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 1.176 2004/10/02 22:39:46 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 "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)) != NIL)
 324                 plan = inheritance_planner(parse, lst);
 325         else
 326                 plan = grouping_planner(parse, tuple_fraction);
 327
 328         /*
 329          * If any subplans were generated, or if we're inside a subplan, build
 330          * initPlan list and extParam/allParam sets for plan nodes.
 331          */
 332         if (PlannerPlanId != saved_planid || PlannerQueryLevel > 1)
 333         {
 334                 Cost            initplan_cost = 0;
 335
 336                 /* Prepare extParam/allParam sets for all nodes in tree */
 337                 SS_finalize_plan(plan, parse->rtable);
 338
 339                 /*
 340                  * SS_finalize_plan doesn't handle initPlans, so we have to
 341                  * manually attach them to the topmost plan node, and add their
 342                  * extParams to the topmost node's, too.
 343                  *
 344                  * We also add the total_cost of each initPlan to the startup cost of
 345                  * the top node.  This is a conservative overestimate, since in
 346                  * fact each initPlan might be executed later than plan startup,
 347                  * or even not at all.
 348                  */
 349                 plan->initPlan = PlannerInitPlan;
 350
 351                 foreach(l, plan->initPlan)
 352                 {
 353                         SubPlan    *initplan = (SubPlan *) lfirst(l);
 354
 355                         plan->extParam = bms_add_members(plan->extParam,
 356                                                                                          initplan->plan->extParam);
 357                         /* allParam must include all members of extParam */
 358                         plan->allParam = bms_add_members(plan->allParam,
 359                                                                                          plan->extParam);
 360                         initplan_cost += initplan->plan->total_cost;
 361                 }
 362
 363                 plan->startup_cost += initplan_cost;
 364                 plan->total_cost += initplan_cost;
 365         }
 366
 367         /* Return to outer subquery context */
 368         PlannerQueryLevel--;
 369         PlannerInitPlan = saved_initplan;
 370         /* we do NOT restore PlannerPlanId; that's not an oversight! */
 371
 372         return plan;
 373 }
 374
 375 /*
 376  * preprocess_expression
 377  *              Do subquery_planner's preprocessing work for an expression,
 378  *              which can be a targetlist, a WHERE clause (including JOIN/ON
 379  *              conditions), or a HAVING clause.
 380  */
 381 static Node *
 382 preprocess_expression(Query *parse, Node *expr, int kind)
 383 {
 384         /*
 385          * If the query has any join RTEs, replace join alias variables with
 386          * base-relation variables. We must do this before sublink processing,
 387          * else sublinks expanded out from join aliases wouldn't get
 388          * processed.
 389          */
 390         if (parse->hasJoinRTEs)
 391                 expr = flatten_join_alias_vars(parse, expr);
 392
 393         /*
 394          * If it's a qual or havingQual, canonicalize it.  It seems most
 395          * useful to do this before applying eval_const_expressions, since the
 396          * latter can optimize flattened AND/ORs better than unflattened ones.
 397          *
 398          * Note: all processing of a qual expression after this point must be
 399          * careful to maintain AND/OR flatness --- that is, do not generate a
 400          * tree with AND directly under AND, nor OR directly under OR.
 401          */
 402         if (kind == EXPRKIND_QUAL)
 403         {
 404                 expr = (Node *) canonicalize_qual((Expr *) expr);
 405
 406 #ifdef OPTIMIZER_DEBUG
 407                 printf("After canonicalize_qual()\n");
 408                 pprint(expr);
 409 #endif
 410         }
 411
 412         /*
 413          * Simplify constant expressions.
 414          */
 415         expr = eval_const_expressions(expr);
 416
 417         /* Expand SubLinks to SubPlans */
 418         if (parse->hasSubLinks)
 419                 expr = SS_process_sublinks(expr, (kind == EXPRKIND_QUAL));
 420
 421         /*
 422          * XXX do not insert anything here unless you have grokked the
 423          * comments in SS_replace_correlation_vars ...
 424          */
 425
 426         /* Replace uplevel vars with Param nodes */
 427         if (PlannerQueryLevel > 1)
 428                 expr = SS_replace_correlation_vars(expr);
 429
 430         /*
 431          * If it's a qual or havingQual, convert it to implicit-AND format.
 432          * (We don't want to do this before eval_const_expressions, since the
 433          * latter would be unable to simplify a top-level AND correctly. Also,
 434          * SS_process_sublinks expects explicit-AND format.)
 435          */
 436         if (kind == EXPRKIND_QUAL)
 437                 expr = (Node *) make_ands_implicit((Expr *) expr);
 438
 439         return expr;
 440 }
 441
 442 /*
 443  * preprocess_qual_conditions
 444  *              Recursively scan the query's jointree and do subquery_planner's
 445  *              preprocessing work on each qual condition found therein.
 446  */
 447 static void
 448 preprocess_qual_conditions(Query *parse, Node *jtnode)
 449 {
 450         if (jtnode == NULL)
 451                 return;
 452         if (IsA(jtnode, RangeTblRef))
 453         {
 454                 /* nothing to do here */
 455         }
 456         else if (IsA(jtnode, FromExpr))
 457         {
 458                 FromExpr   *f = (FromExpr *) jtnode;
 459                 ListCell   *l;
 460
 461                 foreach(l, f->fromlist)
 462                         preprocess_qual_conditions(parse, lfirst(l));
 463
 464                 f->quals = preprocess_expression(parse, f->quals, EXPRKIND_QUAL);
 465         }
 466         else if (IsA(jtnode, JoinExpr))
 467         {
 468                 JoinExpr   *j = (JoinExpr *) jtnode;
 469
 470                 preprocess_qual_conditions(parse, j->larg);
 471                 preprocess_qual_conditions(parse, j->rarg);
 472
 473                 j->quals = preprocess_expression(parse, j->quals, EXPRKIND_QUAL);
 474         }
 475         else
 476                 elog(ERROR, "unrecognized node type: %d",
 477                          (int) nodeTag(jtnode));
 478 }
 479
 480 /*--------------------
 481  * inheritance_planner
 482  *        Generate a plan in the case where the result relation is an
 483  *        inheritance set.
 484  *
 485  * We have to handle this case differently from cases where a source
 486  * relation is an inheritance set.      Source inheritance is expanded at
 487  * the bottom of the plan tree (see allpaths.c), but target inheritance
 488  * has to be expanded at the top.  The reason is that for UPDATE, each
 489  * target relation needs a different targetlist matching its own column
 490  * set.  (This is not so critical for DELETE, but for simplicity we treat
 491  * inherited DELETE the same way.)      Fortunately, the UPDATE/DELETE target
 492  * can never be the nullable side of an outer join, so it's OK to generate
 493  * the plan this way.
 494  *
 495  * parse is the querytree produced by the parser & rewriter.
 496  * inheritlist is an integer list of RT indexes for the result relation set.
 497  *
 498  * Returns a query plan.
 499  *--------------------
 500  */
 501 static Plan *
 502 inheritance_planner(Query *parse, List *inheritlist)
 503 {
 504         int                     parentRTindex = parse->resultRelation;
 505         Oid                     parentOID = getrelid(parentRTindex, parse->rtable);
 506         int                     mainrtlength = list_length(parse->rtable);
 507         List       *subplans = NIL;
 508         List       *tlist = NIL;
 509         ListCell   *l;
 510
 511         foreach(l, inheritlist)
 512         {
 513                 int                     childRTindex = lfirst_int(l);
 514                 Oid                     childOID = getrelid(childRTindex, parse->rtable);
 515                 Query      *subquery;
 516                 Plan       *subplan;
 517
 518                 /* Generate modified query with this rel as target */
 519                 subquery = (Query *) adjust_inherited_attrs((Node *) parse,
 520                                                                                                 parentRTindex, parentOID,
 521                                                                                                  childRTindex, childOID);
 522                 /* Generate plan */
 523                 subplan = grouping_planner(subquery, 0.0 /* retrieve all tuples */ );
 524                 subplans = lappend(subplans, subplan);
 525
 526                 /*
 527                  * XXX my goodness this next bit is ugly.  Really need to think about
 528                  * ways to rein in planner's habit of scribbling on its input.
 529                  *
 530                  * Planning of the subquery might have modified the rangetable,
 531                  * either by addition of RTEs due to expansion of inherited source
 532                  * tables, or by changes of the Query structures inside subquery
 533                  * RTEs.  We have to ensure that this gets propagated back to the
 534                  * master copy.  However, if we aren't done planning yet, we also
 535                  * need to ensure that subsequent calls to grouping_planner have
 536                  * virgin sub-Queries to work from.  So, if we are at the last
 537                  * list entry, just copy the subquery rangetable back to the master
 538                  * copy; if we are not, then extend the master copy by adding
 539                  * whatever the subquery added.  (We assume these added entries
 540                  * will go untouched by the future grouping_planner calls.  We are
 541                  * also effectively assuming that sub-Queries will get planned
 542                  * identically each time, or at least that the impacts on their
 543                  * rangetables will be the same each time.  Did I say this is ugly?)
 544                  */
 545                 if (lnext(l) == NULL)
 546                         parse->rtable = subquery->rtable;
 547                 else
 548                 {
 549                         int             subrtlength = list_length(subquery->rtable);
 550
 551                         if (subrtlength > mainrtlength)
 552                         {
 553                                 List       *subrt;
 554
 555                                 subrt = list_copy_tail(subquery->rtable, mainrtlength);
 556                                 parse->rtable = list_concat(parse->rtable, subrt);
 557                                 mainrtlength = subrtlength;
 558                         }
 559                 }
 560
 561                 /* Save preprocessed tlist from first rel for use in Append */
 562                 if (tlist == NIL)
 563                         tlist = subplan->targetlist;
 564         }
 565
 566         /* Save the target-relations list for the executor, too */
 567         parse->resultRelations = inheritlist;
 568
 569         /* Mark result as unordered (probably unnecessary) */
 570         parse->query_pathkeys = NIL;
 571
 572         return (Plan *) make_append(subplans, true, tlist);
 573 }
 574
 575 /*--------------------
 576  * grouping_planner
 577  *        Perform planning steps related to grouping, aggregation, etc.
 578  *        This primarily means adding top-level processing to the basic
 579  *        query plan produced by query_planner.
 580  *
 581  * parse is the querytree produced by the parser & rewriter.
 582  * tuple_fraction is the fraction of tuples we expect will be retrieved
 583  *
 584  * tuple_fraction is interpreted as follows:
 585  *        0: expect all tuples to be retrieved (normal case)
 586  *        0 < tuple_fraction < 1: expect the given fraction of tuples available
 587  *              from the plan to be retrieved
 588  *        tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
 589  *              expected to be retrieved (ie, a LIMIT specification)
 590  *
 591  * Returns a query plan.  Also, parse->query_pathkeys is returned as the
 592  * actual output ordering of the plan (in pathkey format).
 593  *--------------------
 594  */
 595 static Plan *
 596 grouping_planner(Query *parse, double tuple_fraction)
 597 {
 598         List       *tlist = parse->targetList;
 599         Plan       *result_plan;
 600         List       *current_pathkeys;
 601         List       *sort_pathkeys;
 602
 603         if (parse->setOperations)
 604         {
 605                 List       *set_sortclauses;
 606
 607                 /*
 608                  * Construct the plan for set operations.  The result will not
 609                  * need any work except perhaps a top-level sort and/or LIMIT.
 610                  */
 611                 result_plan = plan_set_operations(parse,
 612                                                                                   &set_sortclauses);
 613
 614                 /*
 615                  * Calculate pathkeys representing the sort order (if any) of the
 616                  * set operation's result.  We have to do this before overwriting
 617                  * the sort key information...
 618                  */
 619                 current_pathkeys = make_pathkeys_for_sortclauses(set_sortclauses,
 620                                                                                                 result_plan->targetlist);
 621                 current_pathkeys = canonicalize_pathkeys(parse, current_pathkeys);
 622
 623                 /*
 624                  * We should not need to call preprocess_targetlist, since we must
 625                  * be in a SELECT query node.  Instead, use the targetlist
 626                  * returned by plan_set_operations (since this tells whether it
 627                  * returned any resjunk columns!), and transfer any sort key
 628                  * information from the original tlist.
 629                  */
 630                 Assert(parse->commandType == CMD_SELECT);
 631
 632                 tlist = postprocess_setop_tlist(result_plan->targetlist, tlist);
 633
 634                 /*
 635                  * Can't handle FOR UPDATE here (parser should have checked
 636                  * already, but let's make sure).
 637                  */
 638                 if (parse->rowMarks)
 639                         ereport(ERROR,
 640                                         (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
 641                                          errmsg("SELECT FOR UPDATE is not allowed with UNION/INTERSECT/EXCEPT")));
 642
 643                 /*
 644                  * Calculate pathkeys that represent result ordering requirements
 645                  */
 646                 sort_pathkeys = make_pathkeys_for_sortclauses(parse->sortClause,
 647                                                                                                           tlist);
 648                 sort_pathkeys = canonicalize_pathkeys(parse, sort_pathkeys);
 649         }
 650         else
 651         {
 652                 /* No set operations, do regular planning */
 653                 List       *sub_tlist;
 654                 List       *group_pathkeys;
 655                 AttrNumber *groupColIdx = NULL;
 656                 bool            need_tlist_eval = true;
 657                 QualCost        tlist_cost;
 658                 double          sub_tuple_fraction;
 659                 Path       *cheapest_path;
 660                 Path       *sorted_path;
 661                 double          dNumGroups = 0;
 662                 long            numGroups = 0;
 663                 int                     numAggs = 0;
 664                 int                     numGroupCols = list_length(parse->groupClause);
 665                 bool            use_hashed_grouping = false;
 666
 667                 /* Preprocess targetlist in case we are inside an INSERT/UPDATE. */
 668                 tlist = preprocess_targetlist(tlist,
 669                                                                           parse->commandType,
 670                                                                           parse->resultRelation,
 671                                                                           parse->rtable);
 672
 673                 /*
 674                  * Add TID targets for rels selected FOR UPDATE (should this be
 675                  * done in preprocess_targetlist?).  The executor uses the TID to
 676                  * know which rows to lock, much as for UPDATE or DELETE.
 677                  */
 678                 if (parse->rowMarks)
 679                 {
 680                         ListCell   *l;
 681
 682                         /*
 683                          * We've got trouble if the FOR UPDATE appears inside
 684                          * grouping, since grouping renders a reference to individual
 685                          * tuple CTIDs invalid.  This is also checked at parse time,
 686                          * but that's insufficient because of rule substitution, query
 687                          * pullup, etc.
 688                          */
 689                         CheckSelectForUpdate(parse);
 690
 691                         /*
 692                          * Currently the executor only supports FOR UPDATE at top
 693                          * level
 694                          */
 695                         if (PlannerQueryLevel > 1)
 696                                 ereport(ERROR,
 697                                                 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
 698                                                  errmsg("SELECT FOR UPDATE is not allowed in subqueries")));
 699
 700                         foreach(l, parse->rowMarks)
 701                         {
 702                                 Index           rti = lfirst_int(l);
 703                                 char       *resname;
 704                                 Resdom     *resdom;
 705                                 Var                *var;
 706                                 TargetEntry *ctid;
 707
 708                                 resname = (char *) palloc(32);
 709                                 snprintf(resname, 32, "ctid%u", rti);
 710                                 resdom = makeResdom(list_length(tlist) + 1,
 711                                                                         TIDOID,
 712                                                                         -1,
 713                                                                         resname,
 714                                                                         true);
 715
 716                                 var = makeVar(rti,
 717                                                           SelfItemPointerAttributeNumber,
 718                                                           TIDOID,
 719                                                           -1,
 720                                                           0);
 721
 722                                 ctid = makeTargetEntry(resdom, (Expr *) var);
 723                                 tlist = lappend(tlist, ctid);
 724                         }
 725                 }
 726
 727                 /*
 728                  * Generate appropriate target list for subplan; may be different
 729                  * from tlist if grouping or aggregation is needed.
 730                  */
 731                 sub_tlist = make_subplanTargetList(parse, tlist,
 732                                                                                  &groupColIdx, &need_tlist_eval);
 733
 734                 /*
 735                  * Calculate pathkeys that represent grouping/ordering
 736                  * requirements
 737                  */
 738                 group_pathkeys = make_pathkeys_for_sortclauses(parse->groupClause,
 739                                                                                                            tlist);
 740                 sort_pathkeys = make_pathkeys_for_sortclauses(parse->sortClause,
 741                                                                                                           tlist);
 742
 743                 /*
 744                  * Will need actual number of aggregates for estimating costs.
 745                  *
 746                  * Note: we do not attempt to detect duplicate aggregates here; a
 747                  * somewhat-overestimated count is okay for our present purposes.
 748                  *
 749                  * Note: think not that we can turn off hasAggs if we find no aggs.
 750                  * It is possible for constant-expression simplification to remove
 751                  * all explicit references to aggs, but we still have to follow
 752                  * the aggregate semantics (eg, producing only one output row).
 753                  */
 754                 if (parse->hasAggs)
 755                         numAggs = count_agg_clause((Node *) tlist) +
 756                                 count_agg_clause(parse->havingQual);
 757
 758                 /*
 759                  * Figure out whether we need a sorted result from query_planner.
 760                  *
 761                  * If we have a GROUP BY clause, then we want a result sorted
 762                  * properly for grouping.  Otherwise, if there is an ORDER BY
 763                  * clause, we want to sort by the ORDER BY clause.      (Note: if we
 764                  * have both, and ORDER BY is a superset of GROUP BY, it would be
 765                  * tempting to request sort by ORDER BY --- but that might just
 766                  * leave us failing to exploit an available sort order at all.
 767                  * Needs more thought...)
 768                  */
 769                 if (parse->groupClause)
 770                         parse->query_pathkeys = group_pathkeys;
 771                 else if (parse->sortClause)
 772                         parse->query_pathkeys = sort_pathkeys;
 773                 else
 774                         parse->query_pathkeys = NIL;
 775
 776                 /*
 777                  * Adjust tuple_fraction if we see that we are going to apply
 778                  * limiting/grouping/aggregation/etc.  This is not overridable by
 779                  * the caller, since it reflects plan actions that this routine
 780                  * will certainly take, not assumptions about context.
 781                  */
 782                 if (parse->limitCount != NULL)
 783                 {
 784                         /*
 785                          * A LIMIT clause limits the absolute number of tuples
 786                          * returned. However, if it's not a constant LIMIT then we
 787                          * have to punt; for lack of a better idea, assume 10% of the
 788                          * plan's result is wanted.
 789                          */
 790                         double          limit_fraction = 0.0;
 791
 792                         if (IsA(parse->limitCount, Const))
 793                         {
 794                                 Const      *limitc = (Const *) parse->limitCount;
 795                                 int32           count = DatumGetInt32(limitc->constvalue);
 796
 797                                 /*
 798                                  * A NULL-constant LIMIT represents "LIMIT ALL", which we
 799                                  * treat the same as no limit (ie, expect to retrieve all
 800                                  * the tuples).
 801                                  */
 802                                 if (!limitc->constisnull && count > 0)
 803                                 {
 804                                         limit_fraction = (double) count;
 805                                         /* We must also consider the OFFSET, if present */
 806                                         if (parse->limitOffset != NULL)
 807                                         {
 808                                                 if (IsA(parse->limitOffset, Const))
 809                                                 {
 810                                                         int32           offset;
 811
 812                                                         limitc = (Const *) parse->limitOffset;
 813                                                         offset = DatumGetInt32(limitc->constvalue);
 814                                                         if (!limitc->constisnull && offset > 0)
 815                                                                 limit_fraction += (double) offset;
 816                                                 }
 817                                                 else
 818                                                 {
 819                                                         /* OFFSET is an expression ... punt ... */
 820                                                         limit_fraction = 0.10;
 821                                                 }
 822                                         }
 823                                 }
 824                         }
 825                         else
 826                         {
 827                                 /* LIMIT is an expression ... punt ... */
 828                                 limit_fraction = 0.10;
 829                         }
 830
 831                         if (limit_fraction > 0.0)
 832                         {
 833                                 /*
 834                                  * If we have absolute limits from both caller and LIMIT,
 835                                  * use the smaller value; if one is fractional and the
 836                                  * other absolute, treat the fraction as a fraction of the
 837                                  * absolute value; else we can multiply the two fractions
 838                                  * together.
 839                                  */
 840                                 if (tuple_fraction >= 1.0)
 841                                 {
 842                                         if (limit_fraction >= 1.0)
 843                                         {
 844                                                 /* both absolute */
 845                                                 tuple_fraction = Min(tuple_fraction, limit_fraction);
 846                                         }
 847                                         else
 848                                         {
 849                                                 /* caller absolute, limit fractional */
 850                                                 tuple_fraction *= limit_fraction;
 851                                                 if (tuple_fraction < 1.0)
 852                                                         tuple_fraction = 1.0;
 853                                         }
 854                                 }
 855                                 else if (tuple_fraction > 0.0)
 856                                 {
 857                                         if (limit_fraction >= 1.0)
 858                                         {
 859                                                 /* caller fractional, limit absolute */
 860                                                 tuple_fraction *= limit_fraction;
 861                                                 if (tuple_fraction < 1.0)
 862                                                         tuple_fraction = 1.0;
 863                                         }
 864                                         else
 865                                         {
 866                                                 /* both fractional */
 867                                                 tuple_fraction *= limit_fraction;
 868                                         }
 869                                 }
 870                                 else
 871                                 {
 872                                         /* no info from caller, just use limit */
 873                                         tuple_fraction = limit_fraction;
 874                                 }
 875                         }
 876                 }
 877
 878                 /*
 879                  * With grouping or aggregation, the tuple fraction to pass to
 880                  * query_planner() may be different from what it is at top level.
 881                  */
 882                 sub_tuple_fraction = tuple_fraction;
 883
 884                 if (parse->groupClause)
 885                 {
 886                         /*
 887                          * In GROUP BY mode, we have the little problem that we don't
 888                          * really know how many input tuples will be needed to make a
 889                          * group, so we can't translate an output LIMIT count into an
 890                          * input count.  For lack of a better idea, assume 25% of the
 891                          * input data will be processed if there is any output limit.
 892                          * However, if the caller gave us a fraction rather than an
 893                          * absolute count, we can keep using that fraction (which
 894                          * amounts to assuming that all the groups are about the same
 895                          * size).
 896                          */
 897                         if (sub_tuple_fraction >= 1.0)
 898                                 sub_tuple_fraction = 0.25;
 899
 900                         /*
 901                          * If both GROUP BY and ORDER BY are specified, we will need
 902                          * two levels of sort --- and, therefore, certainly need to
 903                          * read all the input tuples --- unless ORDER BY is a subset
 904                          * of GROUP BY.  (We have not yet canonicalized the pathkeys,
 905                          * so must use the slower noncanonical comparison method.)
 906                          */
 907                         if (parse->groupClause && parse->sortClause &&
 908                                 !noncanonical_pathkeys_contained_in(sort_pathkeys,
 909                                                                                                         group_pathkeys))
 910                                 sub_tuple_fraction = 0.0;
 911                 }
 912                 else if (parse->hasAggs)
 913                 {
 914                         /*
 915                          * Ungrouped aggregate will certainly want all the input
 916                          * tuples.
 917                          */
 918                         sub_tuple_fraction = 0.0;
 919                 }
 920                 else if (parse->distinctClause)
 921                 {
 922                         /*
 923                          * SELECT DISTINCT, like GROUP, will absorb an unpredictable
 924                          * number of input tuples per output tuple.  Handle the same
 925                          * way.
 926                          */
 927                         if (sub_tuple_fraction >= 1.0)
 928                                 sub_tuple_fraction = 0.25;
 929                 }
 930
 931                 /*
 932                  * Generate the best unsorted and presorted paths for this Query
 933                  * (but note there may not be any presorted path).
 934                  */
 935                 query_planner(parse, sub_tlist, sub_tuple_fraction,
 936                                           &cheapest_path, &sorted_path);
 937
 938                 /*
 939                  * We couldn't canonicalize group_pathkeys and sort_pathkeys
 940                  * before running query_planner(), so do it now.
 941                  */
 942                 group_pathkeys = canonicalize_pathkeys(parse, group_pathkeys);
 943                 sort_pathkeys = canonicalize_pathkeys(parse, sort_pathkeys);
 944
 945                 /*
 946                  * Consider whether we might want to use hashed grouping.
 947                  */
 948                 if (parse->groupClause)
 949                 {
 950                         List       *groupExprs;
 951                         double          cheapest_path_rows;
 952                         int                     cheapest_path_width;
 953
 954                         /*
 955                          * Beware in this section of the possibility that
 956                          * cheapest_path->parent is NULL.  This could happen if user
 957                          * does something silly like SELECT 'foo' GROUP BY 1;
 958                          */
 959                         if (cheapest_path->parent)
 960                         {
 961                                 cheapest_path_rows = cheapest_path->parent->rows;
 962                                 cheapest_path_width = cheapest_path->parent->width;
 963                         }
 964                         else
 965                         {
 966                                 cheapest_path_rows = 1; /* assume non-set result */
 967                                 cheapest_path_width = 100;              /* arbitrary */
 968                         }
 969
 970                         /*
 971                          * Always estimate the number of groups.  We can't do this
 972                          * until after running query_planner(), either.
 973                          */
 974                         groupExprs = get_sortgrouplist_exprs(parse->groupClause,
 975                                                                                                  parse->targetList);
 976                         dNumGroups = estimate_num_groups(parse,
 977                                                                                          groupExprs,
 978                                                                                          cheapest_path_rows);
 979                         /* Also want it as a long int --- but 'ware overflow! */
 980                         numGroups = (long) Min(dNumGroups, (double) LONG_MAX);
 981
 982                         /*
 983                          * Check can't-do-it conditions, including whether the
 984                          * grouping operators are hashjoinable.
 985                          *
 986                          * Executor doesn't support hashed aggregation with DISTINCT
 987                          * aggregates.  (Doing so would imply storing *all* the input
 988                          * values in the hash table, which seems like a certain
 989                          * loser.)
 990                          */
 991                         if (!enable_hashagg || !hash_safe_grouping(parse))
 992                                 use_hashed_grouping = false;
 993                         else if (parse->hasAggs &&
 994                                          (contain_distinct_agg_clause((Node *) tlist) ||
 995                                           contain_distinct_agg_clause(parse->havingQual)))
 996                                 use_hashed_grouping = false;
 997                         else
 998                         {
 999                                 /*
1000                                  * Use hashed grouping if (a) we think we can fit the
1001                                  * hashtable into work_mem, *and* (b) the estimated cost
1002                                  * is no more than doing it the other way.      While avoiding
1003                                  * the need for sorted input is usually a win, the fact
1004                                  * that the output won't be sorted may be a loss; so we
1005                                  * need to do an actual cost comparison.
1006                                  *
1007                                  * In most cases we have no good way to estimate the size of
1008                                  * the transition value needed by an aggregate;
1009                                  * arbitrarily assume it is 100 bytes.  Also set the
1010                                  * overhead per hashtable entry at 64 bytes.
1011                                  */
1012                                 int                     hashentrysize = cheapest_path_width + 64 + numAggs * 100;
1013
1014                                 if (hashentrysize * dNumGroups <= work_mem * 1024L)
1015                                 {
1016                                         /*
1017                                          * Okay, do the cost comparison.  We need to consider
1018                                          * cheapest_path + hashagg [+ final sort] versus
1019                                          * either cheapest_path [+ sort] + group or agg [+
1020                                          * final sort] or presorted_path + group or agg [+
1021                                          * final sort] where brackets indicate a step that may
1022                                          * not be needed. We assume query_planner() will have
1023                                          * returned a presorted path only if it's a winner
1024                                          * compared to cheapest_path for this purpose.
1025                                          *
1026                                          * These path variables are dummies that just hold cost
1027                                          * fields; we don't make actual Paths for these steps.
1028                                          */
1029                                         Path            hashed_p;
1030                                         Path            sorted_p;
1031
1032                                         cost_agg(&hashed_p, parse,
1033                                                          AGG_HASHED, numAggs,
1034                                                          numGroupCols, dNumGroups,
1035                                                          cheapest_path->startup_cost,
1036                                                          cheapest_path->total_cost,
1037                                                          cheapest_path_rows);
1038                                         /* Result of hashed agg is always unsorted */
1039                                         if (sort_pathkeys)
1040                                                 cost_sort(&hashed_p, parse, sort_pathkeys,
1041                                                                   hashed_p.total_cost,
1042                                                                   dNumGroups,
1043                                                                   cheapest_path_width);
1044
1045                                         if (sorted_path)
1046                                         {
1047                                                 sorted_p.startup_cost = sorted_path->startup_cost;
1048                                                 sorted_p.total_cost = sorted_path->total_cost;
1049                                                 current_pathkeys = sorted_path->pathkeys;
1050                                         }
1051                                         else
1052                                         {
1053                                                 sorted_p.startup_cost = cheapest_path->startup_cost;
1054                                                 sorted_p.total_cost = cheapest_path->total_cost;
1055                                                 current_pathkeys = cheapest_path->pathkeys;
1056                                         }
1057                                         if (!pathkeys_contained_in(group_pathkeys,
1058                                                                                            current_pathkeys))
1059                                         {
1060                                                 cost_sort(&sorted_p, parse, group_pathkeys,
1061                                                                   sorted_p.total_cost,
1062                                                                   cheapest_path_rows,
1063                                                                   cheapest_path_width);
1064                                                 current_pathkeys = group_pathkeys;
1065                                         }
1066                                         if (parse->hasAggs)
1067                                                 cost_agg(&sorted_p, parse,
1068                                                                  AGG_SORTED, numAggs,
1069                                                                  numGroupCols, dNumGroups,
1070                                                                  sorted_p.startup_cost,
1071                                                                  sorted_p.total_cost,
1072                                                                  cheapest_path_rows);
1073                                         else
1074                                                 cost_group(&sorted_p, parse,
1075                                                                    numGroupCols, dNumGroups,
1076                                                                    sorted_p.startup_cost,
1077                                                                    sorted_p.total_cost,
1078                                                                    cheapest_path_rows);
1079                                         /* The Agg or Group node will preserve ordering */
1080                                         if (sort_pathkeys &&
1081                                                 !pathkeys_contained_in(sort_pathkeys,
1082                                                                                            current_pathkeys))
1083                                         {
1084                                                 cost_sort(&sorted_p, parse, sort_pathkeys,
1085                                                                   sorted_p.total_cost,
1086                                                                   dNumGroups,
1087                                                                   cheapest_path_width);
1088                                         }
1089
1090                                         /*
1091                                          * Now make the decision using the top-level tuple
1092                                          * fraction.  First we have to convert an absolute
1093                                          * count (LIMIT) into fractional form.
1094                                          */
1095                                         if (tuple_fraction >= 1.0)
1096                                                 tuple_fraction /= dNumGroups;
1097
1098                                         if (compare_fractional_path_costs(&hashed_p, &sorted_p,
1099                                                                                                           tuple_fraction) < 0)
1100                                         {
1101                                                 /* Hashed is cheaper, so use it */
1102                                                 use_hashed_grouping = true;
1103                                         }
1104                                 }
1105                         }
1106                 }
1107
1108                 /*
1109                  * Select the best path and create a plan to execute it.
1110                  *
1111                  * If we are doing hashed grouping, we will always read all the input
1112                  * tuples, so use the cheapest-total path.      Otherwise, trust
1113                  * query_planner's decision about which to use.
1114                  */
1115                 if (sorted_path && !use_hashed_grouping)
1116                 {
1117                         result_plan = create_plan(parse, sorted_path);
1118                         current_pathkeys = sorted_path->pathkeys;
1119                 }
1120                 else
1121                 {
1122                         result_plan = create_plan(parse, cheapest_path);
1123                         current_pathkeys = cheapest_path->pathkeys;
1124                 }
1125
1126                 /*
1127                  * create_plan() returns a plan with just a "flat" tlist of
1128                  * required Vars.  Usually we need to insert the sub_tlist as the
1129                  * tlist of the top plan node.  However, we can skip that if we
1130                  * determined that whatever query_planner chose to return will be
1131                  * good enough.
1132                  */
1133                 if (need_tlist_eval)
1134                 {
1135                         /*
1136                          * If the top-level plan node is one that cannot do expression
1137                          * evaluation, we must insert a Result node to project the
1138                          * desired tlist.
1139                          */
1140                         if (!is_projection_capable_plan(result_plan))
1141                         {
1142                                 result_plan = (Plan *) make_result(sub_tlist, NULL,
1143                                                                                                    result_plan);
1144                         }
1145                         else
1146                         {
1147                                 /*
1148                                  * Otherwise, just replace the subplan's flat tlist with
1149                                  * the desired tlist.
1150                                  */
1151                                 result_plan->targetlist = sub_tlist;
1152                         }
1153
1154                         /*
1155                          * Also, account for the cost of evaluation of the sub_tlist.
1156                          *
1157                          * Up to now, we have only been dealing with "flat" tlists,
1158                          * containing just Vars.  So their evaluation cost is zero
1159                          * according to the model used by cost_qual_eval() (or if you
1160                          * prefer, the cost is factored into cpu_tuple_cost).  Thus we
1161                          * can avoid accounting for tlist cost throughout
1162                          * query_planner() and subroutines.  But now we've inserted a
1163                          * tlist that might contain actual operators, sub-selects, etc
1164                          * --- so we'd better account for its cost.
1165                          *
1166                          * Below this point, any tlist eval cost for added-on nodes
1167                          * should be accounted for as we create those nodes.
1168                          * Presently, of the node types we can add on, only Agg and
1169                          * Group project new tlists (the rest just copy their input
1170                          * tuples) --- so make_agg() and make_group() are responsible
1171                          * for computing the added cost.
1172                          */
1173                         cost_qual_eval(&tlist_cost, sub_tlist);
1174                         result_plan->startup_cost += tlist_cost.startup;
1175                         result_plan->total_cost += tlist_cost.startup +
1176                                 tlist_cost.per_tuple * result_plan->plan_rows;
1177                 }
1178                 else
1179                 {
1180                         /*
1181                          * Since we're using query_planner's tlist and not the one
1182                          * make_subplanTargetList calculated, we have to refigure any
1183                          * grouping-column indexes make_subplanTargetList computed.
1184                          */
1185                         locate_grouping_columns(parse, tlist, result_plan->targetlist,
1186                                                                         groupColIdx);
1187                 }
1188
1189                 /*
1190                  * Insert AGG or GROUP node if needed, plus an explicit sort step
1191                  * if necessary.
1192                  *
1193                  * HAVING clause, if any, becomes qual of the Agg node
1194                  */
1195                 if (use_hashed_grouping)
1196                 {
1197                         /* Hashed aggregate plan --- no sort needed */
1198                         result_plan = (Plan *) make_agg(parse,
1199                                                                                         tlist,
1200                                                                                         (List *) parse->havingQual,
1201                                                                                         AGG_HASHED,
1202                                                                                         numGroupCols,
1203                                                                                         groupColIdx,
1204                                                                                         numGroups,
1205                                                                                         numAggs,
1206                                                                                         result_plan);
1207                         /* Hashed aggregation produces randomly-ordered results */
1208                         current_pathkeys = NIL;
1209                 }
1210                 else if (parse->hasAggs)
1211                 {
1212                         /* Plain aggregate plan --- sort if needed */
1213                         AggStrategy aggstrategy;
1214
1215                         if (parse->groupClause)
1216                         {
1217                                 if (!pathkeys_contained_in(group_pathkeys, current_pathkeys))
1218                                 {
1219                                         result_plan = (Plan *)
1220                                                 make_sort_from_groupcols(parse,
1221                                                                                                  parse->groupClause,
1222                                                                                                  groupColIdx,
1223                                                                                                  result_plan);
1224                                         current_pathkeys = group_pathkeys;
1225                                 }
1226                                 aggstrategy = AGG_SORTED;
1227
1228                                 /*
1229                                  * The AGG node will not change the sort ordering of its
1230                                  * groups, so current_pathkeys describes the result too.
1231                                  */
1232                         }
1233                         else
1234                         {
1235                                 aggstrategy = AGG_PLAIN;
1236                                 /* Result will be only one row anyway; no sort order */
1237                                 current_pathkeys = NIL;
1238                         }
1239
1240                         result_plan = (Plan *) make_agg(parse,
1241                                                                                         tlist,
1242                                                                                         (List *) parse->havingQual,
1243                                                                                         aggstrategy,
1244                                                                                         numGroupCols,
1245                                                                                         groupColIdx,
1246                                                                                         numGroups,
1247                                                                                         numAggs,
1248                                                                                         result_plan);
1249                 }
1250                 else
1251                 {
1252                         /*
1253                          * If there are no Aggs, we shouldn't have any HAVING qual
1254                          * anymore
1255                          */
1256                         Assert(parse->havingQual == NULL);
1257
1258                         /*
1259                          * If we have a GROUP BY clause, insert a group node (plus the
1260                          * appropriate sort node, if necessary).
1261                          */
1262                         if (parse->groupClause)
1263                         {
1264                                 /*
1265                                  * Add an explicit sort if we couldn't make the path come
1266                                  * out the way the GROUP node needs it.
1267                                  */
1268                                 if (!pathkeys_contained_in(group_pathkeys, current_pathkeys))
1269                                 {
1270                                         result_plan = (Plan *)
1271                                                 make_sort_from_groupcols(parse,
1272                                                                                                  parse->groupClause,
1273                                                                                                  groupColIdx,
1274                                                                                                  result_plan);
1275                                         current_pathkeys = group_pathkeys;
1276                                 }
1277
1278                                 result_plan = (Plan *) make_group(parse,
1279                                                                                                   tlist,
1280                                                                                                   numGroupCols,
1281                                                                                                   groupColIdx,
1282                                                                                                   dNumGroups,
1283                                                                                                   result_plan);
1284                                 /* The Group node won't change sort ordering */
1285                         }
1286                 }
1287         }                                                       /* end of if (setOperations) */
1288
1289         /*
1290          * If we were not able to make the plan come out in the right order,
1291          * add an explicit sort step.
1292          */
1293         if (parse->sortClause)
1294         {
1295                 if (!pathkeys_contained_in(sort_pathkeys, current_pathkeys))
1296                 {
1297                         result_plan = (Plan *)
1298                                 make_sort_from_sortclauses(parse,
1299                                                                                    parse->sortClause,
1300                                                                                    result_plan);
1301                         current_pathkeys = sort_pathkeys;
1302                 }
1303         }
1304
1305         /*
1306          * If there is a DISTINCT clause, add the UNIQUE node.
1307          */
1308         if (parse->distinctClause)
1309         {
1310                 result_plan = (Plan *) make_unique(result_plan, parse->distinctClause);
1311
1312                 /*
1313                  * If there was grouping or aggregation, leave plan_rows as-is
1314                  * (ie, assume the result was already mostly unique).  If not,
1315                  * it's reasonable to assume the UNIQUE filter has effects
1316                  * comparable to GROUP BY.
1317                  */
1318                 if (!parse->groupClause && !parse->hasAggs)
1319                 {
1320                         List       *distinctExprs;
1321
1322                         distinctExprs = get_sortgrouplist_exprs(parse->distinctClause,
1323                                                                                                         parse->targetList);
1324                         result_plan->plan_rows = estimate_num_groups(parse,
1325                                                                                                                  distinctExprs,
1326                                                                                                  result_plan->plan_rows);
1327                 }
1328         }
1329
1330         /*
1331          * Finally, if there is a LIMIT/OFFSET clause, add the LIMIT node.
1332          */
1333         if (parse->limitOffset || parse->limitCount)
1334         {
1335                 result_plan = (Plan *) make_limit(result_plan,
1336                                                                                   parse->limitOffset,
1337                                                                                   parse->limitCount);
1338         }
1339
1340         /*
1341          * Return the actual output ordering in query_pathkeys for possible
1342          * use by an outer query level.
1343          */
1344         parse->query_pathkeys = current_pathkeys;
1345
1346         return result_plan;
1347 }
1348
1349 /*
1350  * hash_safe_grouping - are grouping operators hashable?
1351  *
1352  * We assume hashed aggregation will work if the datatype's equality operator
1353  * is marked hashjoinable.
1354  */
1355 static bool
1356 hash_safe_grouping(Query *parse)
1357 {
1358         ListCell   *gl;
1359
1360         foreach(gl, parse->groupClause)
1361         {
1362                 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1363                 TargetEntry *tle = get_sortgroupclause_tle(grpcl, parse->targetList);
1364                 Operator        optup;
1365                 bool            oprcanhash;
1366
1367                 optup = equality_oper(tle->resdom->restype, true);
1368                 if (!optup)
1369                         return false;
1370                 oprcanhash = ((Form_pg_operator) GETSTRUCT(optup))->oprcanhash;
1371                 ReleaseSysCache(optup);
1372                 if (!oprcanhash)
1373                         return false;
1374         }
1375         return true;
1376 }
1377
1378 /*---------------
1379  * make_subplanTargetList
1380  *        Generate appropriate target list when grouping is required.
1381  *
1382  * When grouping_planner inserts Aggregate or Group plan nodes above
1383  * the result of query_planner, we typically want to pass a different
1384  * target list to query_planner than the outer plan nodes should have.
1385  * This routine generates the correct target list for the subplan.
1386  *
1387  * The initial target list passed from the parser already contains entries
1388  * for all ORDER BY and GROUP BY expressions, but it will not have entries
1389  * for variables used only in HAVING clauses; so we need to add those
1390  * variables to the subplan target list.  Also, if we are doing either
1391  * grouping or aggregation, we flatten all expressions except GROUP BY items
1392  * into their component variables; the other expressions will be computed by
1393  * the inserted nodes rather than by the subplan.  For example,
1394  * given a query like
1395  *              SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
1396  * we want to pass this targetlist to the subplan:
1397  *              a,b,c,d,a+b
1398  * where the a+b target will be used by the Sort/Group steps, and the
1399  * other targets will be used for computing the final results.  (In the
1400  * above example we could theoretically suppress the a and b targets and
1401  * pass down only c,d,a+b, but it's not really worth the trouble to
1402  * eliminate simple var references from the subplan.  We will avoid doing
1403  * the extra computation to recompute a+b at the outer level; see
1404  * replace_vars_with_subplan_refs() in setrefs.c.)
1405  *
1406  * If we are grouping or aggregating, *and* there are no non-Var grouping
1407  * expressions, then the returned tlist is effectively dummy; we do not
1408  * need to force it to be evaluated, because all the Vars it contains
1409  * should be present in the output of query_planner anyway.
1410  *
1411  * 'parse' is the query being processed.
1412  * 'tlist' is the query's target list.
1413  * 'groupColIdx' receives an array of column numbers for the GROUP BY
1414  *                      expressions (if there are any) in the subplan's target list.
1415  * 'need_tlist_eval' is set true if we really need to evaluate the
1416  *                      result tlist.
1417  *
1418  * The result is the targetlist to be passed to the subplan.
1419  *---------------
1420  */
1421 static List *
1422 make_subplanTargetList(Query *parse,
1423                                            List *tlist,
1424                                            AttrNumber **groupColIdx,
1425                                            bool *need_tlist_eval)
1426 {
1427         List       *sub_tlist;
1428         List       *extravars;
1429         int                     numCols;
1430
1431         *groupColIdx = NULL;
1432
1433         /*
1434          * If we're not grouping or aggregating, nothing to do here;
1435          * query_planner should receive the unmodified target list.
1436          */
1437         if (!parse->hasAggs && !parse->groupClause)
1438         {
1439                 *need_tlist_eval = true;
1440                 return tlist;
1441         }
1442
1443         /*
1444          * Otherwise, start with a "flattened" tlist (having just the vars
1445          * mentioned in the targetlist and HAVING qual --- but not upper-
1446          * level Vars; they will be replaced by Params later on).
1447          */
1448         sub_tlist = flatten_tlist(tlist);
1449         extravars = pull_var_clause(parse->havingQual, false);
1450         sub_tlist = add_to_flat_tlist(sub_tlist, extravars);
1451         list_free(extravars);
1452         *need_tlist_eval = false;       /* only eval if not flat tlist */
1453
1454         /*
1455          * If grouping, create sub_tlist entries for all GROUP BY expressions
1456          * (GROUP BY items that are simple Vars should be in the list
1457          * already), and make an array showing where the group columns are in
1458          * the sub_tlist.
1459          */
1460         numCols = list_length(parse->groupClause);
1461         if (numCols > 0)
1462         {
1463                 int                     keyno = 0;
1464                 AttrNumber *grpColIdx;
1465                 ListCell   *gl;
1466
1467                 grpColIdx = (AttrNumber *) palloc(sizeof(AttrNumber) * numCols);
1468                 *groupColIdx = grpColIdx;
1469
1470                 foreach(gl, parse->groupClause)
1471                 {
1472                         GroupClause *grpcl = (GroupClause *) lfirst(gl);
1473                         Node       *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1474                         TargetEntry *te = NULL;
1475                         ListCell   *sl;
1476
1477                         /* Find or make a matching sub_tlist entry */
1478                         foreach(sl, sub_tlist)
1479                         {
1480                                 te = (TargetEntry *) lfirst(sl);
1481                                 if (equal(groupexpr, te->expr))
1482                                         break;
1483                         }
1484                         if (!sl)
1485                         {
1486                                 te = makeTargetEntry(makeResdom(list_length(sub_tlist) + 1,
1487                                                                                                 exprType(groupexpr),
1488                                                                                                 exprTypmod(groupexpr),
1489                                                                                                 NULL,
1490                                                                                                 false),
1491                                                                          (Expr *) groupexpr);
1492                                 sub_tlist = lappend(sub_tlist, te);
1493                                 *need_tlist_eval = true;                /* it's not flat anymore */
1494                         }
1495
1496                         /* and save its resno */
1497                         grpColIdx[keyno++] = te->resdom->resno;
1498                 }
1499         }
1500
1501         return sub_tlist;
1502 }
1503
1504 /*
1505  * locate_grouping_columns
1506  *              Locate grouping columns in the tlist chosen by query_planner.
1507  *
1508  * This is only needed if we don't use the sub_tlist chosen by
1509  * make_subplanTargetList.      We have to forget the column indexes found
1510  * by that routine and re-locate the grouping vars in the real sub_tlist.
1511  */
1512 static void
1513 locate_grouping_columns(Query *parse,
1514                                                 List *tlist,
1515                                                 List *sub_tlist,
1516                                                 AttrNumber *groupColIdx)
1517 {
1518         int                     keyno = 0;
1519         ListCell   *gl;
1520
1521         /*
1522          * No work unless grouping.
1523          */
1524         if (!parse->groupClause)
1525         {
1526                 Assert(groupColIdx == NULL);
1527                 return;
1528         }
1529         Assert(groupColIdx != NULL);
1530
1531         foreach(gl, parse->groupClause)
1532         {
1533                 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1534                 Node       *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1535                 TargetEntry *te = NULL;
1536                 ListCell   *sl;
1537
1538                 foreach(sl, sub_tlist)
1539                 {
1540                         te = (TargetEntry *) lfirst(sl);
1541                         if (equal(groupexpr, te->expr))
1542                                 break;
1543                 }
1544                 if (!sl)
1545                         elog(ERROR, "failed to locate grouping columns");
1546
1547                 groupColIdx[keyno++] = te->resdom->resno;
1548         }
1549 }
1550
1551 /*
1552  * postprocess_setop_tlist
1553  *        Fix up targetlist returned by plan_set_operations().
1554  *
1555  * We need to transpose sort key info from the orig_tlist into new_tlist.
1556  * NOTE: this would not be good enough if we supported resjunk sort keys
1557  * for results of set operations --- then, we'd need to project a whole
1558  * new tlist to evaluate the resjunk columns.  For now, just ereport if we
1559  * find any resjunk columns in orig_tlist.
1560  */
1561 static List *
1562 postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
1563 {
1564         ListCell   *l;
1565         ListCell   *orig_tlist_item = list_head(orig_tlist);
1566
1567         foreach(l, new_tlist)
1568         {
1569                 TargetEntry *new_tle = (TargetEntry *) lfirst(l);
1570                 TargetEntry *orig_tle;
1571
1572                 /* ignore resjunk columns in setop result */
1573                 if (new_tle->resdom->resjunk)
1574                         continue;
1575
1576                 Assert(orig_tlist_item != NULL);
1577                 orig_tle = (TargetEntry *) lfirst(orig_tlist_item);
1578                 orig_tlist_item = lnext(orig_tlist_item);
1579                 if (orig_tle->resdom->resjunk)  /* should not happen */
1580                         elog(ERROR, "resjunk output columns are not implemented");
1581                 Assert(new_tle->resdom->resno == orig_tle->resdom->resno);
1582                 Assert(new_tle->resdom->restype == orig_tle->resdom->restype);
1583                 new_tle->resdom->ressortgroupref = orig_tle->resdom->ressortgroupref;
1584         }
1585         if (orig_tlist_item != NULL)
1586                 elog(ERROR, "resjunk output columns are not implemented");
1587         return new_tlist;
1588 }