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) 1996-2004, 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 1.175 2004/08/30 02:54:38 momjian 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 "miscadmin.h"
  24 #include "nodes/makefuncs.h"
  25 #ifdef OPTIMIZER_DEBUG
  26 #include "nodes/print.h"
  27 #endif
  28 #include "optimizer/clauses.h"
  29 #include "optimizer/cost.h"
  30 #include "optimizer/pathnode.h"
  31 #include "optimizer/paths.h"
  32 #include "optimizer/planmain.h"
  33 #include "optimizer/planner.h"
  34 #include "optimizer/prep.h"
  35 #include "optimizer/subselect.h"
  36 #include "optimizer/tlist.h"
  37 #include "optimizer/var.h"
  38 #include "parser/analyze.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 hash_safe_grouping(Query *parse);
  62 static List *make_subplanTargetList(Query *parse, List *tlist,
  63                                            AttrNumber **groupColIdx, bool *need_tlist_eval);
  64 static void locate_grouping_columns(Query *parse,
  65                                                 List *tlist,
  66                                                 List *sub_tlist,
  67                                                 AttrNumber *groupColIdx);
  68 static List *postprocess_setop_tlist(List *new_tlist, List *orig_tlist);
  69
  70
  71 /*****************************************************************************
  72  *
  73  *         Query optimizer entry point
  74  *
  75  *****************************************************************************/
  76 Plan *
  77 planner(Query *parse, bool isCursor, int cursorOptions,
  78                 ParamListInfo boundParams)
  79 {
  80         double          tuple_fraction;
  81         Plan       *result_plan;
  82         Index           save_PlannerQueryLevel;
  83         List       *save_PlannerParamList;
  84         ParamListInfo save_PlannerBoundParamList;
  85
  86         /*
  87          * The planner can be called recursively (an example is when
  88          * eval_const_expressions tries to pre-evaluate an SQL function). So,
  89          * these global state variables must be saved and restored.
  90          *
  91          * Query level and the param list cannot be moved into the Query
  92          * structure since their whole purpose is communication across
  93          * multiple sub-Queries. Also, boundParams is explicitly info from
  94          * outside the Query, and so is likewise better handled as a global
  95          * variable.
  96          *
  97          * Note we do NOT save and restore PlannerPlanId: it exists to assign
  98          * unique IDs to SubPlan nodes, and we want those IDs to be unique for
  99          * the life of a backend.  Also, PlannerInitPlan is saved/restored in
 100          * subquery_planner, not here.
 101          */
 102         save_PlannerQueryLevel = PlannerQueryLevel;
 103         save_PlannerParamList = PlannerParamList;
 104         save_PlannerBoundParamList = PlannerBoundParamList;
 105
 106         /* Initialize state for handling outer-level references and params */
 107         PlannerQueryLevel = 0;          /* will be 1 in top-level subquery_planner */
 108         PlannerParamList = NIL;
 109         PlannerBoundParamList = boundParams;
 110
 111         /* Determine what fraction of the plan is likely to be scanned */
 112         if (isCursor)
 113         {
 114                 /*
 115                  * We have no real idea how many tuples the user will ultimately
 116                  * FETCH from a cursor, but it seems a good bet that he doesn't
 117                  * want 'em all.  Optimize for 10% retrieval (you gotta better
 118                  * number?      Should this be a SETtable parameter?)
 119                  */
 120                 tuple_fraction = 0.10;
 121         }
 122         else
 123         {
 124                 /* Default assumption is we need all the tuples */
 125                 tuple_fraction = 0.0;
 126         }
 127
 128         /* primary planning entry point (may recurse for subqueries) */
 129         result_plan = subquery_planner(parse, tuple_fraction);
 130
 131         Assert(PlannerQueryLevel == 0);
 132
 133         /*
 134          * If creating a plan for a scrollable cursor, make sure it can run
 135          * backwards on demand.  Add a Material node at the top at need.
 136          */
 137         if (isCursor && (cursorOptions & CURSOR_OPT_SCROLL))
 138         {
 139                 if (!ExecSupportsBackwardScan(result_plan))
 140                         result_plan = materialize_finished_plan(result_plan);
 141         }
 142
 143         /* executor wants to know total number of Params used overall */
 144         result_plan->nParamExec = list_length(PlannerParamList);
 145
 146         /* final cleanup of the plan */
 147         set_plan_references(result_plan, parse->rtable);
 148
 149         /* restore state for outer planner, if any */
 150         PlannerQueryLevel = save_PlannerQueryLevel;
 151         PlannerParamList = save_PlannerParamList;
 152         PlannerBoundParamList = save_PlannerBoundParamList;
 153
 154         return result_plan;
 155 }
 156
 157
 158 /*--------------------
 159  * subquery_planner
 160  *        Invokes the planner on a subquery.  We recurse to here for each
 161  *        sub-SELECT found in the query tree.
 162  *
 163  * parse is the querytree produced by the parser & rewriter.
 164  * tuple_fraction is the fraction of tuples we expect will be retrieved.
 165  * tuple_fraction is interpreted as explained for grouping_planner, below.
 166  *
 167  * Basically, this routine does the stuff that should only be done once
 168  * per Query object.  It then calls grouping_planner.  At one time,
 169  * grouping_planner could be invoked recursively on the same Query object;
 170  * that's not currently true, but we keep the separation between the two
 171  * routines anyway, in case we need it again someday.
 172  *
 173  * subquery_planner will be called recursively to handle sub-Query nodes
 174  * found within the query's expressions and rangetable.
 175  *
 176  * Returns a query plan.
 177  *--------------------
 178  */
 179 Plan *
 180 subquery_planner(Query *parse, double tuple_fraction)
 181 {
 182         List       *saved_initplan = PlannerInitPlan;
 183         int                     saved_planid = PlannerPlanId;
 184         bool            hasOuterJoins;
 185         Plan       *plan;
 186         List       *newHaving;
 187         List       *lst;
 188         ListCell   *l;
 189
 190         /* Set up for a new level of subquery */
 191         PlannerQueryLevel++;
 192         PlannerInitPlan = NIL;
 193
 194         /*
 195          * Look for IN clauses at the top level of WHERE, and transform them
 196          * into joins.  Note that this step only handles IN clauses originally
 197          * at top level of WHERE; if we pull up any subqueries in the next
 198          * step, their INs are processed just before pulling them up.
 199          */
 200         parse->in_info_list = NIL;
 201         if (parse->hasSubLinks)
 202                 parse->jointree->quals = pull_up_IN_clauses(parse,
 203                                                                                                  parse->jointree->quals);
 204
 205         /*
 206          * Check to see if any subqueries in the rangetable can be merged into
 207          * this query.
 208          */
 209         parse->jointree = (FromExpr *)
 210                 pull_up_subqueries(parse, (Node *) parse->jointree, false);
 211
 212         /*
 213          * Detect whether any rangetable entries are RTE_JOIN kind; if not, we
 214          * can avoid the expense of doing flatten_join_alias_vars().  Also
 215          * check for outer joins --- if none, we can skip
 216          * reduce_outer_joins(). This must be done after we have done
 217          * pull_up_subqueries, of course.
 218          */
 219         parse->hasJoinRTEs = false;
 220         hasOuterJoins = false;
 221         foreach(l, parse->rtable)
 222         {
 223                 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
 224
 225                 if (rte->rtekind == RTE_JOIN)
 226                 {
 227                         parse->hasJoinRTEs = true;
 228                         if (IS_OUTER_JOIN(rte->jointype))
 229                         {
 230                                 hasOuterJoins = true;
 231                                 /* Can quit scanning once we find an outer join */
 232                                 break;
 233                         }
 234                 }
 235         }
 236
 237         /*
 238          * Do expression preprocessing on targetlist and quals.
 239          */
 240         parse->targetList = (List *)
 241                 preprocess_expression(parse, (Node *) parse->targetList,
 242                                                           EXPRKIND_TARGET);
 243
 244         preprocess_qual_conditions(parse, (Node *) parse->jointree);
 245
 246         parse->havingQual = preprocess_expression(parse, parse->havingQual,
 247                                                                                           EXPRKIND_QUAL);
 248
 249         parse->limitOffset = preprocess_expression(parse, parse->limitOffset,
 250                                                                                            EXPRKIND_LIMIT);
 251         parse->limitCount = preprocess_expression(parse, parse->limitCount,
 252                                                                                           EXPRKIND_LIMIT);
 253
 254         parse->in_info_list = (List *)
 255                 preprocess_expression(parse, (Node *) parse->in_info_list,
 256                                                           EXPRKIND_ININFO);
 257
 258         /* Also need to preprocess expressions for function RTEs */
 259         foreach(l, parse->rtable)
 260         {
 261                 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
 262
 263                 if (rte->rtekind == RTE_FUNCTION)
 264                         rte->funcexpr = preprocess_expression(parse, rte->funcexpr,
 265                                                                                                   EXPRKIND_RTFUNC);
 266         }
 267
 268         /*
 269          * A HAVING clause without aggregates is equivalent to a WHERE clause
 270          * (except it can only refer to grouped fields).  Transfer any
 271          * agg-free clauses of the HAVING qual into WHERE.      This may seem like
 272          * wasting cycles to cater to stupidly-written queries, but there are
 273          * other reasons for doing it.  Firstly, if the query contains no aggs
 274          * at all, then we aren't going to generate an Agg plan node, and so
 275          * there'll be no place to execute HAVING conditions; without this
 276          * transfer, we'd lose the HAVING condition entirely, which is wrong.
 277          * Secondly, when we push down a qual condition into a sub-query, it's
 278          * easiest to push the qual into HAVING always, in case it contains
 279          * aggs, and then let this code sort it out.
 280          *
 281          * Note that both havingQual and parse->jointree->quals are in
 282          * implicitly-ANDed-list form at this point, even though they are
 283          * declared as Node *.
 284          */
 285         newHaving = NIL;
 286         foreach(l, (List *) parse->havingQual)
 287         {
 288                 Node       *havingclause = (Node *) lfirst(l);
 289
 290                 if (contain_agg_clause(havingclause))
 291                         newHaving = lappend(newHaving, havingclause);
 292                 else
 293                         parse->jointree->quals = (Node *)
 294                                 lappend((List *) parse->jointree->quals, havingclause);
 295         }
 296         parse->havingQual = (Node *) newHaving;
 297
 298         /*
 299          * If we have any outer joins, try to reduce them to plain inner
 300          * joins. This step is most easily done after we've done expression
 301          * preprocessing.
 302          */
 303         if (hasOuterJoins)
 304                 reduce_outer_joins(parse);
 305
 306         /*
 307          * See if we can simplify the jointree; opportunities for this may
 308          * come from having pulled up subqueries, or from flattening explicit
 309          * JOIN syntax.  We must do this after flattening JOIN alias
 310          * variables, since eliminating explicit JOIN nodes from the jointree
 311          * will cause get_relids_for_join() to fail.  But it should happen
 312          * after reduce_outer_joins, anyway.
 313          */
 314         parse->jointree = (FromExpr *)
 315                 simplify_jointree(parse, (Node *) parse->jointree);
 316
 317         /*
 318          * Do the main planning.  If we have an inherited target relation,
 319          * that needs special processing, else go straight to
 320          * grouping_planner.
 321          */
 322         if (parse->resultRelation &&
 323                 (lst = expand_inherited_rtentry(parse, parse->resultRelation,
 324                                                                                 false)) != NIL)
 325                 plan = inheritance_planner(parse, lst);
 326         else
 327                 plan = grouping_planner(parse, tuple_fraction);
 328
 329         /*
 330          * If any subplans were generated, or if we're inside a subplan, build
 331          * initPlan list and extParam/allParam sets for plan nodes.
 332          */
 333         if (PlannerPlanId != saved_planid || PlannerQueryLevel > 1)
 334         {
 335                 Cost            initplan_cost = 0;
 336
 337                 /* Prepare extParam/allParam sets for all nodes in tree */
 338                 SS_finalize_plan(plan, parse->rtable);
 339
 340                 /*
 341                  * SS_finalize_plan doesn't handle initPlans, so we have to
 342                  * manually attach them to the topmost plan node, and add their
 343                  * extParams to the topmost node's, too.
 344                  *
 345                  * We also add the total_cost of each initPlan to the startup cost of
 346                  * the top node.  This is a conservative overestimate, since in
 347                  * fact each initPlan might be executed later than plan startup,
 348                  * or even not at all.
 349                  */
 350                 plan->initPlan = PlannerInitPlan;
 351
 352                 foreach(l, plan->initPlan)
 353                 {
 354                         SubPlan    *initplan = (SubPlan *) lfirst(l);
 355
 356                         plan->extParam = bms_add_members(plan->extParam,
 357                                                                                          initplan->plan->extParam);
 358                         /* allParam must include all members of extParam */
 359                         plan->allParam = bms_add_members(plan->allParam,
 360                                                                                          plan->extParam);
 361                         initplan_cost += initplan->plan->total_cost;
 362                 }
 363
 364                 plan->startup_cost += initplan_cost;
 365                 plan->total_cost += initplan_cost;
 366         }
 367
 368         /* Return to outer subquery context */
 369         PlannerQueryLevel--;
 370         PlannerInitPlan = saved_initplan;
 371         /* we do NOT restore PlannerPlanId; that's not an oversight! */
 372
 373         return plan;
 374 }
 375
 376 /*
 377  * preprocess_expression
 378  *              Do subquery_planner's preprocessing work for an expression,
 379  *              which can be a targetlist, a WHERE clause (including JOIN/ON
 380  *              conditions), or a HAVING clause.
 381  */
 382 static Node *
 383 preprocess_expression(Query *parse, Node *expr, int kind)
 384 {
 385         /*
 386          * If the query has any join RTEs, replace join alias variables with
 387          * base-relation variables. We must do this before sublink processing,
 388          * else sublinks expanded out from join aliases wouldn't get
 389          * processed.
 390          */
 391         if (parse->hasJoinRTEs)
 392                 expr = flatten_join_alias_vars(parse, expr);
 393
 394         /*
 395          * If it's a qual or havingQual, canonicalize it.  It seems most
 396          * useful to do this before applying eval_const_expressions, since the
 397          * latter can optimize flattened AND/ORs better than unflattened ones.
 398          *
 399          * Note: all processing of a qual expression after this point must be
 400          * careful to maintain AND/OR flatness --- that is, do not generate a
 401          * tree with AND directly under AND, nor OR directly under OR.
 402          */
 403         if (kind == EXPRKIND_QUAL)
 404         {
 405                 expr = (Node *) canonicalize_qual((Expr *) expr);
 406
 407 #ifdef OPTIMIZER_DEBUG
 408                 printf("After canonicalize_qual()\n");
 409                 pprint(expr);
 410 #endif
 411         }
 412
 413         /*
 414          * Simplify constant expressions.
 415          */
 416         expr = eval_const_expressions(expr);
 417
 418         /* Expand SubLinks to SubPlans */
 419         if (parse->hasSubLinks)
 420                 expr = SS_process_sublinks(expr, (kind == EXPRKIND_QUAL));
 421
 422         /*
 423          * XXX do not insert anything here unless you have grokked the
 424          * comments in SS_replace_correlation_vars ...
 425          */
 426
 427         /* Replace uplevel vars with Param nodes */
 428         if (PlannerQueryLevel > 1)
 429                 expr = SS_replace_correlation_vars(expr);
 430
 431         /*
 432          * If it's a qual or havingQual, convert it to implicit-AND format.
 433          * (We don't want to do this before eval_const_expressions, since the
 434          * latter would be unable to simplify a top-level AND correctly. Also,
 435          * SS_process_sublinks expects explicit-AND format.)
 436          */
 437         if (kind == EXPRKIND_QUAL)
 438                 expr = (Node *) make_ands_implicit((Expr *) expr);
 439
 440         return expr;
 441 }
 442
 443 /*
 444  * preprocess_qual_conditions
 445  *              Recursively scan the query's jointree and do subquery_planner's
 446  *              preprocessing work on each qual condition found therein.
 447  */
 448 static void
 449 preprocess_qual_conditions(Query *parse, Node *jtnode)
 450 {
 451         if (jtnode == NULL)
 452                 return;
 453         if (IsA(jtnode, RangeTblRef))
 454         {
 455                 /* nothing to do here */
 456         }
 457         else if (IsA(jtnode, FromExpr))
 458         {
 459                 FromExpr   *f = (FromExpr *) jtnode;
 460                 ListCell   *l;
 461
 462                 foreach(l, f->fromlist)
 463                         preprocess_qual_conditions(parse, lfirst(l));
 464
 465                 f->quals = preprocess_expression(parse, f->quals, EXPRKIND_QUAL);
 466         }
 467         else if (IsA(jtnode, JoinExpr))
 468         {
 469                 JoinExpr   *j = (JoinExpr *) jtnode;
 470
 471                 preprocess_qual_conditions(parse, j->larg);
 472                 preprocess_qual_conditions(parse, j->rarg);
 473
 474                 j->quals = preprocess_expression(parse, j->quals, EXPRKIND_QUAL);
 475         }
 476         else
 477                 elog(ERROR, "unrecognized node type: %d",
 478                          (int) nodeTag(jtnode));
 479 }
 480
 481 /*--------------------
 482  * inheritance_planner
 483  *        Generate a plan in the case where the result relation is an
 484  *        inheritance set.
 485  *
 486  * We have to handle this case differently from cases where a source
 487  * relation is an inheritance set.      Source inheritance is expanded at
 488  * the bottom of the plan tree (see allpaths.c), but target inheritance
 489  * has to be expanded at the top.  The reason is that for UPDATE, each
 490  * target relation needs a different targetlist matching its own column
 491  * set.  (This is not so critical for DELETE, but for simplicity we treat
 492  * inherited DELETE the same way.)      Fortunately, the UPDATE/DELETE target
 493  * can never be the nullable side of an outer join, so it's OK to generate
 494  * the plan this way.
 495  *
 496  * parse is the querytree produced by the parser & rewriter.
 497  * inheritlist is an integer list of RT indexes for the result relation set.
 498  *
 499  * Returns a query plan.
 500  *--------------------
 501  */
 502 static Plan *
 503 inheritance_planner(Query *parse, List *inheritlist)
 504 {
 505         int                     parentRTindex = parse->resultRelation;
 506         Oid                     parentOID = getrelid(parentRTindex, parse->rtable);
 507         int                     mainrtlength = list_length(parse->rtable);
 508         List       *subplans = NIL;
 509         List       *tlist = NIL;
 510         ListCell   *l;
 511
 512         foreach(l, inheritlist)
 513         {
 514                 int                     childRTindex = lfirst_int(l);
 515                 Oid                     childOID = getrelid(childRTindex, parse->rtable);
 516                 int                     subrtlength;
 517                 Query      *subquery;
 518                 Plan       *subplan;
 519
 520                 /* Generate modified query with this rel as target */
 521                 subquery = (Query *) adjust_inherited_attrs((Node *) parse,
 522                                                                                                 parentRTindex, parentOID,
 523                                                                                                  childRTindex, childOID);
 524                 /* Generate plan */
 525                 subplan = grouping_planner(subquery, 0.0 /* retrieve all tuples */ );
 526                 subplans = lappend(subplans, subplan);
 527
 528                 /*
 529                  * It's possible that additional RTEs got added to the rangetable
 530                  * due to expansion of inherited source tables (see allpaths.c).
 531                  * If so, we must copy 'em back to the main parse tree's rtable.
 532                  *
 533                  * XXX my goodness this is ugly.  Really need to think about ways to
 534                  * rein in planner's habit of scribbling on its input.
 535                  */
 536                 subrtlength = list_length(subquery->rtable);
 537                 if (subrtlength > mainrtlength)
 538                 {
 539                         List       *subrt;
 540
 541                         subrt = list_copy_tail(subquery->rtable, mainrtlength);
 542                         parse->rtable = list_concat(parse->rtable, subrt);
 543                         mainrtlength = subrtlength;
 544                 }
 545                 /* Save preprocessed tlist from first rel for use in Append */
 546                 if (tlist == NIL)
 547                         tlist = subplan->targetlist;
 548         }
 549
 550         /* Save the target-relations list for the executor, too */
 551         parse->resultRelations = inheritlist;
 552
 553         /* Mark result as unordered (probably unnecessary) */
 554         parse->query_pathkeys = NIL;
 555
 556         return (Plan *) make_append(subplans, true, tlist);
 557 }
 558
 559 /*--------------------
 560  * grouping_planner
 561  *        Perform planning steps related to grouping, aggregation, etc.
 562  *        This primarily means adding top-level processing to the basic
 563  *        query plan produced by query_planner.
 564  *
 565  * parse is the querytree produced by the parser & rewriter.
 566  * tuple_fraction is the fraction of tuples we expect will be retrieved
 567  *
 568  * tuple_fraction is interpreted as follows:
 569  *        0: expect all tuples to be retrieved (normal case)
 570  *        0 < tuple_fraction < 1: expect the given fraction of tuples available
 571  *              from the plan to be retrieved
 572  *        tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
 573  *              expected to be retrieved (ie, a LIMIT specification)
 574  *
 575  * Returns a query plan.  Also, parse->query_pathkeys is returned as the
 576  * actual output ordering of the plan (in pathkey format).
 577  *--------------------
 578  */
 579 static Plan *
 580 grouping_planner(Query *parse, double tuple_fraction)
 581 {
 582         List       *tlist = parse->targetList;
 583         Plan       *result_plan;
 584         List       *current_pathkeys;
 585         List       *sort_pathkeys;
 586
 587         if (parse->setOperations)
 588         {
 589                 List       *set_sortclauses;
 590
 591                 /*
 592                  * Construct the plan for set operations.  The result will not
 593                  * need any work except perhaps a top-level sort and/or LIMIT.
 594                  */
 595                 result_plan = plan_set_operations(parse,
 596                                                                                   &set_sortclauses);
 597
 598                 /*
 599                  * Calculate pathkeys representing the sort order (if any) of the
 600                  * set operation's result.  We have to do this before overwriting
 601                  * the sort key information...
 602                  */
 603                 current_pathkeys = make_pathkeys_for_sortclauses(set_sortclauses,
 604                                                                                                 result_plan->targetlist);
 605                 current_pathkeys = canonicalize_pathkeys(parse, current_pathkeys);
 606
 607                 /*
 608                  * We should not need to call preprocess_targetlist, since we must
 609                  * be in a SELECT query node.  Instead, use the targetlist
 610                  * returned by plan_set_operations (since this tells whether it
 611                  * returned any resjunk columns!), and transfer any sort key
 612                  * information from the original tlist.
 613                  */
 614                 Assert(parse->commandType == CMD_SELECT);
 615
 616                 tlist = postprocess_setop_tlist(result_plan->targetlist, tlist);
 617
 618                 /*
 619                  * Can't handle FOR UPDATE here (parser should have checked
 620                  * already, but let's make sure).
 621                  */
 622                 if (parse->rowMarks)
 623                         ereport(ERROR,
 624                                         (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
 625                                          errmsg("SELECT FOR UPDATE is not allowed with UNION/INTERSECT/EXCEPT")));
 626
 627                 /*
 628                  * Calculate pathkeys that represent result ordering requirements
 629                  */
 630                 sort_pathkeys = make_pathkeys_for_sortclauses(parse->sortClause,
 631                                                                                                           tlist);
 632                 sort_pathkeys = canonicalize_pathkeys(parse, sort_pathkeys);
 633         }
 634         else
 635         {
 636                 /* No set operations, do regular planning */
 637                 List       *sub_tlist;
 638                 List       *group_pathkeys;
 639                 AttrNumber *groupColIdx = NULL;
 640                 bool            need_tlist_eval = true;
 641                 QualCost        tlist_cost;
 642                 double          sub_tuple_fraction;
 643                 Path       *cheapest_path;
 644                 Path       *sorted_path;
 645                 double          dNumGroups = 0;
 646                 long            numGroups = 0;
 647                 int                     numAggs = 0;
 648                 int                     numGroupCols = list_length(parse->groupClause);
 649                 bool            use_hashed_grouping = false;
 650
 651                 /* Preprocess targetlist in case we are inside an INSERT/UPDATE. */
 652                 tlist = preprocess_targetlist(tlist,
 653                                                                           parse->commandType,
 654                                                                           parse->resultRelation,
 655                                                                           parse->rtable);
 656
 657                 /*
 658                  * Add TID targets for rels selected FOR UPDATE (should this be
 659                  * done in preprocess_targetlist?).  The executor uses the TID to
 660                  * know which rows to lock, much as for UPDATE or DELETE.
 661                  */
 662                 if (parse->rowMarks)
 663                 {
 664                         ListCell   *l;
 665
 666                         /*
 667                          * We've got trouble if the FOR UPDATE appears inside
 668                          * grouping, since grouping renders a reference to individual
 669                          * tuple CTIDs invalid.  This is also checked at parse time,
 670                          * but that's insufficient because of rule substitution, query
 671                          * pullup, etc.
 672                          */
 673                         CheckSelectForUpdate(parse);
 674
 675                         /*
 676                          * Currently the executor only supports FOR UPDATE at top
 677                          * level
 678                          */
 679                         if (PlannerQueryLevel > 1)
 680                                 ereport(ERROR,
 681                                                 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
 682                                                  errmsg("SELECT FOR UPDATE is not allowed in subqueries")));
 683
 684                         foreach(l, parse->rowMarks)
 685                         {
 686                                 Index           rti = lfirst_int(l);
 687                                 char       *resname;
 688                                 Resdom     *resdom;
 689                                 Var                *var;
 690                                 TargetEntry *ctid;
 691
 692                                 resname = (char *) palloc(32);
 693                                 snprintf(resname, 32, "ctid%u", rti);
 694                                 resdom = makeResdom(list_length(tlist) + 1,
 695                                                                         TIDOID,
 696                                                                         -1,
 697                                                                         resname,
 698                                                                         true);
 699
 700                                 var = makeVar(rti,
 701                                                           SelfItemPointerAttributeNumber,
 702                                                           TIDOID,
 703                                                           -1,
 704                                                           0);
 705
 706                                 ctid = makeTargetEntry(resdom, (Expr *) var);
 707                                 tlist = lappend(tlist, ctid);
 708                         }
 709                 }
 710
 711                 /*
 712                  * Generate appropriate target list for subplan; may be different
 713                  * from tlist if grouping or aggregation is needed.
 714                  */
 715                 sub_tlist = make_subplanTargetList(parse, tlist,
 716                                                                                  &groupColIdx, &need_tlist_eval);
 717
 718                 /*
 719                  * Calculate pathkeys that represent grouping/ordering
 720                  * requirements
 721                  */
 722                 group_pathkeys = make_pathkeys_for_sortclauses(parse->groupClause,
 723                                                                                                            tlist);
 724                 sort_pathkeys = make_pathkeys_for_sortclauses(parse->sortClause,
 725                                                                                                           tlist);
 726
 727                 /*
 728                  * Will need actual number of aggregates for estimating costs.
 729                  *
 730                  * Note: we do not attempt to detect duplicate aggregates here; a
 731                  * somewhat-overestimated count is okay for our present purposes.
 732                  *
 733                  * Note: think not that we can turn off hasAggs if we find no aggs.
 734                  * It is possible for constant-expression simplification to remove
 735                  * all explicit references to aggs, but we still have to follow
 736                  * the aggregate semantics (eg, producing only one output row).
 737                  */
 738                 if (parse->hasAggs)
 739                         numAggs = count_agg_clause((Node *) tlist) +
 740                                 count_agg_clause(parse->havingQual);
 741
 742                 /*
 743                  * Figure out whether we need a sorted result from query_planner.
 744                  *
 745                  * If we have a GROUP BY clause, then we want a result sorted
 746                  * properly for grouping.  Otherwise, if there is an ORDER BY
 747                  * clause, we want to sort by the ORDER BY clause.      (Note: if we
 748                  * have both, and ORDER BY is a superset of GROUP BY, it would be
 749                  * tempting to request sort by ORDER BY --- but that might just
 750                  * leave us failing to exploit an available sort order at all.
 751                  * Needs more thought...)
 752                  */
 753                 if (parse->groupClause)
 754                         parse->query_pathkeys = group_pathkeys;
 755                 else if (parse->sortClause)
 756                         parse->query_pathkeys = sort_pathkeys;
 757                 else
 758                         parse->query_pathkeys = NIL;
 759
 760                 /*
 761                  * Adjust tuple_fraction if we see that we are going to apply
 762                  * limiting/grouping/aggregation/etc.  This is not overridable by
 763                  * the caller, since it reflects plan actions that this routine
 764                  * will certainly take, not assumptions about context.
 765                  */
 766                 if (parse->limitCount != NULL)
 767                 {
 768                         /*
 769                          * A LIMIT clause limits the absolute number of tuples
 770                          * returned. However, if it's not a constant LIMIT then we
 771                          * have to punt; for lack of a better idea, assume 10% of the
 772                          * plan's result is wanted.
 773                          */
 774                         double          limit_fraction = 0.0;
 775
 776                         if (IsA(parse->limitCount, Const))
 777                         {
 778                                 Const      *limitc = (Const *) parse->limitCount;
 779                                 int32           count = DatumGetInt32(limitc->constvalue);
 780
 781                                 /*
 782                                  * A NULL-constant LIMIT represents "LIMIT ALL", which we
 783                                  * treat the same as no limit (ie, expect to retrieve all
 784                                  * the tuples).
 785                                  */
 786                                 if (!limitc->constisnull && count > 0)
 787                                 {
 788                                         limit_fraction = (double) count;
 789                                         /* We must also consider the OFFSET, if present */
 790                                         if (parse->limitOffset != NULL)
 791                                         {
 792                                                 if (IsA(parse->limitOffset, Const))
 793                                                 {
 794                                                         int32           offset;
 795
 796                                                         limitc = (Const *) parse->limitOffset;
 797                                                         offset = DatumGetInt32(limitc->constvalue);
 798                                                         if (!limitc->constisnull && offset > 0)
 799                                                                 limit_fraction += (double) offset;
 800                                                 }
 801                                                 else
 802                                                 {
 803                                                         /* OFFSET is an expression ... punt ... */
 804                                                         limit_fraction = 0.10;
 805                                                 }
 806                                         }
 807                                 }
 808                         }
 809                         else
 810                         {
 811                                 /* LIMIT is an expression ... punt ... */
 812                                 limit_fraction = 0.10;
 813                         }
 814
 815                         if (limit_fraction > 0.0)
 816                         {
 817                                 /*
 818                                  * If we have absolute limits from both caller and LIMIT,
 819                                  * use the smaller value; if one is fractional and the
 820                                  * other absolute, treat the fraction as a fraction of the
 821                                  * absolute value; else we can multiply the two fractions
 822                                  * together.
 823                                  */
 824                                 if (tuple_fraction >= 1.0)
 825                                 {
 826                                         if (limit_fraction >= 1.0)
 827                                         {
 828                                                 /* both absolute */
 829                                                 tuple_fraction = Min(tuple_fraction, limit_fraction);
 830                                         }
 831                                         else
 832                                         {
 833                                                 /* caller absolute, limit fractional */
 834                                                 tuple_fraction *= limit_fraction;
 835                                                 if (tuple_fraction < 1.0)
 836                                                         tuple_fraction = 1.0;
 837                                         }
 838                                 }
 839                                 else if (tuple_fraction > 0.0)
 840                                 {
 841                                         if (limit_fraction >= 1.0)
 842                                         {
 843                                                 /* caller fractional, limit absolute */
 844                                                 tuple_fraction *= limit_fraction;
 845                                                 if (tuple_fraction < 1.0)
 846                                                         tuple_fraction = 1.0;
 847                                         }
 848                                         else
 849                                         {
 850                                                 /* both fractional */
 851                                                 tuple_fraction *= limit_fraction;
 852                                         }
 853                                 }
 854                                 else
 855                                 {
 856                                         /* no info from caller, just use limit */
 857                                         tuple_fraction = limit_fraction;
 858                                 }
 859                         }
 860                 }
 861
 862                 /*
 863                  * With grouping or aggregation, the tuple fraction to pass to
 864                  * query_planner() may be different from what it is at top level.
 865                  */
 866                 sub_tuple_fraction = tuple_fraction;
 867
 868                 if (parse->groupClause)
 869                 {
 870                         /*
 871                          * In GROUP BY mode, we have the little problem that we don't
 872                          * really know how many input tuples will be needed to make a
 873                          * group, so we can't translate an output LIMIT count into an
 874                          * input count.  For lack of a better idea, assume 25% of the
 875                          * input data will be processed if there is any output limit.
 876                          * However, if the caller gave us a fraction rather than an
 877                          * absolute count, we can keep using that fraction (which
 878                          * amounts to assuming that all the groups are about the same
 879                          * size).
 880                          */
 881                         if (sub_tuple_fraction >= 1.0)
 882                                 sub_tuple_fraction = 0.25;
 883
 884                         /*
 885                          * If both GROUP BY and ORDER BY are specified, we will need
 886                          * two levels of sort --- and, therefore, certainly need to
 887                          * read all the input tuples --- unless ORDER BY is a subset
 888                          * of GROUP BY.  (We have not yet canonicalized the pathkeys,
 889                          * so must use the slower noncanonical comparison method.)
 890                          */
 891                         if (parse->groupClause && parse->sortClause &&
 892                                 !noncanonical_pathkeys_contained_in(sort_pathkeys,
 893                                                                                                         group_pathkeys))
 894                                 sub_tuple_fraction = 0.0;
 895                 }
 896                 else if (parse->hasAggs)
 897                 {
 898                         /*
 899                          * Ungrouped aggregate will certainly want all the input
 900                          * tuples.
 901                          */
 902                         sub_tuple_fraction = 0.0;
 903                 }
 904                 else if (parse->distinctClause)
 905                 {
 906                         /*
 907                          * SELECT DISTINCT, like GROUP, will absorb an unpredictable
 908                          * number of input tuples per output tuple.  Handle the same
 909                          * way.
 910                          */
 911                         if (sub_tuple_fraction >= 1.0)
 912                                 sub_tuple_fraction = 0.25;
 913                 }
 914
 915                 /*
 916                  * Generate the best unsorted and presorted paths for this Query
 917                  * (but note there may not be any presorted path).
 918                  */
 919                 query_planner(parse, sub_tlist, sub_tuple_fraction,
 920                                           &cheapest_path, &sorted_path);
 921
 922                 /*
 923                  * We couldn't canonicalize group_pathkeys and sort_pathkeys
 924                  * before running query_planner(), so do it now.
 925                  */
 926                 group_pathkeys = canonicalize_pathkeys(parse, group_pathkeys);
 927                 sort_pathkeys = canonicalize_pathkeys(parse, sort_pathkeys);
 928
 929                 /*
 930                  * Consider whether we might want to use hashed grouping.
 931                  */
 932                 if (parse->groupClause)
 933                 {
 934                         List       *groupExprs;
 935                         double          cheapest_path_rows;
 936                         int                     cheapest_path_width;
 937
 938                         /*
 939                          * Beware in this section of the possibility that
 940                          * cheapest_path->parent is NULL.  This could happen if user
 941                          * does something silly like SELECT 'foo' GROUP BY 1;
 942                          */
 943                         if (cheapest_path->parent)
 944                         {
 945                                 cheapest_path_rows = cheapest_path->parent->rows;
 946                                 cheapest_path_width = cheapest_path->parent->width;
 947                         }
 948                         else
 949                         {
 950                                 cheapest_path_rows = 1; /* assume non-set result */
 951                                 cheapest_path_width = 100;              /* arbitrary */
 952                         }
 953
 954                         /*
 955                          * Always estimate the number of groups.  We can't do this
 956                          * until after running query_planner(), either.
 957                          */
 958                         groupExprs = get_sortgrouplist_exprs(parse->groupClause,
 959                                                                                                  parse->targetList);
 960                         dNumGroups = estimate_num_groups(parse,
 961                                                                                          groupExprs,
 962                                                                                          cheapest_path_rows);
 963                         /* Also want it as a long int --- but 'ware overflow! */
 964                         numGroups = (long) Min(dNumGroups, (double) LONG_MAX);
 965
 966                         /*
 967                          * Check can't-do-it conditions, including whether the
 968                          * grouping operators are hashjoinable.
 969                          *
 970                          * Executor doesn't support hashed aggregation with DISTINCT
 971                          * aggregates.  (Doing so would imply storing *all* the input
 972                          * values in the hash table, which seems like a certain
 973                          * loser.)
 974                          */
 975                         if (!enable_hashagg || !hash_safe_grouping(parse))
 976                                 use_hashed_grouping = false;
 977                         else if (parse->hasAggs &&
 978                                          (contain_distinct_agg_clause((Node *) tlist) ||
 979                                           contain_distinct_agg_clause(parse->havingQual)))
 980                                 use_hashed_grouping = false;
 981                         else
 982                         {
 983                                 /*
 984                                  * Use hashed grouping if (a) we think we can fit the
 985                                  * hashtable into work_mem, *and* (b) the estimated cost
 986                                  * is no more than doing it the other way.      While avoiding
 987                                  * the need for sorted input is usually a win, the fact
 988                                  * that the output won't be sorted may be a loss; so we
 989                                  * need to do an actual cost comparison.
 990                                  *
 991                                  * In most cases we have no good way to estimate the size of
 992                                  * the transition value needed by an aggregate;
 993                                  * arbitrarily assume it is 100 bytes.  Also set the
 994                                  * overhead per hashtable entry at 64 bytes.
 995                                  */
 996                                 int                     hashentrysize = cheapest_path_width + 64 + numAggs * 100;
 997
 998                                 if (hashentrysize * dNumGroups <= work_mem * 1024L)
 999                                 {
1000                                         /*
1001                                          * Okay, do the cost comparison.  We need to consider
1002                                          * cheapest_path + hashagg [+ final sort] versus
1003                                          * either cheapest_path [+ sort] + group or agg [+
1004                                          * final sort] or presorted_path + group or agg [+
1005                                          * final sort] where brackets indicate a step that may
1006                                          * not be needed. We assume query_planner() will have
1007                                          * returned a presorted path only if it's a winner
1008                                          * compared to cheapest_path for this purpose.
1009                                          *
1010                                          * These path variables are dummies that just hold cost
1011                                          * fields; we don't make actual Paths for these steps.
1012                                          */
1013                                         Path            hashed_p;
1014                                         Path            sorted_p;
1015
1016                                         cost_agg(&hashed_p, parse,
1017                                                          AGG_HASHED, numAggs,
1018                                                          numGroupCols, dNumGroups,
1019                                                          cheapest_path->startup_cost,
1020                                                          cheapest_path->total_cost,
1021                                                          cheapest_path_rows);
1022                                         /* Result of hashed agg is always unsorted */
1023                                         if (sort_pathkeys)
1024                                                 cost_sort(&hashed_p, parse, sort_pathkeys,
1025                                                                   hashed_p.total_cost,
1026                                                                   dNumGroups,
1027                                                                   cheapest_path_width);
1028
1029                                         if (sorted_path)
1030                                         {
1031                                                 sorted_p.startup_cost = sorted_path->startup_cost;
1032                                                 sorted_p.total_cost = sorted_path->total_cost;
1033                                                 current_pathkeys = sorted_path->pathkeys;
1034                                         }
1035                                         else
1036                                         {
1037                                                 sorted_p.startup_cost = cheapest_path->startup_cost;
1038                                                 sorted_p.total_cost = cheapest_path->total_cost;
1039                                                 current_pathkeys = cheapest_path->pathkeys;
1040                                         }
1041                                         if (!pathkeys_contained_in(group_pathkeys,
1042                                                                                            current_pathkeys))
1043                                         {
1044                                                 cost_sort(&sorted_p, parse, group_pathkeys,
1045                                                                   sorted_p.total_cost,
1046                                                                   cheapest_path_rows,
1047                                                                   cheapest_path_width);
1048                                                 current_pathkeys = group_pathkeys;
1049                                         }
1050                                         if (parse->hasAggs)
1051                                                 cost_agg(&sorted_p, parse,
1052                                                                  AGG_SORTED, numAggs,
1053                                                                  numGroupCols, dNumGroups,
1054                                                                  sorted_p.startup_cost,
1055                                                                  sorted_p.total_cost,
1056                                                                  cheapest_path_rows);
1057                                         else
1058                                                 cost_group(&sorted_p, parse,
1059                                                                    numGroupCols, dNumGroups,
1060                                                                    sorted_p.startup_cost,
1061                                                                    sorted_p.total_cost,
1062                                                                    cheapest_path_rows);
1063                                         /* The Agg or Group node will preserve ordering */
1064                                         if (sort_pathkeys &&
1065                                                 !pathkeys_contained_in(sort_pathkeys,
1066                                                                                            current_pathkeys))
1067                                         {
1068                                                 cost_sort(&sorted_p, parse, sort_pathkeys,
1069                                                                   sorted_p.total_cost,
1070                                                                   dNumGroups,
1071                                                                   cheapest_path_width);
1072                                         }
1073
1074                                         /*
1075                                          * Now make the decision using the top-level tuple
1076                                          * fraction.  First we have to convert an absolute
1077                                          * count (LIMIT) into fractional form.
1078                                          */
1079                                         if (tuple_fraction >= 1.0)
1080                                                 tuple_fraction /= dNumGroups;
1081
1082                                         if (compare_fractional_path_costs(&hashed_p, &sorted_p,
1083                                                                                                           tuple_fraction) < 0)
1084                                         {
1085                                                 /* Hashed is cheaper, so use it */
1086                                                 use_hashed_grouping = true;
1087                                         }
1088                                 }
1089                         }
1090                 }
1091
1092                 /*
1093                  * Select the best path and create a plan to execute it.
1094                  *
1095                  * If we are doing hashed grouping, we will always read all the input
1096                  * tuples, so use the cheapest-total path.      Otherwise, trust
1097                  * query_planner's decision about which to use.
1098                  */
1099                 if (sorted_path && !use_hashed_grouping)
1100                 {
1101                         result_plan = create_plan(parse, sorted_path);
1102                         current_pathkeys = sorted_path->pathkeys;
1103                 }
1104                 else
1105                 {
1106                         result_plan = create_plan(parse, cheapest_path);
1107                         current_pathkeys = cheapest_path->pathkeys;
1108                 }
1109
1110                 /*
1111                  * create_plan() returns a plan with just a "flat" tlist of
1112                  * required Vars.  Usually we need to insert the sub_tlist as the
1113                  * tlist of the top plan node.  However, we can skip that if we
1114                  * determined that whatever query_planner chose to return will be
1115                  * good enough.
1116                  */
1117                 if (need_tlist_eval)
1118                 {
1119                         /*
1120                          * If the top-level plan node is one that cannot do expression
1121                          * evaluation, we must insert a Result node to project the
1122                          * desired tlist.
1123                          */
1124                         if (!is_projection_capable_plan(result_plan))
1125                         {
1126                                 result_plan = (Plan *) make_result(sub_tlist, NULL,
1127                                                                                                    result_plan);
1128                         }
1129                         else
1130                         {
1131                                 /*
1132                                  * Otherwise, just replace the subplan's flat tlist with
1133                                  * the desired tlist.
1134                                  */
1135                                 result_plan->targetlist = sub_tlist;
1136                         }
1137
1138                         /*
1139                          * Also, account for the cost of evaluation of the sub_tlist.
1140                          *
1141                          * Up to now, we have only been dealing with "flat" tlists,
1142                          * containing just Vars.  So their evaluation cost is zero
1143                          * according to the model used by cost_qual_eval() (or if you
1144                          * prefer, the cost is factored into cpu_tuple_cost).  Thus we
1145                          * can avoid accounting for tlist cost throughout
1146                          * query_planner() and subroutines.  But now we've inserted a
1147                          * tlist that might contain actual operators, sub-selects, etc
1148                          * --- so we'd better account for its cost.
1149                          *
1150                          * Below this point, any tlist eval cost for added-on nodes
1151                          * should be accounted for as we create those nodes.
1152                          * Presently, of the node types we can add on, only Agg and
1153                          * Group project new tlists (the rest just copy their input
1154                          * tuples) --- so make_agg() and make_group() are responsible
1155                          * for computing the added cost.
1156                          */
1157                         cost_qual_eval(&tlist_cost, sub_tlist);
1158                         result_plan->startup_cost += tlist_cost.startup;
1159                         result_plan->total_cost += tlist_cost.startup +
1160                                 tlist_cost.per_tuple * result_plan->plan_rows;
1161                 }
1162                 else
1163                 {
1164                         /*
1165                          * Since we're using query_planner's tlist and not the one
1166                          * make_subplanTargetList calculated, we have to refigure any
1167                          * grouping-column indexes make_subplanTargetList computed.
1168                          */
1169                         locate_grouping_columns(parse, tlist, result_plan->targetlist,
1170                                                                         groupColIdx);
1171                 }
1172
1173                 /*
1174                  * Insert AGG or GROUP node if needed, plus an explicit sort step
1175                  * if necessary.
1176                  *
1177                  * HAVING clause, if any, becomes qual of the Agg node
1178                  */
1179                 if (use_hashed_grouping)
1180                 {
1181                         /* Hashed aggregate plan --- no sort needed */
1182                         result_plan = (Plan *) make_agg(parse,
1183                                                                                         tlist,
1184                                                                                         (List *) parse->havingQual,
1185                                                                                         AGG_HASHED,
1186                                                                                         numGroupCols,
1187                                                                                         groupColIdx,
1188                                                                                         numGroups,
1189                                                                                         numAggs,
1190                                                                                         result_plan);
1191                         /* Hashed aggregation produces randomly-ordered results */
1192                         current_pathkeys = NIL;
1193                 }
1194                 else if (parse->hasAggs)
1195                 {
1196                         /* Plain aggregate plan --- sort if needed */
1197                         AggStrategy aggstrategy;
1198
1199                         if (parse->groupClause)
1200                         {
1201                                 if (!pathkeys_contained_in(group_pathkeys, current_pathkeys))
1202                                 {
1203                                         result_plan = (Plan *)
1204                                                 make_sort_from_groupcols(parse,
1205                                                                                                  parse->groupClause,
1206                                                                                                  groupColIdx,
1207                                                                                                  result_plan);
1208                                         current_pathkeys = group_pathkeys;
1209                                 }
1210                                 aggstrategy = AGG_SORTED;
1211
1212                                 /*
1213                                  * The AGG node will not change the sort ordering of its
1214                                  * groups, so current_pathkeys describes the result too.
1215                                  */
1216                         }
1217                         else
1218                         {
1219                                 aggstrategy = AGG_PLAIN;
1220                                 /* Result will be only one row anyway; no sort order */
1221                                 current_pathkeys = NIL;
1222                         }
1223
1224                         result_plan = (Plan *) make_agg(parse,
1225                                                                                         tlist,
1226                                                                                         (List *) parse->havingQual,
1227                                                                                         aggstrategy,
1228                                                                                         numGroupCols,
1229                                                                                         groupColIdx,
1230                                                                                         numGroups,
1231                                                                                         numAggs,
1232                                                                                         result_plan);
1233                 }
1234                 else
1235                 {
1236                         /*
1237                          * If there are no Aggs, we shouldn't have any HAVING qual
1238                          * anymore
1239                          */
1240                         Assert(parse->havingQual == NULL);
1241
1242                         /*
1243                          * If we have a GROUP BY clause, insert a group node (plus the
1244                          * appropriate sort node, if necessary).
1245                          */
1246                         if (parse->groupClause)
1247                         {
1248                                 /*
1249                                  * Add an explicit sort if we couldn't make the path come
1250                                  * out the way the GROUP node needs it.
1251                                  */
1252                                 if (!pathkeys_contained_in(group_pathkeys, current_pathkeys))
1253                                 {
1254                                         result_plan = (Plan *)
1255                                                 make_sort_from_groupcols(parse,
1256                                                                                                  parse->groupClause,
1257                                                                                                  groupColIdx,
1258                                                                                                  result_plan);
1259                                         current_pathkeys = group_pathkeys;
1260                                 }
1261
1262                                 result_plan = (Plan *) make_group(parse,
1263                                                                                                   tlist,
1264                                                                                                   numGroupCols,
1265                                                                                                   groupColIdx,
1266                                                                                                   dNumGroups,
1267                                                                                                   result_plan);
1268                                 /* The Group node won't change sort ordering */
1269                         }
1270                 }
1271         }                                                       /* end of if (setOperations) */
1272
1273         /*
1274          * If we were not able to make the plan come out in the right order,
1275          * add an explicit sort step.
1276          */
1277         if (parse->sortClause)
1278         {
1279                 if (!pathkeys_contained_in(sort_pathkeys, current_pathkeys))
1280                 {
1281                         result_plan = (Plan *)
1282                                 make_sort_from_sortclauses(parse,
1283                                                                                    parse->sortClause,
1284                                                                                    result_plan);
1285                         current_pathkeys = sort_pathkeys;
1286                 }
1287         }
1288
1289         /*
1290          * If there is a DISTINCT clause, add the UNIQUE node.
1291          */
1292         if (parse->distinctClause)
1293         {
1294                 result_plan = (Plan *) make_unique(result_plan, parse->distinctClause);
1295
1296                 /*
1297                  * If there was grouping or aggregation, leave plan_rows as-is
1298                  * (ie, assume the result was already mostly unique).  If not,
1299                  * it's reasonable to assume the UNIQUE filter has effects
1300                  * comparable to GROUP BY.
1301                  */
1302                 if (!parse->groupClause && !parse->hasAggs)
1303                 {
1304                         List       *distinctExprs;
1305
1306                         distinctExprs = get_sortgrouplist_exprs(parse->distinctClause,
1307                                                                                                         parse->targetList);
1308                         result_plan->plan_rows = estimate_num_groups(parse,
1309                                                                                                                  distinctExprs,
1310                                                                                                  result_plan->plan_rows);
1311                 }
1312         }
1313
1314         /*
1315          * Finally, if there is a LIMIT/OFFSET clause, add the LIMIT node.
1316          */
1317         if (parse->limitOffset || parse->limitCount)
1318         {
1319                 result_plan = (Plan *) make_limit(result_plan,
1320                                                                                   parse->limitOffset,
1321                                                                                   parse->limitCount);
1322         }
1323
1324         /*
1325          * Return the actual output ordering in query_pathkeys for possible
1326          * use by an outer query level.
1327          */
1328         parse->query_pathkeys = current_pathkeys;
1329
1330         return result_plan;
1331 }
1332
1333 /*
1334  * hash_safe_grouping - are grouping operators hashable?
1335  *
1336  * We assume hashed aggregation will work if the datatype's equality operator
1337  * is marked hashjoinable.
1338  */
1339 static bool
1340 hash_safe_grouping(Query *parse)
1341 {
1342         ListCell   *gl;
1343
1344         foreach(gl, parse->groupClause)
1345         {
1346                 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1347                 TargetEntry *tle = get_sortgroupclause_tle(grpcl, parse->targetList);
1348                 Operator        optup;
1349                 bool            oprcanhash;
1350
1351                 optup = equality_oper(tle->resdom->restype, true);
1352                 if (!optup)
1353                         return false;
1354                 oprcanhash = ((Form_pg_operator) GETSTRUCT(optup))->oprcanhash;
1355                 ReleaseSysCache(optup);
1356                 if (!oprcanhash)
1357                         return false;
1358         }
1359         return true;
1360 }
1361
1362 /*---------------
1363  * make_subplanTargetList
1364  *        Generate appropriate target list when grouping is required.
1365  *
1366  * When grouping_planner inserts Aggregate or Group plan nodes above
1367  * the result of query_planner, we typically want to pass a different
1368  * target list to query_planner than the outer plan nodes should have.
1369  * This routine generates the correct target list for the subplan.
1370  *
1371  * The initial target list passed from the parser already contains entries
1372  * for all ORDER BY and GROUP BY expressions, but it will not have entries
1373  * for variables used only in HAVING clauses; so we need to add those
1374  * variables to the subplan target list.  Also, if we are doing either
1375  * grouping or aggregation, we flatten all expressions except GROUP BY items
1376  * into their component variables; the other expressions will be computed by
1377  * the inserted nodes rather than by the subplan.  For example,
1378  * given a query like
1379  *              SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
1380  * we want to pass this targetlist to the subplan:
1381  *              a,b,c,d,a+b
1382  * where the a+b target will be used by the Sort/Group steps, and the
1383  * other targets will be used for computing the final results.  (In the
1384  * above example we could theoretically suppress the a and b targets and
1385  * pass down only c,d,a+b, but it's not really worth the trouble to
1386  * eliminate simple var references from the subplan.  We will avoid doing
1387  * the extra computation to recompute a+b at the outer level; see
1388  * replace_vars_with_subplan_refs() in setrefs.c.)
1389  *
1390  * If we are grouping or aggregating, *and* there are no non-Var grouping
1391  * expressions, then the returned tlist is effectively dummy; we do not
1392  * need to force it to be evaluated, because all the Vars it contains
1393  * should be present in the output of query_planner anyway.
1394  *
1395  * 'parse' is the query being processed.
1396  * 'tlist' is the query's target list.
1397  * 'groupColIdx' receives an array of column numbers for the GROUP BY
1398  *                      expressions (if there are any) in the subplan's target list.
1399  * 'need_tlist_eval' is set true if we really need to evaluate the
1400  *                      result tlist.
1401  *
1402  * The result is the targetlist to be passed to the subplan.
1403  *---------------
1404  */
1405 static List *
1406 make_subplanTargetList(Query *parse,
1407                                            List *tlist,
1408                                            AttrNumber **groupColIdx,
1409                                            bool *need_tlist_eval)
1410 {
1411         List       *sub_tlist;
1412         List       *extravars;
1413         int                     numCols;
1414
1415         *groupColIdx = NULL;
1416
1417         /*
1418          * If we're not grouping or aggregating, nothing to do here;
1419          * query_planner should receive the unmodified target list.
1420          */
1421         if (!parse->hasAggs && !parse->groupClause)
1422         {
1423                 *need_tlist_eval = true;
1424                 return tlist;
1425         }
1426
1427         /*
1428          * Otherwise, start with a "flattened" tlist (having just the vars
1429          * mentioned in the targetlist and HAVING qual --- but not upper-
1430          * level Vars; they will be replaced by Params later on).
1431          */
1432         sub_tlist = flatten_tlist(tlist);
1433         extravars = pull_var_clause(parse->havingQual, false);
1434         sub_tlist = add_to_flat_tlist(sub_tlist, extravars);
1435         list_free(extravars);
1436         *need_tlist_eval = false;       /* only eval if not flat tlist */
1437
1438         /*
1439          * If grouping, create sub_tlist entries for all GROUP BY expressions
1440          * (GROUP BY items that are simple Vars should be in the list
1441          * already), and make an array showing where the group columns are in
1442          * the sub_tlist.
1443          */
1444         numCols = list_length(parse->groupClause);
1445         if (numCols > 0)
1446         {
1447                 int                     keyno = 0;
1448                 AttrNumber *grpColIdx;
1449                 ListCell   *gl;
1450
1451                 grpColIdx = (AttrNumber *) palloc(sizeof(AttrNumber) * numCols);
1452                 *groupColIdx = grpColIdx;
1453
1454                 foreach(gl, parse->groupClause)
1455                 {
1456                         GroupClause *grpcl = (GroupClause *) lfirst(gl);
1457                         Node       *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1458                         TargetEntry *te = NULL;
1459                         ListCell   *sl;
1460
1461                         /* Find or make a matching sub_tlist entry */
1462                         foreach(sl, sub_tlist)
1463                         {
1464                                 te = (TargetEntry *) lfirst(sl);
1465                                 if (equal(groupexpr, te->expr))
1466                                         break;
1467                         }
1468                         if (!sl)
1469                         {
1470                                 te = makeTargetEntry(makeResdom(list_length(sub_tlist) + 1,
1471                                                                                                 exprType(groupexpr),
1472                                                                                                 exprTypmod(groupexpr),
1473                                                                                                 NULL,
1474                                                                                                 false),
1475                                                                          (Expr *) groupexpr);
1476                                 sub_tlist = lappend(sub_tlist, te);
1477                                 *need_tlist_eval = true;                /* it's not flat anymore */
1478                         }
1479
1480                         /* and save its resno */
1481                         grpColIdx[keyno++] = te->resdom->resno;
1482                 }
1483         }
1484
1485         return sub_tlist;
1486 }
1487
1488 /*
1489  * locate_grouping_columns
1490  *              Locate grouping columns in the tlist chosen by query_planner.
1491  *
1492  * This is only needed if we don't use the sub_tlist chosen by
1493  * make_subplanTargetList.      We have to forget the column indexes found
1494  * by that routine and re-locate the grouping vars in the real sub_tlist.
1495  */
1496 static void
1497 locate_grouping_columns(Query *parse,
1498                                                 List *tlist,
1499                                                 List *sub_tlist,
1500                                                 AttrNumber *groupColIdx)
1501 {
1502         int                     keyno = 0;
1503         ListCell   *gl;
1504
1505         /*
1506          * No work unless grouping.
1507          */
1508         if (!parse->groupClause)
1509         {
1510                 Assert(groupColIdx == NULL);
1511                 return;
1512         }
1513         Assert(groupColIdx != NULL);
1514
1515         foreach(gl, parse->groupClause)
1516         {
1517                 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1518                 Node       *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1519                 TargetEntry *te = NULL;
1520                 ListCell   *sl;
1521
1522                 foreach(sl, sub_tlist)
1523                 {
1524                         te = (TargetEntry *) lfirst(sl);
1525                         if (equal(groupexpr, te->expr))
1526                                 break;
1527                 }
1528                 if (!sl)
1529                         elog(ERROR, "failed to locate grouping columns");
1530
1531                 groupColIdx[keyno++] = te->resdom->resno;
1532         }
1533 }
1534
1535 /*
1536  * postprocess_setop_tlist
1537  *        Fix up targetlist returned by plan_set_operations().
1538  *
1539  * We need to transpose sort key info from the orig_tlist into new_tlist.
1540  * NOTE: this would not be good enough if we supported resjunk sort keys
1541  * for results of set operations --- then, we'd need to project a whole
1542  * new tlist to evaluate the resjunk columns.  For now, just ereport if we
1543  * find any resjunk columns in orig_tlist.
1544  */
1545 static List *
1546 postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
1547 {
1548         ListCell   *l;
1549         ListCell   *orig_tlist_item = list_head(orig_tlist);
1550
1551         foreach(l, new_tlist)
1552         {
1553                 TargetEntry *new_tle = (TargetEntry *) lfirst(l);
1554                 TargetEntry *orig_tle;
1555
1556                 /* ignore resjunk columns in setop result */
1557                 if (new_tle->resdom->resjunk)
1558                         continue;
1559
1560                 Assert(orig_tlist_item != NULL);
1561                 orig_tle = (TargetEntry *) lfirst(orig_tlist_item);
1562                 orig_tlist_item = lnext(orig_tlist_item);
1563                 if (orig_tle->resdom->resjunk)  /* should not happen */
1564                         elog(ERROR, "resjunk output columns are not implemented");
1565                 Assert(new_tle->resdom->resno == orig_tle->resdom->resno);
1566                 Assert(new_tle->resdom->restype == orig_tle->resdom->restype);
1567                 new_tle->resdom->ressortgroupref = orig_tle->resdom->ressortgroupref;
1568         }
1569         if (orig_tlist_item != NULL)
1570                 elog(ERROR, "resjunk output columns are not implemented");
1571         return new_tlist;
1572 }