src/backend/optimizer/plan/planner.c

   1 /*-------------------------------------------------------------------------
   2  *
   3  * planner.c
   4  *        The query optimizer external interface.
   5  *
   6  * Portions Copyright (c) 1996-2001, 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 1.110 2001/10/25 05:49:33 momjian Exp $
  12  *
  13  *-------------------------------------------------------------------------
  14  */
  15
  16 #include "postgres.h"
  17
  18 #include "catalog/pg_type.h"
  19 #include "nodes/makefuncs.h"
  20 #ifdef OPTIMIZER_DEBUG
  21 #include "nodes/print.h"
  22 #endif
  23 #include "optimizer/clauses.h"
  24 #include "optimizer/paths.h"
  25 #include "optimizer/planmain.h"
  26 #include "optimizer/planner.h"
  27 #include "optimizer/prep.h"
  28 #include "optimizer/subselect.h"
  29 #include "optimizer/tlist.h"
  30 #include "optimizer/var.h"
  31 #include "parser/analyze.h"
  32 #include "parser/parsetree.h"
  33 #include "parser/parse_expr.h"
  34 #include "rewrite/rewriteManip.h"
  35 #include "utils/lsyscache.h"
  36
  37
  38 /* Expression kind codes for preprocess_expression */
  39 #define EXPRKIND_TARGET 0
  40 #define EXPRKIND_WHERE  1
  41 #define EXPRKIND_HAVING 2
  42
  43
  44 static Node *pull_up_subqueries(Query *parse, Node *jtnode);
  45 static bool is_simple_subquery(Query *subquery);
  46 static void resolvenew_in_jointree(Node *jtnode, int varno, List *subtlist);
  47 static Node *preprocess_jointree(Query *parse, Node *jtnode);
  48 static Node *preprocess_expression(Query *parse, Node *expr, int kind);
  49 static void preprocess_qual_conditions(Query *parse, Node *jtnode);
  50 static Plan *inheritance_planner(Query *parse, List *inheritlist);
  51 static Plan *grouping_planner(Query *parse, double tuple_fraction);
  52 static List *make_subplanTargetList(Query *parse, List *tlist,
  53                                            AttrNumber **groupColIdx);
  54 static Plan *make_groupplan(Query *parse,
  55                            List *group_tlist, bool tuplePerGroup,
  56                            List *groupClause, AttrNumber *grpColIdx,
  57                            bool is_presorted, Plan *subplan);
  58 static List *postprocess_setop_tlist(List *new_tlist, List *orig_tlist);
  59
  60
  61 /*****************************************************************************
  62  *
  63  *         Query optimizer entry point
  64  *
  65  *****************************************************************************/
  66 Plan *
  67 planner(Query *parse)
  68 {
  69         Plan       *result_plan;
  70         Index           save_PlannerQueryLevel;
  71         List       *save_PlannerParamVar;
  72
  73         /*
  74          * The planner can be called recursively (an example is when
  75          * eval_const_expressions tries to pre-evaluate an SQL function). So,
  76          * these global state variables must be saved and restored.
  77          *
  78          * These vars cannot be moved into the Query structure since their whole
  79          * purpose is communication across multiple sub-Queries.
  80          *
  81          * Note we do NOT save and restore PlannerPlanId: it exists to assign
  82          * unique IDs to SubPlan nodes, and we want those IDs to be unique for
  83          * the life of a backend.  Also, PlannerInitPlan is saved/restored in
  84          * subquery_planner, not here.
  85          */
  86         save_PlannerQueryLevel = PlannerQueryLevel;
  87         save_PlannerParamVar = PlannerParamVar;
  88
  89         /* Initialize state for handling outer-level references and params */
  90         PlannerQueryLevel = 0;          /* will be 1 in top-level subquery_planner */
  91         PlannerParamVar = NIL;
  92
  93         /* primary planning entry point (may recurse for subqueries) */
  94         result_plan = subquery_planner(parse, -1.0 /* default case */ );
  95
  96         Assert(PlannerQueryLevel == 0);
  97
  98         /* executor wants to know total number of Params used overall */
  99         result_plan->nParamExec = length(PlannerParamVar);
 100
 101         /* final cleanup of the plan */
 102         set_plan_references(result_plan);
 103
 104         /* restore state for outer planner, if any */
 105         PlannerQueryLevel = save_PlannerQueryLevel;
 106         PlannerParamVar = save_PlannerParamVar;
 107
 108         return result_plan;
 109 }
 110
 111
 112 /*--------------------
 113  * subquery_planner
 114  *        Invokes the planner on a subquery.  We recurse to here for each
 115  *        sub-SELECT found in the query tree.
 116  *
 117  * parse is the querytree produced by the parser & rewriter.
 118  * tuple_fraction is the fraction of tuples we expect will be retrieved.
 119  * tuple_fraction is interpreted as explained for grouping_planner, below.
 120  *
 121  * Basically, this routine does the stuff that should only be done once
 122  * per Query object.  It then calls grouping_planner.  At one time,
 123  * grouping_planner could be invoked recursively on the same Query object;
 124  * that's not currently true, but we keep the separation between the two
 125  * routines anyway, in case we need it again someday.
 126  *
 127  * subquery_planner will be called recursively to handle sub-Query nodes
 128  * found within the query's expressions and rangetable.
 129  *
 130  * Returns a query plan.
 131  *--------------------
 132  */
 133 Plan *
 134 subquery_planner(Query *parse, double tuple_fraction)
 135 {
 136         List       *saved_initplan = PlannerInitPlan;
 137         int                     saved_planid = PlannerPlanId;
 138         Plan       *plan;
 139         List       *newHaving;
 140         List       *lst;
 141
 142         /* Set up for a new level of subquery */
 143         PlannerQueryLevel++;
 144         PlannerInitPlan = NIL;
 145
 146 #ifdef ENABLE_KEY_SET_QUERY
 147         /* this should go away sometime soon */
 148         transformKeySetQuery(parse);
 149 #endif
 150
 151         /*
 152          * Check to see if any subqueries in the rangetable can be merged into
 153          * this query.
 154          */
 155         parse->jointree = (FromExpr *)
 156                 pull_up_subqueries(parse, (Node *) parse->jointree);
 157
 158         /*
 159          * If so, we may have created opportunities to simplify the jointree.
 160          */
 161         parse->jointree = (FromExpr *)
 162                 preprocess_jointree(parse, (Node *) parse->jointree);
 163
 164         /*
 165          * Do expression preprocessing on targetlist and quals.
 166          */
 167         parse->targetList = (List *)
 168                 preprocess_expression(parse, (Node *) parse->targetList,
 169                                                           EXPRKIND_TARGET);
 170
 171         preprocess_qual_conditions(parse, (Node *) parse->jointree);
 172
 173         parse->havingQual = preprocess_expression(parse, parse->havingQual,
 174                                                                                           EXPRKIND_HAVING);
 175
 176         /*
 177          * A HAVING clause without aggregates is equivalent to a WHERE clause
 178          * (except it can only refer to grouped fields).  Transfer any
 179          * agg-free clauses of the HAVING qual into WHERE.      This may seem like
 180          * wasting cycles to cater to stupidly-written queries, but there are
 181          * other reasons for doing it.  Firstly, if the query contains no aggs
 182          * at all, then we aren't going to generate an Agg plan node, and so
 183          * there'll be no place to execute HAVING conditions; without this
 184          * transfer, we'd lose the HAVING condition entirely, which is wrong.
 185          * Secondly, when we push down a qual condition into a sub-query, it's
 186          * easiest to push the qual into HAVING always, in case it contains
 187          * aggs, and then let this code sort it out.
 188          *
 189          * Note that both havingQual and parse->jointree->quals are in
 190          * implicitly-ANDed-list form at this point, even though they are
 191          * declared as Node *.  Also note that contain_agg_clause does not
 192          * recurse into sub-selects, which is exactly what we need here.
 193          */
 194         newHaving = NIL;
 195         foreach(lst, (List *) parse->havingQual)
 196         {
 197                 Node       *havingclause = (Node *) lfirst(lst);
 198
 199                 if (contain_agg_clause(havingclause))
 200                         newHaving = lappend(newHaving, havingclause);
 201                 else
 202                         parse->jointree->quals = (Node *)
 203                                 lappend((List *) parse->jointree->quals, havingclause);
 204         }
 205         parse->havingQual = (Node *) newHaving;
 206
 207         /*
 208          * Do the main planning.  If we have an inherited target relation,
 209          * that needs special processing, else go straight to
 210          * grouping_planner.
 211          */
 212         if (parse->resultRelation &&
 213          (lst = expand_inherted_rtentry(parse, parse->resultRelation, false))
 214                 != NIL)
 215                 plan = inheritance_planner(parse, lst);
 216         else
 217                 plan = grouping_planner(parse, tuple_fraction);
 218
 219         /*
 220          * If any subplans were generated, or if we're inside a subplan, build
 221          * subPlan, extParam and locParam lists for plan nodes.
 222          */
 223         if (PlannerPlanId != saved_planid || PlannerQueryLevel > 1)
 224         {
 225                 (void) SS_finalize_plan(plan);
 226
 227                 /*
 228                  * At the moment, SS_finalize_plan doesn't handle initPlans and so
 229                  * we assign them to the topmost plan node.
 230                  */
 231                 plan->initPlan = PlannerInitPlan;
 232                 /* Must add the initPlans' extParams to the topmost node's, too */
 233                 foreach(lst, plan->initPlan)
 234                 {
 235                         SubPlan    *subplan = (SubPlan *) lfirst(lst);
 236
 237                         plan->extParam = set_unioni(plan->extParam,
 238                                                                                 subplan->plan->extParam);
 239                 }
 240         }
 241
 242         /* Return to outer subquery context */
 243         PlannerQueryLevel--;
 244         PlannerInitPlan = saved_initplan;
 245         /* we do NOT restore PlannerPlanId; that's not an oversight! */
 246
 247         return plan;
 248 }
 249
 250 /*
 251  * pull_up_subqueries
 252  *              Look for subqueries in the rangetable that can be pulled up into
 253  *              the parent query.  If the subquery has no special features like
 254  *              grouping/aggregation then we can merge it into the parent's jointree.
 255  *
 256  * A tricky aspect of this code is that if we pull up a subquery we have
 257  * to replace Vars that reference the subquery's outputs throughout the
 258  * parent query, including quals attached to jointree nodes above the one
 259  * we are currently processing!  We handle this by being careful not to
 260  * change the jointree structure while recursing: no nodes other than
 261  * subquery RangeTblRef entries will be replaced.  Also, we can't turn
 262  * ResolveNew loose on the whole jointree, because it'll return a mutated
 263  * copy of the tree; we have to invoke it just on the quals, instead.
 264  */
 265 static Node *
 266 pull_up_subqueries(Query *parse, Node *jtnode)
 267 {
 268         if (jtnode == NULL)
 269                 return NULL;
 270         if (IsA(jtnode, RangeTblRef))
 271         {
 272                 int                     varno = ((RangeTblRef *) jtnode)->rtindex;
 273                 RangeTblEntry *rte = rt_fetch(varno, parse->rtable);
 274                 Query      *subquery = rte->subquery;
 275
 276                 /*
 277                  * Is this a subquery RTE, and if so, is the subquery simple
 278                  * enough to pull up?  (If not, do nothing at this node.)
 279                  *
 280                  * Note: even if the subquery itself is simple enough, we can't pull
 281                  * it up if there is a reference to its whole tuple result.
 282                  */
 283                 if (subquery && is_simple_subquery(subquery) &&
 284                         !contain_whole_tuple_var((Node *) parse, varno, 0))
 285                 {
 286                         int                     rtoffset;
 287                         Node       *subjointree;
 288                         List       *subtlist;
 289                         List       *l;
 290
 291                         /*
 292                          * First, recursively pull up the subquery's subqueries, so
 293                          * that this routine's processing is complete for its jointree
 294                          * and rangetable.      NB: if the same subquery is referenced
 295                          * from multiple jointree items (which can't happen normally,
 296                          * but might after rule rewriting), then we will invoke this
 297                          * processing multiple times on that subquery.  OK because
 298                          * nothing will happen after the first time.  We do have to be
 299                          * careful to copy everything we pull up, however, or risk
 300                          * having chunks of structure multiply linked.
 301                          */
 302                         subquery->jointree = (FromExpr *)
 303                                 pull_up_subqueries(subquery, (Node *) subquery->jointree);
 304
 305                         /*
 306                          * Append the subquery's rangetable to mine (currently, no
 307                          * adjustments will be needed in the subquery's rtable).
 308                          */
 309                         rtoffset = length(parse->rtable);
 310                         parse->rtable = nconc(parse->rtable,
 311                                                                   copyObject(subquery->rtable));
 312
 313                         /*
 314                          * Make copies of the subquery's jointree and targetlist with
 315                          * varnos adjusted to match the merged rangetable.
 316                          */
 317                         subjointree = copyObject(subquery->jointree);
 318                         OffsetVarNodes(subjointree, rtoffset, 0);
 319                         subtlist = copyObject(subquery->targetList);
 320                         OffsetVarNodes((Node *) subtlist, rtoffset, 0);
 321
 322                         /*
 323                          * Replace all of the top query's references to the subquery's
 324                          * outputs with copies of the adjusted subtlist items, being
 325                          * careful not to replace any of the jointree structure.
 326                          */
 327                         parse->targetList = (List *)
 328                                 ResolveNew((Node *) parse->targetList,
 329                                                    varno, 0, subtlist, CMD_SELECT, 0);
 330                         resolvenew_in_jointree((Node *) parse->jointree, varno, subtlist);
 331                         parse->havingQual =
 332                                 ResolveNew(parse->havingQual,
 333                                                    varno, 0, subtlist, CMD_SELECT, 0);
 334
 335                         /*
 336                          * Pull up any FOR UPDATE markers, too.
 337                          */
 338                         foreach(l, subquery->rowMarks)
 339                         {
 340                                 int                     submark = lfirsti(l);
 341
 342                                 parse->rowMarks = lappendi(parse->rowMarks,
 343                                                                                    submark + rtoffset);
 344                         }
 345
 346                         /*
 347                          * Miscellaneous housekeeping.
 348                          */
 349                         parse->hasSubLinks |= subquery->hasSubLinks;
 350                         /* subquery won't be pulled up if it hasAggs, so no work there */
 351
 352                         /*
 353                          * Return the adjusted subquery jointree to replace the
 354                          * RangeTblRef entry in my jointree.
 355                          */
 356                         return subjointree;
 357                 }
 358         }
 359         else if (IsA(jtnode, FromExpr))
 360         {
 361                 FromExpr   *f = (FromExpr *) jtnode;
 362                 List       *l;
 363
 364                 foreach(l, f->fromlist)
 365                         lfirst(l) = pull_up_subqueries(parse, lfirst(l));
 366         }
 367         else if (IsA(jtnode, JoinExpr))
 368         {
 369                 JoinExpr   *j = (JoinExpr *) jtnode;
 370
 371                 /*
 372                  * At the moment, we can't pull up subqueries that are inside the
 373                  * nullable side of an outer join, because substituting their
 374                  * target list entries for upper Var references wouldn't do the
 375                  * right thing (the entries wouldn't go to NULL when they're
 376                  * supposed to). Suppressing the pullup is an ugly,
 377                  * performance-losing hack, but I see no alternative for now.
 378                  * Find a better way to handle this when we redesign query trees
 379                  * --- tgl 4/30/01.
 380                  */
 381                 switch (j->jointype)
 382                 {
 383                         case JOIN_INNER:
 384                                 j->larg = pull_up_subqueries(parse, j->larg);
 385                                 j->rarg = pull_up_subqueries(parse, j->rarg);
 386                                 break;
 387                         case JOIN_LEFT:
 388                                 j->larg = pull_up_subqueries(parse, j->larg);
 389                                 break;
 390                         case JOIN_FULL:
 391                                 break;
 392                         case JOIN_RIGHT:
 393                                 j->rarg = pull_up_subqueries(parse, j->rarg);
 394                                 break;
 395                         case JOIN_UNION:
 396
 397                                 /*
 398                                  * This is where we fail if upper levels of planner
 399                                  * haven't rewritten UNION JOIN as an Append ...
 400                                  */
 401                                 elog(ERROR, "UNION JOIN is not implemented yet");
 402                                 break;
 403                         default:
 404                                 elog(ERROR, "pull_up_subqueries: unexpected join type %d",
 405                                          j->jointype);
 406                                 break;
 407                 }
 408         }
 409         else
 410                 elog(ERROR, "pull_up_subqueries: unexpected node type %d",
 411                          nodeTag(jtnode));
 412         return jtnode;
 413 }
 414
 415 /*
 416  * is_simple_subquery
 417  *        Check a subquery in the range table to see if it's simple enough
 418  *        to pull up into the parent query.
 419  */
 420 static bool
 421 is_simple_subquery(Query *subquery)
 422 {
 423         /*
 424          * Let's just make sure it's a valid subselect ...
 425          */
 426         if (!IsA(subquery, Query) ||
 427                 subquery->commandType != CMD_SELECT ||
 428                 subquery->resultRelation != 0 ||
 429                 subquery->into != NULL ||
 430                 subquery->isPortal)
 431                 elog(ERROR, "is_simple_subquery: subquery is bogus");
 432
 433         /*
 434          * Can't currently pull up a query with setops. Maybe after querytree
 435          * redesign...
 436          */
 437         if (subquery->setOperations)
 438                 return false;
 439
 440         /*
 441          * Can't pull up a subquery involving grouping, aggregation, sorting,
 442          * or limiting.
 443          */
 444         if (subquery->hasAggs ||
 445                 subquery->groupClause ||
 446                 subquery->havingQual ||
 447                 subquery->sortClause ||
 448                 subquery->distinctClause ||
 449                 subquery->limitOffset ||
 450                 subquery->limitCount)
 451                 return false;
 452
 453         /*
 454          * Hack: don't try to pull up a subquery with an empty jointree.
 455          * query_planner() will correctly generate a Result plan for a
 456          * jointree that's totally empty, but I don't think the right things
 457          * happen if an empty FromExpr appears lower down in a jointree. Not
 458          * worth working hard on this, just to collapse SubqueryScan/Result
 459          * into Result...
 460          */
 461         if (subquery->jointree->fromlist == NIL)
 462                 return false;
 463
 464         return true;
 465 }
 466
 467 /*
 468  * Helper routine for pull_up_subqueries: do ResolveNew on every expression
 469  * in the jointree, without changing the jointree structure itself.  Ugly,
 470  * but there's no other way...
 471  */
 472 static void
 473 resolvenew_in_jointree(Node *jtnode, int varno, List *subtlist)
 474 {
 475         if (jtnode == NULL)
 476                 return;
 477         if (IsA(jtnode, RangeTblRef))
 478         {
 479                 /* nothing to do here */
 480         }
 481         else if (IsA(jtnode, FromExpr))
 482         {
 483                 FromExpr   *f = (FromExpr *) jtnode;
 484                 List       *l;
 485
 486                 foreach(l, f->fromlist)
 487                         resolvenew_in_jointree(lfirst(l), varno, subtlist);
 488                 f->quals = ResolveNew(f->quals,
 489                                                           varno, 0, subtlist, CMD_SELECT, 0);
 490         }
 491         else if (IsA(jtnode, JoinExpr))
 492         {
 493                 JoinExpr   *j = (JoinExpr *) jtnode;
 494
 495                 resolvenew_in_jointree(j->larg, varno, subtlist);
 496                 resolvenew_in_jointree(j->rarg, varno, subtlist);
 497                 j->quals = ResolveNew(j->quals,
 498                                                           varno, 0, subtlist, CMD_SELECT, 0);
 499
 500                 /*
 501                  * We don't bother to update the colvars list, since it won't be
 502                  * used again ...
 503                  */
 504         }
 505         else
 506                 elog(ERROR, "resolvenew_in_jointree: unexpected node type %d",
 507                          nodeTag(jtnode));
 508 }
 509
 510 /*
 511  * preprocess_jointree
 512  *              Attempt to simplify a query's jointree.
 513  *
 514  * If we succeed in pulling up a subquery then we might form a jointree
 515  * in which a FromExpr is a direct child of another FromExpr.  In that
 516  * case we can consider collapsing the two FromExprs into one.  This is
 517  * an optional conversion, since the planner will work correctly either
 518  * way.  But we may find a better plan (at the cost of more planning time)
 519  * if we merge the two nodes.
 520  *
 521  * NOTE: don't try to do this in the same jointree scan that does subquery
 522  * pullup!      Since we're changing the jointree structure here, that wouldn't
 523  * work reliably --- see comments for pull_up_subqueries().
 524  */
 525 static Node *
 526 preprocess_jointree(Query *parse, Node *jtnode)
 527 {
 528         if (jtnode == NULL)
 529                 return NULL;
 530         if (IsA(jtnode, RangeTblRef))
 531         {
 532                 /* nothing to do here... */
 533         }
 534         else if (IsA(jtnode, FromExpr))
 535         {
 536                 FromExpr   *f = (FromExpr *) jtnode;
 537                 List       *newlist = NIL;
 538                 List       *l;
 539
 540                 foreach(l, f->fromlist)
 541                 {
 542                         Node       *child = (Node *) lfirst(l);
 543
 544                         /* Recursively simplify the child... */
 545                         child = preprocess_jointree(parse, child);
 546                         /* Now, is it a FromExpr? */
 547                         if (child && IsA(child, FromExpr))
 548                         {
 549                                 /*
 550                                  * Yes, so do we want to merge it into parent?  Always do
 551                                  * so if child has just one element (since that doesn't
 552                                  * make the parent's list any longer).  Otherwise we have
 553                                  * to be careful about the increase in planning time
 554                                  * caused by combining the two join search spaces into
 555                                  * one.  Our heuristic is to merge if the merge will
 556                                  * produce a join list no longer than GEQO_RELS/2.
 557                                  * (Perhaps need an additional user parameter?)
 558                                  */
 559                                 FromExpr   *subf = (FromExpr *) child;
 560                                 int                     childlen = length(subf->fromlist);
 561                                 int                     myothers = length(newlist) + length(lnext(l));
 562
 563                                 if (childlen <= 1 || (childlen + myothers) <= geqo_rels / 2)
 564                                 {
 565                                         newlist = nconc(newlist, subf->fromlist);
 566                                         f->quals = make_and_qual(f->quals, subf->quals);
 567                                 }
 568                                 else
 569                                         newlist = lappend(newlist, child);
 570                         }
 571                         else
 572                                 newlist = lappend(newlist, child);
 573                 }
 574                 f->fromlist = newlist;
 575         }
 576         else if (IsA(jtnode, JoinExpr))
 577         {
 578                 JoinExpr   *j = (JoinExpr *) jtnode;
 579
 580                 /* Can't usefully change the JoinExpr, but recurse on children */
 581                 j->larg = preprocess_jointree(parse, j->larg);
 582                 j->rarg = preprocess_jointree(parse, j->rarg);
 583         }
 584         else
 585                 elog(ERROR, "preprocess_jointree: unexpected node type %d",
 586                          nodeTag(jtnode));
 587         return jtnode;
 588 }
 589
 590 /*
 591  * preprocess_expression
 592  *              Do subquery_planner's preprocessing work for an expression,
 593  *              which can be a targetlist, a WHERE clause (including JOIN/ON
 594  *              conditions), or a HAVING clause.
 595  */
 596 static Node *
 597 preprocess_expression(Query *parse, Node *expr, int kind)
 598 {
 599         /*
 600          * Simplify constant expressions.
 601          *
 602          * Note that at this point quals have not yet been converted to
 603          * implicit-AND form, so we can apply eval_const_expressions directly.
 604          * Also note that we need to do this before SS_process_sublinks,
 605          * because that routine inserts bogus "Const" nodes.
 606          */
 607         expr = eval_const_expressions(expr);
 608
 609         /*
 610          * If it's a qual or havingQual, canonicalize it, and convert it to
 611          * implicit-AND format.
 612          *
 613          * XXX Is there any value in re-applying eval_const_expressions after
 614          * canonicalize_qual?
 615          */
 616         if (kind != EXPRKIND_TARGET)
 617         {
 618                 expr = (Node *) canonicalize_qual((Expr *) expr, true);
 619
 620 #ifdef OPTIMIZER_DEBUG
 621                 printf("After canonicalize_qual()\n");
 622                 pprint(expr);
 623 #endif
 624         }
 625
 626         if (parse->hasSubLinks)
 627         {
 628                 /* Expand SubLinks to SubPlans */
 629                 expr = SS_process_sublinks(expr);
 630
 631                 if (kind != EXPRKIND_WHERE &&
 632                         (parse->groupClause != NIL || parse->hasAggs))
 633                 {
 634                         /*
 635                          * Check for ungrouped variables passed to subplans.  Note we
 636                          * do NOT do this for subplans in WHERE (or JOIN/ON); it's
 637                          * legal there because WHERE is evaluated pre-GROUP.
 638                          */
 639                         check_subplans_for_ungrouped_vars(expr, parse);
 640                 }
 641         }
 642
 643         /* Replace uplevel vars with Param nodes */
 644         if (PlannerQueryLevel > 1)
 645                 expr = SS_replace_correlation_vars(expr);
 646
 647         return expr;
 648 }
 649
 650 /*
 651  * preprocess_qual_conditions
 652  *              Recursively scan the query's jointree and do subquery_planner's
 653  *              preprocessing work on each qual condition found therein.
 654  */
 655 static void
 656 preprocess_qual_conditions(Query *parse, Node *jtnode)
 657 {
 658         if (jtnode == NULL)
 659                 return;
 660         if (IsA(jtnode, RangeTblRef))
 661         {
 662                 /* nothing to do here */
 663         }
 664         else if (IsA(jtnode, FromExpr))
 665         {
 666                 FromExpr   *f = (FromExpr *) jtnode;
 667                 List       *l;
 668
 669                 foreach(l, f->fromlist)
 670                         preprocess_qual_conditions(parse, lfirst(l));
 671
 672                 f->quals = preprocess_expression(parse, f->quals, EXPRKIND_WHERE);
 673         }
 674         else if (IsA(jtnode, JoinExpr))
 675         {
 676                 JoinExpr   *j = (JoinExpr *) jtnode;
 677
 678                 preprocess_qual_conditions(parse, j->larg);
 679                 preprocess_qual_conditions(parse, j->rarg);
 680
 681                 j->quals = preprocess_expression(parse, j->quals, EXPRKIND_WHERE);
 682         }
 683         else
 684                 elog(ERROR, "preprocess_qual_conditions: unexpected node type %d",
 685                          nodeTag(jtnode));
 686 }
 687
 688 /*--------------------
 689  * inheritance_planner
 690  *        Generate a plan in the case where the result relation is an
 691  *        inheritance set.
 692  *
 693  * We have to handle this case differently from cases where a source
 694  * relation is an inheritance set.      Source inheritance is expanded at
 695  * the bottom of the plan tree (see allpaths.c), but target inheritance
 696  * has to be expanded at the top.  The reason is that for UPDATE, each
 697  * target relation needs a different targetlist matching its own column
 698  * set.  (This is not so critical for DELETE, but for simplicity we treat
 699  * inherited DELETE the same way.)      Fortunately, the UPDATE/DELETE target
 700  * can never be the nullable side of an outer join, so it's OK to generate
 701  * the plan this way.
 702  *
 703  * parse is the querytree produced by the parser & rewriter.
 704  * inheritlist is an integer list of RT indexes for the result relation set.
 705  *
 706  * Returns a query plan.
 707  *--------------------
 708  */
 709 static Plan *
 710 inheritance_planner(Query *parse, List *inheritlist)
 711 {
 712         int                     parentRTindex = parse->resultRelation;
 713         Oid                     parentOID = getrelid(parentRTindex, parse->rtable);
 714         List       *subplans = NIL;
 715         List       *tlist = NIL;
 716         List       *l;
 717
 718         foreach(l, inheritlist)
 719         {
 720                 int                     childRTindex = lfirsti(l);
 721                 Oid                     childOID = getrelid(childRTindex, parse->rtable);
 722                 Query      *subquery;
 723                 Plan       *subplan;
 724
 725                 /* Generate modified query with this rel as target */
 726                 subquery = (Query *) adjust_inherited_attrs((Node *) parse,
 727                                                                                                 parentRTindex, parentOID,
 728                                                                                                  childRTindex, childOID);
 729                 /* Generate plan */
 730                 subplan = grouping_planner(subquery, 0.0 /* retrieve all tuples */ );
 731                 subplans = lappend(subplans, subplan);
 732                 /* Save preprocessed tlist from first rel for use in Append */
 733                 if (tlist == NIL)
 734                         tlist = subplan->targetlist;
 735         }
 736
 737         /* Save the target-relations list for the executor, too */
 738         parse->resultRelations = inheritlist;
 739
 740         return (Plan *) make_append(subplans, true, tlist);
 741 }
 742
 743 /*--------------------
 744  * grouping_planner
 745  *        Perform planning steps related to grouping, aggregation, etc.
 746  *        This primarily means adding top-level processing to the basic
 747  *        query plan produced by query_planner.
 748  *
 749  * parse is the querytree produced by the parser & rewriter.
 750  * tuple_fraction is the fraction of tuples we expect will be retrieved
 751  *
 752  * tuple_fraction is interpreted as follows:
 753  *        < 0: determine fraction by inspection of query (normal case)
 754  *        0: expect all tuples to be retrieved
 755  *        0 < tuple_fraction < 1: expect the given fraction of tuples available
 756  *              from the plan to be retrieved
 757  *        tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
 758  *              expected to be retrieved (ie, a LIMIT specification)
 759  * The normal case is to pass -1, but some callers pass values >= 0 to
 760  * override this routine's determination of the appropriate fraction.
 761  *
 762  * Returns a query plan.
 763  *--------------------
 764  */
 765 static Plan *
 766 grouping_planner(Query *parse, double tuple_fraction)
 767 {
 768         List       *tlist = parse->targetList;
 769         Plan       *result_plan;
 770         List       *current_pathkeys;
 771         List       *group_pathkeys;
 772         List       *sort_pathkeys;
 773         AttrNumber *groupColIdx = NULL;
 774
 775         if (parse->setOperations)
 776         {
 777                 /*
 778                  * Construct the plan for set operations.  The result will not
 779                  * need any work except perhaps a top-level sort and/or LIMIT.
 780                  */
 781                 result_plan = plan_set_operations(parse);
 782
 783                 /*
 784                  * We should not need to call preprocess_targetlist, since we must
 785                  * be in a SELECT query node.  Instead, use the targetlist
 786                  * returned by plan_set_operations (since this tells whether it
 787                  * returned any resjunk columns!), and transfer any sort key
 788                  * information from the original tlist.
 789                  */
 790                 Assert(parse->commandType == CMD_SELECT);
 791
 792                 tlist = postprocess_setop_tlist(result_plan->targetlist, tlist);
 793
 794                 /*
 795                  * Can't handle FOR UPDATE here (parser should have checked
 796                  * already, but let's make sure).
 797                  */
 798                 if (parse->rowMarks)
 799                         elog(ERROR, "SELECT FOR UPDATE is not allowed with UNION/INTERSECT/EXCEPT");
 800
 801                 /*
 802                  * We set current_pathkeys NIL indicating we do not know sort
 803                  * order.  This is correct when the top set operation is UNION
 804                  * ALL, since the appended-together results are unsorted even if
 805                  * the subplans were sorted.  For other set operations we could be
 806                  * smarter --- room for future improvement!
 807                  */
 808                 current_pathkeys = NIL;
 809
 810                 /*
 811                  * Calculate pathkeys that represent grouping/ordering
 812                  * requirements (grouping should always be null, but...)
 813                  */
 814                 group_pathkeys = make_pathkeys_for_sortclauses(parse->groupClause,
 815                                                                                                            tlist);
 816                 sort_pathkeys = make_pathkeys_for_sortclauses(parse->sortClause,
 817                                                                                                           tlist);
 818         }
 819         else
 820         {
 821                 List       *sub_tlist;
 822
 823                 /* Preprocess targetlist in case we are inside an INSERT/UPDATE. */
 824                 tlist = preprocess_targetlist(tlist,
 825                                                                           parse->commandType,
 826                                                                           parse->resultRelation,
 827                                                                           parse->rtable);
 828
 829                 /*
 830                  * Add TID targets for rels selected FOR UPDATE (should this be
 831                  * done in preprocess_targetlist?).  The executor uses the TID to
 832                  * know which rows to lock, much as for UPDATE or DELETE.
 833                  */
 834                 if (parse->rowMarks)
 835                 {
 836                         List       *l;
 837
 838                         /*
 839                          * We've got trouble if the FOR UPDATE appears inside
 840                          * grouping, since grouping renders a reference to individual
 841                          * tuple CTIDs invalid.  This is also checked at parse time,
 842                          * but that's insufficient because of rule substitution, query
 843                          * pullup, etc.
 844                          */
 845                         CheckSelectForUpdate(parse);
 846
 847                         /*
 848                          * Currently the executor only supports FOR UPDATE at top
 849                          * level
 850                          */
 851                         if (PlannerQueryLevel > 1)
 852                                 elog(ERROR, "SELECT FOR UPDATE is not allowed in subselects");
 853
 854                         foreach(l, parse->rowMarks)
 855                         {
 856                                 Index           rti = lfirsti(l);
 857                                 char       *resname;
 858                                 Resdom     *resdom;
 859                                 Var                *var;
 860                                 TargetEntry *ctid;
 861
 862                                 resname = (char *) palloc(32);
 863                                 sprintf(resname, "ctid%u", rti);
 864                                 resdom = makeResdom(length(tlist) + 1,
 865                                                                         TIDOID,
 866                                                                         -1,
 867                                                                         resname,
 868                                                                         true);
 869
 870                                 var = makeVar(rti,
 871                                                           SelfItemPointerAttributeNumber,
 872                                                           TIDOID,
 873                                                           -1,
 874                                                           0);
 875
 876                                 ctid = makeTargetEntry(resdom, (Node *) var);
 877                                 tlist = lappend(tlist, ctid);
 878                         }
 879                 }
 880
 881                 /*
 882                  * Generate appropriate target list for subplan; may be different
 883                  * from tlist if grouping or aggregation is needed.
 884                  */
 885                 sub_tlist = make_subplanTargetList(parse, tlist, &groupColIdx);
 886
 887                 /*
 888                  * Calculate pathkeys that represent grouping/ordering
 889                  * requirements
 890                  */
 891                 group_pathkeys = make_pathkeys_for_sortclauses(parse->groupClause,
 892                                                                                                            tlist);
 893                 sort_pathkeys = make_pathkeys_for_sortclauses(parse->sortClause,
 894                                                                                                           tlist);
 895
 896                 /*
 897                  * Figure out whether we need a sorted result from query_planner.
 898                  *
 899                  * If we have a GROUP BY clause, then we want a result sorted
 900                  * properly for grouping.  Otherwise, if there is an ORDER BY
 901                  * clause, we want to sort by the ORDER BY clause.      (Note: if we
 902                  * have both, and ORDER BY is a superset of GROUP BY, it would be
 903                  * tempting to request sort by ORDER BY --- but that might just
 904                  * leave us failing to exploit an available sort order at all.
 905                  * Needs more thought...)
 906                  */
 907                 if (parse->groupClause)
 908                         parse->query_pathkeys = group_pathkeys;
 909                 else if (parse->sortClause)
 910                         parse->query_pathkeys = sort_pathkeys;
 911                 else
 912                         parse->query_pathkeys = NIL;
 913
 914                 /*
 915                  * Figure out whether we expect to retrieve all the tuples that
 916                  * the plan can generate, or to stop early due to outside factors
 917                  * such as a cursor.  If the caller passed a value >= 0, believe
 918                  * that value, else do our own examination of the query context.
 919                  */
 920                 if (tuple_fraction < 0.0)
 921                 {
 922                         /* Initial assumption is we need all the tuples */
 923                         tuple_fraction = 0.0;
 924
 925                         /*
 926                          * Check for retrieve-into-portal, ie DECLARE CURSOR.
 927                          *
 928                          * We have no real idea how many tuples the user will ultimately
 929                          * FETCH from a cursor, but it seems a good bet that he
 930                          * doesn't want 'em all.  Optimize for 10% retrieval (you
 931                          * gotta better number?  Should this be a SETtable parameter?)
 932                          */
 933                         if (parse->isPortal)
 934                                 tuple_fraction = 0.10;
 935                 }
 936
 937                 /*
 938                  * Adjust tuple_fraction if we see that we are going to apply
 939                  * limiting/grouping/aggregation/etc.  This is not overridable by
 940                  * the caller, since it reflects plan actions that this routine
 941                  * will certainly take, not assumptions about context.
 942                  */
 943                 if (parse->limitCount != NULL)
 944                 {
 945                         /*
 946                          * A LIMIT clause limits the absolute number of tuples
 947                          * returned. However, if it's not a constant LIMIT then we
 948                          * have to punt; for lack of a better idea, assume 10% of the
 949                          * plan's result is wanted.
 950                          */
 951                         double          limit_fraction = 0.0;
 952
 953                         if (IsA(parse->limitCount, Const))
 954                         {
 955                                 Const      *limitc = (Const *) parse->limitCount;
 956                                 int32           count = DatumGetInt32(limitc->constvalue);
 957
 958                                 /*
 959                                  * A NULL-constant LIMIT represents "LIMIT ALL", which we
 960                                  * treat the same as no limit (ie, expect to retrieve all
 961                                  * the tuples).
 962                                  */
 963                                 if (!limitc->constisnull && count > 0)
 964                                 {
 965                                         limit_fraction = (double) count;
 966                                         /* We must also consider the OFFSET, if present */
 967                                         if (parse->limitOffset != NULL)
 968                                         {
 969                                                 if (IsA(parse->limitOffset, Const))
 970                                                 {
 971                                                         int32           offset;
 972
 973                                                         limitc = (Const *) parse->limitOffset;
 974                                                         offset = DatumGetInt32(limitc->constvalue);
 975                                                         if (!limitc->constisnull && offset > 0)
 976                                                                 limit_fraction += (double) offset;
 977                                                 }
 978                                                 else
 979                                                 {
 980                                                         /* OFFSET is an expression ... punt ... */
 981                                                         limit_fraction = 0.10;
 982                                                 }
 983                                         }
 984                                 }
 985                         }
 986                         else
 987                         {
 988                                 /* LIMIT is an expression ... punt ... */
 989                                 limit_fraction = 0.10;
 990                         }
 991
 992                         if (limit_fraction > 0.0)
 993                         {
 994                                 /*
 995                                  * If we have absolute limits from both caller and LIMIT,
 996                                  * use the smaller value; if one is fractional and the
 997                                  * other absolute, treat the fraction as a fraction of the
 998                                  * absolute value; else we can multiply the two fractions
 999                                  * together.
1000                                  */
1001                                 if (tuple_fraction >= 1.0)
1002                                 {
1003                                         if (limit_fraction >= 1.0)
1004                                         {
1005                                                 /* both absolute */
1006                                                 tuple_fraction = Min(tuple_fraction, limit_fraction);
1007                                         }
1008                                         else
1009                                         {
1010                                                 /* caller absolute, limit fractional */
1011                                                 tuple_fraction *= limit_fraction;
1012                                                 if (tuple_fraction < 1.0)
1013                                                         tuple_fraction = 1.0;
1014                                         }
1015                                 }
1016                                 else if (tuple_fraction > 0.0)
1017                                 {
1018                                         if (limit_fraction >= 1.0)
1019                                         {
1020                                                 /* caller fractional, limit absolute */
1021                                                 tuple_fraction *= limit_fraction;
1022                                                 if (tuple_fraction < 1.0)
1023                                                         tuple_fraction = 1.0;
1024                                         }
1025                                         else
1026                                         {
1027                                                 /* both fractional */
1028                                                 tuple_fraction *= limit_fraction;
1029                                         }
1030                                 }
1031                                 else
1032                                 {
1033                                         /* no info from caller, just use limit */
1034                                         tuple_fraction = limit_fraction;
1035                                 }
1036                         }
1037                 }
1038
1039                 if (parse->groupClause)
1040                 {
1041                         /*
1042                          * In GROUP BY mode, we have the little problem that we don't
1043                          * really know how many input tuples will be needed to make a
1044                          * group, so we can't translate an output LIMIT count into an
1045                          * input count.  For lack of a better idea, assume 25% of the
1046                          * input data will be processed if there is any output limit.
1047                          * However, if the caller gave us a fraction rather than an
1048                          * absolute count, we can keep using that fraction (which
1049                          * amounts to assuming that all the groups are about the same
1050                          * size).
1051                          */
1052                         if (tuple_fraction >= 1.0)
1053                                 tuple_fraction = 0.25;
1054
1055                         /*
1056                          * If both GROUP BY and ORDER BY are specified, we will need
1057                          * two levels of sort --- and, therefore, certainly need to
1058                          * read all the input tuples --- unless ORDER BY is a subset
1059                          * of GROUP BY.  (We have not yet canonicalized the pathkeys,
1060                          * so must use the slower noncanonical comparison method.)
1061                          */
1062                         if (parse->groupClause && parse->sortClause &&
1063                                 !noncanonical_pathkeys_contained_in(sort_pathkeys,
1064                                                                                                         group_pathkeys))
1065                                 tuple_fraction = 0.0;
1066                 }
1067                 else if (parse->hasAggs)
1068                 {
1069                         /*
1070                          * Ungrouped aggregate will certainly want all the input
1071                          * tuples.
1072                          */
1073                         tuple_fraction = 0.0;
1074                 }
1075                 else if (parse->distinctClause)
1076                 {
1077                         /*
1078                          * SELECT DISTINCT, like GROUP, will absorb an unpredictable
1079                          * number of input tuples per output tuple.  Handle the same
1080                          * way.
1081                          */
1082                         if (tuple_fraction >= 1.0)
1083                                 tuple_fraction = 0.25;
1084                 }
1085
1086                 /* Generate the basic plan for this Query */
1087                 result_plan = query_planner(parse,
1088                                                                         sub_tlist,
1089                                                                         tuple_fraction);
1090
1091                 /*
1092                  * query_planner returns actual sort order (which is not
1093                  * necessarily what we requested) in query_pathkeys.
1094                  */
1095                 current_pathkeys = parse->query_pathkeys;
1096         }
1097
1098         /*
1099          * We couldn't canonicalize group_pathkeys and sort_pathkeys before
1100          * running query_planner(), so do it now.
1101          */
1102         group_pathkeys = canonicalize_pathkeys(parse, group_pathkeys);
1103         sort_pathkeys = canonicalize_pathkeys(parse, sort_pathkeys);
1104
1105         /*
1106          * If we have a GROUP BY clause, insert a group node (plus the
1107          * appropriate sort node, if necessary).
1108          */
1109         if (parse->groupClause)
1110         {
1111                 bool            tuplePerGroup;
1112                 List       *group_tlist;
1113                 bool            is_sorted;
1114
1115                 /*
1116                  * Decide whether how many tuples per group the Group node needs
1117                  * to return. (Needs only one tuple per group if no aggregate is
1118                  * present. Otherwise, need every tuple from the group to do the
1119                  * aggregation.)  Note tuplePerGroup is named backwards :-(
1120                  */
1121                 tuplePerGroup = parse->hasAggs;
1122
1123                 /*
1124                  * If there are aggregates then the Group node should just return
1125                  * the same set of vars as the subplan did (but we can exclude any
1126                  * GROUP BY expressions).  If there are no aggregates then the
1127                  * Group node had better compute the final tlist.
1128                  */
1129                 if (parse->hasAggs)
1130                         group_tlist = flatten_tlist(result_plan->targetlist);
1131                 else
1132                         group_tlist = tlist;
1133
1134                 /*
1135                  * Figure out whether the path result is already ordered the way
1136                  * we need it --- if so, no need for an explicit sort step.
1137                  */
1138                 if (pathkeys_contained_in(group_pathkeys, current_pathkeys))
1139                 {
1140                         is_sorted = true;       /* no sort needed now */
1141                         /* current_pathkeys remains unchanged */
1142                 }
1143                 else
1144                 {
1145                         /*
1146                          * We will need to do an explicit sort by the GROUP BY clause.
1147                          * make_groupplan will do the work, but set current_pathkeys
1148                          * to indicate the resulting order.
1149                          */
1150                         is_sorted = false;
1151                         current_pathkeys = group_pathkeys;
1152                 }
1153
1154                 result_plan = make_groupplan(parse,
1155                                                                          group_tlist,
1156                                                                          tuplePerGroup,
1157                                                                          parse->groupClause,
1158                                                                          groupColIdx,
1159                                                                          is_sorted,
1160                                                                          result_plan);
1161         }
1162
1163         /*
1164          * If aggregate is present, insert the Agg node
1165          *
1166          * HAVING clause, if any, becomes qual of the Agg node
1167          */
1168         if (parse->hasAggs)
1169         {
1170                 result_plan = (Plan *) make_agg(tlist,
1171                                                                                 (List *) parse->havingQual,
1172                                                                                 result_plan);
1173                 /* Note: Agg does not affect any existing sort order of the tuples */
1174         }
1175         else
1176         {
1177                 /* If there are no Aggs, we shouldn't have any HAVING qual anymore */
1178                 Assert(parse->havingQual == NULL);
1179         }
1180
1181         /*
1182          * If we were not able to make the plan come out in the right order,
1183          * add an explicit sort step.
1184          */
1185         if (parse->sortClause)
1186         {
1187                 if (!pathkeys_contained_in(sort_pathkeys, current_pathkeys))
1188                         result_plan = make_sortplan(parse, tlist, result_plan,
1189                                                                                 parse->sortClause);
1190         }
1191
1192         /*
1193          * If there is a DISTINCT clause, add the UNIQUE node.
1194          */
1195         if (parse->distinctClause)
1196         {
1197                 result_plan = (Plan *) make_unique(tlist, result_plan,
1198                                                                                    parse->distinctClause);
1199         }
1200
1201         /*
1202          * Finally, if there is a LIMIT/OFFSET clause, add the LIMIT node.
1203          */
1204         if (parse->limitOffset || parse->limitCount)
1205         {
1206                 result_plan = (Plan *) make_limit(tlist, result_plan,
1207                                                                                   parse->limitOffset,
1208                                                                                   parse->limitCount);
1209         }
1210
1211         return result_plan;
1212 }
1213
1214 /*---------------
1215  * make_subplanTargetList
1216  *        Generate appropriate target list when grouping is required.
1217  *
1218  * When grouping_planner inserts Aggregate and/or Group plan nodes above
1219  * the result of query_planner, we typically want to pass a different
1220  * target list to query_planner than the outer plan nodes should have.
1221  * This routine generates the correct target list for the subplan.
1222  *
1223  * The initial target list passed from the parser already contains entries
1224  * for all ORDER BY and GROUP BY expressions, but it will not have entries
1225  * for variables used only in HAVING clauses; so we need to add those
1226  * variables to the subplan target list.  Also, if we are doing either
1227  * grouping or aggregation, we flatten all expressions except GROUP BY items
1228  * into their component variables; the other expressions will be computed by
1229  * the inserted nodes rather than by the subplan.  For example,
1230  * given a query like
1231  *              SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
1232  * we want to pass this targetlist to the subplan:
1233  *              a,b,c,d,a+b
1234  * where the a+b target will be used by the Sort/Group steps, and the
1235  * other targets will be used for computing the final results.  (In the
1236  * above example we could theoretically suppress the a and b targets and
1237  * use only a+b, but it's not really worth the trouble.)
1238  *
1239  * 'parse' is the query being processed.
1240  * 'tlist' is the query's target list.
1241  * 'groupColIdx' receives an array of column numbers for the GROUP BY
1242  * expressions (if there are any) in the subplan's target list.
1243  *
1244  * The result is the targetlist to be passed to the subplan.
1245  *---------------
1246  */
1247 static List *
1248 make_subplanTargetList(Query *parse,
1249                                            List *tlist,
1250                                            AttrNumber **groupColIdx)
1251 {
1252         List       *sub_tlist;
1253         List       *extravars;
1254         int                     numCols;
1255
1256         *groupColIdx = NULL;
1257
1258         /*
1259          * If we're not grouping or aggregating, nothing to do here;
1260          * query_planner should receive the unmodified target list.
1261          */
1262         if (!parse->hasAggs && !parse->groupClause && !parse->havingQual)
1263                 return tlist;
1264
1265         /*
1266          * Otherwise, start with a "flattened" tlist (having just the vars
1267          * mentioned in the targetlist and HAVING qual --- but not upper-
1268          * level Vars; they will be replaced by Params later on).
1269          */
1270         sub_tlist = flatten_tlist(tlist);
1271         extravars = pull_var_clause(parse->havingQual, false);
1272         sub_tlist = add_to_flat_tlist(sub_tlist, extravars);
1273         freeList(extravars);
1274
1275         /*
1276          * If grouping, create sub_tlist entries for all GROUP BY expressions
1277          * (GROUP BY items that are simple Vars should be in the list
1278          * already), and make an array showing where the group columns are in
1279          * the sub_tlist.
1280          */
1281         numCols = length(parse->groupClause);
1282         if (numCols > 0)
1283         {
1284                 int                     keyno = 0;
1285                 AttrNumber *grpColIdx;
1286                 List       *gl;
1287
1288                 grpColIdx = (AttrNumber *) palloc(sizeof(AttrNumber) * numCols);
1289                 *groupColIdx = grpColIdx;
1290
1291                 foreach(gl, parse->groupClause)
1292                 {
1293                         GroupClause *grpcl = (GroupClause *) lfirst(gl);
1294                         Node       *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1295                         TargetEntry *te = NULL;
1296                         List       *sl;
1297
1298                         /* Find or make a matching sub_tlist entry */
1299                         foreach(sl, sub_tlist)
1300                         {
1301                                 te = (TargetEntry *) lfirst(sl);
1302                                 if (equal(groupexpr, te->expr))
1303                                         break;
1304                         }
1305                         if (!sl)
1306                         {
1307                                 te = makeTargetEntry(makeResdom(length(sub_tlist) + 1,
1308                                                                                                 exprType(groupexpr),
1309                                                                                                 exprTypmod(groupexpr),
1310                                                                                                 NULL,
1311                                                                                                 false),
1312                                                                          groupexpr);
1313                                 sub_tlist = lappend(sub_tlist, te);
1314                         }
1315
1316                         /* and save its resno */
1317                         grpColIdx[keyno++] = te->resdom->resno;
1318                 }
1319         }
1320
1321         return sub_tlist;
1322 }
1323
1324 /*
1325  * make_groupplan
1326  *              Add a Group node for GROUP BY processing.
1327  *              If we couldn't make the subplan produce presorted output for grouping,
1328  *              first add an explicit Sort node.
1329  */
1330 static Plan *
1331 make_groupplan(Query *parse,
1332                            List *group_tlist,
1333                            bool tuplePerGroup,
1334                            List *groupClause,
1335                            AttrNumber *grpColIdx,
1336                            bool is_presorted,
1337                            Plan *subplan)
1338 {
1339         int                     numCols = length(groupClause);
1340
1341         if (!is_presorted)
1342         {
1343                 /*
1344                  * The Sort node always just takes a copy of the subplan's tlist
1345                  * plus ordering information.  (This might seem inefficient if the
1346                  * subplan contains complex GROUP BY expressions, but in fact Sort
1347                  * does not evaluate its targetlist --- it only outputs the same
1348                  * tuples in a new order.  So the expressions we might be copying
1349                  * are just dummies with no extra execution cost.)
1350                  */
1351                 List       *sort_tlist = new_unsorted_tlist(subplan->targetlist);
1352                 int                     keyno = 0;
1353                 List       *gl;
1354
1355                 foreach(gl, groupClause)
1356                 {
1357                         GroupClause *grpcl = (GroupClause *) lfirst(gl);
1358                         TargetEntry *te = nth(grpColIdx[keyno] - 1, sort_tlist);
1359                         Resdom     *resdom = te->resdom;
1360
1361                         /*
1362                          * Check for the possibility of duplicate group-by clauses ---
1363                          * the parser should have removed 'em, but the Sort executor
1364                          * will get terribly confused if any get through!
1365                          */
1366                         if (resdom->reskey == 0)
1367                         {
1368                                 /* OK, insert the ordering info needed by the executor. */
1369                                 resdom->reskey = ++keyno;
1370                                 resdom->reskeyop = grpcl->sortop;
1371                         }
1372                 }
1373
1374                 Assert(keyno > 0);
1375
1376                 subplan = (Plan *) make_sort(parse, sort_tlist, subplan, keyno);
1377         }
1378
1379         return (Plan *) make_group(group_tlist, tuplePerGroup, numCols,
1380                                                            grpColIdx, subplan);
1381 }
1382
1383 /*
1384  * make_sortplan
1385  *        Add a Sort node to implement an explicit ORDER BY clause.
1386  */
1387 Plan *
1388 make_sortplan(Query *parse, List *tlist, Plan *plannode, List *sortcls)
1389 {
1390         List       *sort_tlist;
1391         List       *i;
1392         int                     keyno = 0;
1393
1394         /*
1395          * First make a copy of the tlist so that we don't corrupt the
1396          * original.
1397          */
1398         sort_tlist = new_unsorted_tlist(tlist);
1399
1400         foreach(i, sortcls)
1401         {
1402                 SortClause *sortcl = (SortClause *) lfirst(i);
1403                 TargetEntry *tle = get_sortgroupclause_tle(sortcl, sort_tlist);
1404                 Resdom     *resdom = tle->resdom;
1405
1406                 /*
1407                  * Check for the possibility of duplicate order-by clauses --- the
1408                  * parser should have removed 'em, but the executor will get
1409                  * terribly confused if any get through!
1410                  */
1411                 if (resdom->reskey == 0)
1412                 {
1413                         /* OK, insert the ordering info needed by the executor. */
1414                         resdom->reskey = ++keyno;
1415                         resdom->reskeyop = sortcl->sortop;
1416                 }
1417         }
1418
1419         Assert(keyno > 0);
1420
1421         return (Plan *) make_sort(parse, sort_tlist, plannode, keyno);
1422 }
1423
1424 /*
1425  * postprocess_setop_tlist
1426  *        Fix up targetlist returned by plan_set_operations().
1427  *
1428  * We need to transpose sort key info from the orig_tlist into new_tlist.
1429  * NOTE: this would not be good enough if we supported resjunk sort keys
1430  * for results of set operations --- then, we'd need to project a whole
1431  * new tlist to evaluate the resjunk columns.  For now, just elog if we
1432  * find any resjunk columns in orig_tlist.
1433  */
1434 static List *
1435 postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
1436 {
1437         List       *l;
1438
1439         foreach(l, new_tlist)
1440         {
1441                 TargetEntry *new_tle = (TargetEntry *) lfirst(l);
1442                 TargetEntry *orig_tle;
1443
1444                 /* ignore resjunk columns in setop result */
1445                 if (new_tle->resdom->resjunk)
1446                         continue;
1447
1448                 Assert(orig_tlist != NIL);
1449                 orig_tle = (TargetEntry *) lfirst(orig_tlist);
1450                 orig_tlist = lnext(orig_tlist);
1451                 if (orig_tle->resdom->resjunk)
1452                         elog(ERROR, "postprocess_setop_tlist: resjunk output columns not implemented");
1453                 Assert(new_tle->resdom->resno == orig_tle->resdom->resno);
1454                 Assert(new_tle->resdom->restype == orig_tle->resdom->restype);
1455                 new_tle->resdom->ressortgroupref = orig_tle->resdom->ressortgroupref;
1456         }
1457         if (orig_tlist != NIL)
1458                 elog(ERROR, "postprocess_setop_tlist: resjunk output columns not implemented");
1459         return new_tlist;
1460 }