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 1.212 2007/01/20 20:45:39 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->planner_cxt = CurrentMemoryContext;
 210         root->eq_classes = NIL;
 211         root->in_info_list = NIL;
 212         root->append_rel_list = NIL;
 213
 214         /*
 215          * Look for IN clauses at the top level of WHERE, and transform them into
 216          * joins.  Note that this step only handles IN clauses originally at top
 217          * level of WHERE; if we pull up any subqueries in the next step, their
 218          * INs are processed just before pulling them up.
 219          */
 220         if (parse->hasSubLinks)
 221                 parse->jointree->quals = pull_up_IN_clauses(root,
 222                                                                                                         parse->jointree->quals);
 223
 224         /*
 225          * Check to see if any subqueries in the rangetable can be merged into
 226          * this query.
 227          */
 228         parse->jointree = (FromExpr *)
 229                 pull_up_subqueries(root, (Node *) parse->jointree, false, false);
 230
 231         /*
 232          * Detect whether any rangetable entries are RTE_JOIN kind; if not, we can
 233          * avoid the expense of doing flatten_join_alias_vars().  Also check for
 234          * outer joins --- if none, we can skip reduce_outer_joins() and some
 235          * other processing.  This must be done after we have done
 236          * pull_up_subqueries, of course.
 237          *
 238          * Note: if reduce_outer_joins manages to eliminate all outer joins,
 239          * root->hasOuterJoins is not reset currently.  This is OK since its
 240          * purpose is merely to suppress unnecessary processing in simple cases.
 241          */
 242         root->hasJoinRTEs = false;
 243         root->hasOuterJoins = false;
 244         foreach(l, parse->rtable)
 245         {
 246                 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
 247
 248                 if (rte->rtekind == RTE_JOIN)
 249                 {
 250                         root->hasJoinRTEs = true;
 251                         if (IS_OUTER_JOIN(rte->jointype))
 252                         {
 253                                 root->hasOuterJoins = true;
 254                                 /* Can quit scanning once we find an outer join */
 255                                 break;
 256                         }
 257                 }
 258         }
 259
 260         /*
 261          * Expand any rangetable entries that are inheritance sets into "append
 262          * relations".  This can add entries to the rangetable, but they must be
 263          * plain base relations not joins, so it's OK (and marginally more
 264          * efficient) to do it after checking for join RTEs.  We must do it after
 265          * pulling up subqueries, else we'd fail to handle inherited tables in
 266          * subqueries.
 267          */
 268         expand_inherited_tables(root);
 269
 270         /*
 271          * Set hasHavingQual to remember if HAVING clause is present.  Needed
 272          * because preprocess_expression will reduce a constant-true condition to
 273          * an empty qual list ... but "HAVING TRUE" is not a semantic no-op.
 274          */
 275         root->hasHavingQual = (parse->havingQual != NULL);
 276
 277         /* Clear this flag; might get set in distribute_qual_to_rels */
 278         root->hasPseudoConstantQuals = false;
 279
 280         /*
 281          * Do expression preprocessing on targetlist and quals.
 282          */
 283         parse->targetList = (List *)
 284                 preprocess_expression(root, (Node *) parse->targetList,
 285                                                           EXPRKIND_TARGET);
 286
 287         parse->returningList = (List *)
 288                 preprocess_expression(root, (Node *) parse->returningList,
 289                                                           EXPRKIND_TARGET);
 290
 291         preprocess_qual_conditions(root, (Node *) parse->jointree);
 292
 293         parse->havingQual = preprocess_expression(root, parse->havingQual,
 294                                                                                           EXPRKIND_QUAL);
 295
 296         parse->limitOffset = preprocess_expression(root, parse->limitOffset,
 297                                                                                            EXPRKIND_LIMIT);
 298         parse->limitCount = preprocess_expression(root, parse->limitCount,
 299                                                                                           EXPRKIND_LIMIT);
 300
 301         root->in_info_list = (List *)
 302                 preprocess_expression(root, (Node *) root->in_info_list,
 303                                                           EXPRKIND_ININFO);
 304         root->append_rel_list = (List *)
 305                 preprocess_expression(root, (Node *) root->append_rel_list,
 306                                                           EXPRKIND_APPINFO);
 307
 308         /* Also need to preprocess expressions for function and values RTEs */
 309         foreach(l, parse->rtable)
 310         {
 311                 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
 312
 313                 if (rte->rtekind == RTE_FUNCTION)
 314                         rte->funcexpr = preprocess_expression(root, rte->funcexpr,
 315                                                                                                   EXPRKIND_RTFUNC);
 316                 else if (rte->rtekind == RTE_VALUES)
 317                         rte->values_lists = (List *)
 318                                 preprocess_expression(root, (Node *) rte->values_lists,
 319                                                                           EXPRKIND_VALUES);
 320         }
 321
 322         /*
 323          * In some cases we may want to transfer a HAVING clause into WHERE. We
 324          * cannot do so if the HAVING clause contains aggregates (obviously) or
 325          * volatile functions (since a HAVING clause is supposed to be executed
 326          * only once per group).  Also, it may be that the clause is so expensive
 327          * to execute that we're better off doing it only once per group, despite
 328          * the loss of selectivity.  This is hard to estimate short of doing the
 329          * entire planning process twice, so we use a heuristic: clauses
 330          * containing subplans are left in HAVING.      Otherwise, we move or copy the
 331          * HAVING clause into WHERE, in hopes of eliminating tuples before
 332          * aggregation instead of after.
 333          *
 334          * If the query has explicit grouping then we can simply move such a
 335          * clause into WHERE; any group that fails the clause will not be in the
 336          * output because none of its tuples will reach the grouping or
 337          * aggregation stage.  Otherwise we must have a degenerate (variable-free)
 338          * HAVING clause, which we put in WHERE so that query_planner() can use it
 339          * in a gating Result node, but also keep in HAVING to ensure that we
 340          * don't emit a bogus aggregated row. (This could be done better, but it
 341          * seems not worth optimizing.)
 342          *
 343          * Note that both havingQual and parse->jointree->quals are in
 344          * implicitly-ANDed-list form at this point, even though they are declared
 345          * as Node *.
 346          */
 347         newHaving = NIL;
 348         foreach(l, (List *) parse->havingQual)
 349         {
 350                 Node       *havingclause = (Node *) lfirst(l);
 351
 352                 if (contain_agg_clause(havingclause) ||
 353                         contain_volatile_functions(havingclause) ||
 354                         contain_subplans(havingclause))
 355                 {
 356                         /* keep it in HAVING */
 357                         newHaving = lappend(newHaving, havingclause);
 358                 }
 359                 else if (parse->groupClause)
 360                 {
 361                         /* move it to WHERE */
 362                         parse->jointree->quals = (Node *)
 363                                 lappend((List *) parse->jointree->quals, havingclause);
 364                 }
 365                 else
 366                 {
 367                         /* put a copy in WHERE, keep it in HAVING */
 368                         parse->jointree->quals = (Node *)
 369                                 lappend((List *) parse->jointree->quals,
 370                                                 copyObject(havingclause));
 371                         newHaving = lappend(newHaving, havingclause);
 372                 }
 373         }
 374         parse->havingQual = (Node *) newHaving;
 375
 376         /*
 377          * If we have any outer joins, try to reduce them to plain inner joins.
 378          * This step is most easily done after we've done expression
 379          * preprocessing.
 380          */
 381         if (root->hasOuterJoins)
 382                 reduce_outer_joins(root);
 383
 384         /*
 385          * Do the main planning.  If we have an inherited target relation, that
 386          * needs special processing, else go straight to grouping_planner.
 387          */
 388         if (parse->resultRelation &&
 389                 rt_fetch(parse->resultRelation, parse->rtable)->inh)
 390                 plan = inheritance_planner(root);
 391         else
 392                 plan = grouping_planner(root, tuple_fraction);
 393
 394         /*
 395          * If any subplans were generated, or if we're inside a subplan, build
 396          * initPlan list and extParam/allParam sets for plan nodes, and attach the
 397          * initPlans to the top plan node.
 398          */
 399         if (PlannerPlanId != saved_planid || PlannerQueryLevel > 1)
 400                 SS_finalize_plan(plan, parse->rtable);
 401
 402         /* Return sort ordering info if caller wants it */
 403         if (subquery_pathkeys)
 404                 *subquery_pathkeys = root->query_pathkeys;
 405
 406         /* Return to outer subquery context */
 407         PlannerQueryLevel--;
 408         PlannerInitPlan = saved_initplan;
 409         /* we do NOT restore PlannerPlanId; that's not an oversight! */
 410
 411         return plan;
 412 }
 413
 414 /*
 415  * preprocess_expression
 416  *              Do subquery_planner's preprocessing work for an expression,
 417  *              which can be a targetlist, a WHERE clause (including JOIN/ON
 418  *              conditions), or a HAVING clause.
 419  */
 420 static Node *
 421 preprocess_expression(PlannerInfo *root, Node *expr, int kind)
 422 {
 423         /*
 424          * Fall out quickly if expression is empty.  This occurs often enough to
 425          * be worth checking.  Note that null->null is the correct conversion for
 426          * implicit-AND result format, too.
 427          */
 428         if (expr == NULL)
 429                 return NULL;
 430
 431         /*
 432          * If the query has any join RTEs, replace join alias variables with
 433          * base-relation variables. We must do this before sublink processing,
 434          * else sublinks expanded out from join aliases wouldn't get processed. We
 435          * can skip it in VALUES lists, however, since they can't contain any Vars
 436          * at all.
 437          */
 438         if (root->hasJoinRTEs && kind != EXPRKIND_VALUES)
 439                 expr = flatten_join_alias_vars(root, expr);
 440
 441         /*
 442          * Simplify constant expressions.
 443          *
 444          * Note: this also flattens nested AND and OR expressions into N-argument
 445          * form.  All processing of a qual expression after this point must be
 446          * careful to maintain AND/OR flatness --- that is, do not generate a tree
 447          * with AND directly under AND, nor OR directly under OR.
 448          *
 449          * Because this is a relatively expensive process, we skip it when the
 450          * query is trivial, such as "SELECT 2+2;" or "INSERT ... VALUES()". The
 451          * expression will only be evaluated once anyway, so no point in
 452          * pre-simplifying; we can't execute it any faster than the executor can,
 453          * and we will waste cycles copying the tree.  Notice however that we
 454          * still must do it for quals (to get AND/OR flatness); and if we are in a
 455          * subquery we should not assume it will be done only once.
 456          *
 457          * For VALUES lists we never do this at all, again on the grounds that we
 458          * should optimize for one-time evaluation.
 459          */
 460         if (kind != EXPRKIND_VALUES &&
 461                 (root->parse->jointree->fromlist != NIL ||
 462                  kind == EXPRKIND_QUAL ||
 463                  PlannerQueryLevel > 1))
 464                 expr = eval_const_expressions(expr);
 465
 466         /*
 467          * If it's a qual or havingQual, canonicalize it.
 468          */
 469         if (kind == EXPRKIND_QUAL)
 470         {
 471                 expr = (Node *) canonicalize_qual((Expr *) expr);
 472
 473 #ifdef OPTIMIZER_DEBUG
 474                 printf("After canonicalize_qual()\n");
 475                 pprint(expr);
 476 #endif
 477         }
 478
 479         /* Expand SubLinks to SubPlans */
 480         if (root->parse->hasSubLinks)
 481                 expr = SS_process_sublinks(expr, (kind == EXPRKIND_QUAL));
 482
 483         /*
 484          * XXX do not insert anything here unless you have grokked the comments in
 485          * SS_replace_correlation_vars ...
 486          */
 487
 488         /* Replace uplevel vars with Param nodes (this IS possible in VALUES) */
 489         if (PlannerQueryLevel > 1)
 490                 expr = SS_replace_correlation_vars(expr);
 491
 492         /*
 493          * If it's a qual or havingQual, convert it to implicit-AND format. (We
 494          * don't want to do this before eval_const_expressions, since the latter
 495          * would be unable to simplify a top-level AND correctly. Also,
 496          * SS_process_sublinks expects explicit-AND format.)
 497          */
 498         if (kind == EXPRKIND_QUAL)
 499                 expr = (Node *) make_ands_implicit((Expr *) expr);
 500
 501         return expr;
 502 }
 503
 504 /*
 505  * preprocess_qual_conditions
 506  *              Recursively scan the query's jointree and do subquery_planner's
 507  *              preprocessing work on each qual condition found therein.
 508  */
 509 static void
 510 preprocess_qual_conditions(PlannerInfo *root, Node *jtnode)
 511 {
 512         if (jtnode == NULL)
 513                 return;
 514         if (IsA(jtnode, RangeTblRef))
 515         {
 516                 /* nothing to do here */
 517         }
 518         else if (IsA(jtnode, FromExpr))
 519         {
 520                 FromExpr   *f = (FromExpr *) jtnode;
 521                 ListCell   *l;
 522
 523                 foreach(l, f->fromlist)
 524                         preprocess_qual_conditions(root, lfirst(l));
 525
 526                 f->quals = preprocess_expression(root, f->quals, EXPRKIND_QUAL);
 527         }
 528         else if (IsA(jtnode, JoinExpr))
 529         {
 530                 JoinExpr   *j = (JoinExpr *) jtnode;
 531
 532                 preprocess_qual_conditions(root, j->larg);
 533                 preprocess_qual_conditions(root, j->rarg);
 534
 535                 j->quals = preprocess_expression(root, j->quals, EXPRKIND_QUAL);
 536         }
 537         else
 538                 elog(ERROR, "unrecognized node type: %d",
 539                          (int) nodeTag(jtnode));
 540 }
 541
 542 /*
 543  * inheritance_planner
 544  *        Generate a plan in the case where the result relation is an
 545  *        inheritance set.
 546  *
 547  * We have to handle this case differently from cases where a source relation
 548  * is an inheritance set. Source inheritance is expanded at the bottom of the
 549  * plan tree (see allpaths.c), but target inheritance has to be expanded at
 550  * the top.  The reason is that for UPDATE, each target relation needs a
 551  * different targetlist matching its own column set.  Also, for both UPDATE
 552  * and DELETE, the executor needs the Append plan node at the top, else it
 553  * can't keep track of which table is the current target table.  Fortunately,
 554  * the UPDATE/DELETE target can never be the nullable side of an outer join,
 555  * so it's OK to generate the plan this way.
 556  *
 557  * Returns a query plan.
 558  */
 559 static Plan *
 560 inheritance_planner(PlannerInfo *root)
 561 {
 562         Query      *parse = root->parse;
 563         int                     parentRTindex = parse->resultRelation;
 564         List       *subplans = NIL;
 565         List       *resultRelations = NIL;
 566         List       *returningLists = NIL;
 567         List       *rtable = NIL;
 568         List       *tlist = NIL;
 569         PlannerInfo subroot;
 570         ListCell   *l;
 571
 572         foreach(l, root->append_rel_list)
 573         {
 574                 AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
 575                 Plan       *subplan;
 576
 577                 /* append_rel_list contains all append rels; ignore others */
 578                 if (appinfo->parent_relid != parentRTindex)
 579                         continue;
 580
 581                 /*
 582                  * Generate modified query with this rel as target.  We have to be
 583                  * prepared to translate varnos in in_info_list as well as in the
 584                  * Query proper.
 585                  */
 586                 memcpy(&subroot, root, sizeof(PlannerInfo));
 587                 subroot.parse = (Query *)
 588                         adjust_appendrel_attrs((Node *) parse,
 589                                                                    appinfo);
 590                 subroot.in_info_list = (List *)
 591                         adjust_appendrel_attrs((Node *) root->in_info_list,
 592                                                                    appinfo);
 593                 /* There shouldn't be any OJ info to translate, as yet */
 594                 Assert(subroot.oj_info_list == NIL);
 595
 596                 /* Generate plan */
 597                 subplan = grouping_planner(&subroot, 0.0 /* retrieve all tuples */ );
 598
 599                 /*
 600                  * If this child rel was excluded by constraint exclusion, exclude it
 601                  * from the plan.
 602                  */
 603                 if (is_dummy_plan(subplan))
 604                         continue;
 605
 606                 /* Save rtable and tlist from first rel for use below */
 607                 if (subplans == NIL)
 608                 {
 609                         rtable = subroot.parse->rtable;
 610                         tlist = subplan->targetlist;
 611                 }
 612
 613                 subplans = lappend(subplans, subplan);
 614
 615                 /* Build target-relations list for the executor */
 616                 resultRelations = lappend_int(resultRelations, appinfo->child_relid);
 617
 618                 /* Build list of per-relation RETURNING targetlists */
 619                 if (parse->returningList)
 620                 {
 621                         Assert(list_length(subroot.parse->returningLists) == 1);
 622                         returningLists = list_concat(returningLists,
 623                                                                                  subroot.parse->returningLists);
 624                 }
 625         }
 626
 627         parse->resultRelations = resultRelations;
 628         parse->returningLists = returningLists;
 629
 630         /* Mark result as unordered (probably unnecessary) */
 631         root->query_pathkeys = NIL;
 632
 633         /*
 634          * If we managed to exclude every child rel, return a dummy plan
 635          */
 636         if (subplans == NIL)
 637                 return (Plan *) make_result(tlist,
 638                                                                         (Node *) list_make1(makeBoolConst(false,
 639                                                                                                                                           false)),
 640                                                                         NULL);
 641
 642         /*
 643          * Planning might have modified the rangetable, due to changes of the
 644          * Query structures inside subquery RTEs.  We have to ensure that this
 645          * gets propagated back to the master copy.  But can't do this until we
 646          * are done planning, because all the calls to grouping_planner need
 647          * virgin sub-Queries to work from.  (We are effectively assuming that
 648          * sub-Queries will get planned identically each time, or at least that
 649          * the impacts on their rangetables will be the same each time.)
 650          *
 651          * XXX should clean this up someday
 652          */
 653         parse->rtable = rtable;
 654
 655         return (Plan *) make_append(subplans, true, tlist);
 656 }
 657
 658 /*--------------------
 659  * grouping_planner
 660  *        Perform planning steps related to grouping, aggregation, etc.
 661  *        This primarily means adding top-level processing to the basic
 662  *        query plan produced by query_planner.
 663  *
 664  * tuple_fraction is the fraction of tuples we expect will be retrieved
 665  *
 666  * tuple_fraction is interpreted as follows:
 667  *        0: expect all tuples to be retrieved (normal case)
 668  *        0 < tuple_fraction < 1: expect the given fraction of tuples available
 669  *              from the plan to be retrieved
 670  *        tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
 671  *              expected to be retrieved (ie, a LIMIT specification)
 672  *
 673  * Returns a query plan.  Also, root->query_pathkeys is returned as the
 674  * actual output ordering of the plan (in pathkey format).
 675  *--------------------
 676  */
 677 static Plan *
 678 grouping_planner(PlannerInfo *root, double tuple_fraction)
 679 {
 680         Query      *parse = root->parse;
 681         List       *tlist = parse->targetList;
 682         int64           offset_est = 0;
 683         int64           count_est = 0;
 684         Plan       *result_plan;
 685         List       *current_pathkeys;
 686         List       *sort_pathkeys;
 687         double          dNumGroups = 0;
 688
 689         /* Tweak caller-supplied tuple_fraction if have LIMIT/OFFSET */
 690         if (parse->limitCount || parse->limitOffset)
 691                 tuple_fraction = preprocess_limit(root, tuple_fraction,
 692                                                                                   &offset_est, &count_est);
 693
 694         if (parse->setOperations)
 695         {
 696                 List       *set_sortclauses;
 697
 698                 /*
 699                  * If there's a top-level ORDER BY, assume we have to fetch all the
 700                  * tuples.      This might seem too simplistic given all the hackery below
 701                  * to possibly avoid the sort ... but a nonzero tuple_fraction is only
 702                  * of use to plan_set_operations() when the setop is UNION ALL, and
 703                  * the result of UNION ALL is always unsorted.
 704                  */
 705                 if (parse->sortClause)
 706                         tuple_fraction = 0.0;
 707
 708                 /*
 709                  * Construct the plan for set operations.  The result will not need
 710                  * any work except perhaps a top-level sort and/or LIMIT.
 711                  */
 712                 result_plan = plan_set_operations(root, tuple_fraction,
 713                                                                                   &set_sortclauses);
 714
 715                 /*
 716                  * Calculate pathkeys representing the sort order (if any) of the set
 717                  * operation's result.  We have to do this before overwriting the sort
 718                  * key information...
 719                  */
 720                 current_pathkeys = make_pathkeys_for_sortclauses(root,
 721                                                                                                                  set_sortclauses,
 722                                                                                                         result_plan->targetlist,
 723                                                                                                                  true);
 724
 725                 /*
 726                  * We should not need to call preprocess_targetlist, since we must be
 727                  * in a SELECT query node.      Instead, use the targetlist returned by
 728                  * plan_set_operations (since this tells whether it returned any
 729                  * resjunk columns!), and transfer any sort key information from the
 730                  * original tlist.
 731                  */
 732                 Assert(parse->commandType == CMD_SELECT);
 733
 734                 tlist = postprocess_setop_tlist(result_plan->targetlist, tlist);
 735
 736                 /*
 737                  * Can't handle FOR UPDATE/SHARE here (parser should have checked
 738                  * already, but let's make sure).
 739                  */
 740                 if (parse->rowMarks)
 741                         ereport(ERROR,
 742                                         (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
 743                                          errmsg("SELECT FOR UPDATE/SHARE is not allowed with UNION/INTERSECT/EXCEPT")));
 744
 745                 /*
 746                  * Calculate pathkeys that represent result ordering requirements
 747                  */
 748                 sort_pathkeys = make_pathkeys_for_sortclauses(root,
 749                                                                                                           parse->sortClause,
 750                                                                                                           tlist,
 751                                                                                                           true);
 752         }
 753         else
 754         {
 755                 /* No set operations, do regular planning */
 756                 List       *sub_tlist;
 757                 List       *group_pathkeys;
 758                 AttrNumber *groupColIdx = NULL;
 759                 Oid                *groupOperators = NULL;
 760                 bool            need_tlist_eval = true;
 761                 QualCost        tlist_cost;
 762                 Path       *cheapest_path;
 763                 Path       *sorted_path;
 764                 Path       *best_path;
 765                 long            numGroups = 0;
 766                 AggClauseCounts agg_counts;
 767                 int                     numGroupCols = list_length(parse->groupClause);
 768                 bool            use_hashed_grouping = false;
 769
 770                 MemSet(&agg_counts, 0, sizeof(AggClauseCounts));
 771
 772                 /* Preprocess targetlist */
 773                 tlist = preprocess_targetlist(root, tlist);
 774
 775                 /*
 776                  * Generate appropriate target list for subplan; may be different from
 777                  * tlist if grouping or aggregation is needed.
 778                  */
 779                 sub_tlist = make_subplanTargetList(root, tlist,
 780                                                                                    &groupColIdx, &need_tlist_eval);
 781
 782                 /*
 783                  * Calculate pathkeys that represent grouping/ordering requirements.
 784                  * Stash them in PlannerInfo so that query_planner can canonicalize
 785                  * them after EquivalenceClasses have been formed.
 786                  */
 787                 root->group_pathkeys =
 788                         make_pathkeys_for_sortclauses(root,
 789                                                                                   parse->groupClause,
 790                                                                                   tlist,
 791                                                                                   false);
 792                 root->sort_pathkeys =
 793                         make_pathkeys_for_sortclauses(root,
 794                                                                                   parse->sortClause,
 795                                                                                   tlist,
 796                                                                                   false);
 797
 798                 /*
 799                  * Will need actual number of aggregates for estimating costs.
 800                  *
 801                  * Note: we do not attempt to detect duplicate aggregates here; a
 802                  * somewhat-overestimated count is okay for our present purposes.
 803                  *
 804                  * Note: think not that we can turn off hasAggs if we find no aggs. It
 805                  * is possible for constant-expression simplification to remove all
 806                  * explicit references to aggs, but we still have to follow the
 807                  * aggregate semantics (eg, producing only one output row).
 808                  */
 809                 if (parse->hasAggs)
 810                 {
 811                         count_agg_clauses((Node *) tlist, &agg_counts);
 812                         count_agg_clauses(parse->havingQual, &agg_counts);
 813                 }
 814
 815                 /*
 816                  * Figure out whether we need a sorted result from query_planner.
 817                  *
 818                  * If we have a GROUP BY clause, then we want a result sorted properly
 819                  * for grouping.  Otherwise, if there is an ORDER BY clause, we want
 820                  * to sort by the ORDER BY clause.      (Note: if we have both, and ORDER
 821                  * BY is a superset of GROUP BY, it would be tempting to request sort
 822                  * by ORDER BY --- but that might just leave us failing to exploit an
 823                  * available sort order at all. Needs more thought...)
 824                  */
 825                 if (parse->groupClause)
 826                         root->query_pathkeys = root->group_pathkeys;
 827                 else if (parse->sortClause)
 828                         root->query_pathkeys = root->sort_pathkeys;
 829                 else
 830                         root->query_pathkeys = NIL;
 831
 832                 /*
 833                  * Generate the best unsorted and presorted paths for this Query (but
 834                  * note there may not be any presorted path).  query_planner will also
 835                  * estimate the number of groups in the query, and canonicalize all
 836                  * the pathkeys.
 837                  */
 838                 query_planner(root, sub_tlist, tuple_fraction,
 839                                           &cheapest_path, &sorted_path, &dNumGroups);
 840
 841                 group_pathkeys = root->group_pathkeys;
 842                 sort_pathkeys = root->sort_pathkeys;
 843
 844                 /*
 845                  * If grouping, extract the grouping operators and decide whether we
 846                  * want to use hashed grouping.
 847                  */
 848                 if (parse->groupClause)
 849                 {
 850                         groupOperators = extract_grouping_ops(parse->groupClause);
 851                         use_hashed_grouping =
 852                                 choose_hashed_grouping(root, tuple_fraction,
 853                                                                            cheapest_path, sorted_path,
 854                                                                            groupOperators, dNumGroups,
 855                                                                            &agg_counts);
 856
 857                         /* Also convert # groups to long int --- but 'ware overflow! */
 858                         numGroups = (long) Min(dNumGroups, (double) LONG_MAX);
 859                 }
 860
 861                 /*
 862                  * Select the best path.  If we are doing hashed grouping, we will
 863                  * always read all the input tuples, so use the cheapest-total path.
 864                  * Otherwise, trust query_planner's decision about which to use.
 865                  */
 866                 if (use_hashed_grouping || !sorted_path)
 867                         best_path = cheapest_path;
 868                 else
 869                         best_path = sorted_path;
 870
 871                 /*
 872                  * Check to see if it's possible to optimize MIN/MAX aggregates. If
 873                  * so, we will forget all the work we did so far to choose a "regular"
 874                  * path ... but we had to do it anyway to be able to tell which way is
 875                  * cheaper.
 876                  */
 877                 result_plan = optimize_minmax_aggregates(root,
 878                                                                                                  tlist,
 879                                                                                                  best_path);
 880                 if (result_plan != NULL)
 881                 {
 882                         /*
 883                          * optimize_minmax_aggregates generated the full plan, with the
 884                          * right tlist, and it has no sort order.
 885                          */
 886                         current_pathkeys = NIL;
 887                 }
 888                 else
 889                 {
 890                         /*
 891                          * Normal case --- create a plan according to query_planner's
 892                          * results.
 893                          */
 894                         result_plan = create_plan(root, best_path);
 895                         current_pathkeys = best_path->pathkeys;
 896
 897                         /*
 898                          * create_plan() returns a plan with just a "flat" tlist of
 899                          * required Vars.  Usually we need to insert the sub_tlist as the
 900                          * tlist of the top plan node.  However, we can skip that if we
 901                          * determined that whatever query_planner chose to return will be
 902                          * good enough.
 903                          */
 904                         if (need_tlist_eval)
 905                         {
 906                                 /*
 907                                  * If the top-level plan node is one that cannot do expression
 908                                  * evaluation, we must insert a Result node to project the
 909                                  * desired tlist.
 910                                  */
 911                                 if (!is_projection_capable_plan(result_plan))
 912                                 {
 913                                         result_plan = (Plan *) make_result(sub_tlist, NULL,
 914                                                                                                            result_plan);
 915                                 }
 916                                 else
 917                                 {
 918                                         /*
 919                                          * Otherwise, just replace the subplan's flat tlist with
 920                                          * the desired tlist.
 921                                          */
 922                                         result_plan->targetlist = sub_tlist;
 923                                 }
 924
 925                                 /*
 926                                  * Also, account for the cost of evaluation of the sub_tlist.
 927                                  *
 928                                  * Up to now, we have only been dealing with "flat" tlists,
 929                                  * containing just Vars.  So their evaluation cost is zero
 930                                  * according to the model used by cost_qual_eval() (or if you
 931                                  * prefer, the cost is factored into cpu_tuple_cost).  Thus we
 932                                  * can avoid accounting for tlist cost throughout
 933                                  * query_planner() and subroutines.  But now we've inserted a
 934                                  * tlist that might contain actual operators, sub-selects, etc
 935                                  * --- so we'd better account for its cost.
 936                                  *
 937                                  * Below this point, any tlist eval cost for added-on nodes
 938                                  * should be accounted for as we create those nodes.
 939                                  * Presently, of the node types we can add on, only Agg and
 940                                  * Group project new tlists (the rest just copy their input
 941                                  * tuples) --- so make_agg() and make_group() are responsible
 942                                  * for computing the added cost.
 943                                  */
 944                                 cost_qual_eval(&tlist_cost, sub_tlist);
 945                                 result_plan->startup_cost += tlist_cost.startup;
 946                                 result_plan->total_cost += tlist_cost.startup +
 947                                         tlist_cost.per_tuple * result_plan->plan_rows;
 948                         }
 949                         else
 950                         {
 951                                 /*
 952                                  * Since we're using query_planner's tlist and not the one
 953                                  * make_subplanTargetList calculated, we have to refigure any
 954                                  * grouping-column indexes make_subplanTargetList computed.
 955                                  */
 956                                 locate_grouping_columns(root, tlist, result_plan->targetlist,
 957                                                                                 groupColIdx);
 958                         }
 959
 960                         /*
 961                          * Insert AGG or GROUP node if needed, plus an explicit sort step
 962                          * if necessary.
 963                          *
 964                          * HAVING clause, if any, becomes qual of the Agg or Group node.
 965                          */
 966                         if (use_hashed_grouping)
 967                         {
 968                                 /* Hashed aggregate plan --- no sort needed */
 969                                 result_plan = (Plan *) make_agg(root,
 970                                                                                                 tlist,
 971                                                                                                 (List *) parse->havingQual,
 972                                                                                                 AGG_HASHED,
 973                                                                                                 numGroupCols,
 974                                                                                                 groupColIdx,
 975                                                                                                 groupOperators,
 976                                                                                                 numGroups,
 977                                                                                                 agg_counts.numAggs,
 978                                                                                                 result_plan);
 979                                 /* Hashed aggregation produces randomly-ordered results */
 980                                 current_pathkeys = NIL;
 981                         }
 982                         else if (parse->hasAggs)
 983                         {
 984                                 /* Plain aggregate plan --- sort if needed */
 985                                 AggStrategy aggstrategy;
 986
 987                                 if (parse->groupClause)
 988                                 {
 989                                         if (!pathkeys_contained_in(group_pathkeys,
 990                                                                                            current_pathkeys))
 991                                         {
 992                                                 result_plan = (Plan *)
 993                                                         make_sort_from_groupcols(root,
 994                                                                                                          parse->groupClause,
 995                                                                                                          groupColIdx,
 996                                                                                                          result_plan);
 997                                                 current_pathkeys = group_pathkeys;
 998                                         }
 999                                         aggstrategy = AGG_SORTED;
1000
1001                                         /*
1002                                          * The AGG node will not change the sort ordering of its
1003                                          * groups, so current_pathkeys describes the result too.
1004                                          */
1005                                 }
1006                                 else
1007                                 {
1008                                         aggstrategy = AGG_PLAIN;
1009                                         /* Result will be only one row anyway; no sort order */
1010                                         current_pathkeys = NIL;
1011                                 }
1012
1013                                 result_plan = (Plan *) make_agg(root,
1014                                                                                                 tlist,
1015                                                                                                 (List *) parse->havingQual,
1016                                                                                                 aggstrategy,
1017                                                                                                 numGroupCols,
1018                                                                                                 groupColIdx,
1019                                                                                                 groupOperators,
1020                                                                                                 numGroups,
1021                                                                                                 agg_counts.numAggs,
1022                                                                                                 result_plan);
1023                         }
1024                         else if (parse->groupClause)
1025                         {
1026                                 /*
1027                                  * GROUP BY without aggregation, so insert a group node (plus
1028                                  * the appropriate sort node, if necessary).
1029                                  *
1030                                  * Add an explicit sort if we couldn't make the path come out
1031                                  * the way the GROUP node needs it.
1032                                  */
1033                                 if (!pathkeys_contained_in(group_pathkeys, current_pathkeys))
1034                                 {
1035                                         result_plan = (Plan *)
1036                                                 make_sort_from_groupcols(root,
1037                                                                                                  parse->groupClause,
1038                                                                                                  groupColIdx,
1039                                                                                                  result_plan);
1040                                         current_pathkeys = group_pathkeys;
1041                                 }
1042
1043                                 result_plan = (Plan *) make_group(root,
1044                                                                                                   tlist,
1045                                                                                                   (List *) parse->havingQual,
1046                                                                                                   numGroupCols,
1047                                                                                                   groupColIdx,
1048                                                                                                   groupOperators,
1049                                                                                                   dNumGroups,
1050                                                                                                   result_plan);
1051                                 /* The Group node won't change sort ordering */
1052                         }
1053                         else if (root->hasHavingQual)
1054                         {
1055                                 /*
1056                                  * No aggregates, and no GROUP BY, but we have a HAVING qual.
1057                                  * This is a degenerate case in which we are supposed to emit
1058                                  * either 0 or 1 row depending on whether HAVING succeeds.
1059                                  * Furthermore, there cannot be any variables in either HAVING
1060                                  * or the targetlist, so we actually do not need the FROM
1061                                  * table at all!  We can just throw away the plan-so-far and
1062                                  * generate a Result node.      This is a sufficiently unusual
1063                                  * corner case that it's not worth contorting the structure of
1064                                  * this routine to avoid having to generate the plan in the
1065                                  * first place.
1066                                  */
1067                                 result_plan = (Plan *) make_result(tlist,
1068                                                                                                    parse->havingQual,
1069                                                                                                    NULL);
1070                         }
1071                 }                                               /* end of non-minmax-aggregate case */
1072         }                                                       /* end of if (setOperations) */
1073
1074         /*
1075          * If we were not able to make the plan come out in the right order, add
1076          * an explicit sort step.
1077          */
1078         if (parse->sortClause)
1079         {
1080                 if (!pathkeys_contained_in(sort_pathkeys, current_pathkeys))
1081                 {
1082                         result_plan = (Plan *) make_sort_from_pathkeys(root,
1083                                                                                                                    result_plan,
1084                                                                                                                    sort_pathkeys);
1085                         current_pathkeys = sort_pathkeys;
1086                 }
1087         }
1088
1089         /*
1090          * If there is a DISTINCT clause, add the UNIQUE node.
1091          */
1092         if (parse->distinctClause)
1093         {
1094                 result_plan = (Plan *) make_unique(result_plan, parse->distinctClause);
1095
1096                 /*
1097                  * If there was grouping or aggregation, leave plan_rows as-is (ie,
1098                  * assume the result was already mostly unique).  If not, use the
1099                  * number of distinct-groups calculated by query_planner.
1100                  */
1101                 if (!parse->groupClause && !root->hasHavingQual && !parse->hasAggs)
1102                         result_plan->plan_rows = dNumGroups;
1103         }
1104
1105         /*
1106          * Finally, if there is a LIMIT/OFFSET clause, add the LIMIT node.
1107          */
1108         if (parse->limitCount || parse->limitOffset)
1109         {
1110                 result_plan = (Plan *) make_limit(result_plan,
1111                                                                                   parse->limitOffset,
1112                                                                                   parse->limitCount,
1113                                                                                   offset_est,
1114                                                                                   count_est);
1115         }
1116
1117         /*
1118          * Deal with the RETURNING clause if any.  It's convenient to pass the
1119          * returningList through setrefs.c now rather than at top level (if we
1120          * waited, handling inherited UPDATE/DELETE would be much harder).
1121          */
1122         if (parse->returningList)
1123         {
1124                 List       *rlist;
1125
1126                 rlist = set_returning_clause_references(parse->returningList,
1127                                                                                                 result_plan,
1128                                                                                                 parse->resultRelation);
1129                 parse->returningLists = list_make1(rlist);
1130         }
1131
1132         /*
1133          * Return the actual output ordering in query_pathkeys for possible use by
1134          * an outer query level.
1135          */
1136         root->query_pathkeys = current_pathkeys;
1137
1138         return result_plan;
1139 }
1140
1141 /*
1142  * Detect whether a plan node is a "dummy" plan created when a relation
1143  * is deemed not to need scanning due to constraint exclusion.
1144  *
1145  * Currently, such dummy plans are Result nodes with constant FALSE
1146  * filter quals.
1147  */
1148 static bool
1149 is_dummy_plan(Plan *plan)
1150 {
1151         if (IsA(plan, Result))
1152         {
1153                 List       *rcqual = (List *) ((Result *) plan)->resconstantqual;
1154
1155                 if (list_length(rcqual) == 1)
1156                 {
1157                         Const      *constqual = (Const *) linitial(rcqual);
1158
1159                         if (constqual && IsA(constqual, Const))
1160                         {
1161                                 if (!constqual->constisnull &&
1162                                         !DatumGetBool(constqual->constvalue))
1163                                         return true;
1164                         }
1165                 }
1166         }
1167         return false;
1168 }
1169
1170 /*
1171  * preprocess_limit - do pre-estimation for LIMIT and/or OFFSET clauses
1172  *
1173  * We try to estimate the values of the LIMIT/OFFSET clauses, and pass the
1174  * results back in *count_est and *offset_est.  These variables are set to
1175  * 0 if the corresponding clause is not present, and -1 if it's present
1176  * but we couldn't estimate the value for it.  (The "0" convention is OK
1177  * for OFFSET but a little bit bogus for LIMIT: effectively we estimate
1178  * LIMIT 0 as though it were LIMIT 1.  But this is in line with the planner's
1179  * usual practice of never estimating less than one row.)  These values will
1180  * be passed to make_limit, which see if you change this code.
1181  *
1182  * The return value is the suitably adjusted tuple_fraction to use for
1183  * planning the query.  This adjustment is not overridable, since it reflects
1184  * plan actions that grouping_planner() will certainly take, not assumptions
1185  * about context.
1186  */
1187 static double
1188 preprocess_limit(PlannerInfo *root, double tuple_fraction,
1189                                  int64 *offset_est, int64 *count_est)
1190 {
1191         Query      *parse = root->parse;
1192         Node       *est;
1193         double          limit_fraction;
1194
1195         /* Should not be called unless LIMIT or OFFSET */
1196         Assert(parse->limitCount || parse->limitOffset);
1197
1198         /*
1199          * Try to obtain the clause values.  We use estimate_expression_value
1200          * primarily because it can sometimes do something useful with Params.
1201          */
1202         if (parse->limitCount)
1203         {
1204                 est = estimate_expression_value(parse->limitCount);
1205                 if (est && IsA(est, Const))
1206                 {
1207                         if (((Const *) est)->constisnull)
1208                         {
1209                                 /* NULL indicates LIMIT ALL, ie, no limit */
1210                                 *count_est = 0; /* treat as not present */
1211                         }
1212                         else
1213                         {
1214                                 *count_est = DatumGetInt64(((Const *) est)->constvalue);
1215                                 if (*count_est <= 0)
1216                                         *count_est = 1;         /* force to at least 1 */
1217                         }
1218                 }
1219                 else
1220                         *count_est = -1;        /* can't estimate */
1221         }
1222         else
1223                 *count_est = 0;                 /* not present */
1224
1225         if (parse->limitOffset)
1226         {
1227                 est = estimate_expression_value(parse->limitOffset);
1228                 if (est && IsA(est, Const))
1229                 {
1230                         if (((Const *) est)->constisnull)
1231                         {
1232                                 /* Treat NULL as no offset; the executor will too */
1233                                 *offset_est = 0;        /* treat as not present */
1234                         }
1235                         else
1236                         {
1237                                 *offset_est = DatumGetInt64(((Const *) est)->constvalue);
1238                                 if (*offset_est < 0)
1239                                         *offset_est = 0;        /* less than 0 is same as 0 */
1240                         }
1241                 }
1242                 else
1243                         *offset_est = -1;       /* can't estimate */
1244         }
1245         else
1246                 *offset_est = 0;                /* not present */
1247
1248         if (*count_est != 0)
1249         {
1250                 /*
1251                  * A LIMIT clause limits the absolute number of tuples returned.
1252                  * However, if it's not a constant LIMIT then we have to guess; for
1253                  * lack of a better idea, assume 10% of the plan's result is wanted.
1254                  */
1255                 if (*count_est < 0 || *offset_est < 0)
1256                 {
1257                         /* LIMIT or OFFSET is an expression ... punt ... */
1258                         limit_fraction = 0.10;
1259                 }
1260                 else
1261                 {
1262                         /* LIMIT (plus OFFSET, if any) is max number of tuples needed */
1263                         limit_fraction = (double) *count_est + (double) *offset_est;
1264                 }
1265
1266                 /*
1267                  * If we have absolute limits from both caller and LIMIT, use the
1268                  * smaller value; likewise if they are both fractional.  If one is
1269                  * fractional and the other absolute, we can't easily determine which
1270                  * is smaller, but we use the heuristic that the absolute will usually
1271                  * be smaller.
1272                  */
1273                 if (tuple_fraction >= 1.0)
1274                 {
1275                         if (limit_fraction >= 1.0)
1276                         {
1277                                 /* both absolute */
1278                                 tuple_fraction = Min(tuple_fraction, limit_fraction);
1279                         }
1280                         else
1281                         {
1282                                 /* caller absolute, limit fractional; use caller's value */
1283                         }
1284                 }
1285                 else if (tuple_fraction > 0.0)
1286                 {
1287                         if (limit_fraction >= 1.0)
1288                         {
1289                                 /* caller fractional, limit absolute; use limit */
1290                                 tuple_fraction = limit_fraction;
1291                         }
1292                         else
1293                         {
1294                                 /* both fractional */
1295                                 tuple_fraction = Min(tuple_fraction, limit_fraction);
1296                         }
1297                 }
1298                 else
1299                 {
1300                         /* no info from caller, just use limit */
1301                         tuple_fraction = limit_fraction;
1302                 }
1303         }
1304         else if (*offset_est != 0 && tuple_fraction > 0.0)
1305         {
1306                 /*
1307                  * We have an OFFSET but no LIMIT.      This acts entirely differently
1308                  * from the LIMIT case: here, we need to increase rather than decrease
1309                  * the caller's tuple_fraction, because the OFFSET acts to cause more
1310                  * tuples to be fetched instead of fewer.  This only matters if we got
1311                  * a tuple_fraction > 0, however.
1312                  *
1313                  * As above, use 10% if OFFSET is present but unestimatable.
1314                  */
1315                 if (*offset_est < 0)
1316                         limit_fraction = 0.10;
1317                 else
1318                         limit_fraction = (double) *offset_est;
1319
1320                 /*
1321                  * If we have absolute counts from both caller and OFFSET, add them
1322                  * together; likewise if they are both fractional.      If one is
1323                  * fractional and the other absolute, we want to take the larger, and
1324                  * we heuristically assume that's the fractional one.
1325                  */
1326                 if (tuple_fraction >= 1.0)
1327                 {
1328                         if (limit_fraction >= 1.0)
1329                         {
1330                                 /* both absolute, so add them together */
1331                                 tuple_fraction += limit_fraction;
1332                         }
1333                         else
1334                         {
1335                                 /* caller absolute, limit fractional; use limit */
1336                                 tuple_fraction = limit_fraction;
1337                         }
1338                 }
1339                 else
1340                 {
1341                         if (limit_fraction >= 1.0)
1342                         {
1343                                 /* caller fractional, limit absolute; use caller's value */
1344                         }
1345                         else
1346                         {
1347                                 /* both fractional, so add them together */
1348                                 tuple_fraction += limit_fraction;
1349                                 if (tuple_fraction >= 1.0)
1350                                         tuple_fraction = 0.0;           /* assume fetch all */
1351                         }
1352                 }
1353         }
1354
1355         return tuple_fraction;
1356 }
1357
1358 /*
1359  * extract_grouping_ops - make an array of the equality operator OIDs
1360  *              for the GROUP BY clause
1361  */
1362 static Oid *
1363 extract_grouping_ops(List *groupClause)
1364 {
1365         int                     numCols = list_length(groupClause);
1366         int                     colno = 0;
1367         Oid                *groupOperators;
1368         ListCell   *glitem;
1369
1370         groupOperators = (Oid *) palloc(sizeof(Oid) * numCols);
1371
1372         foreach(glitem, groupClause)
1373         {
1374                 GroupClause *groupcl = (GroupClause *) lfirst(glitem);
1375
1376                 groupOperators[colno] = get_equality_op_for_ordering_op(groupcl->sortop);
1377                 if (!OidIsValid(groupOperators[colno]))         /* shouldn't happen */
1378                         elog(ERROR, "could not find equality operator for ordering operator %u",
1379                                  groupcl->sortop);
1380                 colno++;
1381         }
1382
1383         return groupOperators;
1384 }
1385
1386 /*
1387  * choose_hashed_grouping - should we use hashed grouping?
1388  */
1389 static bool
1390 choose_hashed_grouping(PlannerInfo *root, double tuple_fraction,
1391                                            Path *cheapest_path, Path *sorted_path,
1392                                            Oid *groupOperators, double dNumGroups,
1393                                            AggClauseCounts *agg_counts)
1394 {
1395         int                     numGroupCols = list_length(root->parse->groupClause);
1396         double          cheapest_path_rows;
1397         int                     cheapest_path_width;
1398         Size            hashentrysize;
1399         List       *current_pathkeys;
1400         Path            hashed_p;
1401         Path            sorted_p;
1402         int                     i;
1403
1404         /*
1405          * Check can't-do-it conditions, including whether the grouping operators
1406          * are hashjoinable.  (We assume hashing is OK if they are marked
1407          * oprcanhash.  If there isn't actually a supporting hash function,
1408          * the executor will complain at runtime.)
1409          *
1410          * Executor doesn't support hashed aggregation with DISTINCT aggregates.
1411          * (Doing so would imply storing *all* the input values in the hash table,
1412          * which seems like a certain loser.)
1413          */
1414         if (!enable_hashagg)
1415                 return false;
1416         if (agg_counts->numDistinctAggs != 0)
1417                 return false;
1418         for (i = 0; i < numGroupCols; i++)
1419         {
1420                 if (!op_hashjoinable(groupOperators[i]))
1421                         return false;
1422         }
1423
1424         /*
1425          * Don't do it if it doesn't look like the hashtable will fit into
1426          * work_mem.
1427          *
1428          * Beware here of the possibility that cheapest_path->parent is NULL. This
1429          * could happen if user does something silly like SELECT 'foo' GROUP BY 1;
1430          */
1431         if (cheapest_path->parent)
1432         {
1433                 cheapest_path_rows = cheapest_path->parent->rows;
1434                 cheapest_path_width = cheapest_path->parent->width;
1435         }
1436         else
1437         {
1438                 cheapest_path_rows = 1; /* assume non-set result */
1439                 cheapest_path_width = 100;              /* arbitrary */
1440         }
1441
1442         /* Estimate per-hash-entry space at tuple width... */
1443         hashentrysize = MAXALIGN(cheapest_path_width) + MAXALIGN(sizeof(MinimalTupleData));
1444         /* plus space for pass-by-ref transition values... */
1445         hashentrysize += agg_counts->transitionSpace;
1446         /* plus the per-hash-entry overhead */
1447         hashentrysize += hash_agg_entry_size(agg_counts->numAggs);
1448
1449         if (hashentrysize * dNumGroups > work_mem * 1024L)
1450                 return false;
1451
1452         /*
1453          * See if the estimated cost is no more than doing it the other way. While
1454          * avoiding the need for sorted input is usually a win, the fact that the
1455          * output won't be sorted may be a loss; so we need to do an actual cost
1456          * comparison.
1457          *
1458          * We need to consider cheapest_path + hashagg [+ final sort] versus
1459          * either cheapest_path [+ sort] + group or agg [+ final sort] or
1460          * presorted_path + group or agg [+ final sort] where brackets indicate a
1461          * step that may not be needed. We assume query_planner() will have
1462          * returned a presorted path only if it's a winner compared to
1463          * cheapest_path for this purpose.
1464          *
1465          * These path variables are dummies that just hold cost fields; we don't
1466          * make actual Paths for these steps.
1467          */
1468         cost_agg(&hashed_p, root, AGG_HASHED, agg_counts->numAggs,
1469                          numGroupCols, dNumGroups,
1470                          cheapest_path->startup_cost, cheapest_path->total_cost,
1471                          cheapest_path_rows);
1472         /* Result of hashed agg is always unsorted */
1473         if (root->sort_pathkeys)
1474                 cost_sort(&hashed_p, root, root->sort_pathkeys, hashed_p.total_cost,
1475                                   dNumGroups, cheapest_path_width);
1476
1477         if (sorted_path)
1478         {
1479                 sorted_p.startup_cost = sorted_path->startup_cost;
1480                 sorted_p.total_cost = sorted_path->total_cost;
1481                 current_pathkeys = sorted_path->pathkeys;
1482         }
1483         else
1484         {
1485                 sorted_p.startup_cost = cheapest_path->startup_cost;
1486                 sorted_p.total_cost = cheapest_path->total_cost;
1487                 current_pathkeys = cheapest_path->pathkeys;
1488         }
1489         if (!pathkeys_contained_in(root->group_pathkeys, current_pathkeys))
1490         {
1491                 cost_sort(&sorted_p, root, root->group_pathkeys, sorted_p.total_cost,
1492                                   cheapest_path_rows, cheapest_path_width);
1493                 current_pathkeys = root->group_pathkeys;
1494         }
1495
1496         if (root->parse->hasAggs)
1497                 cost_agg(&sorted_p, root, AGG_SORTED, agg_counts->numAggs,
1498                                  numGroupCols, dNumGroups,
1499                                  sorted_p.startup_cost, sorted_p.total_cost,
1500                                  cheapest_path_rows);
1501         else
1502                 cost_group(&sorted_p, root, numGroupCols, dNumGroups,
1503                                    sorted_p.startup_cost, sorted_p.total_cost,
1504                                    cheapest_path_rows);
1505         /* The Agg or Group node will preserve ordering */
1506         if (root->sort_pathkeys &&
1507                 !pathkeys_contained_in(root->sort_pathkeys, current_pathkeys))
1508                 cost_sort(&sorted_p, root, root->sort_pathkeys, sorted_p.total_cost,
1509                                   dNumGroups, cheapest_path_width);
1510
1511         /*
1512          * Now make the decision using the top-level tuple fraction.  First we
1513          * have to convert an absolute count (LIMIT) into fractional form.
1514          */
1515         if (tuple_fraction >= 1.0)
1516                 tuple_fraction /= dNumGroups;
1517
1518         if (compare_fractional_path_costs(&hashed_p, &sorted_p,
1519                                                                           tuple_fraction) < 0)
1520         {
1521                 /* Hashed is cheaper, so use it */
1522                 return true;
1523         }
1524         return false;
1525 }
1526
1527 /*---------------
1528  * make_subplanTargetList
1529  *        Generate appropriate target list when grouping is required.
1530  *
1531  * When grouping_planner inserts Aggregate, Group, or Result plan nodes
1532  * above the result of query_planner, we typically want to pass a different
1533  * target list to query_planner than the outer plan nodes should have.
1534  * This routine generates the correct target list for the subplan.
1535  *
1536  * The initial target list passed from the parser already contains entries
1537  * for all ORDER BY and GROUP BY expressions, but it will not have entries
1538  * for variables used only in HAVING clauses; so we need to add those
1539  * variables to the subplan target list.  Also, we flatten all expressions
1540  * except GROUP BY items into their component variables; the other expressions
1541  * will be computed by the inserted nodes rather than by the subplan.
1542  * For example, given a query like
1543  *              SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
1544  * we want to pass this targetlist to the subplan:
1545  *              a,b,c,d,a+b
1546  * where the a+b target will be used by the Sort/Group steps, and the
1547  * other targets will be used for computing the final results.  (In the
1548  * above example we could theoretically suppress the a and b targets and
1549  * pass down only c,d,a+b, but it's not really worth the trouble to
1550  * eliminate simple var references from the subplan.  We will avoid doing
1551  * the extra computation to recompute a+b at the outer level; see
1552  * replace_vars_with_subplan_refs() in setrefs.c.)
1553  *
1554  * If we are grouping or aggregating, *and* there are no non-Var grouping
1555  * expressions, then the returned tlist is effectively dummy; we do not
1556  * need to force it to be evaluated, because all the Vars it contains
1557  * should be present in the output of query_planner anyway.
1558  *
1559  * 'tlist' is the query's target list.
1560  * 'groupColIdx' receives an array of column numbers for the GROUP BY
1561  *                      expressions (if there are any) in the subplan's target list.
1562  * 'need_tlist_eval' is set true if we really need to evaluate the
1563  *                      result tlist.
1564  *
1565  * The result is the targetlist to be passed to the subplan.
1566  *---------------
1567  */
1568 static List *
1569 make_subplanTargetList(PlannerInfo *root,
1570                                            List *tlist,
1571                                            AttrNumber **groupColIdx,
1572                                            bool *need_tlist_eval)
1573 {
1574         Query      *parse = root->parse;
1575         List       *sub_tlist;
1576         List       *extravars;
1577         int                     numCols;
1578
1579         *groupColIdx = NULL;
1580
1581         /*
1582          * If we're not grouping or aggregating, there's nothing to do here;
1583          * query_planner should receive the unmodified target list.
1584          */
1585         if (!parse->hasAggs && !parse->groupClause && !root->hasHavingQual)
1586         {
1587                 *need_tlist_eval = true;
1588                 return tlist;
1589         }
1590
1591         /*
1592          * Otherwise, start with a "flattened" tlist (having just the vars
1593          * mentioned in the targetlist and HAVING qual --- but not upper- level
1594          * Vars; they will be replaced by Params later on).
1595          */
1596         sub_tlist = flatten_tlist(tlist);
1597         extravars = pull_var_clause(parse->havingQual, false);
1598         sub_tlist = add_to_flat_tlist(sub_tlist, extravars);
1599         list_free(extravars);
1600         *need_tlist_eval = false;       /* only eval if not flat tlist */
1601
1602         /*
1603          * If grouping, create sub_tlist entries for all GROUP BY expressions
1604          * (GROUP BY items that are simple Vars should be in the list already),
1605          * and make an array showing where the group columns are in the sub_tlist.
1606          */
1607         numCols = list_length(parse->groupClause);
1608         if (numCols > 0)
1609         {
1610                 int                     keyno = 0;
1611                 AttrNumber *grpColIdx;
1612                 ListCell   *gl;
1613
1614                 grpColIdx = (AttrNumber *) palloc(sizeof(AttrNumber) * numCols);
1615                 *groupColIdx = grpColIdx;
1616
1617                 foreach(gl, parse->groupClause)
1618                 {
1619                         GroupClause *grpcl = (GroupClause *) lfirst(gl);
1620                         Node       *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1621                         TargetEntry *te = NULL;
1622                         ListCell   *sl;
1623
1624                         /* Find or make a matching sub_tlist entry */
1625                         foreach(sl, sub_tlist)
1626                         {
1627                                 te = (TargetEntry *) lfirst(sl);
1628                                 if (equal(groupexpr, te->expr))
1629                                         break;
1630                         }
1631                         if (!sl)
1632                         {
1633                                 te = makeTargetEntry((Expr *) groupexpr,
1634                                                                          list_length(sub_tlist) + 1,
1635                                                                          NULL,
1636                                                                          false);
1637                                 sub_tlist = lappend(sub_tlist, te);
1638                                 *need_tlist_eval = true;                /* it's not flat anymore */
1639                         }
1640
1641                         /* and save its resno */
1642                         grpColIdx[keyno++] = te->resno;
1643                 }
1644         }
1645
1646         return sub_tlist;
1647 }
1648
1649 /*
1650  * locate_grouping_columns
1651  *              Locate grouping columns in the tlist chosen by query_planner.
1652  *
1653  * This is only needed if we don't use the sub_tlist chosen by
1654  * make_subplanTargetList.      We have to forget the column indexes found
1655  * by that routine and re-locate the grouping vars in the real sub_tlist.
1656  */
1657 static void
1658 locate_grouping_columns(PlannerInfo *root,
1659                                                 List *tlist,
1660                                                 List *sub_tlist,
1661                                                 AttrNumber *groupColIdx)
1662 {
1663         int                     keyno = 0;
1664         ListCell   *gl;
1665
1666         /*
1667          * No work unless grouping.
1668          */
1669         if (!root->parse->groupClause)
1670         {
1671                 Assert(groupColIdx == NULL);
1672                 return;
1673         }
1674         Assert(groupColIdx != NULL);
1675
1676         foreach(gl, root->parse->groupClause)
1677         {
1678                 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1679                 Node       *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1680                 TargetEntry *te = NULL;
1681                 ListCell   *sl;
1682
1683                 foreach(sl, sub_tlist)
1684                 {
1685                         te = (TargetEntry *) lfirst(sl);
1686                         if (equal(groupexpr, te->expr))
1687                                 break;
1688                 }
1689                 if (!sl)
1690                         elog(ERROR, "failed to locate grouping columns");
1691
1692                 groupColIdx[keyno++] = te->resno;
1693         }
1694 }
1695
1696 /*
1697  * postprocess_setop_tlist
1698  *        Fix up targetlist returned by plan_set_operations().
1699  *
1700  * We need to transpose sort key info from the orig_tlist into new_tlist.
1701  * NOTE: this would not be good enough if we supported resjunk sort keys
1702  * for results of set operations --- then, we'd need to project a whole
1703  * new tlist to evaluate the resjunk columns.  For now, just ereport if we
1704  * find any resjunk columns in orig_tlist.
1705  */
1706 static List *
1707 postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
1708 {
1709         ListCell   *l;
1710         ListCell   *orig_tlist_item = list_head(orig_tlist);
1711
1712         foreach(l, new_tlist)
1713         {
1714                 TargetEntry *new_tle = (TargetEntry *) lfirst(l);
1715                 TargetEntry *orig_tle;
1716
1717                 /* ignore resjunk columns in setop result */
1718                 if (new_tle->resjunk)
1719                         continue;
1720
1721                 Assert(orig_tlist_item != NULL);
1722                 orig_tle = (TargetEntry *) lfirst(orig_tlist_item);
1723                 orig_tlist_item = lnext(orig_tlist_item);
1724                 if (orig_tle->resjunk)  /* should not happen */
1725                         elog(ERROR, "resjunk output columns are not implemented");
1726                 Assert(new_tle->resno == orig_tle->resno);
1727                 new_tle->ressortgroupref = orig_tle->ressortgroupref;
1728         }
1729         if (orig_tlist_item != NULL)
1730                 elog(ERROR, "resjunk output columns are not implemented");
1731         return new_tlist;
1732 }