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 1.179 2005/03/10 23:21:22 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/analyze.h"
  40 #include "parser/parsetree.h"
  41 #include "parser/parse_expr.h"
  42 #include "parser/parse_oper.h"
  43 #include "utils/selfuncs.h"
  44 #include "utils/syscache.h"
  45
  46
  47 ParamListInfo PlannerBoundParamList = NULL;             /* current boundParams */
  48
  49
  50 /* Expression kind codes for preprocess_expression */
  51 #define EXPRKIND_QUAL   0
  52 #define EXPRKIND_TARGET 1
  53 #define EXPRKIND_RTFUNC 2
  54 #define EXPRKIND_LIMIT  3
  55 #define EXPRKIND_ININFO 4
  56
  57
  58 static Node *preprocess_expression(Query *parse, Node *expr, int kind);
  59 static void preprocess_qual_conditions(Query *parse, Node *jtnode);
  60 static Plan *inheritance_planner(Query *parse, List *inheritlist);
  61 static Plan *grouping_planner(Query *parse, double tuple_fraction);
  62 static bool hash_safe_grouping(Query *parse);
  63 static List *make_subplanTargetList(Query *parse, List *tlist,
  64                                            AttrNumber **groupColIdx, bool *need_tlist_eval);
  65 static void locate_grouping_columns(Query *parse,
  66                                                 List *tlist,
  67                                                 List *sub_tlist,
  68                                                 AttrNumber *groupColIdx);
  69 static List *postprocess_setop_tlist(List *new_tlist, List *orig_tlist);
  70
  71
  72 /*****************************************************************************
  73  *
  74  *         Query optimizer entry point
  75  *
  76  *****************************************************************************/
  77 Plan *
  78 planner(Query *parse, bool isCursor, int cursorOptions,
  79                 ParamListInfo boundParams)
  80 {
  81         double          tuple_fraction;
  82         Plan       *result_plan;
  83         Index           save_PlannerQueryLevel;
  84         List       *save_PlannerParamList;
  85         ParamListInfo save_PlannerBoundParamList;
  86
  87         /*
  88          * The planner can be called recursively (an example is when
  89          * eval_const_expressions tries to pre-evaluate an SQL function). So,
  90          * these global state variables must be saved and restored.
  91          *
  92          * Query level and the param list cannot be moved into the Query
  93          * structure since their whole purpose is communication across
  94          * multiple sub-Queries. Also, boundParams is explicitly info from
  95          * outside the Query, and so is likewise better handled as a global
  96          * variable.
  97          *
  98          * Note we do NOT save and restore PlannerPlanId: it exists to assign
  99          * unique IDs to SubPlan nodes, and we want those IDs to be unique for
 100          * the life of a backend.  Also, PlannerInitPlan is saved/restored in
 101          * subquery_planner, not here.
 102          */
 103         save_PlannerQueryLevel = PlannerQueryLevel;
 104         save_PlannerParamList = PlannerParamList;
 105         save_PlannerBoundParamList = PlannerBoundParamList;
 106
 107         /* Initialize state for handling outer-level references and params */
 108         PlannerQueryLevel = 0;          /* will be 1 in top-level subquery_planner */
 109         PlannerParamList = NIL;
 110         PlannerBoundParamList = boundParams;
 111
 112         /* Determine what fraction of the plan is likely to be scanned */
 113         if (isCursor)
 114         {
 115                 /*
 116                  * We have no real idea how many tuples the user will ultimately
 117                  * FETCH from a cursor, but it seems a good bet that he doesn't
 118                  * want 'em all.  Optimize for 10% retrieval (you gotta better
 119                  * number?      Should this be a SETtable parameter?)
 120                  */
 121                 tuple_fraction = 0.10;
 122         }
 123         else
 124         {
 125                 /* Default assumption is we need all the tuples */
 126                 tuple_fraction = 0.0;
 127         }
 128
 129         /* primary planning entry point (may recurse for subqueries) */
 130         result_plan = subquery_planner(parse, tuple_fraction);
 131
 132         Assert(PlannerQueryLevel == 0);
 133
 134         /*
 135          * If creating a plan for a scrollable cursor, make sure it can run
 136          * backwards on demand.  Add a Material node at the top at need.
 137          */
 138         if (isCursor && (cursorOptions & CURSOR_OPT_SCROLL))
 139         {
 140                 if (!ExecSupportsBackwardScan(result_plan))
 141                         result_plan = materialize_finished_plan(result_plan);
 142         }
 143
 144         /* executor wants to know total number of Params used overall */
 145         result_plan->nParamExec = list_length(PlannerParamList);
 146
 147         /* final cleanup of the plan */
 148         set_plan_references(result_plan, parse->rtable);
 149
 150         /* restore state for outer planner, if any */
 151         PlannerQueryLevel = save_PlannerQueryLevel;
 152         PlannerParamList = save_PlannerParamList;
 153         PlannerBoundParamList = save_PlannerBoundParamList;
 154
 155         return result_plan;
 156 }
 157
 158
 159 /*--------------------
 160  * subquery_planner
 161  *        Invokes the planner on a subquery.  We recurse to here for each
 162  *        sub-SELECT found in the query tree.
 163  *
 164  * parse is the querytree produced by the parser & rewriter.
 165  * tuple_fraction is the fraction of tuples we expect will be retrieved.
 166  * tuple_fraction is interpreted as explained for grouping_planner, below.
 167  *
 168  * Basically, this routine does the stuff that should only be done once
 169  * per Query object.  It then calls grouping_planner.  At one time,
 170  * grouping_planner could be invoked recursively on the same Query object;
 171  * that's not currently true, but we keep the separation between the two
 172  * routines anyway, in case we need it again someday.
 173  *
 174  * subquery_planner will be called recursively to handle sub-Query nodes
 175  * found within the query's expressions and rangetable.
 176  *
 177  * Returns a query plan.
 178  *--------------------
 179  */
 180 Plan *
 181 subquery_planner(Query *parse, double tuple_fraction)
 182 {
 183         List       *saved_initplan = PlannerInitPlan;
 184         int                     saved_planid = PlannerPlanId;
 185         bool            hasOuterJoins;
 186         Plan       *plan;
 187         List       *newHaving;
 188         List       *lst;
 189         ListCell   *l;
 190
 191         /* Set up for a new level of subquery */
 192         PlannerQueryLevel++;
 193         PlannerInitPlan = NIL;
 194
 195         /*
 196          * Look for IN clauses at the top level of WHERE, and transform them
 197          * into joins.  Note that this step only handles IN clauses originally
 198          * at top level of WHERE; if we pull up any subqueries in the next
 199          * step, their INs are processed just before pulling them up.
 200          */
 201         parse->in_info_list = NIL;
 202         if (parse->hasSubLinks)
 203                 parse->jointree->quals = pull_up_IN_clauses(parse,
 204                                                                                                  parse->jointree->quals);
 205
 206         /*
 207          * Check to see if any subqueries in the rangetable can be merged into
 208          * this query.
 209          */
 210         parse->jointree = (FromExpr *)
 211                 pull_up_subqueries(parse, (Node *) parse->jointree, false);
 212
 213         /*
 214          * Detect whether any rangetable entries are RTE_JOIN kind; if not, we
 215          * can avoid the expense of doing flatten_join_alias_vars().  Also
 216          * check for outer joins --- if none, we can skip
 217          * reduce_outer_joins(). This must be done after we have done
 218          * pull_up_subqueries, of course.
 219          */
 220         parse->hasJoinRTEs = false;
 221         hasOuterJoins = false;
 222         foreach(l, parse->rtable)
 223         {
 224                 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
 225
 226                 if (rte->rtekind == RTE_JOIN)
 227                 {
 228                         parse->hasJoinRTEs = true;
 229                         if (IS_OUTER_JOIN(rte->jointype))
 230                         {
 231                                 hasOuterJoins = true;
 232                                 /* Can quit scanning once we find an outer join */
 233                                 break;
 234                         }
 235                 }
 236         }
 237
 238         /*
 239          * Set hasHavingQual to remember if HAVING clause is present.  Needed
 240          * because preprocess_expression will reduce a constant-true condition
 241          * to an empty qual list ... but "HAVING TRUE" is not a semantic no-op.
 242          */
 243         parse->hasHavingQual = (parse->havingQual != NULL);
 244
 245         /*
 246          * Do expression preprocessing on targetlist and quals.
 247          */
 248         parse->targetList = (List *)
 249                 preprocess_expression(parse, (Node *) parse->targetList,
 250                                                           EXPRKIND_TARGET);
 251
 252         preprocess_qual_conditions(parse, (Node *) parse->jointree);
 253
 254         parse->havingQual = preprocess_expression(parse, parse->havingQual,
 255                                                                                           EXPRKIND_QUAL);
 256
 257         parse->limitOffset = preprocess_expression(parse, parse->limitOffset,
 258                                                                                            EXPRKIND_LIMIT);
 259         parse->limitCount = preprocess_expression(parse, parse->limitCount,
 260                                                                                           EXPRKIND_LIMIT);
 261
 262         parse->in_info_list = (List *)
 263                 preprocess_expression(parse, (Node *) parse->in_info_list,
 264                                                           EXPRKIND_ININFO);
 265
 266         /* Also need to preprocess expressions for function RTEs */
 267         foreach(l, parse->rtable)
 268         {
 269                 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
 270
 271                 if (rte->rtekind == RTE_FUNCTION)
 272                         rte->funcexpr = preprocess_expression(parse, rte->funcexpr,
 273                                                                                                   EXPRKIND_RTFUNC);
 274         }
 275
 276         /*
 277          * In some cases we may want to transfer a HAVING clause into WHERE.
 278          * We cannot do so if the HAVING clause contains aggregates (obviously)
 279          * or volatile functions (since a HAVING clause is supposed to be executed
 280          * only once per group).  Also, it may be that the clause is so expensive
 281          * to execute that we're better off doing it only once per group, despite
 282          * the loss of selectivity.  This is hard to estimate short of doing the
 283          * entire planning process twice, so we use a heuristic: clauses
 284          * containing subplans are left in HAVING.  Otherwise, we move or copy
 285          * the HAVING clause into WHERE, in hopes of eliminating tuples before
 286          * aggregation instead of after.
 287          *
 288          * If the query has explicit grouping then we can simply move such a
 289          * clause into WHERE; any group that fails the clause will not be
 290          * in the output because none of its tuples will reach the grouping
 291          * or aggregation stage.  Otherwise we must have a degenerate
 292          * (variable-free) HAVING clause, which we put in WHERE so that
 293          * query_planner() can use it in a gating Result node, but also keep
 294          * in HAVING to ensure that we don't emit a bogus aggregated row.
 295          * (This could be done better, but it seems not worth optimizing.)
 296          *
 297          * Note that both havingQual and parse->jointree->quals are in
 298          * implicitly-ANDed-list form at this point, even though they are
 299          * declared as Node *.
 300          */
 301         newHaving = NIL;
 302         foreach(l, (List *) parse->havingQual)
 303         {
 304                 Node       *havingclause = (Node *) lfirst(l);
 305
 306                 if (contain_agg_clause(havingclause) ||
 307                         contain_volatile_functions(havingclause) ||
 308                         contain_subplans(havingclause))
 309                 {
 310                         /* keep it in HAVING */
 311                         newHaving = lappend(newHaving, havingclause);
 312                 }
 313                 else if (parse->groupClause)
 314                 {
 315                         /* move it to WHERE */
 316                         parse->jointree->quals = (Node *)
 317                                 lappend((List *) parse->jointree->quals, havingclause);
 318                 }
 319                 else
 320                 {
 321                         /* put a copy in WHERE, keep it in HAVING */
 322                         parse->jointree->quals = (Node *)
 323                                 lappend((List *) parse->jointree->quals,
 324                                                 copyObject(havingclause));
 325                         newHaving = lappend(newHaving, havingclause);
 326                 }
 327         }
 328         parse->havingQual = (Node *) newHaving;
 329
 330         /*
 331          * If we have any outer joins, try to reduce them to plain inner
 332          * joins. This step is most easily done after we've done expression
 333          * preprocessing.
 334          */
 335         if (hasOuterJoins)
 336                 reduce_outer_joins(parse);
 337
 338         /*
 339          * See if we can simplify the jointree; opportunities for this may
 340          * come from having pulled up subqueries, or from flattening explicit
 341          * JOIN syntax.  We must do this after flattening JOIN alias
 342          * variables, since eliminating explicit JOIN nodes from the jointree
 343          * will cause get_relids_for_join() to fail.  But it should happen
 344          * after reduce_outer_joins, anyway.
 345          */
 346         parse->jointree = (FromExpr *)
 347                 simplify_jointree(parse, (Node *) parse->jointree);
 348
 349         /*
 350          * Do the main planning.  If we have an inherited target relation,
 351          * that needs special processing, else go straight to
 352          * grouping_planner.
 353          */
 354         if (parse->resultRelation &&
 355                 (lst = expand_inherited_rtentry(parse, parse->resultRelation)) != NIL)
 356                 plan = inheritance_planner(parse, lst);
 357         else
 358                 plan = grouping_planner(parse, tuple_fraction);
 359
 360         /*
 361          * If any subplans were generated, or if we're inside a subplan, build
 362          * initPlan list and extParam/allParam sets for plan nodes.
 363          */
 364         if (PlannerPlanId != saved_planid || PlannerQueryLevel > 1)
 365         {
 366                 Cost            initplan_cost = 0;
 367
 368                 /* Prepare extParam/allParam sets for all nodes in tree */
 369                 SS_finalize_plan(plan, parse->rtable);
 370
 371                 /*
 372                  * SS_finalize_plan doesn't handle initPlans, so we have to
 373                  * manually attach them to the topmost plan node, and add their
 374                  * extParams to the topmost node's, too.
 375                  *
 376                  * We also add the total_cost of each initPlan to the startup cost of
 377                  * the top node.  This is a conservative overestimate, since in
 378                  * fact each initPlan might be executed later than plan startup,
 379                  * or even not at all.
 380                  */
 381                 plan->initPlan = PlannerInitPlan;
 382
 383                 foreach(l, plan->initPlan)
 384                 {
 385                         SubPlan    *initplan = (SubPlan *) lfirst(l);
 386
 387                         plan->extParam = bms_add_members(plan->extParam,
 388                                                                                          initplan->plan->extParam);
 389                         /* allParam must include all members of extParam */
 390                         plan->allParam = bms_add_members(plan->allParam,
 391                                                                                          plan->extParam);
 392                         initplan_cost += initplan->plan->total_cost;
 393                 }
 394
 395                 plan->startup_cost += initplan_cost;
 396                 plan->total_cost += initplan_cost;
 397         }
 398
 399         /* Return to outer subquery context */
 400         PlannerQueryLevel--;
 401         PlannerInitPlan = saved_initplan;
 402         /* we do NOT restore PlannerPlanId; that's not an oversight! */
 403
 404         return plan;
 405 }
 406
 407 /*
 408  * preprocess_expression
 409  *              Do subquery_planner's preprocessing work for an expression,
 410  *              which can be a targetlist, a WHERE clause (including JOIN/ON
 411  *              conditions), or a HAVING clause.
 412  */
 413 static Node *
 414 preprocess_expression(Query *parse, Node *expr, int kind)
 415 {
 416         /*
 417          * If the query has any join RTEs, replace join alias variables with
 418          * base-relation variables. We must do this before sublink processing,
 419          * else sublinks expanded out from join aliases wouldn't get
 420          * processed.
 421          */
 422         if (parse->hasJoinRTEs)
 423                 expr = flatten_join_alias_vars(parse, expr);
 424
 425         /*
 426          * If it's a qual or havingQual, canonicalize it.  It seems most
 427          * useful to do this before applying eval_const_expressions, since the
 428          * latter can optimize flattened AND/ORs better than unflattened ones.
 429          *
 430          * Note: all processing of a qual expression after this point must be
 431          * careful to maintain AND/OR flatness --- that is, do not generate a
 432          * tree with AND directly under AND, nor OR directly under OR.
 433          */
 434         if (kind == EXPRKIND_QUAL)
 435         {
 436                 expr = (Node *) canonicalize_qual((Expr *) expr);
 437
 438 #ifdef OPTIMIZER_DEBUG
 439                 printf("After canonicalize_qual()\n");
 440                 pprint(expr);
 441 #endif
 442         }
 443
 444         /*
 445          * Simplify constant expressions.
 446          */
 447         expr = eval_const_expressions(expr);
 448
 449         /* Expand SubLinks to SubPlans */
 450         if (parse->hasSubLinks)
 451                 expr = SS_process_sublinks(expr, (kind == EXPRKIND_QUAL));
 452
 453         /*
 454          * XXX do not insert anything here unless you have grokked the
 455          * comments in SS_replace_correlation_vars ...
 456          */
 457
 458         /* Replace uplevel vars with Param nodes */
 459         if (PlannerQueryLevel > 1)
 460                 expr = SS_replace_correlation_vars(expr);
 461
 462         /*
 463          * If it's a qual or havingQual, convert it to implicit-AND format.
 464          * (We don't want to do this before eval_const_expressions, since the
 465          * latter would be unable to simplify a top-level AND correctly. Also,
 466          * SS_process_sublinks expects explicit-AND format.)
 467          */
 468         if (kind == EXPRKIND_QUAL)
 469                 expr = (Node *) make_ands_implicit((Expr *) expr);
 470
 471         return expr;
 472 }
 473
 474 /*
 475  * preprocess_qual_conditions
 476  *              Recursively scan the query's jointree and do subquery_planner's
 477  *              preprocessing work on each qual condition found therein.
 478  */
 479 static void
 480 preprocess_qual_conditions(Query *parse, Node *jtnode)
 481 {
 482         if (jtnode == NULL)
 483                 return;
 484         if (IsA(jtnode, RangeTblRef))
 485         {
 486                 /* nothing to do here */
 487         }
 488         else if (IsA(jtnode, FromExpr))
 489         {
 490                 FromExpr   *f = (FromExpr *) jtnode;
 491                 ListCell   *l;
 492
 493                 foreach(l, f->fromlist)
 494                         preprocess_qual_conditions(parse, lfirst(l));
 495
 496                 f->quals = preprocess_expression(parse, f->quals, EXPRKIND_QUAL);
 497         }
 498         else if (IsA(jtnode, JoinExpr))
 499         {
 500                 JoinExpr   *j = (JoinExpr *) jtnode;
 501
 502                 preprocess_qual_conditions(parse, j->larg);
 503                 preprocess_qual_conditions(parse, j->rarg);
 504
 505                 j->quals = preprocess_expression(parse, j->quals, EXPRKIND_QUAL);
 506         }
 507         else
 508                 elog(ERROR, "unrecognized node type: %d",
 509                          (int) nodeTag(jtnode));
 510 }
 511
 512 /*--------------------
 513  * inheritance_planner
 514  *        Generate a plan in the case where the result relation is an
 515  *        inheritance set.
 516  *
 517  * We have to handle this case differently from cases where a source
 518  * relation is an inheritance set.      Source inheritance is expanded at
 519  * the bottom of the plan tree (see allpaths.c), but target inheritance
 520  * has to be expanded at the top.  The reason is that for UPDATE, each
 521  * target relation needs a different targetlist matching its own column
 522  * set.  (This is not so critical for DELETE, but for simplicity we treat
 523  * inherited DELETE the same way.)      Fortunately, the UPDATE/DELETE target
 524  * can never be the nullable side of an outer join, so it's OK to generate
 525  * the plan this way.
 526  *
 527  * parse is the querytree produced by the parser & rewriter.
 528  * inheritlist is an integer list of RT indexes for the result relation set.
 529  *
 530  * Returns a query plan.
 531  *--------------------
 532  */
 533 static Plan *
 534 inheritance_planner(Query *parse, List *inheritlist)
 535 {
 536         int                     parentRTindex = parse->resultRelation;
 537         Oid                     parentOID = getrelid(parentRTindex, parse->rtable);
 538         int                     mainrtlength = list_length(parse->rtable);
 539         List       *subplans = NIL;
 540         List       *tlist = NIL;
 541         ListCell   *l;
 542
 543         foreach(l, inheritlist)
 544         {
 545                 int                     childRTindex = lfirst_int(l);
 546                 Oid                     childOID = getrelid(childRTindex, parse->rtable);
 547                 Query      *subquery;
 548                 Plan       *subplan;
 549
 550                 /* Generate modified query with this rel as target */
 551                 subquery = (Query *) adjust_inherited_attrs((Node *) parse,
 552                                                                                                 parentRTindex, parentOID,
 553                                                                                                  childRTindex, childOID);
 554                 /* Generate plan */
 555                 subplan = grouping_planner(subquery, 0.0 /* retrieve all tuples */ );
 556                 subplans = lappend(subplans, subplan);
 557
 558                 /*
 559                  * XXX my goodness this next bit is ugly.  Really need to think about
 560                  * ways to rein in planner's habit of scribbling on its input.
 561                  *
 562                  * Planning of the subquery might have modified the rangetable,
 563                  * either by addition of RTEs due to expansion of inherited source
 564                  * tables, or by changes of the Query structures inside subquery
 565                  * RTEs.  We have to ensure that this gets propagated back to the
 566                  * master copy.  However, if we aren't done planning yet, we also
 567                  * need to ensure that subsequent calls to grouping_planner have
 568                  * virgin sub-Queries to work from.  So, if we are at the last
 569                  * list entry, just copy the subquery rangetable back to the master
 570                  * copy; if we are not, then extend the master copy by adding
 571                  * whatever the subquery added.  (We assume these added entries
 572                  * will go untouched by the future grouping_planner calls.  We are
 573                  * also effectively assuming that sub-Queries will get planned
 574                  * identically each time, or at least that the impacts on their
 575                  * rangetables will be the same each time.  Did I say this is ugly?)
 576                  */
 577                 if (lnext(l) == NULL)
 578                         parse->rtable = subquery->rtable;
 579                 else
 580                 {
 581                         int             subrtlength = list_length(subquery->rtable);
 582
 583                         if (subrtlength > mainrtlength)
 584                         {
 585                                 List       *subrt;
 586
 587                                 subrt = list_copy_tail(subquery->rtable, mainrtlength);
 588                                 parse->rtable = list_concat(parse->rtable, subrt);
 589                                 mainrtlength = subrtlength;
 590                         }
 591                 }
 592
 593                 /* Save preprocessed tlist from first rel for use in Append */
 594                 if (tlist == NIL)
 595                         tlist = subplan->targetlist;
 596         }
 597
 598         /* Save the target-relations list for the executor, too */
 599         parse->resultRelations = inheritlist;
 600
 601         /* Mark result as unordered (probably unnecessary) */
 602         parse->query_pathkeys = NIL;
 603
 604         return (Plan *) make_append(subplans, true, tlist);
 605 }
 606
 607 /*--------------------
 608  * grouping_planner
 609  *        Perform planning steps related to grouping, aggregation, etc.
 610  *        This primarily means adding top-level processing to the basic
 611  *        query plan produced by query_planner.
 612  *
 613  * parse is the querytree produced by the parser & rewriter.
 614  * tuple_fraction is the fraction of tuples we expect will be retrieved
 615  *
 616  * tuple_fraction is interpreted as follows:
 617  *        0: expect all tuples to be retrieved (normal case)
 618  *        0 < tuple_fraction < 1: expect the given fraction of tuples available
 619  *              from the plan to be retrieved
 620  *        tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
 621  *              expected to be retrieved (ie, a LIMIT specification)
 622  *
 623  * Returns a query plan.  Also, parse->query_pathkeys is returned as the
 624  * actual output ordering of the plan (in pathkey format).
 625  *--------------------
 626  */
 627 static Plan *
 628 grouping_planner(Query *parse, double tuple_fraction)
 629 {
 630         List       *tlist = parse->targetList;
 631         Plan       *result_plan;
 632         List       *current_pathkeys;
 633         List       *sort_pathkeys;
 634
 635         if (parse->setOperations)
 636         {
 637                 List       *set_sortclauses;
 638
 639                 /*
 640                  * Construct the plan for set operations.  The result will not
 641                  * need any work except perhaps a top-level sort and/or LIMIT.
 642                  */
 643                 result_plan = plan_set_operations(parse,
 644                                                                                   &set_sortclauses);
 645
 646                 /*
 647                  * Calculate pathkeys representing the sort order (if any) of the
 648                  * set operation's result.  We have to do this before overwriting
 649                  * the sort key information...
 650                  */
 651                 current_pathkeys = make_pathkeys_for_sortclauses(set_sortclauses,
 652                                                                                                 result_plan->targetlist);
 653                 current_pathkeys = canonicalize_pathkeys(parse, current_pathkeys);
 654
 655                 /*
 656                  * We should not need to call preprocess_targetlist, since we must
 657                  * be in a SELECT query node.  Instead, use the targetlist
 658                  * returned by plan_set_operations (since this tells whether it
 659                  * returned any resjunk columns!), and transfer any sort key
 660                  * information from the original tlist.
 661                  */
 662                 Assert(parse->commandType == CMD_SELECT);
 663
 664                 tlist = postprocess_setop_tlist(result_plan->targetlist, tlist);
 665
 666                 /*
 667                  * Can't handle FOR UPDATE here (parser should have checked
 668                  * already, but let's make sure).
 669                  */
 670                 if (parse->rowMarks)
 671                         ereport(ERROR,
 672                                         (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
 673                                          errmsg("SELECT FOR UPDATE is not allowed with UNION/INTERSECT/EXCEPT")));
 674
 675                 /*
 676                  * Calculate pathkeys that represent result ordering requirements
 677                  */
 678                 sort_pathkeys = make_pathkeys_for_sortclauses(parse->sortClause,
 679                                                                                                           tlist);
 680                 sort_pathkeys = canonicalize_pathkeys(parse, sort_pathkeys);
 681         }
 682         else
 683         {
 684                 /* No set operations, do regular planning */
 685                 List       *sub_tlist;
 686                 List       *group_pathkeys;
 687                 AttrNumber *groupColIdx = NULL;
 688                 bool            need_tlist_eval = true;
 689                 QualCost        tlist_cost;
 690                 double          sub_tuple_fraction;
 691                 Path       *cheapest_path;
 692                 Path       *sorted_path;
 693                 double          dNumGroups = 0;
 694                 long            numGroups = 0;
 695                 AggClauseCounts agg_counts;
 696                 int                     numGroupCols = list_length(parse->groupClause);
 697                 bool            use_hashed_grouping = false;
 698
 699                 MemSet(&agg_counts, 0, sizeof(AggClauseCounts));
 700
 701                 /* Preprocess targetlist in case we are inside an INSERT/UPDATE. */
 702                 tlist = preprocess_targetlist(tlist,
 703                                                                           parse->commandType,
 704                                                                           parse->resultRelation,
 705                                                                           parse->rtable);
 706
 707                 /*
 708                  * Add TID targets for rels selected FOR UPDATE (should this be
 709                  * done in preprocess_targetlist?).  The executor uses the TID to
 710                  * know which rows to lock, much as for UPDATE or DELETE.
 711                  */
 712                 if (parse->rowMarks)
 713                 {
 714                         ListCell   *l;
 715
 716                         /*
 717                          * We've got trouble if the FOR UPDATE appears inside
 718                          * grouping, since grouping renders a reference to individual
 719                          * tuple CTIDs invalid.  This is also checked at parse time,
 720                          * but that's insufficient because of rule substitution, query
 721                          * pullup, etc.
 722                          */
 723                         CheckSelectForUpdate(parse);
 724
 725                         /*
 726                          * Currently the executor only supports FOR UPDATE at top
 727                          * level
 728                          */
 729                         if (PlannerQueryLevel > 1)
 730                                 ereport(ERROR,
 731                                                 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
 732                                                  errmsg("SELECT FOR UPDATE is not allowed in subqueries")));
 733
 734                         foreach(l, parse->rowMarks)
 735                         {
 736                                 Index           rti = lfirst_int(l);
 737                                 char       *resname;
 738                                 Resdom     *resdom;
 739                                 Var                *var;
 740                                 TargetEntry *ctid;
 741
 742                                 resname = (char *) palloc(32);
 743                                 snprintf(resname, 32, "ctid%u", rti);
 744                                 resdom = makeResdom(list_length(tlist) + 1,
 745                                                                         TIDOID,
 746                                                                         -1,
 747                                                                         resname,
 748                                                                         true);
 749
 750                                 var = makeVar(rti,
 751                                                           SelfItemPointerAttributeNumber,
 752                                                           TIDOID,
 753                                                           -1,
 754                                                           0);
 755
 756                                 ctid = makeTargetEntry(resdom, (Expr *) var);
 757                                 tlist = lappend(tlist, ctid);
 758                         }
 759                 }
 760
 761                 /*
 762                  * Generate appropriate target list for subplan; may be different
 763                  * from tlist if grouping or aggregation is needed.
 764                  */
 765                 sub_tlist = make_subplanTargetList(parse, tlist,
 766                                                                                  &groupColIdx, &need_tlist_eval);
 767
 768                 /*
 769                  * Calculate pathkeys that represent grouping/ordering
 770                  * requirements
 771                  */
 772                 group_pathkeys = make_pathkeys_for_sortclauses(parse->groupClause,
 773                                                                                                            tlist);
 774                 sort_pathkeys = make_pathkeys_for_sortclauses(parse->sortClause,
 775                                                                                                           tlist);
 776
 777                 /*
 778                  * Will need actual number of aggregates for estimating costs.
 779                  *
 780                  * Note: we do not attempt to detect duplicate aggregates here; a
 781                  * somewhat-overestimated count is okay for our present purposes.
 782                  *
 783                  * Note: think not that we can turn off hasAggs if we find no aggs.
 784                  * It is possible for constant-expression simplification to remove
 785                  * all explicit references to aggs, but we still have to follow
 786                  * the aggregate semantics (eg, producing only one output row).
 787                  */
 788                 if (parse->hasAggs)
 789                 {
 790                         count_agg_clauses((Node *) tlist, &agg_counts);
 791                         count_agg_clauses(parse->havingQual, &agg_counts);
 792                 }
 793
 794                 /*
 795                  * Figure out whether we need a sorted result from query_planner.
 796                  *
 797                  * If we have a GROUP BY clause, then we want a result sorted
 798                  * properly for grouping.  Otherwise, if there is an ORDER BY
 799                  * clause, we want to sort by the ORDER BY clause.      (Note: if we
 800                  * have both, and ORDER BY is a superset of GROUP BY, it would be
 801                  * tempting to request sort by ORDER BY --- but that might just
 802                  * leave us failing to exploit an available sort order at all.
 803                  * Needs more thought...)
 804                  */
 805                 if (parse->groupClause)
 806                         parse->query_pathkeys = group_pathkeys;
 807                 else if (parse->sortClause)
 808                         parse->query_pathkeys = sort_pathkeys;
 809                 else
 810                         parse->query_pathkeys = NIL;
 811
 812                 /*
 813                  * Adjust tuple_fraction if we see that we are going to apply
 814                  * limiting/grouping/aggregation/etc.  This is not overridable by
 815                  * the caller, since it reflects plan actions that this routine
 816                  * will certainly take, not assumptions about context.
 817                  */
 818                 if (parse->limitCount != NULL)
 819                 {
 820                         /*
 821                          * A LIMIT clause limits the absolute number of tuples
 822                          * returned. However, if it's not a constant LIMIT then we
 823                          * have to punt; for lack of a better idea, assume 10% of the
 824                          * plan's result is wanted.
 825                          */
 826                         double          limit_fraction = 0.0;
 827
 828                         if (IsA(parse->limitCount, Const))
 829                         {
 830                                 Const      *limitc = (Const *) parse->limitCount;
 831                                 int32           count = DatumGetInt32(limitc->constvalue);
 832
 833                                 /*
 834                                  * A NULL-constant LIMIT represents "LIMIT ALL", which we
 835                                  * treat the same as no limit (ie, expect to retrieve all
 836                                  * the tuples).
 837                                  */
 838                                 if (!limitc->constisnull && count > 0)
 839                                 {
 840                                         limit_fraction = (double) count;
 841                                         /* We must also consider the OFFSET, if present */
 842                                         if (parse->limitOffset != NULL)
 843                                         {
 844                                                 if (IsA(parse->limitOffset, Const))
 845                                                 {
 846                                                         int32           offset;
 847
 848                                                         limitc = (Const *) parse->limitOffset;
 849                                                         offset = DatumGetInt32(limitc->constvalue);
 850                                                         if (!limitc->constisnull && offset > 0)
 851                                                                 limit_fraction += (double) offset;
 852                                                 }
 853                                                 else
 854                                                 {
 855                                                         /* OFFSET is an expression ... punt ... */
 856                                                         limit_fraction = 0.10;
 857                                                 }
 858                                         }
 859                                 }
 860                         }
 861                         else
 862                         {
 863                                 /* LIMIT is an expression ... punt ... */
 864                                 limit_fraction = 0.10;
 865                         }
 866
 867                         if (limit_fraction > 0.0)
 868                         {
 869                                 /*
 870                                  * If we have absolute limits from both caller and LIMIT,
 871                                  * use the smaller value; if one is fractional and the
 872                                  * other absolute, treat the fraction as a fraction of the
 873                                  * absolute value; else we can multiply the two fractions
 874                                  * together.
 875                                  */
 876                                 if (tuple_fraction >= 1.0)
 877                                 {
 878                                         if (limit_fraction >= 1.0)
 879                                         {
 880                                                 /* both absolute */
 881                                                 tuple_fraction = Min(tuple_fraction, limit_fraction);
 882                                         }
 883                                         else
 884                                         {
 885                                                 /* caller absolute, limit fractional */
 886                                                 tuple_fraction *= limit_fraction;
 887                                                 if (tuple_fraction < 1.0)
 888                                                         tuple_fraction = 1.0;
 889                                         }
 890                                 }
 891                                 else if (tuple_fraction > 0.0)
 892                                 {
 893                                         if (limit_fraction >= 1.0)
 894                                         {
 895                                                 /* caller fractional, limit absolute */
 896                                                 tuple_fraction *= limit_fraction;
 897                                                 if (tuple_fraction < 1.0)
 898                                                         tuple_fraction = 1.0;
 899                                         }
 900                                         else
 901                                         {
 902                                                 /* both fractional */
 903                                                 tuple_fraction *= limit_fraction;
 904                                         }
 905                                 }
 906                                 else
 907                                 {
 908                                         /* no info from caller, just use limit */
 909                                         tuple_fraction = limit_fraction;
 910                                 }
 911                         }
 912                 }
 913
 914                 /*
 915                  * With grouping or aggregation, the tuple fraction to pass to
 916                  * query_planner() may be different from what it is at top level.
 917                  */
 918                 sub_tuple_fraction = tuple_fraction;
 919
 920                 if (parse->groupClause)
 921                 {
 922                         /*
 923                          * In GROUP BY mode, we have the little problem that we don't
 924                          * really know how many input tuples will be needed to make a
 925                          * group, so we can't translate an output LIMIT count into an
 926                          * input count.  For lack of a better idea, assume 25% of the
 927                          * input data will be processed if there is any output limit.
 928                          * However, if the caller gave us a fraction rather than an
 929                          * absolute count, we can keep using that fraction (which
 930                          * amounts to assuming that all the groups are about the same
 931                          * size).
 932                          */
 933                         if (sub_tuple_fraction >= 1.0)
 934                                 sub_tuple_fraction = 0.25;
 935
 936                         /*
 937                          * If both GROUP BY and ORDER BY are specified, we will need
 938                          * two levels of sort --- and, therefore, certainly need to
 939                          * read all the input tuples --- unless ORDER BY is a subset
 940                          * of GROUP BY.  (We have not yet canonicalized the pathkeys,
 941                          * so must use the slower noncanonical comparison method.)
 942                          */
 943                         if (parse->groupClause && parse->sortClause &&
 944                                 !noncanonical_pathkeys_contained_in(sort_pathkeys,
 945                                                                                                         group_pathkeys))
 946                                 sub_tuple_fraction = 0.0;
 947                 }
 948                 else if (parse->hasAggs)
 949                 {
 950                         /*
 951                          * Ungrouped aggregate will certainly want all the input
 952                          * tuples.
 953                          */
 954                         sub_tuple_fraction = 0.0;
 955                 }
 956                 else if (parse->distinctClause)
 957                 {
 958                         /*
 959                          * SELECT DISTINCT, like GROUP, will absorb an unpredictable
 960                          * number of input tuples per output tuple.  Handle the same
 961                          * way.
 962                          */
 963                         if (sub_tuple_fraction >= 1.0)
 964                                 sub_tuple_fraction = 0.25;
 965                 }
 966
 967                 /*
 968                  * Generate the best unsorted and presorted paths for this Query
 969                  * (but note there may not be any presorted path).
 970                  */
 971                 query_planner(parse, sub_tlist, sub_tuple_fraction,
 972                                           &cheapest_path, &sorted_path);
 973
 974                 /*
 975                  * We couldn't canonicalize group_pathkeys and sort_pathkeys
 976                  * before running query_planner(), so do it now.
 977                  */
 978                 group_pathkeys = canonicalize_pathkeys(parse, group_pathkeys);
 979                 sort_pathkeys = canonicalize_pathkeys(parse, sort_pathkeys);
 980
 981                 /*
 982                  * Consider whether we might want to use hashed grouping.
 983                  */
 984                 if (parse->groupClause)
 985                 {
 986                         List       *groupExprs;
 987                         double          cheapest_path_rows;
 988                         int                     cheapest_path_width;
 989
 990                         /*
 991                          * Beware in this section of the possibility that
 992                          * cheapest_path->parent is NULL.  This could happen if user
 993                          * does something silly like SELECT 'foo' GROUP BY 1;
 994                          */
 995                         if (cheapest_path->parent)
 996                         {
 997                                 cheapest_path_rows = cheapest_path->parent->rows;
 998                                 cheapest_path_width = cheapest_path->parent->width;
 999                         }
1000                         else
1001                         {
1002                                 cheapest_path_rows = 1; /* assume non-set result */
1003                                 cheapest_path_width = 100;              /* arbitrary */
1004                         }
1005
1006                         /*
1007                          * Always estimate the number of groups.  We can't do this
1008                          * until after running query_planner(), either.
1009                          */
1010                         groupExprs = get_sortgrouplist_exprs(parse->groupClause,
1011                                                                                                  parse->targetList);
1012                         dNumGroups = estimate_num_groups(parse,
1013                                                                                          groupExprs,
1014                                                                                          cheapest_path_rows);
1015                         /* Also want it as a long int --- but 'ware overflow! */
1016                         numGroups = (long) Min(dNumGroups, (double) LONG_MAX);
1017
1018                         /*
1019                          * Check can't-do-it conditions, including whether the
1020                          * grouping operators are hashjoinable.
1021                          *
1022                          * Executor doesn't support hashed aggregation with DISTINCT
1023                          * aggregates.  (Doing so would imply storing *all* the input
1024                          * values in the hash table, which seems like a certain
1025                          * loser.)
1026                          */
1027                         if (!enable_hashagg || !hash_safe_grouping(parse))
1028                                 use_hashed_grouping = false;
1029                         else if (agg_counts.numDistinctAggs != 0)
1030                                 use_hashed_grouping = false;
1031                         else
1032                         {
1033                                 /*
1034                                  * Use hashed grouping if (a) we think we can fit the
1035                                  * hashtable into work_mem, *and* (b) the estimated cost
1036                                  * is no more than doing it the other way.      While avoiding
1037                                  * the need for sorted input is usually a win, the fact
1038                                  * that the output won't be sorted may be a loss; so we
1039                                  * need to do an actual cost comparison.
1040                                  */
1041                                 Size            hashentrysize;
1042
1043                                 /* Estimate per-hash-entry space at tuple width... */
1044                                 hashentrysize = cheapest_path_width;
1045                                 /* plus space for pass-by-ref transition values... */
1046                                 hashentrysize += agg_counts.transitionSpace;
1047                                 /* plus the per-hash-entry overhead */
1048                                 hashentrysize += hash_agg_entry_size(agg_counts.numAggs);
1049
1050                                 if (hashentrysize * dNumGroups <= work_mem * 1024L)
1051                                 {
1052                                         /*
1053                                          * Okay, do the cost comparison.  We need to consider
1054                                          * cheapest_path + hashagg [+ final sort] versus
1055                                          * either cheapest_path [+ sort] + group or agg [+
1056                                          * final sort] or presorted_path + group or agg [+
1057                                          * final sort] where brackets indicate a step that may
1058                                          * not be needed. We assume query_planner() will have
1059                                          * returned a presorted path only if it's a winner
1060                                          * compared to cheapest_path for this purpose.
1061                                          *
1062                                          * These path variables are dummies that just hold cost
1063                                          * fields; we don't make actual Paths for these steps.
1064                                          */
1065                                         Path            hashed_p;
1066                                         Path            sorted_p;
1067
1068                                         cost_agg(&hashed_p, parse,
1069                                                          AGG_HASHED, agg_counts.numAggs,
1070                                                          numGroupCols, dNumGroups,
1071                                                          cheapest_path->startup_cost,
1072                                                          cheapest_path->total_cost,
1073                                                          cheapest_path_rows);
1074                                         /* Result of hashed agg is always unsorted */
1075                                         if (sort_pathkeys)
1076                                                 cost_sort(&hashed_p, parse, sort_pathkeys,
1077                                                                   hashed_p.total_cost,
1078                                                                   dNumGroups,
1079                                                                   cheapest_path_width);
1080
1081                                         if (sorted_path)
1082                                         {
1083                                                 sorted_p.startup_cost = sorted_path->startup_cost;
1084                                                 sorted_p.total_cost = sorted_path->total_cost;
1085                                                 current_pathkeys = sorted_path->pathkeys;
1086                                         }
1087                                         else
1088                                         {
1089                                                 sorted_p.startup_cost = cheapest_path->startup_cost;
1090                                                 sorted_p.total_cost = cheapest_path->total_cost;
1091                                                 current_pathkeys = cheapest_path->pathkeys;
1092                                         }
1093                                         if (!pathkeys_contained_in(group_pathkeys,
1094                                                                                            current_pathkeys))
1095                                         {
1096                                                 cost_sort(&sorted_p, parse, group_pathkeys,
1097                                                                   sorted_p.total_cost,
1098                                                                   cheapest_path_rows,
1099                                                                   cheapest_path_width);
1100                                                 current_pathkeys = group_pathkeys;
1101                                         }
1102                                         if (parse->hasAggs)
1103                                                 cost_agg(&sorted_p, parse,
1104                                                                  AGG_SORTED, agg_counts.numAggs,
1105                                                                  numGroupCols, dNumGroups,
1106                                                                  sorted_p.startup_cost,
1107                                                                  sorted_p.total_cost,
1108                                                                  cheapest_path_rows);
1109                                         else
1110                                                 cost_group(&sorted_p, parse,
1111                                                                    numGroupCols, dNumGroups,
1112                                                                    sorted_p.startup_cost,
1113                                                                    sorted_p.total_cost,
1114                                                                    cheapest_path_rows);
1115                                         /* The Agg or Group node will preserve ordering */
1116                                         if (sort_pathkeys &&
1117                                                 !pathkeys_contained_in(sort_pathkeys,
1118                                                                                            current_pathkeys))
1119                                         {
1120                                                 cost_sort(&sorted_p, parse, sort_pathkeys,
1121                                                                   sorted_p.total_cost,
1122                                                                   dNumGroups,
1123                                                                   cheapest_path_width);
1124                                         }
1125
1126                                         /*
1127                                          * Now make the decision using the top-level tuple
1128                                          * fraction.  First we have to convert an absolute
1129                                          * count (LIMIT) into fractional form.
1130                                          */
1131                                         if (tuple_fraction >= 1.0)
1132                                                 tuple_fraction /= dNumGroups;
1133
1134                                         if (compare_fractional_path_costs(&hashed_p, &sorted_p,
1135                                                                                                           tuple_fraction) < 0)
1136                                         {
1137                                                 /* Hashed is cheaper, so use it */
1138                                                 use_hashed_grouping = true;
1139                                         }
1140                                 }
1141                         }
1142                 }
1143
1144                 /*
1145                  * Select the best path and create a plan to execute it.
1146                  *
1147                  * If we are doing hashed grouping, we will always read all the input
1148                  * tuples, so use the cheapest-total path.      Otherwise, trust
1149                  * query_planner's decision about which to use.
1150                  */
1151                 if (sorted_path && !use_hashed_grouping)
1152                 {
1153                         result_plan = create_plan(parse, sorted_path);
1154                         current_pathkeys = sorted_path->pathkeys;
1155                 }
1156                 else
1157                 {
1158                         result_plan = create_plan(parse, cheapest_path);
1159                         current_pathkeys = cheapest_path->pathkeys;
1160                 }
1161
1162                 /*
1163                  * create_plan() returns a plan with just a "flat" tlist of
1164                  * required Vars.  Usually we need to insert the sub_tlist as the
1165                  * tlist of the top plan node.  However, we can skip that if we
1166                  * determined that whatever query_planner chose to return will be
1167                  * good enough.
1168                  */
1169                 if (need_tlist_eval)
1170                 {
1171                         /*
1172                          * If the top-level plan node is one that cannot do expression
1173                          * evaluation, we must insert a Result node to project the
1174                          * desired tlist.
1175                          */
1176                         if (!is_projection_capable_plan(result_plan))
1177                         {
1178                                 result_plan = (Plan *) make_result(sub_tlist, NULL,
1179                                                                                                    result_plan);
1180                         }
1181                         else
1182                         {
1183                                 /*
1184                                  * Otherwise, just replace the subplan's flat tlist with
1185                                  * the desired tlist.
1186                                  */
1187                                 result_plan->targetlist = sub_tlist;
1188                         }
1189
1190                         /*
1191                          * Also, account for the cost of evaluation of the sub_tlist.
1192                          *
1193                          * Up to now, we have only been dealing with "flat" tlists,
1194                          * containing just Vars.  So their evaluation cost is zero
1195                          * according to the model used by cost_qual_eval() (or if you
1196                          * prefer, the cost is factored into cpu_tuple_cost).  Thus we
1197                          * can avoid accounting for tlist cost throughout
1198                          * query_planner() and subroutines.  But now we've inserted a
1199                          * tlist that might contain actual operators, sub-selects, etc
1200                          * --- so we'd better account for its cost.
1201                          *
1202                          * Below this point, any tlist eval cost for added-on nodes
1203                          * should be accounted for as we create those nodes.
1204                          * Presently, of the node types we can add on, only Agg and
1205                          * Group project new tlists (the rest just copy their input
1206                          * tuples) --- so make_agg() and make_group() are responsible
1207                          * for computing the added cost.
1208                          */
1209                         cost_qual_eval(&tlist_cost, sub_tlist);
1210                         result_plan->startup_cost += tlist_cost.startup;
1211                         result_plan->total_cost += tlist_cost.startup +
1212                                 tlist_cost.per_tuple * result_plan->plan_rows;
1213                 }
1214                 else
1215                 {
1216                         /*
1217                          * Since we're using query_planner's tlist and not the one
1218                          * make_subplanTargetList calculated, we have to refigure any
1219                          * grouping-column indexes make_subplanTargetList computed.
1220                          */
1221                         locate_grouping_columns(parse, tlist, result_plan->targetlist,
1222                                                                         groupColIdx);
1223                 }
1224
1225                 /*
1226                  * Insert AGG or GROUP node if needed, plus an explicit sort step
1227                  * if necessary.
1228                  *
1229                  * HAVING clause, if any, becomes qual of the Agg or Group node.
1230                  */
1231                 if (use_hashed_grouping)
1232                 {
1233                         /* Hashed aggregate plan --- no sort needed */
1234                         result_plan = (Plan *) make_agg(parse,
1235                                                                                         tlist,
1236                                                                                         (List *) parse->havingQual,
1237                                                                                         AGG_HASHED,
1238                                                                                         numGroupCols,
1239                                                                                         groupColIdx,
1240                                                                                         numGroups,
1241                                                                                         agg_counts.numAggs,
1242                                                                                         result_plan);
1243                         /* Hashed aggregation produces randomly-ordered results */
1244                         current_pathkeys = NIL;
1245                 }
1246                 else if (parse->hasAggs)
1247                 {
1248                         /* Plain aggregate plan --- sort if needed */
1249                         AggStrategy aggstrategy;
1250
1251                         if (parse->groupClause)
1252                         {
1253                                 if (!pathkeys_contained_in(group_pathkeys, current_pathkeys))
1254                                 {
1255                                         result_plan = (Plan *)
1256                                                 make_sort_from_groupcols(parse,
1257                                                                                                  parse->groupClause,
1258                                                                                                  groupColIdx,
1259                                                                                                  result_plan);
1260                                         current_pathkeys = group_pathkeys;
1261                                 }
1262                                 aggstrategy = AGG_SORTED;
1263
1264                                 /*
1265                                  * The AGG node will not change the sort ordering of its
1266                                  * groups, so current_pathkeys describes the result too.
1267                                  */
1268                         }
1269                         else
1270                         {
1271                                 aggstrategy = AGG_PLAIN;
1272                                 /* Result will be only one row anyway; no sort order */
1273                                 current_pathkeys = NIL;
1274                         }
1275
1276                         result_plan = (Plan *) make_agg(parse,
1277                                                                                         tlist,
1278                                                                                         (List *) parse->havingQual,
1279                                                                                         aggstrategy,
1280                                                                                         numGroupCols,
1281                                                                                         groupColIdx,
1282                                                                                         numGroups,
1283                                                                                         agg_counts.numAggs,
1284                                                                                         result_plan);
1285                 }
1286                 else if (parse->groupClause)
1287                 {
1288                         /*
1289                          * GROUP BY without aggregation, so insert a group node (plus the
1290                          * appropriate sort node, if necessary).
1291                          *
1292                          * Add an explicit sort if we couldn't make the path come
1293                          * out the way the GROUP node needs it.
1294                          */
1295                         if (!pathkeys_contained_in(group_pathkeys, current_pathkeys))
1296                         {
1297                                 result_plan = (Plan *)
1298                                         make_sort_from_groupcols(parse,
1299                                                                                          parse->groupClause,
1300                                                                                          groupColIdx,
1301                                                                                          result_plan);
1302                                 current_pathkeys = group_pathkeys;
1303                         }
1304
1305                         result_plan = (Plan *) make_group(parse,
1306                                                                                           tlist,
1307                                                                                           (List *) parse->havingQual,
1308                                                                                           numGroupCols,
1309                                                                                           groupColIdx,
1310                                                                                           dNumGroups,
1311                                                                                           result_plan);
1312                         /* The Group node won't change sort ordering */
1313                 }
1314                 else if (parse->hasHavingQual)
1315                 {
1316                         /*
1317                          * No aggregates, and no GROUP BY, but we have a HAVING qual.
1318                          * This is a degenerate case in which we are supposed to emit
1319                          * either 0 or 1 row depending on whether HAVING succeeds.
1320                          * Furthermore, there cannot be any variables in either HAVING
1321                          * or the targetlist, so we actually do not need the FROM table
1322                          * at all!  We can just throw away the plan-so-far and generate
1323                          * a Result node.  This is a sufficiently unusual corner case
1324                          * that it's not worth contorting the structure of this routine
1325                          * to avoid having to generate the plan in the first place.
1326                          */
1327                         result_plan = (Plan *) make_result(tlist,
1328                                                                                            parse->havingQual,
1329                                                                                            NULL);
1330                 }
1331         }                                                       /* end of if (setOperations) */
1332
1333         /*
1334          * If we were not able to make the plan come out in the right order,
1335          * add an explicit sort step.
1336          */
1337         if (parse->sortClause)
1338         {
1339                 if (!pathkeys_contained_in(sort_pathkeys, current_pathkeys))
1340                 {
1341                         result_plan = (Plan *)
1342                                 make_sort_from_sortclauses(parse,
1343                                                                                    parse->sortClause,
1344                                                                                    result_plan);
1345                         current_pathkeys = sort_pathkeys;
1346                 }
1347         }
1348
1349         /*
1350          * If there is a DISTINCT clause, add the UNIQUE node.
1351          */
1352         if (parse->distinctClause)
1353         {
1354                 result_plan = (Plan *) make_unique(result_plan, parse->distinctClause);
1355
1356                 /*
1357                  * If there was grouping or aggregation, leave plan_rows as-is
1358                  * (ie, assume the result was already mostly unique).  If not,
1359                  * it's reasonable to assume the UNIQUE filter has effects
1360                  * comparable to GROUP BY.
1361                  */
1362                 if (!parse->groupClause && !parse->hasHavingQual && !parse->hasAggs)
1363                 {
1364                         List       *distinctExprs;
1365
1366                         distinctExprs = get_sortgrouplist_exprs(parse->distinctClause,
1367                                                                                                         parse->targetList);
1368                         result_plan->plan_rows = estimate_num_groups(parse,
1369                                                                                                                  distinctExprs,
1370                                                                                                  result_plan->plan_rows);
1371                 }
1372         }
1373
1374         /*
1375          * Finally, if there is a LIMIT/OFFSET clause, add the LIMIT node.
1376          */
1377         if (parse->limitOffset || parse->limitCount)
1378         {
1379                 result_plan = (Plan *) make_limit(result_plan,
1380                                                                                   parse->limitOffset,
1381                                                                                   parse->limitCount);
1382         }
1383
1384         /*
1385          * Return the actual output ordering in query_pathkeys for possible
1386          * use by an outer query level.
1387          */
1388         parse->query_pathkeys = current_pathkeys;
1389
1390         return result_plan;
1391 }
1392
1393 /*
1394  * hash_safe_grouping - are grouping operators hashable?
1395  *
1396  * We assume hashed aggregation will work if the datatype's equality operator
1397  * is marked hashjoinable.
1398  */
1399 static bool
1400 hash_safe_grouping(Query *parse)
1401 {
1402         ListCell   *gl;
1403
1404         foreach(gl, parse->groupClause)
1405         {
1406                 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1407                 TargetEntry *tle = get_sortgroupclause_tle(grpcl, parse->targetList);
1408                 Operator        optup;
1409                 bool            oprcanhash;
1410
1411                 optup = equality_oper(tle->resdom->restype, true);
1412                 if (!optup)
1413                         return false;
1414                 oprcanhash = ((Form_pg_operator) GETSTRUCT(optup))->oprcanhash;
1415                 ReleaseSysCache(optup);
1416                 if (!oprcanhash)
1417                         return false;
1418         }
1419         return true;
1420 }
1421
1422 /*---------------
1423  * make_subplanTargetList
1424  *        Generate appropriate target list when grouping is required.
1425  *
1426  * When grouping_planner inserts Aggregate, Group, or Result plan nodes
1427  * above the result of query_planner, we typically want to pass a different
1428  * target list to query_planner than the outer plan nodes should have.
1429  * This routine generates the correct target list for the subplan.
1430  *
1431  * The initial target list passed from the parser already contains entries
1432  * for all ORDER BY and GROUP BY expressions, but it will not have entries
1433  * for variables used only in HAVING clauses; so we need to add those
1434  * variables to the subplan target list.  Also, we flatten all expressions
1435  * except GROUP BY items into their component variables; the other expressions
1436  * will be computed by the inserted nodes rather than by the subplan.
1437  * For example, given a query like
1438  *              SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
1439  * we want to pass this targetlist to the subplan:
1440  *              a,b,c,d,a+b
1441  * where the a+b target will be used by the Sort/Group steps, and the
1442  * other targets will be used for computing the final results.  (In the
1443  * above example we could theoretically suppress the a and b targets and
1444  * pass down only c,d,a+b, but it's not really worth the trouble to
1445  * eliminate simple var references from the subplan.  We will avoid doing
1446  * the extra computation to recompute a+b at the outer level; see
1447  * replace_vars_with_subplan_refs() in setrefs.c.)
1448  *
1449  * If we are grouping or aggregating, *and* there are no non-Var grouping
1450  * expressions, then the returned tlist is effectively dummy; we do not
1451  * need to force it to be evaluated, because all the Vars it contains
1452  * should be present in the output of query_planner anyway.
1453  *
1454  * 'parse' is the query being processed.
1455  * 'tlist' is the query's target list.
1456  * 'groupColIdx' receives an array of column numbers for the GROUP BY
1457  *                      expressions (if there are any) in the subplan's target list.
1458  * 'need_tlist_eval' is set true if we really need to evaluate the
1459  *                      result tlist.
1460  *
1461  * The result is the targetlist to be passed to the subplan.
1462  *---------------
1463  */
1464 static List *
1465 make_subplanTargetList(Query *parse,
1466                                            List *tlist,
1467                                            AttrNumber **groupColIdx,
1468                                            bool *need_tlist_eval)
1469 {
1470         List       *sub_tlist;
1471         List       *extravars;
1472         int                     numCols;
1473
1474         *groupColIdx = NULL;
1475
1476         /*
1477          * If we're not grouping or aggregating, there's nothing to do here;
1478          * query_planner should receive the unmodified target list.
1479          */
1480         if (!parse->hasAggs && !parse->groupClause && !parse->hasHavingQual)
1481         {
1482                 *need_tlist_eval = true;
1483                 return tlist;
1484         }
1485
1486         /*
1487          * Otherwise, start with a "flattened" tlist (having just the vars
1488          * mentioned in the targetlist and HAVING qual --- but not upper-
1489          * level Vars; they will be replaced by Params later on).
1490          */
1491         sub_tlist = flatten_tlist(tlist);
1492         extravars = pull_var_clause(parse->havingQual, false);
1493         sub_tlist = add_to_flat_tlist(sub_tlist, extravars);
1494         list_free(extravars);
1495         *need_tlist_eval = false;       /* only eval if not flat tlist */
1496
1497         /*
1498          * If grouping, create sub_tlist entries for all GROUP BY expressions
1499          * (GROUP BY items that are simple Vars should be in the list
1500          * already), and make an array showing where the group columns are in
1501          * the sub_tlist.
1502          */
1503         numCols = list_length(parse->groupClause);
1504         if (numCols > 0)
1505         {
1506                 int                     keyno = 0;
1507                 AttrNumber *grpColIdx;
1508                 ListCell   *gl;
1509
1510                 grpColIdx = (AttrNumber *) palloc(sizeof(AttrNumber) * numCols);
1511                 *groupColIdx = grpColIdx;
1512
1513                 foreach(gl, parse->groupClause)
1514                 {
1515                         GroupClause *grpcl = (GroupClause *) lfirst(gl);
1516                         Node       *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1517                         TargetEntry *te = NULL;
1518                         ListCell   *sl;
1519
1520                         /* Find or make a matching sub_tlist entry */
1521                         foreach(sl, sub_tlist)
1522                         {
1523                                 te = (TargetEntry *) lfirst(sl);
1524                                 if (equal(groupexpr, te->expr))
1525                                         break;
1526                         }
1527                         if (!sl)
1528                         {
1529                                 te = makeTargetEntry(makeResdom(list_length(sub_tlist) + 1,
1530                                                                                                 exprType(groupexpr),
1531                                                                                                 exprTypmod(groupexpr),
1532                                                                                                 NULL,
1533                                                                                                 false),
1534                                                                          (Expr *) groupexpr);
1535                                 sub_tlist = lappend(sub_tlist, te);
1536                                 *need_tlist_eval = true;                /* it's not flat anymore */
1537                         }
1538
1539                         /* and save its resno */
1540                         grpColIdx[keyno++] = te->resdom->resno;
1541                 }
1542         }
1543
1544         return sub_tlist;
1545 }
1546
1547 /*
1548  * locate_grouping_columns
1549  *              Locate grouping columns in the tlist chosen by query_planner.
1550  *
1551  * This is only needed if we don't use the sub_tlist chosen by
1552  * make_subplanTargetList.      We have to forget the column indexes found
1553  * by that routine and re-locate the grouping vars in the real sub_tlist.
1554  */
1555 static void
1556 locate_grouping_columns(Query *parse,
1557                                                 List *tlist,
1558                                                 List *sub_tlist,
1559                                                 AttrNumber *groupColIdx)
1560 {
1561         int                     keyno = 0;
1562         ListCell   *gl;
1563
1564         /*
1565          * No work unless grouping.
1566          */
1567         if (!parse->groupClause)
1568         {
1569                 Assert(groupColIdx == NULL);
1570                 return;
1571         }
1572         Assert(groupColIdx != NULL);
1573
1574         foreach(gl, parse->groupClause)
1575         {
1576                 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1577                 Node       *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1578                 TargetEntry *te = NULL;
1579                 ListCell   *sl;
1580
1581                 foreach(sl, sub_tlist)
1582                 {
1583                         te = (TargetEntry *) lfirst(sl);
1584                         if (equal(groupexpr, te->expr))
1585                                 break;
1586                 }
1587                 if (!sl)
1588                         elog(ERROR, "failed to locate grouping columns");
1589
1590                 groupColIdx[keyno++] = te->resdom->resno;
1591         }
1592 }
1593
1594 /*
1595  * postprocess_setop_tlist
1596  *        Fix up targetlist returned by plan_set_operations().
1597  *
1598  * We need to transpose sort key info from the orig_tlist into new_tlist.
1599  * NOTE: this would not be good enough if we supported resjunk sort keys
1600  * for results of set operations --- then, we'd need to project a whole
1601  * new tlist to evaluate the resjunk columns.  For now, just ereport if we
1602  * find any resjunk columns in orig_tlist.
1603  */
1604 static List *
1605 postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
1606 {
1607         ListCell   *l;
1608         ListCell   *orig_tlist_item = list_head(orig_tlist);
1609
1610         foreach(l, new_tlist)
1611         {
1612                 TargetEntry *new_tle = (TargetEntry *) lfirst(l);
1613                 TargetEntry *orig_tle;
1614
1615                 /* ignore resjunk columns in setop result */
1616                 if (new_tle->resdom->resjunk)
1617                         continue;
1618
1619                 Assert(orig_tlist_item != NULL);
1620                 orig_tle = (TargetEntry *) lfirst(orig_tlist_item);
1621                 orig_tlist_item = lnext(orig_tlist_item);
1622                 if (orig_tle->resdom->resjunk)  /* should not happen */
1623                         elog(ERROR, "resjunk output columns are not implemented");
1624                 Assert(new_tle->resdom->resno == orig_tle->resdom->resno);
1625                 Assert(new_tle->resdom->restype == orig_tle->resdom->restype);
1626                 new_tle->resdom->ressortgroupref = orig_tle->resdom->ressortgroupref;
1627         }
1628         if (orig_tlist_item != NULL)
1629                 elog(ERROR, "resjunk output columns are not implemented");
1630         return new_tlist;
1631 }