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) 1996-2005, 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 1.183 2005/04/10 19:50:08 tgl Exp $
  12  *
  13  *-------------------------------------------------------------------------
  14  */
  15
  16 #include "postgres.h"
  17
  18 #include <limits.h>
  19
  20 #include "catalog/pg_operator.h"
  21 #include "catalog/pg_type.h"
  22 #include "executor/executor.h"
  23 #include "executor/nodeAgg.h"
  24 #include "miscadmin.h"
  25 #include "nodes/makefuncs.h"
  26 #ifdef OPTIMIZER_DEBUG
  27 #include "nodes/print.h"
  28 #endif
  29 #include "optimizer/clauses.h"
  30 #include "optimizer/cost.h"
  31 #include "optimizer/pathnode.h"
  32 #include "optimizer/paths.h"
  33 #include "optimizer/planmain.h"
  34 #include "optimizer/planner.h"
  35 #include "optimizer/prep.h"
  36 #include "optimizer/subselect.h"
  37 #include "optimizer/tlist.h"
  38 #include "optimizer/var.h"
  39 #include "parser/parsetree.h"
  40 #include "parser/parse_expr.h"
  41 #include "parser/parse_oper.h"
  42 #include "utils/selfuncs.h"
  43 #include "utils/syscache.h"
  44
  45
  46 ParamListInfo PlannerBoundParamList = NULL;             /* current boundParams */
  47
  48
  49 /* Expression kind codes for preprocess_expression */
  50 #define EXPRKIND_QUAL   0
  51 #define EXPRKIND_TARGET 1
  52 #define EXPRKIND_RTFUNC 2
  53 #define EXPRKIND_LIMIT  3
  54 #define EXPRKIND_ININFO 4
  55
  56
  57 static Node *preprocess_expression(Query *parse, Node *expr, int kind);
  58 static void preprocess_qual_conditions(Query *parse, Node *jtnode);
  59 static Plan *inheritance_planner(Query *parse, List *inheritlist);
  60 static Plan *grouping_planner(Query *parse, double tuple_fraction);
  61 static bool choose_hashed_grouping(Query *parse, double tuple_fraction,
  62                                            Path *cheapest_path, Path *sorted_path,
  63                                            List *sort_pathkeys, List *group_pathkeys,
  64                                            double dNumGroups, AggClauseCounts *agg_counts);
  65 static bool hash_safe_grouping(Query *parse);
  66 static List *make_subplanTargetList(Query *parse, List *tlist,
  67                                            AttrNumber **groupColIdx, bool *need_tlist_eval);
  68 static void locate_grouping_columns(Query *parse,
  69                                                 List *tlist,
  70                                                 List *sub_tlist,
  71                                                 AttrNumber *groupColIdx);
  72 static List *postprocess_setop_tlist(List *new_tlist, List *orig_tlist);
  73
  74
  75 /*****************************************************************************
  76  *
  77  *         Query optimizer entry point
  78  *
  79  *****************************************************************************/
  80 Plan *
  81 planner(Query *parse, bool isCursor, int cursorOptions,
  82                 ParamListInfo boundParams)
  83 {
  84         double          tuple_fraction;
  85         Plan       *result_plan;
  86         Index           save_PlannerQueryLevel;
  87         List       *save_PlannerParamList;
  88         ParamListInfo save_PlannerBoundParamList;
  89
  90         /*
  91          * The planner can be called recursively (an example is when
  92          * eval_const_expressions tries to pre-evaluate an SQL function). So,
  93          * these global state variables must be saved and restored.
  94          *
  95          * Query level and the param list cannot be moved into the Query
  96          * structure since their whole purpose is communication across
  97          * multiple sub-Queries. Also, boundParams is explicitly info from
  98          * outside the Query, and so is likewise better handled as a global
  99          * variable.
 100          *
 101          * Note we do NOT save and restore PlannerPlanId: it exists to assign
 102          * unique IDs to SubPlan nodes, and we want those IDs to be unique for
 103          * the life of a backend.  Also, PlannerInitPlan is saved/restored in
 104          * subquery_planner, not here.
 105          */
 106         save_PlannerQueryLevel = PlannerQueryLevel;
 107         save_PlannerParamList = PlannerParamList;
 108         save_PlannerBoundParamList = PlannerBoundParamList;
 109
 110         /* Initialize state for handling outer-level references and params */
 111         PlannerQueryLevel = 0;          /* will be 1 in top-level subquery_planner */
 112         PlannerParamList = NIL;
 113         PlannerBoundParamList = boundParams;
 114
 115         /* Determine what fraction of the plan is likely to be scanned */
 116         if (isCursor)
 117         {
 118                 /*
 119                  * We have no real idea how many tuples the user will ultimately
 120                  * FETCH from a cursor, but it seems a good bet that he doesn't
 121                  * want 'em all.  Optimize for 10% retrieval (you gotta better
 122                  * number?      Should this be a SETtable parameter?)
 123                  */
 124                 tuple_fraction = 0.10;
 125         }
 126         else
 127         {
 128                 /* Default assumption is we need all the tuples */
 129                 tuple_fraction = 0.0;
 130         }
 131
 132         /* primary planning entry point (may recurse for subqueries) */
 133         result_plan = subquery_planner(parse, tuple_fraction);
 134
 135         Assert(PlannerQueryLevel == 0);
 136
 137         /*
 138          * If creating a plan for a scrollable cursor, make sure it can run
 139          * backwards on demand.  Add a Material node at the top at need.
 140          */
 141         if (isCursor && (cursorOptions & CURSOR_OPT_SCROLL))
 142         {
 143                 if (!ExecSupportsBackwardScan(result_plan))
 144                         result_plan = materialize_finished_plan(result_plan);
 145         }
 146
 147         /* executor wants to know total number of Params used overall */
 148         result_plan->nParamExec = list_length(PlannerParamList);
 149
 150         /* final cleanup of the plan */
 151         set_plan_references(result_plan, parse->rtable);
 152
 153         /* restore state for outer planner, if any */
 154         PlannerQueryLevel = save_PlannerQueryLevel;
 155         PlannerParamList = save_PlannerParamList;
 156         PlannerBoundParamList = save_PlannerBoundParamList;
 157
 158         return result_plan;
 159 }
 160
 161
 162 /*--------------------
 163  * subquery_planner
 164  *        Invokes the planner on a subquery.  We recurse to here for each
 165  *        sub-SELECT found in the query tree.
 166  *
 167  * parse is the querytree produced by the parser & rewriter.
 168  * tuple_fraction is the fraction of tuples we expect will be retrieved.
 169  * tuple_fraction is interpreted as explained for grouping_planner, below.
 170  *
 171  * Basically, this routine does the stuff that should only be done once
 172  * per Query object.  It then calls grouping_planner.  At one time,
 173  * grouping_planner could be invoked recursively on the same Query object;
 174  * that's not currently true, but we keep the separation between the two
 175  * routines anyway, in case we need it again someday.
 176  *
 177  * subquery_planner will be called recursively to handle sub-Query nodes
 178  * found within the query's expressions and rangetable.
 179  *
 180  * Returns a query plan.
 181  *--------------------
 182  */
 183 Plan *
 184 subquery_planner(Query *parse, double tuple_fraction)
 185 {
 186         List       *saved_initplan = PlannerInitPlan;
 187         int                     saved_planid = PlannerPlanId;
 188         bool            hasOuterJoins;
 189         Plan       *plan;
 190         List       *newHaving;
 191         List       *lst;
 192         ListCell   *l;
 193
 194         /* Set up for a new level of subquery */
 195         PlannerQueryLevel++;
 196         PlannerInitPlan = NIL;
 197
 198         /*
 199          * Look for IN clauses at the top level of WHERE, and transform them
 200          * into joins.  Note that this step only handles IN clauses originally
 201          * at top level of WHERE; if we pull up any subqueries in the next
 202          * step, their INs are processed just before pulling them up.
 203          */
 204         parse->in_info_list = NIL;
 205         if (parse->hasSubLinks)
 206                 parse->jointree->quals = pull_up_IN_clauses(parse,
 207                                                                                                  parse->jointree->quals);
 208
 209         /*
 210          * Check to see if any subqueries in the rangetable can be merged into
 211          * this query.
 212          */
 213         parse->jointree = (FromExpr *)
 214                 pull_up_subqueries(parse, (Node *) parse->jointree, false);
 215
 216         /*
 217          * Detect whether any rangetable entries are RTE_JOIN kind; if not, we
 218          * can avoid the expense of doing flatten_join_alias_vars().  Also
 219          * check for outer joins --- if none, we can skip
 220          * reduce_outer_joins(). This must be done after we have done
 221          * pull_up_subqueries, of course.
 222          */
 223         parse->hasJoinRTEs = false;
 224         hasOuterJoins = false;
 225         foreach(l, parse->rtable)
 226         {
 227                 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
 228
 229                 if (rte->rtekind == RTE_JOIN)
 230                 {
 231                         parse->hasJoinRTEs = true;
 232                         if (IS_OUTER_JOIN(rte->jointype))
 233                         {
 234                                 hasOuterJoins = true;
 235                                 /* Can quit scanning once we find an outer join */
 236                                 break;
 237                         }
 238                 }
 239         }
 240
 241         /*
 242          * Set hasHavingQual to remember if HAVING clause is present.  Needed
 243          * because preprocess_expression will reduce a constant-true condition
 244          * to an empty qual list ... but "HAVING TRUE" is not a semantic no-op.
 245          */
 246         parse->hasHavingQual = (parse->havingQual != NULL);
 247
 248         /*
 249          * Do expression preprocessing on targetlist and quals.
 250          */
 251         parse->targetList = (List *)
 252                 preprocess_expression(parse, (Node *) parse->targetList,
 253                                                           EXPRKIND_TARGET);
 254
 255         preprocess_qual_conditions(parse, (Node *) parse->jointree);
 256
 257         parse->havingQual = preprocess_expression(parse, parse->havingQual,
 258                                                                                           EXPRKIND_QUAL);
 259
 260         parse->limitOffset = preprocess_expression(parse, parse->limitOffset,
 261                                                                                            EXPRKIND_LIMIT);
 262         parse->limitCount = preprocess_expression(parse, parse->limitCount,
 263                                                                                           EXPRKIND_LIMIT);
 264
 265         parse->in_info_list = (List *)
 266                 preprocess_expression(parse, (Node *) parse->in_info_list,
 267                                                           EXPRKIND_ININFO);
 268
 269         /* Also need to preprocess expressions for function RTEs */
 270         foreach(l, parse->rtable)
 271         {
 272                 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
 273
 274                 if (rte->rtekind == RTE_FUNCTION)
 275                         rte->funcexpr = preprocess_expression(parse, rte->funcexpr,
 276                                                                                                   EXPRKIND_RTFUNC);
 277         }
 278
 279         /*
 280          * In some cases we may want to transfer a HAVING clause into WHERE.
 281          * We cannot do so if the HAVING clause contains aggregates (obviously)
 282          * or volatile functions (since a HAVING clause is supposed to be executed
 283          * only once per group).  Also, it may be that the clause is so expensive
 284          * to execute that we're better off doing it only once per group, despite
 285          * the loss of selectivity.  This is hard to estimate short of doing the
 286          * entire planning process twice, so we use a heuristic: clauses
 287          * containing subplans are left in HAVING.  Otherwise, we move or copy
 288          * the HAVING clause into WHERE, in hopes of eliminating tuples before
 289          * aggregation instead of after.
 290          *
 291          * If the query has explicit grouping then we can simply move such a
 292          * clause into WHERE; any group that fails the clause will not be
 293          * in the output because none of its tuples will reach the grouping
 294          * or aggregation stage.  Otherwise we must have a degenerate
 295          * (variable-free) HAVING clause, which we put in WHERE so that
 296          * query_planner() can use it in a gating Result node, but also keep
 297          * in HAVING to ensure that we don't emit a bogus aggregated row.
 298          * (This could be done better, but it seems not worth optimizing.)
 299          *
 300          * Note that both havingQual and parse->jointree->quals are in
 301          * implicitly-ANDed-list form at this point, even though they are
 302          * declared as Node *.
 303          */
 304         newHaving = NIL;
 305         foreach(l, (List *) parse->havingQual)
 306         {
 307                 Node       *havingclause = (Node *) lfirst(l);
 308
 309                 if (contain_agg_clause(havingclause) ||
 310                         contain_volatile_functions(havingclause) ||
 311                         contain_subplans(havingclause))
 312                 {
 313                         /* keep it in HAVING */
 314                         newHaving = lappend(newHaving, havingclause);
 315                 }
 316                 else if (parse->groupClause)
 317                 {
 318                         /* move it to WHERE */
 319                         parse->jointree->quals = (Node *)
 320                                 lappend((List *) parse->jointree->quals, havingclause);
 321                 }
 322                 else
 323                 {
 324                         /* put a copy in WHERE, keep it in HAVING */
 325                         parse->jointree->quals = (Node *)
 326                                 lappend((List *) parse->jointree->quals,
 327                                                 copyObject(havingclause));
 328                         newHaving = lappend(newHaving, havingclause);
 329                 }
 330         }
 331         parse->havingQual = (Node *) newHaving;
 332
 333         /*
 334          * If we have any outer joins, try to reduce them to plain inner
 335          * joins. This step is most easily done after we've done expression
 336          * preprocessing.
 337          */
 338         if (hasOuterJoins)
 339                 reduce_outer_joins(parse);
 340
 341         /*
 342          * See if we can simplify the jointree; opportunities for this may
 343          * come from having pulled up subqueries, or from flattening explicit
 344          * JOIN syntax.  We must do this after flattening JOIN alias
 345          * variables, since eliminating explicit JOIN nodes from the jointree
 346          * will cause get_relids_for_join() to fail.  But it should happen
 347          * after reduce_outer_joins, anyway.
 348          */
 349         parse->jointree = (FromExpr *)
 350                 simplify_jointree(parse, (Node *) parse->jointree);
 351
 352         /*
 353          * Do the main planning.  If we have an inherited target relation,
 354          * that needs special processing, else go straight to
 355          * grouping_planner.
 356          */
 357         if (parse->resultRelation &&
 358                 (lst = expand_inherited_rtentry(parse, parse->resultRelation)) != NIL)
 359                 plan = inheritance_planner(parse, lst);
 360         else
 361                 plan = grouping_planner(parse, tuple_fraction);
 362
 363         /*
 364          * If any subplans were generated, or if we're inside a subplan, build
 365          * initPlan list and extParam/allParam sets for plan nodes.
 366          */
 367         if (PlannerPlanId != saved_planid || PlannerQueryLevel > 1)
 368         {
 369                 Cost            initplan_cost = 0;
 370
 371                 /* Prepare extParam/allParam sets for all nodes in tree */
 372                 SS_finalize_plan(plan, parse->rtable);
 373
 374                 /*
 375                  * SS_finalize_plan doesn't handle initPlans, so we have to
 376                  * manually attach them to the topmost plan node, and add their
 377                  * extParams to the topmost node's, too.
 378                  *
 379                  * We also add the total_cost of each initPlan to the startup cost of
 380                  * the top node.  This is a conservative overestimate, since in
 381                  * fact each initPlan might be executed later than plan startup,
 382                  * or even not at all.
 383                  */
 384                 plan->initPlan = PlannerInitPlan;
 385
 386                 foreach(l, plan->initPlan)
 387                 {
 388                         SubPlan    *initplan = (SubPlan *) lfirst(l);
 389
 390                         plan->extParam = bms_add_members(plan->extParam,
 391                                                                                          initplan->plan->extParam);
 392                         /* allParam must include all members of extParam */
 393                         plan->allParam = bms_add_members(plan->allParam,
 394                                                                                          plan->extParam);
 395                         initplan_cost += initplan->plan->total_cost;
 396                 }
 397
 398                 plan->startup_cost += initplan_cost;
 399                 plan->total_cost += initplan_cost;
 400         }
 401
 402         /* Return to outer subquery context */
 403         PlannerQueryLevel--;
 404         PlannerInitPlan = saved_initplan;
 405         /* we do NOT restore PlannerPlanId; that's not an oversight! */
 406
 407         return plan;
 408 }
 409
 410 /*
 411  * preprocess_expression
 412  *              Do subquery_planner's preprocessing work for an expression,
 413  *              which can be a targetlist, a WHERE clause (including JOIN/ON
 414  *              conditions), or a HAVING clause.
 415  */
 416 static Node *
 417 preprocess_expression(Query *parse, Node *expr, int kind)
 418 {
 419         /*
 420          * If the query has any join RTEs, replace join alias variables with
 421          * base-relation variables. We must do this before sublink processing,
 422          * else sublinks expanded out from join aliases wouldn't get
 423          * processed.
 424          */
 425         if (parse->hasJoinRTEs)
 426                 expr = flatten_join_alias_vars(parse, expr);
 427
 428         /*
 429          * Simplify constant expressions.
 430          *
 431          * Note: this also flattens nested AND and OR expressions into N-argument
 432          * form.  All processing of a qual expression after this point must be
 433          * careful to maintain AND/OR flatness --- that is, do not generate a tree
 434          * with AND directly under AND, nor OR directly under OR.
 435          */
 436         expr = eval_const_expressions(expr);
 437
 438         /*
 439          * If it's a qual or havingQual, canonicalize it.
 440          */
 441         if (kind == EXPRKIND_QUAL)
 442         {
 443                 expr = (Node *) canonicalize_qual((Expr *) expr);
 444
 445 #ifdef OPTIMIZER_DEBUG
 446                 printf("After canonicalize_qual()\n");
 447                 pprint(expr);
 448 #endif
 449         }
 450
 451         /* Expand SubLinks to SubPlans */
 452         if (parse->hasSubLinks)
 453                 expr = SS_process_sublinks(expr, (kind == EXPRKIND_QUAL));
 454
 455         /*
 456          * XXX do not insert anything here unless you have grokked the
 457          * comments in SS_replace_correlation_vars ...
 458          */
 459
 460         /* Replace uplevel vars with Param nodes */
 461         if (PlannerQueryLevel > 1)
 462                 expr = SS_replace_correlation_vars(expr);
 463
 464         /*
 465          * If it's a qual or havingQual, convert it to implicit-AND format.
 466          * (We don't want to do this before eval_const_expressions, since the
 467          * latter would be unable to simplify a top-level AND correctly. Also,
 468          * SS_process_sublinks expects explicit-AND format.)
 469          */
 470         if (kind == EXPRKIND_QUAL)
 471                 expr = (Node *) make_ands_implicit((Expr *) expr);
 472
 473         return expr;
 474 }
 475
 476 /*
 477  * preprocess_qual_conditions
 478  *              Recursively scan the query's jointree and do subquery_planner's
 479  *              preprocessing work on each qual condition found therein.
 480  */
 481 static void
 482 preprocess_qual_conditions(Query *parse, Node *jtnode)
 483 {
 484         if (jtnode == NULL)
 485                 return;
 486         if (IsA(jtnode, RangeTblRef))
 487         {
 488                 /* nothing to do here */
 489         }
 490         else if (IsA(jtnode, FromExpr))
 491         {
 492                 FromExpr   *f = (FromExpr *) jtnode;
 493                 ListCell   *l;
 494
 495                 foreach(l, f->fromlist)
 496                         preprocess_qual_conditions(parse, lfirst(l));
 497
 498                 f->quals = preprocess_expression(parse, f->quals, EXPRKIND_QUAL);
 499         }
 500         else if (IsA(jtnode, JoinExpr))
 501         {
 502                 JoinExpr   *j = (JoinExpr *) jtnode;
 503
 504                 preprocess_qual_conditions(parse, j->larg);
 505                 preprocess_qual_conditions(parse, j->rarg);
 506
 507                 j->quals = preprocess_expression(parse, j->quals, EXPRKIND_QUAL);
 508         }
 509         else
 510                 elog(ERROR, "unrecognized node type: %d",
 511                          (int) nodeTag(jtnode));
 512 }
 513
 514 /*--------------------
 515  * inheritance_planner
 516  *        Generate a plan in the case where the result relation is an
 517  *        inheritance set.
 518  *
 519  * We have to handle this case differently from cases where a source
 520  * relation is an inheritance set.      Source inheritance is expanded at
 521  * the bottom of the plan tree (see allpaths.c), but target inheritance
 522  * has to be expanded at the top.  The reason is that for UPDATE, each
 523  * target relation needs a different targetlist matching its own column
 524  * set.  (This is not so critical for DELETE, but for simplicity we treat
 525  * inherited DELETE the same way.)      Fortunately, the UPDATE/DELETE target
 526  * can never be the nullable side of an outer join, so it's OK to generate
 527  * the plan this way.
 528  *
 529  * parse is the querytree produced by the parser & rewriter.
 530  * inheritlist is an integer list of RT indexes for the result relation set.
 531  *
 532  * Returns a query plan.
 533  *--------------------
 534  */
 535 static Plan *
 536 inheritance_planner(Query *parse, List *inheritlist)
 537 {
 538         int                     parentRTindex = parse->resultRelation;
 539         Oid                     parentOID = getrelid(parentRTindex, parse->rtable);
 540         int                     mainrtlength = list_length(parse->rtable);
 541         List       *subplans = NIL;
 542         List       *tlist = NIL;
 543         ListCell   *l;
 544
 545         foreach(l, inheritlist)
 546         {
 547                 int                     childRTindex = lfirst_int(l);
 548                 Oid                     childOID = getrelid(childRTindex, parse->rtable);
 549                 Query      *subquery;
 550                 Plan       *subplan;
 551
 552                 /* Generate modified query with this rel as target */
 553                 subquery = (Query *) adjust_inherited_attrs((Node *) parse,
 554                                                                                                 parentRTindex, parentOID,
 555                                                                                                  childRTindex, childOID);
 556                 /* Generate plan */
 557                 subplan = grouping_planner(subquery, 0.0 /* retrieve all tuples */ );
 558                 subplans = lappend(subplans, subplan);
 559
 560                 /*
 561                  * XXX my goodness this next bit is ugly.  Really need to think about
 562                  * ways to rein in planner's habit of scribbling on its input.
 563                  *
 564                  * Planning of the subquery might have modified the rangetable,
 565                  * either by addition of RTEs due to expansion of inherited source
 566                  * tables, or by changes of the Query structures inside subquery
 567                  * RTEs.  We have to ensure that this gets propagated back to the
 568                  * master copy.  However, if we aren't done planning yet, we also
 569                  * need to ensure that subsequent calls to grouping_planner have
 570                  * virgin sub-Queries to work from.  So, if we are at the last
 571                  * list entry, just copy the subquery rangetable back to the master
 572                  * copy; if we are not, then extend the master copy by adding
 573                  * whatever the subquery added.  (We assume these added entries
 574                  * will go untouched by the future grouping_planner calls.  We are
 575                  * also effectively assuming that sub-Queries will get planned
 576                  * identically each time, or at least that the impacts on their
 577                  * rangetables will be the same each time.  Did I say this is ugly?)
 578                  */
 579                 if (lnext(l) == NULL)
 580                         parse->rtable = subquery->rtable;
 581                 else
 582                 {
 583                         int             subrtlength = list_length(subquery->rtable);
 584
 585                         if (subrtlength > mainrtlength)
 586                         {
 587                                 List       *subrt;
 588
 589                                 subrt = list_copy_tail(subquery->rtable, mainrtlength);
 590                                 parse->rtable = list_concat(parse->rtable, subrt);
 591                                 mainrtlength = subrtlength;
 592                         }
 593                 }
 594
 595                 /* Save preprocessed tlist from first rel for use in Append */
 596                 if (tlist == NIL)
 597                         tlist = subplan->targetlist;
 598         }
 599
 600         /* Save the target-relations list for the executor, too */
 601         parse->resultRelations = inheritlist;
 602
 603         /* Mark result as unordered (probably unnecessary) */
 604         parse->query_pathkeys = NIL;
 605
 606         return (Plan *) make_append(subplans, true, tlist);
 607 }
 608
 609 /*--------------------
 610  * grouping_planner
 611  *        Perform planning steps related to grouping, aggregation, etc.
 612  *        This primarily means adding top-level processing to the basic
 613  *        query plan produced by query_planner.
 614  *
 615  * parse is the querytree produced by the parser & rewriter.
 616  * tuple_fraction is the fraction of tuples we expect will be retrieved
 617  *
 618  * tuple_fraction is interpreted as follows:
 619  *        0: expect all tuples to be retrieved (normal case)
 620  *        0 < tuple_fraction < 1: expect the given fraction of tuples available
 621  *              from the plan to be retrieved
 622  *        tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
 623  *              expected to be retrieved (ie, a LIMIT specification)
 624  *
 625  * Returns a query plan.  Also, parse->query_pathkeys is returned as the
 626  * actual output ordering of the plan (in pathkey format).
 627  *--------------------
 628  */
 629 static Plan *
 630 grouping_planner(Query *parse, double tuple_fraction)
 631 {
 632         List       *tlist = parse->targetList;
 633         Plan       *result_plan;
 634         List       *current_pathkeys;
 635         List       *sort_pathkeys;
 636
 637         if (parse->setOperations)
 638         {
 639                 List       *set_sortclauses;
 640
 641                 /*
 642                  * Construct the plan for set operations.  The result will not
 643                  * need any work except perhaps a top-level sort and/or LIMIT.
 644                  */
 645                 result_plan = plan_set_operations(parse,
 646                                                                                   &set_sortclauses);
 647
 648                 /*
 649                  * Calculate pathkeys representing the sort order (if any) of the
 650                  * set operation's result.  We have to do this before overwriting
 651                  * the sort key information...
 652                  */
 653                 current_pathkeys = make_pathkeys_for_sortclauses(set_sortclauses,
 654                                                                                                 result_plan->targetlist);
 655                 current_pathkeys = canonicalize_pathkeys(parse, current_pathkeys);
 656
 657                 /*
 658                  * We should not need to call preprocess_targetlist, since we must
 659                  * be in a SELECT query node.  Instead, use the targetlist
 660                  * returned by plan_set_operations (since this tells whether it
 661                  * returned any resjunk columns!), and transfer any sort key
 662                  * information from the original tlist.
 663                  */
 664                 Assert(parse->commandType == CMD_SELECT);
 665
 666                 tlist = postprocess_setop_tlist(result_plan->targetlist, tlist);
 667
 668                 /*
 669                  * Can't handle FOR UPDATE here (parser should have checked
 670                  * already, but let's make sure).
 671                  */
 672                 if (parse->rowMarks)
 673                         ereport(ERROR,
 674                                         (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
 675                                          errmsg("SELECT FOR UPDATE is not allowed with UNION/INTERSECT/EXCEPT")));
 676
 677                 /*
 678                  * Calculate pathkeys that represent result ordering requirements
 679                  */
 680                 sort_pathkeys = make_pathkeys_for_sortclauses(parse->sortClause,
 681                                                                                                           tlist);
 682                 sort_pathkeys = canonicalize_pathkeys(parse, sort_pathkeys);
 683         }
 684         else
 685         {
 686                 /* No set operations, do regular planning */
 687                 List       *sub_tlist;
 688                 List       *group_pathkeys;
 689                 AttrNumber *groupColIdx = NULL;
 690                 bool            need_tlist_eval = true;
 691                 QualCost        tlist_cost;
 692                 double          sub_tuple_fraction;
 693                 Path       *cheapest_path;
 694                 Path       *sorted_path;
 695                 double          dNumGroups = 0;
 696                 long            numGroups = 0;
 697                 AggClauseCounts agg_counts;
 698                 int                     numGroupCols = list_length(parse->groupClause);
 699                 bool            use_hashed_grouping = false;
 700
 701                 MemSet(&agg_counts, 0, sizeof(AggClauseCounts));
 702
 703                 /* Preprocess targetlist */
 704                 tlist = preprocess_targetlist(parse, tlist);
 705
 706                 /*
 707                  * Generate appropriate target list for subplan; may be different
 708                  * from tlist if grouping or aggregation is needed.
 709                  */
 710                 sub_tlist = make_subplanTargetList(parse, tlist,
 711                                                                                  &groupColIdx, &need_tlist_eval);
 712
 713                 /*
 714                  * Calculate pathkeys that represent grouping/ordering
 715                  * requirements
 716                  */
 717                 group_pathkeys = make_pathkeys_for_sortclauses(parse->groupClause,
 718                                                                                                            tlist);
 719                 sort_pathkeys = make_pathkeys_for_sortclauses(parse->sortClause,
 720                                                                                                           tlist);
 721
 722                 /*
 723                  * Will need actual number of aggregates for estimating costs.
 724                  *
 725                  * Note: we do not attempt to detect duplicate aggregates here; a
 726                  * somewhat-overestimated count is okay for our present purposes.
 727                  *
 728                  * Note: think not that we can turn off hasAggs if we find no aggs.
 729                  * It is possible for constant-expression simplification to remove
 730                  * all explicit references to aggs, but we still have to follow
 731                  * the aggregate semantics (eg, producing only one output row).
 732                  */
 733                 if (parse->hasAggs)
 734                 {
 735                         count_agg_clauses((Node *) tlist, &agg_counts);
 736                         count_agg_clauses(parse->havingQual, &agg_counts);
 737                 }
 738
 739                 /*
 740                  * Figure out whether we need a sorted result from query_planner.
 741                  *
 742                  * If we have a GROUP BY clause, then we want a result sorted
 743                  * properly for grouping.  Otherwise, if there is an ORDER BY
 744                  * clause, we want to sort by the ORDER BY clause.      (Note: if we
 745                  * have both, and ORDER BY is a superset of GROUP BY, it would be
 746                  * tempting to request sort by ORDER BY --- but that might just
 747                  * leave us failing to exploit an available sort order at all.
 748                  * Needs more thought...)
 749                  */
 750                 if (parse->groupClause)
 751                         parse->query_pathkeys = group_pathkeys;
 752                 else if (parse->sortClause)
 753                         parse->query_pathkeys = sort_pathkeys;
 754                 else
 755                         parse->query_pathkeys = NIL;
 756
 757                 /*
 758                  * Adjust tuple_fraction if we see that we are going to apply
 759                  * limiting/grouping/aggregation/etc.  This is not overridable by
 760                  * the caller, since it reflects plan actions that this routine
 761                  * will certainly take, not assumptions about context.
 762                  */
 763                 if (parse->limitCount != NULL)
 764                 {
 765                         /*
 766                          * A LIMIT clause limits the absolute number of tuples
 767                          * returned. However, if it's not a constant LIMIT then we
 768                          * have to punt; for lack of a better idea, assume 10% of the
 769                          * plan's result is wanted.
 770                          */
 771                         double          limit_fraction = 0.0;
 772
 773                         if (IsA(parse->limitCount, Const))
 774                         {
 775                                 Const      *limitc = (Const *) parse->limitCount;
 776                                 int32           count = DatumGetInt32(limitc->constvalue);
 777
 778                                 /*
 779                                  * A NULL-constant LIMIT represents "LIMIT ALL", which we
 780                                  * treat the same as no limit (ie, expect to retrieve all
 781                                  * the tuples).
 782                                  */
 783                                 if (!limitc->constisnull && count > 0)
 784                                 {
 785                                         limit_fraction = (double) count;
 786                                         /* We must also consider the OFFSET, if present */
 787                                         if (parse->limitOffset != NULL)
 788                                         {
 789                                                 if (IsA(parse->limitOffset, Const))
 790                                                 {
 791                                                         int32           offset;
 792
 793                                                         limitc = (Const *) parse->limitOffset;
 794                                                         offset = DatumGetInt32(limitc->constvalue);
 795                                                         if (!limitc->constisnull && offset > 0)
 796                                                                 limit_fraction += (double) offset;
 797                                                 }
 798                                                 else
 799                                                 {
 800                                                         /* OFFSET is an expression ... punt ... */
 801                                                         limit_fraction = 0.10;
 802                                                 }
 803                                         }
 804                                 }
 805                         }
 806                         else
 807                         {
 808                                 /* LIMIT is an expression ... punt ... */
 809                                 limit_fraction = 0.10;
 810                         }
 811
 812                         if (limit_fraction > 0.0)
 813                         {
 814                                 /*
 815                                  * If we have absolute limits from both caller and LIMIT,
 816                                  * use the smaller value; if one is fractional and the
 817                                  * other absolute, treat the fraction as a fraction of the
 818                                  * absolute value; else we can multiply the two fractions
 819                                  * together.
 820                                  */
 821                                 if (tuple_fraction >= 1.0)
 822                                 {
 823                                         if (limit_fraction >= 1.0)
 824                                         {
 825                                                 /* both absolute */
 826                                                 tuple_fraction = Min(tuple_fraction, limit_fraction);
 827                                         }
 828                                         else
 829                                         {
 830                                                 /* caller absolute, limit fractional */
 831                                                 tuple_fraction *= limit_fraction;
 832                                                 if (tuple_fraction < 1.0)
 833                                                         tuple_fraction = 1.0;
 834                                         }
 835                                 }
 836                                 else if (tuple_fraction > 0.0)
 837                                 {
 838                                         if (limit_fraction >= 1.0)
 839                                         {
 840                                                 /* caller fractional, limit absolute */
 841                                                 tuple_fraction *= limit_fraction;
 842                                                 if (tuple_fraction < 1.0)
 843                                                         tuple_fraction = 1.0;
 844                                         }
 845                                         else
 846                                         {
 847                                                 /* both fractional */
 848                                                 tuple_fraction *= limit_fraction;
 849                                         }
 850                                 }
 851                                 else
 852                                 {
 853                                         /* no info from caller, just use limit */
 854                                         tuple_fraction = limit_fraction;
 855                                 }
 856                         }
 857                 }
 858
 859                 /*
 860                  * With grouping or aggregation, the tuple fraction to pass to
 861                  * query_planner() may be different from what it is at top level.
 862                  */
 863                 sub_tuple_fraction = tuple_fraction;
 864
 865                 if (parse->groupClause)
 866                 {
 867                         /*
 868                          * In GROUP BY mode, we have the little problem that we don't
 869                          * really know how many input tuples will be needed to make a
 870                          * group, so we can't translate an output LIMIT count into an
 871                          * input count.  For lack of a better idea, assume 25% of the
 872                          * input data will be processed if there is any output limit.
 873                          * However, if the caller gave us a fraction rather than an
 874                          * absolute count, we can keep using that fraction (which
 875                          * amounts to assuming that all the groups are about the same
 876                          * size).
 877                          */
 878                         if (sub_tuple_fraction >= 1.0)
 879                                 sub_tuple_fraction = 0.25;
 880
 881                         /*
 882                          * If both GROUP BY and ORDER BY are specified, we will need
 883                          * two levels of sort --- and, therefore, certainly need to
 884                          * read all the input tuples --- unless ORDER BY is a subset
 885                          * of GROUP BY.  (We have not yet canonicalized the pathkeys,
 886                          * so must use the slower noncanonical comparison method.)
 887                          */
 888                         if (parse->groupClause && parse->sortClause &&
 889                                 !noncanonical_pathkeys_contained_in(sort_pathkeys,
 890                                                                                                         group_pathkeys))
 891                                 sub_tuple_fraction = 0.0;
 892                 }
 893                 else if (parse->hasAggs)
 894                 {
 895                         /*
 896                          * Ungrouped aggregate will certainly want all the input
 897                          * tuples.
 898                          */
 899                         sub_tuple_fraction = 0.0;
 900                 }
 901                 else if (parse->distinctClause)
 902                 {
 903                         /*
 904                          * SELECT DISTINCT, like GROUP, will absorb an unpredictable
 905                          * number of input tuples per output tuple.  Handle the same
 906                          * way.
 907                          */
 908                         if (sub_tuple_fraction >= 1.0)
 909                                 sub_tuple_fraction = 0.25;
 910                 }
 911
 912                 /*
 913                  * Generate the best unsorted and presorted paths for this Query
 914                  * (but note there may not be any presorted path).
 915                  */
 916                 query_planner(parse, sub_tlist, sub_tuple_fraction,
 917                                           &cheapest_path, &sorted_path);
 918
 919                 /*
 920                  * We couldn't canonicalize group_pathkeys and sort_pathkeys
 921                  * before running query_planner(), so do it now.
 922                  */
 923                 group_pathkeys = canonicalize_pathkeys(parse, group_pathkeys);
 924                 sort_pathkeys = canonicalize_pathkeys(parse, sort_pathkeys);
 925
 926                 /*
 927                  * If grouping, estimate the number of groups.  (We can't do this
 928                  * until after running query_planner(), either.)  Then decide
 929                  * whether we want to use hashed grouping.
 930                  */
 931                 if (parse->groupClause)
 932                 {
 933                         List       *groupExprs;
 934                         double          cheapest_path_rows;
 935
 936                         /*
 937                          * Beware of the possibility that cheapest_path->parent is NULL.
 938                          * This could happen if user does something silly like
 939                          *              SELECT 'foo' GROUP BY 1;
 940                          */
 941                         if (cheapest_path->parent)
 942                                 cheapest_path_rows = cheapest_path->parent->rows;
 943                         else
 944                                 cheapest_path_rows = 1; /* assume non-set result */
 945
 946                         groupExprs = get_sortgrouplist_exprs(parse->groupClause,
 947                                                                                                  parse->targetList);
 948                         dNumGroups = estimate_num_groups(parse,
 949                                                                                          groupExprs,
 950                                                                                          cheapest_path_rows);
 951                         /* Also want it as a long int --- but 'ware overflow! */
 952                         numGroups = (long) Min(dNumGroups, (double) LONG_MAX);
 953
 954                         use_hashed_grouping =
 955                                 choose_hashed_grouping(parse, tuple_fraction,
 956                                                                            cheapest_path, sorted_path,
 957                                                                            sort_pathkeys, group_pathkeys,
 958                                                                            dNumGroups, &agg_counts);
 959                 }
 960
 961                 /*
 962                  * Select the best path and create a plan to execute it.
 963                  *
 964                  * If we are doing hashed grouping, we will always read all the input
 965                  * tuples, so use the cheapest-total path.      Otherwise, trust
 966                  * query_planner's decision about which to use.
 967                  */
 968                 if (sorted_path && !use_hashed_grouping)
 969                 {
 970                         result_plan = create_plan(parse, sorted_path);
 971                         current_pathkeys = sorted_path->pathkeys;
 972                 }
 973                 else
 974                 {
 975                         result_plan = create_plan(parse, cheapest_path);
 976                         current_pathkeys = cheapest_path->pathkeys;
 977                 }
 978
 979                 /*
 980                  * create_plan() returns a plan with just a "flat" tlist of
 981                  * required Vars.  Usually we need to insert the sub_tlist as the
 982                  * tlist of the top plan node.  However, we can skip that if we
 983                  * determined that whatever query_planner chose to return will be
 984                  * good enough.
 985                  */
 986                 if (need_tlist_eval)
 987                 {
 988                         /*
 989                          * If the top-level plan node is one that cannot do expression
 990                          * evaluation, we must insert a Result node to project the
 991                          * desired tlist.
 992                          */
 993                         if (!is_projection_capable_plan(result_plan))
 994                         {
 995                                 result_plan = (Plan *) make_result(sub_tlist, NULL,
 996                                                                                                    result_plan);
 997                         }
 998                         else
 999                         {
1000                                 /*
1001                                  * Otherwise, just replace the subplan's flat tlist with
1002                                  * the desired tlist.
1003                                  */
1004                                 result_plan->targetlist = sub_tlist;
1005                         }
1006
1007                         /*
1008                          * Also, account for the cost of evaluation of the sub_tlist.
1009                          *
1010                          * Up to now, we have only been dealing with "flat" tlists,
1011                          * containing just Vars.  So their evaluation cost is zero
1012                          * according to the model used by cost_qual_eval() (or if you
1013                          * prefer, the cost is factored into cpu_tuple_cost).  Thus we
1014                          * can avoid accounting for tlist cost throughout
1015                          * query_planner() and subroutines.  But now we've inserted a
1016                          * tlist that might contain actual operators, sub-selects, etc
1017                          * --- so we'd better account for its cost.
1018                          *
1019                          * Below this point, any tlist eval cost for added-on nodes
1020                          * should be accounted for as we create those nodes.
1021                          * Presently, of the node types we can add on, only Agg and
1022                          * Group project new tlists (the rest just copy their input
1023                          * tuples) --- so make_agg() and make_group() are responsible
1024                          * for computing the added cost.
1025                          */
1026                         cost_qual_eval(&tlist_cost, sub_tlist);
1027                         result_plan->startup_cost += tlist_cost.startup;
1028                         result_plan->total_cost += tlist_cost.startup +
1029                                 tlist_cost.per_tuple * result_plan->plan_rows;
1030                 }
1031                 else
1032                 {
1033                         /*
1034                          * Since we're using query_planner's tlist and not the one
1035                          * make_subplanTargetList calculated, we have to refigure any
1036                          * grouping-column indexes make_subplanTargetList computed.
1037                          */
1038                         locate_grouping_columns(parse, tlist, result_plan->targetlist,
1039                                                                         groupColIdx);
1040                 }
1041
1042                 /*
1043                  * Insert AGG or GROUP node if needed, plus an explicit sort step
1044                  * if necessary.
1045                  *
1046                  * HAVING clause, if any, becomes qual of the Agg or Group node.
1047                  */
1048                 if (use_hashed_grouping)
1049                 {
1050                         /* Hashed aggregate plan --- no sort needed */
1051                         result_plan = (Plan *) make_agg(parse,
1052                                                                                         tlist,
1053                                                                                         (List *) parse->havingQual,
1054                                                                                         AGG_HASHED,
1055                                                                                         numGroupCols,
1056                                                                                         groupColIdx,
1057                                                                                         numGroups,
1058                                                                                         agg_counts.numAggs,
1059                                                                                         result_plan);
1060                         /* Hashed aggregation produces randomly-ordered results */
1061                         current_pathkeys = NIL;
1062                 }
1063                 else if (parse->hasAggs)
1064                 {
1065                         /* Plain aggregate plan --- sort if needed */
1066                         AggStrategy aggstrategy;
1067
1068                         if (parse->groupClause)
1069                         {
1070                                 if (!pathkeys_contained_in(group_pathkeys, current_pathkeys))
1071                                 {
1072                                         result_plan = (Plan *)
1073                                                 make_sort_from_groupcols(parse,
1074                                                                                                  parse->groupClause,
1075                                                                                                  groupColIdx,
1076                                                                                                  result_plan);
1077                                         current_pathkeys = group_pathkeys;
1078                                 }
1079                                 aggstrategy = AGG_SORTED;
1080
1081                                 /*
1082                                  * The AGG node will not change the sort ordering of its
1083                                  * groups, so current_pathkeys describes the result too.
1084                                  */
1085                         }
1086                         else
1087                         {
1088                                 aggstrategy = AGG_PLAIN;
1089                                 /* Result will be only one row anyway; no sort order */
1090                                 current_pathkeys = NIL;
1091                         }
1092
1093                         result_plan = (Plan *) make_agg(parse,
1094                                                                                         tlist,
1095                                                                                         (List *) parse->havingQual,
1096                                                                                         aggstrategy,
1097                                                                                         numGroupCols,
1098                                                                                         groupColIdx,
1099                                                                                         numGroups,
1100                                                                                         agg_counts.numAggs,
1101                                                                                         result_plan);
1102                 }
1103                 else if (parse->groupClause)
1104                 {
1105                         /*
1106                          * GROUP BY without aggregation, so insert a group node (plus the
1107                          * appropriate sort node, if necessary).
1108                          *
1109                          * Add an explicit sort if we couldn't make the path come
1110                          * out the way the GROUP node needs it.
1111                          */
1112                         if (!pathkeys_contained_in(group_pathkeys, current_pathkeys))
1113                         {
1114                                 result_plan = (Plan *)
1115                                         make_sort_from_groupcols(parse,
1116                                                                                          parse->groupClause,
1117                                                                                          groupColIdx,
1118                                                                                          result_plan);
1119                                 current_pathkeys = group_pathkeys;
1120                         }
1121
1122                         result_plan = (Plan *) make_group(parse,
1123                                                                                           tlist,
1124                                                                                           (List *) parse->havingQual,
1125                                                                                           numGroupCols,
1126                                                                                           groupColIdx,
1127                                                                                           dNumGroups,
1128                                                                                           result_plan);
1129                         /* The Group node won't change sort ordering */
1130                 }
1131                 else if (parse->hasHavingQual)
1132                 {
1133                         /*
1134                          * No aggregates, and no GROUP BY, but we have a HAVING qual.
1135                          * This is a degenerate case in which we are supposed to emit
1136                          * either 0 or 1 row depending on whether HAVING succeeds.
1137                          * Furthermore, there cannot be any variables in either HAVING
1138                          * or the targetlist, so we actually do not need the FROM table
1139                          * at all!  We can just throw away the plan-so-far and generate
1140                          * a Result node.  This is a sufficiently unusual corner case
1141                          * that it's not worth contorting the structure of this routine
1142                          * to avoid having to generate the plan in the first place.
1143                          */
1144                         result_plan = (Plan *) make_result(tlist,
1145                                                                                            parse->havingQual,
1146                                                                                            NULL);
1147                 }
1148         }                                                       /* end of if (setOperations) */
1149
1150         /*
1151          * If we were not able to make the plan come out in the right order,
1152          * add an explicit sort step.
1153          */
1154         if (parse->sortClause)
1155         {
1156                 if (!pathkeys_contained_in(sort_pathkeys, current_pathkeys))
1157                 {
1158                         result_plan = (Plan *)
1159                                 make_sort_from_sortclauses(parse,
1160                                                                                    parse->sortClause,
1161                                                                                    result_plan);
1162                         current_pathkeys = sort_pathkeys;
1163                 }
1164         }
1165
1166         /*
1167          * If there is a DISTINCT clause, add the UNIQUE node.
1168          */
1169         if (parse->distinctClause)
1170         {
1171                 result_plan = (Plan *) make_unique(result_plan, parse->distinctClause);
1172
1173                 /*
1174                  * If there was grouping or aggregation, leave plan_rows as-is
1175                  * (ie, assume the result was already mostly unique).  If not,
1176                  * it's reasonable to assume the UNIQUE filter has effects
1177                  * comparable to GROUP BY.
1178                  */
1179                 if (!parse->groupClause && !parse->hasHavingQual && !parse->hasAggs)
1180                 {
1181                         List       *distinctExprs;
1182
1183                         distinctExprs = get_sortgrouplist_exprs(parse->distinctClause,
1184                                                                                                         parse->targetList);
1185                         result_plan->plan_rows = estimate_num_groups(parse,
1186                                                                                                                  distinctExprs,
1187                                                                                                  result_plan->plan_rows);
1188                 }
1189         }
1190
1191         /*
1192          * Finally, if there is a LIMIT/OFFSET clause, add the LIMIT node.
1193          */
1194         if (parse->limitOffset || parse->limitCount)
1195         {
1196                 result_plan = (Plan *) make_limit(result_plan,
1197                                                                                   parse->limitOffset,
1198                                                                                   parse->limitCount);
1199         }
1200
1201         /*
1202          * Return the actual output ordering in query_pathkeys for possible
1203          * use by an outer query level.
1204          */
1205         parse->query_pathkeys = current_pathkeys;
1206
1207         return result_plan;
1208 }
1209
1210 /*
1211  * choose_hashed_grouping - should we use hashed grouping?
1212  */
1213 static bool
1214 choose_hashed_grouping(Query *parse, double tuple_fraction,
1215                                            Path *cheapest_path, Path *sorted_path,
1216                                            List *sort_pathkeys, List *group_pathkeys,
1217                                            double dNumGroups, AggClauseCounts *agg_counts)
1218 {
1219         int                     numGroupCols = list_length(parse->groupClause);
1220         double          cheapest_path_rows;
1221         int                     cheapest_path_width;
1222         Size            hashentrysize;
1223         List       *current_pathkeys;
1224         Path            hashed_p;
1225         Path            sorted_p;
1226
1227         /*
1228          * Check can't-do-it conditions, including whether the grouping operators
1229          * are hashjoinable.
1230          *
1231          * Executor doesn't support hashed aggregation with DISTINCT aggregates.
1232          * (Doing so would imply storing *all* the input values in the hash table,
1233          * which seems like a certain loser.)
1234          */
1235         if (!enable_hashagg)
1236                 return false;
1237         if (agg_counts->numDistinctAggs != 0)
1238                 return false;
1239         if (!hash_safe_grouping(parse))
1240                 return false;
1241
1242         /*
1243          * Don't do it if it doesn't look like the hashtable will fit into
1244          * work_mem.
1245          *
1246          * Beware here of the possibility that cheapest_path->parent is NULL.
1247          * This could happen if user does something silly like
1248          *              SELECT 'foo' GROUP BY 1;
1249          */
1250         if (cheapest_path->parent)
1251         {
1252                 cheapest_path_rows = cheapest_path->parent->rows;
1253                 cheapest_path_width = cheapest_path->parent->width;
1254         }
1255         else
1256         {
1257                 cheapest_path_rows = 1;                         /* assume non-set result */
1258                 cheapest_path_width = 100;                      /* arbitrary */
1259         }
1260
1261         /* Estimate per-hash-entry space at tuple width... */
1262         hashentrysize = cheapest_path_width;
1263         /* plus space for pass-by-ref transition values... */
1264         hashentrysize += agg_counts->transitionSpace;
1265         /* plus the per-hash-entry overhead */
1266         hashentrysize += hash_agg_entry_size(agg_counts->numAggs);
1267
1268         if (hashentrysize * dNumGroups > work_mem * 1024L)
1269                 return false;
1270
1271         /*
1272          * See if the estimated cost is no more than doing it the other way.
1273          * While avoiding the need for sorted input is usually a win, the fact
1274          * that the output won't be sorted may be a loss; so we need to do an
1275          * actual cost comparison.
1276          *
1277          * We need to consider
1278          *              cheapest_path + hashagg [+ final sort]
1279          * versus either
1280          *              cheapest_path [+ sort] + group or agg [+ final sort]
1281          * or
1282          *              presorted_path + group or agg [+ final sort]
1283          * where brackets indicate a step that may not be needed. We assume
1284          * query_planner() will have returned a presorted path only if it's a
1285          * winner compared to cheapest_path for this purpose.
1286          *
1287          * These path variables are dummies that just hold cost fields; we don't
1288          * make actual Paths for these steps.
1289          */
1290         cost_agg(&hashed_p, parse, AGG_HASHED, agg_counts->numAggs,
1291                          numGroupCols, dNumGroups,
1292                          cheapest_path->startup_cost, cheapest_path->total_cost,
1293                          cheapest_path_rows);
1294         /* Result of hashed agg is always unsorted */
1295         if (sort_pathkeys)
1296                 cost_sort(&hashed_p, parse, sort_pathkeys, hashed_p.total_cost,
1297                                   dNumGroups, cheapest_path_width);
1298
1299         if (sorted_path)
1300         {
1301                 sorted_p.startup_cost = sorted_path->startup_cost;
1302                 sorted_p.total_cost = sorted_path->total_cost;
1303                 current_pathkeys = sorted_path->pathkeys;
1304         }
1305         else
1306         {
1307                 sorted_p.startup_cost = cheapest_path->startup_cost;
1308                 sorted_p.total_cost = cheapest_path->total_cost;
1309                 current_pathkeys = cheapest_path->pathkeys;
1310         }
1311         if (!pathkeys_contained_in(group_pathkeys,
1312                                                            current_pathkeys))
1313         {
1314                 cost_sort(&sorted_p, parse, group_pathkeys, sorted_p.total_cost,
1315                                   cheapest_path_rows, cheapest_path_width);
1316                 current_pathkeys = group_pathkeys;
1317         }
1318
1319         if (parse->hasAggs)
1320                 cost_agg(&sorted_p, parse, AGG_SORTED, agg_counts->numAggs,
1321                                  numGroupCols, dNumGroups,
1322                                  sorted_p.startup_cost, sorted_p.total_cost,
1323                                  cheapest_path_rows);
1324         else
1325                 cost_group(&sorted_p, parse, numGroupCols, dNumGroups,
1326                                    sorted_p.startup_cost, sorted_p.total_cost,
1327                                    cheapest_path_rows);
1328         /* The Agg or Group node will preserve ordering */
1329         if (sort_pathkeys &&
1330                 !pathkeys_contained_in(sort_pathkeys, current_pathkeys))
1331                 cost_sort(&sorted_p, parse, sort_pathkeys, sorted_p.total_cost,
1332                                   dNumGroups, cheapest_path_width);
1333
1334         /*
1335          * Now make the decision using the top-level tuple fraction.  First we
1336          * have to convert an absolute count (LIMIT) into fractional form.
1337          */
1338         if (tuple_fraction >= 1.0)
1339                 tuple_fraction /= dNumGroups;
1340
1341         if (compare_fractional_path_costs(&hashed_p, &sorted_p,
1342                                                                           tuple_fraction) < 0)
1343         {
1344                 /* Hashed is cheaper, so use it */
1345                 return true;
1346         }
1347         return false;
1348 }
1349
1350 /*
1351  * hash_safe_grouping - are grouping operators hashable?
1352  *
1353  * We assume hashed aggregation will work if the datatype's equality operator
1354  * is marked hashjoinable.
1355  */
1356 static bool
1357 hash_safe_grouping(Query *parse)
1358 {
1359         ListCell   *gl;
1360
1361         foreach(gl, parse->groupClause)
1362         {
1363                 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1364                 TargetEntry *tle = get_sortgroupclause_tle(grpcl, parse->targetList);
1365                 Operator        optup;
1366                 bool            oprcanhash;
1367
1368                 optup = equality_oper(exprType((Node *) tle->expr), true);
1369                 if (!optup)
1370                         return false;
1371                 oprcanhash = ((Form_pg_operator) GETSTRUCT(optup))->oprcanhash;
1372                 ReleaseSysCache(optup);
1373                 if (!oprcanhash)
1374                         return false;
1375         }
1376         return true;
1377 }
1378
1379 /*---------------
1380  * make_subplanTargetList
1381  *        Generate appropriate target list when grouping is required.
1382  *
1383  * When grouping_planner inserts Aggregate, Group, or Result plan nodes
1384  * above the result of query_planner, we typically want to pass a different
1385  * target list to query_planner than the outer plan nodes should have.
1386  * This routine generates the correct target list for the subplan.
1387  *
1388  * The initial target list passed from the parser already contains entries
1389  * for all ORDER BY and GROUP BY expressions, but it will not have entries
1390  * for variables used only in HAVING clauses; so we need to add those
1391  * variables to the subplan target list.  Also, we flatten all expressions
1392  * except GROUP BY items into their component variables; the other expressions
1393  * will be computed by the inserted nodes rather than by the subplan.
1394  * For example, given a query like
1395  *              SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
1396  * we want to pass this targetlist to the subplan:
1397  *              a,b,c,d,a+b
1398  * where the a+b target will be used by the Sort/Group steps, and the
1399  * other targets will be used for computing the final results.  (In the
1400  * above example we could theoretically suppress the a and b targets and
1401  * pass down only c,d,a+b, but it's not really worth the trouble to
1402  * eliminate simple var references from the subplan.  We will avoid doing
1403  * the extra computation to recompute a+b at the outer level; see
1404  * replace_vars_with_subplan_refs() in setrefs.c.)
1405  *
1406  * If we are grouping or aggregating, *and* there are no non-Var grouping
1407  * expressions, then the returned tlist is effectively dummy; we do not
1408  * need to force it to be evaluated, because all the Vars it contains
1409  * should be present in the output of query_planner anyway.
1410  *
1411  * 'parse' is the query being processed.
1412  * 'tlist' is the query's target list.
1413  * 'groupColIdx' receives an array of column numbers for the GROUP BY
1414  *                      expressions (if there are any) in the subplan's target list.
1415  * 'need_tlist_eval' is set true if we really need to evaluate the
1416  *                      result tlist.
1417  *
1418  * The result is the targetlist to be passed to the subplan.
1419  *---------------
1420  */
1421 static List *
1422 make_subplanTargetList(Query *parse,
1423                                            List *tlist,
1424                                            AttrNumber **groupColIdx,
1425                                            bool *need_tlist_eval)
1426 {
1427         List       *sub_tlist;
1428         List       *extravars;
1429         int                     numCols;
1430
1431         *groupColIdx = NULL;
1432
1433         /*
1434          * If we're not grouping or aggregating, there's nothing to do here;
1435          * query_planner should receive the unmodified target list.
1436          */
1437         if (!parse->hasAggs && !parse->groupClause && !parse->hasHavingQual)
1438         {
1439                 *need_tlist_eval = true;
1440                 return tlist;
1441         }
1442
1443         /*
1444          * Otherwise, start with a "flattened" tlist (having just the vars
1445          * mentioned in the targetlist and HAVING qual --- but not upper-
1446          * level Vars; they will be replaced by Params later on).
1447          */
1448         sub_tlist = flatten_tlist(tlist);
1449         extravars = pull_var_clause(parse->havingQual, false);
1450         sub_tlist = add_to_flat_tlist(sub_tlist, extravars);
1451         list_free(extravars);
1452         *need_tlist_eval = false;       /* only eval if not flat tlist */
1453
1454         /*
1455          * If grouping, create sub_tlist entries for all GROUP BY expressions
1456          * (GROUP BY items that are simple Vars should be in the list
1457          * already), and make an array showing where the group columns are in
1458          * the sub_tlist.
1459          */
1460         numCols = list_length(parse->groupClause);
1461         if (numCols > 0)
1462         {
1463                 int                     keyno = 0;
1464                 AttrNumber *grpColIdx;
1465                 ListCell   *gl;
1466
1467                 grpColIdx = (AttrNumber *) palloc(sizeof(AttrNumber) * numCols);
1468                 *groupColIdx = grpColIdx;
1469
1470                 foreach(gl, parse->groupClause)
1471                 {
1472                         GroupClause *grpcl = (GroupClause *) lfirst(gl);
1473                         Node       *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1474                         TargetEntry *te = NULL;
1475                         ListCell   *sl;
1476
1477                         /* Find or make a matching sub_tlist entry */
1478                         foreach(sl, sub_tlist)
1479                         {
1480                                 te = (TargetEntry *) lfirst(sl);
1481                                 if (equal(groupexpr, te->expr))
1482                                         break;
1483                         }
1484                         if (!sl)
1485                         {
1486                                 te = makeTargetEntry((Expr *) groupexpr,
1487                                                                          list_length(sub_tlist) + 1,
1488                                                                          NULL,
1489                                                                          false);
1490                                 sub_tlist = lappend(sub_tlist, te);
1491                                 *need_tlist_eval = true;                /* it's not flat anymore */
1492                         }
1493
1494                         /* and save its resno */
1495                         grpColIdx[keyno++] = te->resno;
1496                 }
1497         }
1498
1499         return sub_tlist;
1500 }
1501
1502 /*
1503  * locate_grouping_columns
1504  *              Locate grouping columns in the tlist chosen by query_planner.
1505  *
1506  * This is only needed if we don't use the sub_tlist chosen by
1507  * make_subplanTargetList.      We have to forget the column indexes found
1508  * by that routine and re-locate the grouping vars in the real sub_tlist.
1509  */
1510 static void
1511 locate_grouping_columns(Query *parse,
1512                                                 List *tlist,
1513                                                 List *sub_tlist,
1514                                                 AttrNumber *groupColIdx)
1515 {
1516         int                     keyno = 0;
1517         ListCell   *gl;
1518
1519         /*
1520          * No work unless grouping.
1521          */
1522         if (!parse->groupClause)
1523         {
1524                 Assert(groupColIdx == NULL);
1525                 return;
1526         }
1527         Assert(groupColIdx != NULL);
1528
1529         foreach(gl, parse->groupClause)
1530         {
1531                 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1532                 Node       *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1533                 TargetEntry *te = NULL;
1534                 ListCell   *sl;
1535
1536                 foreach(sl, sub_tlist)
1537                 {
1538                         te = (TargetEntry *) lfirst(sl);
1539                         if (equal(groupexpr, te->expr))
1540                                 break;
1541                 }
1542                 if (!sl)
1543                         elog(ERROR, "failed to locate grouping columns");
1544
1545                 groupColIdx[keyno++] = te->resno;
1546         }
1547 }
1548
1549 /*
1550  * postprocess_setop_tlist
1551  *        Fix up targetlist returned by plan_set_operations().
1552  *
1553  * We need to transpose sort key info from the orig_tlist into new_tlist.
1554  * NOTE: this would not be good enough if we supported resjunk sort keys
1555  * for results of set operations --- then, we'd need to project a whole
1556  * new tlist to evaluate the resjunk columns.  For now, just ereport if we
1557  * find any resjunk columns in orig_tlist.
1558  */
1559 static List *
1560 postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
1561 {
1562         ListCell   *l;
1563         ListCell   *orig_tlist_item = list_head(orig_tlist);
1564
1565         foreach(l, new_tlist)
1566         {
1567                 TargetEntry *new_tle = (TargetEntry *) lfirst(l);
1568                 TargetEntry *orig_tle;
1569
1570                 /* ignore resjunk columns in setop result */
1571                 if (new_tle->resjunk)
1572                         continue;
1573
1574                 Assert(orig_tlist_item != NULL);
1575                 orig_tle = (TargetEntry *) lfirst(orig_tlist_item);
1576                 orig_tlist_item = lnext(orig_tlist_item);
1577                 if (orig_tle->resjunk)                  /* should not happen */
1578                         elog(ERROR, "resjunk output columns are not implemented");
1579                 Assert(new_tle->resno == orig_tle->resno);
1580                 new_tle->ressortgroupref = orig_tle->ressortgroupref;
1581         }
1582         if (orig_tlist_item != NULL)
1583                 elog(ERROR, "resjunk output columns are not implemented");
1584         return new_tlist;
1585 }