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 1.216 2007/02/27 01:11:25 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, bool isCursor, int cursorOptions,
  84                 ParamListInfo boundParams)
  85 {
  86         PlannedStmt *result;
  87         PlannerGlobal *glob;
  88         double          tuple_fraction;
  89         PlannerInfo *root;
  90         Plan       *top_plan;
  91         ListCell   *lp,
  92                            *lr;
  93
  94         /*
  95          * Set up global state for this planner invocation.  This data is needed
  96          * across all levels of sub-Query that might exist in the given command,
  97          * so we keep it in a separate struct that's linked to by each per-Query
  98          * PlannerInfo.
  99          */
 100         glob = makeNode(PlannerGlobal);
 101
 102         glob->boundParams = boundParams;
 103         glob->paramlist = NIL;
 104         glob->subplans = NIL;
 105         glob->subrtables = NIL;
 106         glob->rewindPlanIDs = NULL;
 107         glob->finalrtable = NIL;
 108
 109         /* Determine what fraction of the plan is likely to be scanned */
 110         if (isCursor)
 111         {
 112                 /*
 113                  * We have no real idea how many tuples the user will ultimately FETCH
 114                  * from a cursor, but it seems a good bet that he doesn't want 'em
 115                  * all.  Optimize for 10% retrieval (you gotta better number?  Should
 116                  * this be a SETtable parameter?)
 117                  */
 118                 tuple_fraction = 0.10;
 119         }
 120         else
 121         {
 122                 /* Default assumption is we need all the tuples */
 123                 tuple_fraction = 0.0;
 124         }
 125
 126         /* primary planning entry point (may recurse for subqueries) */
 127         top_plan = subquery_planner(glob, parse, 1, tuple_fraction, &root);
 128
 129         /*
 130          * If creating a plan for a scrollable cursor, make sure it can run
 131          * backwards on demand.  Add a Material node at the top at need.
 132          */
 133         if (isCursor && (cursorOptions & CURSOR_OPT_SCROLL))
 134         {
 135                 if (!ExecSupportsBackwardScan(top_plan))
 136                         top_plan = materialize_finished_plan(top_plan);
 137         }
 138
 139         /* final cleanup of the plan */
 140         Assert(glob->finalrtable == NIL);
 141         top_plan = set_plan_references(glob, top_plan, root->parse->rtable);
 142         /* ... and the subplans (both regular subplans and initplans) */
 143         Assert(list_length(glob->subplans) == list_length(glob->subrtables));
 144         forboth(lp, glob->subplans, lr, glob->subrtables)
 145         {
 146                 Plan   *subplan = (Plan *) lfirst(lp);
 147                 List   *subrtable = (List *) lfirst(lr);
 148
 149                 lfirst(lp) = set_plan_references(glob, subplan, subrtable);
 150         }
 151
 152         /* build the PlannedStmt result */
 153         result = makeNode(PlannedStmt);
 154
 155         result->commandType = parse->commandType;
 156         result->canSetTag = parse->canSetTag;
 157         result->planTree = top_plan;
 158         result->rtable = glob->finalrtable;
 159         result->resultRelations = root->resultRelations;
 160         result->into = parse->into;
 161         result->subplans = glob->subplans;
 162         result->rewindPlanIDs = glob->rewindPlanIDs;
 163         result->returningLists = root->returningLists;
 164         result->rowMarks = parse->rowMarks;
 165         result->nParamExec = list_length(glob->paramlist);
 166
 167         return result;
 168 }
 169
 170
 171 /*--------------------
 172  * subquery_planner
 173  *        Invokes the planner on a subquery.  We recurse to here for each
 174  *        sub-SELECT found in the query tree.
 175  *
 176  * glob is the global state for the current planner run.
 177  * parse is the querytree produced by the parser & rewriter.
 178  * level is the current recursion depth (1 at the top-level Query).
 179  * tuple_fraction is the fraction of tuples we expect will be retrieved.
 180  * tuple_fraction is interpreted as explained for grouping_planner, below.
 181  *
 182  * If subroot isn't NULL, we pass back the query's final PlannerInfo struct;
 183  * among other things this tells the output sort ordering of the plan.
 184  *
 185  * Basically, this routine does the stuff that should only be done once
 186  * per Query object.  It then calls grouping_planner.  At one time,
 187  * grouping_planner could be invoked recursively on the same Query object;
 188  * that's not currently true, but we keep the separation between the two
 189  * routines anyway, in case we need it again someday.
 190  *
 191  * subquery_planner will be called recursively to handle sub-Query nodes
 192  * found within the query's expressions and rangetable.
 193  *
 194  * Returns a query plan.
 195  *--------------------
 196  */
 197 Plan *
 198 subquery_planner(PlannerGlobal *glob, Query *parse,
 199                                  Index level, double tuple_fraction,
 200                                  PlannerInfo **subroot)
 201 {
 202         int                     num_old_subplans = list_length(glob->subplans);
 203         PlannerInfo *root;
 204         Plan       *plan;
 205         List       *newHaving;
 206         ListCell   *l;
 207
 208         /* Create a PlannerInfo data structure for this subquery */
 209         root = makeNode(PlannerInfo);
 210         root->parse = parse;
 211         root->glob = glob;
 212         root->query_level = level;
 213         root->planner_cxt = CurrentMemoryContext;
 214         root->init_plans = NIL;
 215         root->eq_classes = NIL;
 216         root->in_info_list = NIL;
 217         root->append_rel_list = NIL;
 218
 219         /*
 220          * Look for IN clauses at the top level of WHERE, and transform them into
 221          * joins.  Note that this step only handles IN clauses originally at top
 222          * level of WHERE; if we pull up any subqueries in the next step, their
 223          * INs are processed just before pulling them up.
 224          */
 225         if (parse->hasSubLinks)
 226                 parse->jointree->quals = pull_up_IN_clauses(root,
 227                                                                                                         parse->jointree->quals);
 228
 229         /*
 230          * Check to see if any subqueries in the rangetable can be merged into
 231          * this query.
 232          */
 233         parse->jointree = (FromExpr *)
 234                 pull_up_subqueries(root, (Node *) parse->jointree, false, false);
 235
 236         /*
 237          * Detect whether any rangetable entries are RTE_JOIN kind; if not, we can
 238          * avoid the expense of doing flatten_join_alias_vars().  Also check for
 239          * outer joins --- if none, we can skip reduce_outer_joins() and some
 240          * other processing.  This must be done after we have done
 241          * pull_up_subqueries, of course.
 242          *
 243          * Note: if reduce_outer_joins manages to eliminate all outer joins,
 244          * root->hasOuterJoins is not reset currently.  This is OK since its
 245          * purpose is merely to suppress unnecessary processing in simple cases.
 246          */
 247         root->hasJoinRTEs = false;
 248         root->hasOuterJoins = false;
 249         foreach(l, parse->rtable)
 250         {
 251                 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
 252
 253                 if (rte->rtekind == RTE_JOIN)
 254                 {
 255                         root->hasJoinRTEs = true;
 256                         if (IS_OUTER_JOIN(rte->jointype))
 257                         {
 258                                 root->hasOuterJoins = true;
 259                                 /* Can quit scanning once we find an outer join */
 260                                 break;
 261                         }
 262                 }
 263         }
 264
 265         /*
 266          * Expand any rangetable entries that are inheritance sets into "append
 267          * relations".  This can add entries to the rangetable, but they must be
 268          * plain base relations not joins, so it's OK (and marginally more
 269          * efficient) to do it after checking for join RTEs.  We must do it after
 270          * pulling up subqueries, else we'd fail to handle inherited tables in
 271          * subqueries.
 272          */
 273         expand_inherited_tables(root);
 274
 275         /*
 276          * Set hasHavingQual to remember if HAVING clause is present.  Needed
 277          * because preprocess_expression will reduce a constant-true condition to
 278          * an empty qual list ... but "HAVING TRUE" is not a semantic no-op.
 279          */
 280         root->hasHavingQual = (parse->havingQual != NULL);
 281
 282         /* Clear this flag; might get set in distribute_qual_to_rels */
 283         root->hasPseudoConstantQuals = false;
 284
 285         /*
 286          * Do expression preprocessing on targetlist and quals.
 287          */
 288         parse->targetList = (List *)
 289                 preprocess_expression(root, (Node *) parse->targetList,
 290                                                           EXPRKIND_TARGET);
 291
 292         parse->returningList = (List *)
 293                 preprocess_expression(root, (Node *) parse->returningList,
 294                                                           EXPRKIND_TARGET);
 295
 296         preprocess_qual_conditions(root, (Node *) parse->jointree);
 297
 298         parse->havingQual = preprocess_expression(root, parse->havingQual,
 299                                                                                           EXPRKIND_QUAL);
 300
 301         parse->limitOffset = preprocess_expression(root, parse->limitOffset,
 302                                                                                            EXPRKIND_LIMIT);
 303         parse->limitCount = preprocess_expression(root, parse->limitCount,
 304                                                                                           EXPRKIND_LIMIT);
 305
 306         root->in_info_list = (List *)
 307                 preprocess_expression(root, (Node *) root->in_info_list,
 308                                                           EXPRKIND_ININFO);
 309         root->append_rel_list = (List *)
 310                 preprocess_expression(root, (Node *) root->append_rel_list,
 311                                                           EXPRKIND_APPINFO);
 312
 313         /* Also need to preprocess expressions for function and values RTEs */
 314         foreach(l, parse->rtable)
 315         {
 316                 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
 317
 318                 if (rte->rtekind == RTE_FUNCTION)
 319                         rte->funcexpr = preprocess_expression(root, rte->funcexpr,
 320                                                                                                   EXPRKIND_RTFUNC);
 321                 else if (rte->rtekind == RTE_VALUES)
 322                         rte->values_lists = (List *)
 323                                 preprocess_expression(root, (Node *) rte->values_lists,
 324                                                                           EXPRKIND_VALUES);
 325         }
 326
 327         /*
 328          * In some cases we may want to transfer a HAVING clause into WHERE. We
 329          * cannot do so if the HAVING clause contains aggregates (obviously) or
 330          * volatile functions (since a HAVING clause is supposed to be executed
 331          * only once per group).  Also, it may be that the clause is so expensive
 332          * to execute that we're better off doing it only once per group, despite
 333          * the loss of selectivity.  This is hard to estimate short of doing the
 334          * entire planning process twice, so we use a heuristic: clauses
 335          * containing subplans are left in HAVING.      Otherwise, we move or copy the
 336          * HAVING clause into WHERE, in hopes of eliminating tuples before
 337          * aggregation instead of after.
 338          *
 339          * If the query has explicit grouping then we can simply move such a
 340          * clause into WHERE; any group that fails the clause will not be in the
 341          * output because none of its tuples will reach the grouping or
 342          * aggregation stage.  Otherwise we must have a degenerate (variable-free)
 343          * HAVING clause, which we put in WHERE so that query_planner() can use it
 344          * in a gating Result node, but also keep in HAVING to ensure that we
 345          * don't emit a bogus aggregated row. (This could be done better, but it
 346          * seems not worth optimizing.)
 347          *
 348          * Note that both havingQual and parse->jointree->quals are in
 349          * implicitly-ANDed-list form at this point, even though they are declared
 350          * as Node *.
 351          */
 352         newHaving = NIL;
 353         foreach(l, (List *) parse->havingQual)
 354         {
 355                 Node       *havingclause = (Node *) lfirst(l);
 356
 357                 if (contain_agg_clause(havingclause) ||
 358                         contain_volatile_functions(havingclause) ||
 359                         contain_subplans(havingclause))
 360                 {
 361                         /* keep it in HAVING */
 362                         newHaving = lappend(newHaving, havingclause);
 363                 }
 364                 else if (parse->groupClause)
 365                 {
 366                         /* move it to WHERE */
 367                         parse->jointree->quals = (Node *)
 368                                 lappend((List *) parse->jointree->quals, havingclause);
 369                 }
 370                 else
 371                 {
 372                         /* put a copy in WHERE, keep it in HAVING */
 373                         parse->jointree->quals = (Node *)
 374                                 lappend((List *) parse->jointree->quals,
 375                                                 copyObject(havingclause));
 376                         newHaving = lappend(newHaving, havingclause);
 377                 }
 378         }
 379         parse->havingQual = (Node *) newHaving;
 380
 381         /*
 382          * If we have any outer joins, try to reduce them to plain inner joins.
 383          * This step is most easily done after we've done expression
 384          * preprocessing.
 385          */
 386         if (root->hasOuterJoins)
 387                 reduce_outer_joins(root);
 388
 389         /*
 390          * Do the main planning.  If we have an inherited target relation, that
 391          * needs special processing, else go straight to grouping_planner.
 392          */
 393         if (parse->resultRelation &&
 394                 rt_fetch(parse->resultRelation, parse->rtable)->inh)
 395                 plan = inheritance_planner(root);
 396         else
 397                 plan = grouping_planner(root, tuple_fraction);
 398
 399         /*
 400          * If any subplans were generated, or if we're inside a subplan, build
 401          * initPlan list and extParam/allParam sets for plan nodes, and attach the
 402          * initPlans to the top plan node.
 403          */
 404         if (list_length(glob->subplans) != num_old_subplans ||
 405                 root->query_level > 1)
 406                 SS_finalize_plan(root, plan);
 407
 408         /* Return internal info if caller wants it */
 409         if (subroot)
 410                 *subroot = root;
 411
 412         return plan;
 413 }
 414
 415 /*
 416  * preprocess_expression
 417  *              Do subquery_planner's preprocessing work for an expression,
 418  *              which can be a targetlist, a WHERE clause (including JOIN/ON
 419  *              conditions), or a HAVING clause.
 420  */
 421 static Node *
 422 preprocess_expression(PlannerInfo *root, Node *expr, int kind)
 423 {
 424         /*
 425          * Fall out quickly if expression is empty.  This occurs often enough to
 426          * be worth checking.  Note that null->null is the correct conversion for
 427          * implicit-AND result format, too.
 428          */
 429         if (expr == NULL)
 430                 return NULL;
 431
 432         /*
 433          * If the query has any join RTEs, replace join alias variables with
 434          * base-relation variables. We must do this before sublink processing,
 435          * else sublinks expanded out from join aliases wouldn't get processed. We
 436          * can skip it in VALUES lists, however, since they can't contain any Vars
 437          * at all.
 438          */
 439         if (root->hasJoinRTEs && kind != EXPRKIND_VALUES)
 440                 expr = flatten_join_alias_vars(root, expr);
 441
 442         /*
 443          * Simplify constant expressions.
 444          *
 445          * Note: this also flattens nested AND and OR expressions into N-argument
 446          * form.  All processing of a qual expression after this point must be
 447          * careful to maintain AND/OR flatness --- that is, do not generate a tree
 448          * with AND directly under AND, nor OR directly under OR.
 449          *
 450          * Because this is a relatively expensive process, we skip it when the
 451          * query is trivial, such as "SELECT 2+2;" or "INSERT ... VALUES()". The
 452          * expression will only be evaluated once anyway, so no point in
 453          * pre-simplifying; we can't execute it any faster than the executor can,
 454          * and we will waste cycles copying the tree.  Notice however that we
 455          * still must do it for quals (to get AND/OR flatness); and if we are in a
 456          * subquery we should not assume it will be done only once.
 457          *
 458          * For VALUES lists we never do this at all, again on the grounds that we
 459          * should optimize for one-time evaluation.
 460          */
 461         if (kind != EXPRKIND_VALUES &&
 462                 (root->parse->jointree->fromlist != NIL ||
 463                  kind == EXPRKIND_QUAL ||
 464                  root->query_level > 1))
 465                 expr = eval_const_expressions(expr);
 466
 467         /*
 468          * If it's a qual or havingQual, canonicalize it.
 469          */
 470         if (kind == EXPRKIND_QUAL)
 471         {
 472                 expr = (Node *) canonicalize_qual((Expr *) expr);
 473
 474 #ifdef OPTIMIZER_DEBUG
 475                 printf("After canonicalize_qual()\n");
 476                 pprint(expr);
 477 #endif
 478         }
 479
 480         /* Expand SubLinks to SubPlans */
 481         if (root->parse->hasSubLinks)
 482                 expr = SS_process_sublinks(root, expr, (kind == EXPRKIND_QUAL));
 483
 484         /*
 485          * XXX do not insert anything here unless you have grokked the comments in
 486          * SS_replace_correlation_vars ...
 487          */
 488
 489         /* Replace uplevel vars with Param nodes (this IS possible in VALUES) */
 490         if (root->query_level > 1)
 491                 expr = SS_replace_correlation_vars(root, expr);
 492
 493         /*
 494          * If it's a qual or havingQual, convert it to implicit-AND format. (We
 495          * don't want to do this before eval_const_expressions, since the latter
 496          * would be unable to simplify a top-level AND correctly. Also,
 497          * SS_process_sublinks expects explicit-AND format.)
 498          */
 499         if (kind == EXPRKIND_QUAL)
 500                 expr = (Node *) make_ands_implicit((Expr *) expr);
 501
 502         return expr;
 503 }
 504
 505 /*
 506  * preprocess_qual_conditions
 507  *              Recursively scan the query's jointree and do subquery_planner's
 508  *              preprocessing work on each qual condition found therein.
 509  */
 510 static void
 511 preprocess_qual_conditions(PlannerInfo *root, Node *jtnode)
 512 {
 513         if (jtnode == NULL)
 514                 return;
 515         if (IsA(jtnode, RangeTblRef))
 516         {
 517                 /* nothing to do here */
 518         }
 519         else if (IsA(jtnode, FromExpr))
 520         {
 521                 FromExpr   *f = (FromExpr *) jtnode;
 522                 ListCell   *l;
 523
 524                 foreach(l, f->fromlist)
 525                         preprocess_qual_conditions(root, lfirst(l));
 526
 527                 f->quals = preprocess_expression(root, f->quals, EXPRKIND_QUAL);
 528         }
 529         else if (IsA(jtnode, JoinExpr))
 530         {
 531                 JoinExpr   *j = (JoinExpr *) jtnode;
 532
 533                 preprocess_qual_conditions(root, j->larg);
 534                 preprocess_qual_conditions(root, j->rarg);
 535
 536                 j->quals = preprocess_expression(root, j->quals, EXPRKIND_QUAL);
 537         }
 538         else
 539                 elog(ERROR, "unrecognized node type: %d",
 540                          (int) nodeTag(jtnode));
 541 }
 542
 543 /*
 544  * inheritance_planner
 545  *        Generate a plan in the case where the result relation is an
 546  *        inheritance set.
 547  *
 548  * We have to handle this case differently from cases where a source relation
 549  * is an inheritance set. Source inheritance is expanded at the bottom of the
 550  * plan tree (see allpaths.c), but target inheritance has to be expanded at
 551  * the top.  The reason is that for UPDATE, each target relation needs a
 552  * different targetlist matching its own column set.  Also, for both UPDATE
 553  * and DELETE, the executor needs the Append plan node at the top, else it
 554  * can't keep track of which table is the current target table.  Fortunately,
 555  * the UPDATE/DELETE target can never be the nullable side of an outer join,
 556  * so it's OK to generate the plan this way.
 557  *
 558  * Returns a query plan.
 559  */
 560 static Plan *
 561 inheritance_planner(PlannerInfo *root)
 562 {
 563         Query      *parse = root->parse;
 564         int                     parentRTindex = parse->resultRelation;
 565         List       *subplans = NIL;
 566         List       *resultRelations = NIL;
 567         List       *returningLists = NIL;
 568         List       *rtable = NIL;
 569         List       *tlist = NIL;
 570         PlannerInfo subroot;
 571         ListCell   *l;
 572
 573         foreach(l, root->append_rel_list)
 574         {
 575                 AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
 576                 Plan       *subplan;
 577
 578                 /* append_rel_list contains all append rels; ignore others */
 579                 if (appinfo->parent_relid != parentRTindex)
 580                         continue;
 581
 582                 /*
 583                  * Generate modified query with this rel as target.  We have to be
 584                  * prepared to translate varnos in in_info_list as well as in the
 585                  * Query proper.
 586                  */
 587                 memcpy(&subroot, root, sizeof(PlannerInfo));
 588                 subroot.parse = (Query *)
 589                         adjust_appendrel_attrs((Node *) parse,
 590                                                                    appinfo);
 591                 subroot.in_info_list = (List *)
 592                         adjust_appendrel_attrs((Node *) root->in_info_list,
 593                                                                    appinfo);
 594                 subroot.init_plans = NIL;
 595                 /* There shouldn't be any OJ info to translate, as yet */
 596                 Assert(subroot.oj_info_list == NIL);
 597
 598                 /* Generate plan */
 599                 subplan = grouping_planner(&subroot, 0.0 /* retrieve all tuples */ );
 600
 601                 /*
 602                  * If this child rel was excluded by constraint exclusion, exclude it
 603                  * from the plan.
 604                  */
 605                 if (is_dummy_plan(subplan))
 606                         continue;
 607
 608                 /* Save rtable and tlist from first rel for use below */
 609                 if (subplans == NIL)
 610                 {
 611                         rtable = subroot.parse->rtable;
 612                         tlist = subplan->targetlist;
 613                 }
 614
 615                 subplans = lappend(subplans, subplan);
 616
 617                 /* Make sure any initplans from this rel get into the outer list */
 618                 root->init_plans = list_concat(root->init_plans, subroot.init_plans);
 619
 620                 /* Build target-relations list for the executor */
 621                 resultRelations = lappend_int(resultRelations, appinfo->child_relid);
 622
 623                 /* Build list of per-relation RETURNING targetlists */
 624                 if (parse->returningList)
 625                 {
 626                         Assert(list_length(subroot.returningLists) == 1);
 627                         returningLists = list_concat(returningLists,
 628                                                                                  subroot.returningLists);
 629                 }
 630         }
 631
 632         root->resultRelations = resultRelations;
 633         root->returningLists = returningLists;
 634
 635         /* Mark result as unordered (probably unnecessary) */
 636         root->query_pathkeys = NIL;
 637
 638         /*
 639          * If we managed to exclude every child rel, return a dummy plan
 640          */
 641         if (subplans == NIL)
 642                 return (Plan *) make_result(root,
 643                                                                         tlist,
 644                                                                         (Node *) list_make1(makeBoolConst(false,
 645                                                                                                                                           false)),
 646                                                                         NULL);
 647
 648         /*
 649          * Planning might have modified the rangetable, due to changes of the
 650          * Query structures inside subquery RTEs.  We have to ensure that this
 651          * gets propagated back to the master copy.  But can't do this until we
 652          * are done planning, because all the calls to grouping_planner need
 653          * virgin sub-Queries to work from.  (We are effectively assuming that
 654          * sub-Queries will get planned identically each time, or at least that
 655          * the impacts on their rangetables will be the same each time.)
 656          *
 657          * XXX should clean this up someday
 658          */
 659         parse->rtable = rtable;
 660
 661         /* Suppress Append if there's only one surviving child rel */
 662         if (list_length(subplans) == 1)
 663                 return (Plan *) linitial(subplans);
 664
 665         return (Plan *) make_append(subplans, true, tlist);
 666 }
 667
 668 /*--------------------
 669  * grouping_planner
 670  *        Perform planning steps related to grouping, aggregation, etc.
 671  *        This primarily means adding top-level processing to the basic
 672  *        query plan produced by query_planner.
 673  *
 674  * tuple_fraction is the fraction of tuples we expect will be retrieved
 675  *
 676  * tuple_fraction is interpreted as follows:
 677  *        0: expect all tuples to be retrieved (normal case)
 678  *        0 < tuple_fraction < 1: expect the given fraction of tuples available
 679  *              from the plan to be retrieved
 680  *        tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
 681  *              expected to be retrieved (ie, a LIMIT specification)
 682  *
 683  * Returns a query plan.  Also, root->query_pathkeys is returned as the
 684  * actual output ordering of the plan (in pathkey format).
 685  *--------------------
 686  */
 687 static Plan *
 688 grouping_planner(PlannerInfo *root, double tuple_fraction)
 689 {
 690         Query      *parse = root->parse;
 691         List       *tlist = parse->targetList;
 692         int64           offset_est = 0;
 693         int64           count_est = 0;
 694         Plan       *result_plan;
 695         List       *current_pathkeys;
 696         List       *sort_pathkeys;
 697         double          dNumGroups = 0;
 698
 699         /* Tweak caller-supplied tuple_fraction if have LIMIT/OFFSET */
 700         if (parse->limitCount || parse->limitOffset)
 701                 tuple_fraction = preprocess_limit(root, tuple_fraction,
 702                                                                                   &offset_est, &count_est);
 703
 704         if (parse->setOperations)
 705         {
 706                 List       *set_sortclauses;
 707
 708                 /*
 709                  * If there's a top-level ORDER BY, assume we have to fetch all the
 710                  * tuples.      This might seem too simplistic given all the hackery below
 711                  * to possibly avoid the sort ... but a nonzero tuple_fraction is only
 712                  * of use to plan_set_operations() when the setop is UNION ALL, and
 713                  * the result of UNION ALL is always unsorted.
 714                  */
 715                 if (parse->sortClause)
 716                         tuple_fraction = 0.0;
 717
 718                 /*
 719                  * Construct the plan for set operations.  The result will not need
 720                  * any work except perhaps a top-level sort and/or LIMIT.
 721                  */
 722                 result_plan = plan_set_operations(root, tuple_fraction,
 723                                                                                   &set_sortclauses);
 724
 725                 /*
 726                  * Calculate pathkeys representing the sort order (if any) of the set
 727                  * operation's result.  We have to do this before overwriting the sort
 728                  * key information...
 729                  */
 730                 current_pathkeys = make_pathkeys_for_sortclauses(root,
 731                                                                                                                  set_sortclauses,
 732                                                                                                         result_plan->targetlist,
 733                                                                                                                  true);
 734
 735                 /*
 736                  * We should not need to call preprocess_targetlist, since we must be
 737                  * in a SELECT query node.      Instead, use the targetlist returned by
 738                  * plan_set_operations (since this tells whether it returned any
 739                  * resjunk columns!), and transfer any sort key information from the
 740                  * original tlist.
 741                  */
 742                 Assert(parse->commandType == CMD_SELECT);
 743
 744                 tlist = postprocess_setop_tlist(result_plan->targetlist, tlist);
 745
 746                 /*
 747                  * Can't handle FOR UPDATE/SHARE here (parser should have checked
 748                  * already, but let's make sure).
 749                  */
 750                 if (parse->rowMarks)
 751                         ereport(ERROR,
 752                                         (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
 753                                          errmsg("SELECT FOR UPDATE/SHARE is not allowed with UNION/INTERSECT/EXCEPT")));
 754
 755                 /*
 756                  * Calculate pathkeys that represent result ordering requirements
 757                  */
 758                 sort_pathkeys = make_pathkeys_for_sortclauses(root,
 759                                                                                                           parse->sortClause,
 760                                                                                                           tlist,
 761                                                                                                           true);
 762         }
 763         else
 764         {
 765                 /* No set operations, do regular planning */
 766                 List       *sub_tlist;
 767                 List       *group_pathkeys;
 768                 AttrNumber *groupColIdx = NULL;
 769                 Oid                *groupOperators = NULL;
 770                 bool            need_tlist_eval = true;
 771                 QualCost        tlist_cost;
 772                 Path       *cheapest_path;
 773                 Path       *sorted_path;
 774                 Path       *best_path;
 775                 long            numGroups = 0;
 776                 AggClauseCounts agg_counts;
 777                 int                     numGroupCols = list_length(parse->groupClause);
 778                 bool            use_hashed_grouping = false;
 779
 780                 MemSet(&agg_counts, 0, sizeof(AggClauseCounts));
 781
 782                 /* Preprocess targetlist */
 783                 tlist = preprocess_targetlist(root, tlist);
 784
 785                 /*
 786                  * Generate appropriate target list for subplan; may be different from
 787                  * tlist if grouping or aggregation is needed.
 788                  */
 789                 sub_tlist = make_subplanTargetList(root, tlist,
 790                                                                                    &groupColIdx, &need_tlist_eval);
 791
 792                 /*
 793                  * Calculate pathkeys that represent grouping/ordering requirements.
 794                  * Stash them in PlannerInfo so that query_planner can canonicalize
 795                  * them after EquivalenceClasses have been formed.
 796                  */
 797                 root->group_pathkeys =
 798                         make_pathkeys_for_sortclauses(root,
 799                                                                                   parse->groupClause,
 800                                                                                   tlist,
 801                                                                                   false);
 802                 root->sort_pathkeys =
 803                         make_pathkeys_for_sortclauses(root,
 804                                                                                   parse->sortClause,
 805                                                                                   tlist,
 806                                                                                   false);
 807
 808                 /*
 809                  * Will need actual number of aggregates for estimating costs.
 810                  *
 811                  * Note: we do not attempt to detect duplicate aggregates here; a
 812                  * somewhat-overestimated count is okay for our present purposes.
 813                  *
 814                  * Note: think not that we can turn off hasAggs if we find no aggs. It
 815                  * is possible for constant-expression simplification to remove all
 816                  * explicit references to aggs, but we still have to follow the
 817                  * aggregate semantics (eg, producing only one output row).
 818                  */
 819                 if (parse->hasAggs)
 820                 {
 821                         count_agg_clauses((Node *) tlist, &agg_counts);
 822                         count_agg_clauses(parse->havingQual, &agg_counts);
 823                 }
 824
 825                 /*
 826                  * Figure out whether we need a sorted result from query_planner.
 827                  *
 828                  * If we have a GROUP BY clause, then we want a result sorted properly
 829                  * for grouping.  Otherwise, if there is an ORDER BY clause, we want
 830                  * to sort by the ORDER BY clause.      (Note: if we have both, and ORDER
 831                  * BY is a superset of GROUP BY, it would be tempting to request sort
 832                  * by ORDER BY --- but that might just leave us failing to exploit an
 833                  * available sort order at all. Needs more thought...)
 834                  */
 835                 if (parse->groupClause)
 836                         root->query_pathkeys = root->group_pathkeys;
 837                 else if (parse->sortClause)
 838                         root->query_pathkeys = root->sort_pathkeys;
 839                 else
 840                         root->query_pathkeys = NIL;
 841
 842                 /*
 843                  * Generate the best unsorted and presorted paths for this Query (but
 844                  * note there may not be any presorted path).  query_planner will also
 845                  * estimate the number of groups in the query, and canonicalize all
 846                  * the pathkeys.
 847                  */
 848                 query_planner(root, sub_tlist, tuple_fraction,
 849                                           &cheapest_path, &sorted_path, &dNumGroups);
 850
 851                 group_pathkeys = root->group_pathkeys;
 852                 sort_pathkeys = root->sort_pathkeys;
 853
 854                 /*
 855                  * If grouping, extract the grouping operators and decide whether we
 856                  * want to use hashed grouping.
 857                  */
 858                 if (parse->groupClause)
 859                 {
 860                         groupOperators = extract_grouping_ops(parse->groupClause);
 861                         use_hashed_grouping =
 862                                 choose_hashed_grouping(root, tuple_fraction,
 863                                                                            cheapest_path, sorted_path,
 864                                                                            groupOperators, dNumGroups,
 865                                                                            &agg_counts);
 866
 867                         /* Also convert # groups to long int --- but 'ware overflow! */
 868                         numGroups = (long) Min(dNumGroups, (double) LONG_MAX);
 869                 }
 870
 871                 /*
 872                  * Select the best path.  If we are doing hashed grouping, we will
 873                  * always read all the input tuples, so use the cheapest-total path.
 874                  * Otherwise, trust query_planner's decision about which to use.
 875                  */
 876                 if (use_hashed_grouping || !sorted_path)
 877                         best_path = cheapest_path;
 878                 else
 879                         best_path = sorted_path;
 880
 881                 /*
 882                  * Check to see if it's possible to optimize MIN/MAX aggregates. If
 883                  * so, we will forget all the work we did so far to choose a "regular"
 884                  * path ... but we had to do it anyway to be able to tell which way is
 885                  * cheaper.
 886                  */
 887                 result_plan = optimize_minmax_aggregates(root,
 888                                                                                                  tlist,
 889                                                                                                  best_path);
 890                 if (result_plan != NULL)
 891                 {
 892                         /*
 893                          * optimize_minmax_aggregates generated the full plan, with the
 894                          * right tlist, and it has no sort order.
 895                          */
 896                         current_pathkeys = NIL;
 897                 }
 898                 else
 899                 {
 900                         /*
 901                          * Normal case --- create a plan according to query_planner's
 902                          * results.
 903                          */
 904                         result_plan = create_plan(root, best_path);
 905                         current_pathkeys = best_path->pathkeys;
 906
 907                         /*
 908                          * create_plan() returns a plan with just a "flat" tlist of
 909                          * required Vars.  Usually we need to insert the sub_tlist as the
 910                          * tlist of the top plan node.  However, we can skip that if we
 911                          * determined that whatever query_planner chose to return will be
 912                          * good enough.
 913                          */
 914                         if (need_tlist_eval)
 915                         {
 916                                 /*
 917                                  * If the top-level plan node is one that cannot do expression
 918                                  * evaluation, we must insert a Result node to project the
 919                                  * desired tlist.
 920                                  */
 921                                 if (!is_projection_capable_plan(result_plan))
 922                                 {
 923                                         result_plan = (Plan *) make_result(root,
 924                                                                                                            sub_tlist,
 925                                                                                                            NULL,
 926                                                                                                            result_plan);
 927                                 }
 928                                 else
 929                                 {
 930                                         /*
 931                                          * Otherwise, just replace the subplan's flat tlist with
 932                                          * the desired tlist.
 933                                          */
 934                                         result_plan->targetlist = sub_tlist;
 935                                 }
 936
 937                                 /*
 938                                  * Also, account for the cost of evaluation of the sub_tlist.
 939                                  *
 940                                  * Up to now, we have only been dealing with "flat" tlists,
 941                                  * containing just Vars.  So their evaluation cost is zero
 942                                  * according to the model used by cost_qual_eval() (or if you
 943                                  * prefer, the cost is factored into cpu_tuple_cost).  Thus we
 944                                  * can avoid accounting for tlist cost throughout
 945                                  * query_planner() and subroutines.  But now we've inserted a
 946                                  * tlist that might contain actual operators, sub-selects, etc
 947                                  * --- so we'd better account for its cost.
 948                                  *
 949                                  * Below this point, any tlist eval cost for added-on nodes
 950                                  * should be accounted for as we create those nodes.
 951                                  * Presently, of the node types we can add on, only Agg and
 952                                  * Group project new tlists (the rest just copy their input
 953                                  * tuples) --- so make_agg() and make_group() are responsible
 954                                  * for computing the added cost.
 955                                  */
 956                                 cost_qual_eval(&tlist_cost, sub_tlist, root);
 957                                 result_plan->startup_cost += tlist_cost.startup;
 958                                 result_plan->total_cost += tlist_cost.startup +
 959                                         tlist_cost.per_tuple * result_plan->plan_rows;
 960                         }
 961                         else
 962                         {
 963                                 /*
 964                                  * Since we're using query_planner's tlist and not the one
 965                                  * make_subplanTargetList calculated, we have to refigure any
 966                                  * grouping-column indexes make_subplanTargetList computed.
 967                                  */
 968                                 locate_grouping_columns(root, tlist, result_plan->targetlist,
 969                                                                                 groupColIdx);
 970                         }
 971
 972                         /*
 973                          * Insert AGG or GROUP node if needed, plus an explicit sort step
 974                          * if necessary.
 975                          *
 976                          * HAVING clause, if any, becomes qual of the Agg or Group node.
 977                          */
 978                         if (use_hashed_grouping)
 979                         {
 980                                 /* Hashed aggregate plan --- no sort needed */
 981                                 result_plan = (Plan *) make_agg(root,
 982                                                                                                 tlist,
 983                                                                                                 (List *) parse->havingQual,
 984                                                                                                 AGG_HASHED,
 985                                                                                                 numGroupCols,
 986                                                                                                 groupColIdx,
 987                                                                                                 groupOperators,
 988                                                                                                 numGroups,
 989                                                                                                 agg_counts.numAggs,
 990                                                                                                 result_plan);
 991                                 /* Hashed aggregation produces randomly-ordered results */
 992                                 current_pathkeys = NIL;
 993                         }
 994                         else if (parse->hasAggs)
 995                         {
 996                                 /* Plain aggregate plan --- sort if needed */
 997                                 AggStrategy aggstrategy;
 998
 999                                 if (parse->groupClause)
1000                                 {
1001                                         if (!pathkeys_contained_in(group_pathkeys,
1002                                                                                            current_pathkeys))
1003                                         {
1004                                                 result_plan = (Plan *)
1005                                                         make_sort_from_groupcols(root,
1006                                                                                                          parse->groupClause,
1007                                                                                                          groupColIdx,
1008                                                                                                          result_plan);
1009                                                 current_pathkeys = group_pathkeys;
1010                                         }
1011                                         aggstrategy = AGG_SORTED;
1012
1013                                         /*
1014                                          * The AGG node will not change the sort ordering of its
1015                                          * groups, so current_pathkeys describes the result too.
1016                                          */
1017                                 }
1018                                 else
1019                                 {
1020                                         aggstrategy = AGG_PLAIN;
1021                                         /* Result will be only one row anyway; no sort order */
1022                                         current_pathkeys = NIL;
1023                                 }
1024
1025                                 result_plan = (Plan *) make_agg(root,
1026                                                                                                 tlist,
1027                                                                                                 (List *) parse->havingQual,
1028                                                                                                 aggstrategy,
1029                                                                                                 numGroupCols,
1030                                                                                                 groupColIdx,
1031                                                                                                 groupOperators,
1032                                                                                                 numGroups,
1033                                                                                                 agg_counts.numAggs,
1034                                                                                                 result_plan);
1035                         }
1036                         else if (parse->groupClause)
1037                         {
1038                                 /*
1039                                  * GROUP BY without aggregation, so insert a group node (plus
1040                                  * the appropriate sort node, if necessary).
1041                                  *
1042                                  * Add an explicit sort if we couldn't make the path come out
1043                                  * the way the GROUP node needs it.
1044                                  */
1045                                 if (!pathkeys_contained_in(group_pathkeys, current_pathkeys))
1046                                 {
1047                                         result_plan = (Plan *)
1048                                                 make_sort_from_groupcols(root,
1049                                                                                                  parse->groupClause,
1050                                                                                                  groupColIdx,
1051                                                                                                  result_plan);
1052                                         current_pathkeys = group_pathkeys;
1053                                 }
1054
1055                                 result_plan = (Plan *) make_group(root,
1056                                                                                                   tlist,
1057                                                                                                   (List *) parse->havingQual,
1058                                                                                                   numGroupCols,
1059                                                                                                   groupColIdx,
1060                                                                                                   groupOperators,
1061                                                                                                   dNumGroups,
1062                                                                                                   result_plan);
1063                                 /* The Group node won't change sort ordering */
1064                         }
1065                         else if (root->hasHavingQual)
1066                         {
1067                                 /*
1068                                  * No aggregates, and no GROUP BY, but we have a HAVING qual.
1069                                  * This is a degenerate case in which we are supposed to emit
1070                                  * either 0 or 1 row depending on whether HAVING succeeds.
1071                                  * Furthermore, there cannot be any variables in either HAVING
1072                                  * or the targetlist, so we actually do not need the FROM
1073                                  * table at all!  We can just throw away the plan-so-far and
1074                                  * generate a Result node.      This is a sufficiently unusual
1075                                  * corner case that it's not worth contorting the structure of
1076                                  * this routine to avoid having to generate the plan in the
1077                                  * first place.
1078                                  */
1079                                 result_plan = (Plan *) make_result(root,
1080                                                                                                    tlist,
1081                                                                                                    parse->havingQual,
1082                                                                                                    NULL);
1083                         }
1084                 }                                               /* end of non-minmax-aggregate case */
1085         }                                                       /* end of if (setOperations) */
1086
1087         /*
1088          * If we were not able to make the plan come out in the right order, add
1089          * an explicit sort step.
1090          */
1091         if (parse->sortClause)
1092         {
1093                 if (!pathkeys_contained_in(sort_pathkeys, current_pathkeys))
1094                 {
1095                         result_plan = (Plan *) make_sort_from_pathkeys(root,
1096                                                                                                                    result_plan,
1097                                                                                                                    sort_pathkeys);
1098                         current_pathkeys = sort_pathkeys;
1099                 }
1100         }
1101
1102         /*
1103          * If there is a DISTINCT clause, add the UNIQUE node.
1104          */
1105         if (parse->distinctClause)
1106         {
1107                 result_plan = (Plan *) make_unique(result_plan, parse->distinctClause);
1108
1109                 /*
1110                  * If there was grouping or aggregation, leave plan_rows as-is (ie,
1111                  * assume the result was already mostly unique).  If not, use the
1112                  * number of distinct-groups calculated by query_planner.
1113                  */
1114                 if (!parse->groupClause && !root->hasHavingQual && !parse->hasAggs)
1115                         result_plan->plan_rows = dNumGroups;
1116         }
1117
1118         /*
1119          * Finally, if there is a LIMIT/OFFSET clause, add the LIMIT node.
1120          */
1121         if (parse->limitCount || parse->limitOffset)
1122         {
1123                 result_plan = (Plan *) make_limit(result_plan,
1124                                                                                   parse->limitOffset,
1125                                                                                   parse->limitCount,
1126                                                                                   offset_est,
1127                                                                                   count_est);
1128         }
1129
1130         /*
1131          * Deal with the RETURNING clause if any.  It's convenient to pass the
1132          * returningList through setrefs.c now rather than at top level (if we
1133          * waited, handling inherited UPDATE/DELETE would be much harder).
1134          */
1135         if (parse->returningList)
1136         {
1137                 List       *rlist;
1138
1139                 Assert(parse->resultRelation);
1140                 rlist = set_returning_clause_references(parse->returningList,
1141                                                                                                 result_plan,
1142                                                                                                 parse->resultRelation);
1143                 root->returningLists = list_make1(rlist);
1144         }
1145         else
1146                 root->returningLists = NIL;
1147
1148         /* Compute result-relations list if needed */
1149         if (parse->resultRelation)
1150                 root->resultRelations = list_make1_int(parse->resultRelation);
1151         else
1152                 root->resultRelations = NIL;
1153
1154         /*
1155          * Return the actual output ordering in query_pathkeys for possible use by
1156          * an outer query level.
1157          */
1158         root->query_pathkeys = current_pathkeys;
1159
1160         return result_plan;
1161 }
1162
1163 /*
1164  * Detect whether a plan node is a "dummy" plan created when a relation
1165  * is deemed not to need scanning due to constraint exclusion.
1166  *
1167  * Currently, such dummy plans are Result nodes with constant FALSE
1168  * filter quals.
1169  */
1170 static bool
1171 is_dummy_plan(Plan *plan)
1172 {
1173         if (IsA(plan, Result))
1174         {
1175                 List       *rcqual = (List *) ((Result *) plan)->resconstantqual;
1176
1177                 if (list_length(rcqual) == 1)
1178                 {
1179                         Const      *constqual = (Const *) linitial(rcqual);
1180
1181                         if (constqual && IsA(constqual, Const))
1182                         {
1183                                 if (!constqual->constisnull &&
1184                                         !DatumGetBool(constqual->constvalue))
1185                                         return true;
1186                         }
1187                 }
1188         }
1189         return false;
1190 }
1191
1192 /*
1193  * preprocess_limit - do pre-estimation for LIMIT and/or OFFSET clauses
1194  *
1195  * We try to estimate the values of the LIMIT/OFFSET clauses, and pass the
1196  * results back in *count_est and *offset_est.  These variables are set to
1197  * 0 if the corresponding clause is not present, and -1 if it's present
1198  * but we couldn't estimate the value for it.  (The "0" convention is OK
1199  * for OFFSET but a little bit bogus for LIMIT: effectively we estimate
1200  * LIMIT 0 as though it were LIMIT 1.  But this is in line with the planner's
1201  * usual practice of never estimating less than one row.)  These values will
1202  * be passed to make_limit, which see if you change this code.
1203  *
1204  * The return value is the suitably adjusted tuple_fraction to use for
1205  * planning the query.  This adjustment is not overridable, since it reflects
1206  * plan actions that grouping_planner() will certainly take, not assumptions
1207  * about context.
1208  */
1209 static double
1210 preprocess_limit(PlannerInfo *root, double tuple_fraction,
1211                                  int64 *offset_est, int64 *count_est)
1212 {
1213         Query      *parse = root->parse;
1214         Node       *est;
1215         double          limit_fraction;
1216
1217         /* Should not be called unless LIMIT or OFFSET */
1218         Assert(parse->limitCount || parse->limitOffset);
1219
1220         /*
1221          * Try to obtain the clause values.  We use estimate_expression_value
1222          * primarily because it can sometimes do something useful with Params.
1223          */
1224         if (parse->limitCount)
1225         {
1226                 est = estimate_expression_value(root, parse->limitCount);
1227                 if (est && IsA(est, Const))
1228                 {
1229                         if (((Const *) est)->constisnull)
1230                         {
1231                                 /* NULL indicates LIMIT ALL, ie, no limit */
1232                                 *count_est = 0; /* treat as not present */
1233                         }
1234                         else
1235                         {
1236                                 *count_est = DatumGetInt64(((Const *) est)->constvalue);
1237                                 if (*count_est <= 0)
1238                                         *count_est = 1;         /* force to at least 1 */
1239                         }
1240                 }
1241                 else
1242                         *count_est = -1;        /* can't estimate */
1243         }
1244         else
1245                 *count_est = 0;                 /* not present */
1246
1247         if (parse->limitOffset)
1248         {
1249                 est = estimate_expression_value(root, parse->limitOffset);
1250                 if (est && IsA(est, Const))
1251                 {
1252                         if (((Const *) est)->constisnull)
1253                         {
1254                                 /* Treat NULL as no offset; the executor will too */
1255                                 *offset_est = 0;        /* treat as not present */
1256                         }
1257                         else
1258                         {
1259                                 *offset_est = DatumGetInt64(((Const *) est)->constvalue);
1260                                 if (*offset_est < 0)
1261                                         *offset_est = 0;        /* less than 0 is same as 0 */
1262                         }
1263                 }
1264                 else
1265                         *offset_est = -1;       /* can't estimate */
1266         }
1267         else
1268                 *offset_est = 0;                /* not present */
1269
1270         if (*count_est != 0)
1271         {
1272                 /*
1273                  * A LIMIT clause limits the absolute number of tuples returned.
1274                  * However, if it's not a constant LIMIT then we have to guess; for
1275                  * lack of a better idea, assume 10% of the plan's result is wanted.
1276                  */
1277                 if (*count_est < 0 || *offset_est < 0)
1278                 {
1279                         /* LIMIT or OFFSET is an expression ... punt ... */
1280                         limit_fraction = 0.10;
1281                 }
1282                 else
1283                 {
1284                         /* LIMIT (plus OFFSET, if any) is max number of tuples needed */
1285                         limit_fraction = (double) *count_est + (double) *offset_est;
1286                 }
1287
1288                 /*
1289                  * If we have absolute limits from both caller and LIMIT, use the
1290                  * smaller value; likewise if they are both fractional.  If one is
1291                  * fractional and the other absolute, we can't easily determine which
1292                  * is smaller, but we use the heuristic that the absolute will usually
1293                  * be smaller.
1294                  */
1295                 if (tuple_fraction >= 1.0)
1296                 {
1297                         if (limit_fraction >= 1.0)
1298                         {
1299                                 /* both absolute */
1300                                 tuple_fraction = Min(tuple_fraction, limit_fraction);
1301                         }
1302                         else
1303                         {
1304                                 /* caller absolute, limit fractional; use caller's value */
1305                         }
1306                 }
1307                 else if (tuple_fraction > 0.0)
1308                 {
1309                         if (limit_fraction >= 1.0)
1310                         {
1311                                 /* caller fractional, limit absolute; use limit */
1312                                 tuple_fraction = limit_fraction;
1313                         }
1314                         else
1315                         {
1316                                 /* both fractional */
1317                                 tuple_fraction = Min(tuple_fraction, limit_fraction);
1318                         }
1319                 }
1320                 else
1321                 {
1322                         /* no info from caller, just use limit */
1323                         tuple_fraction = limit_fraction;
1324                 }
1325         }
1326         else if (*offset_est != 0 && tuple_fraction > 0.0)
1327         {
1328                 /*
1329                  * We have an OFFSET but no LIMIT.      This acts entirely differently
1330                  * from the LIMIT case: here, we need to increase rather than decrease
1331                  * the caller's tuple_fraction, because the OFFSET acts to cause more
1332                  * tuples to be fetched instead of fewer.  This only matters if we got
1333                  * a tuple_fraction > 0, however.
1334                  *
1335                  * As above, use 10% if OFFSET is present but unestimatable.
1336                  */
1337                 if (*offset_est < 0)
1338                         limit_fraction = 0.10;
1339                 else
1340                         limit_fraction = (double) *offset_est;
1341
1342                 /*
1343                  * If we have absolute counts from both caller and OFFSET, add them
1344                  * together; likewise if they are both fractional.      If one is
1345                  * fractional and the other absolute, we want to take the larger, and
1346                  * we heuristically assume that's the fractional one.
1347                  */
1348                 if (tuple_fraction >= 1.0)
1349                 {
1350                         if (limit_fraction >= 1.0)
1351                         {
1352                                 /* both absolute, so add them together */
1353                                 tuple_fraction += limit_fraction;
1354                         }
1355                         else
1356                         {
1357                                 /* caller absolute, limit fractional; use limit */
1358                                 tuple_fraction = limit_fraction;
1359                         }
1360                 }
1361                 else
1362                 {
1363                         if (limit_fraction >= 1.0)
1364                         {
1365                                 /* caller fractional, limit absolute; use caller's value */
1366                         }
1367                         else
1368                         {
1369                                 /* both fractional, so add them together */
1370                                 tuple_fraction += limit_fraction;
1371                                 if (tuple_fraction >= 1.0)
1372                                         tuple_fraction = 0.0;           /* assume fetch all */
1373                         }
1374                 }
1375         }
1376
1377         return tuple_fraction;
1378 }
1379
1380 /*
1381  * extract_grouping_ops - make an array of the equality operator OIDs
1382  *              for the GROUP BY clause
1383  */
1384 static Oid *
1385 extract_grouping_ops(List *groupClause)
1386 {
1387         int                     numCols = list_length(groupClause);
1388         int                     colno = 0;
1389         Oid                *groupOperators;
1390         ListCell   *glitem;
1391
1392         groupOperators = (Oid *) palloc(sizeof(Oid) * numCols);
1393
1394         foreach(glitem, groupClause)
1395         {
1396                 GroupClause *groupcl = (GroupClause *) lfirst(glitem);
1397
1398                 groupOperators[colno] = get_equality_op_for_ordering_op(groupcl->sortop);
1399                 if (!OidIsValid(groupOperators[colno]))         /* shouldn't happen */
1400                         elog(ERROR, "could not find equality operator for ordering operator %u",
1401                                  groupcl->sortop);
1402                 colno++;
1403         }
1404
1405         return groupOperators;
1406 }
1407
1408 /*
1409  * choose_hashed_grouping - should we use hashed grouping?
1410  */
1411 static bool
1412 choose_hashed_grouping(PlannerInfo *root, double tuple_fraction,
1413                                            Path *cheapest_path, Path *sorted_path,
1414                                            Oid *groupOperators, double dNumGroups,
1415                                            AggClauseCounts *agg_counts)
1416 {
1417         int                     numGroupCols = list_length(root->parse->groupClause);
1418         double          cheapest_path_rows;
1419         int                     cheapest_path_width;
1420         Size            hashentrysize;
1421         List       *current_pathkeys;
1422         Path            hashed_p;
1423         Path            sorted_p;
1424         int                     i;
1425
1426         /*
1427          * Check can't-do-it conditions, including whether the grouping operators
1428          * are hashjoinable.  (We assume hashing is OK if they are marked
1429          * oprcanhash.  If there isn't actually a supporting hash function,
1430          * the executor will complain at runtime.)
1431          *
1432          * Executor doesn't support hashed aggregation with DISTINCT aggregates.
1433          * (Doing so would imply storing *all* the input values in the hash table,
1434          * which seems like a certain loser.)
1435          */
1436         if (!enable_hashagg)
1437                 return false;
1438         if (agg_counts->numDistinctAggs != 0)
1439                 return false;
1440         for (i = 0; i < numGroupCols; i++)
1441         {
1442                 if (!op_hashjoinable(groupOperators[i]))
1443                         return false;
1444         }
1445
1446         /*
1447          * Don't do it if it doesn't look like the hashtable will fit into
1448          * work_mem.
1449          *
1450          * Beware here of the possibility that cheapest_path->parent is NULL. This
1451          * could happen if user does something silly like SELECT 'foo' GROUP BY 1;
1452          */
1453         if (cheapest_path->parent)
1454         {
1455                 cheapest_path_rows = cheapest_path->parent->rows;
1456                 cheapest_path_width = cheapest_path->parent->width;
1457         }
1458         else
1459         {
1460                 cheapest_path_rows = 1; /* assume non-set result */
1461                 cheapest_path_width = 100;              /* arbitrary */
1462         }
1463
1464         /* Estimate per-hash-entry space at tuple width... */
1465         hashentrysize = MAXALIGN(cheapest_path_width) + MAXALIGN(sizeof(MinimalTupleData));
1466         /* plus space for pass-by-ref transition values... */
1467         hashentrysize += agg_counts->transitionSpace;
1468         /* plus the per-hash-entry overhead */
1469         hashentrysize += hash_agg_entry_size(agg_counts->numAggs);
1470
1471         if (hashentrysize * dNumGroups > work_mem * 1024L)
1472                 return false;
1473
1474         /*
1475          * See if the estimated cost is no more than doing it the other way. While
1476          * avoiding the need for sorted input is usually a win, the fact that the
1477          * output won't be sorted may be a loss; so we need to do an actual cost
1478          * comparison.
1479          *
1480          * We need to consider cheapest_path + hashagg [+ final sort] versus
1481          * either cheapest_path [+ sort] + group or agg [+ final sort] or
1482          * presorted_path + group or agg [+ final sort] where brackets indicate a
1483          * step that may not be needed. We assume query_planner() will have
1484          * returned a presorted path only if it's a winner compared to
1485          * cheapest_path for this purpose.
1486          *
1487          * These path variables are dummies that just hold cost fields; we don't
1488          * make actual Paths for these steps.
1489          */
1490         cost_agg(&hashed_p, root, AGG_HASHED, agg_counts->numAggs,
1491                          numGroupCols, dNumGroups,
1492                          cheapest_path->startup_cost, cheapest_path->total_cost,
1493                          cheapest_path_rows);
1494         /* Result of hashed agg is always unsorted */
1495         if (root->sort_pathkeys)
1496                 cost_sort(&hashed_p, root, root->sort_pathkeys, hashed_p.total_cost,
1497                                   dNumGroups, cheapest_path_width);
1498
1499         if (sorted_path)
1500         {
1501                 sorted_p.startup_cost = sorted_path->startup_cost;
1502                 sorted_p.total_cost = sorted_path->total_cost;
1503                 current_pathkeys = sorted_path->pathkeys;
1504         }
1505         else
1506         {
1507                 sorted_p.startup_cost = cheapest_path->startup_cost;
1508                 sorted_p.total_cost = cheapest_path->total_cost;
1509                 current_pathkeys = cheapest_path->pathkeys;
1510         }
1511         if (!pathkeys_contained_in(root->group_pathkeys, current_pathkeys))
1512         {
1513                 cost_sort(&sorted_p, root, root->group_pathkeys, sorted_p.total_cost,
1514                                   cheapest_path_rows, cheapest_path_width);
1515                 current_pathkeys = root->group_pathkeys;
1516         }
1517
1518         if (root->parse->hasAggs)
1519                 cost_agg(&sorted_p, root, AGG_SORTED, agg_counts->numAggs,
1520                                  numGroupCols, dNumGroups,
1521                                  sorted_p.startup_cost, sorted_p.total_cost,
1522                                  cheapest_path_rows);
1523         else
1524                 cost_group(&sorted_p, root, numGroupCols, dNumGroups,
1525                                    sorted_p.startup_cost, sorted_p.total_cost,
1526                                    cheapest_path_rows);
1527         /* The Agg or Group node will preserve ordering */
1528         if (root->sort_pathkeys &&
1529                 !pathkeys_contained_in(root->sort_pathkeys, current_pathkeys))
1530                 cost_sort(&sorted_p, root, root->sort_pathkeys, sorted_p.total_cost,
1531                                   dNumGroups, cheapest_path_width);
1532
1533         /*
1534          * Now make the decision using the top-level tuple fraction.  First we
1535          * have to convert an absolute count (LIMIT) into fractional form.
1536          */
1537         if (tuple_fraction >= 1.0)
1538                 tuple_fraction /= dNumGroups;
1539
1540         if (compare_fractional_path_costs(&hashed_p, &sorted_p,
1541                                                                           tuple_fraction) < 0)
1542         {
1543                 /* Hashed is cheaper, so use it */
1544                 return true;
1545         }
1546         return false;
1547 }
1548
1549 /*---------------
1550  * make_subplanTargetList
1551  *        Generate appropriate target list when grouping is required.
1552  *
1553  * When grouping_planner inserts Aggregate, Group, or Result plan nodes
1554  * above the result of query_planner, we typically want to pass a different
1555  * target list to query_planner than the outer plan nodes should have.
1556  * This routine generates the correct target list for the subplan.
1557  *
1558  * The initial target list passed from the parser already contains entries
1559  * for all ORDER BY and GROUP BY expressions, but it will not have entries
1560  * for variables used only in HAVING clauses; so we need to add those
1561  * variables to the subplan target list.  Also, we flatten all expressions
1562  * except GROUP BY items into their component variables; the other expressions
1563  * will be computed by the inserted nodes rather than by the subplan.
1564  * For example, given a query like
1565  *              SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
1566  * we want to pass this targetlist to the subplan:
1567  *              a,b,c,d,a+b
1568  * where the a+b target will be used by the Sort/Group steps, and the
1569  * other targets will be used for computing the final results.  (In the
1570  * above example we could theoretically suppress the a and b targets and
1571  * pass down only c,d,a+b, but it's not really worth the trouble to
1572  * eliminate simple var references from the subplan.  We will avoid doing
1573  * the extra computation to recompute a+b at the outer level; see
1574  * fix_upper_expr() in setrefs.c.)
1575  *
1576  * If we are grouping or aggregating, *and* there are no non-Var grouping
1577  * expressions, then the returned tlist is effectively dummy; we do not
1578  * need to force it to be evaluated, because all the Vars it contains
1579  * should be present in the output of query_planner anyway.
1580  *
1581  * 'tlist' is the query's target list.
1582  * 'groupColIdx' receives an array of column numbers for the GROUP BY
1583  *                      expressions (if there are any) in the subplan's target list.
1584  * 'need_tlist_eval' is set true if we really need to evaluate the
1585  *                      result tlist.
1586  *
1587  * The result is the targetlist to be passed to the subplan.
1588  *---------------
1589  */
1590 static List *
1591 make_subplanTargetList(PlannerInfo *root,
1592                                            List *tlist,
1593                                            AttrNumber **groupColIdx,
1594                                            bool *need_tlist_eval)
1595 {
1596         Query      *parse = root->parse;
1597         List       *sub_tlist;
1598         List       *extravars;
1599         int                     numCols;
1600
1601         *groupColIdx = NULL;
1602
1603         /*
1604          * If we're not grouping or aggregating, there's nothing to do here;
1605          * query_planner should receive the unmodified target list.
1606          */
1607         if (!parse->hasAggs && !parse->groupClause && !root->hasHavingQual)
1608         {
1609                 *need_tlist_eval = true;
1610                 return tlist;
1611         }
1612
1613         /*
1614          * Otherwise, start with a "flattened" tlist (having just the vars
1615          * mentioned in the targetlist and HAVING qual --- but not upper- level
1616          * Vars; they will be replaced by Params later on).
1617          */
1618         sub_tlist = flatten_tlist(tlist);
1619         extravars = pull_var_clause(parse->havingQual, false);
1620         sub_tlist = add_to_flat_tlist(sub_tlist, extravars);
1621         list_free(extravars);
1622         *need_tlist_eval = false;       /* only eval if not flat tlist */
1623
1624         /*
1625          * If grouping, create sub_tlist entries for all GROUP BY expressions
1626          * (GROUP BY items that are simple Vars should be in the list already),
1627          * and make an array showing where the group columns are in the sub_tlist.
1628          */
1629         numCols = list_length(parse->groupClause);
1630         if (numCols > 0)
1631         {
1632                 int                     keyno = 0;
1633                 AttrNumber *grpColIdx;
1634                 ListCell   *gl;
1635
1636                 grpColIdx = (AttrNumber *) palloc(sizeof(AttrNumber) * numCols);
1637                 *groupColIdx = grpColIdx;
1638
1639                 foreach(gl, parse->groupClause)
1640                 {
1641                         GroupClause *grpcl = (GroupClause *) lfirst(gl);
1642                         Node       *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1643                         TargetEntry *te = NULL;
1644                         ListCell   *sl;
1645
1646                         /* Find or make a matching sub_tlist entry */
1647                         foreach(sl, sub_tlist)
1648                         {
1649                                 te = (TargetEntry *) lfirst(sl);
1650                                 if (equal(groupexpr, te->expr))
1651                                         break;
1652                         }
1653                         if (!sl)
1654                         {
1655                                 te = makeTargetEntry((Expr *) groupexpr,
1656                                                                          list_length(sub_tlist) + 1,
1657                                                                          NULL,
1658                                                                          false);
1659                                 sub_tlist = lappend(sub_tlist, te);
1660                                 *need_tlist_eval = true;                /* it's not flat anymore */
1661                         }
1662
1663                         /* and save its resno */
1664                         grpColIdx[keyno++] = te->resno;
1665                 }
1666         }
1667
1668         return sub_tlist;
1669 }
1670
1671 /*
1672  * locate_grouping_columns
1673  *              Locate grouping columns in the tlist chosen by query_planner.
1674  *
1675  * This is only needed if we don't use the sub_tlist chosen by
1676  * make_subplanTargetList.      We have to forget the column indexes found
1677  * by that routine and re-locate the grouping vars in the real sub_tlist.
1678  */
1679 static void
1680 locate_grouping_columns(PlannerInfo *root,
1681                                                 List *tlist,
1682                                                 List *sub_tlist,
1683                                                 AttrNumber *groupColIdx)
1684 {
1685         int                     keyno = 0;
1686         ListCell   *gl;
1687
1688         /*
1689          * No work unless grouping.
1690          */
1691         if (!root->parse->groupClause)
1692         {
1693                 Assert(groupColIdx == NULL);
1694                 return;
1695         }
1696         Assert(groupColIdx != NULL);
1697
1698         foreach(gl, root->parse->groupClause)
1699         {
1700                 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1701                 Node       *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1702                 TargetEntry *te = NULL;
1703                 ListCell   *sl;
1704
1705                 foreach(sl, sub_tlist)
1706                 {
1707                         te = (TargetEntry *) lfirst(sl);
1708                         if (equal(groupexpr, te->expr))
1709                                 break;
1710                 }
1711                 if (!sl)
1712                         elog(ERROR, "failed to locate grouping columns");
1713
1714                 groupColIdx[keyno++] = te->resno;
1715         }
1716 }
1717
1718 /*
1719  * postprocess_setop_tlist
1720  *        Fix up targetlist returned by plan_set_operations().
1721  *
1722  * We need to transpose sort key info from the orig_tlist into new_tlist.
1723  * NOTE: this would not be good enough if we supported resjunk sort keys
1724  * for results of set operations --- then, we'd need to project a whole
1725  * new tlist to evaluate the resjunk columns.  For now, just ereport if we
1726  * find any resjunk columns in orig_tlist.
1727  */
1728 static List *
1729 postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
1730 {
1731         ListCell   *l;
1732         ListCell   *orig_tlist_item = list_head(orig_tlist);
1733
1734         foreach(l, new_tlist)
1735         {
1736                 TargetEntry *new_tle = (TargetEntry *) lfirst(l);
1737                 TargetEntry *orig_tle;
1738
1739                 /* ignore resjunk columns in setop result */
1740                 if (new_tle->resjunk)
1741                         continue;
1742
1743                 Assert(orig_tlist_item != NULL);
1744                 orig_tle = (TargetEntry *) lfirst(orig_tlist_item);
1745                 orig_tlist_item = lnext(orig_tlist_item);
1746                 if (orig_tle->resjunk)  /* should not happen */
1747                         elog(ERROR, "resjunk output columns are not implemented");
1748                 Assert(new_tle->resno == orig_tle->resno);
1749                 new_tle->ressortgroupref = orig_tle->ressortgroupref;
1750         }
1751         if (orig_tlist_item != NULL)
1752                 elog(ERROR, "resjunk output columns are not implemented");
1753         return new_tlist;
1754 }