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 1.217 2007/04/16 01:14:56 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         /*
  94          * Set up global state for this planner invocation.  This data is needed
  95          * across all levels of sub-Query that might exist in the given command,
  96          * so we keep it in a separate struct that's linked to by each per-Query
  97          * PlannerInfo.
  98          */
  99         glob = makeNode(PlannerGlobal);
 100
 101         glob->boundParams = boundParams;
 102         glob->paramlist = NIL;
 103         glob->subplans = NIL;
 104         glob->subrtables = NIL;
 105         glob->rewindPlanIDs = NULL;
 106         glob->finalrtable = NIL;
 107
 108         /* Determine what fraction of the plan is likely to be scanned */
 109         if (cursorOptions & CURSOR_OPT_FAST_PLAN)
 110         {
 111                 /*
 112                  * We have no real idea how many tuples the user will ultimately FETCH
 113                  * from a cursor, but it seems a good bet that he doesn't want 'em
 114                  * all.  Optimize for 10% retrieval (you gotta better number?  Should
 115                  * this be a SETtable parameter?)
 116                  */
 117                 tuple_fraction = 0.10;
 118         }
 119         else
 120         {
 121                 /* Default assumption is we need all the tuples */
 122                 tuple_fraction = 0.0;
 123         }
 124
 125         /* primary planning entry point (may recurse for subqueries) */
 126         top_plan = subquery_planner(glob, parse, 1, tuple_fraction, &root);
 127
 128         /*
 129          * If creating a plan for a scrollable cursor, make sure it can run
 130          * backwards on demand.  Add a Material node at the top at need.
 131          */
 132         if (cursorOptions & CURSOR_OPT_SCROLL)
 133         {
 134                 if (!ExecSupportsBackwardScan(top_plan))
 135                         top_plan = materialize_finished_plan(top_plan);
 136         }
 137
 138         /* final cleanup of the plan */
 139         Assert(glob->finalrtable == NIL);
 140         top_plan = set_plan_references(glob, top_plan, root->parse->rtable);
 141         /* ... and the subplans (both regular subplans and initplans) */
 142         Assert(list_length(glob->subplans) == list_length(glob->subrtables));
 143         forboth(lp, glob->subplans, lr, glob->subrtables)
 144         {
 145                 Plan   *subplan = (Plan *) lfirst(lp);
 146                 List   *subrtable = (List *) lfirst(lr);
 147
 148                 lfirst(lp) = set_plan_references(glob, subplan, subrtable);
 149         }
 150
 151         /* build the PlannedStmt result */
 152         result = makeNode(PlannedStmt);
 153
 154         result->commandType = parse->commandType;
 155         result->canSetTag = parse->canSetTag;
 156         result->planTree = top_plan;
 157         result->rtable = glob->finalrtable;
 158         result->resultRelations = root->resultRelations;
 159         result->into = parse->into;
 160         result->subplans = glob->subplans;
 161         result->rewindPlanIDs = glob->rewindPlanIDs;
 162         result->returningLists = root->returningLists;
 163         result->rowMarks = parse->rowMarks;
 164         result->nParamExec = list_length(glob->paramlist);
 165
 166         return result;
 167 }
 168
 169
 170 /*--------------------
 171  * subquery_planner
 172  *        Invokes the planner on a subquery.  We recurse to here for each
 173  *        sub-SELECT found in the query tree.
 174  *
 175  * glob is the global state for the current planner run.
 176  * parse is the querytree produced by the parser & rewriter.
 177  * level is the current recursion depth (1 at the top-level Query).
 178  * tuple_fraction is the fraction of tuples we expect will be retrieved.
 179  * tuple_fraction is interpreted as explained for grouping_planner, below.
 180  *
 181  * If subroot isn't NULL, we pass back the query's final PlannerInfo struct;
 182  * among other things this tells the output sort ordering of the plan.
 183  *
 184  * Basically, this routine does the stuff that should only be done once
 185  * per Query object.  It then calls grouping_planner.  At one time,
 186  * grouping_planner could be invoked recursively on the same Query object;
 187  * that's not currently true, but we keep the separation between the two
 188  * routines anyway, in case we need it again someday.
 189  *
 190  * subquery_planner will be called recursively to handle sub-Query nodes
 191  * found within the query's expressions and rangetable.
 192  *
 193  * Returns a query plan.
 194  *--------------------
 195  */
 196 Plan *
 197 subquery_planner(PlannerGlobal *glob, Query *parse,
 198                                  Index level, double tuple_fraction,
 199                                  PlannerInfo **subroot)
 200 {
 201         int                     num_old_subplans = list_length(glob->subplans);
 202         PlannerInfo *root;
 203         Plan       *plan;
 204         List       *newHaving;
 205         ListCell   *l;
 206
 207         /* Create a PlannerInfo data structure for this subquery */
 208         root = makeNode(PlannerInfo);
 209         root->parse = parse;
 210         root->glob = glob;
 211         root->query_level = level;
 212         root->planner_cxt = CurrentMemoryContext;
 213         root->init_plans = NIL;
 214         root->eq_classes = NIL;
 215         root->in_info_list = NIL;
 216         root->append_rel_list = NIL;
 217
 218         /*
 219          * Look for IN clauses at the top level of WHERE, and transform them into
 220          * joins.  Note that this step only handles IN clauses originally at top
 221          * level of WHERE; if we pull up any subqueries in the next step, their
 222          * INs are processed just before pulling them up.
 223          */
 224         if (parse->hasSubLinks)
 225                 parse->jointree->quals = pull_up_IN_clauses(root,
 226                                                                                                         parse->jointree->quals);
 227
 228         /*
 229          * Check to see if any subqueries in the rangetable can be merged into
 230          * this query.
 231          */
 232         parse->jointree = (FromExpr *)
 233                 pull_up_subqueries(root, (Node *) parse->jointree, false, false);
 234
 235         /*
 236          * Detect whether any rangetable entries are RTE_JOIN kind; if not, we can
 237          * avoid the expense of doing flatten_join_alias_vars().  Also check for
 238          * outer joins --- if none, we can skip reduce_outer_joins() and some
 239          * other processing.  This must be done after we have done
 240          * pull_up_subqueries, of course.
 241          *
 242          * Note: if reduce_outer_joins manages to eliminate all outer joins,
 243          * root->hasOuterJoins is not reset currently.  This is OK since its
 244          * purpose is merely to suppress unnecessary processing in simple cases.
 245          */
 246         root->hasJoinRTEs = false;
 247         root->hasOuterJoins = false;
 248         foreach(l, parse->rtable)
 249         {
 250                 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
 251
 252                 if (rte->rtekind == RTE_JOIN)
 253                 {
 254                         root->hasJoinRTEs = true;
 255                         if (IS_OUTER_JOIN(rte->jointype))
 256                         {
 257                                 root->hasOuterJoins = true;
 258                                 /* Can quit scanning once we find an outer join */
 259                                 break;
 260                         }
 261                 }
 262         }
 263
 264         /*
 265          * Expand any rangetable entries that are inheritance sets into "append
 266          * relations".  This can add entries to the rangetable, but they must be
 267          * plain base relations not joins, so it's OK (and marginally more
 268          * efficient) to do it after checking for join RTEs.  We must do it after
 269          * pulling up subqueries, else we'd fail to handle inherited tables in
 270          * subqueries.
 271          */
 272         expand_inherited_tables(root);
 273
 274         /*
 275          * Set hasHavingQual to remember if HAVING clause is present.  Needed
 276          * because preprocess_expression will reduce a constant-true condition to
 277          * an empty qual list ... but "HAVING TRUE" is not a semantic no-op.
 278          */
 279         root->hasHavingQual = (parse->havingQual != NULL);
 280
 281         /* Clear this flag; might get set in distribute_qual_to_rels */
 282         root->hasPseudoConstantQuals = false;
 283
 284         /*
 285          * Do expression preprocessing on targetlist and quals.
 286          */
 287         parse->targetList = (List *)
 288                 preprocess_expression(root, (Node *) parse->targetList,
 289                                                           EXPRKIND_TARGET);
 290
 291         parse->returningList = (List *)
 292                 preprocess_expression(root, (Node *) parse->returningList,
 293                                                           EXPRKIND_TARGET);
 294
 295         preprocess_qual_conditions(root, (Node *) parse->jointree);
 296
 297         parse->havingQual = preprocess_expression(root, parse->havingQual,
 298                                                                                           EXPRKIND_QUAL);
 299
 300         parse->limitOffset = preprocess_expression(root, parse->limitOffset,
 301                                                                                            EXPRKIND_LIMIT);
 302         parse->limitCount = preprocess_expression(root, parse->limitCount,
 303                                                                                           EXPRKIND_LIMIT);
 304
 305         root->in_info_list = (List *)
 306                 preprocess_expression(root, (Node *) root->in_info_list,
 307                                                           EXPRKIND_ININFO);
 308         root->append_rel_list = (List *)
 309                 preprocess_expression(root, (Node *) root->append_rel_list,
 310                                                           EXPRKIND_APPINFO);
 311
 312         /* Also need to preprocess expressions for function and values RTEs */
 313         foreach(l, parse->rtable)
 314         {
 315                 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
 316
 317                 if (rte->rtekind == RTE_FUNCTION)
 318                         rte->funcexpr = preprocess_expression(root, rte->funcexpr,
 319                                                                                                   EXPRKIND_RTFUNC);
 320                 else if (rte->rtekind == RTE_VALUES)
 321                         rte->values_lists = (List *)
 322                                 preprocess_expression(root, (Node *) rte->values_lists,
 323                                                                           EXPRKIND_VALUES);
 324         }
 325
 326         /*
 327          * In some cases we may want to transfer a HAVING clause into WHERE. We
 328          * cannot do so if the HAVING clause contains aggregates (obviously) or
 329          * volatile functions (since a HAVING clause is supposed to be executed
 330          * only once per group).  Also, it may be that the clause is so expensive
 331          * to execute that we're better off doing it only once per group, despite
 332          * the loss of selectivity.  This is hard to estimate short of doing the
 333          * entire planning process twice, so we use a heuristic: clauses
 334          * containing subplans are left in HAVING.      Otherwise, we move or copy the
 335          * HAVING clause into WHERE, in hopes of eliminating tuples before
 336          * aggregation instead of after.
 337          *
 338          * If the query has explicit grouping then we can simply move such a
 339          * clause into WHERE; any group that fails the clause will not be in the
 340          * output because none of its tuples will reach the grouping or
 341          * aggregation stage.  Otherwise we must have a degenerate (variable-free)
 342          * HAVING clause, which we put in WHERE so that query_planner() can use it
 343          * in a gating Result node, but also keep in HAVING to ensure that we
 344          * don't emit a bogus aggregated row. (This could be done better, but it
 345          * seems not worth optimizing.)
 346          *
 347          * Note that both havingQual and parse->jointree->quals are in
 348          * implicitly-ANDed-list form at this point, even though they are declared
 349          * as Node *.
 350          */
 351         newHaving = NIL;
 352         foreach(l, (List *) parse->havingQual)
 353         {
 354                 Node       *havingclause = (Node *) lfirst(l);
 355
 356                 if (contain_agg_clause(havingclause) ||
 357                         contain_volatile_functions(havingclause) ||
 358                         contain_subplans(havingclause))
 359                 {
 360                         /* keep it in HAVING */
 361                         newHaving = lappend(newHaving, havingclause);
 362                 }
 363                 else if (parse->groupClause)
 364                 {
 365                         /* move it to WHERE */
 366                         parse->jointree->quals = (Node *)
 367                                 lappend((List *) parse->jointree->quals, havingclause);
 368                 }
 369                 else
 370                 {
 371                         /* put a copy in WHERE, keep it in HAVING */
 372                         parse->jointree->quals = (Node *)
 373                                 lappend((List *) parse->jointree->quals,
 374                                                 copyObject(havingclause));
 375                         newHaving = lappend(newHaving, havingclause);
 376                 }
 377         }
 378         parse->havingQual = (Node *) newHaving;
 379
 380         /*
 381          * If we have any outer joins, try to reduce them to plain inner joins.
 382          * This step is most easily done after we've done expression
 383          * preprocessing.
 384          */
 385         if (root->hasOuterJoins)
 386                 reduce_outer_joins(root);
 387
 388         /*
 389          * Do the main planning.  If we have an inherited target relation, that
 390          * needs special processing, else go straight to grouping_planner.
 391          */
 392         if (parse->resultRelation &&
 393                 rt_fetch(parse->resultRelation, parse->rtable)->inh)
 394                 plan = inheritance_planner(root);
 395         else
 396                 plan = grouping_planner(root, tuple_fraction);
 397
 398         /*
 399          * If any subplans were generated, or if we're inside a subplan, build
 400          * initPlan list and extParam/allParam sets for plan nodes, and attach the
 401          * initPlans to the top plan node.
 402          */
 403         if (list_length(glob->subplans) != num_old_subplans ||
 404                 root->query_level > 1)
 405                 SS_finalize_plan(root, plan);
 406
 407         /* Return internal info if caller wants it */
 408         if (subroot)
 409                 *subroot = root;
 410
 411         return plan;
 412 }
 413
 414 /*
 415  * preprocess_expression
 416  *              Do subquery_planner's preprocessing work for an expression,
 417  *              which can be a targetlist, a WHERE clause (including JOIN/ON
 418  *              conditions), or a HAVING clause.
 419  */
 420 static Node *
 421 preprocess_expression(PlannerInfo *root, Node *expr, int kind)
 422 {
 423         /*
 424          * Fall out quickly if expression is empty.  This occurs often enough to
 425          * be worth checking.  Note that null->null is the correct conversion for
 426          * implicit-AND result format, too.
 427          */
 428         if (expr == NULL)
 429                 return NULL;
 430
 431         /*
 432          * If the query has any join RTEs, replace join alias variables with
 433          * base-relation variables. We must do this before sublink processing,
 434          * else sublinks expanded out from join aliases wouldn't get processed. We
 435          * can skip it in VALUES lists, however, since they can't contain any Vars
 436          * at all.
 437          */
 438         if (root->hasJoinRTEs && kind != EXPRKIND_VALUES)
 439                 expr = flatten_join_alias_vars(root, expr);
 440
 441         /*
 442          * Simplify constant expressions.
 443          *
 444          * Note: this also flattens nested AND and OR expressions into N-argument
 445          * form.  All processing of a qual expression after this point must be
 446          * careful to maintain AND/OR flatness --- that is, do not generate a tree
 447          * with AND directly under AND, nor OR directly under OR.
 448          *
 449          * Because this is a relatively expensive process, we skip it when the
 450          * query is trivial, such as "SELECT 2+2;" or "INSERT ... VALUES()". The
 451          * expression will only be evaluated once anyway, so no point in
 452          * pre-simplifying; we can't execute it any faster than the executor can,
 453          * and we will waste cycles copying the tree.  Notice however that we
 454          * still must do it for quals (to get AND/OR flatness); and if we are in a
 455          * subquery we should not assume it will be done only once.
 456          *
 457          * For VALUES lists we never do this at all, again on the grounds that we
 458          * should optimize for one-time evaluation.
 459          */
 460         if (kind != EXPRKIND_VALUES &&
 461                 (root->parse->jointree->fromlist != NIL ||
 462                  kind == EXPRKIND_QUAL ||
 463                  root->query_level > 1))
 464                 expr = eval_const_expressions(expr);
 465
 466         /*
 467          * If it's a qual or havingQual, canonicalize it.
 468          */
 469         if (kind == EXPRKIND_QUAL)
 470         {
 471                 expr = (Node *) canonicalize_qual((Expr *) expr);
 472
 473 #ifdef OPTIMIZER_DEBUG
 474                 printf("After canonicalize_qual()\n");
 475                 pprint(expr);
 476 #endif
 477         }
 478
 479         /* Expand SubLinks to SubPlans */
 480         if (root->parse->hasSubLinks)
 481                 expr = SS_process_sublinks(root, expr, (kind == EXPRKIND_QUAL));
 482
 483         /*
 484          * XXX do not insert anything here unless you have grokked the comments in
 485          * SS_replace_correlation_vars ...
 486          */
 487
 488         /* Replace uplevel vars with Param nodes (this IS possible in VALUES) */
 489         if (root->query_level > 1)
 490                 expr = SS_replace_correlation_vars(root, expr);
 491
 492         /*
 493          * If it's a qual or havingQual, convert it to implicit-AND format. (We
 494          * don't want to do this before eval_const_expressions, since the latter
 495          * would be unable to simplify a top-level AND correctly. Also,
 496          * SS_process_sublinks expects explicit-AND format.)
 497          */
 498         if (kind == EXPRKIND_QUAL)
 499                 expr = (Node *) make_ands_implicit((Expr *) expr);
 500
 501         return expr;
 502 }
 503
 504 /*
 505  * preprocess_qual_conditions
 506  *              Recursively scan the query's jointree and do subquery_planner's
 507  *              preprocessing work on each qual condition found therein.
 508  */
 509 static void
 510 preprocess_qual_conditions(PlannerInfo *root, Node *jtnode)
 511 {
 512         if (jtnode == NULL)
 513                 return;
 514         if (IsA(jtnode, RangeTblRef))
 515         {
 516                 /* nothing to do here */
 517         }
 518         else if (IsA(jtnode, FromExpr))
 519         {
 520                 FromExpr   *f = (FromExpr *) jtnode;
 521                 ListCell   *l;
 522
 523                 foreach(l, f->fromlist)
 524                         preprocess_qual_conditions(root, lfirst(l));
 525
 526                 f->quals = preprocess_expression(root, f->quals, EXPRKIND_QUAL);
 527         }
 528         else if (IsA(jtnode, JoinExpr))
 529         {
 530                 JoinExpr   *j = (JoinExpr *) jtnode;
 531
 532                 preprocess_qual_conditions(root, j->larg);
 533                 preprocess_qual_conditions(root, j->rarg);
 534
 535                 j->quals = preprocess_expression(root, j->quals, EXPRKIND_QUAL);
 536         }
 537         else
 538                 elog(ERROR, "unrecognized node type: %d",
 539                          (int) nodeTag(jtnode));
 540 }
 541
 542 /*
 543  * inheritance_planner
 544  *        Generate a plan in the case where the result relation is an
 545  *        inheritance set.
 546  *
 547  * We have to handle this case differently from cases where a source relation
 548  * is an inheritance set. Source inheritance is expanded at the bottom of the
 549  * plan tree (see allpaths.c), but target inheritance has to be expanded at
 550  * the top.  The reason is that for UPDATE, each target relation needs a
 551  * different targetlist matching its own column set.  Also, for both UPDATE
 552  * and DELETE, the executor needs the Append plan node at the top, else it
 553  * can't keep track of which table is the current target table.  Fortunately,
 554  * the UPDATE/DELETE target can never be the nullable side of an outer join,
 555  * so it's OK to generate the plan this way.
 556  *
 557  * Returns a query plan.
 558  */
 559 static Plan *
 560 inheritance_planner(PlannerInfo *root)
 561 {
 562         Query      *parse = root->parse;
 563         int                     parentRTindex = parse->resultRelation;
 564         List       *subplans = NIL;
 565         List       *resultRelations = NIL;
 566         List       *returningLists = NIL;
 567         List       *rtable = NIL;
 568         List       *tlist = NIL;
 569         PlannerInfo subroot;
 570         ListCell   *l;
 571
 572         foreach(l, root->append_rel_list)
 573         {
 574                 AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
 575                 Plan       *subplan;
 576
 577                 /* append_rel_list contains all append rels; ignore others */
 578                 if (appinfo->parent_relid != parentRTindex)
 579                         continue;
 580
 581                 /*
 582                  * Generate modified query with this rel as target.  We have to be
 583                  * prepared to translate varnos in in_info_list as well as in the
 584                  * Query proper.
 585                  */
 586                 memcpy(&subroot, root, sizeof(PlannerInfo));
 587                 subroot.parse = (Query *)
 588                         adjust_appendrel_attrs((Node *) parse,
 589                                                                    appinfo);
 590                 subroot.in_info_list = (List *)
 591                         adjust_appendrel_attrs((Node *) root->in_info_list,
 592                                                                    appinfo);
 593                 subroot.init_plans = NIL;
 594                 /* There shouldn't be any OJ info to translate, as yet */
 595                 Assert(subroot.oj_info_list == NIL);
 596
 597                 /* Generate plan */
 598                 subplan = grouping_planner(&subroot, 0.0 /* retrieve all tuples */ );
 599
 600                 /*
 601                  * If this child rel was excluded by constraint exclusion, exclude it
 602                  * from the plan.
 603                  */
 604                 if (is_dummy_plan(subplan))
 605                         continue;
 606
 607                 /* Save rtable and tlist from first rel for use below */
 608                 if (subplans == NIL)
 609                 {
 610                         rtable = subroot.parse->rtable;
 611                         tlist = subplan->targetlist;
 612                 }
 613
 614                 subplans = lappend(subplans, subplan);
 615
 616                 /* Make sure any initplans from this rel get into the outer list */
 617                 root->init_plans = list_concat(root->init_plans, subroot.init_plans);
 618
 619                 /* Build target-relations list for the executor */
 620                 resultRelations = lappend_int(resultRelations, appinfo->child_relid);
 621
 622                 /* Build list of per-relation RETURNING targetlists */
 623                 if (parse->returningList)
 624                 {
 625                         Assert(list_length(subroot.returningLists) == 1);
 626                         returningLists = list_concat(returningLists,
 627                                                                                  subroot.returningLists);
 628                 }
 629         }
 630
 631         root->resultRelations = resultRelations;
 632         root->returningLists = returningLists;
 633
 634         /* Mark result as unordered (probably unnecessary) */
 635         root->query_pathkeys = NIL;
 636
 637         /*
 638          * If we managed to exclude every child rel, return a dummy plan
 639          */
 640         if (subplans == NIL)
 641                 return (Plan *) make_result(root,
 642                                                                         tlist,
 643                                                                         (Node *) list_make1(makeBoolConst(false,
 644                                                                                                                                           false)),
 645                                                                         NULL);
 646
 647         /*
 648          * Planning might have modified the rangetable, due to changes of the
 649          * Query structures inside subquery RTEs.  We have to ensure that this
 650          * gets propagated back to the master copy.  But can't do this until we
 651          * are done planning, because all the calls to grouping_planner need
 652          * virgin sub-Queries to work from.  (We are effectively assuming that
 653          * sub-Queries will get planned identically each time, or at least that
 654          * the impacts on their rangetables will be the same each time.)
 655          *
 656          * XXX should clean this up someday
 657          */
 658         parse->rtable = rtable;
 659
 660         /* Suppress Append if there's only one surviving child rel */
 661         if (list_length(subplans) == 1)
 662                 return (Plan *) linitial(subplans);
 663
 664         return (Plan *) make_append(subplans, true, tlist);
 665 }
 666
 667 /*--------------------
 668  * grouping_planner
 669  *        Perform planning steps related to grouping, aggregation, etc.
 670  *        This primarily means adding top-level processing to the basic
 671  *        query plan produced by query_planner.
 672  *
 673  * tuple_fraction is the fraction of tuples we expect will be retrieved
 674  *
 675  * tuple_fraction is interpreted as follows:
 676  *        0: expect all tuples to be retrieved (normal case)
 677  *        0 < tuple_fraction < 1: expect the given fraction of tuples available
 678  *              from the plan to be retrieved
 679  *        tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
 680  *              expected to be retrieved (ie, a LIMIT specification)
 681  *
 682  * Returns a query plan.  Also, root->query_pathkeys is returned as the
 683  * actual output ordering of the plan (in pathkey format).
 684  *--------------------
 685  */
 686 static Plan *
 687 grouping_planner(PlannerInfo *root, double tuple_fraction)
 688 {
 689         Query      *parse = root->parse;
 690         List       *tlist = parse->targetList;
 691         int64           offset_est = 0;
 692         int64           count_est = 0;
 693         Plan       *result_plan;
 694         List       *current_pathkeys;
 695         List       *sort_pathkeys;
 696         double          dNumGroups = 0;
 697
 698         /* Tweak caller-supplied tuple_fraction if have LIMIT/OFFSET */
 699         if (parse->limitCount || parse->limitOffset)
 700                 tuple_fraction = preprocess_limit(root, tuple_fraction,
 701                                                                                   &offset_est, &count_est);
 702
 703         if (parse->setOperations)
 704         {
 705                 List       *set_sortclauses;
 706
 707                 /*
 708                  * If there's a top-level ORDER BY, assume we have to fetch all the
 709                  * tuples.      This might seem too simplistic given all the hackery below
 710                  * to possibly avoid the sort ... but a nonzero tuple_fraction is only
 711                  * of use to plan_set_operations() when the setop is UNION ALL, and
 712                  * the result of UNION ALL is always unsorted.
 713                  */
 714                 if (parse->sortClause)
 715                         tuple_fraction = 0.0;
 716
 717                 /*
 718                  * Construct the plan for set operations.  The result will not need
 719                  * any work except perhaps a top-level sort and/or LIMIT.
 720                  */
 721                 result_plan = plan_set_operations(root, tuple_fraction,
 722                                                                                   &set_sortclauses);
 723
 724                 /*
 725                  * Calculate pathkeys representing the sort order (if any) of the set
 726                  * operation's result.  We have to do this before overwriting the sort
 727                  * key information...
 728                  */
 729                 current_pathkeys = make_pathkeys_for_sortclauses(root,
 730                                                                                                                  set_sortclauses,
 731                                                                                                         result_plan->targetlist,
 732                                                                                                                  true);
 733
 734                 /*
 735                  * We should not need to call preprocess_targetlist, since we must be
 736                  * in a SELECT query node.      Instead, use the targetlist returned by
 737                  * plan_set_operations (since this tells whether it returned any
 738                  * resjunk columns!), and transfer any sort key information from the
 739                  * original tlist.
 740                  */
 741                 Assert(parse->commandType == CMD_SELECT);
 742
 743                 tlist = postprocess_setop_tlist(result_plan->targetlist, tlist);
 744
 745                 /*
 746                  * Can't handle FOR UPDATE/SHARE here (parser should have checked
 747                  * already, but let's make sure).
 748                  */
 749                 if (parse->rowMarks)
 750                         ereport(ERROR,
 751                                         (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
 752                                          errmsg("SELECT FOR UPDATE/SHARE is not allowed with UNION/INTERSECT/EXCEPT")));
 753
 754                 /*
 755                  * Calculate pathkeys that represent result ordering requirements
 756                  */
 757                 sort_pathkeys = make_pathkeys_for_sortclauses(root,
 758                                                                                                           parse->sortClause,
 759                                                                                                           tlist,
 760                                                                                                           true);
 761         }
 762         else
 763         {
 764                 /* No set operations, do regular planning */
 765                 List       *sub_tlist;
 766                 List       *group_pathkeys;
 767                 AttrNumber *groupColIdx = NULL;
 768                 Oid                *groupOperators = NULL;
 769                 bool            need_tlist_eval = true;
 770                 QualCost        tlist_cost;
 771                 Path       *cheapest_path;
 772                 Path       *sorted_path;
 773                 Path       *best_path;
 774                 long            numGroups = 0;
 775                 AggClauseCounts agg_counts;
 776                 int                     numGroupCols = list_length(parse->groupClause);
 777                 bool            use_hashed_grouping = false;
 778
 779                 MemSet(&agg_counts, 0, sizeof(AggClauseCounts));
 780
 781                 /* Preprocess targetlist */
 782                 tlist = preprocess_targetlist(root, tlist);
 783
 784                 /*
 785                  * Generate appropriate target list for subplan; may be different from
 786                  * tlist if grouping or aggregation is needed.
 787                  */
 788                 sub_tlist = make_subplanTargetList(root, tlist,
 789                                                                                    &groupColIdx, &need_tlist_eval);
 790
 791                 /*
 792                  * Calculate pathkeys that represent grouping/ordering requirements.
 793                  * Stash them in PlannerInfo so that query_planner can canonicalize
 794                  * them after EquivalenceClasses have been formed.
 795                  */
 796                 root->group_pathkeys =
 797                         make_pathkeys_for_sortclauses(root,
 798                                                                                   parse->groupClause,
 799                                                                                   tlist,
 800                                                                                   false);
 801                 root->sort_pathkeys =
 802                         make_pathkeys_for_sortclauses(root,
 803                                                                                   parse->sortClause,
 804                                                                                   tlist,
 805                                                                                   false);
 806
 807                 /*
 808                  * Will need actual number of aggregates for estimating costs.
 809                  *
 810                  * Note: we do not attempt to detect duplicate aggregates here; a
 811                  * somewhat-overestimated count is okay for our present purposes.
 812                  *
 813                  * Note: think not that we can turn off hasAggs if we find no aggs. It
 814                  * is possible for constant-expression simplification to remove all
 815                  * explicit references to aggs, but we still have to follow the
 816                  * aggregate semantics (eg, producing only one output row).
 817                  */
 818                 if (parse->hasAggs)
 819                 {
 820                         count_agg_clauses((Node *) tlist, &agg_counts);
 821                         count_agg_clauses(parse->havingQual, &agg_counts);
 822                 }
 823
 824                 /*
 825                  * Figure out whether we need a sorted result from query_planner.
 826                  *
 827                  * If we have a GROUP BY clause, then we want a result sorted properly
 828                  * for grouping.  Otherwise, if there is an ORDER BY clause, we want
 829                  * to sort by the ORDER BY clause.      (Note: if we have both, and ORDER
 830                  * BY is a superset of GROUP BY, it would be tempting to request sort
 831                  * by ORDER BY --- but that might just leave us failing to exploit an
 832                  * available sort order at all. Needs more thought...)
 833                  */
 834                 if (parse->groupClause)
 835                         root->query_pathkeys = root->group_pathkeys;
 836                 else if (parse->sortClause)
 837                         root->query_pathkeys = root->sort_pathkeys;
 838                 else
 839                         root->query_pathkeys = NIL;
 840
 841                 /*
 842                  * Generate the best unsorted and presorted paths for this Query (but
 843                  * note there may not be any presorted path).  query_planner will also
 844                  * estimate the number of groups in the query, and canonicalize all
 845                  * the pathkeys.
 846                  */
 847                 query_planner(root, sub_tlist, tuple_fraction,
 848                                           &cheapest_path, &sorted_path, &dNumGroups);
 849
 850                 group_pathkeys = root->group_pathkeys;
 851                 sort_pathkeys = root->sort_pathkeys;
 852
 853                 /*
 854                  * If grouping, extract the grouping operators and decide whether we
 855                  * want to use hashed grouping.
 856                  */
 857                 if (parse->groupClause)
 858                 {
 859                         groupOperators = extract_grouping_ops(parse->groupClause);
 860                         use_hashed_grouping =
 861                                 choose_hashed_grouping(root, tuple_fraction,
 862                                                                            cheapest_path, sorted_path,
 863                                                                            groupOperators, dNumGroups,
 864                                                                            &agg_counts);
 865
 866                         /* Also convert # groups to long int --- but 'ware overflow! */
 867                         numGroups = (long) Min(dNumGroups, (double) LONG_MAX);
 868                 }
 869
 870                 /*
 871                  * Select the best path.  If we are doing hashed grouping, we will
 872                  * always read all the input tuples, so use the cheapest-total path.
 873                  * Otherwise, trust query_planner's decision about which to use.
 874                  */
 875                 if (use_hashed_grouping || !sorted_path)
 876                         best_path = cheapest_path;
 877                 else
 878                         best_path = sorted_path;
 879
 880                 /*
 881                  * Check to see if it's possible to optimize MIN/MAX aggregates. If
 882                  * so, we will forget all the work we did so far to choose a "regular"
 883                  * path ... but we had to do it anyway to be able to tell which way is
 884                  * cheaper.
 885                  */
 886                 result_plan = optimize_minmax_aggregates(root,
 887                                                                                                  tlist,
 888                                                                                                  best_path);
 889                 if (result_plan != NULL)
 890                 {
 891                         /*
 892                          * optimize_minmax_aggregates generated the full plan, with the
 893                          * right tlist, and it has no sort order.
 894                          */
 895                         current_pathkeys = NIL;
 896                 }
 897                 else
 898                 {
 899                         /*
 900                          * Normal case --- create a plan according to query_planner's
 901                          * results.
 902                          */
 903                         result_plan = create_plan(root, best_path);
 904                         current_pathkeys = best_path->pathkeys;
 905
 906                         /*
 907                          * create_plan() returns a plan with just a "flat" tlist of
 908                          * required Vars.  Usually we need to insert the sub_tlist as the
 909                          * tlist of the top plan node.  However, we can skip that if we
 910                          * determined that whatever query_planner chose to return will be
 911                          * good enough.
 912                          */
 913                         if (need_tlist_eval)
 914                         {
 915                                 /*
 916                                  * If the top-level plan node is one that cannot do expression
 917                                  * evaluation, we must insert a Result node to project the
 918                                  * desired tlist.
 919                                  */
 920                                 if (!is_projection_capable_plan(result_plan))
 921                                 {
 922                                         result_plan = (Plan *) make_result(root,
 923                                                                                                            sub_tlist,
 924                                                                                                            NULL,
 925                                                                                                            result_plan);
 926                                 }
 927                                 else
 928                                 {
 929                                         /*
 930                                          * Otherwise, just replace the subplan's flat tlist with
 931                                          * the desired tlist.
 932                                          */
 933                                         result_plan->targetlist = sub_tlist;
 934                                 }
 935
 936                                 /*
 937                                  * Also, account for the cost of evaluation of the sub_tlist.
 938                                  *
 939                                  * Up to now, we have only been dealing with "flat" tlists,
 940                                  * containing just Vars.  So their evaluation cost is zero
 941                                  * according to the model used by cost_qual_eval() (or if you
 942                                  * prefer, the cost is factored into cpu_tuple_cost).  Thus we
 943                                  * can avoid accounting for tlist cost throughout
 944                                  * query_planner() and subroutines.  But now we've inserted a
 945                                  * tlist that might contain actual operators, sub-selects, etc
 946                                  * --- so we'd better account for its cost.
 947                                  *
 948                                  * Below this point, any tlist eval cost for added-on nodes
 949                                  * should be accounted for as we create those nodes.
 950                                  * Presently, of the node types we can add on, only Agg and
 951                                  * Group project new tlists (the rest just copy their input
 952                                  * tuples) --- so make_agg() and make_group() are responsible
 953                                  * for computing the added cost.
 954                                  */
 955                                 cost_qual_eval(&tlist_cost, sub_tlist, root);
 956                                 result_plan->startup_cost += tlist_cost.startup;
 957                                 result_plan->total_cost += tlist_cost.startup +
 958                                         tlist_cost.per_tuple * result_plan->plan_rows;
 959                         }
 960                         else
 961                         {
 962                                 /*
 963                                  * Since we're using query_planner's tlist and not the one
 964                                  * make_subplanTargetList calculated, we have to refigure any
 965                                  * grouping-column indexes make_subplanTargetList computed.
 966                                  */
 967                                 locate_grouping_columns(root, tlist, result_plan->targetlist,
 968                                                                                 groupColIdx);
 969                         }
 970
 971                         /*
 972                          * Insert AGG or GROUP node if needed, plus an explicit sort step
 973                          * if necessary.
 974                          *
 975                          * HAVING clause, if any, becomes qual of the Agg or Group node.
 976                          */
 977                         if (use_hashed_grouping)
 978                         {
 979                                 /* Hashed aggregate plan --- no sort needed */
 980                                 result_plan = (Plan *) make_agg(root,
 981                                                                                                 tlist,
 982                                                                                                 (List *) parse->havingQual,
 983                                                                                                 AGG_HASHED,
 984                                                                                                 numGroupCols,
 985                                                                                                 groupColIdx,
 986                                                                                                 groupOperators,
 987                                                                                                 numGroups,
 988                                                                                                 agg_counts.numAggs,
 989                                                                                                 result_plan);
 990                                 /* Hashed aggregation produces randomly-ordered results */
 991                                 current_pathkeys = NIL;
 992                         }
 993                         else if (parse->hasAggs)
 994                         {
 995                                 /* Plain aggregate plan --- sort if needed */
 996                                 AggStrategy aggstrategy;
 997
 998                                 if (parse->groupClause)
 999                                 {
1000                                         if (!pathkeys_contained_in(group_pathkeys,
1001                                                                                            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                                         aggstrategy = AGG_SORTED;
1011
1012                                         /*
1013                                          * The AGG node will not change the sort ordering of its
1014                                          * groups, so current_pathkeys describes the result too.
1015                                          */
1016                                 }
1017                                 else
1018                                 {
1019                                         aggstrategy = AGG_PLAIN;
1020                                         /* Result will be only one row anyway; no sort order */
1021                                         current_pathkeys = NIL;
1022                                 }
1023
1024                                 result_plan = (Plan *) make_agg(root,
1025                                                                                                 tlist,
1026                                                                                                 (List *) parse->havingQual,
1027                                                                                                 aggstrategy,
1028                                                                                                 numGroupCols,
1029                                                                                                 groupColIdx,
1030                                                                                                 groupOperators,
1031                                                                                                 numGroups,
1032                                                                                                 agg_counts.numAggs,
1033                                                                                                 result_plan);
1034                         }
1035                         else if (parse->groupClause)
1036                         {
1037                                 /*
1038                                  * GROUP BY without aggregation, so insert a group node (plus
1039                                  * the appropriate sort node, if necessary).
1040                                  *
1041                                  * Add an explicit sort if we couldn't make the path come out
1042                                  * the way the GROUP node needs it.
1043                                  */
1044                                 if (!pathkeys_contained_in(group_pathkeys, current_pathkeys))
1045                                 {
1046                                         result_plan = (Plan *)
1047                                                 make_sort_from_groupcols(root,
1048                                                                                                  parse->groupClause,
1049                                                                                                  groupColIdx,
1050                                                                                                  result_plan);
1051                                         current_pathkeys = group_pathkeys;
1052                                 }
1053
1054                                 result_plan = (Plan *) make_group(root,
1055                                                                                                   tlist,
1056                                                                                                   (List *) parse->havingQual,
1057                                                                                                   numGroupCols,
1058                                                                                                   groupColIdx,
1059                                                                                                   groupOperators,
1060                                                                                                   dNumGroups,
1061                                                                                                   result_plan);
1062                                 /* The Group node won't change sort ordering */
1063                         }
1064                         else if (root->hasHavingQual)
1065                         {
1066                                 /*
1067                                  * No aggregates, and no GROUP BY, but we have a HAVING qual.
1068                                  * This is a degenerate case in which we are supposed to emit
1069                                  * either 0 or 1 row depending on whether HAVING succeeds.
1070                                  * Furthermore, there cannot be any variables in either HAVING
1071                                  * or the targetlist, so we actually do not need the FROM
1072                                  * table at all!  We can just throw away the plan-so-far and
1073                                  * generate a Result node.      This is a sufficiently unusual
1074                                  * corner case that it's not worth contorting the structure of
1075                                  * this routine to avoid having to generate the plan in the
1076                                  * first place.
1077                                  */
1078                                 result_plan = (Plan *) make_result(root,
1079                                                                                                    tlist,
1080                                                                                                    parse->havingQual,
1081                                                                                                    NULL);
1082                         }
1083                 }                                               /* end of non-minmax-aggregate case */
1084         }                                                       /* end of if (setOperations) */
1085
1086         /*
1087          * If we were not able to make the plan come out in the right order, add
1088          * an explicit sort step.
1089          */
1090         if (parse->sortClause)
1091         {
1092                 if (!pathkeys_contained_in(sort_pathkeys, current_pathkeys))
1093                 {
1094                         result_plan = (Plan *) make_sort_from_pathkeys(root,
1095                                                                                                                    result_plan,
1096                                                                                                                    sort_pathkeys);
1097                         current_pathkeys = sort_pathkeys;
1098                 }
1099         }
1100
1101         /*
1102          * If there is a DISTINCT clause, add the UNIQUE node.
1103          */
1104         if (parse->distinctClause)
1105         {
1106                 result_plan = (Plan *) make_unique(result_plan, parse->distinctClause);
1107
1108                 /*
1109                  * If there was grouping or aggregation, leave plan_rows as-is (ie,
1110                  * assume the result was already mostly unique).  If not, use the
1111                  * number of distinct-groups calculated by query_planner.
1112                  */
1113                 if (!parse->groupClause && !root->hasHavingQual && !parse->hasAggs)
1114                         result_plan->plan_rows = dNumGroups;
1115         }
1116
1117         /*
1118          * Finally, if there is a LIMIT/OFFSET clause, add the LIMIT node.
1119          */
1120         if (parse->limitCount || parse->limitOffset)
1121         {
1122                 result_plan = (Plan *) make_limit(result_plan,
1123                                                                                   parse->limitOffset,
1124                                                                                   parse->limitCount,
1125                                                                                   offset_est,
1126                                                                                   count_est);
1127         }
1128
1129         /*
1130          * Deal with the RETURNING clause if any.  It's convenient to pass the
1131          * returningList through setrefs.c now rather than at top level (if we
1132          * waited, handling inherited UPDATE/DELETE would be much harder).
1133          */
1134         if (parse->returningList)
1135         {
1136                 List       *rlist;
1137
1138                 Assert(parse->resultRelation);
1139                 rlist = set_returning_clause_references(parse->returningList,
1140                                                                                                 result_plan,
1141                                                                                                 parse->resultRelation);
1142                 root->returningLists = list_make1(rlist);
1143         }
1144         else
1145                 root->returningLists = NIL;
1146
1147         /* Compute result-relations list if needed */
1148         if (parse->resultRelation)
1149                 root->resultRelations = list_make1_int(parse->resultRelation);
1150         else
1151                 root->resultRelations = NIL;
1152
1153         /*
1154          * Return the actual output ordering in query_pathkeys for possible use by
1155          * an outer query level.
1156          */
1157         root->query_pathkeys = current_pathkeys;
1158
1159         return result_plan;
1160 }
1161
1162 /*
1163  * Detect whether a plan node is a "dummy" plan created when a relation
1164  * is deemed not to need scanning due to constraint exclusion.
1165  *
1166  * Currently, such dummy plans are Result nodes with constant FALSE
1167  * filter quals.
1168  */
1169 static bool
1170 is_dummy_plan(Plan *plan)
1171 {
1172         if (IsA(plan, Result))
1173         {
1174                 List       *rcqual = (List *) ((Result *) plan)->resconstantqual;
1175
1176                 if (list_length(rcqual) == 1)
1177                 {
1178                         Const      *constqual = (Const *) linitial(rcqual);
1179
1180                         if (constqual && IsA(constqual, Const))
1181                         {
1182                                 if (!constqual->constisnull &&
1183                                         !DatumGetBool(constqual->constvalue))
1184                                         return true;
1185                         }
1186                 }
1187         }
1188         return false;
1189 }
1190
1191 /*
1192  * preprocess_limit - do pre-estimation for LIMIT and/or OFFSET clauses
1193  *
1194  * We try to estimate the values of the LIMIT/OFFSET clauses, and pass the
1195  * results back in *count_est and *offset_est.  These variables are set to
1196  * 0 if the corresponding clause is not present, and -1 if it's present
1197  * but we couldn't estimate the value for it.  (The "0" convention is OK
1198  * for OFFSET but a little bit bogus for LIMIT: effectively we estimate
1199  * LIMIT 0 as though it were LIMIT 1.  But this is in line with the planner's
1200  * usual practice of never estimating less than one row.)  These values will
1201  * be passed to make_limit, which see if you change this code.
1202  *
1203  * The return value is the suitably adjusted tuple_fraction to use for
1204  * planning the query.  This adjustment is not overridable, since it reflects
1205  * plan actions that grouping_planner() will certainly take, not assumptions
1206  * about context.
1207  */
1208 static double
1209 preprocess_limit(PlannerInfo *root, double tuple_fraction,
1210                                  int64 *offset_est, int64 *count_est)
1211 {
1212         Query      *parse = root->parse;
1213         Node       *est;
1214         double          limit_fraction;
1215
1216         /* Should not be called unless LIMIT or OFFSET */
1217         Assert(parse->limitCount || parse->limitOffset);
1218
1219         /*
1220          * Try to obtain the clause values.  We use estimate_expression_value
1221          * primarily because it can sometimes do something useful with Params.
1222          */
1223         if (parse->limitCount)
1224         {
1225                 est = estimate_expression_value(root, parse->limitCount);
1226                 if (est && IsA(est, Const))
1227                 {
1228                         if (((Const *) est)->constisnull)
1229                         {
1230                                 /* NULL indicates LIMIT ALL, ie, no limit */
1231                                 *count_est = 0; /* treat as not present */
1232                         }
1233                         else
1234                         {
1235                                 *count_est = DatumGetInt64(((Const *) est)->constvalue);
1236                                 if (*count_est <= 0)
1237                                         *count_est = 1;         /* force to at least 1 */
1238                         }
1239                 }
1240                 else
1241                         *count_est = -1;        /* can't estimate */
1242         }
1243         else
1244                 *count_est = 0;                 /* not present */
1245
1246         if (parse->limitOffset)
1247         {
1248                 est = estimate_expression_value(root, parse->limitOffset);
1249                 if (est && IsA(est, Const))
1250                 {
1251                         if (((Const *) est)->constisnull)
1252                         {
1253                                 /* Treat NULL as no offset; the executor will too */
1254                                 *offset_est = 0;        /* treat as not present */
1255                         }
1256                         else
1257                         {
1258                                 *offset_est = DatumGetInt64(((Const *) est)->constvalue);
1259                                 if (*offset_est < 0)
1260                                         *offset_est = 0;        /* less than 0 is same as 0 */
1261                         }
1262                 }
1263                 else
1264                         *offset_est = -1;       /* can't estimate */
1265         }
1266         else
1267                 *offset_est = 0;                /* not present */
1268
1269         if (*count_est != 0)
1270         {
1271                 /*
1272                  * A LIMIT clause limits the absolute number of tuples returned.
1273                  * However, if it's not a constant LIMIT then we have to guess; for
1274                  * lack of a better idea, assume 10% of the plan's result is wanted.
1275                  */
1276                 if (*count_est < 0 || *offset_est < 0)
1277                 {
1278                         /* LIMIT or OFFSET is an expression ... punt ... */
1279                         limit_fraction = 0.10;
1280                 }
1281                 else
1282                 {
1283                         /* LIMIT (plus OFFSET, if any) is max number of tuples needed */
1284                         limit_fraction = (double) *count_est + (double) *offset_est;
1285                 }
1286
1287                 /*
1288                  * If we have absolute limits from both caller and LIMIT, use the
1289                  * smaller value; likewise if they are both fractional.  If one is
1290                  * fractional and the other absolute, we can't easily determine which
1291                  * is smaller, but we use the heuristic that the absolute will usually
1292                  * be smaller.
1293                  */
1294                 if (tuple_fraction >= 1.0)
1295                 {
1296                         if (limit_fraction >= 1.0)
1297                         {
1298                                 /* both absolute */
1299                                 tuple_fraction = Min(tuple_fraction, limit_fraction);
1300                         }
1301                         else
1302                         {
1303                                 /* caller absolute, limit fractional; use caller's value */
1304                         }
1305                 }
1306                 else if (tuple_fraction > 0.0)
1307                 {
1308                         if (limit_fraction >= 1.0)
1309                         {
1310                                 /* caller fractional, limit absolute; use limit */
1311                                 tuple_fraction = limit_fraction;
1312                         }
1313                         else
1314                         {
1315                                 /* both fractional */
1316                                 tuple_fraction = Min(tuple_fraction, limit_fraction);
1317                         }
1318                 }
1319                 else
1320                 {
1321                         /* no info from caller, just use limit */
1322                         tuple_fraction = limit_fraction;
1323                 }
1324         }
1325         else if (*offset_est != 0 && tuple_fraction > 0.0)
1326         {
1327                 /*
1328                  * We have an OFFSET but no LIMIT.      This acts entirely differently
1329                  * from the LIMIT case: here, we need to increase rather than decrease
1330                  * the caller's tuple_fraction, because the OFFSET acts to cause more
1331                  * tuples to be fetched instead of fewer.  This only matters if we got
1332                  * a tuple_fraction > 0, however.
1333                  *
1334                  * As above, use 10% if OFFSET is present but unestimatable.
1335                  */
1336                 if (*offset_est < 0)
1337                         limit_fraction = 0.10;
1338                 else
1339                         limit_fraction = (double) *offset_est;
1340
1341                 /*
1342                  * If we have absolute counts from both caller and OFFSET, add them
1343                  * together; likewise if they are both fractional.      If one is
1344                  * fractional and the other absolute, we want to take the larger, and
1345                  * we heuristically assume that's the fractional one.
1346                  */
1347                 if (tuple_fraction >= 1.0)
1348                 {
1349                         if (limit_fraction >= 1.0)
1350                         {
1351                                 /* both absolute, so add them together */
1352                                 tuple_fraction += limit_fraction;
1353                         }
1354                         else
1355                         {
1356                                 /* caller absolute, limit fractional; use limit */
1357                                 tuple_fraction = limit_fraction;
1358                         }
1359                 }
1360                 else
1361                 {
1362                         if (limit_fraction >= 1.0)
1363                         {
1364                                 /* caller fractional, limit absolute; use caller's value */
1365                         }
1366                         else
1367                         {
1368                                 /* both fractional, so add them together */
1369                                 tuple_fraction += limit_fraction;
1370                                 if (tuple_fraction >= 1.0)
1371                                         tuple_fraction = 0.0;           /* assume fetch all */
1372                         }
1373                 }
1374         }
1375
1376         return tuple_fraction;
1377 }
1378
1379 /*
1380  * extract_grouping_ops - make an array of the equality operator OIDs
1381  *              for the GROUP BY clause
1382  */
1383 static Oid *
1384 extract_grouping_ops(List *groupClause)
1385 {
1386         int                     numCols = list_length(groupClause);
1387         int                     colno = 0;
1388         Oid                *groupOperators;
1389         ListCell   *glitem;
1390
1391         groupOperators = (Oid *) palloc(sizeof(Oid) * numCols);
1392
1393         foreach(glitem, groupClause)
1394         {
1395                 GroupClause *groupcl = (GroupClause *) lfirst(glitem);
1396
1397                 groupOperators[colno] = get_equality_op_for_ordering_op(groupcl->sortop);
1398                 if (!OidIsValid(groupOperators[colno]))         /* shouldn't happen */
1399                         elog(ERROR, "could not find equality operator for ordering operator %u",
1400                                  groupcl->sortop);
1401                 colno++;
1402         }
1403
1404         return groupOperators;
1405 }
1406
1407 /*
1408  * choose_hashed_grouping - should we use hashed grouping?
1409  */
1410 static bool
1411 choose_hashed_grouping(PlannerInfo *root, double tuple_fraction,
1412                                            Path *cheapest_path, Path *sorted_path,
1413                                            Oid *groupOperators, double dNumGroups,
1414                                            AggClauseCounts *agg_counts)
1415 {
1416         int                     numGroupCols = list_length(root->parse->groupClause);
1417         double          cheapest_path_rows;
1418         int                     cheapest_path_width;
1419         Size            hashentrysize;
1420         List       *current_pathkeys;
1421         Path            hashed_p;
1422         Path            sorted_p;
1423         int                     i;
1424
1425         /*
1426          * Check can't-do-it conditions, including whether the grouping operators
1427          * are hashjoinable.  (We assume hashing is OK if they are marked
1428          * oprcanhash.  If there isn't actually a supporting hash function,
1429          * the executor will complain at runtime.)
1430          *
1431          * Executor doesn't support hashed aggregation with DISTINCT aggregates.
1432          * (Doing so would imply storing *all* the input values in the hash table,
1433          * which seems like a certain loser.)
1434          */
1435         if (!enable_hashagg)
1436                 return false;
1437         if (agg_counts->numDistinctAggs != 0)
1438                 return false;
1439         for (i = 0; i < numGroupCols; i++)
1440         {
1441                 if (!op_hashjoinable(groupOperators[i]))
1442                         return false;
1443         }
1444
1445         /*
1446          * Don't do it if it doesn't look like the hashtable will fit into
1447          * work_mem.
1448          *
1449          * Beware here of the possibility that cheapest_path->parent is NULL. This
1450          * could happen if user does something silly like SELECT 'foo' GROUP BY 1;
1451          */
1452         if (cheapest_path->parent)
1453         {
1454                 cheapest_path_rows = cheapest_path->parent->rows;
1455                 cheapest_path_width = cheapest_path->parent->width;
1456         }
1457         else
1458         {
1459                 cheapest_path_rows = 1; /* assume non-set result */
1460                 cheapest_path_width = 100;              /* arbitrary */
1461         }
1462
1463         /* Estimate per-hash-entry space at tuple width... */
1464         hashentrysize = MAXALIGN(cheapest_path_width) + MAXALIGN(sizeof(MinimalTupleData));
1465         /* plus space for pass-by-ref transition values... */
1466         hashentrysize += agg_counts->transitionSpace;
1467         /* plus the per-hash-entry overhead */
1468         hashentrysize += hash_agg_entry_size(agg_counts->numAggs);
1469
1470         if (hashentrysize * dNumGroups > work_mem * 1024L)
1471                 return false;
1472
1473         /*
1474          * See if the estimated cost is no more than doing it the other way. While
1475          * avoiding the need for sorted input is usually a win, the fact that the
1476          * output won't be sorted may be a loss; so we need to do an actual cost
1477          * comparison.
1478          *
1479          * We need to consider cheapest_path + hashagg [+ final sort] versus
1480          * either cheapest_path [+ sort] + group or agg [+ final sort] or
1481          * presorted_path + group or agg [+ final sort] where brackets indicate a
1482          * step that may not be needed. We assume query_planner() will have
1483          * returned a presorted path only if it's a winner compared to
1484          * cheapest_path for this purpose.
1485          *
1486          * These path variables are dummies that just hold cost fields; we don't
1487          * make actual Paths for these steps.
1488          */
1489         cost_agg(&hashed_p, root, AGG_HASHED, agg_counts->numAggs,
1490                          numGroupCols, dNumGroups,
1491                          cheapest_path->startup_cost, cheapest_path->total_cost,
1492                          cheapest_path_rows);
1493         /* Result of hashed agg is always unsorted */
1494         if (root->sort_pathkeys)
1495                 cost_sort(&hashed_p, root, root->sort_pathkeys, hashed_p.total_cost,
1496                                   dNumGroups, cheapest_path_width);
1497
1498         if (sorted_path)
1499         {
1500                 sorted_p.startup_cost = sorted_path->startup_cost;
1501                 sorted_p.total_cost = sorted_path->total_cost;
1502                 current_pathkeys = sorted_path->pathkeys;
1503         }
1504         else
1505         {
1506                 sorted_p.startup_cost = cheapest_path->startup_cost;
1507                 sorted_p.total_cost = cheapest_path->total_cost;
1508                 current_pathkeys = cheapest_path->pathkeys;
1509         }
1510         if (!pathkeys_contained_in(root->group_pathkeys, current_pathkeys))
1511         {
1512                 cost_sort(&sorted_p, root, root->group_pathkeys, sorted_p.total_cost,
1513                                   cheapest_path_rows, cheapest_path_width);
1514                 current_pathkeys = root->group_pathkeys;
1515         }
1516
1517         if (root->parse->hasAggs)
1518                 cost_agg(&sorted_p, root, AGG_SORTED, agg_counts->numAggs,
1519                                  numGroupCols, dNumGroups,
1520                                  sorted_p.startup_cost, sorted_p.total_cost,
1521                                  cheapest_path_rows);
1522         else
1523                 cost_group(&sorted_p, root, numGroupCols, dNumGroups,
1524                                    sorted_p.startup_cost, sorted_p.total_cost,
1525                                    cheapest_path_rows);
1526         /* The Agg or Group node will preserve ordering */
1527         if (root->sort_pathkeys &&
1528                 !pathkeys_contained_in(root->sort_pathkeys, current_pathkeys))
1529                 cost_sort(&sorted_p, root, root->sort_pathkeys, sorted_p.total_cost,
1530                                   dNumGroups, cheapest_path_width);
1531
1532         /*
1533          * Now make the decision using the top-level tuple fraction.  First we
1534          * have to convert an absolute count (LIMIT) into fractional form.
1535          */
1536         if (tuple_fraction >= 1.0)
1537                 tuple_fraction /= dNumGroups;
1538
1539         if (compare_fractional_path_costs(&hashed_p, &sorted_p,
1540                                                                           tuple_fraction) < 0)
1541         {
1542                 /* Hashed is cheaper, so use it */
1543                 return true;
1544         }
1545         return false;
1546 }
1547
1548 /*---------------
1549  * make_subplanTargetList
1550  *        Generate appropriate target list when grouping is required.
1551  *
1552  * When grouping_planner inserts Aggregate, Group, or Result plan nodes
1553  * above the result of query_planner, we typically want to pass a different
1554  * target list to query_planner than the outer plan nodes should have.
1555  * This routine generates the correct target list for the subplan.
1556  *
1557  * The initial target list passed from the parser already contains entries
1558  * for all ORDER BY and GROUP BY expressions, but it will not have entries
1559  * for variables used only in HAVING clauses; so we need to add those
1560  * variables to the subplan target list.  Also, we flatten all expressions
1561  * except GROUP BY items into their component variables; the other expressions
1562  * will be computed by the inserted nodes rather than by the subplan.
1563  * For example, given a query like
1564  *              SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
1565  * we want to pass this targetlist to the subplan:
1566  *              a,b,c,d,a+b
1567  * where the a+b target will be used by the Sort/Group steps, and the
1568  * other targets will be used for computing the final results.  (In the
1569  * above example we could theoretically suppress the a and b targets and
1570  * pass down only c,d,a+b, but it's not really worth the trouble to
1571  * eliminate simple var references from the subplan.  We will avoid doing
1572  * the extra computation to recompute a+b at the outer level; see
1573  * fix_upper_expr() in setrefs.c.)
1574  *
1575  * If we are grouping or aggregating, *and* there are no non-Var grouping
1576  * expressions, then the returned tlist is effectively dummy; we do not
1577  * need to force it to be evaluated, because all the Vars it contains
1578  * should be present in the output of query_planner anyway.
1579  *
1580  * 'tlist' is the query's target list.
1581  * 'groupColIdx' receives an array of column numbers for the GROUP BY
1582  *                      expressions (if there are any) in the subplan's target list.
1583  * 'need_tlist_eval' is set true if we really need to evaluate the
1584  *                      result tlist.
1585  *
1586  * The result is the targetlist to be passed to the subplan.
1587  *---------------
1588  */
1589 static List *
1590 make_subplanTargetList(PlannerInfo *root,
1591                                            List *tlist,
1592                                            AttrNumber **groupColIdx,
1593                                            bool *need_tlist_eval)
1594 {
1595         Query      *parse = root->parse;
1596         List       *sub_tlist;
1597         List       *extravars;
1598         int                     numCols;
1599
1600         *groupColIdx = NULL;
1601
1602         /*
1603          * If we're not grouping or aggregating, there's nothing to do here;
1604          * query_planner should receive the unmodified target list.
1605          */
1606         if (!parse->hasAggs && !parse->groupClause && !root->hasHavingQual)
1607         {
1608                 *need_tlist_eval = true;
1609                 return tlist;
1610         }
1611
1612         /*
1613          * Otherwise, start with a "flattened" tlist (having just the vars
1614          * mentioned in the targetlist and HAVING qual --- but not upper- level
1615          * Vars; they will be replaced by Params later on).
1616          */
1617         sub_tlist = flatten_tlist(tlist);
1618         extravars = pull_var_clause(parse->havingQual, false);
1619         sub_tlist = add_to_flat_tlist(sub_tlist, extravars);
1620         list_free(extravars);
1621         *need_tlist_eval = false;       /* only eval if not flat tlist */
1622
1623         /*
1624          * If grouping, create sub_tlist entries for all GROUP BY expressions
1625          * (GROUP BY items that are simple Vars should be in the list already),
1626          * and make an array showing where the group columns are in the sub_tlist.
1627          */
1628         numCols = list_length(parse->groupClause);
1629         if (numCols > 0)
1630         {
1631                 int                     keyno = 0;
1632                 AttrNumber *grpColIdx;
1633                 ListCell   *gl;
1634
1635                 grpColIdx = (AttrNumber *) palloc(sizeof(AttrNumber) * numCols);
1636                 *groupColIdx = grpColIdx;
1637
1638                 foreach(gl, parse->groupClause)
1639                 {
1640                         GroupClause *grpcl = (GroupClause *) lfirst(gl);
1641                         Node       *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1642                         TargetEntry *te = NULL;
1643                         ListCell   *sl;
1644
1645                         /* Find or make a matching sub_tlist entry */
1646                         foreach(sl, sub_tlist)
1647                         {
1648                                 te = (TargetEntry *) lfirst(sl);
1649                                 if (equal(groupexpr, te->expr))
1650                                         break;
1651                         }
1652                         if (!sl)
1653                         {
1654                                 te = makeTargetEntry((Expr *) groupexpr,
1655                                                                          list_length(sub_tlist) + 1,
1656                                                                          NULL,
1657                                                                          false);
1658                                 sub_tlist = lappend(sub_tlist, te);
1659                                 *need_tlist_eval = true;                /* it's not flat anymore */
1660                         }
1661
1662                         /* and save its resno */
1663                         grpColIdx[keyno++] = te->resno;
1664                 }
1665         }
1666
1667         return sub_tlist;
1668 }
1669
1670 /*
1671  * locate_grouping_columns
1672  *              Locate grouping columns in the tlist chosen by query_planner.
1673  *
1674  * This is only needed if we don't use the sub_tlist chosen by
1675  * make_subplanTargetList.      We have to forget the column indexes found
1676  * by that routine and re-locate the grouping vars in the real sub_tlist.
1677  */
1678 static void
1679 locate_grouping_columns(PlannerInfo *root,
1680                                                 List *tlist,
1681                                                 List *sub_tlist,
1682                                                 AttrNumber *groupColIdx)
1683 {
1684         int                     keyno = 0;
1685         ListCell   *gl;
1686
1687         /*
1688          * No work unless grouping.
1689          */
1690         if (!root->parse->groupClause)
1691         {
1692                 Assert(groupColIdx == NULL);
1693                 return;
1694         }
1695         Assert(groupColIdx != NULL);
1696
1697         foreach(gl, root->parse->groupClause)
1698         {
1699                 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1700                 Node       *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1701                 TargetEntry *te = NULL;
1702                 ListCell   *sl;
1703
1704                 foreach(sl, sub_tlist)
1705                 {
1706                         te = (TargetEntry *) lfirst(sl);
1707                         if (equal(groupexpr, te->expr))
1708                                 break;
1709                 }
1710                 if (!sl)
1711                         elog(ERROR, "failed to locate grouping columns");
1712
1713                 groupColIdx[keyno++] = te->resno;
1714         }
1715 }
1716
1717 /*
1718  * postprocess_setop_tlist
1719  *        Fix up targetlist returned by plan_set_operations().
1720  *
1721  * We need to transpose sort key info from the orig_tlist into new_tlist.
1722  * NOTE: this would not be good enough if we supported resjunk sort keys
1723  * for results of set operations --- then, we'd need to project a whole
1724  * new tlist to evaluate the resjunk columns.  For now, just ereport if we
1725  * find any resjunk columns in orig_tlist.
1726  */
1727 static List *
1728 postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
1729 {
1730         ListCell   *l;
1731         ListCell   *orig_tlist_item = list_head(orig_tlist);
1732
1733         foreach(l, new_tlist)
1734         {
1735                 TargetEntry *new_tle = (TargetEntry *) lfirst(l);
1736                 TargetEntry *orig_tle;
1737
1738                 /* ignore resjunk columns in setop result */
1739                 if (new_tle->resjunk)
1740                         continue;
1741
1742                 Assert(orig_tlist_item != NULL);
1743                 orig_tle = (TargetEntry *) lfirst(orig_tlist_item);
1744                 orig_tlist_item = lnext(orig_tlist_item);
1745                 if (orig_tle->resjunk)  /* should not happen */
1746                         elog(ERROR, "resjunk output columns are not implemented");
1747                 Assert(new_tle->resno == orig_tle->resno);
1748                 new_tle->ressortgroupref = orig_tle->ressortgroupref;
1749         }
1750         if (orig_tlist_item != NULL)
1751                 elog(ERROR, "resjunk output columns are not implemented");
1752         return new_tlist;
1753 }