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) 1996-2003, PostgreSQL Global Development Group
   7  * Portions Copyright (c) 1994, Regents of the University of California
   8  *
   9  *
  10  * IDENTIFICATION
  11  *        $Header: /cvsroot/pgsql/src/backend/optimizer/plan/planner.c,v 1.161 2003/09/25 06:58:00 petere 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 /* Expression kind codes for preprocess_expression */
  47 #define EXPRKIND_QUAL   0
  48 #define EXPRKIND_TARGET 1
  49 #define EXPRKIND_RTFUNC 2
  50 #define EXPRKIND_LIMIT  3
  51 #define EXPRKIND_ININFO 4
  52
  53
  54 static Node *preprocess_expression(Query *parse, Node *expr, int kind);
  55 static void preprocess_qual_conditions(Query *parse, Node *jtnode);
  56 static Plan *inheritance_planner(Query *parse, List *inheritlist);
  57 static Plan *grouping_planner(Query *parse, double tuple_fraction);
  58 static bool hash_safe_grouping(Query *parse);
  59 static List *make_subplanTargetList(Query *parse, List *tlist,
  60                                            AttrNumber **groupColIdx, bool *need_tlist_eval);
  61 static void locate_grouping_columns(Query *parse,
  62                                                 List *tlist,
  63                                                 List *sub_tlist,
  64                                                 AttrNumber *groupColIdx);
  65 static List *postprocess_setop_tlist(List *new_tlist, List *orig_tlist);
  66
  67
  68 /*****************************************************************************
  69  *
  70  *         Query optimizer entry point
  71  *
  72  *****************************************************************************/
  73 Plan *
  74 planner(Query *parse, bool isCursor, int cursorOptions)
  75 {
  76         double          tuple_fraction;
  77         Plan       *result_plan;
  78         Index           save_PlannerQueryLevel;
  79         List       *save_PlannerParamList;
  80
  81         /*
  82          * The planner can be called recursively (an example is when
  83          * eval_const_expressions tries to pre-evaluate an SQL function). So,
  84          * these global state variables must be saved and restored.
  85          *
  86          * These vars cannot be moved into the Query structure since their whole
  87          * purpose is communication across multiple sub-Queries.
  88          *
  89          * Note we do NOT save and restore PlannerPlanId: it exists to assign
  90          * unique IDs to SubPlan nodes, and we want those IDs to be unique for
  91          * the life of a backend.  Also, PlannerInitPlan is saved/restored in
  92          * subquery_planner, not here.
  93          */
  94         save_PlannerQueryLevel = PlannerQueryLevel;
  95         save_PlannerParamList = PlannerParamList;
  96
  97         /* Initialize state for handling outer-level references and params */
  98         PlannerQueryLevel = 0;          /* will be 1 in top-level subquery_planner */
  99         PlannerParamList = NIL;
 100
 101         /* Determine what fraction of the plan is likely to be scanned */
 102         if (isCursor)
 103         {
 104                 /*
 105                  * We have no real idea how many tuples the user will ultimately
 106                  * FETCH from a cursor, but it seems a good bet that he doesn't
 107                  * want 'em all.  Optimize for 10% retrieval (you gotta better
 108                  * number?      Should this be a SETtable parameter?)
 109                  */
 110                 tuple_fraction = 0.10;
 111         }
 112         else
 113         {
 114                 /* Default assumption is we need all the tuples */
 115                 tuple_fraction = 0.0;
 116         }
 117
 118         /* primary planning entry point (may recurse for subqueries) */
 119         result_plan = subquery_planner(parse, tuple_fraction);
 120
 121         Assert(PlannerQueryLevel == 0);
 122
 123         /*
 124          * If creating a plan for a scrollable cursor, make sure it can run
 125          * backwards on demand.  Add a Material node at the top at need.
 126          */
 127         if (isCursor && (cursorOptions & CURSOR_OPT_SCROLL))
 128         {
 129                 if (!ExecSupportsBackwardScan(result_plan))
 130                         result_plan = materialize_finished_plan(result_plan);
 131         }
 132
 133         /* executor wants to know total number of Params used overall */
 134         result_plan->nParamExec = length(PlannerParamList);
 135
 136         /* final cleanup of the plan */
 137         set_plan_references(result_plan, parse->rtable);
 138
 139         /* restore state for outer planner, if any */
 140         PlannerQueryLevel = save_PlannerQueryLevel;
 141         PlannerParamList = save_PlannerParamList;
 142
 143         return result_plan;
 144 }
 145
 146
 147 /*--------------------
 148  * subquery_planner
 149  *        Invokes the planner on a subquery.  We recurse to here for each
 150  *        sub-SELECT found in the query tree.
 151  *
 152  * parse is the querytree produced by the parser & rewriter.
 153  * tuple_fraction is the fraction of tuples we expect will be retrieved.
 154  * tuple_fraction is interpreted as explained for grouping_planner, below.
 155  *
 156  * Basically, this routine does the stuff that should only be done once
 157  * per Query object.  It then calls grouping_planner.  At one time,
 158  * grouping_planner could be invoked recursively on the same Query object;
 159  * that's not currently true, but we keep the separation between the two
 160  * routines anyway, in case we need it again someday.
 161  *
 162  * subquery_planner will be called recursively to handle sub-Query nodes
 163  * found within the query's expressions and rangetable.
 164  *
 165  * Returns a query plan.
 166  *--------------------
 167  */
 168 Plan *
 169 subquery_planner(Query *parse, double tuple_fraction)
 170 {
 171         List       *saved_initplan = PlannerInitPlan;
 172         int                     saved_planid = PlannerPlanId;
 173         bool            hasOuterJoins;
 174         Plan       *plan;
 175         List       *newHaving;
 176         List       *lst;
 177
 178         /* Set up for a new level of subquery */
 179         PlannerQueryLevel++;
 180         PlannerInitPlan = NIL;
 181
 182         /*
 183          * Look for IN clauses at the top level of WHERE, and transform them
 184          * into joins.  Note that this step only handles IN clauses originally
 185          * at top level of WHERE; if we pull up any subqueries in the next
 186          * step, their INs are processed just before pulling them up.
 187          */
 188         parse->in_info_list = NIL;
 189         if (parse->hasSubLinks)
 190                 parse->jointree->quals = pull_up_IN_clauses(parse,
 191                                                                                                  parse->jointree->quals);
 192
 193         /*
 194          * Check to see if any subqueries in the rangetable can be merged into
 195          * this query.
 196          */
 197         parse->jointree = (FromExpr *)
 198                 pull_up_subqueries(parse, (Node *) parse->jointree, false);
 199
 200         /*
 201          * Detect whether any rangetable entries are RTE_JOIN kind; if not, we
 202          * can avoid the expense of doing flatten_join_alias_vars().  Also
 203          * check for outer joins --- if none, we can skip
 204          * reduce_outer_joins(). This must be done after we have done
 205          * pull_up_subqueries, of course.
 206          */
 207         parse->hasJoinRTEs = false;
 208         hasOuterJoins = false;
 209         foreach(lst, parse->rtable)
 210         {
 211                 RangeTblEntry *rte = (RangeTblEntry *) lfirst(lst);
 212
 213                 if (rte->rtekind == RTE_JOIN)
 214                 {
 215                         parse->hasJoinRTEs = true;
 216                         if (IS_OUTER_JOIN(rte->jointype))
 217                         {
 218                                 hasOuterJoins = true;
 219                                 /* Can quit scanning once we find an outer join */
 220                                 break;
 221                         }
 222                 }
 223         }
 224
 225         /*
 226          * Do expression preprocessing on targetlist and quals.
 227          */
 228         parse->targetList = (List *)
 229                 preprocess_expression(parse, (Node *) parse->targetList,
 230                                                           EXPRKIND_TARGET);
 231
 232         preprocess_qual_conditions(parse, (Node *) parse->jointree);
 233
 234         parse->havingQual = preprocess_expression(parse, parse->havingQual,
 235                                                                                           EXPRKIND_QUAL);
 236
 237         parse->limitOffset = preprocess_expression(parse, parse->limitOffset,
 238                                                                                            EXPRKIND_LIMIT);
 239         parse->limitCount = preprocess_expression(parse, parse->limitCount,
 240                                                                                           EXPRKIND_LIMIT);
 241
 242         parse->in_info_list = (List *)
 243                 preprocess_expression(parse, (Node *) parse->in_info_list,
 244                                                           EXPRKIND_ININFO);
 245
 246         /* Also need to preprocess expressions for function RTEs */
 247         foreach(lst, parse->rtable)
 248         {
 249                 RangeTblEntry *rte = (RangeTblEntry *) lfirst(lst);
 250
 251                 if (rte->rtekind == RTE_FUNCTION)
 252                         rte->funcexpr = preprocess_expression(parse, rte->funcexpr,
 253                                                                                                   EXPRKIND_RTFUNC);
 254         }
 255
 256         /*
 257          * A HAVING clause without aggregates is equivalent to a WHERE clause
 258          * (except it can only refer to grouped fields).  Transfer any
 259          * agg-free clauses of the HAVING qual into WHERE.      This may seem like
 260          * wasting cycles to cater to stupidly-written queries, but there are
 261          * other reasons for doing it.  Firstly, if the query contains no aggs
 262          * at all, then we aren't going to generate an Agg plan node, and so
 263          * there'll be no place to execute HAVING conditions; without this
 264          * transfer, we'd lose the HAVING condition entirely, which is wrong.
 265          * Secondly, when we push down a qual condition into a sub-query, it's
 266          * easiest to push the qual into HAVING always, in case it contains
 267          * aggs, and then let this code sort it out.
 268          *
 269          * Note that both havingQual and parse->jointree->quals are in
 270          * implicitly-ANDed-list form at this point, even though they are
 271          * declared as Node *.
 272          */
 273         newHaving = NIL;
 274         foreach(lst, (List *) parse->havingQual)
 275         {
 276                 Node       *havingclause = (Node *) lfirst(lst);
 277
 278                 if (contain_agg_clause(havingclause))
 279                         newHaving = lappend(newHaving, havingclause);
 280                 else
 281                         parse->jointree->quals = (Node *)
 282                                 lappend((List *) parse->jointree->quals, havingclause);
 283         }
 284         parse->havingQual = (Node *) newHaving;
 285
 286         /*
 287          * If we have any outer joins, try to reduce them to plain inner
 288          * joins. This step is most easily done after we've done expression
 289          * preprocessing.
 290          */
 291         if (hasOuterJoins)
 292                 reduce_outer_joins(parse);
 293
 294         /*
 295          * See if we can simplify the jointree; opportunities for this may
 296          * come from having pulled up subqueries, or from flattening explicit
 297          * JOIN syntax.  We must do this after flattening JOIN alias
 298          * variables, since eliminating explicit JOIN nodes from the jointree
 299          * will cause get_relids_for_join() to fail.  But it should happen
 300          * after reduce_outer_joins, anyway.
 301          */
 302         parse->jointree = (FromExpr *)
 303                 simplify_jointree(parse, (Node *) parse->jointree);
 304
 305         /*
 306          * Do the main planning.  If we have an inherited target relation,
 307          * that needs special processing, else go straight to
 308          * grouping_planner.
 309          */
 310         if (parse->resultRelation &&
 311                 (lst = expand_inherited_rtentry(parse, parse->resultRelation,
 312                                                                                 false)) != NIL)
 313                 plan = inheritance_planner(parse, lst);
 314         else
 315                 plan = grouping_planner(parse, tuple_fraction);
 316
 317         /*
 318          * If any subplans were generated, or if we're inside a subplan, build
 319          * initPlan list and extParam/allParam sets for plan nodes.
 320          */
 321         if (PlannerPlanId != saved_planid || PlannerQueryLevel > 1)
 322         {
 323                 Cost            initplan_cost = 0;
 324
 325                 /* Prepare extParam/allParam sets for all nodes in tree */
 326                 SS_finalize_plan(plan, parse->rtable);
 327
 328                 /*
 329                  * SS_finalize_plan doesn't handle initPlans, so we have to
 330                  * manually attach them to the topmost plan node, and add their
 331                  * extParams to the topmost node's, too.
 332                  *
 333                  * We also add the total_cost of each initPlan to the startup cost of
 334                  * the top node.  This is a conservative overestimate, since in
 335                  * fact each initPlan might be executed later than plan startup,
 336                  * or even not at all.
 337                  */
 338                 plan->initPlan = PlannerInitPlan;
 339
 340                 foreach(lst, plan->initPlan)
 341                 {
 342                         SubPlan    *initplan = (SubPlan *) lfirst(lst);
 343
 344                         plan->extParam = bms_add_members(plan->extParam,
 345                                                                                          initplan->plan->extParam);
 346                         initplan_cost += initplan->plan->total_cost;
 347                 }
 348
 349                 plan->startup_cost += initplan_cost;
 350                 plan->total_cost += initplan_cost;
 351         }
 352
 353         /* Return to outer subquery context */
 354         PlannerQueryLevel--;
 355         PlannerInitPlan = saved_initplan;
 356         /* we do NOT restore PlannerPlanId; that's not an oversight! */
 357
 358         return plan;
 359 }
 360
 361 /*
 362  * preprocess_expression
 363  *              Do subquery_planner's preprocessing work for an expression,
 364  *              which can be a targetlist, a WHERE clause (including JOIN/ON
 365  *              conditions), or a HAVING clause.
 366  */
 367 static Node *
 368 preprocess_expression(Query *parse, Node *expr, int kind)
 369 {
 370         /*
 371          * If the query has any join RTEs, replace join alias variables with
 372          * base-relation variables. We must do this before sublink processing,
 373          * else sublinks expanded out from join aliases wouldn't get
 374          * processed.
 375          */
 376         if (parse->hasJoinRTEs)
 377                 expr = flatten_join_alias_vars(parse, expr);
 378
 379         /*
 380          * Simplify constant expressions.
 381          *
 382          * Note that at this point quals have not yet been converted to
 383          * implicit-AND form, so we can apply eval_const_expressions directly.
 384          */
 385         expr = eval_const_expressions(expr);
 386
 387         /*
 388          * If it's a qual or havingQual, canonicalize it, and convert it to
 389          * implicit-AND format.
 390          *
 391          * XXX Is there any value in re-applying eval_const_expressions after
 392          * canonicalize_qual?
 393          */
 394         if (kind == EXPRKIND_QUAL)
 395         {
 396                 expr = (Node *) canonicalize_qual((Expr *) expr, true);
 397
 398 #ifdef OPTIMIZER_DEBUG
 399                 printf("After canonicalize_qual()\n");
 400                 pprint(expr);
 401 #endif
 402         }
 403
 404         /* Expand SubLinks to SubPlans */
 405         if (parse->hasSubLinks)
 406                 expr = SS_process_sublinks(expr, (kind == EXPRKIND_QUAL));
 407
 408         /*
 409          * XXX do not insert anything here unless you have grokked the
 410          * comments in SS_replace_correlation_vars ...
 411          */
 412
 413         /* Replace uplevel vars with Param nodes */
 414         if (PlannerQueryLevel > 1)
 415                 expr = SS_replace_correlation_vars(expr);
 416
 417         return expr;
 418 }
 419
 420 /*
 421  * preprocess_qual_conditions
 422  *              Recursively scan the query's jointree and do subquery_planner's
 423  *              preprocessing work on each qual condition found therein.
 424  */
 425 static void
 426 preprocess_qual_conditions(Query *parse, Node *jtnode)
 427 {
 428         if (jtnode == NULL)
 429                 return;
 430         if (IsA(jtnode, RangeTblRef))
 431         {
 432                 /* nothing to do here */
 433         }
 434         else if (IsA(jtnode, FromExpr))
 435         {
 436                 FromExpr   *f = (FromExpr *) jtnode;
 437                 List       *l;
 438
 439                 foreach(l, f->fromlist)
 440                         preprocess_qual_conditions(parse, lfirst(l));
 441
 442                 f->quals = preprocess_expression(parse, f->quals, EXPRKIND_QUAL);
 443         }
 444         else if (IsA(jtnode, JoinExpr))
 445         {
 446                 JoinExpr   *j = (JoinExpr *) jtnode;
 447
 448                 preprocess_qual_conditions(parse, j->larg);
 449                 preprocess_qual_conditions(parse, j->rarg);
 450
 451                 j->quals = preprocess_expression(parse, j->quals, EXPRKIND_QUAL);
 452         }
 453         else
 454                 elog(ERROR, "unrecognized node type: %d",
 455                          (int) nodeTag(jtnode));
 456 }
 457
 458 /*--------------------
 459  * inheritance_planner
 460  *        Generate a plan in the case where the result relation is an
 461  *        inheritance set.
 462  *
 463  * We have to handle this case differently from cases where a source
 464  * relation is an inheritance set.      Source inheritance is expanded at
 465  * the bottom of the plan tree (see allpaths.c), but target inheritance
 466  * has to be expanded at the top.  The reason is that for UPDATE, each
 467  * target relation needs a different targetlist matching its own column
 468  * set.  (This is not so critical for DELETE, but for simplicity we treat
 469  * inherited DELETE the same way.)      Fortunately, the UPDATE/DELETE target
 470  * can never be the nullable side of an outer join, so it's OK to generate
 471  * the plan this way.
 472  *
 473  * parse is the querytree produced by the parser & rewriter.
 474  * inheritlist is an integer list of RT indexes for the result relation set.
 475  *
 476  * Returns a query plan.
 477  *--------------------
 478  */
 479 static Plan *
 480 inheritance_planner(Query *parse, List *inheritlist)
 481 {
 482         int                     parentRTindex = parse->resultRelation;
 483         Oid                     parentOID = getrelid(parentRTindex, parse->rtable);
 484         int                     mainrtlength = length(parse->rtable);
 485         List       *subplans = NIL;
 486         List       *tlist = NIL;
 487         List       *l;
 488
 489         foreach(l, inheritlist)
 490         {
 491                 int                     childRTindex = lfirsti(l);
 492                 Oid                     childOID = getrelid(childRTindex, parse->rtable);
 493                 int                     subrtlength;
 494                 Query      *subquery;
 495                 Plan       *subplan;
 496
 497                 /* Generate modified query with this rel as target */
 498                 subquery = (Query *) adjust_inherited_attrs((Node *) parse,
 499                                                                                                 parentRTindex, parentOID,
 500                                                                                                  childRTindex, childOID);
 501                 /* Generate plan */
 502                 subplan = grouping_planner(subquery, 0.0 /* retrieve all tuples */ );
 503                 subplans = lappend(subplans, subplan);
 504
 505                 /*
 506                  * It's possible that additional RTEs got added to the rangetable
 507                  * due to expansion of inherited source tables (see allpaths.c).
 508                  * If so, we must copy 'em back to the main parse tree's rtable.
 509                  *
 510                  * XXX my goodness this is ugly.  Really need to think about ways to
 511                  * rein in planner's habit of scribbling on its input.
 512                  */
 513                 subrtlength = length(subquery->rtable);
 514                 if (subrtlength > mainrtlength)
 515                 {
 516                         List       *subrt = subquery->rtable;
 517
 518                         while (mainrtlength-- > 0)      /* wish we had nthcdr() */
 519                                 subrt = lnext(subrt);
 520                         parse->rtable = nconc(parse->rtable, subrt);
 521                         mainrtlength = subrtlength;
 522                 }
 523                 /* Save preprocessed tlist from first rel for use in Append */
 524                 if (tlist == NIL)
 525                         tlist = subplan->targetlist;
 526         }
 527
 528         /* Save the target-relations list for the executor, too */
 529         parse->resultRelations = inheritlist;
 530
 531         /* Mark result as unordered (probably unnecessary) */
 532         parse->query_pathkeys = NIL;
 533
 534         return (Plan *) make_append(subplans, true, tlist);
 535 }
 536
 537 /*--------------------
 538  * grouping_planner
 539  *        Perform planning steps related to grouping, aggregation, etc.
 540  *        This primarily means adding top-level processing to the basic
 541  *        query plan produced by query_planner.
 542  *
 543  * parse is the querytree produced by the parser & rewriter.
 544  * tuple_fraction is the fraction of tuples we expect will be retrieved
 545  *
 546  * tuple_fraction is interpreted as follows:
 547  *        0: expect all tuples to be retrieved (normal case)
 548  *        0 < tuple_fraction < 1: expect the given fraction of tuples available
 549  *              from the plan to be retrieved
 550  *        tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
 551  *              expected to be retrieved (ie, a LIMIT specification)
 552  *
 553  * Returns a query plan.  Also, parse->query_pathkeys is returned as the
 554  * actual output ordering of the plan (in pathkey format).
 555  *--------------------
 556  */
 557 static Plan *
 558 grouping_planner(Query *parse, double tuple_fraction)
 559 {
 560         List       *tlist = parse->targetList;
 561         Plan       *result_plan;
 562         List       *current_pathkeys;
 563         List       *sort_pathkeys;
 564
 565         if (parse->setOperations)
 566         {
 567                 /*
 568                  * Construct the plan for set operations.  The result will not
 569                  * need any work except perhaps a top-level sort and/or LIMIT.
 570                  */
 571                 result_plan = plan_set_operations(parse);
 572
 573                 /*
 574                  * We should not need to call preprocess_targetlist, since we must
 575                  * be in a SELECT query node.  Instead, use the targetlist
 576                  * returned by plan_set_operations (since this tells whether it
 577                  * returned any resjunk columns!), and transfer any sort key
 578                  * information from the original tlist.
 579                  */
 580                 Assert(parse->commandType == CMD_SELECT);
 581
 582                 tlist = postprocess_setop_tlist(result_plan->targetlist, tlist);
 583
 584                 /*
 585                  * Can't handle FOR UPDATE here (parser should have checked
 586                  * already, but let's make sure).
 587                  */
 588                 if (parse->rowMarks)
 589                         ereport(ERROR,
 590                                         (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
 591                                          errmsg("SELECT FOR UPDATE is not allowed with UNION/INTERSECT/EXCEPT")));
 592
 593                 /*
 594                  * We set current_pathkeys NIL indicating we do not know sort
 595                  * order.  This is correct when the top set operation is UNION
 596                  * ALL, since the appended-together results are unsorted even if
 597                  * the subplans were sorted.  For other set operations we could be
 598                  * smarter --- room for future improvement!
 599                  */
 600                 current_pathkeys = NIL;
 601
 602                 /*
 603                  * Calculate pathkeys that represent ordering requirements
 604                  */
 605                 sort_pathkeys = make_pathkeys_for_sortclauses(parse->sortClause,
 606                                                                                                           tlist);
 607                 sort_pathkeys = canonicalize_pathkeys(parse, sort_pathkeys);
 608         }
 609         else
 610         {
 611                 /* No set operations, do regular planning */
 612                 List       *sub_tlist;
 613                 List       *group_pathkeys;
 614                 AttrNumber *groupColIdx = NULL;
 615                 bool            need_tlist_eval = true;
 616                 QualCost        tlist_cost;
 617                 double          sub_tuple_fraction;
 618                 Path       *cheapest_path;
 619                 Path       *sorted_path;
 620                 double          dNumGroups = 0;
 621                 long            numGroups = 0;
 622                 int                     numAggs = 0;
 623                 int                     numGroupCols = length(parse->groupClause);
 624                 bool            use_hashed_grouping = false;
 625
 626                 /* Preprocess targetlist in case we are inside an INSERT/UPDATE. */
 627                 tlist = preprocess_targetlist(tlist,
 628                                                                           parse->commandType,
 629                                                                           parse->resultRelation,
 630                                                                           parse->rtable);
 631
 632                 /*
 633                  * Add TID targets for rels selected FOR UPDATE (should this be
 634                  * done in preprocess_targetlist?).  The executor uses the TID to
 635                  * know which rows to lock, much as for UPDATE or DELETE.
 636                  */
 637                 if (parse->rowMarks)
 638                 {
 639                         List       *l;
 640
 641                         /*
 642                          * We've got trouble if the FOR UPDATE appears inside
 643                          * grouping, since grouping renders a reference to individual
 644                          * tuple CTIDs invalid.  This is also checked at parse time,
 645                          * but that's insufficient because of rule substitution, query
 646                          * pullup, etc.
 647                          */
 648                         CheckSelectForUpdate(parse);
 649
 650                         /*
 651                          * Currently the executor only supports FOR UPDATE at top
 652                          * level
 653                          */
 654                         if (PlannerQueryLevel > 1)
 655                                 ereport(ERROR,
 656                                                 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
 657                                                  errmsg("SELECT FOR UPDATE is not allowed in subqueries")));
 658
 659                         foreach(l, parse->rowMarks)
 660                         {
 661                                 Index           rti = lfirsti(l);
 662                                 char       *resname;
 663                                 Resdom     *resdom;
 664                                 Var                *var;
 665                                 TargetEntry *ctid;
 666
 667                                 resname = (char *) palloc(32);
 668                                 snprintf(resname, 32, "ctid%u", rti);
 669                                 resdom = makeResdom(length(tlist) + 1,
 670                                                                         TIDOID,
 671                                                                         -1,
 672                                                                         resname,
 673                                                                         true);
 674
 675                                 var = makeVar(rti,
 676                                                           SelfItemPointerAttributeNumber,
 677                                                           TIDOID,
 678                                                           -1,
 679                                                           0);
 680
 681                                 ctid = makeTargetEntry(resdom, (Expr *) var);
 682                                 tlist = lappend(tlist, ctid);
 683                         }
 684                 }
 685
 686                 /*
 687                  * Generate appropriate target list for subplan; may be different
 688                  * from tlist if grouping or aggregation is needed.
 689                  */
 690                 sub_tlist = make_subplanTargetList(parse, tlist,
 691                                                                                  &groupColIdx, &need_tlist_eval);
 692
 693                 /*
 694                  * Calculate pathkeys that represent grouping/ordering
 695                  * requirements
 696                  */
 697                 group_pathkeys = make_pathkeys_for_sortclauses(parse->groupClause,
 698                                                                                                            tlist);
 699                 sort_pathkeys = make_pathkeys_for_sortclauses(parse->sortClause,
 700                                                                                                           tlist);
 701
 702                 /*
 703                  * Will need actual number of aggregates for estimating costs.
 704                  * Also, it's possible that optimization has eliminated all
 705                  * aggregates, and we may as well check for that here.
 706                  *
 707                  * Note: we do not attempt to detect duplicate aggregates here; a
 708                  * somewhat-overestimated count is okay for our present purposes.
 709                  */
 710                 if (parse->hasAggs)
 711                 {
 712                         numAggs = count_agg_clause((Node *) tlist) +
 713                                 count_agg_clause(parse->havingQual);
 714                         if (numAggs == 0)
 715                                 parse->hasAggs = false;
 716                 }
 717
 718                 /*
 719                  * Figure out whether we need a sorted result from query_planner.
 720                  *
 721                  * If we have a GROUP BY clause, then we want a result sorted
 722                  * properly for grouping.  Otherwise, if there is an ORDER BY
 723                  * clause, we want to sort by the ORDER BY clause.      (Note: if we
 724                  * have both, and ORDER BY is a superset of GROUP BY, it would be
 725                  * tempting to request sort by ORDER BY --- but that might just
 726                  * leave us failing to exploit an available sort order at all.
 727                  * Needs more thought...)
 728                  */
 729                 if (parse->groupClause)
 730                         parse->query_pathkeys = group_pathkeys;
 731                 else if (parse->sortClause)
 732                         parse->query_pathkeys = sort_pathkeys;
 733                 else
 734                         parse->query_pathkeys = NIL;
 735
 736                 /*
 737                  * Adjust tuple_fraction if we see that we are going to apply
 738                  * limiting/grouping/aggregation/etc.  This is not overridable by
 739                  * the caller, since it reflects plan actions that this routine
 740                  * will certainly take, not assumptions about context.
 741                  */
 742                 if (parse->limitCount != NULL)
 743                 {
 744                         /*
 745                          * A LIMIT clause limits the absolute number of tuples
 746                          * returned. However, if it's not a constant LIMIT then we
 747                          * have to punt; for lack of a better idea, assume 10% of the
 748                          * plan's result is wanted.
 749                          */
 750                         double          limit_fraction = 0.0;
 751
 752                         if (IsA(parse->limitCount, Const))
 753                         {
 754                                 Const      *limitc = (Const *) parse->limitCount;
 755                                 int32           count = DatumGetInt32(limitc->constvalue);
 756
 757                                 /*
 758                                  * A NULL-constant LIMIT represents "LIMIT ALL", which we
 759                                  * treat the same as no limit (ie, expect to retrieve all
 760                                  * the tuples).
 761                                  */
 762                                 if (!limitc->constisnull && count > 0)
 763                                 {
 764                                         limit_fraction = (double) count;
 765                                         /* We must also consider the OFFSET, if present */
 766                                         if (parse->limitOffset != NULL)
 767                                         {
 768                                                 if (IsA(parse->limitOffset, Const))
 769                                                 {
 770                                                         int32           offset;
 771
 772                                                         limitc = (Const *) parse->limitOffset;
 773                                                         offset = DatumGetInt32(limitc->constvalue);
 774                                                         if (!limitc->constisnull && offset > 0)
 775                                                                 limit_fraction += (double) offset;
 776                                                 }
 777                                                 else
 778                                                 {
 779                                                         /* OFFSET is an expression ... punt ... */
 780                                                         limit_fraction = 0.10;
 781                                                 }
 782                                         }
 783                                 }
 784                         }
 785                         else
 786                         {
 787                                 /* LIMIT is an expression ... punt ... */
 788                                 limit_fraction = 0.10;
 789                         }
 790
 791                         if (limit_fraction > 0.0)
 792                         {
 793                                 /*
 794                                  * If we have absolute limits from both caller and LIMIT,
 795                                  * use the smaller value; if one is fractional and the
 796                                  * other absolute, treat the fraction as a fraction of the
 797                                  * absolute value; else we can multiply the two fractions
 798                                  * together.
 799                                  */
 800                                 if (tuple_fraction >= 1.0)
 801                                 {
 802                                         if (limit_fraction >= 1.0)
 803                                         {
 804                                                 /* both absolute */
 805                                                 tuple_fraction = Min(tuple_fraction, limit_fraction);
 806                                         }
 807                                         else
 808                                         {
 809                                                 /* caller absolute, limit fractional */
 810                                                 tuple_fraction *= limit_fraction;
 811                                                 if (tuple_fraction < 1.0)
 812                                                         tuple_fraction = 1.0;
 813                                         }
 814                                 }
 815                                 else if (tuple_fraction > 0.0)
 816                                 {
 817                                         if (limit_fraction >= 1.0)
 818                                         {
 819                                                 /* caller fractional, limit absolute */
 820                                                 tuple_fraction *= limit_fraction;
 821                                                 if (tuple_fraction < 1.0)
 822                                                         tuple_fraction = 1.0;
 823                                         }
 824                                         else
 825                                         {
 826                                                 /* both fractional */
 827                                                 tuple_fraction *= limit_fraction;
 828                                         }
 829                                 }
 830                                 else
 831                                 {
 832                                         /* no info from caller, just use limit */
 833                                         tuple_fraction = limit_fraction;
 834                                 }
 835                         }
 836                 }
 837
 838                 /*
 839                  * With grouping or aggregation, the tuple fraction to pass to
 840                  * query_planner() may be different from what it is at top level.
 841                  */
 842                 sub_tuple_fraction = tuple_fraction;
 843
 844                 if (parse->groupClause)
 845                 {
 846                         /*
 847                          * In GROUP BY mode, we have the little problem that we don't
 848                          * really know how many input tuples will be needed to make a
 849                          * group, so we can't translate an output LIMIT count into an
 850                          * input count.  For lack of a better idea, assume 25% of the
 851                          * input data will be processed if there is any output limit.
 852                          * However, if the caller gave us a fraction rather than an
 853                          * absolute count, we can keep using that fraction (which
 854                          * amounts to assuming that all the groups are about the same
 855                          * size).
 856                          */
 857                         if (sub_tuple_fraction >= 1.0)
 858                                 sub_tuple_fraction = 0.25;
 859
 860                         /*
 861                          * If both GROUP BY and ORDER BY are specified, we will need
 862                          * two levels of sort --- and, therefore, certainly need to
 863                          * read all the input tuples --- unless ORDER BY is a subset
 864                          * of GROUP BY.  (We have not yet canonicalized the pathkeys,
 865                          * so must use the slower noncanonical comparison method.)
 866                          */
 867                         if (parse->groupClause && parse->sortClause &&
 868                                 !noncanonical_pathkeys_contained_in(sort_pathkeys,
 869                                                                                                         group_pathkeys))
 870                                 sub_tuple_fraction = 0.0;
 871                 }
 872                 else if (parse->hasAggs)
 873                 {
 874                         /*
 875                          * Ungrouped aggregate will certainly want all the input
 876                          * tuples.
 877                          */
 878                         sub_tuple_fraction = 0.0;
 879                 }
 880                 else if (parse->distinctClause)
 881                 {
 882                         /*
 883                          * SELECT DISTINCT, like GROUP, will absorb an unpredictable
 884                          * number of input tuples per output tuple.  Handle the same
 885                          * way.
 886                          */
 887                         if (sub_tuple_fraction >= 1.0)
 888                                 sub_tuple_fraction = 0.25;
 889                 }
 890
 891                 /*
 892                  * Generate the best unsorted and presorted paths for this Query
 893                  * (but note there may not be any presorted path).
 894                  */
 895                 query_planner(parse, sub_tlist, sub_tuple_fraction,
 896                                           &cheapest_path, &sorted_path);
 897
 898                 /*
 899                  * We couldn't canonicalize group_pathkeys and sort_pathkeys
 900                  * before running query_planner(), so do it now.
 901                  */
 902                 group_pathkeys = canonicalize_pathkeys(parse, group_pathkeys);
 903                 sort_pathkeys = canonicalize_pathkeys(parse, sort_pathkeys);
 904
 905                 /*
 906                  * Consider whether we might want to use hashed grouping.
 907                  */
 908                 if (parse->groupClause)
 909                 {
 910                         List       *groupExprs;
 911                         double          cheapest_path_rows;
 912                         int                     cheapest_path_width;
 913
 914                         /*
 915                          * Beware in this section of the possibility that
 916                          * cheapest_path->parent is NULL.  This could happen if user
 917                          * does something silly like SELECT 'foo' GROUP BY 1;
 918                          */
 919                         if (cheapest_path->parent)
 920                         {
 921                                 cheapest_path_rows = cheapest_path->parent->rows;
 922                                 cheapest_path_width = cheapest_path->parent->width;
 923                         }
 924                         else
 925                         {
 926                                 cheapest_path_rows = 1; /* assume non-set result */
 927                                 cheapest_path_width = 100;              /* arbitrary */
 928                         }
 929
 930                         /*
 931                          * Always estimate the number of groups.  We can't do this
 932                          * until after running query_planner(), either.
 933                          */
 934                         groupExprs = get_sortgrouplist_exprs(parse->groupClause,
 935                                                                                                  parse->targetList);
 936                         dNumGroups = estimate_num_groups(parse,
 937                                                                                          groupExprs,
 938                                                                                          cheapest_path_rows);
 939                         /* Also want it as a long int --- but 'ware overflow! */
 940                         numGroups = (long) Min(dNumGroups, (double) LONG_MAX);
 941
 942                         /*
 943                          * Check can't-do-it conditions, including whether the
 944                          * grouping operators are hashjoinable.
 945                          *
 946                          * Executor doesn't support hashed aggregation with DISTINCT
 947                          * aggregates.  (Doing so would imply storing *all* the input
 948                          * values in the hash table, which seems like a certain
 949                          * loser.)
 950                          */
 951                         if (!enable_hashagg || !hash_safe_grouping(parse))
 952                                 use_hashed_grouping = false;
 953                         else if (parse->hasAggs &&
 954                                          (contain_distinct_agg_clause((Node *) tlist) ||
 955                                           contain_distinct_agg_clause(parse->havingQual)))
 956                                 use_hashed_grouping = false;
 957                         else
 958                         {
 959                                 /*
 960                                  * Use hashed grouping if (a) we think we can fit the
 961                                  * hashtable into SortMem, *and* (b) the estimated cost is
 962                                  * no more than doing it the other way.  While avoiding
 963                                  * the need for sorted input is usually a win, the fact
 964                                  * that the output won't be sorted may be a loss; so we
 965                                  * need to do an actual cost comparison.
 966                                  *
 967                                  * In most cases we have no good way to estimate the size of
 968                                  * the transition value needed by an aggregate;
 969                                  * arbitrarily assume it is 100 bytes.  Also set the
 970                                  * overhead per hashtable entry at 64 bytes.
 971                                  */
 972                                 int                     hashentrysize = cheapest_path_width + 64 + numAggs * 100;
 973
 974                                 if (hashentrysize * dNumGroups <= SortMem * 1024L)
 975                                 {
 976                                         /*
 977                                          * Okay, do the cost comparison.  We need to consider
 978                                          * cheapest_path + hashagg [+ final sort] versus
 979                                          * either cheapest_path [+ sort] + group or agg [+
 980                                          * final sort] or presorted_path + group or agg [+
 981                                          * final sort] where brackets indicate a step that may
 982                                          * not be needed. We assume query_planner() will have
 983                                          * returned a presorted path only if it's a winner
 984                                          * compared to cheapest_path for this purpose.
 985                                          *
 986                                          * These path variables are dummies that just hold cost
 987                                          * fields; we don't make actual Paths for these steps.
 988                                          */
 989                                         Path            hashed_p;
 990                                         Path            sorted_p;
 991
 992                                         cost_agg(&hashed_p, parse,
 993                                                          AGG_HASHED, numAggs,
 994                                                          numGroupCols, dNumGroups,
 995                                                          cheapest_path->startup_cost,
 996                                                          cheapest_path->total_cost,
 997                                                          cheapest_path_rows);
 998                                         /* Result of hashed agg is always unsorted */
 999                                         if (sort_pathkeys)
1000                                                 cost_sort(&hashed_p, parse, sort_pathkeys,
1001                                                                   hashed_p.total_cost,
1002                                                                   dNumGroups,
1003                                                                   cheapest_path_width);
1004
1005                                         if (sorted_path)
1006                                         {
1007                                                 sorted_p.startup_cost = sorted_path->startup_cost;
1008                                                 sorted_p.total_cost = sorted_path->total_cost;
1009                                                 current_pathkeys = sorted_path->pathkeys;
1010                                         }
1011                                         else
1012                                         {
1013                                                 sorted_p.startup_cost = cheapest_path->startup_cost;
1014                                                 sorted_p.total_cost = cheapest_path->total_cost;
1015                                                 current_pathkeys = cheapest_path->pathkeys;
1016                                         }
1017                                         if (!pathkeys_contained_in(group_pathkeys,
1018                                                                                            current_pathkeys))
1019                                         {
1020                                                 cost_sort(&sorted_p, parse, group_pathkeys,
1021                                                                   sorted_p.total_cost,
1022                                                                   cheapest_path_rows,
1023                                                                   cheapest_path_width);
1024                                                 current_pathkeys = group_pathkeys;
1025                                         }
1026                                         if (parse->hasAggs)
1027                                                 cost_agg(&sorted_p, parse,
1028                                                                  AGG_SORTED, numAggs,
1029                                                                  numGroupCols, dNumGroups,
1030                                                                  sorted_p.startup_cost,
1031                                                                  sorted_p.total_cost,
1032                                                                  cheapest_path_rows);
1033                                         else
1034                                                 cost_group(&sorted_p, parse,
1035                                                                    numGroupCols, dNumGroups,
1036                                                                    sorted_p.startup_cost,
1037                                                                    sorted_p.total_cost,
1038                                                                    cheapest_path_rows);
1039                                         /* The Agg or Group node will preserve ordering */
1040                                         if (sort_pathkeys &&
1041                                                 !pathkeys_contained_in(sort_pathkeys,
1042                                                                                            current_pathkeys))
1043                                         {
1044                                                 cost_sort(&sorted_p, parse, sort_pathkeys,
1045                                                                   sorted_p.total_cost,
1046                                                                   dNumGroups,
1047                                                                   cheapest_path_width);
1048                                         }
1049
1050                                         /*
1051                                          * Now make the decision using the top-level tuple
1052                                          * fraction.  First we have to convert an absolute
1053                                          * count (LIMIT) into fractional form.
1054                                          */
1055                                         if (tuple_fraction >= 1.0)
1056                                                 tuple_fraction /= dNumGroups;
1057
1058                                         if (compare_fractional_path_costs(&hashed_p, &sorted_p,
1059                                                                                                           tuple_fraction) < 0)
1060                                         {
1061                                                 /* Hashed is cheaper, so use it */
1062                                                 use_hashed_grouping = true;
1063                                         }
1064                                 }
1065                         }
1066                 }
1067
1068                 /*
1069                  * Select the best path and create a plan to execute it.
1070                  *
1071                  * If we are doing hashed grouping, we will always read all the input
1072                  * tuples, so use the cheapest-total path.      Otherwise, trust
1073                  * query_planner's decision about which to use.
1074                  */
1075                 if (sorted_path && !use_hashed_grouping)
1076                 {
1077                         result_plan = create_plan(parse, sorted_path);
1078                         current_pathkeys = sorted_path->pathkeys;
1079                 }
1080                 else
1081                 {
1082                         result_plan = create_plan(parse, cheapest_path);
1083                         current_pathkeys = cheapest_path->pathkeys;
1084                 }
1085
1086                 /*
1087                  * create_plan() returns a plan with just a "flat" tlist of
1088                  * required Vars.  Usually we need to insert the sub_tlist as the
1089                  * tlist of the top plan node.  However, we can skip that if we
1090                  * determined that whatever query_planner chose to return will be
1091                  * good enough.
1092                  */
1093                 if (need_tlist_eval)
1094                 {
1095                         /*
1096                          * If the top-level plan node is one that cannot do expression
1097                          * evaluation, we must insert a Result node to project the
1098                          * desired tlist. Currently, the only plan node we might see
1099                          * here that falls into that category is Append.
1100                          */
1101                         if (IsA(result_plan, Append))
1102                         {
1103                                 result_plan = (Plan *) make_result(sub_tlist, NULL,
1104                                                                                                    result_plan);
1105                         }
1106                         else
1107                         {
1108                                 /*
1109                                  * Otherwise, just replace the subplan's flat tlist with
1110                                  * the desired tlist.
1111                                  */
1112                                 result_plan->targetlist = sub_tlist;
1113                         }
1114
1115                         /*
1116                          * Also, account for the cost of evaluation of the sub_tlist.
1117                          *
1118                          * Up to now, we have only been dealing with "flat" tlists,
1119                          * containing just Vars.  So their evaluation cost is zero
1120                          * according to the model used by cost_qual_eval() (or if you
1121                          * prefer, the cost is factored into cpu_tuple_cost).  Thus we
1122                          * can avoid accounting for tlist cost throughout
1123                          * query_planner() and subroutines.  But now we've inserted a
1124                          * tlist that might contain actual operators, sub-selects, etc
1125                          * --- so we'd better account for its cost.
1126                          *
1127                          * Below this point, any tlist eval cost for added-on nodes
1128                          * should be accounted for as we create those nodes.
1129                          * Presently, of the node types we can add on, only Agg and
1130                          * Group project new tlists (the rest just copy their input
1131                          * tuples) --- so make_agg() and make_group() are responsible
1132                          * for computing the added cost.
1133                          */
1134                         cost_qual_eval(&tlist_cost, sub_tlist);
1135                         result_plan->startup_cost += tlist_cost.startup;
1136                         result_plan->total_cost += tlist_cost.startup +
1137                                 tlist_cost.per_tuple * result_plan->plan_rows;
1138                 }
1139                 else
1140                 {
1141                         /*
1142                          * Since we're using query_planner's tlist and not the one
1143                          * make_subplanTargetList calculated, we have to refigure any
1144                          * grouping-column indexes make_subplanTargetList computed.
1145                          */
1146                         locate_grouping_columns(parse, tlist, result_plan->targetlist,
1147                                                                         groupColIdx);
1148                 }
1149
1150                 /*
1151                  * Insert AGG or GROUP node if needed, plus an explicit sort step
1152                  * if necessary.
1153                  *
1154                  * HAVING clause, if any, becomes qual of the Agg node
1155                  */
1156                 if (use_hashed_grouping)
1157                 {
1158                         /* Hashed aggregate plan --- no sort needed */
1159                         result_plan = (Plan *) make_agg(parse,
1160                                                                                         tlist,
1161                                                                                         (List *) parse->havingQual,
1162                                                                                         AGG_HASHED,
1163                                                                                         numGroupCols,
1164                                                                                         groupColIdx,
1165                                                                                         numGroups,
1166                                                                                         numAggs,
1167                                                                                         result_plan);
1168                         /* Hashed aggregation produces randomly-ordered results */
1169                         current_pathkeys = NIL;
1170                 }
1171                 else if (parse->hasAggs)
1172                 {
1173                         /* Plain aggregate plan --- sort if needed */
1174                         AggStrategy aggstrategy;
1175
1176                         if (parse->groupClause)
1177                         {
1178                                 if (!pathkeys_contained_in(group_pathkeys, current_pathkeys))
1179                                 {
1180                                         result_plan = (Plan *)
1181                                                 make_sort_from_groupcols(parse,
1182                                                                                                  parse->groupClause,
1183                                                                                                  groupColIdx,
1184                                                                                                  result_plan);
1185                                         current_pathkeys = group_pathkeys;
1186                                 }
1187                                 aggstrategy = AGG_SORTED;
1188
1189                                 /*
1190                                  * The AGG node will not change the sort ordering of its
1191                                  * groups, so current_pathkeys describes the result too.
1192                                  */
1193                         }
1194                         else
1195                         {
1196                                 aggstrategy = AGG_PLAIN;
1197                                 /* Result will be only one row anyway; no sort order */
1198                                 current_pathkeys = NIL;
1199                         }
1200
1201                         result_plan = (Plan *) make_agg(parse,
1202                                                                                         tlist,
1203                                                                                         (List *) parse->havingQual,
1204                                                                                         aggstrategy,
1205                                                                                         numGroupCols,
1206                                                                                         groupColIdx,
1207                                                                                         numGroups,
1208                                                                                         numAggs,
1209                                                                                         result_plan);
1210                 }
1211                 else
1212                 {
1213                         /*
1214                          * If there are no Aggs, we shouldn't have any HAVING qual
1215                          * anymore
1216                          */
1217                         Assert(parse->havingQual == NULL);
1218
1219                         /*
1220                          * If we have a GROUP BY clause, insert a group node (plus the
1221                          * appropriate sort node, if necessary).
1222                          */
1223                         if (parse->groupClause)
1224                         {
1225                                 /*
1226                                  * Add an explicit sort if we couldn't make the path come
1227                                  * out the way the GROUP node needs it.
1228                                  */
1229                                 if (!pathkeys_contained_in(group_pathkeys, current_pathkeys))
1230                                 {
1231                                         result_plan = (Plan *)
1232                                                 make_sort_from_groupcols(parse,
1233                                                                                                  parse->groupClause,
1234                                                                                                  groupColIdx,
1235                                                                                                  result_plan);
1236                                         current_pathkeys = group_pathkeys;
1237                                 }
1238
1239                                 result_plan = (Plan *) make_group(parse,
1240                                                                                                   tlist,
1241                                                                                                   numGroupCols,
1242                                                                                                   groupColIdx,
1243                                                                                                   dNumGroups,
1244                                                                                                   result_plan);
1245                                 /* The Group node won't change sort ordering */
1246                         }
1247                 }
1248         }                                                       /* end of if (setOperations) */
1249
1250         /*
1251          * If we were not able to make the plan come out in the right order,
1252          * add an explicit sort step.
1253          */
1254         if (parse->sortClause)
1255         {
1256                 if (!pathkeys_contained_in(sort_pathkeys, current_pathkeys))
1257                 {
1258                         result_plan = (Plan *)
1259                                 make_sort_from_sortclauses(parse,
1260                                                                                    tlist,
1261                                                                                    result_plan,
1262                                                                                    parse->sortClause);
1263                         current_pathkeys = sort_pathkeys;
1264                 }
1265         }
1266
1267         /*
1268          * If there is a DISTINCT clause, add the UNIQUE node.
1269          */
1270         if (parse->distinctClause)
1271         {
1272                 result_plan = (Plan *) make_unique(tlist, result_plan,
1273                                                                                    parse->distinctClause);
1274
1275                 /*
1276                  * If there was grouping or aggregation, leave plan_rows as-is
1277                  * (ie, assume the result was already mostly unique).  If not,
1278                  * it's reasonable to assume the UNIQUE filter has effects
1279                  * comparable to GROUP BY.
1280                  */
1281                 if (!parse->groupClause && !parse->hasAggs)
1282                 {
1283                         List       *distinctExprs;
1284
1285                         distinctExprs = get_sortgrouplist_exprs(parse->distinctClause,
1286                                                                                                         parse->targetList);
1287                         result_plan->plan_rows = estimate_num_groups(parse,
1288                                                                                                                  distinctExprs,
1289                                                                                                  result_plan->plan_rows);
1290                 }
1291         }
1292
1293         /*
1294          * Finally, if there is a LIMIT/OFFSET clause, add the LIMIT node.
1295          */
1296         if (parse->limitOffset || parse->limitCount)
1297         {
1298                 result_plan = (Plan *) make_limit(tlist, result_plan,
1299                                                                                   parse->limitOffset,
1300                                                                                   parse->limitCount);
1301         }
1302
1303         /*
1304          * Return the actual output ordering in query_pathkeys for possible
1305          * use by an outer query level.
1306          */
1307         parse->query_pathkeys = current_pathkeys;
1308
1309         return result_plan;
1310 }
1311
1312 /*
1313  * hash_safe_grouping - are grouping operators hashable?
1314  *
1315  * We assume hashed aggregation will work if the datatype's equality operator
1316  * is marked hashjoinable.
1317  */
1318 static bool
1319 hash_safe_grouping(Query *parse)
1320 {
1321         List       *gl;
1322
1323         foreach(gl, parse->groupClause)
1324         {
1325                 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1326                 TargetEntry *tle = get_sortgroupclause_tle(grpcl, parse->targetList);
1327                 Operator        optup;
1328                 bool            oprcanhash;
1329
1330                 optup = equality_oper(tle->resdom->restype, true);
1331                 if (!optup)
1332                         return false;
1333                 oprcanhash = ((Form_pg_operator) GETSTRUCT(optup))->oprcanhash;
1334                 ReleaseSysCache(optup);
1335                 if (!oprcanhash)
1336                         return false;
1337         }
1338         return true;
1339 }
1340
1341 /*---------------
1342  * make_subplanTargetList
1343  *        Generate appropriate target list when grouping is required.
1344  *
1345  * When grouping_planner inserts Aggregate or Group plan nodes above
1346  * the result of query_planner, we typically want to pass a different
1347  * target list to query_planner than the outer plan nodes should have.
1348  * This routine generates the correct target list for the subplan.
1349  *
1350  * The initial target list passed from the parser already contains entries
1351  * for all ORDER BY and GROUP BY expressions, but it will not have entries
1352  * for variables used only in HAVING clauses; so we need to add those
1353  * variables to the subplan target list.  Also, if we are doing either
1354  * grouping or aggregation, we flatten all expressions except GROUP BY items
1355  * into their component variables; the other expressions will be computed by
1356  * the inserted nodes rather than by the subplan.  For example,
1357  * given a query like
1358  *              SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
1359  * we want to pass this targetlist to the subplan:
1360  *              a,b,c,d,a+b
1361  * where the a+b target will be used by the Sort/Group steps, and the
1362  * other targets will be used for computing the final results.  (In the
1363  * above example we could theoretically suppress the a and b targets and
1364  * pass down only c,d,a+b, but it's not really worth the trouble to
1365  * eliminate simple var references from the subplan.  We will avoid doing
1366  * the extra computation to recompute a+b at the outer level; see
1367  * replace_vars_with_subplan_refs() in setrefs.c.)
1368  *
1369  * If we are grouping or aggregating, *and* there are no non-Var grouping
1370  * expressions, then the returned tlist is effectively dummy; we do not
1371  * need to force it to be evaluated, because all the Vars it contains
1372  * should be present in the output of query_planner anyway.
1373  *
1374  * 'parse' is the query being processed.
1375  * 'tlist' is the query's target list.
1376  * 'groupColIdx' receives an array of column numbers for the GROUP BY
1377  *                      expressions (if there are any) in the subplan's target list.
1378  * 'need_tlist_eval' is set true if we really need to evaluate the
1379  *                      result tlist.
1380  *
1381  * The result is the targetlist to be passed to the subplan.
1382  *---------------
1383  */
1384 static List *
1385 make_subplanTargetList(Query *parse,
1386                                            List *tlist,
1387                                            AttrNumber **groupColIdx,
1388                                            bool *need_tlist_eval)
1389 {
1390         List       *sub_tlist;
1391         List       *extravars;
1392         int                     numCols;
1393
1394         *groupColIdx = NULL;
1395
1396         /*
1397          * If we're not grouping or aggregating, nothing to do here;
1398          * query_planner should receive the unmodified target list.
1399          */
1400         if (!parse->hasAggs && !parse->groupClause)
1401         {
1402                 *need_tlist_eval = true;
1403                 return tlist;
1404         }
1405
1406         /*
1407          * Otherwise, start with a "flattened" tlist (having just the vars
1408          * mentioned in the targetlist and HAVING qual --- but not upper-
1409          * level Vars; they will be replaced by Params later on).
1410          */
1411         sub_tlist = flatten_tlist(tlist);
1412         extravars = pull_var_clause(parse->havingQual, false);
1413         sub_tlist = add_to_flat_tlist(sub_tlist, extravars);
1414         freeList(extravars);
1415         *need_tlist_eval = false;       /* only eval if not flat tlist */
1416
1417         /*
1418          * If grouping, create sub_tlist entries for all GROUP BY expressions
1419          * (GROUP BY items that are simple Vars should be in the list
1420          * already), and make an array showing where the group columns are in
1421          * the sub_tlist.
1422          */
1423         numCols = length(parse->groupClause);
1424         if (numCols > 0)
1425         {
1426                 int                     keyno = 0;
1427                 AttrNumber *grpColIdx;
1428                 List       *gl;
1429
1430                 grpColIdx = (AttrNumber *) palloc(sizeof(AttrNumber) * numCols);
1431                 *groupColIdx = grpColIdx;
1432
1433                 foreach(gl, parse->groupClause)
1434                 {
1435                         GroupClause *grpcl = (GroupClause *) lfirst(gl);
1436                         Node       *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1437                         TargetEntry *te = NULL;
1438                         List       *sl;
1439
1440                         /* Find or make a matching sub_tlist entry */
1441                         foreach(sl, sub_tlist)
1442                         {
1443                                 te = (TargetEntry *) lfirst(sl);
1444                                 if (equal(groupexpr, te->expr))
1445                                         break;
1446                         }
1447                         if (!sl)
1448                         {
1449                                 te = makeTargetEntry(makeResdom(length(sub_tlist) + 1,
1450                                                                                                 exprType(groupexpr),
1451                                                                                                 exprTypmod(groupexpr),
1452                                                                                                 NULL,
1453                                                                                                 false),
1454                                                                          (Expr *) groupexpr);
1455                                 sub_tlist = lappend(sub_tlist, te);
1456                                 *need_tlist_eval = true;                /* it's not flat anymore */
1457                         }
1458
1459                         /* and save its resno */
1460                         grpColIdx[keyno++] = te->resdom->resno;
1461                 }
1462         }
1463
1464         return sub_tlist;
1465 }
1466
1467 /*
1468  * locate_grouping_columns
1469  *              Locate grouping columns in the tlist chosen by query_planner.
1470  *
1471  * This is only needed if we don't use the sub_tlist chosen by
1472  * make_subplanTargetList.      We have to forget the column indexes found
1473  * by that routine and re-locate the grouping vars in the real sub_tlist.
1474  */
1475 static void
1476 locate_grouping_columns(Query *parse,
1477                                                 List *tlist,
1478                                                 List *sub_tlist,
1479                                                 AttrNumber *groupColIdx)
1480 {
1481         int                     keyno = 0;
1482         List       *gl;
1483
1484         /*
1485          * No work unless grouping.
1486          */
1487         if (!parse->groupClause)
1488         {
1489                 Assert(groupColIdx == NULL);
1490                 return;
1491         }
1492         Assert(groupColIdx != NULL);
1493
1494         foreach(gl, parse->groupClause)
1495         {
1496                 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1497                 Node       *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1498                 TargetEntry *te = NULL;
1499                 List       *sl;
1500
1501                 foreach(sl, sub_tlist)
1502                 {
1503                         te = (TargetEntry *) lfirst(sl);
1504                         if (equal(groupexpr, te->expr))
1505                                 break;
1506                 }
1507                 if (!sl)
1508                         elog(ERROR, "failed to locate grouping columns");
1509
1510                 groupColIdx[keyno++] = te->resdom->resno;
1511         }
1512 }
1513
1514 /*
1515  * postprocess_setop_tlist
1516  *        Fix up targetlist returned by plan_set_operations().
1517  *
1518  * We need to transpose sort key info from the orig_tlist into new_tlist.
1519  * NOTE: this would not be good enough if we supported resjunk sort keys
1520  * for results of set operations --- then, we'd need to project a whole
1521  * new tlist to evaluate the resjunk columns.  For now, just ereport if we
1522  * find any resjunk columns in orig_tlist.
1523  */
1524 static List *
1525 postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
1526 {
1527         List       *l;
1528
1529         foreach(l, new_tlist)
1530         {
1531                 TargetEntry *new_tle = (TargetEntry *) lfirst(l);
1532                 TargetEntry *orig_tle;
1533
1534                 /* ignore resjunk columns in setop result */
1535                 if (new_tle->resdom->resjunk)
1536                         continue;
1537
1538                 Assert(orig_tlist != NIL);
1539                 orig_tle = (TargetEntry *) lfirst(orig_tlist);
1540                 orig_tlist = lnext(orig_tlist);
1541                 if (orig_tle->resdom->resjunk)  /* should not happen */
1542                         elog(ERROR, "resjunk output columns are not implemented");
1543                 Assert(new_tle->resdom->resno == orig_tle->resdom->resno);
1544                 Assert(new_tle->resdom->restype == orig_tle->resdom->restype);
1545                 new_tle->resdom->ressortgroupref = orig_tle->resdom->ressortgroupref;
1546         }
1547         if (orig_tlist != NIL)
1548                 elog(ERROR, "resjunk output columns are not implemented");
1549         return new_tlist;
1550 }