  * found within the query's expressions and rangetable.
 187  *
 188  * Returns a query plan.
 189  *--------------------
 190  */
 191 Plan *
 192 subquery_planner(Query *parse, double tuple_fraction,
 193                                  List **subquery_pathkeys)
 194 {
 195         List       *saved_initplan = PlannerInitPlan;
 196         int                     saved_planid = PlannerPlanId;
 197         PlannerInfo *root;
 198         Plan       *plan;
 199         List       *newHaving;
 200         ListCell   *l;
 201
 202         /* Set up for a new level of subquery */
 203         PlannerQueryLevel++;
 204         PlannerInitPlan = NIL;
 205
 206         /* Create a PlannerInfo data structure for this subquery */
 207         root = makeNode(PlannerInfo);
 208         root->parse = parse;
 209         root->in_info_list = NIL;
 210         root->append_rel_list = NIL;
 211
 212         /*
 213          * Look for IN clauses at the top level of WHERE, and transform them into
 214          * joins.  Note that this step only handles IN clauses originally at top
 215          * level of WHERE; if we pull up any subqueries in the next step, their
 216          * INs are processed just before pulling them up.
 217          */
 218         if (parse->hasSubLinks)
 219                 parse->jointree->quals = pull_up_IN_clauses(root,
 220                                                                                                         parse->jointree->quals);
 221
 222         /*
 223          * Check to see if any subqueries in the rangetable can be merged into
 224          * this query.
 225          */
 226         parse->jointree = (FromExpr *)
 227                 pull_up_subqueries(root, (Node *) parse->jointree, false, false);
 228
 229         /*
 230          * Detect whether any rangetable entries are RTE_JOIN kind; if not, we can
 231          * avoid the expense of doing flatten_join_alias_vars().  Also check for
 232          * outer joins --- if none, we can skip reduce_outer_joins() and some
 233          * other processing.  This must be done after we have done
 234          * pull_up_subqueries, of course.
 235          *
 236          * Note: if reduce_outer_joins manages to eliminate all outer joins,
 237          * root->hasOuterJoins is not reset currently.  This is OK since its
 238          * purpose is merely to suppress unnecessary processing in simple cases.
 239          */
 240         root->hasJoinRTEs = false;
 241         root->hasOuterJoins = false;
 242         foreach(l, parse->rtable)
 243         {
 244                 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
 245
 246                 if (rte->rtekind == RTE_JOIN)
 247                 {
 248                         root->hasJoinRTEs = true;
 249                         if (IS_OUTER_JOIN(rte->jointype))
 250                         {
 251                                 root->hasOuterJoins = true;
 252                                 /* Can quit scanning once we find an outer join */
 253                                 break;
 254                         }
 255                 }
 256         }
 257
 258         /*
 259          * Expand any rangetable entries that are inheritance sets into "append
 260          * relations".  This can add entries to the rangetable, but they must be
 261          * plain base relations not joins, so it's OK (and marginally more
 262          * efficient) to do it after checking for join RTEs.  We must do it after
 263          * pulling up subqueries, else we'd fail to handle inherited tables in
 264          * subqueries.
 265          */
 266         expand_inherited_tables(root);
 267
 268         /*
 269          * Set hasHavingQual to remember if HAVING clause is present.  Needed
 270          * because preprocess_expression will reduce a constant-true condition to
 271          * an empty qual list ... but "HAVING TRUE" is not a semantic no-op.
 272          */
 273         root->hasHavingQual = (parse->havingQual != NULL);
 274
 275         /* Clear this flag; might get set in distribute_qual_to_rels */
 276         root->hasPseudoConstantQuals = false;
 277
 278         /*
 279          * Do expression preprocessing on targetlist and quals.
 280          */
 281         parse->targetList = (List *)
 282                 preprocess_expression(root, (Node *) parse->targetList,
 283                                                           EXPRKIND_TARGET);
 284
 285         parse->returningList = (List *)
 286                 preprocess_expression(root, (Node *) parse->returningList,
 287                                                           EXPRKIND_TARGET);
 288
 289         preprocess_qual_conditions(root, (Node *) parse->jointree);
 290
 291         parse->havingQual = preprocess_expression(root, parse->havingQual,
 292                                                                                           EXPRKIND_QUAL);
 293
 294         parse->limitOffset = preprocess_expression(root, parse->limitOffset,
 295                                                                                            EXPRKIND_LIMIT);
 296         parse->limitCount = preprocess_expression(root, parse->limitCount,
 297                                                                                           EXPRKIND_LIMIT);
 298
 299         root->in_info_list = (List *)
 300                 preprocess_expression(root, (Node *) root->in_info_list,
 301                                                           EXPRKIND_ININFO);
 302         root->append_rel_list = (List *)
 303                 preprocess_expression(root, (Node *) root->append_rel_list,
 304                                                           EXPRKIND_APPINFO);
 305
 306         /* Also need to preprocess expressions for function and values RTEs */
 307         foreach(l, parse->rtable)
 308         {
 309                 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
 310
 311                 if (rte->rtekind == RTE_FUNCTION)
 312                         rte->funcexpr = preprocess_expression(root, rte->funcexpr,
 313                                                                                                   EXPRKIND_RTFUNC);
 314                 else if (rte->rtekind == RTE_VALUES)
 315                         rte->values_lists = (List *)
 316                                 preprocess_expression(root, (Node *) rte->values_lists,
 317                                                                           EXPRKIND_VALUES);
 318         }
 319
 320         /*
 321          * In some cases we may want to transfer a HAVING clause into WHERE. We
 322          * cannot do so if the HAVING clause contains aggregates (obviously) or
 323          * volatile functions (since a HAVING clause is supposed to be executed
 324          * only once per group).  Also, it may be that the clause is so expensive
 325          * to execute that we're better off doing it only once per group, despite
 326          * the loss of selectivity.  This is hard to estimate short of doing the
 327          * entire planning process twice, so we use a heuristic: clauses
 328          * containing subplans are left in HAVING.      Otherwise, we move or copy the
 329          * HAVING clause into WHERE, in hopes of eliminating tuples before
 330          * aggregation instead of after.
 331          *
 332          * If the query has explicit grouping then we can simply move such a
 333          * clause into WHERE; any group that fails the clause will not be in the
 334          * output because none of its tuples will reach the grouping or
 335          * aggregation stage.  Otherwise we must have a degenerate (variable-free)
 336          * HAVING clause, which we put in WHERE so that query_planner() can use it
 337          * in a gating Result node, but also keep in HAVING to ensure that we
 338          * don't emit a bogus aggregated row. (This could be done better, but it
 339          * seems not worth optimizing.)
 340          *
 341          * Note that both havingQual and parse->jointree->quals are in
 342          * implicitly-ANDed-list form at this point, even though they are declared
 343          * as Node *.
 344          */
 345         newHaving = NIL;
 346         foreach(l, (List *) parse->havingQual)
 347         {
 348                 Node       *havingclause = (Node *) lfirst(l);
 349
 350                 if (contain_agg_clause(havingclause) ||
 351                         contain_volatile_functions(havingclause) ||
 352                         contain_subplans(havingclause))
 353                 {
 354                         /* keep it in HAVING */
 355                         newHaving = lappend(newHaving, havingclause);
 356                 }
 357                 else if (parse->groupClause)
 358                 {
 359                         /* move it to WHERE */
 360                         parse->jointree->quals = (Node *)
 361                                 lappend((List *) parse->jointree->quals, havingclause);
 362                 }
 363                 else
 364                 {
 365                         /* put a copy in WHERE, keep it in HAVING */
 366                         parse->jointree->quals = (Node *)
 367                                 lappend((List *) parse->jointree->quals,
 368                                                 copyObject(havingclause));
 369                         newHaving = lappend(newHaving, havingclause);
 370                 }
 371         }
 372         parse->havingQual = (Node *) newHaving;
 373
 374         /*
 375          * If we have any outer joins, try to reduce them to plain inner joins.
 376          * This step is most easily done after we've done expression
 377          * preprocessing.
 378          */
 379         if (root->hasOuterJoins)
 380                 reduce_outer_joins(root);
 381
 382         /*
 383          * Do the main planning.  If we have an inherited target relation, that
 384          * needs special processing, else go straight to grouping_planner.
 385          */
 386         if (parse->resultRelation &&
 387                 rt_fetch(parse->resultRelation, parse->rtable)->inh)
 388                 plan = inheritance_planner(root);
 389         else
 390                 plan = grouping_planner(root, tuple_fraction);
 391
 392         /*
 393          * If any subplans were generated, or if we're inside a subplan, build
 394          * initPlan list and extParam/allParam sets for plan nodes, and attach the
 395          * initPlans to the top plan node.
 396          */
 397         if (PlannerPlanId != saved_planid || PlannerQueryLevel > 1)
 398                 SS_finalize_plan(plan, parse->rtable);
 399
 400         /* Return sort ordering info if caller wants it */
 401         if (subquery_pathkeys)
 402                 *subquery_pathkeys = root->query_pathkeys;
 403
 404         /* Return to outer subquery context */
 405         PlannerQueryLevel--;
 406         PlannerInitPlan = saved_initplan;
 407         /* we do NOT restore PlannerPlanId; that's not an oversight! */
 408
 409         return plan;
 410 }
 411
 412 /*
 413  * preprocess_expression
 414  *              Do subquery_planner's preprocessing work for an expression,
 415  *              which can be a targetlist, a WHERE clause (including JOIN/ON
 416  *              conditions), or a HAVING clause.
 417  */
 418 static Node *
 419 preprocess_expression(PlannerInfo *root, Node *expr, int kind)
 420 {
 421         /*
 422          * Fall out quickly if expression is empty.  This occurs often enough to
 423          * be worth checking.  Note that null->null is the correct conversion for
 424          * implicit-AND result format, too.
 425          */
 426         if (expr == NULL)
 427                 return NULL;
 428
 429         /*
 430          * If the query has any join RTEs, replace join alias variables with
 431          * base-relation variables. We must do this before sublink processing,
 432          * else sublinks expanded out from join aliases wouldn't get processed. We
 433          * can skip it in VALUES lists, however, since they can't contain any Vars
 434          * at all.
 435          */
 436         if (root->hasJoinRTEs && kind != EXPRKIND_VALUES)
 437                 expr = flatten_join_alias_vars(root, expr);
 438
 439         /*
 440          * Simplify constant expressions.
 441          *
 442          * Note: this also flattens nested AND and OR expressions into N-argument
 443          * form.  All processing of a qual expression after this point must be
 444          * careful to maintain AND/OR flatness --- that is, do not generate a tree
 445          * with AND directly under AND, nor OR directly under OR.
 446          *
 447          * Because this is a relatively expensive process, we skip it when the
 448          * query is trivial, such as "SELECT 2+2;" or "INSERT ... VALUES()". The
 449          * expression will only be evaluated once anyway, so no point in
 450          * pre-simplifying; we can't execute it any faster than the executor can,
 451          * and we will waste cycles copying the tree.  Notice however that we
 452          * still must do it for quals (to get AND/OR flatness); and if we are in a
 453          * subquery we should not assume it will be done only once.
 454          *
 455          * For VALUES lists we never do this at all, again on the grounds that we
 456          * should optimize for one-time evaluation.
 457          */
 458         if (kind != EXPRKIND_VALUES &&
 459                 (root->parse->jointree->fromlist != NIL ||
 460                  kind == EXPRKIND_QUAL ||
 461                  PlannerQueryLevel > 1))
 462                 expr = eval_const_expressions(expr);
 463
 464         /*
 465          * If it's a qual or havingQual, canonicalize it.
 466          */
 467         if (kind == EXPRKIND_QUAL)
 468         {
 469                 expr = (Node *) canonicalize_qual((Expr *) expr);
 470
 471 #ifdef OPTIMIZER_DEBUG
 472                 printf("After canonicalize_qual()\n");
 473                 pprint(expr);
 474 #endif
 475         }
 476
 477         /* Expand SubLinks to SubPlans */
 478         if (root->parse->hasSubLinks)
 479                 expr = SS_process_sublinks(expr, (kind == EXPRKIND_QUAL));
 480
 481         /*
 482          * XXX do not insert anything here unless you have grokked the comments in
 483          * SS_replace_correlation_vars ...
 484          */
 485
 486         /* Replace uplevel vars with Param nodes (this IS possible in VALUES) */
 487         if (PlannerQueryLevel > 1)
 488                 expr = SS_replace_correlation_vars(expr);
 489
 490         /*
 491          * If it's a qual or havingQual, convert it to implicit-AND format. (We
 492          * don't want to do this before eval_const_expressions, since the latter
 493          * would be unable to simplify a top-level AND correctly. Also,
 494          * SS_process_sublinks expects explicit-AND format.)
 495          */
 496         if (kind == EXPRKIND_QUAL)
 497                 expr = (Node *) make_ands_implicit((Expr *) expr);
 498
 499         return expr;
 500 }
 501
 502 /*
 503  * preprocess_qual_conditions
 504  *              Recursively scan the query's jointree and do subquery_planner's
 505  *              preprocessing work on each qual condition found therein.
 506  */
 507 static void
 508 preprocess_qual_conditions(PlannerInfo *root, Node *jtnode)
 509 {
 510         if (jtnode == NULL)
 511                 return;
 512         if (IsA(jtnode, RangeTblRef))
 513         {
 514                 /* nothing to do here */
 515         }
 516         else if (IsA(jtnode, FromExpr))
 517         {
 518                 FromExpr   *f = (FromExpr *) jtnode;
 519                 ListCell   *l;
 520
 521                 foreach(l, f->fromlist)
 522                         preprocess_qual_conditions(root, lfirst(l));
 523
 524                 f->quals = preprocess_expression(root, f->quals, EXPRKIND_QUAL);
 525         }
 526         else if (IsA(jtnode, JoinExpr))
 527         {
 528                 JoinExpr   *j = (JoinExpr *) jtnode;
 529
 530                 preprocess_qual_conditions(root, j->larg);
 531                 preprocess_qual_conditions(root, j->rarg);
 532
 533                 j->quals = preprocess_expression(root, j->quals, EXPRKIND_QUAL);
 534         }
 535         else
 536                 elog(ERROR, "unrecognized node type: %d",
 537                          (int) nodeTag(jtnode));
 538 }
 539
 540 /*
 541  * inheritance_planner
 542  *        Generate a plan in the case where the result relation is an
 543  *        inheritance set.
 544  *
 545  * We have to handle this case differently from cases where a source relation
 546  * is an inheritance set. Source inheritance is expanded at the bottom of the
 547  * plan tree (see allpaths.c), but target inheritance has to be expanded at
 548  * the top.  The reason is that for UPDATE, each target relation needs a
 549  * different targetlist matching its own column set.  Also, for both UPDATE
 550  * and DELETE, the executor needs the Append plan node at the top, else it
 551  * can't keep track of which table is the current target table.  Fortunately,
 552  * the UPDATE/DELETE target can never be the nullable side of an outer join,
 553  * so it's OK to generate the plan this way.
 554  *
 555  * Returns a query plan.
 556  */
 557 static Plan *
 558 inheritance_planner(PlannerInfo *root)
 559 {
 560         Query      *parse = root->parse;
 561         int                     parentRTindex = parse->resultRelation;
 562         List       *subplans = NIL;
 563         List       *resultRelations = NIL;
 564         List       *returningLists = NIL;
 565         List       *rtable = NIL;
 566         List       *tlist = NIL;
 567         PlannerInfo subroot;
 568         ListCell   *l;
 569
 570         foreach(l, root->append_rel_list)
 571         {
 572                 AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
 573                 Plan       *subplan;
 574
 575                 /* append_rel_list contains all append rels; ignore others */
 576                 if (appinfo->parent_relid != parentRTindex)
 577                         continue;
 578
 579                 /*
 580                  * Generate modified query with this rel as target.  We have to be
 581                  * prepared to translate varnos in in_info_list as well as in the
 582                  * Query proper.
 583                  */
 584                 memcpy(&subroot, root, sizeof(PlannerInfo));
 585                 subroot.parse = (Query *)
 586                         adjust_appendrel_attrs((Node *) parse,
 587                                                                    appinfo);
 588                 subroot.in_info_list = (List *)
 589                         adjust_appendrel_attrs((Node *) root->in_info_list,
 590                                                                    appinfo);
 591                 /* There shouldn't be any OJ info to translate, as yet */
 592                 Assert(subroot.oj_info_list == NIL);
 593
 594                 /* Generate plan */
 595                 subplan = grouping_planner(&subroot, 0.0 /* retrieve all tuples */ );
 596
 597                 /*
 598                  * If this child rel was excluded by constraint exclusion, exclude it
 599                  * from the plan.
 600                  */
 601                 if (is_dummy_plan(subplan))
 602                         continue;
 603
 604                 /* Save rtable and tlist from first rel for use below */
 605                 if (subplans == NIL)
 606                 {
 607                         rtable = subroot.parse->rtable;
 608                         tlist = subplan->targetlist;
 609                 }
 610
 611                 subplans = lappend(subplans, subplan);
 612
 613                 /* Build target-relations list for the executor */
 614                 resultRelations = lappend_int(resultRelations, appinfo->child_relid);
 615
 616                 /* Build list of per-relation RETURNING targetlists */
 617                 if (parse->returningList)
 618                 {
 619                         Assert(list_length(subroot.parse->returningLists) == 1);
 620                         returningLists = list_concat(returningLists,
 621                                                                                  subroot.parse->returningLists);
 622                 }
 623         }
 624
 625         parse->resultRelations = resultRelations;
 626         parse->returningLists = returningLists;
 627
 628         /* Mark result as unordered (probably unnecessary) */
 629         root->query_pathkeys = NIL;
 630
 631         /*
 632          * If we managed to exclude every child rel, return a dummy plan
 633          */
 634         if (subplans == NIL)
 635                 return (Plan *) make_result(tlist,
 636                                                                         (Node *) list_make1(makeBoolConst(false,
 637                                                                                                                                           false)),
 638                                                                         NULL);
 639
 640         /*
 641          * Planning might have modified the rangetable, due to changes of the
 642          * Query structures inside subquery RTEs.  We have to ensure that this
 643          * gets propagated back to the master copy.  But can't do this until we
 644          * are done planning, because all the calls to grouping_planner need
 645          * virgin sub-Queries to work from.  (We are effectively assuming that
 646          * sub-Queries will get planned identically each time, or at least that
 647          * the impacts on their rangetables will be the same each time.)
 648          *
 649          * XXX should clean this up someday
 650          */
 651         parse->rtable = rtable;
 652
 653         return (Plan *) make_append(subplans, true, tlist);
 654 }
 655
 656 /*--------------------
 657  * grouping_planner
 658  *        Perform planning steps related to grouping, aggregation, etc.
 659  *        This primarily means adding top-level processing to the basic
 660  *        query plan produced by query_planner.
 661  *
 662  * tuple_fraction is the fraction of tuples we expect will be retrieved
 663  *
 664  * tuple_fraction is interpreted as follows:
 665  *        0: expect all tuples to be retrieved (normal case)
 666  *        0 < tuple_fraction < 1: expect the given fraction of tuples available
 667  *              from the plan to be retrieved
 668  *        tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
 669  *              expected to be retrieved (ie, a LIMIT specification)
 670  *
 671  * Returns a query plan.  Also, root->query_pathkeys is returned as the
 672  * actual output ordering of the plan (in pathkey format).
 673  *--------------------
 674  */
 675 static Plan *
 676 grouping_planner(PlannerInfo *root, double tuple_fraction)
 677 {
 678         Query      *parse = root->parse;
 679         List       *tlist = parse->targetList;
 680         int64           offset_est = 0;
 681         int64           count_est = 0;
 682         Plan       *result_plan;
 683         List       *current_pathkeys;
 684         List       *sort_pathkeys;
 685         double          dNumGroups = 0;
 686
 687         /* Tweak caller-supplied tuple_fraction if have LIMIT/OFFSET */
 688         if (parse->limitCount || parse->limitOffset)
 689                 tuple_fraction = preprocess_limit(root, tuple_fraction,
 690                                                                                   &offset_est, &count_est);
 691
 692         if (parse->setOperations)
 693         {
 694                 List       *set_sortclauses;
 695
 696                 /*
 697                  * If there's a top-level ORDER BY, assume we have to fetch all the
 698                  * tuples.      This might seem too simplistic given all the hackery below
 699                  * to possibly avoid the sort ... but a nonzero tuple_fraction is only
 700                  * of use to plan_set_operations() when the setop is UNION ALL, and
 701                  * the result of UNION ALL is always unsorted.
 702                  */
 703                 if (parse->sortClause)
 704                         tuple_fraction = 0.0;
 705
 706                 /*
 707                  * Construct the plan for set operations.  The result will not need
 708                  * any work except perhaps a top-level sort and/or LIMIT.
 709                  */
 710                 result_plan = plan_set_operations(root, tuple_fraction,
 711                                                                                   &set_sortclauses);
 712
 713                 /*
 714                  * Calculate pathkeys representing the sort order (if any) of the set
 715                  * operation's result.  We have to do this before overwriting the sort
 716                  * key information...
 717                  */
 718                 current_pathkeys = make_pathkeys_for_sortclauses(set_sortclauses,
 719                                                                                                         result_plan->targetlist);
 720                 current_pathkeys = canonicalize_pathkeys(root, current_pathkeys);
 721
 722                 /*
 723                  * We should not need to call preprocess_targetlist, since we must be
 724                  * in a SELECT query node.      Instead, use the targetlist returned by
 725                  * plan_set_operations (since this tells whether it returned any
 726                  * resjunk columns!), and transfer any sort key information from the
 727                  * original tlist.
 728                  */
 729                 Assert(parse->commandType == CMD_SELECT);
 730
 731                 tlist = postprocess_setop_tlist(result_plan->targetlist, tlist);
 732
 733                 /*
 734                  * Can't handle FOR UPDATE/SHARE here (parser should have checked
 735                  * already, but let's make sure).
 736                  */
 737                 if (parse->rowMarks)
 738                         ereport(ERROR,
 739                                         (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
 740                                          errmsg("SELECT FOR UPDATE/SHARE is not allowed with UNION/INTERSECT/EXCEPT")));
 741
 742                 /*
 743                  * Calculate pathkeys that represent result ordering requirements
 744                  */
 745                 sort_pathkeys = make_pathkeys_for_sortclauses(parse->sortClause,
 746                                                                                                           tlist);
 747                 sort_pathkeys = canonicalize_pathkeys(root, sort_pathkeys);
 748         }
 749         else
 750         {
 751                 /* No set operations, do regular planning */
 752                 List       *sub_tlist;
 753                 List       *group_pathkeys;
 754                 AttrNumber *groupColIdx = NULL;
 755                 Oid                *groupOperators = NULL;
 756                 bool            need_tlist_eval = true;
 757                 QualCost        tlist_cost;
 758                 Path       *cheapest_path;
 759                 Path       *sorted_path;
 760                 Path       *best_path;
 761                 long            numGroups = 0;
 762                 AggClauseCounts agg_counts;
 763                 int                     numGroupCols = list_length(parse->groupClause);
 764                 bool            use_hashed_grouping = false;
 765
 766                 MemSet(&agg_counts, 0, sizeof(AggClauseCounts));
 767
 768                 /* Preprocess targetlist */
 769                 tlist = preprocess_targetlist(root, tlist);
 770
 771                 /*
 772                  * Generate appropriate target list for subplan; may be different from
 773                  * tlist if grouping or aggregation is needed.
 774                  */
 775                 sub_tlist = make_subplanTargetList(root, tlist,
 776                                                                                    &groupColIdx, &need_tlist_eval);
 777
 778                 /*
 779                  * Calculate pathkeys that represent grouping/ordering requirements.
 780                  * Stash them in PlannerInfo so that query_planner can canonicalize
 781                  * them.
 782                  */
 783                 root->group_pathkeys =
 784                         make_pathkeys_for_sortclauses(parse->groupClause, tlist);
 785                 root->sort_pathkeys =
 786                         make_pathkeys_for_sortclauses(parse->sortClause, tlist);
 787
 788                 /*
 789                  * Will need actual number of aggregates for estimating costs.
 790                  *
 791                  * Note: we do not attempt to detect duplicate aggregates here; a
 792                  * somewhat-overestimated count is okay for our present purposes.
 793                  *
 794                  * Note: think not that we can turn off hasAggs if we find no aggs. It
 795                  * is possible for constant-expression simplification to remove all
 796                  * explicit references to aggs, but we still have to follow the
 797                  * aggregate semantics (eg, producing only one output row).
 798                  */
 799                 if (parse->hasAggs)
 800                 {
 801                         count_agg_clauses((Node *) tlist, &agg_counts);
 802                         count_agg_clauses(parse->havingQual, &agg_counts);
 803                 }
 804
 805                 /*
 806                  * Figure out whether we need a sorted result from query_planner.
 807                  *
 808                  * If we have a GROUP BY clause, then we want a result sorted properly
 809                  * for grouping.  Otherwise, if there is an ORDER BY clause, we want
 810                  * to sort by the ORDER BY clause.      (Note: if we have both, and ORDER
 811                  * BY is a superset of GROUP BY, it would be tempting to request sort
 812                  * by ORDER BY --- but that might just leave us failing to exploit an
 813                  * available sort order at all. Needs more thought...)
 814                  */
 815                 if (parse->groupClause)
 816                         root->query_pathkeys = root->group_pathkeys;
 817                 else if (parse->sortClause)
 818                         root->query_pathkeys = root->sort_pathkeys;
 819                 else
 820                         root->query_pathkeys = NIL;
 821
 822                 /*
 823                  * Generate the best unsorted and presorted paths for this Query (but
 824                  * note there may not be any presorted path).  query_planner will also
 825                  * estimate the number of groups in the query, and canonicalize all
 826                  * the pathkeys.
 827                  */
 828                 query_planner(root, sub_tlist, tuple_fraction,
 829                                           &cheapest_path, &sorted_path, &dNumGroups);
 830
 831                 group_pathkeys = root->group_pathkeys;
 832                 sort_pathkeys = root->sort_pathkeys;
 833
 834                 /*
 835                  * If grouping, extract the grouping operators and decide whether we
 836                  * want to use hashed grouping.
 837                  */
 838                 if (parse->groupClause)
 839                 {
 840                         groupOperators = extract_grouping_ops(parse->groupClause);
 841                         use_hashed_grouping =
 842                                 choose_hashed_grouping(root, tuple_fraction,
 843                                                                            cheapest_path, sorted_path,
 844                                                                            groupOperators, dNumGroups,
 845                                                                            &agg_counts);
 846
 847                         /* Also convert # groups to long int --- but 'ware overflow! */
 848                         numGroups = (long) Min(dNumGroups, (double) LONG_MAX);
 849                 }
 850
 851                 /*
 852                  * Select the best path.  If we are doing hashed grouping, we will
 853                  * always read all the input tuples, so use the cheapest-total path.
 854                  * Otherwise, trust query_planner's decision about which to use.
 855                  */
 856                 if (use_hashed_grouping || !sorted_path)
 857                         best_path = cheapest_path;
 858                 else
 859                         best_path = sorted_path;
 860
 861                 /*
 862                  * Check to see if it's possible to optimize MIN/MAX aggregates. If
 863                  * so, we will forget all the work we did so far to choose a "regular"
 864                  * path ... but we had to do it anyway to be able to tell which way is
 865                  * cheaper.
 866                  */
 867                 result_plan = optimize_minmax_aggregates(root,
 868                                                                                                  tlist,
 869                                                                                                  best_path);
 870                 if (result_plan != NULL)
 871                 {
 872                         /*
 873                          * optimize_minmax_aggregates generated the full plan, with the
 874                          * right tlist, and it has no sort order.
 875                          */
 876                         current_pathkeys = NIL;
 877                 }
 878                 else
 879                 {
 880                         /*
 881                          * Normal case --- create a plan according to query_planner's
 882                          * results.
 883                          */
 884                         result_plan = create_plan(root, best_path);
 885                         current_pathkeys = best_path->pathkeys;
 886
 887                         /*
 888                          * create_plan() returns a plan with just a "flat" tlist of
 889                          * required Vars.  Usually we need to insert the sub_tlist as the
 890                          * tlist of the top plan node.  However, we can skip that if we
 891                          * determined that whatever query_planner chose to return will be
 892                          * good enough.
 893                          */
 894                         if (need_tlist_eval)
 895                         {
 896                                 /*
 897                                  * If the top-level plan node is one that cannot do expression
 898                                  * evaluation, we must insert a Result node to project the
 899                                  * desired tlist.
 900                                  */
 901                                 if (!is_projection_capable_plan(result_plan))
 902                                 {
 903                                         result_plan = (Plan *) make_result(sub_tlist, NULL,
 904                                                                                                            result_plan);
 905                                 }
 906                                 else
 907                                 {
 908                                         /*
 909                                          * Otherwise, just replace the subplan's flat tlist with
 910                                          * the desired tlist.
 911                                          */
 912                                         result_plan->targetlist = sub_tlist;
 913                                 }
 914
 915                                 /*
 916                                  * Also, account for the cost of evaluation of the sub_tlist.
 917                                  *
 918                                  * Up to now, we have only been dealing with "flat" tlists,
 919                                  * containing just Vars.  So their evaluation cost is zero
 920                                  * according to the model used by cost_qual_eval() (or if you
 921                                  * prefer, the cost is factored into cpu_tuple_cost).  Thus we
 922                                  * can avoid accounting for tlist cost throughout
 923                                  * query_planner() and subroutines.  But now we've inserted a
 924                                  * tlist that might contain actual operators, sub-selects, etc
 925                                  * --- so we'd better account for its cost.
 926                                  *
 927                                  * Below this point, any tlist eval cost for added-on nodes
 928                                  * should be accounted for as we create those nodes.
 929                                  * Presently, of the node types we can add on, only Agg and
 930                                  * Group project new tlists (the rest just copy their input
 931                                  * tuples) --- so make_agg() and make_group() are responsible
 932                                  * for computing the added cost.
 933                                  */
 934                                 cost_qual_eval(&tlist_cost, sub_tlist);
 935                                 result_plan->startup_cost += tlist_cost.startup;
 936                                 result_plan->total_cost += tlist_cost.startup +
 937                                         tlist_cost.per_tuple * result_plan->plan_rows;
 938                         }
 939                         else
 940                         {
 941                                 /*
 942                                  * Since we're using query_planner's tlist and not the one
 943                                  * make_subplanTargetList calculated, we have to refigure any
 944                                  * grouping-column indexes make_subplanTargetList computed.
 945                                  */
 946                                 locate_grouping_columns(root, tlist, result_plan->targetlist,
 947                                                                                 groupColIdx);
 948                         }
 949
 950                         /*
 951                          * Insert AGG or GROUP node if needed, plus an explicit sort step
 952                          * if necessary.
 953                          *
 954                          * HAVING clause, if any, becomes qual of the Agg or Group node.
 955                          */
 956                         if (use_hashed_grouping)
 957                         {
 958                                 /* Hashed aggregate plan --- no sort needed */
 959                                 result_plan = (Plan *) make_agg(root,
 960                                                                                                 tlist,
 961                                                                                                 (List *) parse->havingQual,
 962                                                                                                 AGG_HASHED,
 963                                                                                                 numGroupCols,
 964                                                                                                 groupColIdx,
 965                                                                                                 groupOperators,
 966                                                                                                 numGroups,
 967                                                                                                 agg_counts.numAggs,
 968                                                                                                 result_plan);
 969                                 /* Hashed aggregation produces randomly-ordered results */
 970                                 current_pathkeys = NIL;
 971                         }
 972                         else if (parse->hasAggs)
 973                         {
 974                                 /* Plain aggregate plan --- sort if needed */
 975                                 AggStrategy aggstrategy;
 976
 977                                 if (parse->groupClause)
 978                                 {
 979                                         if (!pathkeys_contained_in(group_pathkeys,
 980                                                                                            current_pathkeys))
 981                                         {
 982                                                 result_plan = (Plan *)
 983                                                         make_sort_from_groupcols(root,
 984                                                                                                          parse->groupClause,
 985                                                                                                          groupColIdx,
 986                                                                                                          result_plan);
 987                                                 current_pathkeys = group_pathkeys;
 988                                         }
 989                                         aggstrategy = AGG_SORTED;
 990
 991                                         /*
 992                                          * The AGG node will not change the sort ordering of its
 993                                          * groups, so current_pathkeys describes the result too.
 994                                          */
 995                                 }
 996                                 else
 997                                 {
 998                                         aggstrategy = AGG_PLAIN;
 999                                         /* Result will be only one row anyway; no sort order */
1000                                         current_pathkeys = NIL;
1001                                 }
1002
1003                                 result_plan = (Plan *) make_agg(root,
1004                                                                                                 tlist,
1005                                                                                                 (List *) parse->havingQual,
1006                                                                                                 aggstrategy,
1007                                                                                                 numGroupCols,
1008                                                                                                 groupColIdx,
1009                                                                                                 groupOperators,
1010                                                                                                 numGroups,
1011                                                                                                 agg_counts.numAggs,
1012                                                                                                 result_plan);
1013                         }
1014                         else if (parse->groupClause)
1015                         {
1016                                 /*
1017                                  * GROUP BY without aggregation, so insert a group node (plus
1018                                  * the appropriate sort node, if necessary).
1019                                  *
1020                                  * Add an explicit sort if we couldn't make the path come out
1021                                  * the way the GROUP node needs it.
1022                                  */
1023                                 if (!pathkeys_contained_in(group_pathkeys, current_pathkeys))
1024                                 {
1025                                         result_plan = (Plan *)
1026                                                 make_sort_from_groupcols(root,
1027                                                                                                  parse->groupClause,
1028                                                                                                  groupColIdx,
1029                                                                                                  result_plan);
1030                                         current_pathkeys = group_pathkeys;
1031                                 }
1032
1033                                 result_plan = (Plan *) make_group(root,
1034                                                                                                   tlist,
1035                                                                                                   (List *) parse->havingQual,
1036                                                                                                   numGroupCols,
1037                                                                                                   groupColIdx,
1038                                                                                                   groupOperators,
1039                                                                                                   dNumGroups,
1040                                                                                                   result_plan);
1041                                 /* The Group node won't change sort ordering */
1042                         }
1043                         else if (root->hasHavingQual)
1044                         {
1045                                 /*
1046                                  * No aggregates, and no GROUP BY, but we have a HAVING qual.
1047                                  * This is a degenerate case in which we are supposed to emit
1048                                  * either 0 or 1 row depending on whether HAVING succeeds.
1049                                  * Furthermore, there cannot be any variables in either HAVING
1050                                  * or the targetlist, so we actually do not need the FROM
1051                                  * table at all!  We can just throw away the plan-so-far and
1052                                  * generate a Result node.      This is a sufficiently unusual
1053                                  * corner case that it's not worth contorting the structure of
1054                                  * this routine to avoid having to generate the plan in the
1055                                  * first place.
1056                                  */
1057                                 result_plan = (Plan *) make_result(tlist,
1058                                                                                                    parse->havingQual,
1059                                                                                                    NULL);
1060                         }
1061                 }                                               /* end of non-minmax-aggregate case */
1062         }                                                       /* end of if (setOperations) */
1063
1064         /*
1065          * If we were not able to make the plan come out in the right order, add
1066          * an explicit sort step.
1067          */
1068         if (parse->sortClause)
1069         {
1070                 if (!pathkeys_contained_in(sort_pathkeys, current_pathkeys))
1071                 {
1072                         result_plan = (Plan *)
1073                                 make_sort_from_sortclauses(root,
1074                                                                                    parse->sortClause,
1075                                                                                    result_plan);
1076                         current_pathkeys = sort_pathkeys;
1077                 }
1078         }
1079
1080         /*
1081          * If there is a DISTINCT clause, add the UNIQUE node.
1082          */
1083         if (parse->distinctClause)
1084         {
1085                 result_plan = (Plan *) make_unique(result_plan, parse->distinctClause);
1086
1087                 /*
1088                  * If there was grouping or aggregation, leave plan_rows as-is (ie,
1089                  * assume the result was already mostly unique).  If not, use the
1090                  * number of distinct-groups calculated by query_planner.
1091                  */
1092                 if (!parse->groupClause && !root->hasHavingQual && !parse->hasAggs)
1093                         result_plan->plan_rows = dNumGroups;
1094         }
1095
1096         /*
1097          * Finally, if there is a LIMIT/OFFSET clause, add the LIMIT node.
1098          */
1099         if (parse->limitCount || parse->limitOffset)
1100         {
1101                 result_plan = (Plan *) make_limit(result_plan,
1102                                                                                   parse->limitOffset,
1103                                                                                   parse->limitCount,
1104                                                                                   offset_est,
1105                                                                                   count_est);
1106         }
1107
1108         /*
1109          * Deal with the RETURNING clause if any.  It's convenient to pass the
1110          * returningList through setrefs.c now rather than at top level (if we
1111          * waited, handling inherited UPDATE/DELETE would be much harder).
1112          */
1113         if (parse->returningList)
1114         {
1115                 List       *rlist;
1116
1117                 rlist = set_returning_clause_references(parse->returningList,
1118                                                                                                 result_plan,
1119                                                                                                 parse->resultRelation);
1120                 parse->returningLists = list_make1(rlist);
1121         }
1122
1123         /*
1124          * Return the actual output ordering in query_pathkeys for possible use by
1125          * an outer query level.
1126          */
1127         root->query_pathkeys = current_pathkeys;
1128
1129         return result_plan;
1130 }
1131
1132 /*
1133  * Detect whether a plan node is a "dummy" plan created when a relation
1134  * is deemed not to need scanning due to constraint exclusion.
1135  *
1136  * Currently, such dummy plans are Result nodes with constant FALSE
1137  * filter quals.
1138  */
1139 static bool
1140 is_dummy_plan(Plan *plan)
1141 {
1142         if (IsA(plan, Result))
1143         {
1144                 List       *rcqual = (List *) ((Result *) plan)->resconstantqual;
1145
1146                 if (list_length(rcqual) == 1)
1147                 {
1148                         Const      *constqual = (Const *) linitial(rcqual);
1149
1150                         if (constqual && IsA(constqual, Const))
1151                         {
1152                                 if (!constqual->constisnull &&
1153                                         !DatumGetBool(constqual->constvalue))
1154                                         return true;
1155                         }
1156                 }
1157         }
1158         return false;
1159 }
1160
1161 /*
1162  * preprocess_limit - do pre-estimation for LIMIT and/or OFFSET clauses
1163  *
1164  * We try to estimate the values of the LIMIT/OFFSET clauses, and pass the
1165  * results back in *count_est and *offset_est.  These variables are set to
1166  * 0 if the corresponding clause is not present, and -1 if it's present
1167  * but we couldn't estimate the value for it.  (The "0" convention is OK
1168  * for OFFSET but a little bit bogus for LIMIT: effectively we estimate
1169  * LIMIT 0 as though it were LIMIT 1.  But this is in line with the planner's
1170  * usual practice of never estimating less than one row.)  These values will
1171  * be passed to make_limit, which see if you change this code.
1172  *
1173  * The return value is the suitably adjusted tuple_fraction to use for
1174  * planning the query.  This adjustment is not overridable, since it reflects
1175  * plan actions that grouping_planner() will certainly take, not assumptions
1176  * about context.
1177  */
1178 static double
1179 preprocess_limit(PlannerInfo *root, double tuple_fraction,
1180                                  int64 *offset_est, int64 *count_est)
1181 {
1182         Query      *parse = root->parse;
1183         Node       *est;
1184         double          limit_fraction;
1185
1186         /* Should not be called unless LIMIT or OFFSET */
1187         Assert(parse->limitCount || parse->limitOffset);
1188
1189         /*
1190          * Try to obtain the clause values.  We use estimate_expression_value
1191          * primarily because it can sometimes do something useful with Params.
1192          */
1193         if (parse->limitCount)
1194         {
1195                 est = estimate_expression_value(parse->limitCount);
1196                 if (est && IsA(est, Const))
1197                 {
1198                         if (((Const *) est)->constisnull)
1199                         {
1200                                 /* NULL indicates LIMIT ALL, ie, no limit */
1201                                 *count_est = 0; /* treat as not present */
1202                         }
1203                         else
1204                         {
1205                                 *count_est = DatumGetInt64(((Const *) est)->constvalue);
1206                                 if (*count_est <= 0)
1207                                         *count_est = 1;         /* force to at least 1 */
1208                         }
1209                 }
1210                 else
1211                         *count_est = -1;        /* can't estimate */
1212         }
1213         else
1214                 *count_est = 0;                 /* not present */
1215
1216         if (parse->limitOffset)
1217         {
1218                 est = estimate_expression_value(parse->limitOffset);
1219                 if (est && IsA(est, Const))
1220                 {
1221                         if (((Const *) est)->constisnull)
1222                         {
1223                                 /* Treat NULL as no offset; the executor will too */
1224                                 *offset_est = 0;        /* treat as not present */
1225                         }
1226                         else
1227                         {
1228                                 *offset_est = DatumGetInt64(((Const *) est)->constvalue);
1229                                 if (*offset_est < 0)
1230                                         *offset_est = 0;        /* less than 0 is same as 0 */
1231                         }
1232                 }
1233                 else
1234                         *offset_est = -1;       /* can't estimate */
1235         }
1236         else
1237                 *offset_est = 0;                /* not present */
1238
1239         if (*count_est != 0)
1240         {
1241                 /*
1242                  * A LIMIT clause limits the absolute number of tuples returned.
1243                  * However, if it's not a constant LIMIT then we have to guess; for
1244                  * lack of a better idea, assume 10% of the plan's result is wanted.
1245                  */
1246                 if (*count_est < 0 || *offset_est < 0)
1247                 {
1248                         /* LIMIT or OFFSET is an expression ... punt ... */
1249                         limit_fraction = 0.10;
1250                 }
1251                 else
1252                 {
1253                         /* LIMIT (plus OFFSET, if any) is max number of tuples needed */
1254                         limit_fraction = (double) *count_est + (double) *offset_est;
1255                 }
1256
1257                 /*
1258                  * If we have absolute limits from both caller and LIMIT, use the
1259                  * smaller value; likewise if they are both fractional.  If one is
1260                  * fractional and the other absolute, we can't easily determine which
1261                  * is smaller, but we use the heuristic that the absolute will usually
1262                  * be smaller.
1263                  */
1264                 if (tuple_fraction >= 1.0)
1265                 {
1266                         if (limit_fraction >= 1.0)
1267                         {
1268                                 /* both absolute */
1269                                 tuple_fraction = Min(tuple_fraction, limit_fraction);
1270                         }
1271                         else
1272                         {
1273                                 /* caller absolute, limit fractional; use caller's value */
1274                         }
1275                 }
1276                 else if (tuple_fraction > 0.0)
1277                 {
1278                         if (limit_fraction >= 1.0)
1279                         {
1280                                 /* caller fractional, limit absolute; use limit */
1281                                 tuple_fraction = limit_fraction;
1282                         }
1283                         else
1284                         {
1285                                 /* both fractional */
1286                                 tuple_fraction = Min(tuple_fraction, limit_fraction);
1287                         }
1288                 }
1289                 else
1290                 {
1291                         /* no info from caller, just use limit */
1292                         tuple_fraction = limit_fraction;
1293                 }
1294         }
1295         else if (*offset_est != 0 && tuple_fraction > 0.0)
1296         {
1297                 /*
1298                  * We have an OFFSET but no LIMIT.      This acts entirely differently
1299                  * from the LIMIT case: here, we need to increase rather than decrease
1300                  * the caller's tuple_fraction, because the OFFSET acts to cause more
1301                  * tuples to be fetched instead of fewer.  This only matters if we got
1302                  * a tuple_fraction > 0, however.
1303                  *
1304                  * As above, use 10% if OFFSET is present but unestimatable.
1305                  */
1306                 if (*offset_est < 0)
1307                         limit_fraction = 0.10;
1308                 else
1309                         limit_fraction = (double) *offset_est;
1310
1311                 /*
1312                  * If we have absolute counts from both caller and OFFSET, add them
1313                  * together; likewise if they are both fractional.      If one is
1314                  * fractional and the other absolute, we want to take the larger, and
1315                  * we heuristically assume that's the fractional one.
1316                  */
1317                 if (tuple_fraction >= 1.0)
1318                 {
1319                         if (limit_fraction >= 1.0)
1320                         {
1321                                 /* both absolute, so add them together */
1322                                 tuple_fraction += limit_fraction;
1323                         }
1324                         else
1325                         {
1326                                 /* caller absolute, limit fractional; use limit */
1327                                 tuple_fraction = limit_fraction;
1328                         }
1329                 }
1330                 else
1331                 {
1332                         if (limit_fraction >= 1.0)
1333                         {
1334                                 /* caller fractional, limit absolute; use caller's value */
1335                         }
1336                         else
1337                         {
1338                                 /* both fractional, so add them together */
1339                                 tuple_fraction += limit_fraction;
1340                                 if (tuple_fraction >= 1.0)
1341                                         tuple_fraction = 0.0;           /* assume fetch all */
1342                         }
1343                 }
1344         }
1345
1346         return tuple_fraction;
1347 }
1348
1349 /*
1350  * extract_grouping_ops - make an array of the equality operator OIDs
1351  *              for the GROUP BY clause
1352  */
1353 static Oid *
1354 extract_grouping_ops(List *groupClause)
1355 {
1356         int                     numCols = list_length(groupClause);
1357         int                     colno = 0;
1358         Oid                *groupOperators;
1359         ListCell   *glitem;
1360
1361         groupOperators = (Oid *) palloc(sizeof(Oid) * numCols);
1362
1363         foreach(glitem, groupClause)
1364         {
1365                 GroupClause *groupcl = (GroupClause *) lfirst(glitem);
1366
1367                 groupOperators[colno] = get_equality_op_for_ordering_op(groupcl->sortop);
1368                 if (!OidIsValid(groupOperators[colno]))         /* shouldn't happen */
1369                         elog(ERROR, "could not find equality operator for ordering operator %u",
1370                                  groupcl->sortop);
1371                 colno++;
1372         }
1373
1374         return groupOperators;
1375 }
1376
1377 /*
1378  * choose_hashed_grouping - should we use hashed grouping?
1379  */
1380 static bool
1381 choose_hashed_grouping(PlannerInfo *root, double tuple_fraction,
1382                                            Path *cheapest_path, Path *sorted_path,
1383                                            Oid *groupOperators, double dNumGroups,
1384                                            AggClauseCounts *agg_counts)
1385 {
1386         int                     numGroupCols = list_length(root->parse->groupClause);
1387         double          cheapest_path_rows;
1388         int                     cheapest_path_width;
1389         Size            hashentrysize;
1390         List       *current_pathkeys;
1391         Path            hashed_p;
1392         Path            sorted_p;
1393         int                     i;
1394
1395         /*
1396          * Check can't-do-it conditions, including whether the grouping operators
1397          * are hashjoinable.  (We assume hashing is OK if they are marked
1398          * oprcanhash.  If there isn't actually a supporting hash function,
1399          * the executor will complain at runtime.)
1400          *
1401          * Executor doesn't support hashed aggregation with DISTINCT aggregates.
1402          * (Doing so would imply storing *all* the input values in the hash table,
1403          * which seems like a certain loser.)
1404          */
1405         if (!enable_hashagg)
1406                 return false;
1407         if (agg_counts->numDistinctAggs != 0)
1408                 return false;
1409         for (i = 0; i < numGroupCols; i++)
1410         {
1411                 if (!op_hashjoinable(groupOperators[i]))
1412                         return false;
1413         }
1414
1415         /*
1416          * Don't do it if it doesn't look like the hashtable will fit into
1417          * work_mem.
1418          *
1419          * Beware here of the possibility that cheapest_path->parent is NULL. This
1420          * could happen if user does something silly like SELECT 'foo' GROUP BY 1;
1421          */
1422         if (cheapest_path->parent)
1423         {
1424                 cheapest_path_rows = cheapest_path->parent->rows;
1425                 cheapest_path_width = cheapest_path->parent->width;
1426         }
1427         else
1428         {
1429                 cheapest_path_rows = 1; /* assume non-set result */
1430                 cheapest_path_width = 100;              /* arbitrary */
1431         }
1432
1433         /* Estimate per-hash-entry space at tuple width... */
1434         hashentrysize = MAXALIGN(cheapest_path_width) + MAXALIGN(sizeof(MinimalTupleData));
1435         /* plus space for pass-by-ref transition values... */
1436         hashentrysize += agg_counts->transitionSpace;
1437         /* plus the per-hash-entry overhead */
1438         hashentrysize += hash_agg_entry_size(agg_counts->numAggs);
1439
1440         if (hashentrysize * dNumGroups > work_mem * 1024L)
1441                 return false;
1442
1443         /*
1444          * See if the estimated cost is no more than doing it the other way. While
1445          * avoiding the need for sorted input is usually a win, the fact that the
1446          * output won't be sorted may be a loss; so we need to do an actual cost
1447          * comparison.
1448          *
1449          * We need to consider cheapest_path + hashagg [+ final sort] versus
1450          * either cheapest_path [+ sort] + group or agg [+ final sort] or
1451          * presorted_path + group or agg [+ final sort] where brackets indicate a
1452          * step that may not be needed. We assume query_planner() will have
1453          * returned a presorted path only if it's a winner compared to
1454          * cheapest_path for this purpose.
1455          *
1456          * These path variables are dummies that just hold cost fields; we don't
1457          * make actual Paths for these steps.
1458          */
1459         cost_agg(&hashed_p, root, AGG_HASHED, agg_counts->numAggs,
1460                          numGroupCols, dNumGroups,
1461                          cheapest_path->startup_cost, cheapest_path->total_cost,
1462                          cheapest_path_rows);
1463         /* Result of hashed agg is always unsorted */
1464         if (root->sort_pathkeys)
1465                 cost_sort(&hashed_p, root, root->sort_pathkeys, hashed_p.total_cost,
1466                                   dNumGroups, cheapest_path_width);
1467
1468         if (sorted_path)
1469         {
1470                 sorted_p.startup_cost = sorted_path->startup_cost;
1471                 sorted_p.total_cost = sorted_path->total_cost;
1472                 current_pathkeys = sorted_path->pathkeys;
1473         }
1474         else
1475         {
1476                 sorted_p.startup_cost = cheapest_path->startup_cost;
1477                 sorted_p.total_cost = cheapest_path->total_cost;
1478                 current_pathkeys = cheapest_path->pathkeys;
1479         }
1480         if (!pathkeys_contained_in(root->group_pathkeys, current_pathkeys))
1481         {
1482                 cost_sort(&sorted_p, root, root->group_pathkeys, sorted_p.total_cost,
1483                                   cheapest_path_rows, cheapest_path_width);
1484                 current_pathkeys = root->group_pathkeys;
1485         }
1486
1487         if (root->parse->hasAggs)
1488                 cost_agg(&sorted_p, root, AGG_SORTED, agg_counts->numAggs,
1489                                  numGroupCols, dNumGroups,
1490                                  sorted_p.startup_cost, sorted_p.total_cost,
1491                                  cheapest_path_rows);
1492         else
1493                 cost_group(&sorted_p, root, numGroupCols, dNumGroups,
1494                                    sorted_p.startup_cost, sorted_p.total_cost,
1495                                    cheapest_path_rows);
1496         /* The Agg or Group node will preserve ordering */
1497         if (root->sort_pathkeys &&
1498                 !pathkeys_contained_in(root->sort_pathkeys, current_pathkeys))
1499                 cost_sort(&sorted_p, root, root->sort_pathkeys, sorted_p.total_cost,
1500                                   dNumGroups, cheapest_path_width);
1501
1502         /*
1503          * Now make the decision using the top-level tuple fraction.  First we
1504          * have to convert an absolute count (LIMIT) into fractional form.
1505          */
1506         if (tuple_fraction >= 1.0)
1507                 tuple_fraction /= dNumGroups;
1508
1509         if (compare_fractional_path_costs(&hashed_p, &sorted_p,
1510                                                                           tuple_fraction) < 0)
1511         {
1512                 /* Hashed is cheaper, so use it */
1513                 return true;
1514         }
1515         return false;
1516 }
1517
1518 /*---------------
1519  * make_subplanTargetList
1520  *        Generate appropriate target list when grouping is required.
1521  *
1522  * When grouping_planner inserts Aggregate, Group, or Result plan nodes
1523  * above the result of query_planner, we typically want to pass a different
1524  * target list to query_planner than the outer plan nodes should have.
1525  * This routine generates the correct target list for the subplan.
1526  *
1527  * The initial target list passed from the parser already contains entries
1528  * for all ORDER BY and GROUP BY expressions, but it will not have entries
1529  * for variables used only in HAVING clauses; so we need to add those
1530  * variables to the subplan target list.  Also, we flatten all expressions
1531  * except GROUP BY items into their component variables; the other expressions
1532  * will be computed by the inserted nodes rather than by the subplan.
1533  * For example, given a query like
1534  *              SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
1535  * we want to pass this targetlist to the subplan:
1536  *              a,b,c,d,a+b
1537  * where the a+b target will be used by the Sort/Group steps, and the
1538  * other targets will be used for computing the final results.  (In the
1539  * above example we could theoretically suppress the a and b targets and
1540  * pass down only c,d,a+b, but it's not really worth the trouble to
1541  * eliminate simple var references from the subplan.  We will avoid doing
1542  * the extra computation to recompute a+b at the outer level; see
1543  * replace_vars_with_subplan_refs() in setrefs.c.)
1544  *
1545  * If we are grouping or aggregating, *and* there are no non-Var grouping
1546  * expressions, then the returned tlist is effectively dummy; we do not
1547  * need to force it to be evaluated, because all the Vars it contains
1548  * should be present in the output of query_planner anyway.
1549  *
1550  * 'tlist' is the query's target list.
1551  * 'groupColIdx' receives an array of column numbers for the GROUP BY
1552  *                      expressions (if there are any) in the subplan's target list.
1553  * 'need_tlist_eval' is set true if we really need to evaluate the
1554  *                      result tlist.
1555  *
1556  * The result is the targetlist to be passed to the subplan.
1557  *---------------
1558  */
1559 static List *
1560 make_subplanTargetList(PlannerInfo *root,
1561                                            List *tlist,
1562                                            AttrNumber **groupColIdx,
1563                                            bool *need_tlist_eval)
1564 {
1565         Query      *parse = root->parse;
1566         List       *sub_tlist;
1567         List       *extravars;
1568         int                     numCols;
1569
1570         *groupColIdx = NULL;
1571
1572         /*
1573          * If we're not grouping or aggregating, there's nothing to do here;
1574          * query_planner should receive the unmodified target list.
1575          */
1576         if (!parse->hasAggs && !parse->groupClause && !root->hasHavingQual)
1577         {
1578                 *need_tlist_eval = true;
1579                 return tlist;
1580         }
1581
1582         /*
1583          * Otherwise, start with a "flattened" tlist (having just the vars
1584          * mentioned in the targetlist and HAVING qual --- but not upper- level
1585          * Vars; they will be replaced by Params later on).
1586          */
1587         sub_tlist = flatten_tlist(tlist);
1588         extravars = pull_var_clause(parse->havingQual, false);
1589         sub_tlist = add_to_flat_tlist(sub_tlist, extravars);
1590         list_free(extravars);
1591         *need_tlist_eval = false;       /* only eval if not flat tlist */
1592
1593         /*
1594          * If grouping, create sub_tlist entries for all GROUP BY expressions
1595          * (GROUP BY items that are simple Vars should be in the list already),
1596          * and make an array showing where the group columns are in the sub_tlist.
1597          */
1598         numCols = list_length(parse->groupClause);
1599         if (numCols > 0)
1600         {
1601                 int                     keyno = 0;
1602                 AttrNumber *grpColIdx;
1603                 ListCell   *gl;
1604
1605                 grpColIdx = (AttrNumber *) palloc(sizeof(AttrNumber) * numCols);
1606                 *groupColIdx = grpColIdx;
1607
1608                 foreach(gl, parse->groupClause)
1609                 {
1610                         GroupClause *grpcl = (GroupClause *) lfirst(gl);
1611                         Node       *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1612                         TargetEntry *te = NULL;
1613                         ListCell   *sl;
1614
1615                         /* Find or make a matching sub_tlist entry */
1616                         foreach(sl, sub_tlist)
1617                         {
1618                                 te = (TargetEntry *) lfirst(sl);
1619                                 if (equal(groupexpr, te->expr))
1620                                         break;
1621                         }
1622                         if (!sl)
1623                         {
1624                                 te = makeTargetEntry((Expr *) groupexpr,
1625                                                                          list_length(sub_tlist) + 1,
1626                                                                          NULL,
1627                                                                          false);
1628                                 sub_tlist = lappend(sub_tlist, te);
1629                                 *need_tlist_eval = true;                /* it's not flat anymore */
1630                         }
1631
1632                         /* and save its resno */
1633                         grpColIdx[keyno++] = te->resno;
1634                 }
1635         }
1636
1637         return sub_tlist;
1638 }
1639
1640 /*
1641  * locate_grouping_columns
1642  *              Locate grouping columns in the tlist chosen by query_planner.
1643  *
1644  * This is only needed if we don't use the sub_tlist chosen by
1645  * make_subplanTargetList.      We have to forget the column indexes found
1646  * by that routine and re-locate the grouping vars in the real sub_tlist.
1647  */
1648 static void
1649 locate_grouping_columns(PlannerInfo *root,
1650                                                 List *tlist,
1651                                                 List *sub_tlist,
1652                                                 AttrNumber *groupColIdx)
1653 {
1654         int                     keyno = 0;
1655         ListCell   *gl;
1656
1657         /*
1658          * No work unless grouping.
1659          */
1660         if (!root->parse->groupClause)
1661         {
1662                 Assert(groupColIdx == NULL);
1663                 return;
1664         }
1665         Assert(groupColIdx != NULL);
1666
1667         foreach(gl, root->parse->groupClause)
1668         {
1669                 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1670                 Node       *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1671                 TargetEntry *te = NULL;
1672                 ListCell   *sl;
1673
1674                 foreach(sl, sub_tlist)
1675                 {
1676                         te = (TargetEntry *) lfirst(sl);
1677                         if (equal(groupexpr, te->expr))
1678                                 break;
1679                 }
1680                 if (!sl)
1681                         elog(ERROR, "failed to locate grouping columns");
1682
1683                 groupColIdx[keyno++] = te->resno;
1684         }
1685 }
1686
1687 /*
1688  * postprocess_setop_tlist
1689  *        Fix up targetlist returned by plan_set_operations().
1690  *
1691  * We need to transpose sort key info from the orig_tlist into new_tlist.
1692  * NOTE: this would not be good enough if we supported resjunk sort keys
1693  * for results of set operations --- then, we'd need to project a whole
1694  * new tlist to evaluate the resjunk columns.  For now, just ereport if we
1695  * find any resjunk columns in orig_tlist.
1696  */
1697 static List *
1698 postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
1699 {
1700         ListCell   *l;
1701         ListCell   *orig_tlist_item = list_head(orig_tlist);
1702
1703         foreach(l, new_tlist)
1704         {
1705                 TargetEntry *new_tle = (TargetEntry *) lfirst(l);
1706                 TargetEntry *orig_tle;
1707
1708                 /* ignore resjunk columns in setop result */
1709                 if (new_tle->resjunk)
1710                         continue;
1711
1712                 Assert(orig_tlist_item != NULL);
1713                 orig_tle = (TargetEntry *) lfirst(orig_tlist_item);
1714                 orig_tlist_item = lnext(orig_tlist_item);
1715                 if (orig_tle->resjunk)  /* should not happen */
1716                         elog(ERROR, "resjunk output columns are not implemented");
1717                 Assert(new_tle->resno == orig_tle->resno);
1718                 new_tle->ressortgroupref = orig_tle->ressortgroupref;
1719         }
1720         if (orig_tlist_item != NULL)
1721                 elog(ERROR, "resjunk output columns are not implemented");
1722         return new_tlist;
1723 }