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) 1996-2000, PostgreSQL, Inc
   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 1.79 2000/04/12 17:15:22 momjian Exp $
  12  *
  13  *-------------------------------------------------------------------------
  14  */
  15 #include <sys/types.h>
  16
  17 #include "postgres.h"
  18
  19 #include "access/genam.h"
  20 #include "access/heapam.h"
  21 #include "catalog/pg_type.h"
  22 #include "executor/executor.h"
  23 #include "nodes/makefuncs.h"
  24 #include "optimizer/clauses.h"
  25 #include "optimizer/internal.h"
  26 #include "optimizer/paths.h"
  27 #include "optimizer/planmain.h"
  28 #include "optimizer/planner.h"
  29 #include "optimizer/prep.h"
  30 #include "optimizer/subselect.h"
  31 #include "optimizer/tlist.h"
  32 #include "optimizer/var.h"
  33 #include "parser/parse_expr.h"
  34 #include "parser/parse_oper.h"
  35 #include "utils/builtins.h"
  36 #include "utils/lsyscache.h"
  37 #include "utils/syscache.h"
  38
  39
  40 static List *make_subplanTargetList(Query *parse, List *tlist,
  41                                            AttrNumber **groupColIdx);
  42 static Plan *make_groupplan(List *group_tlist, bool tuplePerGroup,
  43                            List *groupClause, AttrNumber *grpColIdx,
  44                            bool is_presorted, Plan *subplan);
  45 static Plan *make_sortplan(List *tlist, List *sortcls, Plan *plannode);
  46
  47 /*****************************************************************************
  48  *
  49  *         Query optimizer entry point
  50  *
  51  *****************************************************************************/
  52 Plan *
  53 planner(Query *parse)
  54 {
  55         Plan       *result_plan;
  56
  57         /* Initialize state for subselects */
  58         PlannerQueryLevel = 1;
  59         PlannerInitPlan = NULL;
  60         PlannerParamVar = NULL;
  61         PlannerPlanId = 0;
  62
  63         /* this should go away sometime soon */
  64         transformKeySetQuery(parse);
  65
  66         /* primary planning entry point (may recurse for subplans) */
  67         result_plan = subquery_planner(parse, -1.0 /* default case */ );
  68
  69         Assert(PlannerQueryLevel == 1);
  70
  71         /* if top-level query had subqueries, do housekeeping for them */
  72         if (PlannerPlanId > 0)
  73         {
  74                 (void) SS_finalize_plan(result_plan);
  75                 result_plan->initPlan = PlannerInitPlan;
  76         }
  77
  78         /* executor wants to know total number of Params used overall */
  79         result_plan->nParamExec = length(PlannerParamVar);
  80
  81         /* final cleanup of the plan */
  82         set_plan_references(result_plan);
  83
  84         return result_plan;
  85 }
  86
  87
  88 /*--------------------
  89  * subquery_planner
  90  *        Invokes the planner on a subquery.  We recurse to here for each
  91  *        sub-SELECT found in the query tree.
  92  *
  93  * parse is the querytree produced by the parser & rewriter.
  94  * tuple_fraction is the fraction of tuples we expect will be retrieved.
  95  * tuple_fraction is interpreted as explained for union_planner, below.
  96  *
  97  * Basically, this routine does the stuff that should only be done once
  98  * per Query object.  It then calls union_planner, which may be called
  99  * recursively on the same Query node in order to handle UNIONs and/or
 100  * inheritance.  subquery_planner is called recursively from subselect.c.
 101  *
 102  * prepunion.c uses an unholy combination of calling union_planner when
 103  * recursing on the primary Query node, or subquery_planner when recursing
 104  * on a UNION'd Query node that hasn't previously been seen by
 105  * subquery_planner.  That whole chunk of code needs rewritten from scratch.
 106  *
 107  * Returns a query plan.
 108  *--------------------
 109  */
 110 Plan *
 111 subquery_planner(Query *parse, double tuple_fraction)
 112 {
 113
 114         /*
 115          * A HAVING clause without aggregates is equivalent to a WHERE clause
 116          * (except it can only refer to grouped fields).  If there are no aggs
 117          * anywhere in the query, then we don't want to create an Agg plan
 118          * node, so merge the HAVING condition into WHERE.      (We used to
 119          * consider this an error condition, but it seems to be legal SQL.)
 120          */
 121         if (parse->havingQual != NULL && !parse->hasAggs)
 122         {
 123                 if (parse->qual == NULL)
 124                         parse->qual = parse->havingQual;
 125                 else
 126                         parse->qual = (Node *) make_andclause(lappend(lcons(parse->qual,
 127                                                                                                                                 NIL),
 128                                                                                                          parse->havingQual));
 129                 parse->havingQual = NULL;
 130         }
 131
 132         /*
 133          * Simplify constant expressions in targetlist and quals.
 134          *
 135          * Note that at this point the qual has not yet been converted to
 136          * implicit-AND form, so we can apply eval_const_expressions directly.
 137          * Also note that we need to do this before SS_process_sublinks,
 138          * because that routine inserts bogus "Const" nodes.
 139          */
 140         parse->targetList = (List *)
 141                 eval_const_expressions((Node *) parse->targetList);
 142         parse->qual = eval_const_expressions(parse->qual);
 143         parse->havingQual = eval_const_expressions(parse->havingQual);
 144
 145         /*
 146          * Canonicalize the qual, and convert it to implicit-AND format.
 147          *
 148          * XXX Is there any value in re-applying eval_const_expressions after
 149          * canonicalize_qual?
 150          */
 151         parse->qual = (Node *) canonicalize_qual((Expr *) parse->qual, true);
 152 #ifdef OPTIMIZER_DEBUG
 153         printf("After canonicalize_qual()\n");
 154         pprint(parse->qual);
 155 #endif
 156
 157         /*
 158          * Ditto for the havingQual
 159          */
 160         parse->havingQual = (Node *) canonicalize_qual((Expr *) parse->havingQual,
 161                                                                                                    true);
 162
 163         /* Expand SubLinks to SubPlans */
 164         if (parse->hasSubLinks)
 165         {
 166                 parse->targetList = (List *)
 167                         SS_process_sublinks((Node *) parse->targetList);
 168                 parse->qual = SS_process_sublinks(parse->qual);
 169                 parse->havingQual = SS_process_sublinks(parse->havingQual);
 170
 171                 if (parse->groupClause != NIL)
 172                 {
 173
 174                         /*
 175                          * Check for ungrouped variables passed to subplans. Note we
 176                          * do NOT do this for subplans in WHERE; it's legal there
 177                          * because WHERE is evaluated pre-GROUP.
 178                          *
 179                          * An interesting fine point: if we reassigned a HAVING qual into
 180                          * WHERE above, then we will accept references to ungrouped
 181                          * vars from subplans in the HAVING qual.  This is not
 182                          * entirely consistent, but it doesn't seem particularly
 183                          * harmful...
 184                          */
 185                         check_subplans_for_ungrouped_vars((Node *) parse->targetList,
 186                                                                                           parse);
 187                         check_subplans_for_ungrouped_vars(parse->havingQual, parse);
 188                 }
 189         }
 190
 191         /* Replace uplevel vars with Param nodes */
 192         if (PlannerQueryLevel > 1)
 193         {
 194                 parse->targetList = (List *)
 195                         SS_replace_correlation_vars((Node *) parse->targetList);
 196                 parse->qual = SS_replace_correlation_vars(parse->qual);
 197                 parse->havingQual = SS_replace_correlation_vars(parse->havingQual);
 198         }
 199
 200         /* Do the main planning (potentially recursive) */
 201
 202         return union_planner(parse, tuple_fraction);
 203
 204         /*
 205          * XXX should any more of union_planner's activity be moved here?
 206          *
 207          * That would take careful study of the interactions with prepunion.c,
 208          * but I suspect it would pay off in simplicity and avoidance of
 209          * wasted cycles.
 210          */
 211 }
 212
 213
 214 /*--------------------
 215  * union_planner
 216  *        Invokes the planner on union-type queries (both regular UNIONs and
 217  *        appends produced by inheritance), recursing if necessary to get them
 218  *        all, then processes normal plans.
 219  *
 220  * parse is the querytree produced by the parser & rewriter.
 221  * tuple_fraction is the fraction of tuples we expect will be retrieved
 222  *
 223  * tuple_fraction is interpreted as follows:
 224  *        < 0: determine fraction by inspection of query (normal case)
 225  *        0: expect all tuples to be retrieved
 226  *        0 < tuple_fraction < 1: expect the given fraction of tuples available
 227  *              from the plan to be retrieved
 228  *        tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
 229  *              expected to be retrieved (ie, a LIMIT specification)
 230  * The normal case is to pass -1, but some callers pass values >= 0 to
 231  * override this routine's determination of the appropriate fraction.
 232  *
 233  * Returns a query plan.
 234  *--------------------
 235  */
 236 Plan *
 237 union_planner(Query *parse,
 238                           double tuple_fraction)
 239 {
 240         List       *tlist = parse->targetList;
 241         List       *rangetable = parse->rtable;
 242         Plan       *result_plan = (Plan *) NULL;
 243         AttrNumber *groupColIdx = NULL;
 244         List       *current_pathkeys = NIL;
 245         List       *group_pathkeys;
 246         List       *sort_pathkeys;
 247         Index           rt_index;
 248
 249         if (parse->unionClause)
 250         {
 251                 result_plan = (Plan *) plan_union_queries(parse);
 252                 /* XXX do we need to do this? bjm 12/19/97 */
 253                 tlist = preprocess_targetlist(tlist,
 254                                                                           parse->commandType,
 255                                                                           parse->resultRelation,
 256                                                                           parse->rtable);
 257
 258                 /*
 259                  * We leave current_pathkeys NIL indicating we do not know sort
 260                  * order. Actually, for a normal UNION we have done an explicit
 261                  * sort; ought to change interface to plan_union_queries to pass
 262                  * that info back!
 263                  */
 264
 265                 /*
 266                  * Calculate pathkeys that represent grouping/ordering
 267                  * requirements
 268                  */
 269                 group_pathkeys = make_pathkeys_for_sortclauses(parse->groupClause,
 270                                                                                                            tlist);
 271                 sort_pathkeys = make_pathkeys_for_sortclauses(parse->sortClause,
 272                                                                                                           tlist);
 273         }
 274         else if ((rt_index = first_inherit_rt_entry(rangetable)) != -1)
 275         {
 276                 List       *sub_tlist;
 277
 278                 /*
 279                  * Generate appropriate target list for subplan; may be different
 280                  * from tlist if grouping or aggregation is needed.
 281                  */
 282                 sub_tlist = make_subplanTargetList(parse, tlist, &groupColIdx);
 283
 284                 /*
 285                  * Recursively plan the subqueries needed for inheritance
 286                  */
 287                 result_plan = (Plan *) plan_inherit_queries(parse, sub_tlist,
 288                                                                                                         rt_index);
 289
 290                 /*
 291                  * Fix up outer target list.  NOTE: unlike the case for
 292                  * non-inherited query, we pass the unfixed tlist to subplans,
 293                  * which do their own fixing.  But we still want to fix the outer
 294                  * target list afterwards. I *think* this is correct --- doing the
 295                  * fix before recursing is definitely wrong, because
 296                  * preprocess_targetlist() will do the wrong thing if invoked
 297                  * twice on the same list. Maybe that is a bug? tgl 6/6/99
 298                  */
 299                 tlist = preprocess_targetlist(tlist,
 300                                                                           parse->commandType,
 301                                                                           parse->resultRelation,
 302                                                                           parse->rtable);
 303
 304                 if (parse->rowMark != NULL)
 305                         elog(ERROR, "SELECT FOR UPDATE is not supported for inherit queries");
 306
 307                 /*
 308                  * We leave current_pathkeys NIL indicating we do not know sort
 309                  * order of the Append-ed results.
 310                  */
 311
 312                 /*
 313                  * Calculate pathkeys that represent grouping/ordering
 314                  * requirements
 315                  */
 316                 group_pathkeys = make_pathkeys_for_sortclauses(parse->groupClause,
 317                                                                                                            tlist);
 318                 sort_pathkeys = make_pathkeys_for_sortclauses(parse->sortClause,
 319                                                                                                           tlist);
 320         }
 321         else
 322         {
 323                 List       *sub_tlist;
 324
 325                 /* Preprocess targetlist in case we are inside an INSERT/UPDATE. */
 326                 tlist = preprocess_targetlist(tlist,
 327                                                                           parse->commandType,
 328                                                                           parse->resultRelation,
 329                                                                           parse->rtable);
 330
 331                 /*
 332                  * Add row-mark targets for UPDATE (should this be done in
 333                  * preprocess_targetlist?)
 334                  */
 335                 if (parse->rowMark != NULL)
 336                 {
 337                         List       *l;
 338
 339                         foreach(l, parse->rowMark)
 340                         {
 341                                 RowMark    *rowmark = (RowMark *) lfirst(l);
 342                                 TargetEntry *ctid;
 343                                 Resdom     *resdom;
 344                                 Var                *var;
 345                                 char       *resname;
 346
 347                                 if (!(rowmark->info & ROW_MARK_FOR_UPDATE))
 348                                         continue;
 349
 350                                 resname = (char *) palloc(32);
 351                                 sprintf(resname, "ctid%u", rowmark->rti);
 352                                 resdom = makeResdom(length(tlist) + 1,
 353                                                                         TIDOID,
 354                                                                         -1,
 355                                                                         resname,
 356                                                                         0,
 357                                                                         0,
 358                                                                         true);
 359
 360                                 var = makeVar(rowmark->rti, -1, TIDOID, -1, 0);
 361
 362                                 ctid = makeTargetEntry(resdom, (Node *) var);
 363                                 tlist = lappend(tlist, ctid);
 364                         }
 365                 }
 366
 367                 /*
 368                  * Generate appropriate target list for subplan; may be different
 369                  * from tlist if grouping or aggregation is needed.
 370                  */
 371                 sub_tlist = make_subplanTargetList(parse, tlist, &groupColIdx);
 372
 373                 /*
 374                  * Calculate pathkeys that represent grouping/ordering
 375                  * requirements
 376                  */
 377                 group_pathkeys = make_pathkeys_for_sortclauses(parse->groupClause,
 378                                                                                                            tlist);
 379                 sort_pathkeys = make_pathkeys_for_sortclauses(parse->sortClause,
 380                                                                                                           tlist);
 381
 382                 /*
 383                  * Figure out whether we need a sorted result from query_planner.
 384                  *
 385                  * If we have a GROUP BY clause, then we want a result sorted
 386                  * properly for grouping.  Otherwise, if there is an ORDER BY
 387                  * clause, we want to sort by the ORDER BY clause.      (Note: if we
 388                  * have both, and ORDER BY is a superset of GROUP BY, it would be
 389                  * tempting to request sort by ORDER BY --- but that might just
 390                  * leave us failing to exploit an available sort order at all.
 391                  * Needs more thought...)
 392                  */
 393                 if (parse->groupClause)
 394                         parse->query_pathkeys = group_pathkeys;
 395                 else if (parse->sortClause)
 396                         parse->query_pathkeys = sort_pathkeys;
 397                 else
 398                         parse->query_pathkeys = NIL;
 399
 400                 /*
 401                  * Figure out whether we expect to retrieve all the tuples that
 402                  * the plan can generate, or to stop early due to a LIMIT or other
 403                  * factors.  If the caller passed a value >= 0, believe that
 404                  * value, else do our own examination of the query context.
 405                  */
 406                 if (tuple_fraction < 0.0)
 407                 {
 408                         /* Initial assumption is we need all the tuples */
 409                         tuple_fraction = 0.0;
 410
 411                         /*
 412                          * Check for a LIMIT clause.
 413                          */
 414                         if (parse->limitCount != NULL)
 415                         {
 416                                 if (IsA(parse->limitCount, Const))
 417                                 {
 418                                         Const      *limitc = (Const *) parse->limitCount;
 419                                         int                     count = (int) (limitc->constvalue);
 420
 421                                         /*
 422                                          * The constant can legally be either 0 ("ALL") or a
 423                                          * positive integer.  If it is not ALL, we also need
 424                                          * to consider the OFFSET part of LIMIT.
 425                                          */
 426                                         if (count > 0)
 427                                         {
 428                                                 tuple_fraction = (double) count;
 429                                                 if (parse->limitOffset != NULL)
 430                                                 {
 431                                                         if (IsA(parse->limitOffset, Const))
 432                                                         {
 433                                                                 int                     offset;
 434
 435                                                                 limitc = (Const *) parse->limitOffset;
 436                                                                 offset = (int) (limitc->constvalue);
 437                                                                 if (offset > 0)
 438                                                                         tuple_fraction += (double) offset;
 439                                                         }
 440                                                         else
 441                                                         {
 442                                                                 /* It's a PARAM ... punt ... */
 443                                                                 tuple_fraction = 0.10;
 444                                                         }
 445                                                 }
 446                                         }
 447                                 }
 448                                 else
 449                                 {
 450
 451                                         /*
 452                                          * COUNT is a PARAM ... don't know exactly what the
 453                                          * limit will be, but for lack of a better idea assume
 454                                          * 10% of the plan's result is wanted.
 455                                          */
 456                                         tuple_fraction = 0.10;
 457                                 }
 458                         }
 459
 460                         /*
 461                          * Check for a retrieve-into-portal, ie DECLARE CURSOR.
 462                          *
 463                          * We have no real idea how many tuples the user will ultimately
 464                          * FETCH from a cursor, but it seems a good bet that he
 465                          * doesn't want 'em all.  Optimize for 10% retrieval (you
 466                          * gotta better number?)
 467                          */
 468                         if (parse->isPortal)
 469                                 tuple_fraction = 0.10;
 470                 }
 471
 472                 /*
 473                  * Adjust tuple_fraction if we see that we are going to apply
 474                  * grouping/aggregation/etc.  This is not overridable by the
 475                  * caller, since it reflects plan actions that this routine will
 476                  * certainly take, not assumptions about context.
 477                  */
 478                 if (parse->groupClause)
 479                 {
 480
 481                         /*
 482                          * In GROUP BY mode, we have the little problem that we don't
 483                          * really know how many input tuples will be needed to make a
 484                          * group, so we can't translate an output LIMIT count into an
 485                          * input count.  For lack of a better idea, assume 25% of the
 486                          * input data will be processed if there is any output limit.
 487                          * However, if the caller gave us a fraction rather than an
 488                          * absolute count, we can keep using that fraction (which
 489                          * amounts to assuming that all the groups are about the same
 490                          * size).
 491                          */
 492                         if (tuple_fraction >= 1.0)
 493                                 tuple_fraction = 0.25;
 494
 495                         /*
 496                          * If both GROUP BY and ORDER BY are specified, we will need
 497                          * two levels of sort --- and, therefore, certainly need to
 498                          * read all the input tuples --- unless ORDER BY is a subset
 499                          * of GROUP BY.  (Although we are comparing non-canonicalized
 500                          * pathkeys here, it should be OK since they will both contain
 501                          * only single-element sublists at this point.  See
 502                          * pathkeys.c.)
 503                          */
 504                         if (parse->groupClause && parse->sortClause &&
 505                                 !pathkeys_contained_in(sort_pathkeys, group_pathkeys))
 506                                 tuple_fraction = 0.0;
 507                 }
 508                 else if (parse->hasAggs)
 509                 {
 510
 511                         /*
 512                          * Ungrouped aggregate will certainly want all the input
 513                          * tuples.
 514                          */
 515                         tuple_fraction = 0.0;
 516                 }
 517                 else if (parse->distinctClause)
 518                 {
 519
 520                         /*
 521                          * SELECT DISTINCT, like GROUP, will absorb an unpredictable
 522                          * number of input tuples per output tuple.  Handle the same
 523                          * way.
 524                          */
 525                         if (tuple_fraction >= 1.0)
 526                                 tuple_fraction = 0.25;
 527                 }
 528
 529                 /* Generate the (sub) plan */
 530                 result_plan = query_planner(parse,
 531                                                                         sub_tlist,
 532                                                                         (List *) parse->qual,
 533                                                                         tuple_fraction);
 534
 535                 /*
 536                  * query_planner returns actual sort order (which is not
 537                  * necessarily what we requested) in query_pathkeys.
 538                  */
 539                 current_pathkeys = parse->query_pathkeys;
 540         }
 541
 542         /* query_planner returns NULL if it thinks plan is bogus */
 543         if (!result_plan)
 544                 elog(ERROR, "union_planner: failed to create plan");
 545
 546         /*
 547          * We couldn't canonicalize group_pathkeys and sort_pathkeys before
 548          * running query_planner(), so do it now.
 549          */
 550         group_pathkeys = canonicalize_pathkeys(parse, group_pathkeys);
 551         sort_pathkeys = canonicalize_pathkeys(parse, sort_pathkeys);
 552
 553         /*
 554          * If we have a GROUP BY clause, insert a group node (plus the
 555          * appropriate sort node, if necessary).
 556          */
 557         if (parse->groupClause)
 558         {
 559                 bool            tuplePerGroup;
 560                 List       *group_tlist;
 561                 bool            is_sorted;
 562
 563                 /*
 564                  * Decide whether how many tuples per group the Group node needs
 565                  * to return. (Needs only one tuple per group if no aggregate is
 566                  * present. Otherwise, need every tuple from the group to do the
 567                  * aggregation.)  Note tuplePerGroup is named backwards :-(
 568                  */
 569                 tuplePerGroup = parse->hasAggs;
 570
 571                 /*
 572                  * If there are aggregates then the Group node should just return
 573                  * the same set of vars as the subplan did (but we can exclude any
 574                  * GROUP BY expressions).  If there are no aggregates then the
 575                  * Group node had better compute the final tlist.
 576                  */
 577                 if (parse->hasAggs)
 578                         group_tlist = flatten_tlist(result_plan->targetlist);
 579                 else
 580                         group_tlist = tlist;
 581
 582                 /*
 583                  * Figure out whether the path result is already ordered the way
 584                  * we need it --- if so, no need for an explicit sort step.
 585                  */
 586                 if (pathkeys_contained_in(group_pathkeys, current_pathkeys))
 587                 {
 588                         is_sorted = true;       /* no sort needed now */
 589                         /* current_pathkeys remains unchanged */
 590                 }
 591                 else
 592                 {
 593
 594                         /*
 595                          * We will need to do an explicit sort by the GROUP BY clause.
 596                          * make_groupplan will do the work, but set current_pathkeys
 597                          * to indicate the resulting order.
 598                          */
 599                         is_sorted = false;
 600                         current_pathkeys = group_pathkeys;
 601                 }
 602
 603                 result_plan = make_groupplan(group_tlist,
 604                                                                          tuplePerGroup,
 605                                                                          parse->groupClause,
 606                                                                          groupColIdx,
 607                                                                          is_sorted,
 608                                                                          result_plan);
 609         }
 610
 611         /*
 612          * If aggregate is present, insert the Agg node
 613          *
 614          * HAVING clause, if any, becomes qual of the Agg node
 615          */
 616         if (parse->hasAggs)
 617         {
 618                 result_plan = (Plan *) make_agg(tlist,
 619                                                                                 (List *) parse->havingQual,
 620                                                                                 result_plan);
 621                 /* Note: Agg does not affect any existing sort order of the tuples */
 622         }
 623
 624         /*
 625          * If we were not able to make the plan come out in the right order,
 626          * add an explicit sort step.
 627          */
 628         if (parse->sortClause)
 629         {
 630                 if (!pathkeys_contained_in(sort_pathkeys, current_pathkeys))
 631                         result_plan = make_sortplan(tlist, parse->sortClause, result_plan);
 632         }
 633
 634         /*
 635          * Finally, if there is a DISTINCT clause, add the UNIQUE node.
 636          */
 637         if (parse->distinctClause)
 638         {
 639                 result_plan = (Plan *) make_unique(tlist, result_plan,
 640                                                                                    parse->distinctClause);
 641         }
 642
 643         return result_plan;
 644 }
 645
 646 /*---------------
 647  * make_subplanTargetList
 648  *        Generate appropriate target list when grouping is required.
 649  *
 650  * When union_planner inserts Aggregate and/or Group plan nodes above
 651  * the result of query_planner, we typically want to pass a different
 652  * target list to query_planner than the outer plan nodes should have.
 653  * This routine generates the correct target list for the subplan.
 654  *
 655  * The initial target list passed from the parser already contains entries
 656  * for all ORDER BY and GROUP BY expressions, but it will not have entries
 657  * for variables used only in HAVING clauses; so we need to add those
 658  * variables to the subplan target list.  Also, if we are doing either
 659  * grouping or aggregation, we flatten all expressions except GROUP BY items
 660  * into their component variables; the other expressions will be computed by
 661  * the inserted nodes rather than by the subplan.  For example,
 662  * given a query like
 663  *              SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
 664  * we want to pass this targetlist to the subplan:
 665  *              a,b,c,d,a+b
 666  * where the a+b target will be used by the Sort/Group steps, and the
 667  * other targets will be used for computing the final results.  (In the
 668  * above example we could theoretically suppress the a and b targets and
 669  * use only a+b, but it's not really worth the trouble.)
 670  *
 671  * 'parse' is the query being processed.
 672  * 'tlist' is the query's target list.
 673  * 'groupColIdx' receives an array of column numbers for the GROUP BY
 674  * expressions (if there are any) in the subplan's target list.
 675  *
 676  * The result is the targetlist to be passed to the subplan.
 677  *---------------
 678  */
 679 static List *
 680 make_subplanTargetList(Query *parse,
 681                                            List *tlist,
 682                                            AttrNumber **groupColIdx)
 683 {
 684         List       *sub_tlist;
 685         List       *extravars;
 686         int                     numCols;
 687
 688         *groupColIdx = NULL;
 689
 690         /*
 691          * If we're not grouping or aggregating, nothing to do here;
 692          * query_planner should receive the unmodified target list.
 693          */
 694         if (!parse->hasAggs && !parse->groupClause && !parse->havingQual)
 695                 return tlist;
 696
 697         /*
 698          * Otherwise, start with a "flattened" tlist (having just the vars
 699          * mentioned in the targetlist and HAVING qual --- but not upper-
 700          * level Vars; they will be replaced by Params later on).
 701          */
 702         sub_tlist = flatten_tlist(tlist);
 703         extravars = pull_var_clause(parse->havingQual, false);
 704         sub_tlist = add_to_flat_tlist(sub_tlist, extravars);
 705         freeList(extravars);
 706
 707         /*
 708          * If grouping, create sub_tlist entries for all GROUP BY expressions
 709          * (GROUP BY items that are simple Vars should be in the list
 710          * already), and make an array showing where the group columns are in
 711          * the sub_tlist.
 712          */
 713         numCols = length(parse->groupClause);
 714         if (numCols > 0)
 715         {
 716                 int                     keyno = 0;
 717                 AttrNumber *grpColIdx;
 718                 List       *gl;
 719
 720                 grpColIdx = (AttrNumber *) palloc(sizeof(AttrNumber) * numCols);
 721                 *groupColIdx = grpColIdx;
 722
 723                 foreach(gl, parse->groupClause)
 724                 {
 725                         GroupClause *grpcl = (GroupClause *) lfirst(gl);
 726                         Node       *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
 727                         TargetEntry *te = NULL;
 728                         List       *sl;
 729
 730                         /* Find or make a matching sub_tlist entry */
 731                         foreach(sl, sub_tlist)
 732                         {
 733                                 te = (TargetEntry *) lfirst(sl);
 734                                 if (equal(groupexpr, te->expr))
 735                                         break;
 736                         }
 737                         if (!sl)
 738                         {
 739                                 te = makeTargetEntry(makeResdom(length(sub_tlist) + 1,
 740                                                                                                 exprType(groupexpr),
 741                                                                                                 exprTypmod(groupexpr),
 742                                                                                                 NULL,
 743                                                                                                 (Index) 0,
 744                                                                                                 (Oid) 0,
 745                                                                                                 false),
 746                                                                          groupexpr);
 747                                 sub_tlist = lappend(sub_tlist, te);
 748                         }
 749
 750                         /* and save its resno */
 751                         grpColIdx[keyno++] = te->resdom->resno;
 752                 }
 753         }
 754
 755         return sub_tlist;
 756 }
 757
 758 /*
 759  * make_groupplan
 760  *              Add a Group node for GROUP BY processing.
 761  *              If we couldn't make the subplan produce presorted output for grouping,
 762  *              first add an explicit Sort node.
 763  */
 764 static Plan *
 765 make_groupplan(List *group_tlist,
 766                            bool tuplePerGroup,
 767                            List *groupClause,
 768                            AttrNumber *grpColIdx,
 769                            bool is_presorted,
 770                            Plan *subplan)
 771 {
 772         int                     numCols = length(groupClause);
 773
 774         if (!is_presorted)
 775         {
 776
 777                 /*
 778                  * The Sort node always just takes a copy of the subplan's tlist
 779                  * plus ordering information.  (This might seem inefficient if the
 780                  * subplan contains complex GROUP BY expressions, but in fact Sort
 781                  * does not evaluate its targetlist --- it only outputs the same
 782                  * tuples in a new order.  So the expressions we might be copying
 783                  * are just dummies with no extra execution cost.)
 784                  */
 785                 List       *sort_tlist = new_unsorted_tlist(subplan->targetlist);
 786                 int                     keyno = 0;
 787                 List       *gl;
 788
 789                 foreach(gl, groupClause)
 790                 {
 791                         GroupClause *grpcl = (GroupClause *) lfirst(gl);
 792                         TargetEntry *te = nth(grpColIdx[keyno] - 1, sort_tlist);
 793                         Resdom     *resdom = te->resdom;
 794
 795                         /*
 796                          * Check for the possibility of duplicate group-by clauses ---
 797                          * the parser should have removed 'em, but the Sort executor
 798                          * will get terribly confused if any get through!
 799                          */
 800                         if (resdom->reskey == 0)
 801                         {
 802                                 /* OK, insert the ordering info needed by the executor. */
 803                                 resdom->reskey = ++keyno;
 804                                 resdom->reskeyop = get_opcode(grpcl->sortop);
 805                         }
 806                 }
 807
 808                 subplan = (Plan *) make_sort(sort_tlist,
 809                                                                          _NONAME_RELATION_ID_,
 810                                                                          subplan,
 811                                                                          keyno);
 812         }
 813
 814         return (Plan *) make_group(group_tlist, tuplePerGroup, numCols,
 815                                                            grpColIdx, subplan);
 816 }
 817
 818 /*
 819  * make_sortplan
 820  *        Add a Sort node to implement an explicit ORDER BY clause.
 821  */
 822 static Plan *
 823 make_sortplan(List *tlist, List *sortcls, Plan *plannode)
 824 {
 825         List       *temp_tlist;
 826         List       *i;
 827         int                     keyno = 0;
 828
 829         /*
 830          * First make a copy of the tlist so that we don't corrupt the
 831          * original.
 832          */
 833
 834         temp_tlist = new_unsorted_tlist(tlist);
 835
 836         foreach(i, sortcls)
 837         {
 838                 SortClause *sortcl = (SortClause *) lfirst(i);
 839                 TargetEntry *tle = get_sortgroupclause_tle(sortcl, temp_tlist);
 840                 Resdom     *resdom = tle->resdom;
 841
 842                 /*
 843                  * Check for the possibility of duplicate order-by clauses --- the
 844                  * parser should have removed 'em, but the executor will get
 845                  * terribly confused if any get through!
 846                  */
 847                 if (resdom->reskey == 0)
 848                 {
 849                         /* OK, insert the ordering info needed by the executor. */
 850                         resdom->reskey = ++keyno;
 851                         resdom->reskeyop = get_opcode(sortcl->sortop);
 852                 }
 853         }
 854
 855         return (Plan *) make_sort(temp_tlist,
 856                                                           _NONAME_RELATION_ID_,
 857                                                           plannode,
 858                                                           keyno);
 859 }
 860
 861 /*
 862  * pg_checkretval() -- check return value of a list of sql parse
 863  *                                              trees.
 864  *
 865  * The return value of a sql function is the value returned by
 866  * the final query in the function.  We do some ad-hoc define-time
 867  * type checking here to be sure that the user is returning the
 868  * type he claims.
 869  *
 870  * XXX Why is this function in this module?
 871  */
 872 void
 873 pg_checkretval(Oid rettype, List *queryTreeList)
 874 {
 875         Query      *parse;
 876         List       *tlist;
 877         List       *rt;
 878         int                     cmd;
 879         Type            typ;
 880         Resdom     *resnode;
 881         Relation        reln;
 882         Oid                     relid;
 883         int                     relnatts;
 884         int                     i;
 885
 886         /* find the final query */
 887         parse = (Query *) nth(length(queryTreeList) - 1, queryTreeList);
 888
 889         /*
 890          * test 1:      if the last query is a utility invocation, then there had
 891          * better not be a return value declared.
 892          */
 893         if (parse->commandType == CMD_UTILITY)
 894         {
 895                 if (rettype == InvalidOid)
 896                         return;
 897                 else
 898                         elog(ERROR, "return type mismatch in function decl: final query is a catalog utility");
 899         }
 900
 901         /* okay, it's an ordinary query */
 902         tlist = parse->targetList;
 903         rt = parse->rtable;
 904         cmd = parse->commandType;
 905
 906         /*
 907          * test 2:      if the function is declared to return no value, then the
 908          * final query had better not be a retrieve.
 909          */
 910         if (rettype == InvalidOid)
 911         {
 912                 if (cmd == CMD_SELECT)
 913                         elog(ERROR,
 914                                  "function declared with no return type, but final query is a retrieve");
 915                 else
 916                         return;
 917         }
 918
 919         /* by here, the function is declared to return some type */
 920         if ((typ = typeidType(rettype)) == NULL)
 921                 elog(ERROR, "can't find return type %u for function\n", rettype);
 922
 923         /*
 924          * test 3:      if the function is declared to return a value, then the
 925          * final query had better be a retrieve.
 926          */
 927         if (cmd != CMD_SELECT)
 928                 elog(ERROR, "function declared to return type %s, but final query is not a retrieve", typeTypeName(typ));
 929
 930         /*
 931          * test 4:      for base type returns, the target list should have exactly
 932          * one entry, and its type should agree with what the user declared.
 933          */
 934
 935         if (typeTypeRelid(typ) == InvalidOid)
 936         {
 937                 if (ExecTargetListLength(tlist) > 1)
 938                         elog(ERROR, "function declared to return %s returns multiple values in final retrieve", typeTypeName(typ));
 939
 940                 resnode = (Resdom *) ((TargetEntry *) lfirst(tlist))->resdom;
 941                 if (resnode->restype != rettype)
 942                         elog(ERROR, "return type mismatch in function: declared to return %s, returns %s", typeTypeName(typ), typeidTypeName(resnode->restype));
 943
 944                 /* by here, base return types match */
 945                 return;
 946         }
 947
 948         /*
 949          * If the target list is of length 1, and the type of the varnode in
 950          * the target list is the same as the declared return type, this is
 951          * okay.  This can happen, for example, where the body of the function
 952          * is 'retrieve (x = func2())', where func2 has the same return type
 953          * as the function that's calling it.
 954          */
 955         if (ExecTargetListLength(tlist) == 1)
 956         {
 957                 resnode = (Resdom *) ((TargetEntry *) lfirst(tlist))->resdom;
 958                 if (resnode->restype == rettype)
 959                         return;
 960         }
 961
 962         /*
 963          * By here, the procedure returns a (set of) tuples.  This part of the
 964          * typechecking is a hack.      We look up the relation that is the
 965          * declared return type, and be sure that attributes 1 .. n in the
 966          * target list match the declared types.
 967          */
 968         reln = heap_open(typeTypeRelid(typ), AccessShareLock);
 969         relid = reln->rd_id;
 970         relnatts = reln->rd_rel->relnatts;
 971
 972         if (ExecTargetListLength(tlist) != relnatts)
 973                 elog(ERROR, "function declared to return type %s does not retrieve (%s.*)", typeTypeName(typ), typeTypeName(typ));
 974
 975         /* expect attributes 1 .. n in order */
 976         for (i = 1; i <= relnatts; i++)
 977         {
 978                 TargetEntry *tle = lfirst(tlist);
 979                 Node       *thenode = tle->expr;
 980                 Oid                     tletype = exprType(thenode);
 981
 982                 if (tletype != reln->rd_att->attrs[i - 1]->atttypid)
 983                         elog(ERROR, "function declared to return type %s does not retrieve (%s.all)", typeTypeName(typ), typeTypeName(typ));
 984                 tlist = lnext(tlist);
 985         }
 986
 987         heap_close(reln, AccessShareLock);
 988 }