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) 1996-2006, 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 1.207 2006/08/05 17:21:52 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/syscache.h"
  42
  43
  44 ParamListInfo PlannerBoundParamList = NULL;             /* current boundParams */
  45
  46
  47 /* Expression kind codes for preprocess_expression */
  48 #define EXPRKIND_QUAL           0
  49 #define EXPRKIND_TARGET         1
  50 #define EXPRKIND_RTFUNC         2
  51 #define EXPRKIND_VALUES         3
  52 #define EXPRKIND_LIMIT          4
  53 #define EXPRKIND_ININFO         5
  54 #define EXPRKIND_APPINFO        6
  55
  56
  57 static Node *preprocess_expression(PlannerInfo *root, Node *expr, int kind);
  58 static void preprocess_qual_conditions(PlannerInfo *root, Node *jtnode);
  59 static Plan *inheritance_planner(PlannerInfo *root);
  60 static Plan *grouping_planner(PlannerInfo *root, double tuple_fraction);
  61 static bool is_dummy_plan(Plan *plan);
  62 static double preprocess_limit(PlannerInfo *root,
  63                                  double tuple_fraction,
  64                                  int64 *offset_est, int64 *count_est);
  65 static bool choose_hashed_grouping(PlannerInfo *root, double tuple_fraction,
  66                                            Path *cheapest_path, Path *sorted_path,
  67                                            double dNumGroups, AggClauseCounts *agg_counts);
  68 static bool hash_safe_grouping(PlannerInfo *root);
  69 static List *make_subplanTargetList(PlannerInfo *root, List *tlist,
  70                                            AttrNumber **groupColIdx, bool *need_tlist_eval);
  71 static void locate_grouping_columns(PlannerInfo *root,
  72                                                 List *tlist,
  73                                                 List *sub_tlist,
  74                                                 AttrNumber *groupColIdx);
  75 static List *postprocess_setop_tlist(List *new_tlist, List *orig_tlist);
  76
  77
  78 /*****************************************************************************
  79  *
  80  *         Query optimizer entry point
  81  *
  82  *****************************************************************************/
  83 Plan *
  84 planner(Query *parse, bool isCursor, int cursorOptions,
  85                 ParamListInfo boundParams)
  86 {
  87         double          tuple_fraction;
  88         Plan       *result_plan;
  89         Index           save_PlannerQueryLevel;
  90         List       *save_PlannerParamList;
  91         ParamListInfo save_PlannerBoundParamList;
  92
  93         /*
  94          * The planner can be called recursively (an example is when
  95          * eval_const_expressions tries to pre-evaluate an SQL function). So,
  96          * these global state variables must be saved and restored.
  97          *
  98          * Query level and the param list cannot be moved into the per-query
  99          * PlannerInfo structure since their whole purpose is communication across
 100          * multiple sub-queries. Also, boundParams is explicitly info from outside
 101          * the query, and so is likewise better handled as a global variable.
 102          *
 103          * Note we do NOT save and restore PlannerPlanId: it exists to assign
 104          * unique IDs to SubPlan nodes, and we want those IDs to be unique for the
 105          * life of a backend.  Also, PlannerInitPlan is saved/restored in
 106          * subquery_planner, not here.
 107          */
 108         save_PlannerQueryLevel = PlannerQueryLevel;
 109         save_PlannerParamList = PlannerParamList;
 110         save_PlannerBoundParamList = PlannerBoundParamList;
 111
 112         /* Initialize state for handling outer-level references and params */
 113         PlannerQueryLevel = 0;          /* will be 1 in top-level subquery_planner */
 114         PlannerParamList = NIL;
 115         PlannerBoundParamList = boundParams;
 116
 117         /* Determine what fraction of the plan is likely to be scanned */
 118         if (isCursor)
 119         {
 120                 /*
 121                  * We have no real idea how many tuples the user will ultimately FETCH
 122                  * from a cursor, but it seems a good bet that he doesn't want 'em
 123                  * all.  Optimize for 10% retrieval (you gotta better number?  Should
 124                  * this be a SETtable parameter?)
 125                  */
 126                 tuple_fraction = 0.10;
 127         }
 128         else
 129         {
 130                 /* Default assumption is we need all the tuples */
 131                 tuple_fraction = 0.0;
 132         }
 133
 134         /* primary planning entry point (may recurse for subqueries) */
 135         result_plan = subquery_planner(parse, tuple_fraction, NULL);
 136
 137         /* check we popped out the right number of levels */
 138         Assert(PlannerQueryLevel == 0);
 139
 140         /*
 141          * If creating a plan for a scrollable cursor, make sure it can run
 142          * backwards on demand.  Add a Material node at the top at need.
 143          */
 144         if (isCursor && (cursorOptions & CURSOR_OPT_SCROLL))
 145         {
 146                 if (!ExecSupportsBackwardScan(result_plan))
 147                         result_plan = materialize_finished_plan(result_plan);
 148         }
 149
 150         /* final cleanup of the plan */
 151         result_plan = set_plan_references(result_plan, parse->rtable);
 152
 153         /* executor wants to know total number of Params used overall */
 154         result_plan->nParamExec = list_length(PlannerParamList);
 155
 156         /* restore state for outer planner, if any */
 157         PlannerQueryLevel = save_PlannerQueryLevel;
 158         PlannerParamList = save_PlannerParamList;
 159         PlannerBoundParamList = save_PlannerBoundParamList;
 160
 161         return result_plan;
 162 }
 163
 164
 165 /*--------------------
 166  * subquery_planner
 167  *        Invokes the planner on a subquery.  We recurse to here for each
 168  *        sub-SELECT found in the query tree.
 169  *
 170  * parse is the querytree produced by the parser & rewriter.
 171  * tuple_fraction is the fraction of tuples we expect will be retrieved.
 172  * tuple_fraction is interpreted as explained for grouping_planner, below.
 173  *
 174  * If subquery_pathkeys isn't NULL, it receives a list of pathkeys indicating
 175  * the output sort ordering of the completed plan.
 176  *
 177  * Basically, this routine does the stuff that should only be done once
 178  * per Query object.  It then calls grouping_planner.  At one time,
 179  * grouping_planner could be invoked recursively on the same Query object;
 180  * that's not currently true, but we keep the separation between the two
 181  * routines anyway, in case we need it again someday.
 182  *
 183  * subquery_planner will be called recursively to handle sub-Query nodes
 184  * found within the query's expressions and rangetable.
 185  *
 186  * Returns a query plan.
 187  *--------------------
 188  */
 189 Plan *
 190 subquery_planner(Query *parse, double tuple_fraction,
 191                                  List **subquery_pathkeys)
 192 {
 193         List       *saved_initplan = PlannerInitPlan;
 194         int                     saved_planid = PlannerPlanId;
 195         PlannerInfo *root;
 196         Plan       *plan;
 197         List       *newHaving;
 198         ListCell   *l;
 199
 200         /* Set up for a new level of subquery */
 201         PlannerQueryLevel++;
 202         PlannerInitPlan = NIL;
 203
 204         /* Create a PlannerInfo data structure for this subquery */
 205         root = makeNode(PlannerInfo);
 206         root->parse = parse;
 207         root->in_info_list = NIL;
 208         root->append_rel_list = NIL;
 209
 210         /*
 211          * Look for IN clauses at the top level of WHERE, and transform them into
 212          * joins.  Note that this step only handles IN clauses originally at top
 213          * level of WHERE; if we pull up any subqueries in the next step, their
 214          * INs are processed just before pulling them up.
 215          */
 216         if (parse->hasSubLinks)
 217                 parse->jointree->quals = pull_up_IN_clauses(root,
 218                                                                                                         parse->jointree->quals);
 219
 220         /*
 221          * Check to see if any subqueries in the rangetable can be merged into
 222          * this query.
 223          */
 224         parse->jointree = (FromExpr *)
 225                 pull_up_subqueries(root, (Node *) parse->jointree, false, false);
 226
 227         /*
 228          * Detect whether any rangetable entries are RTE_JOIN kind; if not, we can
 229          * avoid the expense of doing flatten_join_alias_vars().  Also check for
 230          * outer joins --- if none, we can skip reduce_outer_joins() and some
 231          * other processing.  This must be done after we have done
 232          * pull_up_subqueries, of course.
 233          *
 234          * Note: if reduce_outer_joins manages to eliminate all outer joins,
 235          * root->hasOuterJoins is not reset currently.  This is OK since its
 236          * purpose is merely to suppress unnecessary processing in simple cases.
 237          */
 238         root->hasJoinRTEs = false;
 239         root->hasOuterJoins = false;
 240         foreach(l, parse->rtable)
 241         {
 242                 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
 243
 244                 if (rte->rtekind == RTE_JOIN)
 245                 {
 246                         root->hasJoinRTEs = true;
 247                         if (IS_OUTER_JOIN(rte->jointype))
 248                         {
 249                                 root->hasOuterJoins = true;
 250                                 /* Can quit scanning once we find an outer join */
 251                                 break;
 252                         }
 253                 }
 254         }
 255
 256         /*
 257          * Expand any rangetable entries that are inheritance sets into "append
 258          * relations".  This can add entries to the rangetable, but they must be
 259          * plain base relations not joins, so it's OK (and marginally more
 260          * efficient) to do it after checking for join RTEs.  We must do it after
 261          * pulling up subqueries, else we'd fail to handle inherited tables in
 262          * subqueries.
 263          */
 264         expand_inherited_tables(root);
 265
 266         /*
 267          * Set hasHavingQual to remember if HAVING clause is present.  Needed
 268          * because preprocess_expression will reduce a constant-true condition to
 269          * an empty qual list ... but "HAVING TRUE" is not a semantic no-op.
 270          */
 271         root->hasHavingQual = (parse->havingQual != NULL);
 272
 273         /* Clear this flag; might get set in distribute_qual_to_rels */
 274         root->hasPseudoConstantQuals = false;
 275
 276         /*
 277          * Do expression preprocessing on targetlist and quals.
 278          */
 279         parse->targetList = (List *)
 280                 preprocess_expression(root, (Node *) parse->targetList,
 281                                                           EXPRKIND_TARGET);
 282
 283         preprocess_qual_conditions(root, (Node *) parse->jointree);
 284
 285         parse->havingQual = preprocess_expression(root, parse->havingQual,
 286                                                                                           EXPRKIND_QUAL);
 287
 288         parse->limitOffset = preprocess_expression(root, parse->limitOffset,
 289                                                                                            EXPRKIND_LIMIT);
 290         parse->limitCount = preprocess_expression(root, parse->limitCount,
 291                                                                                           EXPRKIND_LIMIT);
 292
 293         root->in_info_list = (List *)
 294                 preprocess_expression(root, (Node *) root->in_info_list,
 295                                                           EXPRKIND_ININFO);
 296         root->append_rel_list = (List *)
 297                 preprocess_expression(root, (Node *) root->append_rel_list,
 298                                                           EXPRKIND_APPINFO);
 299
 300         /* Also need to preprocess expressions for function and values RTEs */
 301         foreach(l, parse->rtable)
 302         {
 303                 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
 304
 305                 if (rte->rtekind == RTE_FUNCTION)
 306                         rte->funcexpr = preprocess_expression(root, rte->funcexpr,
 307                                                                                                   EXPRKIND_RTFUNC);
 308                 else if (rte->rtekind == RTE_VALUES)
 309                         rte->values_lists = (List *)
 310                                 preprocess_expression(root, (Node *) rte->values_lists,
 311                                                                           EXPRKIND_VALUES);
 312         }
 313
 314         /*
 315          * In some cases we may want to transfer a HAVING clause into WHERE. We
 316          * cannot do so if the HAVING clause contains aggregates (obviously) or
 317          * volatile functions (since a HAVING clause is supposed to be executed
 318          * only once per group).  Also, it may be that the clause is so expensive
 319          * to execute that we're better off doing it only once per group, despite
 320          * the loss of selectivity.  This is hard to estimate short of doing the
 321          * entire planning process twice, so we use a heuristic: clauses
 322          * containing subplans are left in HAVING.      Otherwise, we move or copy the
 323          * HAVING clause into WHERE, in hopes of eliminating tuples before
 324          * aggregation instead of after.
 325          *
 326          * If the query has explicit grouping then we can simply move such a
 327          * clause into WHERE; any group that fails the clause will not be in the
 328          * output because none of its tuples will reach the grouping or
 329          * aggregation stage.  Otherwise we must have a degenerate (variable-free)
 330          * HAVING clause, which we put in WHERE so that query_planner() can use it
 331          * in a gating Result node, but also keep in HAVING to ensure that we
 332          * don't emit a bogus aggregated row. (This could be done better, but it
 333          * seems not worth optimizing.)
 334          *
 335          * Note that both havingQual and parse->jointree->quals are in
 336          * implicitly-ANDed-list form at this point, even though they are declared
 337          * as Node *.
 338          */
 339         newHaving = NIL;
 340         foreach(l, (List *) parse->havingQual)
 341         {
 342                 Node       *havingclause = (Node *) lfirst(l);
 343
 344                 if (contain_agg_clause(havingclause) ||
 345                         contain_volatile_functions(havingclause) ||
 346                         contain_subplans(havingclause))
 347                 {
 348                         /* keep it in HAVING */
 349                         newHaving = lappend(newHaving, havingclause);
 350                 }
 351                 else if (parse->groupClause)
 352                 {
 353                         /* move it to WHERE */
 354                         parse->jointree->quals = (Node *)
 355                                 lappend((List *) parse->jointree->quals, havingclause);
 356                 }
 357                 else
 358                 {
 359                         /* put a copy in WHERE, keep it in HAVING */
 360                         parse->jointree->quals = (Node *)
 361                                 lappend((List *) parse->jointree->quals,
 362                                                 copyObject(havingclause));
 363                         newHaving = lappend(newHaving, havingclause);
 364                 }
 365         }
 366         parse->havingQual = (Node *) newHaving;
 367
 368         /*
 369          * If we have any outer joins, try to reduce them to plain inner joins.
 370          * This step is most easily done after we've done expression
 371          * preprocessing.
 372          */
 373         if (root->hasOuterJoins)
 374                 reduce_outer_joins(root);
 375
 376         /*
 377          * Do the main planning.  If we have an inherited target relation, that
 378          * needs special processing, else go straight to grouping_planner.
 379          */
 380         if (parse->resultRelation &&
 381                 rt_fetch(parse->resultRelation, parse->rtable)->inh)
 382                 plan = inheritance_planner(root);
 383         else
 384                 plan = grouping_planner(root, tuple_fraction);
 385
 386         /*
 387          * If any subplans were generated, or if we're inside a subplan, build
 388          * initPlan list and extParam/allParam sets for plan nodes, and attach the
 389          * initPlans to the top plan node.
 390          */
 391         if (PlannerPlanId != saved_planid || PlannerQueryLevel > 1)
 392                 SS_finalize_plan(plan, parse->rtable);
 393
 394         /* Return sort ordering info if caller wants it */
 395         if (subquery_pathkeys)
 396                 *subquery_pathkeys = root->query_pathkeys;
 397
 398         /* Return to outer subquery context */
 399         PlannerQueryLevel--;
 400         PlannerInitPlan = saved_initplan;
 401         /* we do NOT restore PlannerPlanId; that's not an oversight! */
 402
 403         return plan;
 404 }
 405
 406 /*
 407  * preprocess_expression
 408  *              Do subquery_planner's preprocessing work for an expression,
 409  *              which can be a targetlist, a WHERE clause (including JOIN/ON
 410  *              conditions), or a HAVING clause.
 411  */
 412 static Node *
 413 preprocess_expression(PlannerInfo *root, Node *expr, int kind)
 414 {
 415         /*
 416          * Fall out quickly if expression is empty.  This occurs often enough to
 417          * be worth checking.  Note that null->null is the correct conversion for
 418          * implicit-AND result format, too.
 419          */
 420         if (expr == NULL)
 421                 return NULL;
 422
 423         /*
 424          * If the query has any join RTEs, replace join alias variables with
 425          * base-relation variables. We must do this before sublink processing,
 426          * else sublinks expanded out from join aliases wouldn't get processed.
 427          * We can skip it in VALUES lists, however, since they can't contain
 428          * any Vars at all.
 429          */
 430         if (root->hasJoinRTEs && kind != EXPRKIND_VALUES)
 431                 expr = flatten_join_alias_vars(root, expr);
 432
 433         /*
 434          * Simplify constant expressions.
 435          *
 436          * Note: this also flattens nested AND and OR expressions into N-argument
 437          * form.  All processing of a qual expression after this point must be
 438          * careful to maintain AND/OR flatness --- that is, do not generate a tree
 439          * with AND directly under AND, nor OR directly under OR.
 440          *
 441          * Because this is a relatively expensive process, we skip it when the
 442          * query is trivial, such as "SELECT 2+2;" or "INSERT ... VALUES()". The
 443          * expression will only be evaluated once anyway, so no point in
 444          * pre-simplifying; we can't execute it any faster than the executor can,
 445          * and we will waste cycles copying the tree.  Notice however that we
 446          * still must do it for quals (to get AND/OR flatness); and if we are in a
 447          * subquery we should not assume it will be done only once.
 448          *
 449          * For VALUES lists we never do this at all, again on the grounds that
 450          * we should optimize for one-time evaluation.
 451          */
 452         if (kind != EXPRKIND_VALUES &&
 453                 (root->parse->jointree->fromlist != NIL ||
 454                  kind == EXPRKIND_QUAL ||
 455                  PlannerQueryLevel > 1))
 456                 expr = eval_const_expressions(expr);
 457
 458         /*
 459          * If it's a qual or havingQual, canonicalize it.
 460          */
 461         if (kind == EXPRKIND_QUAL)
 462         {
 463                 expr = (Node *) canonicalize_qual((Expr *) expr);
 464
 465 #ifdef OPTIMIZER_DEBUG
 466                 printf("After canonicalize_qual()\n");
 467                 pprint(expr);
 468 #endif
 469         }
 470
 471         /* Expand SubLinks to SubPlans */
 472         if (root->parse->hasSubLinks)
 473                 expr = SS_process_sublinks(expr, (kind == EXPRKIND_QUAL));
 474
 475         /*
 476          * XXX do not insert anything here unless you have grokked the comments in
 477          * SS_replace_correlation_vars ...
 478          */
 479
 480         /* Replace uplevel vars with Param nodes (this IS possible in VALUES) */
 481         if (PlannerQueryLevel > 1)
 482                 expr = SS_replace_correlation_vars(expr);
 483
 484         /*
 485          * If it's a qual or havingQual, convert it to implicit-AND format. (We
 486          * don't want to do this before eval_const_expressions, since the latter
 487          * would be unable to simplify a top-level AND correctly. Also,
 488          * SS_process_sublinks expects explicit-AND format.)
 489          */
 490         if (kind == EXPRKIND_QUAL)
 491                 expr = (Node *) make_ands_implicit((Expr *) expr);
 492
 493         return expr;
 494 }
 495
 496 /*
 497  * preprocess_qual_conditions
 498  *              Recursively scan the query's jointree and do subquery_planner's
 499  *              preprocessing work on each qual condition found therein.
 500  */
 501 static void
 502 preprocess_qual_conditions(PlannerInfo *root, Node *jtnode)
 503 {
 504         if (jtnode == NULL)
 505                 return;
 506         if (IsA(jtnode, RangeTblRef))
 507         {
 508                 /* nothing to do here */
 509         }
 510         else if (IsA(jtnode, FromExpr))
 511         {
 512                 FromExpr   *f = (FromExpr *) jtnode;
 513                 ListCell   *l;
 514
 515                 foreach(l, f->fromlist)
 516                         preprocess_qual_conditions(root, lfirst(l));
 517
 518                 f->quals = preprocess_expression(root, f->quals, EXPRKIND_QUAL);
 519         }
 520         else if (IsA(jtnode, JoinExpr))
 521         {
 522                 JoinExpr   *j = (JoinExpr *) jtnode;
 523
 524                 preprocess_qual_conditions(root, j->larg);
 525                 preprocess_qual_conditions(root, j->rarg);
 526
 527                 j->quals = preprocess_expression(root, j->quals, EXPRKIND_QUAL);
 528         }
 529         else
 530                 elog(ERROR, "unrecognized node type: %d",
 531                          (int) nodeTag(jtnode));
 532 }
 533
 534 /*
 535  * inheritance_planner
 536  *        Generate a plan in the case where the result relation is an
 537  *        inheritance set.
 538  *
 539  * We have to handle this case differently from cases where a source relation
 540  * is an inheritance set. Source inheritance is expanded at the bottom of the
 541  * plan tree (see allpaths.c), but target inheritance has to be expanded at
 542  * the top.  The reason is that for UPDATE, each target relation needs a
 543  * different targetlist matching its own column set.  Also, for both UPDATE
 544  * and DELETE, the executor needs the Append plan node at the top, else it
 545  * can't keep track of which table is the current target table.  Fortunately,
 546  * the UPDATE/DELETE target can never be the nullable side of an outer join,
 547  * so it's OK to generate the plan this way.
 548  *
 549  * Returns a query plan.
 550  */
 551 static Plan *
 552 inheritance_planner(PlannerInfo *root)
 553 {
 554         Query      *parse = root->parse;
 555         int                     parentRTindex = parse->resultRelation;
 556         List       *subplans = NIL;
 557         List       *rtable = NIL;
 558         List       *tlist = NIL;
 559         PlannerInfo subroot;
 560         ListCell   *l;
 561
 562         parse->resultRelations = NIL;
 563
 564         foreach(l, root->append_rel_list)
 565         {
 566                 AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
 567                 Plan       *subplan;
 568
 569                 /* append_rel_list contains all append rels; ignore others */
 570                 if (appinfo->parent_relid != parentRTindex)
 571                         continue;
 572
 573                 /*
 574                  * Generate modified query with this rel as target.  We have to be
 575                  * prepared to translate varnos in in_info_list as well as in the
 576                  * Query proper.
 577                  */
 578                 memcpy(&subroot, root, sizeof(PlannerInfo));
 579                 subroot.parse = (Query *)
 580                         adjust_appendrel_attrs((Node *) parse,
 581                                                                    appinfo);
 582                 subroot.in_info_list = (List *)
 583                         adjust_appendrel_attrs((Node *) root->in_info_list,
 584                                                                    appinfo);
 585                 /* There shouldn't be any OJ info to translate, as yet */
 586                 Assert(subroot.oj_info_list == NIL);
 587
 588                 /* Generate plan */
 589                 subplan = grouping_planner(&subroot, 0.0 /* retrieve all tuples */ );
 590
 591                 /*
 592                  * If this child rel was excluded by constraint exclusion, exclude
 593                  * it from the plan.
 594                  */
 595                 if (is_dummy_plan(subplan))
 596                         continue;
 597
 598                 /* Save rtable and tlist from first rel for use below */
 599                 if (subplans == NIL)
 600                 {
 601                         rtable = subroot.parse->rtable;
 602                         tlist = subplan->targetlist;
 603                 }
 604
 605                 subplans = lappend(subplans, subplan);
 606
 607                 /* Build target-relations list for the executor */
 608                 parse->resultRelations = lappend_int(parse->resultRelations,
 609                                                                                          appinfo->child_relid);
 610         }
 611
 612         /* Mark result as unordered (probably unnecessary) */
 613         root->query_pathkeys = NIL;
 614
 615         /*
 616          * If we managed to exclude every child rel, return a dummy plan
 617          */
 618         if (subplans == NIL)
 619                 return (Plan *) make_result(tlist,
 620                                                                         (Node *) list_make1(makeBoolConst(false,
 621                                                                                                                                           false)),
 622                                                                         NULL);
 623
 624         /*
 625          * Planning might have modified the rangetable, due to changes of the
 626          * Query structures inside subquery RTEs.  We have to ensure that this
 627          * gets propagated back to the master copy.  But can't do this until we
 628          * are done planning, because all the calls to grouping_planner need
 629          * virgin sub-Queries to work from.  (We are effectively assuming that
 630          * sub-Queries will get planned identically each time, or at least that
 631          * the impacts on their rangetables will be the same each time.)
 632          *
 633          * XXX should clean this up someday
 634          */
 635         parse->rtable = rtable;
 636
 637         return (Plan *) make_append(subplans, true, tlist);
 638 }
 639
 640 /*--------------------
 641  * grouping_planner
 642  *        Perform planning steps related to grouping, aggregation, etc.
 643  *        This primarily means adding top-level processing to the basic
 644  *        query plan produced by query_planner.
 645  *
 646  * tuple_fraction is the fraction of tuples we expect will be retrieved
 647  *
 648  * tuple_fraction is interpreted as follows:
 649  *        0: expect all tuples to be retrieved (normal case)
 650  *        0 < tuple_fraction < 1: expect the given fraction of tuples available
 651  *              from the plan to be retrieved
 652  *        tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
 653  *              expected to be retrieved (ie, a LIMIT specification)
 654  *
 655  * Returns a query plan.  Also, root->query_pathkeys is returned as the
 656  * actual output ordering of the plan (in pathkey format).
 657  *--------------------
 658  */
 659 static Plan *
 660 grouping_planner(PlannerInfo *root, double tuple_fraction)
 661 {
 662         Query      *parse = root->parse;
 663         List       *tlist = parse->targetList;
 664         int64           offset_est = 0;
 665         int64           count_est = 0;
 666         Plan       *result_plan;
 667         List       *current_pathkeys;
 668         List       *sort_pathkeys;
 669         double          dNumGroups = 0;
 670
 671         /* Tweak caller-supplied tuple_fraction if have LIMIT/OFFSET */
 672         if (parse->limitCount || parse->limitOffset)
 673                 tuple_fraction = preprocess_limit(root, tuple_fraction,
 674                                                                                   &offset_est, &count_est);
 675
 676         if (parse->setOperations)
 677         {
 678                 List       *set_sortclauses;
 679
 680                 /*
 681                  * If there's a top-level ORDER BY, assume we have to fetch all the
 682                  * tuples.      This might seem too simplistic given all the hackery below
 683                  * to possibly avoid the sort ... but a nonzero tuple_fraction is only
 684                  * of use to plan_set_operations() when the setop is UNION ALL, and
 685                  * the result of UNION ALL is always unsorted.
 686                  */
 687                 if (parse->sortClause)
 688                         tuple_fraction = 0.0;
 689
 690                 /*
 691                  * Construct the plan for set operations.  The result will not need
 692                  * any work except perhaps a top-level sort and/or LIMIT.
 693                  */
 694                 result_plan = plan_set_operations(root, tuple_fraction,
 695                                                                                   &set_sortclauses);
 696
 697                 /*
 698                  * Calculate pathkeys representing the sort order (if any) of the set
 699                  * operation's result.  We have to do this before overwriting the sort
 700                  * key information...
 701                  */
 702                 current_pathkeys = make_pathkeys_for_sortclauses(set_sortclauses,
 703                                                                                                         result_plan->targetlist);
 704                 current_pathkeys = canonicalize_pathkeys(root, current_pathkeys);
 705
 706                 /*
 707                  * We should not need to call preprocess_targetlist, since we must be
 708                  * in a SELECT query node.      Instead, use the targetlist returned by
 709                  * plan_set_operations (since this tells whether it returned any
 710                  * resjunk columns!), and transfer any sort key information from the
 711                  * original tlist.
 712                  */
 713                 Assert(parse->commandType == CMD_SELECT);
 714
 715                 tlist = postprocess_setop_tlist(result_plan->targetlist, tlist);
 716
 717                 /*
 718                  * Can't handle FOR UPDATE/SHARE here (parser should have checked
 719                  * already, but let's make sure).
 720                  */
 721                 if (parse->rowMarks)
 722                         ereport(ERROR,
 723                                         (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
 724                                          errmsg("SELECT FOR UPDATE/SHARE is not allowed with UNION/INTERSECT/EXCEPT")));
 725
 726                 /*
 727                  * Calculate pathkeys that represent result ordering requirements
 728                  */
 729                 sort_pathkeys = make_pathkeys_for_sortclauses(parse->sortClause,
 730                                                                                                           tlist);
 731                 sort_pathkeys = canonicalize_pathkeys(root, sort_pathkeys);
 732         }
 733         else
 734         {
 735                 /* No set operations, do regular planning */
 736                 List       *sub_tlist;
 737                 List       *group_pathkeys;
 738                 AttrNumber *groupColIdx = NULL;
 739                 bool            need_tlist_eval = true;
 740                 QualCost        tlist_cost;
 741                 Path       *cheapest_path;
 742                 Path       *sorted_path;
 743                 Path       *best_path;
 744                 long            numGroups = 0;
 745                 AggClauseCounts agg_counts;
 746                 int                     numGroupCols = list_length(parse->groupClause);
 747                 bool            use_hashed_grouping = false;
 748
 749                 MemSet(&agg_counts, 0, sizeof(AggClauseCounts));
 750
 751                 /* Preprocess targetlist */
 752                 tlist = preprocess_targetlist(root, tlist);
 753
 754                 /*
 755                  * Generate appropriate target list for subplan; may be different from
 756                  * tlist if grouping or aggregation is needed.
 757                  */
 758                 sub_tlist = make_subplanTargetList(root, tlist,
 759                                                                                    &groupColIdx, &need_tlist_eval);
 760
 761                 /*
 762                  * Calculate pathkeys that represent grouping/ordering requirements.
 763                  * Stash them in PlannerInfo so that query_planner can canonicalize
 764                  * them.
 765                  */
 766                 root->group_pathkeys =
 767                         make_pathkeys_for_sortclauses(parse->groupClause, tlist);
 768                 root->sort_pathkeys =
 769                         make_pathkeys_for_sortclauses(parse->sortClause, tlist);
 770
 771                 /*
 772                  * Will need actual number of aggregates for estimating costs.
 773                  *
 774                  * Note: we do not attempt to detect duplicate aggregates here; a
 775                  * somewhat-overestimated count is okay for our present purposes.
 776                  *
 777                  * Note: think not that we can turn off hasAggs if we find no aggs. It
 778                  * is possible for constant-expression simplification to remove all
 779                  * explicit references to aggs, but we still have to follow the
 780                  * aggregate semantics (eg, producing only one output row).
 781                  */
 782                 if (parse->hasAggs)
 783                 {
 784                         count_agg_clauses((Node *) tlist, &agg_counts);
 785                         count_agg_clauses(parse->havingQual, &agg_counts);
 786                 }
 787
 788                 /*
 789                  * Figure out whether we need a sorted result from query_planner.
 790                  *
 791                  * If we have a GROUP BY clause, then we want a result sorted properly
 792                  * for grouping.  Otherwise, if there is an ORDER BY clause, we want
 793                  * to sort by the ORDER BY clause.      (Note: if we have both, and ORDER
 794                  * BY is a superset of GROUP BY, it would be tempting to request sort
 795                  * by ORDER BY --- but that might just leave us failing to exploit an
 796                  * available sort order at all. Needs more thought...)
 797                  */
 798                 if (parse->groupClause)
 799                         root->query_pathkeys = root->group_pathkeys;
 800                 else if (parse->sortClause)
 801                         root->query_pathkeys = root->sort_pathkeys;
 802                 else
 803                         root->query_pathkeys = NIL;
 804
 805                 /*
 806                  * Generate the best unsorted and presorted paths for this Query (but
 807                  * note there may not be any presorted path).  query_planner will also
 808                  * estimate the number of groups in the query, and canonicalize all
 809                  * the pathkeys.
 810                  */
 811                 query_planner(root, sub_tlist, tuple_fraction,
 812                                           &cheapest_path, &sorted_path, &dNumGroups);
 813
 814                 group_pathkeys = root->group_pathkeys;
 815                 sort_pathkeys = root->sort_pathkeys;
 816
 817                 /*
 818                  * If grouping, decide whether we want to use hashed grouping.
 819                  */
 820                 if (parse->groupClause)
 821                 {
 822                         use_hashed_grouping =
 823                                 choose_hashed_grouping(root, tuple_fraction,
 824                                                                            cheapest_path, sorted_path,
 825                                                                            dNumGroups, &agg_counts);
 826
 827                         /* Also convert # groups to long int --- but 'ware overflow! */
 828                         numGroups = (long) Min(dNumGroups, (double) LONG_MAX);
 829                 }
 830
 831                 /*
 832                  * Select the best path.  If we are doing hashed grouping, we will
 833                  * always read all the input tuples, so use the cheapest-total path.
 834                  * Otherwise, trust query_planner's decision about which to use.
 835                  */
 836                 if (use_hashed_grouping || !sorted_path)
 837                         best_path = cheapest_path;
 838                 else
 839                         best_path = sorted_path;
 840
 841                 /*
 842                  * Check to see if it's possible to optimize MIN/MAX aggregates. If
 843                  * so, we will forget all the work we did so far to choose a "regular"
 844                  * path ... but we had to do it anyway to be able to tell which way is
 845                  * cheaper.
 846                  */
 847                 result_plan = optimize_minmax_aggregates(root,
 848                                                                                                  tlist,
 849                                                                                                  best_path);
 850                 if (result_plan != NULL)
 851                 {
 852                         /*
 853                          * optimize_minmax_aggregates generated the full plan, with the
 854                          * right tlist, and it has no sort order.
 855                          */
 856                         current_pathkeys = NIL;
 857                 }
 858                 else
 859                 {
 860                         /*
 861                          * Normal case --- create a plan according to query_planner's
 862                          * results.
 863                          */
 864                         result_plan = create_plan(root, best_path);
 865                         current_pathkeys = best_path->pathkeys;
 866
 867                         /*
 868                          * create_plan() returns a plan with just a "flat" tlist of
 869                          * required Vars.  Usually we need to insert the sub_tlist as the
 870                          * tlist of the top plan node.  However, we can skip that if we
 871                          * determined that whatever query_planner chose to return will be
 872                          * good enough.
 873                          */
 874                         if (need_tlist_eval)
 875                         {
 876                                 /*
 877                                  * If the top-level plan node is one that cannot do expression
 878                                  * evaluation, we must insert a Result node to project the
 879                                  * desired tlist.
 880                                  */
 881                                 if (!is_projection_capable_plan(result_plan))
 882                                 {
 883                                         result_plan = (Plan *) make_result(sub_tlist, NULL,
 884                                                                                                            result_plan);
 885                                 }
 886                                 else
 887                                 {
 888                                         /*
 889                                          * Otherwise, just replace the subplan's flat tlist with
 890                                          * the desired tlist.
 891                                          */
 892                                         result_plan->targetlist = sub_tlist;
 893                                 }
 894
 895                                 /*
 896                                  * Also, account for the cost of evaluation of the sub_tlist.
 897                                  *
 898                                  * Up to now, we have only been dealing with "flat" tlists,
 899                                  * containing just Vars.  So their evaluation cost is zero
 900                                  * according to the model used by cost_qual_eval() (or if you
 901                                  * prefer, the cost is factored into cpu_tuple_cost).  Thus we
 902                                  * can avoid accounting for tlist cost throughout
 903                                  * query_planner() and subroutines.  But now we've inserted a
 904                                  * tlist that might contain actual operators, sub-selects, etc
 905                                  * --- so we'd better account for its cost.
 906                                  *
 907                                  * Below this point, any tlist eval cost for added-on nodes
 908                                  * should be accounted for as we create those nodes.
 909                                  * Presently, of the node types we can add on, only Agg and
 910                                  * Group project new tlists (the rest just copy their input
 911                                  * tuples) --- so make_agg() and make_group() are responsible
 912                                  * for computing the added cost.
 913                                  */
 914                                 cost_qual_eval(&tlist_cost, sub_tlist);
 915                                 result_plan->startup_cost += tlist_cost.startup;
 916                                 result_plan->total_cost += tlist_cost.startup +
 917                                         tlist_cost.per_tuple * result_plan->plan_rows;
 918                         }
 919                         else
 920                         {
 921                                 /*
 922                                  * Since we're using query_planner's tlist and not the one
 923                                  * make_subplanTargetList calculated, we have to refigure any
 924                                  * grouping-column indexes make_subplanTargetList computed.
 925                                  */
 926                                 locate_grouping_columns(root, tlist, result_plan->targetlist,
 927                                                                                 groupColIdx);
 928                         }
 929
 930                         /*
 931                          * Insert AGG or GROUP node if needed, plus an explicit sort step
 932                          * if necessary.
 933                          *
 934                          * HAVING clause, if any, becomes qual of the Agg or Group node.
 935                          */
 936                         if (use_hashed_grouping)
 937                         {
 938                                 /* Hashed aggregate plan --- no sort needed */
 939                                 result_plan = (Plan *) make_agg(root,
 940                                                                                                 tlist,
 941                                                                                                 (List *) parse->havingQual,
 942                                                                                                 AGG_HASHED,
 943                                                                                                 numGroupCols,
 944                                                                                                 groupColIdx,
 945                                                                                                 numGroups,
 946                                                                                                 agg_counts.numAggs,
 947                                                                                                 result_plan);
 948                                 /* Hashed aggregation produces randomly-ordered results */
 949                                 current_pathkeys = NIL;
 950                         }
 951                         else if (parse->hasAggs)
 952                         {
 953                                 /* Plain aggregate plan --- sort if needed */
 954                                 AggStrategy aggstrategy;
 955
 956                                 if (parse->groupClause)
 957                                 {
 958                                         if (!pathkeys_contained_in(group_pathkeys,
 959                                                                                            current_pathkeys))
 960                                         {
 961                                                 result_plan = (Plan *)
 962                                                         make_sort_from_groupcols(root,
 963                                                                                                          parse->groupClause,
 964                                                                                                          groupColIdx,
 965                                                                                                          result_plan);
 966                                                 current_pathkeys = group_pathkeys;
 967                                         }
 968                                         aggstrategy = AGG_SORTED;
 969
 970                                         /*
 971                                          * The AGG node will not change the sort ordering of its
 972                                          * groups, so current_pathkeys describes the result too.
 973                                          */
 974                                 }
 975                                 else
 976                                 {
 977                                         aggstrategy = AGG_PLAIN;
 978                                         /* Result will be only one row anyway; no sort order */
 979                                         current_pathkeys = NIL;
 980                                 }
 981
 982                                 result_plan = (Plan *) make_agg(root,
 983                                                                                                 tlist,
 984                                                                                                 (List *) parse->havingQual,
 985                                                                                                 aggstrategy,
 986                                                                                                 numGroupCols,
 987                                                                                                 groupColIdx,
 988                                                                                                 numGroups,
 989                                                                                                 agg_counts.numAggs,
 990                                                                                                 result_plan);
 991                         }
 992                         else if (parse->groupClause)
 993                         {
 994                                 /*
 995                                  * GROUP BY without aggregation, so insert a group node (plus
 996                                  * the appropriate sort node, if necessary).
 997                                  *
 998                                  * Add an explicit sort if we couldn't make the path come out
 999                                  * the way the GROUP node needs it.
1000                                  */
1001                                 if (!pathkeys_contained_in(group_pathkeys, current_pathkeys))
1002                                 {
1003                                         result_plan = (Plan *)
1004                                                 make_sort_from_groupcols(root,
1005                                                                                                  parse->groupClause,
1006                                                                                                  groupColIdx,
1007                                                                                                  result_plan);
1008                                         current_pathkeys = group_pathkeys;
1009                                 }
1010
1011                                 result_plan = (Plan *) make_group(root,
1012                                                                                                   tlist,
1013                                                                                                   (List *) parse->havingQual,
1014                                                                                                   numGroupCols,
1015                                                                                                   groupColIdx,
1016                                                                                                   dNumGroups,
1017                                                                                                   result_plan);
1018                                 /* The Group node won't change sort ordering */
1019                         }
1020                         else if (root->hasHavingQual)
1021                         {
1022                                 /*
1023                                  * No aggregates, and no GROUP BY, but we have a HAVING qual.
1024                                  * This is a degenerate case in which we are supposed to emit
1025                                  * either 0 or 1 row depending on whether HAVING succeeds.
1026                                  * Furthermore, there cannot be any variables in either HAVING
1027                                  * or the targetlist, so we actually do not need the FROM
1028                                  * table at all!  We can just throw away the plan-so-far and
1029                                  * generate a Result node.      This is a sufficiently unusual
1030                                  * corner case that it's not worth contorting the structure of
1031                                  * this routine to avoid having to generate the plan in the
1032                                  * first place.
1033                                  */
1034                                 result_plan = (Plan *) make_result(tlist,
1035                                                                                                    parse->havingQual,
1036                                                                                                    NULL);
1037                         }
1038                 }                                               /* end of non-minmax-aggregate case */
1039         }                                                       /* end of if (setOperations) */
1040
1041         /*
1042          * If we were not able to make the plan come out in the right order, add
1043          * an explicit sort step.
1044          */
1045         if (parse->sortClause)
1046         {
1047                 if (!pathkeys_contained_in(sort_pathkeys, current_pathkeys))
1048                 {
1049                         result_plan = (Plan *)
1050                                 make_sort_from_sortclauses(root,
1051                                                                                    parse->sortClause,
1052                                                                                    result_plan);
1053                         current_pathkeys = sort_pathkeys;
1054                 }
1055         }
1056
1057         /*
1058          * If there is a DISTINCT clause, add the UNIQUE node.
1059          */
1060         if (parse->distinctClause)
1061         {
1062                 result_plan = (Plan *) make_unique(result_plan, parse->distinctClause);
1063
1064                 /*
1065                  * If there was grouping or aggregation, leave plan_rows as-is (ie,
1066                  * assume the result was already mostly unique).  If not, use the
1067                  * number of distinct-groups calculated by query_planner.
1068                  */
1069                 if (!parse->groupClause && !root->hasHavingQual && !parse->hasAggs)
1070                         result_plan->plan_rows = dNumGroups;
1071         }
1072
1073         /*
1074          * Finally, if there is a LIMIT/OFFSET clause, add the LIMIT node.
1075          */
1076         if (parse->limitCount || parse->limitOffset)
1077         {
1078                 result_plan = (Plan *) make_limit(result_plan,
1079                                                                                   parse->limitOffset,
1080                                                                                   parse->limitCount,
1081                                                                                   offset_est,
1082                                                                                   count_est);
1083         }
1084
1085         /*
1086          * Return the actual output ordering in query_pathkeys for possible use by
1087          * an outer query level.
1088          */
1089         root->query_pathkeys = current_pathkeys;
1090
1091         return result_plan;
1092 }
1093
1094 /*
1095  * Detect whether a plan node is a "dummy" plan created when a relation
1096  * is deemed not to need scanning due to constraint exclusion.
1097  *
1098  * Currently, such dummy plans are Result nodes with constant FALSE
1099  * filter quals.
1100  */
1101 static bool
1102 is_dummy_plan(Plan *plan)
1103 {
1104         if (IsA(plan, Result))
1105         {
1106                 List *rcqual = (List *) ((Result *) plan)->resconstantqual;
1107
1108                 if (list_length(rcqual) == 1)
1109                 {
1110                         Const *constqual = (Const *) linitial(rcqual);
1111
1112                         if (constqual && IsA(constqual, Const))
1113                         {
1114                                 if (!constqual->constisnull &&
1115                                         !DatumGetBool(constqual->constvalue))
1116                                         return true;
1117                         }
1118                 }
1119         }
1120         return false;
1121 }
1122
1123 /*
1124  * preprocess_limit - do pre-estimation for LIMIT and/or OFFSET clauses
1125  *
1126  * We try to estimate the values of the LIMIT/OFFSET clauses, and pass the
1127  * results back in *count_est and *offset_est.  These variables are set to
1128  * 0 if the corresponding clause is not present, and -1 if it's present
1129  * but we couldn't estimate the value for it.  (The "0" convention is OK
1130  * for OFFSET but a little bit bogus for LIMIT: effectively we estimate
1131  * LIMIT 0 as though it were LIMIT 1.  But this is in line with the planner's
1132  * usual practice of never estimating less than one row.)  These values will
1133  * be passed to make_limit, which see if you change this code.
1134  *
1135  * The return value is the suitably adjusted tuple_fraction to use for
1136  * planning the query.  This adjustment is not overridable, since it reflects
1137  * plan actions that grouping_planner() will certainly take, not assumptions
1138  * about context.
1139  */
1140 static double
1141 preprocess_limit(PlannerInfo *root, double tuple_fraction,
1142                                  int64 *offset_est, int64 *count_est)
1143 {
1144         Query      *parse = root->parse;
1145         Node       *est;
1146         double          limit_fraction;
1147
1148         /* Should not be called unless LIMIT or OFFSET */
1149         Assert(parse->limitCount || parse->limitOffset);
1150
1151         /*
1152          * Try to obtain the clause values.  We use estimate_expression_value
1153          * primarily because it can sometimes do something useful with Params.
1154          */
1155         if (parse->limitCount)
1156         {
1157                 est = estimate_expression_value(parse->limitCount);
1158                 if (est && IsA(est, Const))
1159                 {
1160                         if (((Const *) est)->constisnull)
1161                         {
1162                                 /* NULL indicates LIMIT ALL, ie, no limit */
1163                                 *count_est = 0; /* treat as not present */
1164                         }
1165                         else
1166                         {
1167                                 *count_est = DatumGetInt64(((Const *) est)->constvalue);
1168                                 if (*count_est <= 0)
1169                                         *count_est = 1;         /* force to at least 1 */
1170                         }
1171                 }
1172                 else
1173                         *count_est = -1;        /* can't estimate */
1174         }
1175         else
1176                 *count_est = 0;                 /* not present */
1177
1178         if (parse->limitOffset)
1179         {
1180                 est = estimate_expression_value(parse->limitOffset);
1181                 if (est && IsA(est, Const))
1182                 {
1183                         if (((Const *) est)->constisnull)
1184                         {
1185                                 /* Treat NULL as no offset; the executor will too */
1186                                 *offset_est = 0;        /* treat as not present */
1187                         }
1188                         else
1189                         {
1190                                 *offset_est = DatumGetInt64(((Const *) est)->constvalue);
1191                                 if (*offset_est < 0)
1192                                         *offset_est = 0;        /* less than 0 is same as 0 */
1193                         }
1194                 }
1195                 else
1196                         *offset_est = -1;       /* can't estimate */
1197         }
1198         else
1199                 *offset_est = 0;                /* not present */
1200
1201         if (*count_est != 0)
1202         {
1203                 /*
1204                  * A LIMIT clause limits the absolute number of tuples returned.
1205                  * However, if it's not a constant LIMIT then we have to guess; for
1206                  * lack of a better idea, assume 10% of the plan's result is wanted.
1207                  */
1208                 if (*count_est < 0 || *offset_est < 0)
1209                 {
1210                         /* LIMIT or OFFSET is an expression ... punt ... */
1211                         limit_fraction = 0.10;
1212                 }
1213                 else
1214                 {
1215                         /* LIMIT (plus OFFSET, if any) is max number of tuples needed */
1216                         limit_fraction = (double) *count_est + (double) *offset_est;
1217                 }
1218
1219                 /*
1220                  * If we have absolute limits from both caller and LIMIT, use the
1221                  * smaller value; likewise if they are both fractional.  If one is
1222                  * fractional and the other absolute, we can't easily determine which
1223                  * is smaller, but we use the heuristic that the absolute will usually
1224                  * be smaller.
1225                  */
1226                 if (tuple_fraction >= 1.0)
1227                 {
1228                         if (limit_fraction >= 1.0)
1229                         {
1230                                 /* both absolute */
1231                                 tuple_fraction = Min(tuple_fraction, limit_fraction);
1232                         }
1233                         else
1234                         {
1235                                 /* caller absolute, limit fractional; use caller's value */
1236                         }
1237                 }
1238                 else if (tuple_fraction > 0.0)
1239                 {
1240                         if (limit_fraction >= 1.0)
1241                         {
1242                                 /* caller fractional, limit absolute; use limit */
1243                                 tuple_fraction = limit_fraction;
1244                         }
1245                         else
1246                         {
1247                                 /* both fractional */
1248                                 tuple_fraction = Min(tuple_fraction, limit_fraction);
1249                         }
1250                 }
1251                 else
1252                 {
1253                         /* no info from caller, just use limit */
1254                         tuple_fraction = limit_fraction;
1255                 }
1256         }
1257         else if (*offset_est != 0 && tuple_fraction > 0.0)
1258         {
1259                 /*
1260                  * We have an OFFSET but no LIMIT.      This acts entirely differently
1261                  * from the LIMIT case: here, we need to increase rather than decrease
1262                  * the caller's tuple_fraction, because the OFFSET acts to cause more
1263                  * tuples to be fetched instead of fewer.  This only matters if we got
1264                  * a tuple_fraction > 0, however.
1265                  *
1266                  * As above, use 10% if OFFSET is present but unestimatable.
1267                  */
1268                 if (*offset_est < 0)
1269                         limit_fraction = 0.10;
1270                 else
1271                         limit_fraction = (double) *offset_est;
1272
1273                 /*
1274                  * If we have absolute counts from both caller and OFFSET, add them
1275                  * together; likewise if they are both fractional.      If one is
1276                  * fractional and the other absolute, we want to take the larger, and
1277                  * we heuristically assume that's the fractional one.
1278                  */
1279                 if (tuple_fraction >= 1.0)
1280                 {
1281                         if (limit_fraction >= 1.0)
1282                         {
1283                                 /* both absolute, so add them together */
1284                                 tuple_fraction += limit_fraction;
1285                         }
1286                         else
1287                         {
1288                                 /* caller absolute, limit fractional; use limit */
1289                                 tuple_fraction = limit_fraction;
1290                         }
1291                 }
1292                 else
1293                 {
1294                         if (limit_fraction >= 1.0)
1295                         {
1296                                 /* caller fractional, limit absolute; use caller's value */
1297                         }
1298                         else
1299                         {
1300                                 /* both fractional, so add them together */
1301                                 tuple_fraction += limit_fraction;
1302                                 if (tuple_fraction >= 1.0)
1303                                         tuple_fraction = 0.0;           /* assume fetch all */
1304                         }
1305                 }
1306         }
1307
1308         return tuple_fraction;
1309 }
1310
1311 /*
1312  * choose_hashed_grouping - should we use hashed grouping?
1313  */
1314 static bool
1315 choose_hashed_grouping(PlannerInfo *root, double tuple_fraction,
1316                                            Path *cheapest_path, Path *sorted_path,
1317                                            double dNumGroups, AggClauseCounts *agg_counts)
1318 {
1319         int                     numGroupCols = list_length(root->parse->groupClause);
1320         double          cheapest_path_rows;
1321         int                     cheapest_path_width;
1322         Size            hashentrysize;
1323         List       *current_pathkeys;
1324         Path            hashed_p;
1325         Path            sorted_p;
1326
1327         /*
1328          * Check can't-do-it conditions, including whether the grouping operators
1329          * are hashjoinable.
1330          *
1331          * Executor doesn't support hashed aggregation with DISTINCT aggregates.
1332          * (Doing so would imply storing *all* the input values in the hash table,
1333          * which seems like a certain loser.)
1334          */
1335         if (!enable_hashagg)
1336                 return false;
1337         if (agg_counts->numDistinctAggs != 0)
1338                 return false;
1339         if (!hash_safe_grouping(root))
1340                 return false;
1341
1342         /*
1343          * Don't do it if it doesn't look like the hashtable will fit into
1344          * work_mem.
1345          *
1346          * Beware here of the possibility that cheapest_path->parent is NULL. This
1347          * could happen if user does something silly like SELECT 'foo' GROUP BY 1;
1348          */
1349         if (cheapest_path->parent)
1350         {
1351                 cheapest_path_rows = cheapest_path->parent->rows;
1352                 cheapest_path_width = cheapest_path->parent->width;
1353         }
1354         else
1355         {
1356                 cheapest_path_rows = 1; /* assume non-set result */
1357                 cheapest_path_width = 100;              /* arbitrary */
1358         }
1359
1360         /* Estimate per-hash-entry space at tuple width... */
1361         hashentrysize = MAXALIGN(cheapest_path_width) + MAXALIGN(sizeof(MinimalTupleData));
1362         /* plus space for pass-by-ref transition values... */
1363         hashentrysize += agg_counts->transitionSpace;
1364         /* plus the per-hash-entry overhead */
1365         hashentrysize += hash_agg_entry_size(agg_counts->numAggs);
1366
1367         if (hashentrysize * dNumGroups > work_mem * 1024L)
1368                 return false;
1369
1370         /*
1371          * See if the estimated cost is no more than doing it the other way. While
1372          * avoiding the need for sorted input is usually a win, the fact that the
1373          * output won't be sorted may be a loss; so we need to do an actual cost
1374          * comparison.
1375          *
1376          * We need to consider cheapest_path + hashagg [+ final sort] versus
1377          * either cheapest_path [+ sort] + group or agg [+ final sort] or
1378          * presorted_path + group or agg [+ final sort] where brackets indicate a
1379          * step that may not be needed. We assume query_planner() will have
1380          * returned a presorted path only if it's a winner compared to
1381          * cheapest_path for this purpose.
1382          *
1383          * These path variables are dummies that just hold cost fields; we don't
1384          * make actual Paths for these steps.
1385          */
1386         cost_agg(&hashed_p, root, AGG_HASHED, agg_counts->numAggs,
1387                          numGroupCols, dNumGroups,
1388                          cheapest_path->startup_cost, cheapest_path->total_cost,
1389                          cheapest_path_rows);
1390         /* Result of hashed agg is always unsorted */
1391         if (root->sort_pathkeys)
1392                 cost_sort(&hashed_p, root, root->sort_pathkeys, hashed_p.total_cost,
1393                                   dNumGroups, cheapest_path_width);
1394
1395         if (sorted_path)
1396         {
1397                 sorted_p.startup_cost = sorted_path->startup_cost;
1398                 sorted_p.total_cost = sorted_path->total_cost;
1399                 current_pathkeys = sorted_path->pathkeys;
1400         }
1401         else
1402         {
1403                 sorted_p.startup_cost = cheapest_path->startup_cost;
1404                 sorted_p.total_cost = cheapest_path->total_cost;
1405                 current_pathkeys = cheapest_path->pathkeys;
1406         }
1407         if (!pathkeys_contained_in(root->group_pathkeys, current_pathkeys))
1408         {
1409                 cost_sort(&sorted_p, root, root->group_pathkeys, sorted_p.total_cost,
1410                                   cheapest_path_rows, cheapest_path_width);
1411                 current_pathkeys = root->group_pathkeys;
1412         }
1413
1414         if (root->parse->hasAggs)
1415                 cost_agg(&sorted_p, root, AGG_SORTED, agg_counts->numAggs,
1416                                  numGroupCols, dNumGroups,
1417                                  sorted_p.startup_cost, sorted_p.total_cost,
1418                                  cheapest_path_rows);
1419         else
1420                 cost_group(&sorted_p, root, numGroupCols, dNumGroups,
1421                                    sorted_p.startup_cost, sorted_p.total_cost,
1422                                    cheapest_path_rows);
1423         /* The Agg or Group node will preserve ordering */
1424         if (root->sort_pathkeys &&
1425                 !pathkeys_contained_in(root->sort_pathkeys, current_pathkeys))
1426                 cost_sort(&sorted_p, root, root->sort_pathkeys, sorted_p.total_cost,
1427                                   dNumGroups, cheapest_path_width);
1428
1429         /*
1430          * Now make the decision using the top-level tuple fraction.  First we
1431          * have to convert an absolute count (LIMIT) into fractional form.
1432          */
1433         if (tuple_fraction >= 1.0)
1434                 tuple_fraction /= dNumGroups;
1435
1436         if (compare_fractional_path_costs(&hashed_p, &sorted_p,
1437                                                                           tuple_fraction) < 0)
1438         {
1439                 /* Hashed is cheaper, so use it */
1440                 return true;
1441         }
1442         return false;
1443 }
1444
1445 /*
1446  * hash_safe_grouping - are grouping operators hashable?
1447  *
1448  * We assume hashed aggregation will work if the datatype's equality operator
1449  * is marked hashjoinable.
1450  */
1451 static bool
1452 hash_safe_grouping(PlannerInfo *root)
1453 {
1454         ListCell   *gl;
1455
1456         foreach(gl, root->parse->groupClause)
1457         {
1458                 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1459                 TargetEntry *tle = get_sortgroupclause_tle(grpcl,
1460                                                                                                    root->parse->targetList);
1461                 Operator        optup;
1462                 bool            oprcanhash;
1463
1464                 optup = equality_oper(exprType((Node *) tle->expr), true);
1465                 if (!optup)
1466                         return false;
1467                 oprcanhash = ((Form_pg_operator) GETSTRUCT(optup))->oprcanhash;
1468                 ReleaseSysCache(optup);
1469                 if (!oprcanhash)
1470                         return false;
1471         }
1472         return true;
1473 }
1474
1475 /*---------------
1476  * make_subplanTargetList
1477  *        Generate appropriate target list when grouping is required.
1478  *
1479  * When grouping_planner inserts Aggregate, Group, or Result plan nodes
1480  * above the result of query_planner, we typically want to pass a different
1481  * target list to query_planner than the outer plan nodes should have.
1482  * This routine generates the correct target list for the subplan.
1483  *
1484  * The initial target list passed from the parser already contains entries
1485  * for all ORDER BY and GROUP BY expressions, but it will not have entries
1486  * for variables used only in HAVING clauses; so we need to add those
1487  * variables to the subplan target list.  Also, we flatten all expressions
1488  * except GROUP BY items into their component variables; the other expressions
1489  * will be computed by the inserted nodes rather than by the subplan.
1490  * For example, given a query like
1491  *              SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
1492  * we want to pass this targetlist to the subplan:
1493  *              a,b,c,d,a+b
1494  * where the a+b target will be used by the Sort/Group steps, and the
1495  * other targets will be used for computing the final results.  (In the
1496  * above example we could theoretically suppress the a and b targets and
1497  * pass down only c,d,a+b, but it's not really worth the trouble to
1498  * eliminate simple var references from the subplan.  We will avoid doing
1499  * the extra computation to recompute a+b at the outer level; see
1500  * replace_vars_with_subplan_refs() in setrefs.c.)
1501  *
1502  * If we are grouping or aggregating, *and* there are no non-Var grouping
1503  * expressions, then the returned tlist is effectively dummy; we do not
1504  * need to force it to be evaluated, because all the Vars it contains
1505  * should be present in the output of query_planner anyway.
1506  *
1507  * 'tlist' is the query's target list.
1508  * 'groupColIdx' receives an array of column numbers for the GROUP BY
1509  *                      expressions (if there are any) in the subplan's target list.
1510  * 'need_tlist_eval' is set true if we really need to evaluate the
1511  *                      result tlist.
1512  *
1513  * The result is the targetlist to be passed to the subplan.
1514  *---------------
1515  */
1516 static List *
1517 make_subplanTargetList(PlannerInfo *root,
1518                                            List *tlist,
1519                                            AttrNumber **groupColIdx,
1520                                            bool *need_tlist_eval)
1521 {
1522         Query      *parse = root->parse;
1523         List       *sub_tlist;
1524         List       *extravars;
1525         int                     numCols;
1526
1527         *groupColIdx = NULL;
1528
1529         /*
1530          * If we're not grouping or aggregating, there's nothing to do here;
1531          * query_planner should receive the unmodified target list.
1532          */
1533         if (!parse->hasAggs && !parse->groupClause && !root->hasHavingQual)
1534         {
1535                 *need_tlist_eval = true;
1536                 return tlist;
1537         }
1538
1539         /*
1540          * Otherwise, start with a "flattened" tlist (having just the vars
1541          * mentioned in the targetlist and HAVING qual --- but not upper- level
1542          * Vars; they will be replaced by Params later on).
1543          */
1544         sub_tlist = flatten_tlist(tlist);
1545         extravars = pull_var_clause(parse->havingQual, false);
1546         sub_tlist = add_to_flat_tlist(sub_tlist, extravars);
1547         list_free(extravars);
1548         *need_tlist_eval = false;       /* only eval if not flat tlist */
1549
1550         /*
1551          * If grouping, create sub_tlist entries for all GROUP BY expressions
1552          * (GROUP BY items that are simple Vars should be in the list already),
1553          * and make an array showing where the group columns are in the sub_tlist.
1554          */
1555         numCols = list_length(parse->groupClause);
1556         if (numCols > 0)
1557         {
1558                 int                     keyno = 0;
1559                 AttrNumber *grpColIdx;
1560                 ListCell   *gl;
1561
1562                 grpColIdx = (AttrNumber *) palloc(sizeof(AttrNumber) * numCols);
1563                 *groupColIdx = grpColIdx;
1564
1565                 foreach(gl, parse->groupClause)
1566                 {
1567                         GroupClause *grpcl = (GroupClause *) lfirst(gl);
1568                         Node       *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1569                         TargetEntry *te = NULL;
1570                         ListCell   *sl;
1571
1572                         /* Find or make a matching sub_tlist entry */
1573                         foreach(sl, sub_tlist)
1574                         {
1575                                 te = (TargetEntry *) lfirst(sl);
1576                                 if (equal(groupexpr, te->expr))
1577                                         break;
1578                         }
1579                         if (!sl)
1580                         {
1581                                 te = makeTargetEntry((Expr *) groupexpr,
1582                                                                          list_length(sub_tlist) + 1,
1583                                                                          NULL,
1584                                                                          false);
1585                                 sub_tlist = lappend(sub_tlist, te);
1586                                 *need_tlist_eval = true;                /* it's not flat anymore */
1587                         }
1588
1589                         /* and save its resno */
1590                         grpColIdx[keyno++] = te->resno;
1591                 }
1592         }
1593
1594         return sub_tlist;
1595 }
1596
1597 /*
1598  * locate_grouping_columns
1599  *              Locate grouping columns in the tlist chosen by query_planner.
1600  *
1601  * This is only needed if we don't use the sub_tlist chosen by
1602  * make_subplanTargetList.      We have to forget the column indexes found
1603  * by that routine and re-locate the grouping vars in the real sub_tlist.
1604  */
1605 static void
1606 locate_grouping_columns(PlannerInfo *root,
1607                                                 List *tlist,
1608                                                 List *sub_tlist,
1609                                                 AttrNumber *groupColIdx)
1610 {
1611         int                     keyno = 0;
1612         ListCell   *gl;
1613
1614         /*
1615          * No work unless grouping.
1616          */
1617         if (!root->parse->groupClause)
1618         {
1619                 Assert(groupColIdx == NULL);
1620                 return;
1621         }
1622         Assert(groupColIdx != NULL);
1623
1624         foreach(gl, root->parse->groupClause)
1625         {
1626                 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1627                 Node       *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1628                 TargetEntry *te = NULL;
1629                 ListCell   *sl;
1630
1631                 foreach(sl, sub_tlist)
1632                 {
1633                         te = (TargetEntry *) lfirst(sl);
1634                         if (equal(groupexpr, te->expr))
1635                                 break;
1636                 }
1637                 if (!sl)
1638                         elog(ERROR, "failed to locate grouping columns");
1639
1640                 groupColIdx[keyno++] = te->resno;
1641         }
1642 }
1643
1644 /*
1645  * postprocess_setop_tlist
1646  *        Fix up targetlist returned by plan_set_operations().
1647  *
1648  * We need to transpose sort key info from the orig_tlist into new_tlist.
1649  * NOTE: this would not be good enough if we supported resjunk sort keys
1650  * for results of set operations --- then, we'd need to project a whole
1651  * new tlist to evaluate the resjunk columns.  For now, just ereport if we
1652  * find any resjunk columns in orig_tlist.
1653  */
1654 static List *
1655 postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
1656 {
1657         ListCell   *l;
1658         ListCell   *orig_tlist_item = list_head(orig_tlist);
1659
1660         foreach(l, new_tlist)
1661         {
1662                 TargetEntry *new_tle = (TargetEntry *) lfirst(l);
1663                 TargetEntry *orig_tle;
1664
1665                 /* ignore resjunk columns in setop result */
1666                 if (new_tle->resjunk)
1667                         continue;
1668
1669                 Assert(orig_tlist_item != NULL);
1670                 orig_tle = (TargetEntry *) lfirst(orig_tlist_item);
1671                 orig_tlist_item = lnext(orig_tlist_item);
1672                 if (orig_tle->resjunk)  /* should not happen */
1673                         elog(ERROR, "resjunk output columns are not implemented");
1674                 Assert(new_tle->resno == orig_tle->resno);
1675                 new_tle->ressortgroupref = orig_tle->ressortgroupref;
1676         }
1677         if (orig_tlist_item != NULL)
1678                 elog(ERROR, "resjunk output columns are not implemented");
1679         return new_tlist;
1680 }