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) 1996-2005, PostgreSQL Global Development Group
   7  * Portions Copyright (c) 1994, Regents of the University of California
   8  *
   9  *
  10  * IDENTIFICATION
  11  *        $PostgreSQL: pgsql/src/backend/optimizer/plan/planner.c,v 1.193 2005/09/24 22:54:37 tgl Exp $
  12  *
  13  *-------------------------------------------------------------------------
  14  */
  15
  16 #include "postgres.h"
  17
  18 #include <limits.h>
  19
  20 #include "catalog/pg_operator.h"
  21 #include "catalog/pg_type.h"
  22 #include "executor/executor.h"
  23 #include "executor/nodeAgg.h"
  24 #include "miscadmin.h"
  25 #include "nodes/makefuncs.h"
  26 #ifdef OPTIMIZER_DEBUG
  27 #include "nodes/print.h"
  28 #endif
  29 #include "optimizer/clauses.h"
  30 #include "optimizer/cost.h"
  31 #include "optimizer/pathnode.h"
  32 #include "optimizer/paths.h"
  33 #include "optimizer/planmain.h"
  34 #include "optimizer/planner.h"
  35 #include "optimizer/prep.h"
  36 #include "optimizer/subselect.h"
  37 #include "optimizer/tlist.h"
  38 #include "optimizer/var.h"
  39 #include "parser/parsetree.h"
  40 #include "parser/parse_expr.h"
  41 #include "parser/parse_oper.h"
  42 #include "utils/selfuncs.h"
  43 #include "utils/syscache.h"
  44
  45
  46 ParamListInfo PlannerBoundParamList = NULL;             /* current boundParams */
  47
  48
  49 /* Expression kind codes for preprocess_expression */
  50 #define EXPRKIND_QUAL   0
  51 #define EXPRKIND_TARGET 1
  52 #define EXPRKIND_RTFUNC 2
  53 #define EXPRKIND_LIMIT  3
  54 #define EXPRKIND_ININFO 4
  55
  56
  57 static Node *preprocess_expression(PlannerInfo *root, Node *expr, int kind);
  58 static void preprocess_qual_conditions(PlannerInfo *root, Node *jtnode);
  59 static Plan *inheritance_planner(PlannerInfo *root, List *inheritlist);
  60 static Plan *grouping_planner(PlannerInfo *root, double tuple_fraction);
  61 static double preprocess_limit(PlannerInfo *root,
  62                                                            double tuple_fraction,
  63                                                            int *offset_est, int *count_est);
  64 static bool choose_hashed_grouping(PlannerInfo *root, double tuple_fraction,
  65                                            Path *cheapest_path, Path *sorted_path,
  66                                            double dNumGroups, AggClauseCounts *agg_counts);
  67 static bool hash_safe_grouping(PlannerInfo *root);
  68 static List *make_subplanTargetList(PlannerInfo *root, List *tlist,
  69                                            AttrNumber **groupColIdx, bool *need_tlist_eval);
  70 static void locate_grouping_columns(PlannerInfo *root,
  71                                                 List *tlist,
  72                                                 List *sub_tlist,
  73                                                 AttrNumber *groupColIdx);
  74 static List *postprocess_setop_tlist(List *new_tlist, List *orig_tlist);
  75
  76
  77 /*****************************************************************************
  78  *
  79  *         Query optimizer entry point
  80  *
  81  *****************************************************************************/
  82 Plan *
  83 planner(Query *parse, bool isCursor, int cursorOptions,
  84                 ParamListInfo boundParams)
  85 {
  86         double          tuple_fraction;
  87         Plan       *result_plan;
  88         Index           save_PlannerQueryLevel;
  89         List       *save_PlannerParamList;
  90         ParamListInfo save_PlannerBoundParamList;
  91
  92         /*
  93          * The planner can be called recursively (an example is when
  94          * eval_const_expressions tries to pre-evaluate an SQL function). So,
  95          * these global state variables must be saved and restored.
  96          *
  97          * Query level and the param list cannot be moved into the per-query
  98          * PlannerInfo structure since their whole purpose is communication
  99          * across multiple sub-queries. Also, boundParams is explicitly info
 100          * from outside the query, and so is likewise better handled as a global
 101          * variable.
 102          *
 103          * Note we do NOT save and restore PlannerPlanId: it exists to assign
 104          * unique IDs to SubPlan nodes, and we want those IDs to be unique for
 105          * the life of a backend.  Also, PlannerInitPlan is saved/restored in
 106          * subquery_planner, not here.
 107          */
 108         save_PlannerQueryLevel = PlannerQueryLevel;
 109         save_PlannerParamList = PlannerParamList;
 110         save_PlannerBoundParamList = PlannerBoundParamList;
 111
 112         /* Initialize state for handling outer-level references and params */
 113         PlannerQueryLevel = 0;          /* will be 1 in top-level subquery_planner */
 114         PlannerParamList = NIL;
 115         PlannerBoundParamList = boundParams;
 116
 117         /* Determine what fraction of the plan is likely to be scanned */
 118         if (isCursor)
 119         {
 120                 /*
 121                  * We have no real idea how many tuples the user will ultimately
 122                  * FETCH from a cursor, but it seems a good bet that he doesn't
 123                  * want 'em all.  Optimize for 10% retrieval (you gotta better
 124                  * number?      Should this be a SETtable parameter?)
 125                  */
 126                 tuple_fraction = 0.10;
 127         }
 128         else
 129         {
 130                 /* Default assumption is we need all the tuples */
 131                 tuple_fraction = 0.0;
 132         }
 133
 134         /* primary planning entry point (may recurse for subqueries) */
 135         result_plan = subquery_planner(parse, tuple_fraction, NULL);
 136
 137         /* check we popped out the right number of levels */
 138         Assert(PlannerQueryLevel == 0);
 139
 140         /*
 141          * If creating a plan for a scrollable cursor, make sure it can run
 142          * backwards on demand.  Add a Material node at the top at need.
 143          */
 144         if (isCursor && (cursorOptions & CURSOR_OPT_SCROLL))
 145         {
 146                 if (!ExecSupportsBackwardScan(result_plan))
 147                         result_plan = materialize_finished_plan(result_plan);
 148         }
 149
 150         /* final cleanup of the plan */
 151         result_plan = set_plan_references(result_plan, parse->rtable);
 152
 153         /* executor wants to know total number of Params used overall */
 154         result_plan->nParamExec = list_length(PlannerParamList);
 155
 156         /* restore state for outer planner, if any */
 157         PlannerQueryLevel = save_PlannerQueryLevel;
 158         PlannerParamList = save_PlannerParamList;
 159         PlannerBoundParamList = save_PlannerBoundParamList;
 160
 161         return result_plan;
 162 }
 163
 164
 165 /*--------------------
 166  * subquery_planner
 167  *        Invokes the planner on a subquery.  We recurse to here for each
 168  *        sub-SELECT found in the query tree.
 169  *
 170  * parse is the querytree produced by the parser & rewriter.
 171  * tuple_fraction is the fraction of tuples we expect will be retrieved.
 172  * tuple_fraction is interpreted as explained for grouping_planner, below.
 173  *
 174  * If subquery_pathkeys isn't NULL, it receives a list of pathkeys indicating
 175  * the output sort ordering of the completed plan.
 176  *
 177  * Basically, this routine does the stuff that should only be done once
 178  * per Query object.  It then calls grouping_planner.  At one time,
 179  * grouping_planner could be invoked recursively on the same Query object;
 180  * that's not currently true, but we keep the separation between the two
 181  * routines anyway, in case we need it again someday.
 182  *
 183  * subquery_planner will be called recursively to handle sub-Query nodes
 184  * found within the query's expressions and rangetable.
 185  *
 186  * Returns a query plan.
 187  *--------------------
 188  */
 189 Plan *
 190 subquery_planner(Query *parse, double tuple_fraction,
 191                                  List **subquery_pathkeys)
 192 {
 193         List       *saved_initplan = PlannerInitPlan;
 194         int                     saved_planid = PlannerPlanId;
 195         PlannerInfo *root;
 196         Plan       *plan;
 197         List       *newHaving;
 198         List       *lst;
 199         ListCell   *l;
 200
 201         /* Set up for a new level of subquery */
 202         PlannerQueryLevel++;
 203         PlannerInitPlan = NIL;
 204
 205         /* Create a PlannerInfo data structure for this subquery */
 206         root = makeNode(PlannerInfo);
 207         root->parse = parse;
 208
 209         /*
 210          * Look for IN clauses at the top level of WHERE, and transform them
 211          * into joins.  Note that this step only handles IN clauses originally
 212          * at top level of WHERE; if we pull up any subqueries in the next
 213          * step, their INs are processed just before pulling them up.
 214          */
 215         root->in_info_list = NIL;
 216         if (parse->hasSubLinks)
 217                 parse->jointree->quals = pull_up_IN_clauses(root,
 218                                                                                                         parse->jointree->quals);
 219
 220         /*
 221          * Check to see if any subqueries in the rangetable can be merged into
 222          * this query.
 223          */
 224         parse->jointree = (FromExpr *)
 225                 pull_up_subqueries(root, (Node *) parse->jointree, false);
 226
 227         /*
 228          * Detect whether any rangetable entries are RTE_JOIN kind; if not, we
 229          * can avoid the expense of doing flatten_join_alias_vars().  Also
 230          * check for outer joins --- if none, we can skip reduce_outer_joins()
 231          * and some other processing.  This must be done after we have done
 232          * pull_up_subqueries, of course.
 233          *
 234          * Note: if reduce_outer_joins manages to eliminate all outer joins,
 235          * root->hasOuterJoins is not reset currently.  This is OK since its
 236          * purpose is merely to suppress unnecessary processing in simple cases.
 237          */
 238         root->hasJoinRTEs = false;
 239         root->hasOuterJoins = false;
 240         foreach(l, parse->rtable)
 241         {
 242                 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
 243
 244                 if (rte->rtekind == RTE_JOIN)
 245                 {
 246                         root->hasJoinRTEs = true;
 247                         if (IS_OUTER_JOIN(rte->jointype))
 248                         {
 249                                 root->hasOuterJoins = true;
 250                                 /* Can quit scanning once we find an outer join */
 251                                 break;
 252                         }
 253                 }
 254         }
 255
 256         /*
 257          * Set hasHavingQual to remember if HAVING clause is present.  Needed
 258          * because preprocess_expression will reduce a constant-true condition
 259          * to an empty qual list ... but "HAVING TRUE" is not a semantic no-op.
 260          */
 261         root->hasHavingQual = (parse->havingQual != NULL);
 262
 263         /*
 264          * Do expression preprocessing on targetlist and quals.
 265          */
 266         parse->targetList = (List *)
 267                 preprocess_expression(root, (Node *) parse->targetList,
 268                                                           EXPRKIND_TARGET);
 269
 270         preprocess_qual_conditions(root, (Node *) parse->jointree);
 271
 272         parse->havingQual = preprocess_expression(root, parse->havingQual,
 273                                                                                           EXPRKIND_QUAL);
 274
 275         parse->limitOffset = preprocess_expression(root, parse->limitOffset,
 276                                                                                            EXPRKIND_LIMIT);
 277         parse->limitCount = preprocess_expression(root, parse->limitCount,
 278                                                                                           EXPRKIND_LIMIT);
 279
 280         root->in_info_list = (List *)
 281                 preprocess_expression(root, (Node *) root->in_info_list,
 282                                                           EXPRKIND_ININFO);
 283
 284         /* Also need to preprocess expressions for function RTEs */
 285         foreach(l, parse->rtable)
 286         {
 287                 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
 288
 289                 if (rte->rtekind == RTE_FUNCTION)
 290                         rte->funcexpr = preprocess_expression(root, rte->funcexpr,
 291                                                                                                   EXPRKIND_RTFUNC);
 292         }
 293
 294         /*
 295          * In some cases we may want to transfer a HAVING clause into WHERE.
 296          * We cannot do so if the HAVING clause contains aggregates (obviously)
 297          * or volatile functions (since a HAVING clause is supposed to be executed
 298          * only once per group).  Also, it may be that the clause is so expensive
 299          * to execute that we're better off doing it only once per group, despite
 300          * the loss of selectivity.  This is hard to estimate short of doing the
 301          * entire planning process twice, so we use a heuristic: clauses
 302          * containing subplans are left in HAVING.  Otherwise, we move or copy
 303          * the HAVING clause into WHERE, in hopes of eliminating tuples before
 304          * aggregation instead of after.
 305          *
 306          * If the query has explicit grouping then we can simply move such a
 307          * clause into WHERE; any group that fails the clause will not be
 308          * in the output because none of its tuples will reach the grouping
 309          * or aggregation stage.  Otherwise we must have a degenerate
 310          * (variable-free) HAVING clause, which we put in WHERE so that
 311          * query_planner() can use it in a gating Result node, but also keep
 312          * in HAVING to ensure that we don't emit a bogus aggregated row.
 313          * (This could be done better, but it seems not worth optimizing.)
 314          *
 315          * Note that both havingQual and parse->jointree->quals are in
 316          * implicitly-ANDed-list form at this point, even though they are
 317          * declared as Node *.
 318          */
 319         newHaving = NIL;
 320         foreach(l, (List *) parse->havingQual)
 321         {
 322                 Node       *havingclause = (Node *) lfirst(l);
 323
 324                 if (contain_agg_clause(havingclause) ||
 325                         contain_volatile_functions(havingclause) ||
 326                         contain_subplans(havingclause))
 327                 {
 328                         /* keep it in HAVING */
 329                         newHaving = lappend(newHaving, havingclause);
 330                 }
 331                 else if (parse->groupClause)
 332                 {
 333                         /* move it to WHERE */
 334                         parse->jointree->quals = (Node *)
 335                                 lappend((List *) parse->jointree->quals, havingclause);
 336                 }
 337                 else
 338                 {
 339                         /* put a copy in WHERE, keep it in HAVING */
 340                         parse->jointree->quals = (Node *)
 341                                 lappend((List *) parse->jointree->quals,
 342                                                 copyObject(havingclause));
 343                         newHaving = lappend(newHaving, havingclause);
 344                 }
 345         }
 346         parse->havingQual = (Node *) newHaving;
 347
 348         /*
 349          * If we have any outer joins, try to reduce them to plain inner
 350          * joins. This step is most easily done after we've done expression
 351          * preprocessing.
 352          */
 353         if (root->hasOuterJoins)
 354                 reduce_outer_joins(root);
 355
 356         /*
 357          * See if we can simplify the jointree; opportunities for this may
 358          * come from having pulled up subqueries, or from flattening explicit
 359          * JOIN syntax.  We must do this after flattening JOIN alias
 360          * variables, since eliminating explicit JOIN nodes from the jointree
 361          * will cause get_relids_for_join() to fail.  But it should happen
 362          * after reduce_outer_joins, anyway.
 363          */
 364         parse->jointree = (FromExpr *)
 365                 simplify_jointree(root, (Node *) parse->jointree);
 366
 367         /*
 368          * Do the main planning.  If we have an inherited target relation,
 369          * that needs special processing, else go straight to
 370          * grouping_planner.
 371          */
 372         if (parse->resultRelation &&
 373                 (lst = expand_inherited_rtentry(root, parse->resultRelation)) != NIL)
 374                 plan = inheritance_planner(root, lst);
 375         else
 376                 plan = grouping_planner(root, tuple_fraction);
 377
 378         /*
 379          * If any subplans were generated, or if we're inside a subplan, build
 380          * initPlan list and extParam/allParam sets for plan nodes, and attach
 381          * the initPlans to the top plan node.
 382          */
 383         if (PlannerPlanId != saved_planid || PlannerQueryLevel > 1)
 384                 SS_finalize_plan(plan, parse->rtable);
 385
 386         /* Return sort ordering info if caller wants it */
 387         if (subquery_pathkeys)
 388                 *subquery_pathkeys = root->query_pathkeys;
 389
 390         /* Return to outer subquery context */
 391         PlannerQueryLevel--;
 392         PlannerInitPlan = saved_initplan;
 393         /* we do NOT restore PlannerPlanId; that's not an oversight! */
 394
 395         return plan;
 396 }
 397
 398 /*
 399  * preprocess_expression
 400  *              Do subquery_planner's preprocessing work for an expression,
 401  *              which can be a targetlist, a WHERE clause (including JOIN/ON
 402  *              conditions), or a HAVING clause.
 403  */
 404 static Node *
 405 preprocess_expression(PlannerInfo *root, Node *expr, int kind)
 406 {
 407         /*
 408          * Fall out quickly if expression is empty.  This occurs often enough
 409          * to be worth checking.  Note that null->null is the correct conversion
 410          * for implicit-AND result format, too.
 411          */
 412         if (expr == NULL)
 413                 return NULL;
 414
 415         /*
 416          * If the query has any join RTEs, replace join alias variables with
 417          * base-relation variables. We must do this before sublink processing,
 418          * else sublinks expanded out from join aliases wouldn't get
 419          * processed.
 420          */
 421         if (root->hasJoinRTEs)
 422                 expr = flatten_join_alias_vars(root, expr);
 423
 424         /*
 425          * Simplify constant expressions.
 426          *
 427          * Note: this also flattens nested AND and OR expressions into N-argument
 428          * form.  All processing of a qual expression after this point must be
 429          * careful to maintain AND/OR flatness --- that is, do not generate a tree
 430          * with AND directly under AND, nor OR directly under OR.
 431          *
 432          * Because this is a relatively expensive process, we skip it when the
 433          * query is trivial, such as "SELECT 2+2;" or "INSERT ... VALUES()".
 434          * The expression will only be evaluated once anyway, so no point in
 435          * pre-simplifying; we can't execute it any faster than the executor can,
 436          * and we will waste cycles copying the tree.  Notice however that we
 437          * still must do it for quals (to get AND/OR flatness); and if we are
 438          * in a subquery we should not assume it will be done only once.
 439          */
 440         if (root->parse->jointree->fromlist != NIL ||
 441                 kind == EXPRKIND_QUAL ||
 442                 PlannerQueryLevel > 1)
 443                 expr = eval_const_expressions(expr);
 444
 445         /*
 446          * If it's a qual or havingQual, canonicalize it.
 447          */
 448         if (kind == EXPRKIND_QUAL)
 449         {
 450                 expr = (Node *) canonicalize_qual((Expr *) expr);
 451
 452 #ifdef OPTIMIZER_DEBUG
 453                 printf("After canonicalize_qual()\n");
 454                 pprint(expr);
 455 #endif
 456         }
 457
 458         /* Expand SubLinks to SubPlans */
 459         if (root->parse->hasSubLinks)
 460                 expr = SS_process_sublinks(expr, (kind == EXPRKIND_QUAL));
 461
 462         /*
 463          * XXX do not insert anything here unless you have grokked the
 464          * comments in SS_replace_correlation_vars ...
 465          */
 466
 467         /* Replace uplevel vars with Param nodes */
 468         if (PlannerQueryLevel > 1)
 469                 expr = SS_replace_correlation_vars(expr);
 470
 471         /*
 472          * If it's a qual or havingQual, convert it to implicit-AND format.
 473          * (We don't want to do this before eval_const_expressions, since the
 474          * latter would be unable to simplify a top-level AND correctly. Also,
 475          * SS_process_sublinks expects explicit-AND format.)
 476          */
 477         if (kind == EXPRKIND_QUAL)
 478                 expr = (Node *) make_ands_implicit((Expr *) expr);
 479
 480         return expr;
 481 }
 482
 483 /*
 484  * preprocess_qual_conditions
 485  *              Recursively scan the query's jointree and do subquery_planner's
 486  *              preprocessing work on each qual condition found therein.
 487  */
 488 static void
 489 preprocess_qual_conditions(PlannerInfo *root, Node *jtnode)
 490 {
 491         if (jtnode == NULL)
 492                 return;
 493         if (IsA(jtnode, RangeTblRef))
 494         {
 495                 /* nothing to do here */
 496         }
 497         else if (IsA(jtnode, FromExpr))
 498         {
 499                 FromExpr   *f = (FromExpr *) jtnode;
 500                 ListCell   *l;
 501
 502                 foreach(l, f->fromlist)
 503                         preprocess_qual_conditions(root, lfirst(l));
 504
 505                 f->quals = preprocess_expression(root, f->quals, EXPRKIND_QUAL);
 506         }
 507         else if (IsA(jtnode, JoinExpr))
 508         {
 509                 JoinExpr   *j = (JoinExpr *) jtnode;
 510
 511                 preprocess_qual_conditions(root, j->larg);
 512                 preprocess_qual_conditions(root, j->rarg);
 513
 514                 j->quals = preprocess_expression(root, j->quals, EXPRKIND_QUAL);
 515         }
 516         else
 517                 elog(ERROR, "unrecognized node type: %d",
 518                          (int) nodeTag(jtnode));
 519 }
 520
 521 /*--------------------
 522  * inheritance_planner
 523  *        Generate a plan in the case where the result relation is an
 524  *        inheritance set.
 525  *
 526  * We have to handle this case differently from cases where a source
 527  * relation is an inheritance set.      Source inheritance is expanded at
 528  * the bottom of the plan tree (see allpaths.c), but target inheritance
 529  * has to be expanded at the top.  The reason is that for UPDATE, each
 530  * target relation needs a different targetlist matching its own column
 531  * set.  (This is not so critical for DELETE, but for simplicity we treat
 532  * inherited DELETE the same way.)      Fortunately, the UPDATE/DELETE target
 533  * can never be the nullable side of an outer join, so it's OK to generate
 534  * the plan this way.
 535  *
 536  * inheritlist is an integer list of RT indexes for the result relation set.
 537  *
 538  * Returns a query plan.
 539  *--------------------
 540  */
 541 static Plan *
 542 inheritance_planner(PlannerInfo *root, List *inheritlist)
 543 {
 544         Query      *parse = root->parse;
 545         int                     parentRTindex = parse->resultRelation;
 546         Oid                     parentOID = getrelid(parentRTindex, parse->rtable);
 547         int                     mainrtlength = list_length(parse->rtable);
 548         List       *subplans = NIL;
 549         List       *tlist = NIL;
 550         ListCell   *l;
 551
 552         foreach(l, inheritlist)
 553         {
 554                 int                     childRTindex = lfirst_int(l);
 555                 Oid                     childOID = getrelid(childRTindex, parse->rtable);
 556                 PlannerInfo subroot;
 557                 Plan       *subplan;
 558
 559                 /*
 560                  * Generate modified query with this rel as target.  We have to
 561                  * be prepared to translate varnos in in_info_list as well as in
 562                  * the Query proper.
 563                  */
 564                 memcpy(&subroot, root, sizeof(PlannerInfo));
 565                 subroot.parse = (Query *)
 566                         adjust_inherited_attrs((Node *) parse,
 567                                                                    parentRTindex, parentOID,
 568                                                                    childRTindex, childOID);
 569                 subroot.in_info_list = (List *)
 570                         adjust_inherited_attrs((Node *) root->in_info_list,
 571                                                                    parentRTindex, parentOID,
 572                                                                    childRTindex, childOID);
 573
 574                 /* Generate plan */
 575                 subplan = grouping_planner(&subroot, 0.0 /* retrieve all tuples */ );
 576
 577                 subplans = lappend(subplans, subplan);
 578
 579                 /*
 580                  * XXX my goodness this next bit is ugly.  Really need to think about
 581                  * ways to rein in planner's habit of scribbling on its input.
 582                  *
 583                  * Planning of the subquery might have modified the rangetable,
 584                  * either by addition of RTEs due to expansion of inherited source
 585                  * tables, or by changes of the Query structures inside subquery
 586                  * RTEs.  We have to ensure that this gets propagated back to the
 587                  * master copy.  However, if we aren't done planning yet, we also
 588                  * need to ensure that subsequent calls to grouping_planner have
 589                  * virgin sub-Queries to work from.  So, if we are at the last
 590                  * list entry, just copy the subquery rangetable back to the master
 591                  * copy; if we are not, then extend the master copy by adding
 592                  * whatever the subquery added.  (We assume these added entries
 593                  * will go untouched by the future grouping_planner calls.  We are
 594                  * also effectively assuming that sub-Queries will get planned
 595                  * identically each time, or at least that the impacts on their
 596                  * rangetables will be the same each time.  Did I say this is ugly?)
 597                  */
 598                 if (lnext(l) == NULL)
 599                         parse->rtable = subroot.parse->rtable;
 600                 else
 601                 {
 602                         int             subrtlength = list_length(subroot.parse->rtable);
 603
 604                         if (subrtlength > mainrtlength)
 605                         {
 606                                 List       *subrt;
 607
 608                                 subrt = list_copy_tail(subroot.parse->rtable, mainrtlength);
 609                                 parse->rtable = list_concat(parse->rtable, subrt);
 610                                 mainrtlength = subrtlength;
 611                         }
 612                 }
 613
 614                 /* Save preprocessed tlist from first rel for use in Append */
 615                 if (tlist == NIL)
 616                         tlist = subplan->targetlist;
 617         }
 618
 619         /* Save the target-relations list for the executor, too */
 620         parse->resultRelations = inheritlist;
 621
 622         /* Mark result as unordered (probably unnecessary) */
 623         root->query_pathkeys = NIL;
 624
 625         return (Plan *) make_append(subplans, true, tlist);
 626 }
 627
 628 /*--------------------
 629  * grouping_planner
 630  *        Perform planning steps related to grouping, aggregation, etc.
 631  *        This primarily means adding top-level processing to the basic
 632  *        query plan produced by query_planner.
 633  *
 634  * tuple_fraction is the fraction of tuples we expect will be retrieved
 635  *
 636  * tuple_fraction is interpreted as follows:
 637  *        0: expect all tuples to be retrieved (normal case)
 638  *        0 < tuple_fraction < 1: expect the given fraction of tuples available
 639  *              from the plan to be retrieved
 640  *        tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
 641  *              expected to be retrieved (ie, a LIMIT specification)
 642  *
 643  * Returns a query plan.  Also, root->query_pathkeys is returned as the
 644  * actual output ordering of the plan (in pathkey format).
 645  *--------------------
 646  */
 647 static Plan *
 648 grouping_planner(PlannerInfo *root, double tuple_fraction)
 649 {
 650         Query      *parse = root->parse;
 651         List       *tlist = parse->targetList;
 652         int                     offset_est = 0;
 653         int                     count_est = 0;
 654         Plan       *result_plan;
 655         List       *current_pathkeys;
 656         List       *sort_pathkeys;
 657         double          dNumGroups = 0;
 658
 659         /* Tweak caller-supplied tuple_fraction if have LIMIT/OFFSET */
 660         if (parse->limitCount || parse->limitOffset)
 661                 tuple_fraction = preprocess_limit(root, tuple_fraction,
 662                                                                                   &offset_est, &count_est);
 663
 664         if (parse->setOperations)
 665         {
 666                 List       *set_sortclauses;
 667
 668                 /*
 669                  * If there's a top-level ORDER BY, assume we have to fetch all
 670                  * the tuples.  This might seem too simplistic given all the
 671                  * hackery below to possibly avoid the sort ... but a nonzero
 672                  * tuple_fraction is only of use to plan_set_operations() when
 673                  * the setop is UNION ALL, and the result of UNION ALL is always
 674                  * unsorted.
 675                  */
 676                 if (parse->sortClause)
 677                         tuple_fraction = 0.0;
 678
 679                 /*
 680                  * Construct the plan for set operations.  The result will not
 681                  * need any work except perhaps a top-level sort and/or LIMIT.
 682                  */
 683                 result_plan = plan_set_operations(root, tuple_fraction,
 684                                                                                   &set_sortclauses);
 685
 686                 /*
 687                  * Calculate pathkeys representing the sort order (if any) of the
 688                  * set operation's result.  We have to do this before overwriting
 689                  * the sort key information...
 690                  */
 691                 current_pathkeys = make_pathkeys_for_sortclauses(set_sortclauses,
 692                                                                                                 result_plan->targetlist);
 693                 current_pathkeys = canonicalize_pathkeys(root, current_pathkeys);
 694
 695                 /*
 696                  * We should not need to call preprocess_targetlist, since we must
 697                  * be in a SELECT query node.  Instead, use the targetlist
 698                  * returned by plan_set_operations (since this tells whether it
 699                  * returned any resjunk columns!), and transfer any sort key
 700                  * information from the original tlist.
 701                  */
 702                 Assert(parse->commandType == CMD_SELECT);
 703
 704                 tlist = postprocess_setop_tlist(result_plan->targetlist, tlist);
 705
 706                 /*
 707                  * Can't handle FOR UPDATE/SHARE here (parser should have checked
 708                  * already, but let's make sure).
 709                  */
 710                 if (parse->rowMarks)
 711                         ereport(ERROR,
 712                                         (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
 713                                          errmsg("SELECT FOR UPDATE/SHARE is not allowed with UNION/INTERSECT/EXCEPT")));
 714
 715                 /*
 716                  * Calculate pathkeys that represent result ordering requirements
 717                  */
 718                 sort_pathkeys = make_pathkeys_for_sortclauses(parse->sortClause,
 719                                                                                                           tlist);
 720                 sort_pathkeys = canonicalize_pathkeys(root, sort_pathkeys);
 721         }
 722         else
 723         {
 724                 /* No set operations, do regular planning */
 725                 List       *sub_tlist;
 726                 List       *group_pathkeys;
 727                 AttrNumber *groupColIdx = NULL;
 728                 bool            need_tlist_eval = true;
 729                 QualCost        tlist_cost;
 730                 Path       *cheapest_path;
 731                 Path       *sorted_path;
 732                 Path       *best_path;
 733                 long            numGroups = 0;
 734                 AggClauseCounts agg_counts;
 735                 int                     numGroupCols = list_length(parse->groupClause);
 736                 bool            use_hashed_grouping = false;
 737
 738                 MemSet(&agg_counts, 0, sizeof(AggClauseCounts));
 739
 740                 /* Preprocess targetlist */
 741                 tlist = preprocess_targetlist(root, tlist);
 742
 743                 /*
 744                  * Generate appropriate target list for subplan; may be different
 745                  * from tlist if grouping or aggregation is needed.
 746                  */
 747                 sub_tlist = make_subplanTargetList(root, tlist,
 748                                                                                  &groupColIdx, &need_tlist_eval);
 749
 750                 /*
 751                  * Calculate pathkeys that represent grouping/ordering requirements.
 752                  * Stash them in PlannerInfo so that query_planner can canonicalize
 753                  * them.
 754                  */
 755                 root->group_pathkeys =
 756                         make_pathkeys_for_sortclauses(parse->groupClause, tlist);
 757                 root->sort_pathkeys =
 758                         make_pathkeys_for_sortclauses(parse->sortClause, tlist);
 759
 760                 /*
 761                  * Will need actual number of aggregates for estimating costs.
 762                  *
 763                  * Note: we do not attempt to detect duplicate aggregates here; a
 764                  * somewhat-overestimated count is okay for our present purposes.
 765                  *
 766                  * Note: think not that we can turn off hasAggs if we find no aggs.
 767                  * It is possible for constant-expression simplification to remove
 768                  * all explicit references to aggs, but we still have to follow
 769                  * the aggregate semantics (eg, producing only one output row).
 770                  */
 771                 if (parse->hasAggs)
 772                 {
 773                         count_agg_clauses((Node *) tlist, &agg_counts);
 774                         count_agg_clauses(parse->havingQual, &agg_counts);
 775                 }
 776
 777                 /*
 778                  * Figure out whether we need a sorted result from query_planner.
 779                  *
 780                  * If we have a GROUP BY clause, then we want a result sorted
 781                  * properly for grouping.  Otherwise, if there is an ORDER BY
 782                  * clause, we want to sort by the ORDER BY clause.      (Note: if we
 783                  * have both, and ORDER BY is a superset of GROUP BY, it would be
 784                  * tempting to request sort by ORDER BY --- but that might just
 785                  * leave us failing to exploit an available sort order at all.
 786                  * Needs more thought...)
 787                  */
 788                 if (parse->groupClause)
 789                         root->query_pathkeys = root->group_pathkeys;
 790                 else if (parse->sortClause)
 791                         root->query_pathkeys = root->sort_pathkeys;
 792                 else
 793                         root->query_pathkeys = NIL;
 794
 795                 /*
 796                  * Generate the best unsorted and presorted paths for this Query
 797                  * (but note there may not be any presorted path).  query_planner
 798                  * will also estimate the number of groups in the query, and
 799                  * canonicalize all the pathkeys.
 800                  */
 801                 query_planner(root, sub_tlist, tuple_fraction,
 802                                           &cheapest_path, &sorted_path, &dNumGroups);
 803
 804                 group_pathkeys = root->group_pathkeys;
 805                 sort_pathkeys = root->sort_pathkeys;
 806
 807                 /*
 808                  * If grouping, decide whether we want to use hashed grouping.
 809                  */
 810                 if (parse->groupClause)
 811                 {
 812                         use_hashed_grouping =
 813                                 choose_hashed_grouping(root, tuple_fraction,
 814                                                                            cheapest_path, sorted_path,
 815                                                                            dNumGroups, &agg_counts);
 816
 817                         /* Also convert # groups to long int --- but 'ware overflow! */
 818                         numGroups = (long) Min(dNumGroups, (double) LONG_MAX);
 819                 }
 820
 821                 /*
 822                  * Select the best path.  If we are doing hashed grouping, we will
 823                  * always read all the input tuples, so use the cheapest-total
 824                  * path. Otherwise, trust query_planner's decision about which to use.
 825                  */
 826                 if (use_hashed_grouping || !sorted_path)
 827                         best_path = cheapest_path;
 828                 else
 829                         best_path = sorted_path;
 830
 831                 /*
 832                  * Check to see if it's possible to optimize MIN/MAX aggregates.
 833                  * If so, we will forget all the work we did so far to choose a
 834                  * "regular" path ... but we had to do it anyway to be able to
 835                  * tell which way is cheaper.
 836                  */
 837                 result_plan = optimize_minmax_aggregates(root,
 838                                                                                                  tlist,
 839                                                                                                  best_path);
 840                 if (result_plan != NULL)
 841                 {
 842                         /*
 843                          * optimize_minmax_aggregates generated the full plan, with
 844                          * the right tlist, and it has no sort order.
 845                          */
 846                         current_pathkeys = NIL;
 847                 }
 848                 else
 849                 {
 850                         /*
 851                          * Normal case --- create a plan according to query_planner's
 852                          * results.
 853                          */
 854                         result_plan = create_plan(root, best_path);
 855                         current_pathkeys = best_path->pathkeys;
 856
 857                         /*
 858                          * create_plan() returns a plan with just a "flat" tlist of
 859                          * required Vars.  Usually we need to insert the sub_tlist as the
 860                          * tlist of the top plan node.  However, we can skip that if we
 861                          * determined that whatever query_planner chose to return will be
 862                          * good enough.
 863                          */
 864                         if (need_tlist_eval)
 865                         {
 866                                 /*
 867                                  * If the top-level plan node is one that cannot do expression
 868                                  * evaluation, we must insert a Result node to project the
 869                                  * desired tlist.
 870                                  */
 871                                 if (!is_projection_capable_plan(result_plan))
 872                                 {
 873                                         result_plan = (Plan *) make_result(sub_tlist, NULL,
 874                                                                                                            result_plan);
 875                                 }
 876                                 else
 877                                 {
 878                                         /*
 879                                          * Otherwise, just replace the subplan's flat tlist with
 880                                          * the desired tlist.
 881                                          */
 882                                         result_plan->targetlist = sub_tlist;
 883                                 }
 884
 885                                 /*
 886                                  * Also, account for the cost of evaluation of the sub_tlist.
 887                                  *
 888                                  * Up to now, we have only been dealing with "flat" tlists,
 889                                  * containing just Vars.  So their evaluation cost is zero
 890                                  * according to the model used by cost_qual_eval() (or if you
 891                                  * prefer, the cost is factored into cpu_tuple_cost).  Thus we
 892                                  * can avoid accounting for tlist cost throughout
 893                                  * query_planner() and subroutines.  But now we've inserted a
 894                                  * tlist that might contain actual operators, sub-selects, etc
 895                                  * --- so we'd better account for its cost.
 896                                  *
 897                                  * Below this point, any tlist eval cost for added-on nodes
 898                                  * should be accounted for as we create those nodes.
 899                                  * Presently, of the node types we can add on, only Agg and
 900                                  * Group project new tlists (the rest just copy their input
 901                                  * tuples) --- so make_agg() and make_group() are responsible
 902                                  * for computing the added cost.
 903                                  */
 904                                 cost_qual_eval(&tlist_cost, sub_tlist);
 905                                 result_plan->startup_cost += tlist_cost.startup;
 906                                 result_plan->total_cost += tlist_cost.startup +
 907                                         tlist_cost.per_tuple * result_plan->plan_rows;
 908                         }
 909                         else
 910                         {
 911                                 /*
 912                                  * Since we're using query_planner's tlist and not the one
 913                                  * make_subplanTargetList calculated, we have to refigure any
 914                                  * grouping-column indexes make_subplanTargetList computed.
 915                                  */
 916                                 locate_grouping_columns(root, tlist, result_plan->targetlist,
 917                                                                                 groupColIdx);
 918                         }
 919
 920                         /*
 921                          * Insert AGG or GROUP node if needed, plus an explicit sort step
 922                          * if necessary.
 923                          *
 924                          * HAVING clause, if any, becomes qual of the Agg or Group node.
 925                          */
 926                         if (use_hashed_grouping)
 927                         {
 928                                 /* Hashed aggregate plan --- no sort needed */
 929                                 result_plan = (Plan *) make_agg(root,
 930                                                                                                 tlist,
 931                                                                                                 (List *) parse->havingQual,
 932                                                                                                 AGG_HASHED,
 933                                                                                                 numGroupCols,
 934                                                                                                 groupColIdx,
 935                                                                                                 numGroups,
 936                                                                                                 agg_counts.numAggs,
 937                                                                                                 result_plan);
 938                                 /* Hashed aggregation produces randomly-ordered results */
 939                                 current_pathkeys = NIL;
 940                         }
 941                         else if (parse->hasAggs)
 942                         {
 943                                 /* Plain aggregate plan --- sort if needed */
 944                                 AggStrategy aggstrategy;
 945
 946                                 if (parse->groupClause)
 947                                 {
 948                                         if (!pathkeys_contained_in(group_pathkeys,
 949                                                                                            current_pathkeys))
 950                                         {
 951                                                 result_plan = (Plan *)
 952                                                         make_sort_from_groupcols(root,
 953                                                                                                          parse->groupClause,
 954                                                                                                          groupColIdx,
 955                                                                                                          result_plan);
 956                                                 current_pathkeys = group_pathkeys;
 957                                         }
 958                                         aggstrategy = AGG_SORTED;
 959
 960                                         /*
 961                                          * The AGG node will not change the sort ordering of its
 962                                          * groups, so current_pathkeys describes the result too.
 963                                          */
 964                                 }
 965                                 else
 966                                 {
 967                                         aggstrategy = AGG_PLAIN;
 968                                         /* Result will be only one row anyway; no sort order */
 969                                         current_pathkeys = NIL;
 970                                 }
 971
 972                                 result_plan = (Plan *) make_agg(root,
 973                                                                                                 tlist,
 974                                                                                                 (List *) parse->havingQual,
 975                                                                                                 aggstrategy,
 976                                                                                                 numGroupCols,
 977                                                                                                 groupColIdx,
 978                                                                                                 numGroups,
 979                                                                                                 agg_counts.numAggs,
 980                                                                                                 result_plan);
 981                         }
 982                         else if (parse->groupClause)
 983                         {
 984                                 /*
 985                                  * GROUP BY without aggregation, so insert a group node (plus
 986                                  * the appropriate sort node, if necessary).
 987                                  *
 988                                  * Add an explicit sort if we couldn't make the path come
 989                                  * out the way the GROUP node needs it.
 990                                  */
 991                                 if (!pathkeys_contained_in(group_pathkeys, current_pathkeys))
 992                                 {
 993                                         result_plan = (Plan *)
 994                                                 make_sort_from_groupcols(root,
 995                                                                                                  parse->groupClause,
 996                                                                                                  groupColIdx,
 997                                                                                                  result_plan);
 998                                         current_pathkeys = group_pathkeys;
 999                                 }
1000
1001                                 result_plan = (Plan *) make_group(root,
1002                                                                                                   tlist,
1003                                                                                                   (List *) parse->havingQual,
1004                                                                                                   numGroupCols,
1005                                                                                                   groupColIdx,
1006                                                                                                   dNumGroups,
1007                                                                                                   result_plan);
1008                                 /* The Group node won't change sort ordering */
1009                         }
1010                         else if (root->hasHavingQual)
1011                         {
1012                                 /*
1013                                  * No aggregates, and no GROUP BY, but we have a HAVING qual.
1014                                  * This is a degenerate case in which we are supposed to emit
1015                                  * either 0 or 1 row depending on whether HAVING succeeds.
1016                                  * Furthermore, there cannot be any variables in either HAVING
1017                                  * or the targetlist, so we actually do not need the FROM table
1018                                  * at all!  We can just throw away the plan-so-far and generate
1019                                  * a Result node.  This is a sufficiently unusual corner case
1020                                  * that it's not worth contorting the structure of this routine
1021                                  * to avoid having to generate the plan in the first place.
1022                                  */
1023                                 result_plan = (Plan *) make_result(tlist,
1024                                                                                                    parse->havingQual,
1025                                                                                                    NULL);
1026                         }
1027                 }                                               /* end of non-minmax-aggregate case */
1028         }                                                       /* end of if (setOperations) */
1029
1030         /*
1031          * If we were not able to make the plan come out in the right order,
1032          * add an explicit sort step.
1033          */
1034         if (parse->sortClause)
1035         {
1036                 if (!pathkeys_contained_in(sort_pathkeys, current_pathkeys))
1037                 {
1038                         result_plan = (Plan *)
1039                                 make_sort_from_sortclauses(root,
1040                                                                                    parse->sortClause,
1041                                                                                    result_plan);
1042                         current_pathkeys = sort_pathkeys;
1043                 }
1044         }
1045
1046         /*
1047          * If there is a DISTINCT clause, add the UNIQUE node.
1048          */
1049         if (parse->distinctClause)
1050         {
1051                 result_plan = (Plan *) make_unique(result_plan, parse->distinctClause);
1052
1053                 /*
1054                  * If there was grouping or aggregation, leave plan_rows as-is
1055                  * (ie, assume the result was already mostly unique).  If not,
1056                  * use the number of distinct-groups calculated by query_planner.
1057                  */
1058                 if (!parse->groupClause && !root->hasHavingQual && !parse->hasAggs)
1059                         result_plan->plan_rows = dNumGroups;
1060         }
1061
1062         /*
1063          * Finally, if there is a LIMIT/OFFSET clause, add the LIMIT node.
1064          */
1065         if (parse->limitCount || parse->limitOffset)
1066         {
1067                 result_plan = (Plan *) make_limit(result_plan,
1068                                                                                   parse->limitOffset,
1069                                                                                   parse->limitCount,
1070                                                                                   offset_est,
1071                                                                                   count_est);
1072         }
1073
1074         /*
1075          * Return the actual output ordering in query_pathkeys for possible
1076          * use by an outer query level.
1077          */
1078         root->query_pathkeys = current_pathkeys;
1079
1080         return result_plan;
1081 }
1082
1083 /*
1084  * preprocess_limit - do pre-estimation for LIMIT and/or OFFSET clauses
1085  *
1086  * We try to estimate the values of the LIMIT/OFFSET clauses, and pass the
1087  * results back in *count_est and *offset_est.  These variables are set to
1088  * 0 if the corresponding clause is not present, and -1 if it's present
1089  * but we couldn't estimate the value for it.  (The "0" convention is OK
1090  * for OFFSET but a little bit bogus for LIMIT: effectively we estimate
1091  * LIMIT 0 as though it were LIMIT 1.  But this is in line with the planner's
1092  * usual practice of never estimating less than one row.)  These values will
1093  * be passed to make_limit, which see if you change this code.
1094  *
1095  * The return value is the suitably adjusted tuple_fraction to use for
1096  * planning the query.  This adjustment is not overridable, since it reflects
1097  * plan actions that grouping_planner() will certainly take, not assumptions
1098  * about context.
1099  */
1100 static double
1101 preprocess_limit(PlannerInfo *root, double tuple_fraction,
1102                                  int *offset_est, int *count_est)
1103 {
1104         Query      *parse = root->parse;
1105         Node       *est;
1106         double          limit_fraction;
1107
1108         /* Should not be called unless LIMIT or OFFSET */
1109         Assert(parse->limitCount || parse->limitOffset);
1110
1111         /*
1112          * Try to obtain the clause values.  We use estimate_expression_value
1113          * primarily because it can sometimes do something useful with Params.
1114          */
1115         if (parse->limitCount)
1116         {
1117                 est = estimate_expression_value(parse->limitCount);
1118                 if (est && IsA(est, Const))
1119                 {
1120                         if (((Const *) est)->constisnull)
1121                         {
1122                                 /* NULL indicates LIMIT ALL, ie, no limit */
1123                                 *count_est = 0;                 /* treat as not present */
1124                         }
1125                         else
1126                         {
1127                                 *count_est = DatumGetInt32(((Const *) est)->constvalue);
1128                                 if (*count_est <= 0)
1129                                         *count_est = 1;         /* force to at least 1 */
1130                         }
1131                 }
1132                 else
1133                         *count_est = -1;        /* can't estimate */
1134         }
1135         else
1136                 *count_est = 0;                 /* not present */
1137
1138         if (parse->limitOffset)
1139         {
1140                 est = estimate_expression_value(parse->limitOffset);
1141                 if (est && IsA(est, Const))
1142                 {
1143                         if (((Const *) est)->constisnull)
1144                         {
1145                                 /* Treat NULL as no offset; the executor will too */
1146                                 *offset_est = 0;                /* treat as not present */
1147                         }
1148                         else
1149                         {
1150                                 *offset_est = DatumGetInt32(((Const *) est)->constvalue);
1151                                 if (*offset_est < 0)
1152                                         *offset_est = 0;        /* less than 0 is same as 0 */
1153                         }
1154                 }
1155                 else
1156                         *offset_est = -1;       /* can't estimate */
1157         }
1158         else
1159                 *offset_est = 0;                /* not present */
1160
1161         if (*count_est != 0)
1162         {
1163                 /*
1164                  * A LIMIT clause limits the absolute number of tuples returned.
1165                  * However, if it's not a constant LIMIT then we have to guess; for
1166                  * lack of a better idea, assume 10% of the plan's result is wanted.
1167                  */
1168                 if (*count_est < 0 || *offset_est < 0)
1169                 {
1170                         /* LIMIT or OFFSET is an expression ... punt ... */
1171                         limit_fraction = 0.10;
1172                 }
1173                 else
1174                 {
1175                         /* LIMIT (plus OFFSET, if any) is max number of tuples needed */
1176                         limit_fraction = (double) *count_est + (double) *offset_est;
1177                 }
1178
1179                 /*
1180                  * If we have absolute limits from both caller and LIMIT, use the
1181                  * smaller value; likewise if they are both fractional.  If one is
1182                  * fractional and the other absolute, we can't easily determine which
1183                  * is smaller, but we use the heuristic that the absolute will usually
1184                  * be smaller.
1185                  */
1186                 if (tuple_fraction >= 1.0)
1187                 {
1188                         if (limit_fraction >= 1.0)
1189                         {
1190                                 /* both absolute */
1191                                 tuple_fraction = Min(tuple_fraction, limit_fraction);
1192                         }
1193                         else
1194                         {
1195                                 /* caller absolute, limit fractional; use caller's value */
1196                         }
1197                 }
1198                 else if (tuple_fraction > 0.0)
1199                 {
1200                         if (limit_fraction >= 1.0)
1201                         {
1202                                 /* caller fractional, limit absolute; use limit */
1203                                 tuple_fraction = limit_fraction;
1204                         }
1205                         else
1206                         {
1207                                 /* both fractional */
1208                                 tuple_fraction = Min(tuple_fraction, limit_fraction);
1209                         }
1210                 }
1211                 else
1212                 {
1213                         /* no info from caller, just use limit */
1214                         tuple_fraction = limit_fraction;
1215                 }
1216         }
1217         else if (*offset_est != 0 && tuple_fraction > 0.0)
1218         {
1219                 /*
1220                  * We have an OFFSET but no LIMIT.  This acts entirely differently
1221                  * from the LIMIT case: here, we need to increase rather than
1222                  * decrease the caller's tuple_fraction, because the OFFSET acts
1223                  * to cause more tuples to be fetched instead of fewer.  This only
1224                  * matters if we got a tuple_fraction > 0, however.
1225                  *
1226                  * As above, use 10% if OFFSET is present but unestimatable.
1227                  */
1228                 if (*offset_est < 0)
1229                         limit_fraction = 0.10;
1230                 else
1231                         limit_fraction = (double) *offset_est;
1232
1233                 /*
1234                  * If we have absolute counts from both caller and OFFSET, add them
1235                  * together; likewise if they are both fractional.  If one is
1236                  * fractional and the other absolute, we want to take the larger,
1237                  * and we heuristically assume that's the fractional one.
1238                  */
1239                 if (tuple_fraction >= 1.0)
1240                 {
1241                         if (limit_fraction >= 1.0)
1242                         {
1243                                 /* both absolute, so add them together */
1244                                 tuple_fraction += limit_fraction;
1245                         }
1246                         else
1247                         {
1248                                 /* caller absolute, limit fractional; use limit */
1249                                 tuple_fraction = limit_fraction;
1250                         }
1251                 }
1252                 else
1253                 {
1254                         if (limit_fraction >= 1.0)
1255                         {
1256                                 /* caller fractional, limit absolute; use caller's value */
1257                         }
1258                         else
1259                         {
1260                                 /* both fractional, so add them together */
1261                                 tuple_fraction += limit_fraction;
1262                                 if (tuple_fraction >= 1.0)
1263                                         tuple_fraction = 0.0; /* assume fetch all */
1264                         }
1265                 }
1266         }
1267
1268         return tuple_fraction;
1269 }
1270
1271 /*
1272  * choose_hashed_grouping - should we use hashed grouping?
1273  */
1274 static bool
1275 choose_hashed_grouping(PlannerInfo *root, double tuple_fraction,
1276                                            Path *cheapest_path, Path *sorted_path,
1277                                            double dNumGroups, AggClauseCounts *agg_counts)
1278 {
1279         int                     numGroupCols = list_length(root->parse->groupClause);
1280         double          cheapest_path_rows;
1281         int                     cheapest_path_width;
1282         Size            hashentrysize;
1283         List       *current_pathkeys;
1284         Path            hashed_p;
1285         Path            sorted_p;
1286
1287         /*
1288          * Check can't-do-it conditions, including whether the grouping operators
1289          * are hashjoinable.
1290          *
1291          * Executor doesn't support hashed aggregation with DISTINCT aggregates.
1292          * (Doing so would imply storing *all* the input values in the hash table,
1293          * which seems like a certain loser.)
1294          */
1295         if (!enable_hashagg)
1296                 return false;
1297         if (agg_counts->numDistinctAggs != 0)
1298                 return false;
1299         if (!hash_safe_grouping(root))
1300                 return false;
1301
1302         /*
1303          * Don't do it if it doesn't look like the hashtable will fit into
1304          * work_mem.
1305          *
1306          * Beware here of the possibility that cheapest_path->parent is NULL.
1307          * This could happen if user does something silly like
1308          *              SELECT 'foo' GROUP BY 1;
1309          */
1310         if (cheapest_path->parent)
1311         {
1312                 cheapest_path_rows = cheapest_path->parent->rows;
1313                 cheapest_path_width = cheapest_path->parent->width;
1314         }
1315         else
1316         {
1317                 cheapest_path_rows = 1;                         /* assume non-set result */
1318                 cheapest_path_width = 100;                      /* arbitrary */
1319         }
1320
1321         /* Estimate per-hash-entry space at tuple width... */
1322         hashentrysize = cheapest_path_width;
1323         /* plus space for pass-by-ref transition values... */
1324         hashentrysize += agg_counts->transitionSpace;
1325         /* plus the per-hash-entry overhead */
1326         hashentrysize += hash_agg_entry_size(agg_counts->numAggs);
1327
1328         if (hashentrysize * dNumGroups > work_mem * 1024L)
1329                 return false;
1330
1331         /*
1332          * See if the estimated cost is no more than doing it the other way.
1333          * While avoiding the need for sorted input is usually a win, the fact
1334          * that the output won't be sorted may be a loss; so we need to do an
1335          * actual cost comparison.
1336          *
1337          * We need to consider
1338          *              cheapest_path + hashagg [+ final sort]
1339          * versus either
1340          *              cheapest_path [+ sort] + group or agg [+ final sort]
1341          * or
1342          *              presorted_path + group or agg [+ final sort]
1343          * where brackets indicate a step that may not be needed. We assume
1344          * query_planner() will have returned a presorted path only if it's a
1345          * winner compared to cheapest_path for this purpose.
1346          *
1347          * These path variables are dummies that just hold cost fields; we don't
1348          * make actual Paths for these steps.
1349          */
1350         cost_agg(&hashed_p, root, AGG_HASHED, agg_counts->numAggs,
1351                          numGroupCols, dNumGroups,
1352                          cheapest_path->startup_cost, cheapest_path->total_cost,
1353                          cheapest_path_rows);
1354         /* Result of hashed agg is always unsorted */
1355         if (root->sort_pathkeys)
1356                 cost_sort(&hashed_p, root, root->sort_pathkeys, hashed_p.total_cost,
1357                                   dNumGroups, cheapest_path_width);
1358
1359         if (sorted_path)
1360         {
1361                 sorted_p.startup_cost = sorted_path->startup_cost;
1362                 sorted_p.total_cost = sorted_path->total_cost;
1363                 current_pathkeys = sorted_path->pathkeys;
1364         }
1365         else
1366         {
1367                 sorted_p.startup_cost = cheapest_path->startup_cost;
1368                 sorted_p.total_cost = cheapest_path->total_cost;
1369                 current_pathkeys = cheapest_path->pathkeys;
1370         }
1371         if (!pathkeys_contained_in(root->group_pathkeys, current_pathkeys))
1372         {
1373                 cost_sort(&sorted_p, root, root->group_pathkeys, sorted_p.total_cost,
1374                                   cheapest_path_rows, cheapest_path_width);
1375                 current_pathkeys = root->group_pathkeys;
1376         }
1377
1378         if (root->parse->hasAggs)
1379                 cost_agg(&sorted_p, root, AGG_SORTED, agg_counts->numAggs,
1380                                  numGroupCols, dNumGroups,
1381                                  sorted_p.startup_cost, sorted_p.total_cost,
1382                                  cheapest_path_rows);
1383         else
1384                 cost_group(&sorted_p, root, numGroupCols, dNumGroups,
1385                                    sorted_p.startup_cost, sorted_p.total_cost,
1386                                    cheapest_path_rows);
1387         /* The Agg or Group node will preserve ordering */
1388         if (root->sort_pathkeys &&
1389                 !pathkeys_contained_in(root->sort_pathkeys, current_pathkeys))
1390                 cost_sort(&sorted_p, root, root->sort_pathkeys, sorted_p.total_cost,
1391                                   dNumGroups, cheapest_path_width);
1392
1393         /*
1394          * Now make the decision using the top-level tuple fraction.  First we
1395          * have to convert an absolute count (LIMIT) into fractional form.
1396          */
1397         if (tuple_fraction >= 1.0)
1398                 tuple_fraction /= dNumGroups;
1399
1400         if (compare_fractional_path_costs(&hashed_p, &sorted_p,
1401                                                                           tuple_fraction) < 0)
1402         {
1403                 /* Hashed is cheaper, so use it */
1404                 return true;
1405         }
1406         return false;
1407 }
1408
1409 /*
1410  * hash_safe_grouping - are grouping operators hashable?
1411  *
1412  * We assume hashed aggregation will work if the datatype's equality operator
1413  * is marked hashjoinable.
1414  */
1415 static bool
1416 hash_safe_grouping(PlannerInfo *root)
1417 {
1418         ListCell   *gl;
1419
1420         foreach(gl, root->parse->groupClause)
1421         {
1422                 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1423                 TargetEntry *tle = get_sortgroupclause_tle(grpcl,
1424                                                                                                    root->parse->targetList);
1425                 Operator        optup;
1426                 bool            oprcanhash;
1427
1428                 optup = equality_oper(exprType((Node *) tle->expr), true);
1429                 if (!optup)
1430                         return false;
1431                 oprcanhash = ((Form_pg_operator) GETSTRUCT(optup))->oprcanhash;
1432                 ReleaseSysCache(optup);
1433                 if (!oprcanhash)
1434                         return false;
1435         }
1436         return true;
1437 }
1438
1439 /*---------------
1440  * make_subplanTargetList
1441  *        Generate appropriate target list when grouping is required.
1442  *
1443  * When grouping_planner inserts Aggregate, Group, or Result plan nodes
1444  * above the result of query_planner, we typically want to pass a different
1445  * target list to query_planner than the outer plan nodes should have.
1446  * This routine generates the correct target list for the subplan.
1447  *
1448  * The initial target list passed from the parser already contains entries
1449  * for all ORDER BY and GROUP BY expressions, but it will not have entries
1450  * for variables used only in HAVING clauses; so we need to add those
1451  * variables to the subplan target list.  Also, we flatten all expressions
1452  * except GROUP BY items into their component variables; the other expressions
1453  * will be computed by the inserted nodes rather than by the subplan.
1454  * For example, given a query like
1455  *              SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
1456  * we want to pass this targetlist to the subplan:
1457  *              a,b,c,d,a+b
1458  * where the a+b target will be used by the Sort/Group steps, and the
1459  * other targets will be used for computing the final results.  (In the
1460  * above example we could theoretically suppress the a and b targets and
1461  * pass down only c,d,a+b, but it's not really worth the trouble to
1462  * eliminate simple var references from the subplan.  We will avoid doing
1463  * the extra computation to recompute a+b at the outer level; see
1464  * replace_vars_with_subplan_refs() in setrefs.c.)
1465  *
1466  * If we are grouping or aggregating, *and* there are no non-Var grouping
1467  * expressions, then the returned tlist is effectively dummy; we do not
1468  * need to force it to be evaluated, because all the Vars it contains
1469  * should be present in the output of query_planner anyway.
1470  *
1471  * 'tlist' is the query's target list.
1472  * 'groupColIdx' receives an array of column numbers for the GROUP BY
1473  *                      expressions (if there are any) in the subplan's target list.
1474  * 'need_tlist_eval' is set true if we really need to evaluate the
1475  *                      result tlist.
1476  *
1477  * The result is the targetlist to be passed to the subplan.
1478  *---------------
1479  */
1480 static List *
1481 make_subplanTargetList(PlannerInfo *root,
1482                                            List *tlist,
1483                                            AttrNumber **groupColIdx,
1484                                            bool *need_tlist_eval)
1485 {
1486         Query      *parse = root->parse;
1487         List       *sub_tlist;
1488         List       *extravars;
1489         int                     numCols;
1490
1491         *groupColIdx = NULL;
1492
1493         /*
1494          * If we're not grouping or aggregating, there's nothing to do here;
1495          * query_planner should receive the unmodified target list.
1496          */
1497         if (!parse->hasAggs && !parse->groupClause && !root->hasHavingQual)
1498         {
1499                 *need_tlist_eval = true;
1500                 return tlist;
1501         }
1502
1503         /*
1504          * Otherwise, start with a "flattened" tlist (having just the vars
1505          * mentioned in the targetlist and HAVING qual --- but not upper-
1506          * level Vars; they will be replaced by Params later on).
1507          */
1508         sub_tlist = flatten_tlist(tlist);
1509         extravars = pull_var_clause(parse->havingQual, false);
1510         sub_tlist = add_to_flat_tlist(sub_tlist, extravars);
1511         list_free(extravars);
1512         *need_tlist_eval = false;       /* only eval if not flat tlist */
1513
1514         /*
1515          * If grouping, create sub_tlist entries for all GROUP BY expressions
1516          * (GROUP BY items that are simple Vars should be in the list
1517          * already), and make an array showing where the group columns are in
1518          * the sub_tlist.
1519          */
1520         numCols = list_length(parse->groupClause);
1521         if (numCols > 0)
1522         {
1523                 int                     keyno = 0;
1524                 AttrNumber *grpColIdx;
1525                 ListCell   *gl;
1526
1527                 grpColIdx = (AttrNumber *) palloc(sizeof(AttrNumber) * numCols);
1528                 *groupColIdx = grpColIdx;
1529
1530                 foreach(gl, parse->groupClause)
1531                 {
1532                         GroupClause *grpcl = (GroupClause *) lfirst(gl);
1533                         Node       *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1534                         TargetEntry *te = NULL;
1535                         ListCell   *sl;
1536
1537                         /* Find or make a matching sub_tlist entry */
1538                         foreach(sl, sub_tlist)
1539                         {
1540                                 te = (TargetEntry *) lfirst(sl);
1541                                 if (equal(groupexpr, te->expr))
1542                                         break;
1543                         }
1544                         if (!sl)
1545                         {
1546                                 te = makeTargetEntry((Expr *) groupexpr,
1547                                                                          list_length(sub_tlist) + 1,
1548                                                                          NULL,
1549                                                                          false);
1550                                 sub_tlist = lappend(sub_tlist, te);
1551                                 *need_tlist_eval = true;                /* it's not flat anymore */
1552                         }
1553
1554                         /* and save its resno */
1555                         grpColIdx[keyno++] = te->resno;
1556                 }
1557         }
1558
1559         return sub_tlist;
1560 }
1561
1562 /*
1563  * locate_grouping_columns
1564  *              Locate grouping columns in the tlist chosen by query_planner.
1565  *
1566  * This is only needed if we don't use the sub_tlist chosen by
1567  * make_subplanTargetList.      We have to forget the column indexes found
1568  * by that routine and re-locate the grouping vars in the real sub_tlist.
1569  */
1570 static void
1571 locate_grouping_columns(PlannerInfo *root,
1572                                                 List *tlist,
1573                                                 List *sub_tlist,
1574                                                 AttrNumber *groupColIdx)
1575 {
1576         int                     keyno = 0;
1577         ListCell   *gl;
1578
1579         /*
1580          * No work unless grouping.
1581          */
1582         if (!root->parse->groupClause)
1583         {
1584                 Assert(groupColIdx == NULL);
1585                 return;
1586         }
1587         Assert(groupColIdx != NULL);
1588
1589         foreach(gl, root->parse->groupClause)
1590         {
1591                 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1592                 Node       *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1593                 TargetEntry *te = NULL;
1594                 ListCell   *sl;
1595
1596                 foreach(sl, sub_tlist)
1597                 {
1598                         te = (TargetEntry *) lfirst(sl);
1599                         if (equal(groupexpr, te->expr))
1600                                 break;
1601                 }
1602                 if (!sl)
1603                         elog(ERROR, "failed to locate grouping columns");
1604
1605                 groupColIdx[keyno++] = te->resno;
1606         }
1607 }
1608
1609 /*
1610  * postprocess_setop_tlist
1611  *        Fix up targetlist returned by plan_set_operations().
1612  *
1613  * We need to transpose sort key info from the orig_tlist into new_tlist.
1614  * NOTE: this would not be good enough if we supported resjunk sort keys
1615  * for results of set operations --- then, we'd need to project a whole
1616  * new tlist to evaluate the resjunk columns.  For now, just ereport if we
1617  * find any resjunk columns in orig_tlist.
1618  */
1619 static List *
1620 postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
1621 {
1622         ListCell   *l;
1623         ListCell   *orig_tlist_item = list_head(orig_tlist);
1624
1625         foreach(l, new_tlist)
1626         {
1627                 TargetEntry *new_tle = (TargetEntry *) lfirst(l);
1628                 TargetEntry *orig_tle;
1629
1630                 /* ignore resjunk columns in setop result */
1631                 if (new_tle->resjunk)
1632                         continue;
1633
1634                 Assert(orig_tlist_item != NULL);
1635                 orig_tle = (TargetEntry *) lfirst(orig_tlist_item);
1636                 orig_tlist_item = lnext(orig_tlist_item);
1637                 if (orig_tle->resjunk)                  /* should not happen */
1638                         elog(ERROR, "resjunk output columns are not implemented");
1639                 Assert(new_tle->resno == orig_tle->resno);
1640                 new_tle->ressortgroupref = orig_tle->ressortgroupref;
1641         }
1642         if (orig_tlist_item != NULL)
1643                 elog(ERROR, "resjunk output columns are not implemented");
1644         return new_tlist;
1645 }