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) 1996-2008, 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 1.231 2008/04/01 00:48:33 tgl Exp $
  12  *
  13  *-------------------------------------------------------------------------
  14  */
  15
  16 #include "postgres.h"
  17
  18 #include <limits.h>
  19
  20 #include "catalog/pg_operator.h"
  21 #include "executor/executor.h"
  22 #include "executor/nodeAgg.h"
  23 #include "miscadmin.h"
  24 #include "nodes/makefuncs.h"
  25 #include "optimizer/clauses.h"
  26 #include "optimizer/cost.h"
  27 #include "optimizer/pathnode.h"
  28 #include "optimizer/paths.h"
  29 #include "optimizer/planmain.h"
  30 #include "optimizer/planner.h"
  31 #include "optimizer/prep.h"
  32 #include "optimizer/subselect.h"
  33 #include "optimizer/tlist.h"
  34 #include "optimizer/var.h"
  35 #ifdef OPTIMIZER_DEBUG
  36 #include "nodes/print.h"
  37 #endif
  38 #include "parser/parse_expr.h"
  39 #include "parser/parse_oper.h"
  40 #include "parser/parsetree.h"
  41 #include "utils/lsyscache.h"
  42 #include "utils/syscache.h"
  43
  44
  45 /* Hook for plugins to get control in planner() */
  46 planner_hook_type planner_hook = NULL;
  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_VALUES         3
  54 #define EXPRKIND_LIMIT          4
  55 #define EXPRKIND_ININFO         5
  56 #define EXPRKIND_APPINFO        6
  57
  58
  59 static Node *preprocess_expression(PlannerInfo *root, Node *expr, int kind);
  60 static void preprocess_qual_conditions(PlannerInfo *root, Node *jtnode);
  61 static Plan *inheritance_planner(PlannerInfo *root);
  62 static Plan *grouping_planner(PlannerInfo *root, double tuple_fraction);
  63 static bool is_dummy_plan(Plan *plan);
  64 static double preprocess_limit(PlannerInfo *root,
  65                                  double tuple_fraction,
  66                                  int64 *offset_est, int64 *count_est);
  67 static Oid *extract_grouping_ops(List *groupClause);
  68 static bool choose_hashed_grouping(PlannerInfo *root,
  69                                            double tuple_fraction, double limit_tuples,
  70                                            Path *cheapest_path, Path *sorted_path,
  71                                            Oid *groupOperators, double dNumGroups,
  72                                            AggClauseCounts *agg_counts);
  73 static List *make_subplanTargetList(PlannerInfo *root, List *tlist,
  74                                            AttrNumber **groupColIdx, bool *need_tlist_eval);
  75 static void locate_grouping_columns(PlannerInfo *root,
  76                                                 List *tlist,
  77                                                 List *sub_tlist,
  78                                                 AttrNumber *groupColIdx);
  79 static List *postprocess_setop_tlist(List *new_tlist, List *orig_tlist);
  80
  81
  82 /*****************************************************************************
  83  *
  84  *         Query optimizer entry point
  85  *
  86  * To support loadable plugins that monitor or modify planner behavior,
  87  * we provide a hook variable that lets a plugin get control before and
  88  * after the standard planning process.  The plugin would normally call
  89  * standard_planner().
  90  *
  91  * Note to plugin authors: standard_planner() scribbles on its Query input,
  92  * so you'd better copy that data structure if you want to plan more than once.
  93  *
  94  *****************************************************************************/
  95 PlannedStmt *
  96 planner(Query *parse, int cursorOptions, ParamListInfo boundParams)
  97 {
  98         PlannedStmt *result;
  99
 100         if (planner_hook)
 101                 result = (*planner_hook) (parse, cursorOptions, boundParams);
 102         else
 103                 result = standard_planner(parse, cursorOptions, boundParams);
 104         return result;
 105 }
 106
 107 PlannedStmt *
 108 standard_planner(Query *parse, int cursorOptions, ParamListInfo boundParams)
 109 {
 110         PlannedStmt *result;
 111         PlannerGlobal *glob;
 112         double          tuple_fraction;
 113         PlannerInfo *root;
 114         Plan       *top_plan;
 115         ListCell   *lp,
 116                            *lr;
 117
 118         /* Cursor options may come from caller or from DECLARE CURSOR stmt */
 119         if (parse->utilityStmt &&
 120                 IsA(parse->utilityStmt, DeclareCursorStmt))
 121                 cursorOptions |= ((DeclareCursorStmt *) parse->utilityStmt)->options;
 122
 123         /*
 124          * Set up global state for this planner invocation.  This data is needed
 125          * across all levels of sub-Query that might exist in the given command,
 126          * so we keep it in a separate struct that's linked to by each per-Query
 127          * PlannerInfo.
 128          */
 129         glob = makeNode(PlannerGlobal);
 130
 131         glob->boundParams = boundParams;
 132         glob->paramlist = NIL;
 133         glob->subplans = NIL;
 134         glob->subrtables = NIL;
 135         glob->rewindPlanIDs = NULL;
 136         glob->finalrtable = NIL;
 137         glob->relationOids = NIL;
 138         glob->transientPlan = false;
 139
 140         /* Determine what fraction of the plan is likely to be scanned */
 141         if (cursorOptions & CURSOR_OPT_FAST_PLAN)
 142         {
 143                 /*
 144                  * We have no real idea how many tuples the user will ultimately FETCH
 145                  * from a cursor, but it seems a good bet that he doesn't want 'em
 146                  * all.  Optimize for 10% retrieval (you gotta better number?  Should
 147                  * this be a SETtable parameter?)
 148                  */
 149                 tuple_fraction = 0.10;
 150         }
 151         else
 152         {
 153                 /* Default assumption is we need all the tuples */
 154                 tuple_fraction = 0.0;
 155         }
 156
 157         /* primary planning entry point (may recurse for subqueries) */
 158         top_plan = subquery_planner(glob, parse, 1, tuple_fraction, &root);
 159
 160         /*
 161          * If creating a plan for a scrollable cursor, make sure it can run
 162          * backwards on demand.  Add a Material node at the top at need.
 163          */
 164         if (cursorOptions & CURSOR_OPT_SCROLL)
 165         {
 166                 if (!ExecSupportsBackwardScan(top_plan))
 167                         top_plan = materialize_finished_plan(top_plan);
 168         }
 169
 170         /* final cleanup of the plan */
 171         Assert(glob->finalrtable == NIL);
 172         top_plan = set_plan_references(glob, top_plan, root->parse->rtable);
 173         /* ... and the subplans (both regular subplans and initplans) */
 174         Assert(list_length(glob->subplans) == list_length(glob->subrtables));
 175         forboth(lp, glob->subplans, lr, glob->subrtables)
 176         {
 177                 Plan       *subplan = (Plan *) lfirst(lp);
 178                 List       *subrtable = (List *) lfirst(lr);
 179
 180                 lfirst(lp) = set_plan_references(glob, subplan, subrtable);
 181         }
 182
 183         /* build the PlannedStmt result */
 184         result = makeNode(PlannedStmt);
 185
 186         result->commandType = parse->commandType;
 187         result->canSetTag = parse->canSetTag;
 188         result->transientPlan = glob->transientPlan;
 189         result->planTree = top_plan;
 190         result->rtable = glob->finalrtable;
 191         result->resultRelations = root->resultRelations;
 192         result->utilityStmt = parse->utilityStmt;
 193         result->intoClause = parse->intoClause;
 194         result->subplans = glob->subplans;
 195         result->rewindPlanIDs = glob->rewindPlanIDs;
 196         result->returningLists = root->returningLists;
 197         result->rowMarks = parse->rowMarks;
 198         result->relationOids = glob->relationOids;
 199         result->nParamExec = list_length(glob->paramlist);
 200
 201         return result;
 202 }
 203
 204
 205 /*--------------------
 206  * subquery_planner
 207  *        Invokes the planner on a subquery.  We recurse to here for each
 208  *        sub-SELECT found in the query tree.
 209  *
 210  * glob is the global state for the current planner run.
 211  * parse is the querytree produced by the parser & rewriter.
 212  * level is the current recursion depth (1 at the top-level Query).
 213  * tuple_fraction is the fraction of tuples we expect will be retrieved.
 214  * tuple_fraction is interpreted as explained for grouping_planner, below.
 215  *
 216  * If subroot isn't NULL, we pass back the query's final PlannerInfo struct;
 217  * among other things this tells the output sort ordering of the plan.
 218  *
 219  * Basically, this routine does the stuff that should only be done once
 220  * per Query object.  It then calls grouping_planner.  At one time,
 221  * grouping_planner could be invoked recursively on the same Query object;
 222  * that's not currently true, but we keep the separation between the two
 223  * routines anyway, in case we need it again someday.
 224  *
 225  * subquery_planner will be called recursively to handle sub-Query nodes
 226  * found within the query's expressions and rangetable.
 227  *
 228  * Returns a query plan.
 229  *--------------------
 230  */
 231 Plan *
 232 subquery_planner(PlannerGlobal *glob, Query *parse,
 233                                  Index level, double tuple_fraction,
 234                                  PlannerInfo **subroot)
 235 {
 236         int                     num_old_subplans = list_length(glob->subplans);
 237         PlannerInfo *root;
 238         Plan       *plan;
 239         List       *newHaving;
 240         ListCell   *l;
 241
 242         /* Create a PlannerInfo data structure for this subquery */
 243         root = makeNode(PlannerInfo);
 244         root->parse = parse;
 245         root->glob = glob;
 246         root->query_level = level;
 247         root->planner_cxt = CurrentMemoryContext;
 248         root->init_plans = NIL;
 249         root->eq_classes = NIL;
 250         root->in_info_list = NIL;
 251         root->append_rel_list = NIL;
 252
 253         /*
 254          * Look for IN clauses at the top level of WHERE, and transform them into
 255          * joins.  Note that this step only handles IN clauses originally at top
 256          * level of WHERE; if we pull up any subqueries below, their INs are
 257          * processed just before pulling them up.
 258          */
 259         if (parse->hasSubLinks)
 260                 parse->jointree->quals = pull_up_IN_clauses(root,
 261                                                                                                         parse->jointree->quals);
 262
 263         /*
 264          * Scan the rangetable for set-returning functions, and inline them
 265          * if possible (producing subqueries that might get pulled up next).
 266          * Recursion issues here are handled in the same way as for IN clauses.
 267          */
 268         inline_set_returning_functions(root);
 269
 270         /*
 271          * Check to see if any subqueries in the rangetable can be merged into
 272          * this query.
 273          */
 274         parse->jointree = (FromExpr *)
 275                 pull_up_subqueries(root, (Node *) parse->jointree, false, false);
 276
 277         /*
 278          * Detect whether any rangetable entries are RTE_JOIN kind; if not, we can
 279          * avoid the expense of doing flatten_join_alias_vars().  Also check for
 280          * outer joins --- if none, we can skip reduce_outer_joins() and some
 281          * other processing.  This must be done after we have done
 282          * pull_up_subqueries, of course.
 283          *
 284          * Note: if reduce_outer_joins manages to eliminate all outer joins,
 285          * root->hasOuterJoins is not reset currently.  This is OK since its
 286          * purpose is merely to suppress unnecessary processing in simple cases.
 287          */
 288         root->hasJoinRTEs = false;
 289         root->hasOuterJoins = false;
 290         foreach(l, parse->rtable)
 291         {
 292                 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
 293
 294                 if (rte->rtekind == RTE_JOIN)
 295                 {
 296                         root->hasJoinRTEs = true;
 297                         if (IS_OUTER_JOIN(rte->jointype))
 298                         {
 299                                 root->hasOuterJoins = true;
 300                                 /* Can quit scanning once we find an outer join */
 301                                 break;
 302                         }
 303                 }
 304         }
 305
 306         /*
 307          * Expand any rangetable entries that are inheritance sets into "append
 308          * relations".  This can add entries to the rangetable, but they must be
 309          * plain base relations not joins, so it's OK (and marginally more
 310          * efficient) to do it after checking for join RTEs.  We must do it after
 311          * pulling up subqueries, else we'd fail to handle inherited tables in
 312          * subqueries.
 313          */
 314         expand_inherited_tables(root);
 315
 316         /*
 317          * Set hasHavingQual to remember if HAVING clause is present.  Needed
 318          * because preprocess_expression will reduce a constant-true condition to
 319          * an empty qual list ... but "HAVING TRUE" is not a semantic no-op.
 320          */
 321         root->hasHavingQual = (parse->havingQual != NULL);
 322
 323         /* Clear this flag; might get set in distribute_qual_to_rels */
 324         root->hasPseudoConstantQuals = false;
 325
 326         /*
 327          * Do expression preprocessing on targetlist and quals.
 328          */
 329         parse->targetList = (List *)
 330                 preprocess_expression(root, (Node *) parse->targetList,
 331                                                           EXPRKIND_TARGET);
 332
 333         parse->returningList = (List *)
 334                 preprocess_expression(root, (Node *) parse->returningList,
 335                                                           EXPRKIND_TARGET);
 336
 337         preprocess_qual_conditions(root, (Node *) parse->jointree);
 338
 339         parse->havingQual = preprocess_expression(root, parse->havingQual,
 340                                                                                           EXPRKIND_QUAL);
 341
 342         parse->limitOffset = preprocess_expression(root, parse->limitOffset,
 343                                                                                            EXPRKIND_LIMIT);
 344         parse->limitCount = preprocess_expression(root, parse->limitCount,
 345                                                                                           EXPRKIND_LIMIT);
 346
 347         root->in_info_list = (List *)
 348                 preprocess_expression(root, (Node *) root->in_info_list,
 349                                                           EXPRKIND_ININFO);
 350         root->append_rel_list = (List *)
 351                 preprocess_expression(root, (Node *) root->append_rel_list,
 352                                                           EXPRKIND_APPINFO);
 353
 354         /* Also need to preprocess expressions for function and values RTEs */
 355         foreach(l, parse->rtable)
 356         {
 357                 RangeTblEntry *rte = (RangeTblEntry *) lfirst(l);
 358
 359                 if (rte->rtekind == RTE_FUNCTION)
 360                         rte->funcexpr = preprocess_expression(root, rte->funcexpr,
 361                                                                                                   EXPRKIND_RTFUNC);
 362                 else if (rte->rtekind == RTE_VALUES)
 363                         rte->values_lists = (List *)
 364                                 preprocess_expression(root, (Node *) rte->values_lists,
 365                                                                           EXPRKIND_VALUES);
 366         }
 367
 368         /*
 369          * In some cases we may want to transfer a HAVING clause into WHERE. We
 370          * cannot do so if the HAVING clause contains aggregates (obviously) or
 371          * volatile functions (since a HAVING clause is supposed to be executed
 372          * only once per group).  Also, it may be that the clause is so expensive
 373          * to execute that we're better off doing it only once per group, despite
 374          * the loss of selectivity.  This is hard to estimate short of doing the
 375          * entire planning process twice, so we use a heuristic: clauses
 376          * containing subplans are left in HAVING.      Otherwise, we move or copy the
 377          * HAVING clause into WHERE, in hopes of eliminating tuples before
 378          * aggregation instead of after.
 379          *
 380          * If the query has explicit grouping then we can simply move such a
 381          * clause into WHERE; any group that fails the clause will not be in the
 382          * output because none of its tuples will reach the grouping or
 383          * aggregation stage.  Otherwise we must have a degenerate (variable-free)
 384          * HAVING clause, which we put in WHERE so that query_planner() can use it
 385          * in a gating Result node, but also keep in HAVING to ensure that we
 386          * don't emit a bogus aggregated row. (This could be done better, but it
 387          * seems not worth optimizing.)
 388          *
 389          * Note that both havingQual and parse->jointree->quals are in
 390          * implicitly-ANDed-list form at this point, even though they are declared
 391          * as Node *.
 392          */
 393         newHaving = NIL;
 394         foreach(l, (List *) parse->havingQual)
 395         {
 396                 Node       *havingclause = (Node *) lfirst(l);
 397
 398                 if (contain_agg_clause(havingclause) ||
 399                         contain_volatile_functions(havingclause) ||
 400                         contain_subplans(havingclause))
 401                 {
 402                         /* keep it in HAVING */
 403                         newHaving = lappend(newHaving, havingclause);
 404                 }
 405                 else if (parse->groupClause)
 406                 {
 407                         /* move it to WHERE */
 408                         parse->jointree->quals = (Node *)
 409                                 lappend((List *) parse->jointree->quals, havingclause);
 410                 }
 411                 else
 412                 {
 413                         /* put a copy in WHERE, keep it in HAVING */
 414                         parse->jointree->quals = (Node *)
 415                                 lappend((List *) parse->jointree->quals,
 416                                                 copyObject(havingclause));
 417                         newHaving = lappend(newHaving, havingclause);
 418                 }
 419         }
 420         parse->havingQual = (Node *) newHaving;
 421
 422         /*
 423          * If we have any outer joins, try to reduce them to plain inner joins.
 424          * This step is most easily done after we've done expression
 425          * preprocessing.
 426          */
 427         if (root->hasOuterJoins)
 428                 reduce_outer_joins(root);
 429
 430         /*
 431          * Do the main planning.  If we have an inherited target relation, that
 432          * needs special processing, else go straight to grouping_planner.
 433          */
 434         if (parse->resultRelation &&
 435                 rt_fetch(parse->resultRelation, parse->rtable)->inh)
 436                 plan = inheritance_planner(root);
 437         else
 438                 plan = grouping_planner(root, tuple_fraction);
 439
 440         /*
 441          * If any subplans were generated, or if we're inside a subplan, build
 442          * initPlan list and extParam/allParam sets for plan nodes, and attach the
 443          * initPlans to the top plan node.
 444          */
 445         if (list_length(glob->subplans) != num_old_subplans ||
 446                 root->query_level > 1)
 447                 SS_finalize_plan(root, plan);
 448
 449         /* Return internal info if caller wants it */
 450         if (subroot)
 451                 *subroot = root;
 452
 453         return plan;
 454 }
 455
 456 /*
 457  * preprocess_expression
 458  *              Do subquery_planner's preprocessing work for an expression,
 459  *              which can be a targetlist, a WHERE clause (including JOIN/ON
 460  *              conditions), or a HAVING clause.
 461  */
 462 static Node *
 463 preprocess_expression(PlannerInfo *root, Node *expr, int kind)
 464 {
 465         /*
 466          * Fall out quickly if expression is empty.  This occurs often enough to
 467          * be worth checking.  Note that null->null is the correct conversion for
 468          * implicit-AND result format, too.
 469          */
 470         if (expr == NULL)
 471                 return NULL;
 472
 473         /*
 474          * If the query has any join RTEs, replace join alias variables with
 475          * base-relation variables. We must do this before sublink processing,
 476          * else sublinks expanded out from join aliases wouldn't get processed. We
 477          * can skip it in VALUES lists, however, since they can't contain any Vars
 478          * at all.
 479          */
 480         if (root->hasJoinRTEs && kind != EXPRKIND_VALUES)
 481                 expr = flatten_join_alias_vars(root, expr);
 482
 483         /*
 484          * Simplify constant expressions.
 485          *
 486          * Note: this also flattens nested AND and OR expressions into N-argument
 487          * form.  All processing of a qual expression after this point must be
 488          * careful to maintain AND/OR flatness --- that is, do not generate a tree
 489          * with AND directly under AND, nor OR directly under OR.
 490          *
 491          * Because this is a relatively expensive process, we skip it when the
 492          * query is trivial, such as "SELECT 2+2;" or "INSERT ... VALUES()". The
 493          * expression will only be evaluated once anyway, so no point in
 494          * pre-simplifying; we can't execute it any faster than the executor can,
 495          * and we will waste cycles copying the tree.  Notice however that we
 496          * still must do it for quals (to get AND/OR flatness); and if we are in a
 497          * subquery we should not assume it will be done only once.
 498          *
 499          * For VALUES lists we never do this at all, again on the grounds that we
 500          * should optimize for one-time evaluation.
 501          */
 502         if (kind != EXPRKIND_VALUES &&
 503                 (root->parse->jointree->fromlist != NIL ||
 504                  kind == EXPRKIND_QUAL ||
 505                  root->query_level > 1))
 506                 expr = eval_const_expressions(root, expr);
 507
 508         /*
 509          * If it's a qual or havingQual, canonicalize it.
 510          */
 511         if (kind == EXPRKIND_QUAL)
 512         {
 513                 expr = (Node *) canonicalize_qual((Expr *) expr);
 514
 515 #ifdef OPTIMIZER_DEBUG
 516                 printf("After canonicalize_qual()\n");
 517                 pprint(expr);
 518 #endif
 519         }
 520
 521         /* Expand SubLinks to SubPlans */
 522         if (root->parse->hasSubLinks)
 523                 expr = SS_process_sublinks(root, expr, (kind == EXPRKIND_QUAL));
 524
 525         /*
 526          * XXX do not insert anything here unless you have grokked the comments in
 527          * SS_replace_correlation_vars ...
 528          */
 529
 530         /* Replace uplevel vars with Param nodes (this IS possible in VALUES) */
 531         if (root->query_level > 1)
 532                 expr = SS_replace_correlation_vars(root, expr);
 533
 534         /*
 535          * If it's a qual or havingQual, convert it to implicit-AND format. (We
 536          * don't want to do this before eval_const_expressions, since the latter
 537          * would be unable to simplify a top-level AND correctly. Also,
 538          * SS_process_sublinks expects explicit-AND format.)
 539          */
 540         if (kind == EXPRKIND_QUAL)
 541                 expr = (Node *) make_ands_implicit((Expr *) expr);
 542
 543         return expr;
 544 }
 545
 546 /*
 547  * preprocess_qual_conditions
 548  *              Recursively scan the query's jointree and do subquery_planner's
 549  *              preprocessing work on each qual condition found therein.
 550  */
 551 static void
 552 preprocess_qual_conditions(PlannerInfo *root, Node *jtnode)
 553 {
 554         if (jtnode == NULL)
 555                 return;
 556         if (IsA(jtnode, RangeTblRef))
 557         {
 558                 /* nothing to do here */
 559         }
 560         else if (IsA(jtnode, FromExpr))
 561         {
 562                 FromExpr   *f = (FromExpr *) jtnode;
 563                 ListCell   *l;
 564
 565                 foreach(l, f->fromlist)
 566                         preprocess_qual_conditions(root, lfirst(l));
 567
 568                 f->quals = preprocess_expression(root, f->quals, EXPRKIND_QUAL);
 569         }
 570         else if (IsA(jtnode, JoinExpr))
 571         {
 572                 JoinExpr   *j = (JoinExpr *) jtnode;
 573
 574                 preprocess_qual_conditions(root, j->larg);
 575                 preprocess_qual_conditions(root, j->rarg);
 576
 577                 j->quals = preprocess_expression(root, j->quals, EXPRKIND_QUAL);
 578         }
 579         else
 580                 elog(ERROR, "unrecognized node type: %d",
 581                          (int) nodeTag(jtnode));
 582 }
 583
 584 /*
 585  * inheritance_planner
 586  *        Generate a plan in the case where the result relation is an
 587  *        inheritance set.
 588  *
 589  * We have to handle this case differently from cases where a source relation
 590  * is an inheritance set. Source inheritance is expanded at the bottom of the
 591  * plan tree (see allpaths.c), but target inheritance has to be expanded at
 592  * the top.  The reason is that for UPDATE, each target relation needs a
 593  * different targetlist matching its own column set.  Also, for both UPDATE
 594  * and DELETE, the executor needs the Append plan node at the top, else it
 595  * can't keep track of which table is the current target table.  Fortunately,
 596  * the UPDATE/DELETE target can never be the nullable side of an outer join,
 597  * so it's OK to generate the plan this way.
 598  *
 599  * Returns a query plan.
 600  */
 601 static Plan *
 602 inheritance_planner(PlannerInfo *root)
 603 {
 604         Query      *parse = root->parse;
 605         int                     parentRTindex = parse->resultRelation;
 606         List       *subplans = NIL;
 607         List       *resultRelations = NIL;
 608         List       *returningLists = NIL;
 609         List       *rtable = NIL;
 610         List       *tlist = NIL;
 611         PlannerInfo subroot;
 612         ListCell   *l;
 613
 614         foreach(l, root->append_rel_list)
 615         {
 616                 AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
 617                 Plan       *subplan;
 618
 619                 /* append_rel_list contains all append rels; ignore others */
 620                 if (appinfo->parent_relid != parentRTindex)
 621                         continue;
 622
 623                 /*
 624                  * Generate modified query with this rel as target.  We have to be
 625                  * prepared to translate varnos in in_info_list as well as in the
 626                  * Query proper.
 627                  */
 628                 memcpy(&subroot, root, sizeof(PlannerInfo));
 629                 subroot.parse = (Query *)
 630                         adjust_appendrel_attrs((Node *) parse,
 631                                                                    appinfo);
 632                 subroot.in_info_list = (List *)
 633                         adjust_appendrel_attrs((Node *) root->in_info_list,
 634                                                                    appinfo);
 635                 subroot.init_plans = NIL;
 636                 /* There shouldn't be any OJ info to translate, as yet */
 637                 Assert(subroot.oj_info_list == NIL);
 638
 639                 /* Generate plan */
 640                 subplan = grouping_planner(&subroot, 0.0 /* retrieve all tuples */ );
 641
 642                 /*
 643                  * If this child rel was excluded by constraint exclusion, exclude it
 644                  * from the plan.
 645                  */
 646                 if (is_dummy_plan(subplan))
 647                         continue;
 648
 649                 /* Save rtable and tlist from first rel for use below */
 650                 if (subplans == NIL)
 651                 {
 652                         rtable = subroot.parse->rtable;
 653                         tlist = subplan->targetlist;
 654                 }
 655
 656                 subplans = lappend(subplans, subplan);
 657
 658                 /* Make sure any initplans from this rel get into the outer list */
 659                 root->init_plans = list_concat(root->init_plans, subroot.init_plans);
 660
 661                 /* Build target-relations list for the executor */
 662                 resultRelations = lappend_int(resultRelations, appinfo->child_relid);
 663
 664                 /* Build list of per-relation RETURNING targetlists */
 665                 if (parse->returningList)
 666                 {
 667                         Assert(list_length(subroot.returningLists) == 1);
 668                         returningLists = list_concat(returningLists,
 669                                                                                  subroot.returningLists);
 670                 }
 671         }
 672
 673         root->resultRelations = resultRelations;
 674         root->returningLists = returningLists;
 675
 676         /* Mark result as unordered (probably unnecessary) */
 677         root->query_pathkeys = NIL;
 678
 679         /*
 680          * If we managed to exclude every child rel, return a dummy plan
 681          */
 682         if (subplans == NIL)
 683         {
 684                 root->resultRelations = list_make1_int(parentRTindex);
 685                 /* although dummy, it must have a valid tlist for executor */
 686                 tlist = preprocess_targetlist(root, parse->targetList);
 687                 return (Plan *) make_result(root,
 688                                                                         tlist,
 689                                                                         (Node *) list_make1(makeBoolConst(false,
 690                                                                                                                                           false)),
 691                                                                         NULL);
 692         }
 693
 694         /*
 695          * Planning might have modified the rangetable, due to changes of the
 696          * Query structures inside subquery RTEs.  We have to ensure that this
 697          * gets propagated back to the master copy.  But can't do this until we
 698          * are done planning, because all the calls to grouping_planner need
 699          * virgin sub-Queries to work from.  (We are effectively assuming that
 700          * sub-Queries will get planned identically each time, or at least that
 701          * the impacts on their rangetables will be the same each time.)
 702          *
 703          * XXX should clean this up someday
 704          */
 705         parse->rtable = rtable;
 706
 707         /* Suppress Append if there's only one surviving child rel */
 708         if (list_length(subplans) == 1)
 709                 return (Plan *) linitial(subplans);
 710
 711         return (Plan *) make_append(subplans, true, tlist);
 712 }
 713
 714 /*--------------------
 715  * grouping_planner
 716  *        Perform planning steps related to grouping, aggregation, etc.
 717  *        This primarily means adding top-level processing to the basic
 718  *        query plan produced by query_planner.
 719  *
 720  * tuple_fraction is the fraction of tuples we expect will be retrieved
 721  *
 722  * tuple_fraction is interpreted as follows:
 723  *        0: expect all tuples to be retrieved (normal case)
 724  *        0 < tuple_fraction < 1: expect the given fraction of tuples available
 725  *              from the plan to be retrieved
 726  *        tuple_fraction >= 1: tuple_fraction is the absolute number of tuples
 727  *              expected to be retrieved (ie, a LIMIT specification)
 728  *
 729  * Returns a query plan.  Also, root->query_pathkeys is returned as the
 730  * actual output ordering of the plan (in pathkey format).
 731  *--------------------
 732  */
 733 static Plan *
 734 grouping_planner(PlannerInfo *root, double tuple_fraction)
 735 {
 736         Query      *parse = root->parse;
 737         List       *tlist = parse->targetList;
 738         int64           offset_est = 0;
 739         int64           count_est = 0;
 740         double          limit_tuples = -1.0;
 741         Plan       *result_plan;
 742         List       *current_pathkeys;
 743         List       *sort_pathkeys;
 744         double          dNumGroups = 0;
 745
 746         /* Tweak caller-supplied tuple_fraction if have LIMIT/OFFSET */
 747         if (parse->limitCount || parse->limitOffset)
 748         {
 749                 tuple_fraction = preprocess_limit(root, tuple_fraction,
 750                                                                                   &offset_est, &count_est);
 751
 752                 /*
 753                  * If we have a known LIMIT, and don't have an unknown OFFSET, we can
 754                  * estimate the effects of using a bounded sort.
 755                  */
 756                 if (count_est > 0 && offset_est >= 0)
 757                         limit_tuples = (double) count_est + (double) offset_est;
 758         }
 759
 760         if (parse->setOperations)
 761         {
 762                 List       *set_sortclauses;
 763
 764                 /*
 765                  * If there's a top-level ORDER BY, assume we have to fetch all the
 766                  * tuples.      This might seem too simplistic given all the hackery below
 767                  * to possibly avoid the sort ... but a nonzero tuple_fraction is only
 768                  * of use to plan_set_operations() when the setop is UNION ALL, and
 769                  * the result of UNION ALL is always unsorted.
 770                  */
 771                 if (parse->sortClause)
 772                         tuple_fraction = 0.0;
 773
 774                 /*
 775                  * Construct the plan for set operations.  The result will not need
 776                  * any work except perhaps a top-level sort and/or LIMIT.
 777                  */
 778                 result_plan = plan_set_operations(root, tuple_fraction,
 779                                                                                   &set_sortclauses);
 780
 781                 /*
 782                  * Calculate pathkeys representing the sort order (if any) of the set
 783                  * operation's result.  We have to do this before overwriting the sort
 784                  * key information...
 785                  */
 786                 current_pathkeys = make_pathkeys_for_sortclauses(root,
 787                                                                                                                  set_sortclauses,
 788                                                                                                          result_plan->targetlist,
 789                                                                                                                  true);
 790
 791                 /*
 792                  * We should not need to call preprocess_targetlist, since we must be
 793                  * in a SELECT query node.      Instead, use the targetlist returned by
 794                  * plan_set_operations (since this tells whether it returned any
 795                  * resjunk columns!), and transfer any sort key information from the
 796                  * original tlist.
 797                  */
 798                 Assert(parse->commandType == CMD_SELECT);
 799
 800                 tlist = postprocess_setop_tlist(result_plan->targetlist, tlist);
 801
 802                 /*
 803                  * Can't handle FOR UPDATE/SHARE here (parser should have checked
 804                  * already, but let's make sure).
 805                  */
 806                 if (parse->rowMarks)
 807                         ereport(ERROR,
 808                                         (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
 809                                          errmsg("SELECT FOR UPDATE/SHARE is not allowed with UNION/INTERSECT/EXCEPT")));
 810
 811                 /*
 812                  * Calculate pathkeys that represent result ordering requirements
 813                  */
 814                 sort_pathkeys = make_pathkeys_for_sortclauses(root,
 815                                                                                                           parse->sortClause,
 816                                                                                                           tlist,
 817                                                                                                           true);
 818         }
 819         else
 820         {
 821                 /* No set operations, do regular planning */
 822                 List       *sub_tlist;
 823                 List       *group_pathkeys;
 824                 AttrNumber *groupColIdx = NULL;
 825                 Oid                *groupOperators = NULL;
 826                 bool            need_tlist_eval = true;
 827                 QualCost        tlist_cost;
 828                 Path       *cheapest_path;
 829                 Path       *sorted_path;
 830                 Path       *best_path;
 831                 long            numGroups = 0;
 832                 AggClauseCounts agg_counts;
 833                 int                     numGroupCols = list_length(parse->groupClause);
 834                 bool            use_hashed_grouping = false;
 835
 836                 MemSet(&agg_counts, 0, sizeof(AggClauseCounts));
 837
 838                 /* Preprocess targetlist */
 839                 tlist = preprocess_targetlist(root, tlist);
 840
 841                 /*
 842                  * Generate appropriate target list for subplan; may be different from
 843                  * tlist if grouping or aggregation is needed.
 844                  */
 845                 sub_tlist = make_subplanTargetList(root, tlist,
 846                                                                                    &groupColIdx, &need_tlist_eval);
 847
 848                 /*
 849                  * Calculate pathkeys that represent grouping/ordering requirements.
 850                  * Stash them in PlannerInfo so that query_planner can canonicalize
 851                  * them after EquivalenceClasses have been formed.
 852                  */
 853                 root->group_pathkeys =
 854                         make_pathkeys_for_sortclauses(root,
 855                                                                                   parse->groupClause,
 856                                                                                   tlist,
 857                                                                                   false);
 858                 root->sort_pathkeys =
 859                         make_pathkeys_for_sortclauses(root,
 860                                                                                   parse->sortClause,
 861                                                                                   tlist,
 862                                                                                   false);
 863
 864                 /*
 865                  * Will need actual number of aggregates for estimating costs.
 866                  *
 867                  * Note: we do not attempt to detect duplicate aggregates here; a
 868                  * somewhat-overestimated count is okay for our present purposes.
 869                  *
 870                  * Note: think not that we can turn off hasAggs if we find no aggs. It
 871                  * is possible for constant-expression simplification to remove all
 872                  * explicit references to aggs, but we still have to follow the
 873                  * aggregate semantics (eg, producing only one output row).
 874                  */
 875                 if (parse->hasAggs)
 876                 {
 877                         count_agg_clauses((Node *) tlist, &agg_counts);
 878                         count_agg_clauses(parse->havingQual, &agg_counts);
 879                 }
 880
 881                 /*
 882                  * Figure out whether we need a sorted result from query_planner.
 883                  *
 884                  * If we have a GROUP BY clause, then we want a result sorted properly
 885                  * for grouping.  Otherwise, if there is an ORDER BY clause, we want
 886                  * to sort by the ORDER BY clause.      (Note: if we have both, and ORDER
 887                  * BY is a superset of GROUP BY, it would be tempting to request sort
 888                  * by ORDER BY --- but that might just leave us failing to exploit an
 889                  * available sort order at all. Needs more thought...)
 890                  */
 891                 if (parse->groupClause)
 892                         root->query_pathkeys = root->group_pathkeys;
 893                 else if (parse->sortClause)
 894                         root->query_pathkeys = root->sort_pathkeys;
 895                 else
 896                         root->query_pathkeys = NIL;
 897
 898                 /*
 899                  * Generate the best unsorted and presorted paths for this Query (but
 900                  * note there may not be any presorted path).  query_planner will also
 901                  * estimate the number of groups in the query, and canonicalize all
 902                  * the pathkeys.
 903                  */
 904                 query_planner(root, sub_tlist, tuple_fraction, limit_tuples,
 905                                           &cheapest_path, &sorted_path, &dNumGroups);
 906
 907                 group_pathkeys = root->group_pathkeys;
 908                 sort_pathkeys = root->sort_pathkeys;
 909
 910                 /*
 911                  * If grouping, extract the grouping operators and decide whether we
 912                  * want to use hashed grouping.
 913                  */
 914                 if (parse->groupClause)
 915                 {
 916                         groupOperators = extract_grouping_ops(parse->groupClause);
 917                         use_hashed_grouping =
 918                                 choose_hashed_grouping(root, tuple_fraction, limit_tuples,
 919                                                                            cheapest_path, sorted_path,
 920                                                                            groupOperators, dNumGroups,
 921                                                                            &agg_counts);
 922
 923                         /* Also convert # groups to long int --- but 'ware overflow! */
 924                         numGroups = (long) Min(dNumGroups, (double) LONG_MAX);
 925                 }
 926
 927                 /*
 928                  * Select the best path.  If we are doing hashed grouping, we will
 929                  * always read all the input tuples, so use the cheapest-total path.
 930                  * Otherwise, trust query_planner's decision about which to use.
 931                  */
 932                 if (use_hashed_grouping || !sorted_path)
 933                         best_path = cheapest_path;
 934                 else
 935                         best_path = sorted_path;
 936
 937                 /*
 938                  * Check to see if it's possible to optimize MIN/MAX aggregates. If
 939                  * so, we will forget all the work we did so far to choose a "regular"
 940                  * path ... but we had to do it anyway to be able to tell which way is
 941                  * cheaper.
 942                  */
 943                 result_plan = optimize_minmax_aggregates(root,
 944                                                                                                  tlist,
 945                                                                                                  best_path);
 946                 if (result_plan != NULL)
 947                 {
 948                         /*
 949                          * optimize_minmax_aggregates generated the full plan, with the
 950                          * right tlist, and it has no sort order.
 951                          */
 952                         current_pathkeys = NIL;
 953                 }
 954                 else
 955                 {
 956                         /*
 957                          * Normal case --- create a plan according to query_planner's
 958                          * results.
 959                          */
 960                         result_plan = create_plan(root, best_path);
 961                         current_pathkeys = best_path->pathkeys;
 962
 963                         /*
 964                          * create_plan() returns a plan with just a "flat" tlist of
 965                          * required Vars.  Usually we need to insert the sub_tlist as the
 966                          * tlist of the top plan node.  However, we can skip that if we
 967                          * determined that whatever query_planner chose to return will be
 968                          * good enough.
 969                          */
 970                         if (need_tlist_eval)
 971                         {
 972                                 /*
 973                                  * If the top-level plan node is one that cannot do expression
 974                                  * evaluation, we must insert a Result node to project the
 975                                  * desired tlist.
 976                                  */
 977                                 if (!is_projection_capable_plan(result_plan))
 978                                 {
 979                                         result_plan = (Plan *) make_result(root,
 980                                                                                                            sub_tlist,
 981                                                                                                            NULL,
 982                                                                                                            result_plan);
 983                                 }
 984                                 else
 985                                 {
 986                                         /*
 987                                          * Otherwise, just replace the subplan's flat tlist with
 988                                          * the desired tlist.
 989                                          */
 990                                         result_plan->targetlist = sub_tlist;
 991                                 }
 992
 993                                 /*
 994                                  * Also, account for the cost of evaluation of the sub_tlist.
 995                                  *
 996                                  * Up to now, we have only been dealing with "flat" tlists,
 997                                  * containing just Vars.  So their evaluation cost is zero
 998                                  * according to the model used by cost_qual_eval() (or if you
 999                                  * prefer, the cost is factored into cpu_tuple_cost).  Thus we
1000                                  * can avoid accounting for tlist cost throughout
1001                                  * query_planner() and subroutines.  But now we've inserted a
1002                                  * tlist that might contain actual operators, sub-selects, etc
1003                                  * --- so we'd better account for its cost.
1004                                  *
1005                                  * Below this point, any tlist eval cost for added-on nodes
1006                                  * should be accounted for as we create those nodes.
1007                                  * Presently, of the node types we can add on, only Agg and
1008                                  * Group project new tlists (the rest just copy their input
1009                                  * tuples) --- so make_agg() and make_group() are responsible
1010                                  * for computing the added cost.
1011                                  */
1012                                 cost_qual_eval(&tlist_cost, sub_tlist, root);
1013                                 result_plan->startup_cost += tlist_cost.startup;
1014                                 result_plan->total_cost += tlist_cost.startup +
1015                                         tlist_cost.per_tuple * result_plan->plan_rows;
1016                         }
1017                         else
1018                         {
1019                                 /*
1020                                  * Since we're using query_planner's tlist and not the one
1021                                  * make_subplanTargetList calculated, we have to refigure any
1022                                  * grouping-column indexes make_subplanTargetList computed.
1023                                  */
1024                                 locate_grouping_columns(root, tlist, result_plan->targetlist,
1025                                                                                 groupColIdx);
1026                         }
1027
1028                         /*
1029                          * Insert AGG or GROUP node if needed, plus an explicit sort step
1030                          * if necessary.
1031                          *
1032                          * HAVING clause, if any, becomes qual of the Agg or Group node.
1033                          */
1034                         if (use_hashed_grouping)
1035                         {
1036                                 /* Hashed aggregate plan --- no sort needed */
1037                                 result_plan = (Plan *) make_agg(root,
1038                                                                                                 tlist,
1039                                                                                                 (List *) parse->havingQual,
1040                                                                                                 AGG_HASHED,
1041                                                                                                 numGroupCols,
1042                                                                                                 groupColIdx,
1043                                                                                                 groupOperators,
1044                                                                                                 numGroups,
1045                                                                                                 agg_counts.numAggs,
1046                                                                                                 result_plan);
1047                                 /* Hashed aggregation produces randomly-ordered results */
1048                                 current_pathkeys = NIL;
1049                         }
1050                         else if (parse->hasAggs)
1051                         {
1052                                 /* Plain aggregate plan --- sort if needed */
1053                                 AggStrategy aggstrategy;
1054
1055                                 if (parse->groupClause)
1056                                 {
1057                                         if (!pathkeys_contained_in(group_pathkeys,
1058                                                                                            current_pathkeys))
1059                                         {
1060                                                 result_plan = (Plan *)
1061                                                         make_sort_from_groupcols(root,
1062                                                                                                          parse->groupClause,
1063                                                                                                          groupColIdx,
1064                                                                                                          result_plan);
1065                                                 current_pathkeys = group_pathkeys;
1066                                         }
1067                                         aggstrategy = AGG_SORTED;
1068
1069                                         /*
1070                                          * The AGG node will not change the sort ordering of its
1071                                          * groups, so current_pathkeys describes the result too.
1072                                          */
1073                                 }
1074                                 else
1075                                 {
1076                                         aggstrategy = AGG_PLAIN;
1077                                         /* Result will be only one row anyway; no sort order */
1078                                         current_pathkeys = NIL;
1079                                 }
1080
1081                                 result_plan = (Plan *) make_agg(root,
1082                                                                                                 tlist,
1083                                                                                                 (List *) parse->havingQual,
1084                                                                                                 aggstrategy,
1085                                                                                                 numGroupCols,
1086                                                                                                 groupColIdx,
1087                                                                                                 groupOperators,
1088                                                                                                 numGroups,
1089                                                                                                 agg_counts.numAggs,
1090                                                                                                 result_plan);
1091                         }
1092                         else if (parse->groupClause)
1093                         {
1094                                 /*
1095                                  * GROUP BY without aggregation, so insert a group node (plus
1096                                  * the appropriate sort node, if necessary).
1097                                  *
1098                                  * Add an explicit sort if we couldn't make the path come out
1099                                  * the way the GROUP node needs it.
1100                                  */
1101                                 if (!pathkeys_contained_in(group_pathkeys, current_pathkeys))
1102                                 {
1103                                         result_plan = (Plan *)
1104                                                 make_sort_from_groupcols(root,
1105                                                                                                  parse->groupClause,
1106                                                                                                  groupColIdx,
1107                                                                                                  result_plan);
1108                                         current_pathkeys = group_pathkeys;
1109                                 }
1110
1111                                 result_plan = (Plan *) make_group(root,
1112                                                                                                   tlist,
1113                                                                                                   (List *) parse->havingQual,
1114                                                                                                   numGroupCols,
1115                                                                                                   groupColIdx,
1116                                                                                                   groupOperators,
1117                                                                                                   dNumGroups,
1118                                                                                                   result_plan);
1119                                 /* The Group node won't change sort ordering */
1120                         }
1121                         else if (root->hasHavingQual)
1122                         {
1123                                 /*
1124                                  * No aggregates, and no GROUP BY, but we have a HAVING qual.
1125                                  * This is a degenerate case in which we are supposed to emit
1126                                  * either 0 or 1 row depending on whether HAVING succeeds.
1127                                  * Furthermore, there cannot be any variables in either HAVING
1128                                  * or the targetlist, so we actually do not need the FROM
1129                                  * table at all!  We can just throw away the plan-so-far and
1130                                  * generate a Result node.      This is a sufficiently unusual
1131                                  * corner case that it's not worth contorting the structure of
1132                                  * this routine to avoid having to generate the plan in the
1133                                  * first place.
1134                                  */
1135                                 result_plan = (Plan *) make_result(root,
1136                                                                                                    tlist,
1137                                                                                                    parse->havingQual,
1138                                                                                                    NULL);
1139                         }
1140                 }                                               /* end of non-minmax-aggregate case */
1141         }                                                       /* end of if (setOperations) */
1142
1143         /*
1144          * If we were not able to make the plan come out in the right order, add
1145          * an explicit sort step.
1146          */
1147         if (parse->sortClause)
1148         {
1149                 if (!pathkeys_contained_in(sort_pathkeys, current_pathkeys))
1150                 {
1151                         result_plan = (Plan *) make_sort_from_pathkeys(root,
1152                                                                                                                    result_plan,
1153                                                                                                                    sort_pathkeys,
1154                                                                                                                    limit_tuples);
1155                         current_pathkeys = sort_pathkeys;
1156                 }
1157         }
1158
1159         /*
1160          * If there is a DISTINCT clause, add the UNIQUE node.
1161          */
1162         if (parse->distinctClause)
1163         {
1164                 result_plan = (Plan *) make_unique(result_plan, parse->distinctClause);
1165
1166                 /*
1167                  * If there was grouping or aggregation, leave plan_rows as-is (ie,
1168                  * assume the result was already mostly unique).  If not, use the
1169                  * number of distinct-groups calculated by query_planner.
1170                  */
1171                 if (!parse->groupClause && !root->hasHavingQual && !parse->hasAggs)
1172                         result_plan->plan_rows = dNumGroups;
1173         }
1174
1175         /*
1176          * Finally, if there is a LIMIT/OFFSET clause, add the LIMIT node.
1177          */
1178         if (parse->limitCount || parse->limitOffset)
1179         {
1180                 result_plan = (Plan *) make_limit(result_plan,
1181                                                                                   parse->limitOffset,
1182                                                                                   parse->limitCount,
1183                                                                                   offset_est,
1184                                                                                   count_est);
1185         }
1186
1187         /*
1188          * Deal with the RETURNING clause if any.  It's convenient to pass the
1189          * returningList through setrefs.c now rather than at top level (if we
1190          * waited, handling inherited UPDATE/DELETE would be much harder).
1191          */
1192         if (parse->returningList)
1193         {
1194                 List       *rlist;
1195
1196                 Assert(parse->resultRelation);
1197                 rlist = set_returning_clause_references(root->glob,
1198                                                                                                 parse->returningList,
1199                                                                                                 result_plan,
1200                                                                                                 parse->resultRelation);
1201                 root->returningLists = list_make1(rlist);
1202         }
1203         else
1204                 root->returningLists = NIL;
1205
1206         /* Compute result-relations list if needed */
1207         if (parse->resultRelation)
1208                 root->resultRelations = list_make1_int(parse->resultRelation);
1209         else
1210                 root->resultRelations = NIL;
1211
1212         /*
1213          * Return the actual output ordering in query_pathkeys for possible use by
1214          * an outer query level.
1215          */
1216         root->query_pathkeys = current_pathkeys;
1217
1218         return result_plan;
1219 }
1220
1221 /*
1222  * Detect whether a plan node is a "dummy" plan created when a relation
1223  * is deemed not to need scanning due to constraint exclusion.
1224  *
1225  * Currently, such dummy plans are Result nodes with constant FALSE
1226  * filter quals.
1227  */
1228 static bool
1229 is_dummy_plan(Plan *plan)
1230 {
1231         if (IsA(plan, Result))
1232         {
1233                 List       *rcqual = (List *) ((Result *) plan)->resconstantqual;
1234
1235                 if (list_length(rcqual) == 1)
1236                 {
1237                         Const      *constqual = (Const *) linitial(rcqual);
1238
1239                         if (constqual && IsA(constqual, Const))
1240                         {
1241                                 if (!constqual->constisnull &&
1242                                         !DatumGetBool(constqual->constvalue))
1243                                         return true;
1244                         }
1245                 }
1246         }
1247         return false;
1248 }
1249
1250 /*
1251  * preprocess_limit - do pre-estimation for LIMIT and/or OFFSET clauses
1252  *
1253  * We try to estimate the values of the LIMIT/OFFSET clauses, and pass the
1254  * results back in *count_est and *offset_est.  These variables are set to
1255  * 0 if the corresponding clause is not present, and -1 if it's present
1256  * but we couldn't estimate the value for it.  (The "0" convention is OK
1257  * for OFFSET but a little bit bogus for LIMIT: effectively we estimate
1258  * LIMIT 0 as though it were LIMIT 1.  But this is in line with the planner's
1259  * usual practice of never estimating less than one row.)  These values will
1260  * be passed to make_limit, which see if you change this code.
1261  *
1262  * The return value is the suitably adjusted tuple_fraction to use for
1263  * planning the query.  This adjustment is not overridable, since it reflects
1264  * plan actions that grouping_planner() will certainly take, not assumptions
1265  * about context.
1266  */
1267 static double
1268 preprocess_limit(PlannerInfo *root, double tuple_fraction,
1269                                  int64 *offset_est, int64 *count_est)
1270 {
1271         Query      *parse = root->parse;
1272         Node       *est;
1273         double          limit_fraction;
1274
1275         /* Should not be called unless LIMIT or OFFSET */
1276         Assert(parse->limitCount || parse->limitOffset);
1277
1278         /*
1279          * Try to obtain the clause values.  We use estimate_expression_value
1280          * primarily because it can sometimes do something useful with Params.
1281          */
1282         if (parse->limitCount)
1283         {
1284                 est = estimate_expression_value(root, parse->limitCount);
1285                 if (est && IsA(est, Const))
1286                 {
1287                         if (((Const *) est)->constisnull)
1288                         {
1289                                 /* NULL indicates LIMIT ALL, ie, no limit */
1290                                 *count_est = 0; /* treat as not present */
1291                         }
1292                         else
1293                         {
1294                                 *count_est = DatumGetInt64(((Const *) est)->constvalue);
1295                                 if (*count_est <= 0)
1296                                         *count_est = 1;         /* force to at least 1 */
1297                         }
1298                 }
1299                 else
1300                         *count_est = -1;        /* can't estimate */
1301         }
1302         else
1303                 *count_est = 0;                 /* not present */
1304
1305         if (parse->limitOffset)
1306         {
1307                 est = estimate_expression_value(root, parse->limitOffset);
1308                 if (est && IsA(est, Const))
1309                 {
1310                         if (((Const *) est)->constisnull)
1311                         {
1312                                 /* Treat NULL as no offset; the executor will too */
1313                                 *offset_est = 0;        /* treat as not present */
1314                         }
1315                         else
1316                         {
1317                                 *offset_est = DatumGetInt64(((Const *) est)->constvalue);
1318                                 if (*offset_est < 0)
1319                                         *offset_est = 0;        /* less than 0 is same as 0 */
1320                         }
1321                 }
1322                 else
1323                         *offset_est = -1;       /* can't estimate */
1324         }
1325         else
1326                 *offset_est = 0;                /* not present */
1327
1328         if (*count_est != 0)
1329         {
1330                 /*
1331                  * A LIMIT clause limits the absolute number of tuples returned.
1332                  * However, if it's not a constant LIMIT then we have to guess; for
1333                  * lack of a better idea, assume 10% of the plan's result is wanted.
1334                  */
1335                 if (*count_est < 0 || *offset_est < 0)
1336                 {
1337                         /* LIMIT or OFFSET is an expression ... punt ... */
1338                         limit_fraction = 0.10;
1339                 }
1340                 else
1341                 {
1342                         /* LIMIT (plus OFFSET, if any) is max number of tuples needed */
1343                         limit_fraction = (double) *count_est + (double) *offset_est;
1344                 }
1345
1346                 /*
1347                  * If we have absolute limits from both caller and LIMIT, use the
1348                  * smaller value; likewise if they are both fractional.  If one is
1349                  * fractional and the other absolute, we can't easily determine which
1350                  * is smaller, but we use the heuristic that the absolute will usually
1351                  * be smaller.
1352                  */
1353                 if (tuple_fraction >= 1.0)
1354                 {
1355                         if (limit_fraction >= 1.0)
1356                         {
1357                                 /* both absolute */
1358                                 tuple_fraction = Min(tuple_fraction, limit_fraction);
1359                         }
1360                         else
1361                         {
1362                                 /* caller absolute, limit fractional; use caller's value */
1363                         }
1364                 }
1365                 else if (tuple_fraction > 0.0)
1366                 {
1367                         if (limit_fraction >= 1.0)
1368                         {
1369                                 /* caller fractional, limit absolute; use limit */
1370                                 tuple_fraction = limit_fraction;
1371                         }
1372                         else
1373                         {
1374                                 /* both fractional */
1375                                 tuple_fraction = Min(tuple_fraction, limit_fraction);
1376                         }
1377                 }
1378                 else
1379                 {
1380                         /* no info from caller, just use limit */
1381                         tuple_fraction = limit_fraction;
1382                 }
1383         }
1384         else if (*offset_est != 0 && tuple_fraction > 0.0)
1385         {
1386                 /*
1387                  * We have an OFFSET but no LIMIT.      This acts entirely differently
1388                  * from the LIMIT case: here, we need to increase rather than decrease
1389                  * the caller's tuple_fraction, because the OFFSET acts to cause more
1390                  * tuples to be fetched instead of fewer.  This only matters if we got
1391                  * a tuple_fraction > 0, however.
1392                  *
1393                  * As above, use 10% if OFFSET is present but unestimatable.
1394                  */
1395                 if (*offset_est < 0)
1396                         limit_fraction = 0.10;
1397                 else
1398                         limit_fraction = (double) *offset_est;
1399
1400                 /*
1401                  * If we have absolute counts from both caller and OFFSET, add them
1402                  * together; likewise if they are both fractional.      If one is
1403                  * fractional and the other absolute, we want to take the larger, and
1404                  * we heuristically assume that's the fractional one.
1405                  */
1406                 if (tuple_fraction >= 1.0)
1407                 {
1408                         if (limit_fraction >= 1.0)
1409                         {
1410                                 /* both absolute, so add them together */
1411                                 tuple_fraction += limit_fraction;
1412                         }
1413                         else
1414                         {
1415                                 /* caller absolute, limit fractional; use limit */
1416                                 tuple_fraction = limit_fraction;
1417                         }
1418                 }
1419                 else
1420                 {
1421                         if (limit_fraction >= 1.0)
1422                         {
1423                                 /* caller fractional, limit absolute; use caller's value */
1424                         }
1425                         else
1426                         {
1427                                 /* both fractional, so add them together */
1428                                 tuple_fraction += limit_fraction;
1429                                 if (tuple_fraction >= 1.0)
1430                                         tuple_fraction = 0.0;           /* assume fetch all */
1431                         }
1432                 }
1433         }
1434
1435         return tuple_fraction;
1436 }
1437
1438 /*
1439  * extract_grouping_ops - make an array of the equality operator OIDs
1440  *              for the GROUP BY clause
1441  */
1442 static Oid *
1443 extract_grouping_ops(List *groupClause)
1444 {
1445         int                     numCols = list_length(groupClause);
1446         int                     colno = 0;
1447         Oid                *groupOperators;
1448         ListCell   *glitem;
1449
1450         groupOperators = (Oid *) palloc(sizeof(Oid) * numCols);
1451
1452         foreach(glitem, groupClause)
1453         {
1454                 GroupClause *groupcl = (GroupClause *) lfirst(glitem);
1455
1456                 groupOperators[colno] = get_equality_op_for_ordering_op(groupcl->sortop);
1457                 if (!OidIsValid(groupOperators[colno])) /* shouldn't happen */
1458                         elog(ERROR, "could not find equality operator for ordering operator %u",
1459                                  groupcl->sortop);
1460                 colno++;
1461         }
1462
1463         return groupOperators;
1464 }
1465
1466 /*
1467  * choose_hashed_grouping - should we use hashed grouping?
1468  */
1469 static bool
1470 choose_hashed_grouping(PlannerInfo *root,
1471                                            double tuple_fraction, double limit_tuples,
1472                                            Path *cheapest_path, Path *sorted_path,
1473                                            Oid *groupOperators, double dNumGroups,
1474                                            AggClauseCounts *agg_counts)
1475 {
1476         int                     numGroupCols = list_length(root->parse->groupClause);
1477         double          cheapest_path_rows;
1478         int                     cheapest_path_width;
1479         Size            hashentrysize;
1480         List       *current_pathkeys;
1481         Path            hashed_p;
1482         Path            sorted_p;
1483         int                     i;
1484
1485         /*
1486          * Check can't-do-it conditions, including whether the grouping operators
1487          * are hashjoinable.  (We assume hashing is OK if they are marked
1488          * oprcanhash.  If there isn't actually a supporting hash function, the
1489          * executor will complain at runtime.)
1490          *
1491          * Executor doesn't support hashed aggregation with DISTINCT aggregates.
1492          * (Doing so would imply storing *all* the input values in the hash table,
1493          * which seems like a certain loser.)
1494          */
1495         if (!enable_hashagg)
1496                 return false;
1497         if (agg_counts->numDistinctAggs != 0)
1498                 return false;
1499         for (i = 0; i < numGroupCols; i++)
1500         {
1501                 if (!op_hashjoinable(groupOperators[i]))
1502                         return false;
1503         }
1504
1505         /*
1506          * Don't do it if it doesn't look like the hashtable will fit into
1507          * work_mem.
1508          *
1509          * Beware here of the possibility that cheapest_path->parent is NULL. This
1510          * could happen if user does something silly like SELECT 'foo' GROUP BY 1;
1511          */
1512         if (cheapest_path->parent)
1513         {
1514                 cheapest_path_rows = cheapest_path->parent->rows;
1515                 cheapest_path_width = cheapest_path->parent->width;
1516         }
1517         else
1518         {
1519                 cheapest_path_rows = 1; /* assume non-set result */
1520                 cheapest_path_width = 100;              /* arbitrary */
1521         }
1522
1523         /* Estimate per-hash-entry space at tuple width... */
1524         hashentrysize = MAXALIGN(cheapest_path_width) + MAXALIGN(sizeof(MinimalTupleData));
1525         /* plus space for pass-by-ref transition values... */
1526         hashentrysize += agg_counts->transitionSpace;
1527         /* plus the per-hash-entry overhead */
1528         hashentrysize += hash_agg_entry_size(agg_counts->numAggs);
1529
1530         if (hashentrysize * dNumGroups > work_mem * 1024L)
1531                 return false;
1532
1533         /*
1534          * See if the estimated cost is no more than doing it the other way. While
1535          * avoiding the need for sorted input is usually a win, the fact that the
1536          * output won't be sorted may be a loss; so we need to do an actual cost
1537          * comparison.
1538          *
1539          * We need to consider cheapest_path + hashagg [+ final sort] versus
1540          * either cheapest_path [+ sort] + group or agg [+ final sort] or
1541          * presorted_path + group or agg [+ final sort] where brackets indicate a
1542          * step that may not be needed. We assume query_planner() will have
1543          * returned a presorted path only if it's a winner compared to
1544          * cheapest_path for this purpose.
1545          *
1546          * These path variables are dummies that just hold cost fields; we don't
1547          * make actual Paths for these steps.
1548          */
1549         cost_agg(&hashed_p, root, AGG_HASHED, agg_counts->numAggs,
1550                          numGroupCols, dNumGroups,
1551                          cheapest_path->startup_cost, cheapest_path->total_cost,
1552                          cheapest_path_rows);
1553         /* Result of hashed agg is always unsorted */
1554         if (root->sort_pathkeys)
1555                 cost_sort(&hashed_p, root, root->sort_pathkeys, hashed_p.total_cost,
1556                                   dNumGroups, cheapest_path_width, limit_tuples);
1557
1558         if (sorted_path)
1559         {
1560                 sorted_p.startup_cost = sorted_path->startup_cost;
1561                 sorted_p.total_cost = sorted_path->total_cost;
1562                 current_pathkeys = sorted_path->pathkeys;
1563         }
1564         else
1565         {
1566                 sorted_p.startup_cost = cheapest_path->startup_cost;
1567                 sorted_p.total_cost = cheapest_path->total_cost;
1568                 current_pathkeys = cheapest_path->pathkeys;
1569         }
1570         if (!pathkeys_contained_in(root->group_pathkeys, current_pathkeys))
1571         {
1572                 cost_sort(&sorted_p, root, root->group_pathkeys, sorted_p.total_cost,
1573                                   cheapest_path_rows, cheapest_path_width, -1.0);
1574                 current_pathkeys = root->group_pathkeys;
1575         }
1576
1577         if (root->parse->hasAggs)
1578                 cost_agg(&sorted_p, root, AGG_SORTED, agg_counts->numAggs,
1579                                  numGroupCols, dNumGroups,
1580                                  sorted_p.startup_cost, sorted_p.total_cost,
1581                                  cheapest_path_rows);
1582         else
1583                 cost_group(&sorted_p, root, numGroupCols, dNumGroups,
1584                                    sorted_p.startup_cost, sorted_p.total_cost,
1585                                    cheapest_path_rows);
1586         /* The Agg or Group node will preserve ordering */
1587         if (root->sort_pathkeys &&
1588                 !pathkeys_contained_in(root->sort_pathkeys, current_pathkeys))
1589                 cost_sort(&sorted_p, root, root->sort_pathkeys, sorted_p.total_cost,
1590                                   dNumGroups, cheapest_path_width, limit_tuples);
1591
1592         /*
1593          * Now make the decision using the top-level tuple fraction.  First we
1594          * have to convert an absolute count (LIMIT) into fractional form.
1595          */
1596         if (tuple_fraction >= 1.0)
1597                 tuple_fraction /= dNumGroups;
1598
1599         if (compare_fractional_path_costs(&hashed_p, &sorted_p,
1600                                                                           tuple_fraction) < 0)
1601         {
1602                 /* Hashed is cheaper, so use it */
1603                 return true;
1604         }
1605         return false;
1606 }
1607
1608 /*---------------
1609  * make_subplanTargetList
1610  *        Generate appropriate target list when grouping is required.
1611  *
1612  * When grouping_planner inserts Aggregate, Group, or Result plan nodes
1613  * above the result of query_planner, we typically want to pass a different
1614  * target list to query_planner than the outer plan nodes should have.
1615  * This routine generates the correct target list for the subplan.
1616  *
1617  * The initial target list passed from the parser already contains entries
1618  * for all ORDER BY and GROUP BY expressions, but it will not have entries
1619  * for variables used only in HAVING clauses; so we need to add those
1620  * variables to the subplan target list.  Also, we flatten all expressions
1621  * except GROUP BY items into their component variables; the other expressions
1622  * will be computed by the inserted nodes rather than by the subplan.
1623  * For example, given a query like
1624  *              SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
1625  * we want to pass this targetlist to the subplan:
1626  *              a,b,c,d,a+b
1627  * where the a+b target will be used by the Sort/Group steps, and the
1628  * other targets will be used for computing the final results.  (In the
1629  * above example we could theoretically suppress the a and b targets and
1630  * pass down only c,d,a+b, but it's not really worth the trouble to
1631  * eliminate simple var references from the subplan.  We will avoid doing
1632  * the extra computation to recompute a+b at the outer level; see
1633  * fix_upper_expr() in setrefs.c.)
1634  *
1635  * If we are grouping or aggregating, *and* there are no non-Var grouping
1636  * expressions, then the returned tlist is effectively dummy; we do not
1637  * need to force it to be evaluated, because all the Vars it contains
1638  * should be present in the output of query_planner anyway.
1639  *
1640  * 'tlist' is the query's target list.
1641  * 'groupColIdx' receives an array of column numbers for the GROUP BY
1642  *                      expressions (if there are any) in the subplan's target list.
1643  * 'need_tlist_eval' is set true if we really need to evaluate the
1644  *                      result tlist.
1645  *
1646  * The result is the targetlist to be passed to the subplan.
1647  *---------------
1648  */
1649 static List *
1650 make_subplanTargetList(PlannerInfo *root,
1651                                            List *tlist,
1652                                            AttrNumber **groupColIdx,
1653                                            bool *need_tlist_eval)
1654 {
1655         Query      *parse = root->parse;
1656         List       *sub_tlist;
1657         List       *extravars;
1658         int                     numCols;
1659
1660         *groupColIdx = NULL;
1661
1662         /*
1663          * If we're not grouping or aggregating, there's nothing to do here;
1664          * query_planner should receive the unmodified target list.
1665          */
1666         if (!parse->hasAggs && !parse->groupClause && !root->hasHavingQual)
1667         {
1668                 *need_tlist_eval = true;
1669                 return tlist;
1670         }
1671
1672         /*
1673          * Otherwise, start with a "flattened" tlist (having just the vars
1674          * mentioned in the targetlist and HAVING qual --- but not upper- level
1675          * Vars; they will be replaced by Params later on).
1676          */
1677         sub_tlist = flatten_tlist(tlist);
1678         extravars = pull_var_clause(parse->havingQual, false);
1679         sub_tlist = add_to_flat_tlist(sub_tlist, extravars);
1680         list_free(extravars);
1681         *need_tlist_eval = false;       /* only eval if not flat tlist */
1682
1683         /*
1684          * If grouping, create sub_tlist entries for all GROUP BY expressions
1685          * (GROUP BY items that are simple Vars should be in the list already),
1686          * and make an array showing where the group columns are in the sub_tlist.
1687          */
1688         numCols = list_length(parse->groupClause);
1689         if (numCols > 0)
1690         {
1691                 int                     keyno = 0;
1692                 AttrNumber *grpColIdx;
1693                 ListCell   *gl;
1694
1695                 grpColIdx = (AttrNumber *) palloc(sizeof(AttrNumber) * numCols);
1696                 *groupColIdx = grpColIdx;
1697
1698                 foreach(gl, parse->groupClause)
1699                 {
1700                         GroupClause *grpcl = (GroupClause *) lfirst(gl);
1701                         Node       *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1702                         TargetEntry *te = NULL;
1703                         ListCell   *sl;
1704
1705                         /* Find or make a matching sub_tlist entry */
1706                         foreach(sl, sub_tlist)
1707                         {
1708                                 te = (TargetEntry *) lfirst(sl);
1709                                 if (equal(groupexpr, te->expr))
1710                                         break;
1711                         }
1712                         if (!sl)
1713                         {
1714                                 te = makeTargetEntry((Expr *) groupexpr,
1715                                                                          list_length(sub_tlist) + 1,
1716                                                                          NULL,
1717                                                                          false);
1718                                 sub_tlist = lappend(sub_tlist, te);
1719                                 *need_tlist_eval = true;                /* it's not flat anymore */
1720                         }
1721
1722                         /* and save its resno */
1723                         grpColIdx[keyno++] = te->resno;
1724                 }
1725         }
1726
1727         return sub_tlist;
1728 }
1729
1730 /*
1731  * locate_grouping_columns
1732  *              Locate grouping columns in the tlist chosen by query_planner.
1733  *
1734  * This is only needed if we don't use the sub_tlist chosen by
1735  * make_subplanTargetList.      We have to forget the column indexes found
1736  * by that routine and re-locate the grouping vars in the real sub_tlist.
1737  */
1738 static void
1739 locate_grouping_columns(PlannerInfo *root,
1740                                                 List *tlist,
1741                                                 List *sub_tlist,
1742                                                 AttrNumber *groupColIdx)
1743 {
1744         int                     keyno = 0;
1745         ListCell   *gl;
1746
1747         /*
1748          * No work unless grouping.
1749          */
1750         if (!root->parse->groupClause)
1751         {
1752                 Assert(groupColIdx == NULL);
1753                 return;
1754         }
1755         Assert(groupColIdx != NULL);
1756
1757         foreach(gl, root->parse->groupClause)
1758         {
1759                 GroupClause *grpcl = (GroupClause *) lfirst(gl);
1760                 Node       *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
1761                 TargetEntry *te = NULL;
1762                 ListCell   *sl;
1763
1764                 foreach(sl, sub_tlist)
1765                 {
1766                         te = (TargetEntry *) lfirst(sl);
1767                         if (equal(groupexpr, te->expr))
1768                                 break;
1769                 }
1770                 if (!sl)
1771                         elog(ERROR, "failed to locate grouping columns");
1772
1773                 groupColIdx[keyno++] = te->resno;
1774         }
1775 }
1776
1777 /*
1778  * postprocess_setop_tlist
1779  *        Fix up targetlist returned by plan_set_operations().
1780  *
1781  * We need to transpose sort key info from the orig_tlist into new_tlist.
1782  * NOTE: this would not be good enough if we supported resjunk sort keys
1783  * for results of set operations --- then, we'd need to project a whole
1784  * new tlist to evaluate the resjunk columns.  For now, just ereport if we
1785  * find any resjunk columns in orig_tlist.
1786  */
1787 static List *
1788 postprocess_setop_tlist(List *new_tlist, List *orig_tlist)
1789 {
1790         ListCell   *l;
1791         ListCell   *orig_tlist_item = list_head(orig_tlist);
1792
1793         foreach(l, new_tlist)
1794         {
1795                 TargetEntry *new_tle = (TargetEntry *) lfirst(l);
1796                 TargetEntry *orig_tle;
1797
1798                 /* ignore resjunk columns in setop result */
1799                 if (new_tle->resjunk)
1800                         continue;
1801
1802                 Assert(orig_tlist_item != NULL);
1803                 orig_tle = (TargetEntry *) lfirst(orig_tlist_item);
1804                 orig_tlist_item = lnext(orig_tlist_item);
1805                 if (orig_tle->resjunk)  /* should not happen */
1806                         elog(ERROR, "resjunk output columns are not implemented");
1807                 Assert(new_tle->resno == orig_tle->resno);
1808                 new_tle->ressortgroupref = orig_tle->ressortgroupref;
1809         }
1810         if (orig_tlist_item != NULL)
1811                 elog(ERROR, "resjunk output columns are not implemented");
1812         return new_tlist;
1813 }