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 1.218 2007/04/27 22:05:47 tgl Exp $
  12  *
  13  *-------------------------------------------------------------------------
  14  */
  15
  16 #include "postgres.h"
  17
  18 #include <limits.h>
  19
  20 #include "catalog/pg_operator.h"
  21 #include "executor/executor.h"
  22 #include "executor/nodeAgg.h"
  23 #include "miscadmin.h"
  24 #include "nodes/makefuncs.h"
  25 #include "optimizer/clauses.h"
  26 #include "optimizer/cost.h"
  27 #include "optimizer/pathnode.h"
  28 #include "optimizer/paths.h"
  29 #include "optimizer/planmain.h"
  30 #include "optimizer/planner.h"
  31 #include "optimizer/prep.h"
  32 #include "optimizer/subselect.h"
  33 #include "optimizer/tlist.h"
  34 #include "optimizer/var.h"
  35 #ifdef OPTIMIZER_DEBUG
  36 #include "nodes/print.h"
  37 #endif
  38 #include "parser/parse_expr.h"
  39 #include "parser/parse_oper.h"
  40 #include "parser/parsetree.h"
  41 #include "utils/lsyscache.h"
  42 #include "utils/syscache.h"
  43
  44
  45 /* Expression kind codes for preprocess_expression */
  46 #define EXPRKIND_QUAL           0
  47 #define EXPRKIND_TARGET         1
  48 #define EXPRKIND_RTFUNC         2
  49 #define EXPRKIND_VALUES         3
  50 #define EXPRKIND_LIMIT          4
  51 #define EXPRKIND_ININFO         5
  52 #define EXPRKIND_APPINFO        6
  53
  54
  55 static Node *preprocess_expression(PlannerInfo *root, Node *expr, int kind);
  56 static void preprocess_qual_conditions(PlannerInfo *root, Node *jtnode);
  57 static Plan *inheritance_planner(PlannerInfo *root);
  58 static Plan *grouping_planner(PlannerInfo *root, double tuple_fraction);
  59 static bool is_dummy_plan(Plan *plan);
  60 static double preprocess_limit(PlannerInfo *root,
  61                                  double tuple_fraction,
  62                                  int64 *offset_est, int64 *count_est);
  63 static Oid *extract_grouping_ops(List *groupClause);
  64 static bool choose_hashed_grouping(PlannerInfo *root, double tuple_fraction,
  65                                            Path *cheapest_path, Path *sorted_path,
  66                                            Oid *groupOperators, double dNumGroups,
  67                                            AggClauseCounts *agg_counts);
  68 static List *make_subplanTargetList(PlannerInfo *root, List *tlist,
  69                                            AttrNumber **groupColIdx, bool *need_tlist_eval);
  70 static void locate_grouping_columns(PlannerInfo *root,
  71                                                 List *tlist,
  72                                                 List *sub_tlist,
  73                                                 AttrNumber *groupColIdx);
  74 static List *postprocess_setop_tlist(List *new_tlist, List *orig_tlist);
  75
  76
  77 /*****************************************************************************
  78  *
  79  *         Query optimizer entry point
  80  *
  81  *****************************************************************************/
  82 PlannedStmt *
  83 planner(Query *parse, int cursorOptions, ParamListInfo boundParams)
  84 {
  85         PlannedStmt *result;
  86         PlannerGlobal *glob;
  87         double          tuple_fraction;
  88         PlannerInfo *root;
  89         Plan       *top_plan;
  90         ListCell   *lp,
  91                            *lr;
  92
  93         /* Cursor options may come from caller or from DECLARE CURSOR stmt */
  94         if (parse->utilityStmt &&
  95                 IsA(parse->utilityStmt, DeclareCursorStmt))
  96                 cursorOptions |= ((DeclareCursorStmt *) parse->utilityStmt)->options;
  97
  98         /*
  99          * Set up global state for this planner invocation.  This data is needed
 100          * across all levels of sub-Query that might exist in the given command,
 101          * so we keep it in a separate struct that's linked to by each per-Query
 102          * PlannerInfo.
 103          */
 104         glob = makeNode(PlannerGlobal);
 105
 106         glob->boundParams = boundParams;
 107         glob->paramlist = NIL;
 108         glob->subplans = NIL;
 109         glob->subrtables = NIL;
 110         glob->rewindPlanIDs = NULL;
 111         glob->finalrtable = NIL;
 112
 113         /* Determine what fraction of the plan is likely to be scanned */
 114         if (cursorOptions & CURSOR_OPT_FAST_PLAN)
 115         {
 116                 /*
 117                  * We have no real idea how many tuples the user will ultimately FETCH
 118                  * from a cursor, but it seems a good bet that he doesn't want 'em
 119                  * all.  Optimize for 10% retrieval (you gotta better number?  Should
 120                  * this be a SETtable parameter?)
 121                  */
 122                 tuple_fraction = 0.10;
 123         }
 124         else
 125         {
 126                 /* Default assumption is we need all the tuples */
 127                 tuple_fraction = 0.0;
 128         }
 129
 130         /* primary planning entry point (may recurse for subqueries) */
 131         top_plan = subquery_planner(glob, parse, 1, tuple_fraction, &root);
 132
 133         /*
 134          * If creating a plan for a scrollable cursor, make sure it can run
 135          * backwards on demand.  Add a Material node at the top at need.
 136          */
 137         if (cursorOptions & CURSOR_OPT_SCROLL)
 138         {
 139                 if (!ExecSupportsBackwardScan(top_plan))
 140                         top_plan = materialize_finished_plan(top_plan);
 141         }
 142
 143         /* final cleanup of the plan */
 144         Assert(glob->finalrtable == NIL);
 145         top_plan = set_plan_references(glob, top_plan, root->parse->rtable);
 146         /* ... and the subplans (both regular subplans and initplans) */
 147         Assert(list_length(glob->subplans) == list_length(glob->subrtables));
 148         forboth(lp, glob->subplans, lr, glob->subrtables)
 149         {
 150                 Plan   *subplan = (Plan *) lfirst(lp);
 151                 List   *subrtable = (List *) lfirst(lr);
 152
 153                 lfirst(lp) = set_plan_references(glob, subplan, subrtable);
 154         }
 155
 156         /* build the PlannedStmt result */
 157         result = makeNode(PlannedStmt);
 158
 159         result->commandType = parse->commandType;
 160         result->canSetTag = parse->canSetTag;
 161         result->planTree = top_plan;
 162         result->rtable = glob->finalrtable;
 163         result->resultRelations = root->resultRelations;
 164         result->utilityStmt = parse->utilityStmt;
 165         result->intoClause = parse->intoClause;
 166         result->subplans = glob->subplans;
 167         result->rewindPlanIDs = glob->rewindPlanIDs;
 168         result->returningLists = root->returningLists;
 169         result->rowMarks = parse->rowMarks;
 170         result->nParamExec = list_length(glob->paramlist);
 171
 172         return result;
 173 }
 174
 175
 176 /*--------------------
 177  * subquery_planner
 178  *        Invokes the planner on a subquery.  We recurse to here for each
 179  *        sub-SELECT found in the query tree.
 180  *
 181  * glob is the global state for the current planner run.
 182  * parse is the querytree produced by the parser & rewriter.
 183  * level is the current recursion depth (1 at the top-level Query).
 184  * tuple_fraction is the fraction of tuples we expect will be retrieved.
 185  * tuple_fraction is interpreted as explained for grouping_planner, below.
 186  *
 187  * If subroot isn't NULL, we pass back the query's final PlannerInfo struct;
 188  * among other things this tells the output sort ordering of the plan.
 189  *
 190  * Basically, this routine does the stuff that should only be done once
 191  * per Query object.  It then calls grouping_planner.  At one time,
 192  * grouping_planner could be invoked recursively on the same Query object;
 193  * that's not currently true, but we keep the separation between the two
 194  * routines anyway, in case we need it again someday.
 195  *
 196  * subquery_planner will be called recursively to handle sub-Query nodes
 197  * found within the query's expressions and rangetable.
 198  *
 199  * Returns a query plan.
 200  *--------------------
 201  */
 202 Plan *
 203 subquery_planner(PlannerGlobal *glob, Query *parse,
 204                                  Index level, double tuple_fraction,
 205                                  PlannerInfo **subroot)
 206 {
 207         int                     num_old_subplans = list_length(glob->subplans);
 208         PlannerInfo *root;
 209         Plan       *plan;
 210         List       *newHaving;
 211         ListCell   *l;
 212
 213         /* Create a PlannerInfo data structure for this subquery */
 214         root = makeNode(PlannerInfo);
 215         root->parse = parse;
 216         root->glob = glob;
 217         root->query_level = level;
 218         root->planner_cxt = CurrentMemoryContext;
 219         root->init_plans = NIL;
 220         root->eq_classes = NIL;
 221         root->in_info_list = NIL;
 222         root->append_rel_list = NIL;
 223
 224         /*
 225          * Look for IN clauses at the top level of WHERE, and transform them into
 226          * joins.  Note that this step only handles IN clauses originally at top
 227          * level of WHERE; if we pull up any subqueries in the next step, their
 228          * INs are processed just before pulling them up.
 229          */
 230         if (parse->hasSubLinks)
 231                 parse->jointree->quals = pull_up_IN_clauses(root,
 232                                                                                                         parse->jointree->quals);
 233
 234         /*
 235          * Check to see if any subqueries in the rangetable can be merged into
 236          * this query.
 237          */
 238         parse->jointree = (FromExpr *)
 239                 pull_up_subqueries(root, (Node *) parse->jointree, false, false);
 240
 241         /*
 242          * Detect whether any rangetable entries are RTE_JOIN kind; if not, we can
 243          * avoid the expense of doing flatten_join_alias_vars().  Also check for
 244          * outer joins --- if none, we can skip reduce_outer_joins() and some
 245          * other processing.  This must be done after we have done
 246          * pull_up_subqueries, of course.
 247          *
 248          * Note: if reduce_outer_joins manages to eliminate all outer joins,
 249          * root->hasOuterJoins is not reset currently.  This is OK since its
 250          * purpose is merely to suppress unnecessary processing in simple cases.
 251          */
 252         root->hasJoinRTEs = false;
 253         root->hasOuterJoins = false;
 254         foreach(l, parse->rtable)
 255         {
 256                 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
 257
 258                 if (rte->rtekind == RTE_JOIN)
 259                 {
 260                         root->hasJoinRTEs = true;
 261                         if (IS_OUTER_JOIN(rte->jointype))
 262                         {
 263                                 root->hasOuterJoins = true;
 264                                 /* Can quit scanning once we find an outer join */
 265                                 break;
 266                         }
 267                 }
 268         }
 269
 270         /*
 271          * Expand any rangetable entries that are inheritance sets into "append
 272          * relations".  This can add entries to the rangetable, but they must be
 273          * plain base relations not joins, so it's OK (and marginally more
 274          * efficient) to do it after checking for join RTEs.  We must do it after
 275          * pulling up subqueries, else we'd fail to handle inherited tables in
 276          * subqueries.
 277          */
 278         expand_inherited_tables(root);
 279
 280         /*
 281          * Set hasHavingQual to remember if HAVING clause is present.  Needed
 282          * because preprocess_expression will reduce a constant-true condition to
 283          * an empty qual list ... but "HAVING TRUE" is not a semantic no-op.
 284          */
 285         root->hasHavingQual = (parse->havingQual != NULL);
 286
 287         /* Clear this flag; might get set in distribute_qual_to_rels */
 288         root->hasPseudoConstantQuals = false;
 289
 290         /*
 291          * Do expression preprocessing on targetlist and quals.
 292          */
 293         parse->targetList = (List *)
 294                 preprocess_expression(root, (Node *) parse->targetList,
 295                                                           EXPRKIND_TARGET);
 296
 297         parse->returningList = (List *)
 298                 preprocess_expression(root, (Node *) parse->returningList,
 299                                                           EXPRKIND_TARGET);
 300
 301         preprocess_qual_conditions(root, (Node *) parse->jointree);
 302
 303         parse->havingQual = preprocess_expression(root, parse->havingQual,
 304                                                                                           EXPRKIND_QUAL);
 305
 306         parse->limitOffset = preprocess_expression(root, parse->limitOffset,
 307                                                                                            EXPRKIND_LIMIT);
 308         parse->limitCount = preprocess_expression(root, parse->limitCount,
 309                                                                                           EXPRKIND_LIMIT);
 310
 311         root->in_info_list = (List *)
 312                 preprocess_expression(root, (Node *) root->in_info_list,
 313                                                           EXPRKIND_ININFO);
 314         root->append_rel_list = (List *)
 315                 preprocess_expression(root, (Node *) root->append_rel_list,
 316                                                           EXPRKIND_APPINFO);
 317
 318         /* Also need to preprocess expressions for function and values RTEs */
 319         foreach(l, parse->rtable)
 320         {
 321                 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
 322
 323                 if (rte->rtekind == RTE_FUNCTION)
 324                         rte->funcexpr = preprocess_expression(root, rte->funcexpr,
 325                                                                                                   EXPRKIND_RTFUNC);
 326                 else if (rte->rtekind == RTE_VALUES)
 327                         rte->values_lists = (List *)
 328                                 preprocess_expression(root, (Node *) rte->values_lists,
 329                                                                           EXPRKIND_VALUES);
 330         }
 331
 332         /*
 333          * In some cases we may want to transfer a HAVING clause into WHERE. We
 334          * cannot do so if the HAVING clause contains aggregates (obviously) or
 335          * volatile functions (since a HAVING clause is supposed to be executed
 336          * only once per group).  Also, it may be that the clause is so expensive
 337          * to execute that we're better off doing it only once per group, despite
 338          * the loss of selectivity.  This is hard to estimate short of doing the
 339          * entire planning process twice, so we use a heuristic: clauses
 340          * containing subplans are left in HAVING.      Otherwise, we move or copy the
 341          * HAVING clause into WHERE, in hopes of eliminating tuples before
 342          * aggregation instead of after.
 343          *
 344          * If the query has explicit grouping then we can simply move such a
 345          * clause into WHERE; any group that fails the clause will not be in the
 346          * output because none of its tuples will reach the grouping or
 347          * aggregation stage.  Otherwise we must have a degenerate (variable-free)
 348          * HAVING clause, which we put in WHERE so that query_planner() can use it
 349          * in a gating Result node, but also keep in HAVING to ensure that we
 350          * don't emit a bogus aggregated row. (This could be done better, but it
 351          * seems not worth optimizing.)
 352          *
 353          * Note that both havingQual and parse->jointree->quals are in
 354          * implicitly-ANDed-list form at this point, even though they are declared
 355          * as Node *.
 356          */
 357         newHaving = NIL;
 358         foreach(l, (List *) parse->havingQual)
 359         {
 360                 Node       *havingclause = (Node *) lfirst(l);
 361
 362                 if (contain_agg_clause(havingclause) ||
 363                         contain_volatile_functions(havingclause) ||
 364                         contain_subplans(havingclause))
 365                 {
 366                         /* keep it in HAVING */
 367                         newHaving = lappend(newHaving, havingclause);
 368                 }
 369                 else if (parse->groupClause)
 370                 {
 371                         /* move it to WHERE */
 372                         parse->jointree->quals = (Node *)
 373                                 lappend((List *) parse->jointree->quals, havingclause);
 374                 }
 375                 else
 376                 {
 377                         /* put a copy in WHERE, keep it in HAVING */
 378                         parse->jointree->quals = (Node *)
 379                                 lappend((List *) parse->jointree->quals,
 380                                                 copyObject(havingclause));
 381                         newHaving = lappend(newHaving, havingclause);
 382                 }
 383         }
 384         parse->havingQual = (Node *) newHaving;
 385
 386         /*
 387          * If we have any outer joins, try to reduce them to plain inner joins.
 388          * This step is most easily done after we've done expression
 389          * preprocessing.
 390          */
 391         if (root->hasOuterJoins)
 392                 reduce_outer_joins(root);
 393
 394         /*
 395          * Do the main planning.  If we have an inherited target relation, that
 396          * needs special processing, else go straight to grouping_planner.
 397          */
 398         if (parse->resultRelation &&
 399                 rt_fetch(parse->resultRelation, parse->rtable)->inh)
 400                 plan = inheritance_planner(root);
 401         else
 402                 plan = grouping_planner(root, tuple_fraction);
 403
 404         /*
 405          * If any subplans were generated, or if we're inside a subplan, build
 406          * initPlan list and extParam/allParam sets for plan nodes, and attach the
 407          * initPlans to the top plan node.
 408          */
 409         if (list_length(glob->subplans) != num_old_subplans ||
 410                 root->query_level > 1)
 411                 SS_finalize_plan(root, plan);
 412
 413         /* Return internal info if caller wants it */
 414         if (subroot)
 415                 *subroot = root;
 416
 417         return plan;
 418 }
 419
 420 /*
 421  * preprocess_expression
 422  *              Do subquery_planner's preprocessing work for an expression,
 423  *              which can be a targetlist, a WHERE clause (including JOIN/ON
 424  *              conditions), or a HAVING clause.
 425  */
 426 static Node *
 427 preprocess_expression(PlannerInfo *root, Node *expr, int kind)
 428 {
 429         /*
 430          * Fall out quickly if expression is empty.  This occurs often enough to
 431          * be worth checking.  Note that null->null is the correct conversion for
 432          * implicit-AND result format, too.
 433          */
 434         if (expr == NULL)
 435                 return NULL;
 436
 437         /*
 438          * If the query has any join RTEs, replace join alias variables with
 439          * base-relation variables. We must do this before sublink processing,
 440          * else sublinks expanded out from join aliases wouldn't get processed. We
 441          * can skip it in VALUES lists, however, since they can't contain any Vars
 442          * at all.
 443          */
 444         if (root->hasJoinRTEs && kind != EXPRKIND_VALUES)
 445                 expr = flatten_join_alias_vars(root, expr);
 446
 447         /*
 448          * Simplify constant expressions.
 449          *
 450          * Note: this also flattens nested AND and OR expressions into N-argument
 451          * form.  All processing of a qual expression after this point must be
 452          * careful to maintain AND/OR flatness --- that is, do not generate a tree
 453          * with AND directly under AND, nor OR directly under OR.
 454          *
 455          * Because this is a relatively expensive process, we skip it when the
 456          * query is trivial, such as "SELECT 2+2;" or "INSERT ... VALUES()". The
 457          * expression will only be evaluated once anyway, so no point in
 458          * pre-simplifying; we can't execute it any faster than the executor can,
 459          * and we will waste cycles copying the tree.  Notice however that we
 460          * still must do it for quals (to get AND/OR flatness); and if we are in a
 461          * subquery we should not assume it will be done only once.
 462          *
 463          * For VALUES lists we never do this at all, again on the grounds that we
 464          * should optimize for one-time evaluation.
 465          */
 466         if (kind != EXPRKIND_VALUES &&
 467                 (root->parse->jointree->fromlist != NIL ||
 468                  kind == EXPRKIND_QUAL ||
 469                  root->query_level > 1))
 470                 expr = eval_const_expressions(expr);
 471
 472         /*
 473          * If it's a qual or havingQual, canonicalize it.
 474          */
 475         if (kind == EXPRKIND_QUAL)
 476         {
 477                 expr = (Node *) canonicalize_qual((Expr *) expr);
 478
 479 #ifdef OPTIMIZER_DEBUG
 480                 printf("After canonicalize_qual()\n");
 481                 pprint(expr);
 482 #endif
 483         }
 484
 485         /* Expand SubLinks to SubPlans */
 486         if (root->parse->hasSubLinks)
 487                 expr = SS_process_sublinks(root, expr, (kind == EXPRKIND_QUAL));
 488
 489         /*
 490          * XXX do not insert anything here unless you have grokked the comments in
 491          * SS_replace_correlation_vars ...
 492          */
 493
 494         /* Replace uplevel vars with Param nodes (this IS possible in VALUES) */
 495         if (root->query_level > 1)
 496                 expr = SS_replace_correlation_vars(root, expr);
 497
 498         /*
 499          * If it's a qual or havingQual, convert it to implicit-AND format. (We
 500          * don't want to do this before eval_const_expressions, since the latter
 501          * would be unable to simplify a top-level AND correctly. Also,
 502          * SS_process_sublinks expects explicit-AND format.)
 503          */
 504         if (kind == EXPRKIND_QUAL)
 505                 expr = (Node *) make_ands_implicit((Expr *) expr);
 506
 507         return expr;
 508 }
 509
 510 /*
 511  * preprocess_qual_conditions
 512  *              Recursively scan the query's jointree and do subquery_planner's
 513  *              preprocessing work on each qual condition found therein.
 514  */
 515 static void
 516 preprocess_qual_conditions(PlannerInfo *root, Node *jtnode)
 517 {
 518         if (jtnode == NULL)
 519                 return;
 520         if (IsA(jtnode, RangeTblRef))
 521         {
 522                 /* nothing to do here */
 523         }
 524         else if (IsA(jtnode, FromExpr))
 525         {
 526                 FromExpr   *f = (FromExpr *) jtnode;
 527                 ListCell   *l;
 528
 529                 foreach(l, f->fromlist)
 530                         preprocess_qual_conditions(root, lfirst(l));
 531
 532                 f->quals = preprocess_expression(root, f->quals, EXPRKIND_QUAL);
 533         }
 534         else if (IsA(jtnode, JoinExpr))
 535         {
 536                 JoinExpr   *j = (JoinExpr *) jtnode;
 537
 538                 preprocess_qual_conditions(root, j->larg);
 539                 preprocess_qual_conditions(root, j->rarg);
 540
 541                 j->quals = preprocess_expression(root, j->quals, EXPRKIND_QUAL);
 542         }
 543         else
 544                 elog(ERROR, "unrecognized node type: %d",
 545                          (int) nodeTag(jtnode));
 546 }
 547
 548 /*
 549  * inheritance_planner
 550  *        Generate a plan in the case where the result relation is an
 551  *        inheritance set.
 552  *
 553  * We have to handle this case differently from cases where a source relation
 554  * is an inheritance set. Source inheritance is expanded at the bottom of the
 555  * plan tree (see allpaths.c), but target inheritance has to be expanded at
 556  * the top.  The reason is that for UPDATE, each target relation needs a
 557  * different targetlist matching its own column set.  Also, for both UPDATE
 558  * and DELETE, the executor needs the Append plan node at the top, else it
 559  * can't keep track of which table is the current target table.  Fortunately,
 560  * the UPDATE/DELETE target can never be the nullable side of an outer join,
 561  * so it's OK to generate the plan this way.
 562  *
 563  * Returns a query plan.
 564  */
 565 static Plan *
 566 inheritance_planner(PlannerInfo *root)
 567 {
 568         Query      *parse = root->parse;
 569         int                     parentRTindex = parse->resultRelation;
 570         List       *subplans = NIL;
 571         List       *resultRelations = NIL;
 572         List       *returningLists = NIL;
 573         List       *rtable = NIL;
 574         List       *tlist = NIL;
 575         PlannerInfo subroot;
 576         ListCell   *l;
 577
 578         foreach(l, root->append_rel_list)
 579         {
 580                 AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
 581                 Plan       *subplan;
 582
 583                 /* append_rel_list contains all append rels; ignore others */
 584                 if (appinfo->parent_relid != parentRTindex)
 585                         continue;
 586
 587                 /*
 588                  * Generate modified query with this rel as target.  We have to be
 589                  * prepared to translate varnos in in_info_list as well as in the
 590                  * Query proper.
 591                  */
 592                 memcpy(&subroot, root, sizeof(PlannerInfo));
 593                 subroot.parse = (Query *)
 594                         adjust_appendrel_attrs((Node *) parse,
 595                                                                    appinfo);
 596                 subroot.in_info_list = (List *)
 597                         adjust_appendrel_attrs((Node *) root->in_info_list,
 598                                                                    appinfo);
 599                 subroot.init_plans = NIL;
 600                 /* There shouldn't be any OJ info to translate, as yet */
 601                 Assert(subroot.oj_info_list == NIL);
 602
 603                 /* Generate plan */
 604                 subplan = grouping_planner(&subroot, 0.0 /* retrieve all tuples */ );
 605
 606                 /*
 607                  * If this child rel was excluded by constraint exclusion, exclude it
 608                  * from the plan.
 609                  */
 610                 if (is_dummy_plan(subplan))
 611                         continue;
 612
 613                 /* Save rtable and tlist from first rel for use below */
 614                 if (subplans == NIL)
 615                 {
 616                         rtable = subroot.parse->rtable;
 617                         tlist = subplan->targetlist;
 618                 }
 619
 620                 subplans = lappend(subplans, subplan);
 621
 622                 /* Make sure any initplans from this rel get into the outer list */
 623                 root->init_plans = list_concat(root->init_plans, subroot.init_plans);
 624
 625                 /* Build target-relations list for the executor */
 626                 resultRelations = lappend_int(resultRelations, appinfo->child_relid);
 627
 628                 /* Build list of per-relation RETURNING targetlists */
 629                 if (parse->returningList)
 630                 {
 631                         Assert(list_length(subroot.returningLists) == 1);
 632                         returningLists = list_concat(returningLists,
 633                                                                                  subroot.returningLists);
 634                 }
 635         }
 636
 637         root->resultRelations = resultRelations;
 638         root->returningLists = returningLists;
 639
 640         /* Mark result as unordered (probably unnecessary) */
 641         root->query_pathkeys = NIL;
 642
 643         /*
 644          * If we managed to exclude every child rel, return a dummy plan
 645          */
 646         if (subplans == NIL)
 647                 return (Plan *) make_result(root,
 648                                                                         tlist,
 649                                                                         (Node *) list_make1(makeBoolConst(false,
 650                                                                                                                                           false)),
 651                                                                         NULL);
 652
 653         /*
 654          * Planning might have modified the rangetable, due to changes of the
 655          * Query structures inside subquery RTEs.  We have to ensure that this
 656          * gets propagated back to the master copy.  But can't do this until we
 657          * are done planning, because all the calls to grouping_planner need
 658          * virgin sub-Queries to work from.  (We are effectively assuming that
 659          * sub-Queries will get planned identically each time, or at least that
 660          * the impacts on their rangetables will be the same each time.)
 661          *
 662          * XXX should clean this up someday
 663          */
 664         parse->rtable = rtable;
 665
 666         /* Suppress Append if there's only one surviving child rel */
 667         if (list_length(subplans) == 1)
 668                 return (Plan *) linitial(subplans);
 669
 670         return (Plan *) make_append(subplans, true, tlist);
 671 }
 672
 673 /*--------------------
 674  * grouping_planner
 675  *        Perform planning steps related to grouping, aggregation, etc.
 676  *        This primarily means adding top-level processing to the basic
 677  *        query plan produced by query_planner.
 678  *
 679  * tuple_fraction is the fraction of tuples we expect will be retrieved
 680  *
 681  * tuple_fraction is interpreted as follows:
 682  *        0: expect all tuples to be retrieved (normal case)
 683  *        0 < tuple_fraction < 1: expect the given fraction of tuples available
 684  *              from the plan to be retrieved
 685  *        tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
 686  *              expected to be retrieved (ie, a LIMIT specification)
 687  *
 688  * Returns a query plan.  Also, root->query_pathkeys is returned as the
 689  * actual output ordering of the plan (in pathkey format).
 690  *--------------------
 691  */
 692 static Plan *
 693 grouping_planner(PlannerInfo *root, double tuple_fraction)
 694 {
 695         Query      *parse = root->parse;
 696         List       *tlist = parse->targetList;
 697         int64           offset_est = 0;
 698         int64           count_est = 0;
 699         Plan       *result_plan;
 700         List       *current_pathkeys;
 701         List       *sort_pathkeys;
 702         double          dNumGroups = 0;
 703
 704         /* Tweak caller-supplied tuple_fraction if have LIMIT/OFFSET */
 705         if (parse->limitCount || parse->limitOffset)
 706                 tuple_fraction = preprocess_limit(root, tuple_fraction,
 707                                                                                   &offset_est, &count_est);
 708
 709         if (parse->setOperations)
 710         {
 711                 List       *set_sortclauses;
 712
 713                 /*
 714                  * If there's a top-level ORDER BY, assume we have to fetch all the
 715                  * tuples.      This might seem too simplistic given all the hackery below
 716                  * to possibly avoid the sort ... but a nonzero tuple_fraction is only
 717                  * of use to plan_set_operations() when the setop is UNION ALL, and
 718                  * the result of UNION ALL is always unsorted.
 719                  */
 720                 if (parse->sortClause)
 721                         tuple_fraction = 0.0;
 722
 723                 /*
 724                  * Construct the plan for set operations.  The result will not need
 725                  * any work except perhaps a top-level sort and/or LIMIT.
 726                  */
 727                 result_plan = plan_set_operations(root, tuple_fraction,
 728                                                                                   &set_sortclauses);
 729
 730                 /*
 731                  * Calculate pathkeys representing the sort order (if any) of the set
 732                  * operation's result.  We have to do this before overwriting the sort
 733                  * key information...
 734                  */
 735                 current_pathkeys = make_pathkeys_for_sortclauses(root,
 736                                                                                                                  set_sortclauses,
 737                                                                                                         result_plan->targetlist,
 738                                                                                                                  true);
 739
 740                 /*
 741                  * We should not need to call preprocess_targetlist, since we must be
 742                  * in a SELECT query node.      Instead, use the targetlist returned by
 743                  * plan_set_operations (since this tells whether it returned any
 744                  * resjunk columns!), and transfer any sort key information from the
 745                  * original tlist.
 746                  */
 747                 Assert(parse->commandType == CMD_SELECT);
 748
 749                 tlist = postprocess_setop_tlist(result_plan->targetlist, tlist);
 750
 751                 /*
 752                  * Can't handle FOR UPDATE/SHARE here (parser should have checked
 753                  * already, but let's make sure).
 754                  */
 755                 if (parse->rowMarks)
 756                         ereport(ERROR,
 757                                         (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
 758                                          errmsg("SELECT FOR UPDATE/SHARE is not allowed with UNION/INTERSECT/EXCEPT")));
 759
 760                 /*
 761                  * Calculate pathkeys that represent result ordering requirements
 762                  */
 763                 sort_pathkeys = make_pathkeys_for_sortclauses(root,
 764                                                                                                           parse->sortClause,
 765                                                                                                           tlist,
 766                                                                                                           true);
 767         }
 768         else
 769         {
 770                 /* No set operations, do regular planning */
 771                 List       *sub_tlist;
 772                 List       *group_pathkeys;
 773                 AttrNumber *groupColIdx = NULL;
 774                 Oid                *groupOperators = NULL;
 775                 bool            need_tlist_eval = true;
 776                 QualCost        tlist_cost;
 777                 Path       *cheapest_path;
 778                 Path       *sorted_path;
 779                 Path       *best_path;
 780                 long            numGroups = 0;
 781                 AggClauseCounts agg_counts;
 782                 int                     numGroupCols = list_length(parse->groupClause);
 783                 bool            use_hashed_grouping = false;
 784
 785                 MemSet(&agg_counts, 0, sizeof(AggClauseCounts));
 786
 787                 /* Preprocess targetlist */
 788                 tlist = preprocess_targetlist(root, tlist);
 789
 790                 /*
 791                  * Generate appropriate target list for subplan; may be different from
 792                  * tlist if grouping or aggregation is needed.
 793                  */
 794                 sub_tlist = make_subplanTargetList(root, tlist,
 795                                                                                    &groupColIdx, &need_tlist_eval);
 796
 797                 /*
 798                  * Calculate pathkeys that represent grouping/ordering requirements.
 799                  * Stash them in PlannerInfo so that query_planner can canonicalize
 800                  * them after EquivalenceClasses have been formed.
 801                  */
 802                 root->group_pathkeys =
 803                         make_pathkeys_for_sortclauses(root,
 804                                                                                   parse->groupClause,
 805                                                                                   tlist,
 806                                                                                   false);
 807                 root->sort_pathkeys =
 808                         make_pathkeys_for_sortclauses(root,
 809                                                                                   parse->sortClause,
 810                                                                                   tlist,
 811                                                                                   false);
 812
 813                 /*
 814                  * Will need actual number of aggregates for estimating costs.
 815                  *
 816                  * Note: we do not attempt to detect duplicate aggregates here; a
 817                  * somewhat-overestimated count is okay for our present purposes.
 818                  *
 819                  * Note: think not that we can turn off hasAggs if we find no aggs. It
 820                  * is possible for constant-expression simplification to remove all
 821                  * explicit references to aggs, but we still have to follow the
 822                  * aggregate semantics (eg, producing only one output row).
 823                  */
 824                 if (parse->hasAggs)
 825                 {
 826                         count_agg_clauses((Node *) tlist, &agg_counts);
 827                         count_agg_clauses(parse->havingQual, &agg_counts);
 828                 }
 829
 830                 /*
 831                  * Figure out whether we need a sorted result from query_planner.
 832                  *
 833                  * If we have a GROUP BY clause, then we want a result sorted properly
 834                  * for grouping.  Otherwise, if there is an ORDER BY clause, we want
 835                  * to sort by the ORDER BY clause.      (Note: if we have both, and ORDER
 836                  * BY is a superset of GROUP BY, it would be tempting to request sort
 837                  * by ORDER BY --- but that might just leave us failing to exploit an
 838                  * available sort order at all. Needs more thought...)
 839                  */
 840                 if (parse->groupClause)
 841                         root->query_pathkeys = root->group_pathkeys;
 842                 else if (parse->sortClause)
 843                         root->query_pathkeys = root->sort_pathkeys;
 844                 else
 845                         root->query_pathkeys = NIL;
 846
 847                 /*
 848                  * Generate the best unsorted and presorted paths for this Query (but
 849                  * note there may not be any presorted path).  query_planner will also
 850                  * estimate the number of groups in the query, and canonicalize all
 851                  * the pathkeys.
 852                  */
 853                 query_planner(root, sub_tlist, tuple_fraction,
 854                                           &cheapest_path, &sorted_path, &dNumGroups);
 855
 856                 group_pathkeys = root->group_pathkeys;
 857                 sort_pathkeys = root->sort_pathkeys;
 858
 859                 /*
 860                  * If grouping, extract the grouping operators and decide whether we
 861                  * want to use hashed grouping.
 862                  */
 863                 if (parse->groupClause)
 864                 {
 865                         groupOperators = extract_grouping_ops(parse->groupClause);
 866                         use_hashed_grouping =
 867                                 choose_hashed_grouping(root, tuple_fraction,
 868                                                                            cheapest_path, sorted_path,
 869                                                                            groupOperators, dNumGroups,
 870                                                                            &agg_counts);
 871
 872                         /* Also convert # groups to long int --- but 'ware overflow! */
 873                         numGroups = (long) Min(dNumGroups, (double) LONG_MAX);
 874                 }
 875
 876                 /*
 877                  * Select the best path.  If we are doing hashed grouping, we will
 878                  * always read all the input tuples, so use the cheapest-total path.
 879                  * Otherwise, trust query_planner's decision about which to use.
 880                  */
 881                 if (use_hashed_grouping || !sorted_path)
 882                         best_path = cheapest_path;
 883                 else
 884                         best_path = sorted_path;
 885
 886                 /*
 887                  * Check to see if it's possible to optimize MIN/MAX aggregates. If
 888                  * so, we will forget all the work we did so far to choose a "regular"
 889                  * path ... but we had to do it anyway to be able to tell which way is
 890                  * cheaper.
 891                  */
 892                 result_plan = optimize_minmax_aggregates(root,
 893                                                                                                  tlist,
 894                                                                                                  best_path);
 895                 if (result_plan != NULL)
 896                 {
 897                         /*
 898                          * optimize_minmax_aggregates generated the full plan, with the
 899                          * right tlist, and it has no sort order.
 900                          */
 901                         current_pathkeys = NIL;
 902                 }
 903                 else
 904                 {
 905                         /*
 906                          * Normal case --- create a plan according to query_planner's
 907                          * results.
 908                          */
 909                         result_plan = create_plan(root, best_path);
 910                         current_pathkeys = best_path->pathkeys;
 911
 912                         /*
 913                          * create_plan() returns a plan with just a "flat" tlist of
 914                          * required Vars.  Usually we need to insert the sub_tlist as the
 915                          * tlist of the top plan node.  However, we can skip that if we
 916                          * determined that whatever query_planner chose to return will be
 917                          * good enough.
 918                          */
 919                         if (need_tlist_eval)
 920                         {
 921                                 /*
 922                                  * If the top-level plan node is one that cannot do expression
 923                                  * evaluation, we must insert a Result node to project the
 924                                  * desired tlist.
 925                                  */
 926                                 if (!is_projection_capable_plan(result_plan))
 927                                 {
 928                                         result_plan = (Plan *) make_result(root,
 929                                                                                                            sub_tlist,
 930                                                                                                            NULL,
 931                                                                                                            result_plan);
 932                                 }
 933                                 else
 934                                 {
 935                                         /*
 936                                          * Otherwise, just replace the subplan's flat tlist with
 937                                          * the desired tlist.
 938                                          */
 939                                         result_plan->targetlist = sub_tlist;
 940                                 }
 941
 942                                 /*
 943                                  * Also, account for the cost of evaluation of the sub_tlist.
 944                                  *
 945                                  * Up to now, we have only been dealing with "flat" tlists,
 946                                  * containing just Vars.  So their evaluation cost is zero
 947                                  * according to the model used by cost_qual_eval() (or if you
 948                                  * prefer, the cost is factored into cpu_tuple_cost).  Thus we
 949                                  * can avoid accounting for tlist cost throughout
 950                                  * query_planner() and subroutines.  But now we've inserted a
 951                                  * tlist that might contain actual operators, sub-selects, etc
 952                                  * --- so we'd better account for its cost.
 953                                  *
 954                                  * Below this point, any tlist eval cost for added-on nodes
 955                                  * should be accounted for as we create those nodes.
 956                                  * Presently, of the node types we can add on, only Agg and
 957                                  * Group project new tlists (the rest just copy their input
 958                                  * tuples) --- so make_agg() and make_group() are responsible
 959                                  * for computing the added cost.
 960                                  */
 961                                 cost_qual_eval(&tlist_cost, sub_tlist, root);
 962                                 result_plan->startup_cost += tlist_cost.startup;
 963                                 result_plan->total_cost += tlist_cost.startup +
 964                                         tlist_cost.per_tuple * result_plan->plan_rows;
 965                         }
 966                         else
 967                         {
 968                                 /*
 969                                  * Since we're using query_planner's tlist and not the one
 970                                  * make_subplanTargetList calculated, we have to refigure any
 971                                  * grouping-column indexes make_subplanTargetList computed.
 972                                  */
 973                                 locate_grouping_columns(root, tlist, result_plan->targetlist,
 974                                                                                 groupColIdx);
 975                         }
 976
 977                         /*
 978                          * Insert AGG or GROUP node if needed, plus an explicit sort step
 979                          * if necessary.
 980                          *
 981                          * HAVING clause, if any, becomes qual of the Agg or Group node.
 982                          */
 983                         if (use_hashed_grouping)
 984                         {
 985                                 /* Hashed aggregate plan --- no sort needed */
 986                                 result_plan = (Plan *) make_agg(root,
 987                                                                                                 tlist,
 988                                                                                                 (List *) parse->havingQual,
 989                                                                                                 AGG_HASHED,
 990                                                                                                 numGroupCols,
 991                                                                                                 groupColIdx,
 992                                                                                                 groupOperators,
 993                                                                                                 numGroups,
 994                                                                                                 agg_counts.numAggs,
 995                                                                                                 result_plan);
 996                                 /* Hashed aggregation produces randomly-ordered results */
 997                                 current_pathkeys = NIL;
 998                         }
 999                         else if (parse->hasAggs)
1000                         {
1001                                 /* Plain aggregate plan --- sort if needed */
1002                                 AggStrategy aggstrategy;
1003
1004                                 if (parse->groupClause)
1005                                 {
1006                                         if (!pathkeys_contained_in(group_pathkeys,
1007                                                                                            current_pathkeys))
1008                                         {
1009                                                 result_plan = (Plan *)
1010                                                         make_sort_from_groupcols(root,
1011                                                                                                          parse->groupClause,
1012                                                                                                          groupColIdx,
1013                                                                                                          result_plan);
1014                                                 current_pathkeys = group_pathkeys;
1015                                         }
1016                                         aggstrategy = AGG_SORTED;
1017
1018                                         /*
1019                                          * The AGG node will not change the sort ordering of its
1020                                          * groups, so current_pathkeys describes the result too.
1021                                          */
1022                                 }
1023                                 else
1024                                 {
1025                                         aggstrategy = AGG_PLAIN;
1026                                         /* Result will be only one row anyway; no sort order */
1027                                         current_pathkeys = NIL;
1028                                 }
1029
1030                                 result_plan = (Plan *) make_agg(root,
1031                                                                                                 tlist,
1032                                                                                                 (List *) parse->havingQual,
1033                                                                                                 aggstrategy,
1034                                                                                                 numGroupCols,
1035                                                                                                 groupColIdx,
1036                                                                                                 groupOperators,
1037                                                                                                 numGroups,
1038                                                                                                 agg_counts.numAggs,
1039                                                                                                 result_plan);
1040                         }
1041                         else if (parse->groupClause)
1042                         {
1043                                 /*
1044                                  * GROUP BY without aggregation, so insert a group node (plus
1045                                  * the appropriate sort node, if necessary).
1046                                  *
1047                                  * Add an explicit sort if we couldn't make the path come out
1048                                  * the way the GROUP node needs it.
1049                                  */
1050                                 if (!pathkeys_contained_in(group_pathkeys, current_pathkeys))
1051                                 {
1052                                         result_plan = (Plan *)
1053                                                 make_sort_from_groupcols(root,
1054                                                                                                  parse->groupClause,
1055                                                                                                  groupColIdx,
1056                                                                                                  result_plan);
1057                                         current_pathkeys = group_pathkeys;
1058                                 }
1059
1060                                 result_plan = (Plan *) make_group(root,
1061                                                                                                   tlist,
1062                                                                                                   (List *) parse->havingQual,
1063                                                                                                   numGroupCols,
1064                                                                                                   groupColIdx,
1065                                                                                                   groupOperators,
1066                                                                                                   dNumGroups,
1067                                                                                                   result_plan);
1068                                 /* The Group node won't change sort ordering */
1069                         }
1070                         else if (root->hasHavingQual)
1071                         {
1072                                 /*
1073                                  * No aggregates, and no GROUP BY, but we have a HAVING qual.
1074                                  * This is a degenerate case in which we are supposed to emit
1075                                  * either 0 or 1 row depending on whether HAVING succeeds.
1076                                  * Furthermore, there cannot be any variables in either HAVING
1077                                  * or the targetlist, so we actually do not need the FROM
1078                                  * table at all!  We can just throw away the plan-so-far and
1079                                  * generate a Result node.      This is a sufficiently unusual
1080                                  * corner case that it's not worth contorting the structure of
1081                                  * this routine to avoid having to generate the plan in the
1082                                  * first place.
1083                                  */
1084                                 result_plan = (Plan *) make_result(root,
1085                                                                                                    tlist,
1086                                                                                                    parse->havingQual,
1087                                                                                                    NULL);
1088                         }
1089                 }                                               /* end of non-minmax-aggregate case */
1090         }                                                       /* end of if (setOperations) */
1091
1092         /*
1093          * If we were not able to make the plan come out in the right order, add
1094          * an explicit sort step.
1095          */
1096         if (parse->sortClause)
1097         {
1098                 if (!pathkeys_contained_in(sort_pathkeys, current_pathkeys))
1099                 {
1100                         result_plan = (Plan *) make_sort_from_pathkeys(root,
1101                                                                                                                    result_plan,
1102                                                                                                                    sort_pathkeys);
1103                         current_pathkeys = sort_pathkeys;
1104                 }
1105         }
1106
1107         /*
1108          * If there is a DISTINCT clause, add the UNIQUE node.
1109          */
1110         if (parse->distinctClause)
1111         {
1112                 result_plan = (Plan *) make_unique(result_plan, parse->distinctClause);
1113
1114                 /*
1115                  * If there was grouping or aggregation, leave plan_rows as-is (ie,
1116                  * assume the result was already mostly unique).  If not, use the
1117                  * number of distinct-groups calculated by query_planner.
1118                  */
1119                 if (!parse->groupClause && !root->hasHavingQual && !parse->hasAggs)
1120                         result_plan->plan_rows = dNumGroups;
1121         }
1122
1123         /*
1124          * Finally, if there is a LIMIT/OFFSET clause, add the LIMIT node.
1125          */
1126         if (parse->limitCount || parse->limitOffset)
1127         {
1128                 result_plan = (Plan *) make_limit(result_plan,
1129                                                                                   parse->limitOffset,
1130                                                                                   parse->limitCount,
1131                                                                                   offset_est,
1132                                                                                   count_est);
1133         }
1134
1135         /*
1136          * Deal with the RETURNING clause if any.  It's convenient to pass the
1137          * returningList through setrefs.c now rather than at top level (if we
1138          * waited, handling inherited UPDATE/DELETE would be much harder).
1139          */
1140         if (parse->returningList)
1141         {
1142                 List       *rlist;
1143
1144                 Assert(parse->resultRelation);
1145                 rlist = set_returning_clause_references(parse->returningList,
1146                                                                                                 result_plan,
1147                                                                                                 parse->resultRelation);
1148                 root->returningLists = list_make1(rlist);
1149         }
1150         else
1151                 root->returningLists = NIL;
1152
1153         /* Compute result-relations list if needed */
1154         if (parse->resultRelation)
1155                 root->resultRelations = list_make1_int(parse->resultRelation);
1156         else
1157                 root->resultRelations = NIL;
1158
1159         /*
1160          * Return the actual output ordering in query_pathkeys for possible use by
1161          * an outer query level.
1162          */
1163         root->query_pathkeys = current_pathkeys;
1164
1165         return result_plan;
1166 }
1167
1168 /*
1169  * Detect whether a plan node is a "dummy" plan created when a relation
1170  * is deemed not to need scanning due to constraint exclusion.
1171  *
1172  * Currently, such dummy plans are Result nodes with constant FALSE
1173  * filter quals.
1174  */
1175 static bool
1176 is_dummy_plan(Plan *plan)
1177 {
1178         if (IsA(plan, Result))
1179         {
1180                 List       *rcqual = (List *) ((Result *) plan)->resconstantqual;
1181
1182                 if (list_length(rcqual) == 1)
1183                 {
1184                         Const      *constqual = (Const *) linitial(rcqual);
1185
1186                         if (constqual && IsA(constqual, Const))
1187                         {
1188                                 if (!constqual->constisnull &&
1189                                         !DatumGetBool(constqual->constvalue))
1190                                         return true;
1191                         }
1192                 }
1193         }
1194         return false;
1195 }
1196
1197 /*
1198  * preprocess_limit - do pre-estimation for LIMIT and/or OFFSET clauses
1199  *
1200  * We try to estimate the values of the LIMIT/OFFSET clauses, and pass the
1201  * results back in *count_est and *offset_est.  These variables are set to
1202  * 0 if the corresponding clause is not present, and -1 if it's present
1203  * but we couldn't estimate the value for it.  (The "0" convention is OK
1204  * for OFFSET but a little bit bogus for LIMIT: effectively we estimate
1205  * LIMIT 0 as though it were LIMIT 1.  But this is in line with the planner's
1206  * usual practice of never estimating less than one row.)  These values will
1207  * be passed to make_limit, which see if you change this code.
1208  *
1209  * The return value is the suitably adjusted tuple_fraction to use for
1210  * planning the query.  This adjustment is not overridable, since it reflects
1211  * plan actions that grouping_planner() will certainly take, not assumptions
1212  * about context.
1213  */
1214 static double
1215 preprocess_limit(PlannerInfo *root, double tuple_fraction,
1216                                  int64 *offset_est, int64 *count_est)
1217 {
1218         Query      *parse = root->parse;
1219         Node       *est;
1220         double          limit_fraction;
1221
1222         /* Should not be called unless LIMIT or OFFSET */
1223         Assert(parse->limitCount || parse->limitOffset);
1224
1225         /*
1226          * Try to obtain the clause values.  We use estimate_expression_value
1227          * primarily because it can sometimes do something useful with Params.
1228          */
1229         if (parse->limitCount)
1230         {
1231                 est = estimate_expression_value(root, parse->limitCount);
1232                 if (est && IsA(est, Const))
1233                 {
1234                         if (((Const *) est)->constisnull)
1235                         {
1236                                 /* NULL indicates LIMIT ALL, ie, no limit */
1237                                 *count_est = 0; /* treat as not present */
1238                         }
1239                         else
1240                         {
1241                                 *count_est = DatumGetInt64(((Const *) est)->constvalue);
1242                                 if (*count_est <= 0)
1243                                         *count_est = 1;         /* force to at least 1 */
1244                         }
1245                 }
1246                 else
1247                         *count_est = -1;        /* can't estimate */
1248         }
1249         else
1250                 *count_est = 0;                 /* not present */
1251
1252         if (parse->limitOffset)
1253         {
1254                 est = estimate_expression_value(root, parse->limitOffset);
1255                 if (est && IsA(est, Const))
1256                 {
1257                         if (((Const *) est)->constisnull)
1258                         {
1259                                 /* Treat NULL as no offset; the executor will too */
1260                                 *offset_est = 0;        /* treat as not present */
1261                         }
1262                         else
1263                         {
1264                                 *offset_est = DatumGetInt64(((Const *) est)->constvalue);
1265                                 if (*offset_est < 0)
1266                                         *offset_est = 0;        /* less than 0 is same as 0 */
1267                         }
1268                 }
1269                 else
1270                         *offset_est = -1;       /* can't estimate */
1271         }
1272         else
1273                 *offset_est = 0;                /* not present */
1274
1275         if (*count_est != 0)
1276         {
1277                 /*
1278                  * A LIMIT clause limits the absolute number of tuples returned.
1279                  * However, if it's not a constant LIMIT then we have to guess; for
1280                  * lack of a better idea, assume 10% of the plan's result is wanted.
1281                  */
1282                 if (*count_est < 0 || *offset_est < 0)
1283                 {
1284                         /* LIMIT or OFFSET is an expression ... punt ... */
1285                         limit_fraction = 0.10;
1286                 }
1287                 else
1288                 {
1289                         /* LIMIT (plus OFFSET, if any) is max number of tuples needed */
1290                         limit_fraction = (double) *count_est + (double) *offset_est;
1291                 }
1292
1293                 /*
1294                  * If we have absolute limits from both caller and LIMIT, use the
1295                  * smaller value; likewise if they are both fractional.  If one is
1296                  * fractional and the other absolute, we can't easily determine which
1297                  * is smaller, but we use the heuristic that the absolute will usually
1298                  * be smaller.
1299                  */
1300                 if (tuple_fraction >= 1.0)
1301                 {
1302                         if (limit_fraction >= 1.0)
1303                         {
1304                                 /* both absolute */
1305                                 tuple_fraction = Min(tuple_fraction, limit_fraction);
1306                         }
1307                         else
1308                         {
1309                                 /* caller absolute, limit fractional; use caller's value */
1310                         }
1311                 }
1312                 else if (tuple_fraction > 0.0)
1313                 {
1314                         if (limit_fraction >= 1.0)
1315                         {
1316                                 /* caller fractional, limit absolute; use limit */
1317                                 tuple_fraction = limit_fraction;
1318                         }
1319                         else
1320                         {
1321                                 /* both fractional */
1322                                 tuple_fraction = Min(tuple_fraction, limit_fraction);
1323                         }
1324                 }
1325                 else
1326                 {
1327                         /* no info from caller, just use limit */
1328                         tuple_fraction = limit_fraction;
1329                 }
1330         }
1331         else if (*offset_est != 0 && tuple_fraction > 0.0)
1332         {
1333                 /*
1334                  * We have an OFFSET but no LIMIT.      This acts entirely differently
1335                  * from the LIMIT case: here, we need to increase rather than decrease
1336                  * the caller's tuple_fraction, because the OFFSET acts to cause more
1337                  * tuples to be fetched instead of fewer.  This only matters if we got
1338                  * a tuple_fraction > 0, however.
1339                  *
1340                  * As above, use 10% if OFFSET is present but unestimatable.
1341                  */
1342                 if (*offset_est < 0)
1343                         limit_fraction = 0.10;
1344                 else
1345                         limit_fraction = (double) *offset_est;
1346
1347                 /*
1348                  * If we have absolute counts from both caller and OFFSET, add them
1349                  * together; likewise if they are both fractional.      If one is
1350                  * fractional and the other absolute, we want to take the larger, and
1351                  * we heuristically assume that's the fractional one.
1352                  */
1353                 if (tuple_fraction >= 1.0)
1354                 {
1355                         if (limit_fraction >= 1.0)
1356                         {
1357                                 /* both absolute, so add them together */
1358                                 tuple_fraction += limit_fraction;
1359                         }
1360                         else
1361                         {
1362                                 /* caller absolute, limit fractional; use limit */
1363                                 tuple_fraction = limit_fraction;
1364                         }
1365                 }
1366                 else
1367                 {
1368                         if (limit_fraction >= 1.0)
1369                         {
1370                                 /* caller fractional, limit absolute; use caller's value */
1371                         }
1372                         else
1373                         {
1374                                 /* both fractional, so add them together */
1375                                 tuple_fraction += limit_fraction;
1376                                 if (tuple_fraction >= 1.0)
1377                                         tuple_fraction = 0.0;           /* assume fetch all */
1378                         }
1379                 }
1380         }
1381
1382         return tuple_fraction;
1383 }
1384
1385 /*
1386  * extract_grouping_ops - make an array of the equality operator OIDs
1387  *              for the GROUP BY clause
1388  */
1389 static Oid *
1390 extract_grouping_ops(List *groupClause)
1391 {
1392         int                     numCols = list_length(groupClause);
1393         int                     colno = 0;
1394         Oid                *groupOperators;
1395         ListCell   *glitem;
1396
1397         groupOperators = (Oid *) palloc(sizeof(Oid) * numCols);
1398
1399         foreach(glitem, groupClause)
1400         {
1401                 GroupClause *groupcl = (GroupClause *) lfirst(glitem);
1402
1403                 groupOperators[colno] = get_equality_op_for_ordering_op(groupcl->sortop);
1404                 if (!OidIsValid(groupOperators[colno]))         /* shouldn't happen */
1405                         elog(ERROR, "could not find equality operator for ordering operator %u",
1406                                  groupcl->sortop);
1407                 colno++;
1408         }
1409
1410         return groupOperators;
1411 }
1412
1413 /*
1414  * choose_hashed_grouping - should we use hashed grouping?
1415  */
1416 static bool
1417 choose_hashed_grouping(PlannerInfo *root, double tuple_fraction,
1418                                            Path *cheapest_path, Path *sorted_path,
1419                                            Oid *groupOperators, double dNumGroups,
1420                                            AggClauseCounts *agg_counts)
1421 {
1422         int                     numGroupCols = list_length(root->parse->groupClause);
1423         double          cheapest_path_rows;
1424         int                     cheapest_path_width;
1425         Size            hashentrysize;
1426         List       *current_pathkeys;
1427         Path            hashed_p;
1428         Path            sorted_p;
1429         int                     i;
1430
1431         /*
1432          * Check can't-do-it conditions, including whether the grouping operators
1433          * are hashjoinable.  (We assume hashing is OK if they are marked
1434          * oprcanhash.  If there isn't actually a supporting hash function,
1435          * the executor will complain at runtime.)
1436          *
1437          * Executor doesn't support hashed aggregation with DISTINCT aggregates.
1438          * (Doing so would imply storing *all* the input values in the hash table,
1439          * which seems like a certain loser.)
1440          */
1441         if (!enable_hashagg)
1442                 return false;
1443         if (agg_counts->numDistinctAggs != 0)
1444                 return false;
1445         for (i = 0; i < numGroupCols; i++)
1446         {
1447                 if (!op_hashjoinable(groupOperators[i]))
1448                         return false;
1449         }
1450
1451         /*
1452          * Don't do it if it doesn't look like the hashtable will fit into
1453          * work_mem.
1454          *
1455          * Beware here of the possibility that cheapest_path->parent is NULL. This
1456          * could happen if user does something silly like SELECT 'foo' GROUP BY 1;
1457          */
1458         if (cheapest_path->parent)
1459         {
1460                 cheapest_path_rows = cheapest_path->parent->rows;
1461                 cheapest_path_width = cheapest_path->parent->width;
1462         }
1463         else
1464         {
1465                 cheapest_path_rows = 1; /* assume non-set result */
1466                 cheapest_path_width = 100;              /* arbitrary */
1467         }
1468
1469         /* Estimate per-hash-entry space at tuple width... */
1470         hashentrysize = MAXALIGN(cheapest_path_width) + MAXALIGN(sizeof(MinimalTupleData));
1471         /* plus space for pass-by-ref transition values... */
1472         hashentrysize += agg_counts->transitionSpace;
1473         /* plus the per-hash-entry overhead */
1474         hashentrysize += hash_agg_entry_size(agg_counts->numAggs);
1475
1476         if (hashentrysize * dNumGroups > work_mem * 1024L)
1477                 return false;
1478
1479         /*
1480          * See if the estimated cost is no more than doing it the other way. While
1481          * avoiding the need for sorted input is usually a win, the fact that the
1482          * output won't be sorted may be a loss; so we need to do an actual cost
1483          * comparison.
1484          *
1485          * We need to consider cheapest_path + hashagg [+ final sort] versus
1486          * either cheapest_path [+ sort] + group or agg [+ final sort] or
1487          * presorted_path + group or agg [+ final sort] where brackets indicate a
1488          * step that may not be needed. We assume query_planner() will have
1489          * returned a presorted path only if it's a winner compared to
1490          * cheapest_path for this purpose.
1491          *
1492          * These path variables are dummies that just hold cost fields; we don't
1493          * make actual Paths for these steps.
1494          */
1495         cost_agg(&hashed_p, root, AGG_HASHED, agg_counts->numAggs,
1496                          numGroupCols, dNumGroups,
1497                          cheapest_path->startup_cost, cheapest_path->total_cost,
1498                          cheapest_path_rows);
1499         /* Result of hashed agg is always unsorted */
1500         if (root->sort_pathkeys)
1501                 cost_sort(&hashed_p, root, root->sort_pathkeys, hashed_p.total_cost,
1502                                   dNumGroups, cheapest_path_width);
1503
1504         if (sorted_path)
1505         {
1506                 sorted_p.startup_cost = sorted_path->startup_cost;
1507                 sorted_p.total_cost = sorted_path->total_cost;
1508                 current_pathkeys = sorted_path->pathkeys;
1509         }
1510         else
1511         {
1512                 sorted_p.startup_cost = cheapest_path->startup_cost;
1513                 sorted_p.total_cost = cheapest_path->total_cost;
1514                 current_pathkeys = cheapest_path->pathkeys;
1515         }
1516         if (!pathkeys_contained_in(root->group_pathkeys, current_pathkeys))
1517         {
1518                 cost_sort(&sorted_p, root, root->group_pathkeys, sorted_p.total_cost,
1519                                   cheapest_path_rows, cheapest_path_width);
1520                 current_pathkeys = root->group_pathkeys;
1521         }
1522
1523         if (root->parse->hasAggs)
1524                 cost_agg(&sorted_p, root, AGG_SORTED, agg_counts->numAggs,
1525                                  numGroupCols, dNumGroups,
1526                                  sorted_p.startup_cost, sorted_p.total_cost,
1527                                  cheapest_path_rows);
1528         else
1529                 cost_group(&sorted_p, root, numGroupCols, dNumGroups,
1530                                    sorted_p.startup_cost, sorted_p.total_cost,
1531                                    cheapest_path_rows);
1532         /* The Agg or Group node will preserve ordering */
1533         if (root->sort_pathkeys &&
1534                 !pathkeys_contained_in(root->sort_pathkeys, current_pathkeys))
1535                 cost_sort(&sorted_p, root, root->sort_pathkeys, sorted_p.total_cost,
1536                                   dNumGroups, cheapest_path_width);
1537
1538         /*
1539          * Now make the decision using the top-level tuple fraction.  First we
1540          * have to convert an absolute count (LIMIT) into fractional form.
1541          */
1542         if (tuple_fraction >= 1.0)
1543                 tuple_fraction /= dNumGroups;
1544
1545         if (compare_fractional_path_costs(&hashed_p, &sorted_p,
1546                                                                           tuple_fraction) < 0)
1547         {
1548                 /* Hashed is cheaper, so use it */
1549                 return true;
1550         }
1551         return false;
1552 }
1553
1554 /*---------------
1555  * make_subplanTargetList
1556  *        Generate appropriate target list when grouping is required.
1557  *
1558  * When grouping_planner inserts Aggregate, Group, or Result plan nodes
1559  * above the result of query_planner, we typically want to pass a different
1560  * target list to query_planner than the outer plan nodes should have.
1561  * This routine generates the correct target list for the subplan.
1562  *
1563  * The initial target list passed from the parser already contains entries
1564  * for all ORDER BY and GROUP BY expressions, but it will not have entries
1565  * for variables used only in HAVING clauses; so we need to add those
1566  * variables to the subplan target list.  Also, we flatten all expressions
1567  * except GROUP BY items into their component variables; the other expressions
1568  * will be computed by the inserted nodes rather than by the subplan.
1569  * For example, given a query like
1570  *              SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
1571  * we want to pass this targetlist to the subplan:
1572  *              a,b,c,d,a+b
1573  * where the a+b target will be used by the Sort/Group steps, and the
1574  * other targets will be used for computing the final results.  (In the
1575  * above example we could theoretically suppress the a and b targets and
1576  * pass down only c,d,a+b, but it's not really worth the trouble to
1577  * eliminate simple var references from the subplan.  We will avoid doing
1578  * the extra computation to recompute a+b at the outer level; see
1579  * fix_upper_expr() in setrefs.c.)
1580  *
1581  * If we are grouping or aggregating, *and* there are no non-Var grouping
1582  * expressions, then the returned tlist is effectively dummy; we do not
1583  * need to force it to be evaluated, because all the Vars it contains
1584  * should be present in the output of query_planner anyway.
1585  *
1586  * 'tlist' is the query's target list.
1587  * 'groupColIdx' receives an array of column numbers for the GROUP BY
1588  *                      expressions (if there are any) in the subplan's target list.
1589  * 'need_tlist_eval' is set true if we really need to evaluate the
1590  *                      result tlist.
1591  *
1592  * The result is the targetlist to be passed to the subplan.
1593  *---------------
1594  */
1595 static List *
1596 make_subplanTargetList(PlannerInfo *root,
1597                                            List *tlist,
1598                                            AttrNumber **groupColIdx,
1599                                            bool *need_tlist_eval)
1600 {
1601         Query      *parse = root->parse;
1602         List       *sub_tlist;
1603         List       *extravars;
1604         int                     numCols;
1605
1606         *groupColIdx = NULL;
1607
1608         /*
1609          * If we're not grouping or aggregating, there's nothing to do here;
1610          * query_planner should receive the unmodified target list.
1611          */
1612         if (!parse->hasAggs && !parse->groupClause && !root->hasHavingQual)
1613         {
1614                 *need_tlist_eval = true;
1615                 return tlist;
1616         }
1617
1618         /*
1619          * Otherwise, start with a "flattened" tlist (having just the vars
1620          * mentioned in the targetlist and HAVING qual --- but not upper- level
1621          * Vars; they will be replaced by Params later on).
1622          */
1623         sub_tlist = flatten_tlist(tlist);
1624         extravars = pull_var_clause(parse->havingQual, false);
1625         sub_tlist = add_to_flat_tlist(sub_tlist, extravars);
1626         list_free(extravars);
1627         *need_tlist_eval = false;       /* only eval if not flat tlist */
1628
1629         /*
1630          * If grouping, create sub_tlist entries for all GROUP BY expressions
1631          * (GROUP BY items that are simple Vars should be in the list already),
1632          * and make an array showing where the group columns are in the sub_tlist.
1633          */
1634         numCols = list_length(parse->groupClause);
1635         if (numCols > 0)
1636         {
1637                 int                     keyno = 0;
1638                 AttrNumber *grpColIdx;
1639                 ListCell   *gl;
1640
1641                 grpColIdx = (AttrNumber *) palloc(sizeof(AttrNumber) * numCols);
1642                 *groupColIdx = grpColIdx;
1643
1644                 foreach(gl, parse->groupClause)
1645                 {
1646                         GroupClause *grpcl = (GroupClause *) lfirst(gl);
1647                         Node       *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1648                         TargetEntry *te = NULL;
1649                         ListCell   *sl;
1650
1651                         /* Find or make a matching sub_tlist entry */
1652                         foreach(sl, sub_tlist)
1653                         {
1654                                 te = (TargetEntry *) lfirst(sl);
1655                                 if (equal(groupexpr, te->expr))
1656                                         break;
1657                         }
1658                         if (!sl)
1659                         {
1660                                 te = makeTargetEntry((Expr *) groupexpr,
1661                                                                          list_length(sub_tlist) + 1,
1662                                                                          NULL,
1663                                                                          false);
1664                                 sub_tlist = lappend(sub_tlist, te);
1665                                 *need_tlist_eval = true;                /* it's not flat anymore */
1666                         }
1667
1668                         /* and save its resno */
1669                         grpColIdx[keyno++] = te->resno;
1670                 }
1671         }
1672
1673         return sub_tlist;
1674 }
1675
1676 /*
1677  * locate_grouping_columns
1678  *              Locate grouping columns in the tlist chosen by query_planner.
1679  *
1680  * This is only needed if we don't use the sub_tlist chosen by
1681  * make_subplanTargetList.      We have to forget the column indexes found
1682  * by that routine and re-locate the grouping vars in the real sub_tlist.
1683  */
1684 static void
1685 locate_grouping_columns(PlannerInfo *root,
1686                                                 List *tlist,
1687                                                 List *sub_tlist,
1688                                                 AttrNumber *groupColIdx)
1689 {
1690         int                     keyno = 0;
1691         ListCell   *gl;
1692
1693         /*
1694          * No work unless grouping.
1695          */
1696         if (!root->parse->groupClause)
1697         {
1698                 Assert(groupColIdx == NULL);
1699                 return;
1700         }
1701         Assert(groupColIdx != NULL);
1702
1703         foreach(gl, root->parse->groupClause)
1704         {
1705                 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1706                 Node       *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1707                 TargetEntry *te = NULL;
1708                 ListCell   *sl;
1709
1710                 foreach(sl, sub_tlist)
1711                 {
1712                         te = (TargetEntry *) lfirst(sl);
1713                         if (equal(groupexpr, te->expr))
1714                                 break;
1715                 }
1716                 if (!sl)
1717                         elog(ERROR, "failed to locate grouping columns");
1718
1719                 groupColIdx[keyno++] = te->resno;
1720         }
1721 }
1722
1723 /*
1724  * postprocess_setop_tlist
1725  *        Fix up targetlist returned by plan_set_operations().
1726  *
1727  * We need to transpose sort key info from the orig_tlist into new_tlist.
1728  * NOTE: this would not be good enough if we supported resjunk sort keys
1729  * for results of set operations --- then, we'd need to project a whole
1730  * new tlist to evaluate the resjunk columns.  For now, just ereport if we
1731  * find any resjunk columns in orig_tlist.
1732  */
1733 static List *
1734 postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
1735 {
1736         ListCell   *l;
1737         ListCell   *orig_tlist_item = list_head(orig_tlist);
1738
1739         foreach(l, new_tlist)
1740         {
1741                 TargetEntry *new_tle = (TargetEntry *) lfirst(l);
1742                 TargetEntry *orig_tle;
1743
1744                 /* ignore resjunk columns in setop result */
1745                 if (new_tle->resjunk)
1746                         continue;
1747
1748                 Assert(orig_tlist_item != NULL);
1749                 orig_tle = (TargetEntry *) lfirst(orig_tlist_item);
1750                 orig_tlist_item = lnext(orig_tlist_item);
1751                 if (orig_tle->resjunk)  /* should not happen */
1752                         elog(ERROR, "resjunk output columns are not implemented");
1753                 Assert(new_tle->resno == orig_tle->resno);
1754                 new_tle->ressortgroupref = orig_tle->ressortgroupref;
1755         }
1756         if (orig_tlist_item != NULL)
1757                 elog(ERROR, "resjunk output columns are not implemented");
1758         return new_tlist;
1759 }