granicus.if.org Git - postgresql/blob - src/backend/parser/parse_clause.c

   1 /*-------------------------------------------------------------------------
   2  *
   3  * parse_clause.c
   4  *        handle clauses in parser
   5  *
   6  * Portions Copyright (c) 1996-2006, PostgreSQL Global Development Group
   7  * Portions Copyright (c) 1994, Regents of the University of California
   8  *
   9  *
  10  * IDENTIFICATION
  11  *        $PostgreSQL: pgsql/src/backend/parser/parse_clause.c,v 1.154 2006/07/26 00:34:48 momjian Exp $
  12  *
  13  *-------------------------------------------------------------------------
  14  */
  15
  16 #include "postgres.h"
  17
  18 #include "access/heapam.h"
  19 #include "catalog/heap.h"
  20 #include "catalog/pg_type.h"
  21 #include "commands/defrem.h"
  22 #include "nodes/makefuncs.h"
  23 #include "optimizer/clauses.h"
  24 #include "optimizer/tlist.h"
  25 #include "optimizer/var.h"
  26 #include "parser/analyze.h"
  27 #include "parser/parsetree.h"
  28 #include "parser/parse_clause.h"
  29 #include "parser/parse_coerce.h"
  30 #include "parser/parse_expr.h"
  31 #include "parser/parse_oper.h"
  32 #include "parser/parse_relation.h"
  33 #include "parser/parse_target.h"
  34 #include "rewrite/rewriteManip.h"
  35 #include "utils/guc.h"
  36
  37
  38 #define ORDER_CLAUSE 0
  39 #define GROUP_CLAUSE 1
  40 #define DISTINCT_ON_CLAUSE 2
  41
  42 static char *clauseText[] = {"ORDER BY", "GROUP BY", "DISTINCT ON"};
  43
  44 static void extractRemainingColumns(List *common_colnames,
  45                                                 List *src_colnames, List *src_colvars,
  46                                                 List **res_colnames, List **res_colvars);
  47 static Node *transformJoinUsingClause(ParseState *pstate,
  48                                                  List *leftVars, List *rightVars);
  49 static Node *transformJoinOnClause(ParseState *pstate, JoinExpr *j,
  50                                           RangeTblEntry *l_rte,
  51                                           RangeTblEntry *r_rte,
  52                                           List *relnamespace,
  53                                           Relids containedRels);
  54 static RangeTblEntry *transformTableEntry(ParseState *pstate, RangeVar *r);
  55 static RangeTblEntry *transformRangeSubselect(ParseState *pstate,
  56                                                 RangeSubselect *r);
  57 static RangeTblEntry *transformRangeFunction(ParseState *pstate,
  58                                            RangeFunction *r);
  59 static Node *transformFromClauseItem(ParseState *pstate, Node *n,
  60                                                 RangeTblEntry **top_rte, int *top_rti,
  61                                                 List **relnamespace,
  62                                                 Relids *containedRels);
  63 static Node *buildMergedJoinVar(ParseState *pstate, JoinType jointype,
  64                                    Var *l_colvar, Var *r_colvar);
  65 static TargetEntry *findTargetlistEntry(ParseState *pstate, Node *node,
  66                                         List **tlist, int clause);
  67
  68
  69 /*
  70  * transformFromClause -
  71  *        Process the FROM clause and add items to the query's range table,
  72  *        joinlist, and namespaces.
  73  *
  74  * Note: we assume that pstate's p_rtable, p_joinlist, p_relnamespace, and
  75  * p_varnamespace lists were initialized to NIL when the pstate was created.
  76  * We will add onto any entries already present --- this is needed for rule
  77  * processing, as well as for UPDATE and DELETE.
  78  *
  79  * The range table may grow still further when we transform the expressions
  80  * in the query's quals and target list. (This is possible because in
  81  * POSTQUEL, we allowed references to relations not specified in the
  82  * from-clause.  PostgreSQL keeps this extension to standard SQL.)
  83  */
  84 void
  85 transformFromClause(ParseState *pstate, List *frmList)
  86 {
  87         ListCell   *fl;
  88
  89         /*
  90          * The grammar will have produced a list of RangeVars, RangeSubselects,
  91          * RangeFunctions, and/or JoinExprs. Transform each one (possibly adding
  92          * entries to the rtable), check for duplicate refnames, and then add it
  93          * to the joinlist and namespaces.
  94          */
  95         foreach(fl, frmList)
  96         {
  97                 Node       *n = lfirst(fl);
  98                 RangeTblEntry *rte;
  99                 int                     rtindex;
 100                 List       *relnamespace;
 101                 Relids          containedRels;
 102
 103                 n = transformFromClauseItem(pstate, n,
 104                                                                         &rte,
 105                                                                         &rtindex,
 106                                                                         &relnamespace,
 107                                                                         &containedRels);
 108                 checkNameSpaceConflicts(pstate, pstate->p_relnamespace, relnamespace);
 109                 pstate->p_joinlist = lappend(pstate->p_joinlist, n);
 110                 pstate->p_relnamespace = list_concat(pstate->p_relnamespace,
 111                                                                                          relnamespace);
 112                 pstate->p_varnamespace = lappend(pstate->p_varnamespace, rte);
 113                 bms_free(containedRels);
 114         }
 115 }
 116
 117 /*
 118  * setTargetTable
 119  *        Add the target relation of INSERT/UPDATE/DELETE to the range table,
 120  *        and make the special links to it in the ParseState.
 121  *
 122  *        We also open the target relation and acquire a write lock on it.
 123  *        This must be done before processing the FROM list, in case the target
 124  *        is also mentioned as a source relation --- we want to be sure to grab
 125  *        the write lock before any read lock.
 126  *
 127  *        If alsoSource is true, add the target to the query's joinlist and
 128  *        namespace.  For INSERT, we don't want the target to be joined to;
 129  *        it's a destination of tuples, not a source.   For UPDATE/DELETE,
 130  *        we do need to scan or join the target.  (NOTE: we do not bother
 131  *        to check for namespace conflict; we assume that the namespace was
 132  *        initially empty in these cases.)
 133  *
 134  *        Finally, we mark the relation as requiring the permissions specified
 135  *        by requiredPerms.
 136  *
 137  *        Returns the rangetable index of the target relation.
 138  */
 139 int
 140 setTargetTable(ParseState *pstate, RangeVar *relation,
 141                            bool inh, bool alsoSource, AclMode requiredPerms)
 142 {
 143         RangeTblEntry *rte;
 144         int                     rtindex;
 145
 146         /* Close old target; this could only happen for multi-action rules */
 147         if (pstate->p_target_relation != NULL)
 148                 heap_close(pstate->p_target_relation, NoLock);
 149
 150         /*
 151          * Open target rel and grab suitable lock (which we will hold till end of
 152          * transaction).
 153          *
 154          * analyze.c will eventually do the corresponding heap_close(), but *not*
 155          * release the lock.
 156          */
 157         pstate->p_target_relation = heap_openrv(relation, RowExclusiveLock);
 158
 159         /*
 160          * Now build an RTE.
 161          */
 162         rte = addRangeTableEntryForRelation(pstate, pstate->p_target_relation,
 163                                                                                 relation->alias, inh, false);
 164         pstate->p_target_rangetblentry = rte;
 165
 166         /* assume new rte is at end */
 167         rtindex = list_length(pstate->p_rtable);
 168         Assert(rte == rt_fetch(rtindex, pstate->p_rtable));
 169
 170         /*
 171          * Override addRangeTableEntry's default ACL_SELECT permissions check, and
 172          * instead mark target table as requiring exactly the specified
 173          * permissions.
 174          *
 175          * If we find an explicit reference to the rel later during parse
 176          * analysis, scanRTEForColumn will add the ACL_SELECT bit back again. That
 177          * can't happen for INSERT but it is possible for UPDATE and DELETE.
 178          */
 179         rte->requiredPerms = requiredPerms;
 180
 181         /*
 182          * If UPDATE/DELETE, add table to joinlist and namespaces.
 183          */
 184         if (alsoSource)
 185                 addRTEtoQuery(pstate, rte, true, true, true);
 186
 187         return rtindex;
 188 }
 189
 190 /*
 191  * Simplify InhOption (yes/no/default) into boolean yes/no.
 192  *
 193  * The reason we do things this way is that we don't want to examine the
 194  * SQL_inheritance option flag until parse_analyze is run.      Otherwise,
 195  * we'd do the wrong thing with query strings that intermix SET commands
 196  * with queries.
 197  */
 198 bool
 199 interpretInhOption(InhOption inhOpt)
 200 {
 201         switch (inhOpt)
 202         {
 203                 case INH_NO:
 204                         return false;
 205                 case INH_YES:
 206                         return true;
 207                 case INH_DEFAULT:
 208                         return SQL_inheritance;
 209         }
 210         elog(ERROR, "bogus InhOption value: %d", inhOpt);
 211         return false;                           /* keep compiler quiet */
 212 }
 213
 214 /*
 215  * Given a relation-options list (of DefElems), return true iff the specified
 216  * table/result set should be created with OIDs. This needs to be done after
 217  * parsing the query string because the return value can depend upon the
 218  * default_with_oids GUC var.
 219  */
 220 bool
 221 interpretOidsOption(List *defList)
 222 {
 223         ListCell   *cell;
 224
 225         /* Scan list to see if OIDS was included */
 226         foreach(cell, defList)
 227         {
 228                 DefElem    *def = (DefElem *) lfirst(cell);
 229
 230                 if (pg_strcasecmp(def->defname, "oids") == 0)
 231                         return defGetBoolean(def);
 232         }
 233
 234         /* OIDS option was not specified, so use default. */
 235         return default_with_oids;
 236 }
 237
 238 /*
 239  * Extract all not-in-common columns from column lists of a source table
 240  */
 241 static void
 242 extractRemainingColumns(List *common_colnames,
 243                                                 List *src_colnames, List *src_colvars,
 244                                                 List **res_colnames, List **res_colvars)
 245 {
 246         List       *new_colnames = NIL;
 247         List       *new_colvars = NIL;
 248         ListCell   *lnames,
 249                            *lvars;
 250
 251         Assert(list_length(src_colnames) == list_length(src_colvars));
 252
 253         forboth(lnames, src_colnames, lvars, src_colvars)
 254         {
 255                 char       *colname = strVal(lfirst(lnames));
 256                 bool            match = false;
 257                 ListCell   *cnames;
 258
 259                 foreach(cnames, common_colnames)
 260                 {
 261                         char       *ccolname = strVal(lfirst(cnames));
 262
 263                         if (strcmp(colname, ccolname) == 0)
 264                         {
 265                                 match = true;
 266                                 break;
 267                         }
 268                 }
 269
 270                 if (!match)
 271                 {
 272                         new_colnames = lappend(new_colnames, lfirst(lnames));
 273                         new_colvars = lappend(new_colvars, lfirst(lvars));
 274                 }
 275         }
 276
 277         *res_colnames = new_colnames;
 278         *res_colvars = new_colvars;
 279 }
 280
 281 /* transformJoinUsingClause()
 282  *        Build a complete ON clause from a partially-transformed USING list.
 283  *        We are given lists of nodes representing left and right match columns.
 284  *        Result is a transformed qualification expression.
 285  */
 286 static Node *
 287 transformJoinUsingClause(ParseState *pstate, List *leftVars, List *rightVars)
 288 {
 289         Node       *result = NULL;
 290         ListCell   *lvars,
 291                            *rvars;
 292
 293         /*
 294          * We cheat a little bit here by building an untransformed operator tree
 295          * whose leaves are the already-transformed Vars.  This is OK because
 296          * transformExpr() won't complain about already-transformed subnodes.
 297          */
 298         forboth(lvars, leftVars, rvars, rightVars)
 299         {
 300                 Node       *lvar = (Node *) lfirst(lvars);
 301                 Node       *rvar = (Node *) lfirst(rvars);
 302                 A_Expr     *e;
 303
 304                 e = makeSimpleA_Expr(AEXPR_OP, "=",
 305                                                          copyObject(lvar), copyObject(rvar),
 306                                                          -1);
 307
 308                 if (result == NULL)
 309                         result = (Node *) e;
 310                 else
 311                 {
 312                         A_Expr     *a;
 313
 314                         a = makeA_Expr(AEXPR_AND, NIL, result, (Node *) e, -1);
 315                         result = (Node *) a;
 316                 }
 317         }
 318
 319         /*
 320          * Since the references are already Vars, and are certainly from the input
 321          * relations, we don't have to go through the same pushups that
 322          * transformJoinOnClause() does.  Just invoke transformExpr() to fix up
 323          * the operators, and we're done.
 324          */
 325         result = transformExpr(pstate, result);
 326
 327         result = coerce_to_boolean(pstate, result, "JOIN/USING");
 328
 329         return result;
 330 }
 331
 332 /* transformJoinOnClause()
 333  *        Transform the qual conditions for JOIN/ON.
 334  *        Result is a transformed qualification expression.
 335  */
 336 static Node *
 337 transformJoinOnClause(ParseState *pstate, JoinExpr *j,
 338                                           RangeTblEntry *l_rte,
 339                                           RangeTblEntry *r_rte,
 340                                           List *relnamespace,
 341                                           Relids containedRels)
 342 {
 343         Node       *result;
 344         List       *save_relnamespace;
 345         List       *save_varnamespace;
 346         Relids          clause_varnos;
 347         int                     varno;
 348
 349         /*
 350          * This is a tad tricky, for two reasons.  First, the namespace that the
 351          * join expression should see is just the two subtrees of the JOIN plus
 352          * any outer references from upper pstate levels.  So, temporarily set
 353          * this pstate's namespace accordingly.  (We need not check for refname
 354          * conflicts, because transformFromClauseItem() already did.) NOTE: this
 355          * code is OK only because the ON clause can't legally alter the namespace
 356          * by causing implicit relation refs to be added.
 357          */
 358         save_relnamespace = pstate->p_relnamespace;
 359         save_varnamespace = pstate->p_varnamespace;
 360
 361         pstate->p_relnamespace = relnamespace;
 362         pstate->p_varnamespace = list_make2(l_rte, r_rte);
 363
 364         result = transformWhereClause(pstate, j->quals, "JOIN/ON");
 365
 366         pstate->p_relnamespace = save_relnamespace;
 367         pstate->p_varnamespace = save_varnamespace;
 368
 369         /*
 370          * Second, we need to check that the ON condition doesn't refer to any
 371          * rels outside the input subtrees of the JOIN.  It could do that despite
 372          * our hack on the namespace if it uses fully-qualified names. So, grovel
 373          * through the transformed clause and make sure there are no bogus
 374          * references.  (Outer references are OK, and are ignored here.)
 375          */
 376         clause_varnos = pull_varnos(result);
 377         clause_varnos = bms_del_members(clause_varnos, containedRels);
 378         if ((varno = bms_first_member(clause_varnos)) >= 0)
 379         {
 380                 ereport(ERROR,
 381                                 (errcode(ERRCODE_INVALID_COLUMN_REFERENCE),
 382                  errmsg("JOIN/ON clause refers to \"%s\", which is not part of JOIN",
 383                                 rt_fetch(varno, pstate->p_rtable)->eref->aliasname)));
 384         }
 385         bms_free(clause_varnos);
 386
 387         return result;
 388 }
 389
 390 /*
 391  * transformTableEntry --- transform a RangeVar (simple relation reference)
 392  */
 393 static RangeTblEntry *
 394 transformTableEntry(ParseState *pstate, RangeVar *r)
 395 {
 396         RangeTblEntry *rte;
 397
 398         /*
 399          * mark this entry to indicate it comes from the FROM clause. In SQL, the
 400          * target list can only refer to range variables specified in the from
 401          * clause but we follow the more powerful POSTQUEL semantics and
 402          * automatically generate the range variable if not specified. However
 403          * there are times we need to know whether the entries are legitimate.
 404          */
 405         rte = addRangeTableEntry(pstate, r, r->alias,
 406                                                          interpretInhOption(r->inhOpt), true);
 407
 408         return rte;
 409 }
 410
 411
 412 /*
 413  * transformRangeSubselect --- transform a sub-SELECT appearing in FROM
 414  */
 415 static RangeTblEntry *
 416 transformRangeSubselect(ParseState *pstate, RangeSubselect *r)
 417 {
 418         List       *parsetrees;
 419         Query      *query;
 420         RangeTblEntry *rte;
 421
 422         /*
 423          * We require user to supply an alias for a subselect, per SQL92. To relax
 424          * this, we'd have to be prepared to gin up a unique alias for an
 425          * unlabeled subselect.
 426          */
 427         if (r->alias == NULL)
 428                 ereport(ERROR,
 429                                 (errcode(ERRCODE_SYNTAX_ERROR),
 430                                  errmsg("subquery in FROM must have an alias")));
 431
 432         /*
 433          * Analyze and transform the subquery.
 434          */
 435         parsetrees = parse_sub_analyze(r->subquery, pstate);
 436
 437         /*
 438          * Check that we got something reasonable.      Most of these conditions are
 439          * probably impossible given restrictions of the grammar, but check 'em
 440          * anyway.
 441          */
 442         if (list_length(parsetrees) != 1)
 443                 elog(ERROR, "unexpected parse analysis result for subquery in FROM");
 444         query = (Query *) linitial(parsetrees);
 445         if (query == NULL || !IsA(query, Query))
 446                 elog(ERROR, "unexpected parse analysis result for subquery in FROM");
 447
 448         if (query->commandType != CMD_SELECT)
 449                 elog(ERROR, "expected SELECT query from subquery in FROM");
 450         if (query->resultRelation != 0 || query->into != NULL)
 451                 ereport(ERROR,
 452                                 (errcode(ERRCODE_SYNTAX_ERROR),
 453                                  errmsg("subquery in FROM may not have SELECT INTO")));
 454
 455         /*
 456          * The subquery cannot make use of any variables from FROM items created
 457          * earlier in the current query.  Per SQL92, the scope of a FROM item does
 458          * not include other FROM items.  Formerly we hacked the namespace so that
 459          * the other variables weren't even visible, but it seems more useful to
 460          * leave them visible and give a specific error message.
 461          *
 462          * XXX this will need further work to support SQL99's LATERAL() feature,
 463          * wherein such references would indeed be legal.
 464          *
 465          * We can skip groveling through the subquery if there's not anything
 466          * visible in the current query.  Also note that outer references are OK.
 467          */
 468         if (pstate->p_relnamespace || pstate->p_varnamespace)
 469         {
 470                 if (contain_vars_of_level((Node *) query, 1))
 471                         ereport(ERROR,
 472                                         (errcode(ERRCODE_INVALID_COLUMN_REFERENCE),
 473                                          errmsg("subquery in FROM may not refer to other relations of same query level")));
 474         }
 475
 476         /*
 477          * OK, build an RTE for the subquery.
 478          */
 479         rte = addRangeTableEntryForSubquery(pstate, query, r->alias, true);
 480
 481         return rte;
 482 }
 483
 484
 485 /*
 486  * transformRangeFunction --- transform a function call appearing in FROM
 487  */
 488 static RangeTblEntry *
 489 transformRangeFunction(ParseState *pstate, RangeFunction *r)
 490 {
 491         Node       *funcexpr;
 492         char       *funcname;
 493         RangeTblEntry *rte;
 494
 495         /*
 496          * Get function name for possible use as alias.  We use the same
 497          * transformation rules as for a SELECT output expression.      For a FuncCall
 498          * node, the result will be the function name, but it is possible for the
 499          * grammar to hand back other node types.
 500          */
 501         funcname = FigureColname(r->funccallnode);
 502
 503         /*
 504          * Transform the raw expression.
 505          */
 506         funcexpr = transformExpr(pstate, r->funccallnode);
 507
 508         /*
 509          * The function parameters cannot make use of any variables from other
 510          * FROM items.  (Compare to transformRangeSubselect(); the coding is
 511          * different though because we didn't parse as a sub-select with its own
 512          * level of namespace.)
 513          *
 514          * XXX this will need further work to support SQL99's LATERAL() feature,
 515          * wherein such references would indeed be legal.
 516          */
 517         if (pstate->p_relnamespace || pstate->p_varnamespace)
 518         {
 519                 if (contain_vars_of_level(funcexpr, 0))
 520                         ereport(ERROR,
 521                                         (errcode(ERRCODE_INVALID_COLUMN_REFERENCE),
 522                                          errmsg("function expression in FROM may not refer to other relations of same query level")));
 523         }
 524
 525         /*
 526          * Disallow aggregate functions in the expression.      (No reason to postpone
 527          * this check until parseCheckAggregates.)
 528          */
 529         if (pstate->p_hasAggs)
 530         {
 531                 if (checkExprHasAggs(funcexpr))
 532                         ereport(ERROR,
 533                                         (errcode(ERRCODE_GROUPING_ERROR),
 534                                          errmsg("cannot use aggregate function in function expression in FROM")));
 535         }
 536
 537         /*
 538          * OK, build an RTE for the function.
 539          */
 540         rte = addRangeTableEntryForFunction(pstate, funcname, funcexpr,
 541                                                                                 r, true);
 542
 543         /*
 544          * If a coldeflist was supplied, ensure it defines a legal set of names
 545          * (no duplicates) and datatypes (no pseudo-types, for instance).
 546          * addRangeTableEntryForFunction looked up the type names but didn't
 547          * check them further than that.
 548          */
 549         if (r->coldeflist)
 550         {
 551                 TupleDesc       tupdesc;
 552
 553                 tupdesc = BuildDescFromLists(rte->eref->colnames,
 554                                                                          rte->funccoltypes,
 555                                                                          rte->funccoltypmods);
 556                 CheckAttributeNamesTypes(tupdesc, RELKIND_COMPOSITE_TYPE);
 557         }
 558
 559         return rte;
 560 }
 561
 562
 563 /*
 564  * transformFromClauseItem -
 565  *        Transform a FROM-clause item, adding any required entries to the
 566  *        range table list being built in the ParseState, and return the
 567  *        transformed item ready to include in the joinlist and namespaces.
 568  *        This routine can recurse to handle SQL92 JOIN expressions.
 569  *
 570  * The function return value is the node to add to the jointree (a
 571  * RangeTblRef or JoinExpr).  Additional output parameters are:
 572  *
 573  * *top_rte: receives the RTE corresponding to the jointree item.
 574  * (We could extract this from the function return node, but it saves cycles
 575  * to pass it back separately.)
 576  *
 577  * *top_rti: receives the rangetable index of top_rte.  (Ditto.)
 578  *
 579  * *relnamespace: receives a List of the RTEs exposed as relation names
 580  * by this item.
 581  *
 582  * *containedRels: receives a bitmap set of the rangetable indexes
 583  * of all the base and join relations represented in this jointree item.
 584  * This is needed for checking JOIN/ON conditions in higher levels.
 585  *
 586  * We do not need to pass back an explicit varnamespace value, because
 587  * in all cases the varnamespace contribution is exactly top_rte.
 588  */
 589 static Node *
 590 transformFromClauseItem(ParseState *pstate, Node *n,
 591                                                 RangeTblEntry **top_rte, int *top_rti,
 592                                                 List **relnamespace,
 593                                                 Relids *containedRels)
 594 {
 595         if (IsA(n, RangeVar))
 596         {
 597                 /* Plain relation reference */
 598                 RangeTblRef *rtr;
 599                 RangeTblEntry *rte;
 600                 int                     rtindex;
 601
 602                 rte = transformTableEntry(pstate, (RangeVar *) n);
 603                 /* assume new rte is at end */
 604                 rtindex = list_length(pstate->p_rtable);
 605                 Assert(rte == rt_fetch(rtindex, pstate->p_rtable));
 606                 *top_rte = rte;
 607                 *top_rti = rtindex;
 608                 *relnamespace = list_make1(rte);
 609                 *containedRels = bms_make_singleton(rtindex);
 610                 rtr = makeNode(RangeTblRef);
 611                 rtr->rtindex = rtindex;
 612                 return (Node *) rtr;
 613         }
 614         else if (IsA(n, RangeSubselect))
 615         {
 616                 /* sub-SELECT is like a plain relation */
 617                 RangeTblRef *rtr;
 618                 RangeTblEntry *rte;
 619                 int                     rtindex;
 620
 621                 rte = transformRangeSubselect(pstate, (RangeSubselect *) n);
 622                 /* assume new rte is at end */
 623                 rtindex = list_length(pstate->p_rtable);
 624                 Assert(rte == rt_fetch(rtindex, pstate->p_rtable));
 625                 *top_rte = rte;
 626                 *top_rti = rtindex;
 627                 *relnamespace = list_make1(rte);
 628                 *containedRels = bms_make_singleton(rtindex);
 629                 rtr = makeNode(RangeTblRef);
 630                 rtr->rtindex = rtindex;
 631                 return (Node *) rtr;
 632         }
 633         else if (IsA(n, RangeFunction))
 634         {
 635                 /* function is like a plain relation */
 636                 RangeTblRef *rtr;
 637                 RangeTblEntry *rte;
 638                 int                     rtindex;
 639
 640                 rte = transformRangeFunction(pstate, (RangeFunction *) n);
 641                 /* assume new rte is at end */
 642                 rtindex = list_length(pstate->p_rtable);
 643                 Assert(rte == rt_fetch(rtindex, pstate->p_rtable));
 644                 *top_rte = rte;
 645                 *top_rti = rtindex;
 646                 *relnamespace = list_make1(rte);
 647                 *containedRels = bms_make_singleton(rtindex);
 648                 rtr = makeNode(RangeTblRef);
 649                 rtr->rtindex = rtindex;
 650                 return (Node *) rtr;
 651         }
 652         else if (IsA(n, JoinExpr))
 653         {
 654                 /* A newfangled join expression */
 655                 JoinExpr   *j = (JoinExpr *) n;
 656                 RangeTblEntry *l_rte;
 657                 RangeTblEntry *r_rte;
 658                 int                     l_rtindex;
 659                 int                     r_rtindex;
 660                 Relids          l_containedRels,
 661                                         r_containedRels,
 662                                         my_containedRels;
 663                 List       *l_relnamespace,
 664                                    *r_relnamespace,
 665                                    *my_relnamespace,
 666                                    *l_colnames,
 667                                    *r_colnames,
 668                                    *res_colnames,
 669                                    *l_colvars,
 670                                    *r_colvars,
 671                                    *res_colvars;
 672                 RangeTblEntry *rte;
 673
 674                 /*
 675                  * Recursively process the left and right subtrees
 676                  */
 677                 j->larg = transformFromClauseItem(pstate, j->larg,
 678                                                                                   &l_rte,
 679                                                                                   &l_rtindex,
 680                                                                                   &l_relnamespace,
 681                                                                                   &l_containedRels);
 682                 j->rarg = transformFromClauseItem(pstate, j->rarg,
 683                                                                                   &r_rte,
 684                                                                                   &r_rtindex,
 685                                                                                   &r_relnamespace,
 686                                                                                   &r_containedRels);
 687
 688                 /*
 689                  * Check for conflicting refnames in left and right subtrees. Must do
 690                  * this because higher levels will assume I hand back a self-
 691                  * consistent namespace subtree.
 692                  */
 693                 checkNameSpaceConflicts(pstate, l_relnamespace, r_relnamespace);
 694
 695                 /*
 696                  * Generate combined relation membership info for possible use by
 697                  * transformJoinOnClause below.
 698                  */
 699                 my_relnamespace = list_concat(l_relnamespace, r_relnamespace);
 700                 my_containedRels = bms_join(l_containedRels, r_containedRels);
 701
 702                 pfree(r_relnamespace);  /* free unneeded list header */
 703
 704                 /*
 705                  * Extract column name and var lists from both subtrees
 706                  *
 707                  * Note: expandRTE returns new lists, safe for me to modify
 708                  */
 709                 expandRTE(l_rte, l_rtindex, 0, false,
 710                                   &l_colnames, &l_colvars);
 711                 expandRTE(r_rte, r_rtindex, 0, false,
 712                                   &r_colnames, &r_colvars);
 713
 714                 /*
 715                  * Natural join does not explicitly specify columns; must generate
 716                  * columns to join. Need to run through the list of columns from each
 717                  * table or join result and match up the column names. Use the first
 718                  * table, and check every column in the second table for a match.
 719                  * (We'll check that the matches were unique later on.) The result of
 720                  * this step is a list of column names just like an explicitly-written
 721                  * USING list.
 722                  */
 723                 if (j->isNatural)
 724                 {
 725                         List       *rlist = NIL;
 726                         ListCell   *lx,
 727                                            *rx;
 728
 729                         Assert(j->using == NIL);        /* shouldn't have USING() too */
 730
 731                         foreach(lx, l_colnames)
 732                         {
 733                                 char       *l_colname = strVal(lfirst(lx));
 734                                 Value      *m_name = NULL;
 735
 736                                 foreach(rx, r_colnames)
 737                                 {
 738                                         char       *r_colname = strVal(lfirst(rx));
 739
 740                                         if (strcmp(l_colname, r_colname) == 0)
 741                                         {
 742                                                 m_name = makeString(l_colname);
 743                                                 break;
 744                                         }
 745                                 }
 746
 747                                 /* matched a right column? then keep as join column... */
 748                                 if (m_name != NULL)
 749                                         rlist = lappend(rlist, m_name);
 750                         }
 751
 752                         j->using = rlist;
 753                 }
 754
 755                 /*
 756                  * Now transform the join qualifications, if any.
 757                  */
 758                 res_colnames = NIL;
 759                 res_colvars = NIL;
 760
 761                 if (j->using)
 762                 {
 763                         /*
 764                          * JOIN/USING (or NATURAL JOIN, as transformed above). Transform
 765                          * the list into an explicit ON-condition, and generate a list of
 766                          * merged result columns.
 767                          */
 768                         List       *ucols = j->using;
 769                         List       *l_usingvars = NIL;
 770                         List       *r_usingvars = NIL;
 771                         ListCell   *ucol;
 772
 773                         Assert(j->quals == NULL);       /* shouldn't have ON() too */
 774
 775                         foreach(ucol, ucols)
 776                         {
 777                                 char       *u_colname = strVal(lfirst(ucol));
 778                                 ListCell   *col;
 779                                 int                     ndx;
 780                                 int                     l_index = -1;
 781                                 int                     r_index = -1;
 782                                 Var                *l_colvar,
 783                                                    *r_colvar;
 784
 785                                 /* Check for USING(foo,foo) */
 786                                 foreach(col, res_colnames)
 787                                 {
 788                                         char       *res_colname = strVal(lfirst(col));
 789
 790                                         if (strcmp(res_colname, u_colname) == 0)
 791                                                 ereport(ERROR,
 792                                                                 (errcode(ERRCODE_DUPLICATE_COLUMN),
 793                                                                  errmsg("column name \"%s\" appears more than once in USING clause",
 794                                                                                 u_colname)));
 795                                 }
 796
 797                                 /* Find it in left input */
 798                                 ndx = 0;
 799                                 foreach(col, l_colnames)
 800                                 {
 801                                         char       *l_colname = strVal(lfirst(col));
 802
 803                                         if (strcmp(l_colname, u_colname) == 0)
 804                                         {
 805                                                 if (l_index >= 0)
 806                                                         ereport(ERROR,
 807                                                                         (errcode(ERRCODE_AMBIGUOUS_COLUMN),
 808                                                                          errmsg("common column name \"%s\" appears more than once in left table",
 809                                                                                         u_colname)));
 810                                                 l_index = ndx;
 811                                         }
 812                                         ndx++;
 813                                 }
 814                                 if (l_index < 0)
 815                                         ereport(ERROR,
 816                                                         (errcode(ERRCODE_UNDEFINED_COLUMN),
 817                                                          errmsg("column \"%s\" specified in USING clause does not exist in left table",
 818                                                                         u_colname)));
 819
 820                                 /* Find it in right input */
 821                                 ndx = 0;
 822                                 foreach(col, r_colnames)
 823                                 {
 824                                         char       *r_colname = strVal(lfirst(col));
 825
 826                                         if (strcmp(r_colname, u_colname) == 0)
 827                                         {
 828                                                 if (r_index >= 0)
 829                                                         ereport(ERROR,
 830                                                                         (errcode(ERRCODE_AMBIGUOUS_COLUMN),
 831                                                                          errmsg("common column name \"%s\" appears more than once in right table",
 832                                                                                         u_colname)));
 833                                                 r_index = ndx;
 834                                         }
 835                                         ndx++;
 836                                 }
 837                                 if (r_index < 0)
 838                                         ereport(ERROR,
 839                                                         (errcode(ERRCODE_UNDEFINED_COLUMN),
 840                                                          errmsg("column \"%s\" specified in USING clause does not exist in right table",
 841                                                                         u_colname)));
 842
 843                                 l_colvar = list_nth(l_colvars, l_index);
 844                                 l_usingvars = lappend(l_usingvars, l_colvar);
 845                                 r_colvar = list_nth(r_colvars, r_index);
 846                                 r_usingvars = lappend(r_usingvars, r_colvar);
 847
 848                                 res_colnames = lappend(res_colnames, lfirst(ucol));
 849                                 res_colvars = lappend(res_colvars,
 850                                                                           buildMergedJoinVar(pstate,
 851                                                                                                                  j->jointype,
 852                                                                                                                  l_colvar,
 853                                                                                                                  r_colvar));
 854                         }
 855
 856                         j->quals = transformJoinUsingClause(pstate,
 857                                                                                                 l_usingvars,
 858                                                                                                 r_usingvars);
 859                 }
 860                 else if (j->quals)
 861                 {
 862                         /* User-written ON-condition; transform it */
 863                         j->quals = transformJoinOnClause(pstate, j,
 864                                                                                          l_rte, r_rte,
 865                                                                                          my_relnamespace,
 866                                                                                          my_containedRels);
 867                 }
 868                 else
 869                 {
 870                         /* CROSS JOIN: no quals */
 871                 }
 872
 873                 /* Add remaining columns from each side to the output columns */
 874                 extractRemainingColumns(res_colnames,
 875                                                                 l_colnames, l_colvars,
 876                                                                 &l_colnames, &l_colvars);
 877                 extractRemainingColumns(res_colnames,
 878                                                                 r_colnames, r_colvars,
 879                                                                 &r_colnames, &r_colvars);
 880                 res_colnames = list_concat(res_colnames, l_colnames);
 881                 res_colvars = list_concat(res_colvars, l_colvars);
 882                 res_colnames = list_concat(res_colnames, r_colnames);
 883                 res_colvars = list_concat(res_colvars, r_colvars);
 884
 885                 /*
 886                  * Check alias (AS clause), if any.
 887                  */
 888                 if (j->alias)
 889                 {
 890                         if (j->alias->colnames != NIL)
 891                         {
 892                                 if (list_length(j->alias->colnames) > list_length(res_colnames))
 893                                         ereport(ERROR,
 894                                                         (errcode(ERRCODE_SYNTAX_ERROR),
 895                                                          errmsg("column alias list for \"%s\" has too many entries",
 896                                                                         j->alias->aliasname)));
 897                         }
 898                 }
 899
 900                 /*
 901                  * Now build an RTE for the result of the join
 902                  */
 903                 rte = addRangeTableEntryForJoin(pstate,
 904                                                                                 res_colnames,
 905                                                                                 j->jointype,
 906                                                                                 res_colvars,
 907                                                                                 j->alias,
 908                                                                                 true);
 909
 910                 /* assume new rte is at end */
 911                 j->rtindex = list_length(pstate->p_rtable);
 912                 Assert(rte == rt_fetch(j->rtindex, pstate->p_rtable));
 913
 914                 *top_rte = rte;
 915                 *top_rti = j->rtindex;
 916
 917                 /*
 918                  * Prepare returned namespace list.  If the JOIN has an alias then it
 919                  * hides the contained RTEs as far as the relnamespace goes;
 920                  * otherwise, put the contained RTEs and *not* the JOIN into
 921                  * relnamespace.
 922                  */
 923                 if (j->alias)
 924                 {
 925                         *relnamespace = list_make1(rte);
 926                         list_free(my_relnamespace);
 927                 }
 928                 else
 929                         *relnamespace = my_relnamespace;
 930
 931                 /*
 932                  * Include join RTE in returned containedRels set
 933                  */
 934                 *containedRels = bms_add_member(my_containedRels, j->rtindex);
 935
 936                 return (Node *) j;
 937         }
 938         else
 939                 elog(ERROR, "unrecognized node type: %d", (int) nodeTag(n));
 940         return NULL;                            /* can't get here, keep compiler quiet */
 941 }
 942
 943 /*
 944  * buildMergedJoinVar -
 945  *        generate a suitable replacement expression for a merged join column
 946  */
 947 static Node *
 948 buildMergedJoinVar(ParseState *pstate, JoinType jointype,
 949                                    Var *l_colvar, Var *r_colvar)
 950 {
 951         Oid                     outcoltype;
 952         int32           outcoltypmod;
 953         Node       *l_node,
 954                            *r_node,
 955                            *res_node;
 956
 957         /*
 958          * Choose output type if input types are dissimilar.
 959          */
 960         outcoltype = l_colvar->vartype;
 961         outcoltypmod = l_colvar->vartypmod;
 962         if (outcoltype != r_colvar->vartype)
 963         {
 964                 outcoltype = select_common_type(list_make2_oid(l_colvar->vartype,
 965                                                                                                            r_colvar->vartype),
 966                                                                                 "JOIN/USING");
 967                 outcoltypmod = -1;              /* ie, unknown */
 968         }
 969         else if (outcoltypmod != r_colvar->vartypmod)
 970         {
 971                 /* same type, but not same typmod */
 972                 outcoltypmod = -1;              /* ie, unknown */
 973         }
 974
 975         /*
 976          * Insert coercion functions if needed.  Note that a difference in typmod
 977          * can only happen if input has typmod but outcoltypmod is -1. In that
 978          * case we insert a RelabelType to clearly mark that result's typmod is
 979          * not same as input.  We never need coerce_type_typmod.
 980          */
 981         if (l_colvar->vartype != outcoltype)
 982                 l_node = coerce_type(pstate, (Node *) l_colvar, l_colvar->vartype,
 983                                                          outcoltype, outcoltypmod,
 984                                                          COERCION_IMPLICIT, COERCE_IMPLICIT_CAST);
 985         else if (l_colvar->vartypmod != outcoltypmod)
 986                 l_node = (Node *) makeRelabelType((Expr *) l_colvar,
 987                                                                                   outcoltype, outcoltypmod,
 988                                                                                   COERCE_IMPLICIT_CAST);
 989         else
 990                 l_node = (Node *) l_colvar;
 991
 992         if (r_colvar->vartype != outcoltype)
 993                 r_node = coerce_type(pstate, (Node *) r_colvar, r_colvar->vartype,
 994                                                          outcoltype, outcoltypmod,
 995                                                          COERCION_IMPLICIT, COERCE_IMPLICIT_CAST);
 996         else if (r_colvar->vartypmod != outcoltypmod)
 997                 r_node = (Node *) makeRelabelType((Expr *) r_colvar,
 998                                                                                   outcoltype, outcoltypmod,
 999                                                                                   COERCE_IMPLICIT_CAST);
1000         else
1001                 r_node = (Node *) r_colvar;
1002
1003         /*
1004          * Choose what to emit
1005          */
1006         switch (jointype)
1007         {
1008                 case JOIN_INNER:
1009
1010                         /*
1011                          * We can use either var; prefer non-coerced one if available.
1012                          */
1013                         if (IsA(l_node, Var))
1014                                 res_node = l_node;
1015                         else if (IsA(r_node, Var))
1016                                 res_node = r_node;
1017                         else
1018                                 res_node = l_node;
1019                         break;
1020                 case JOIN_LEFT:
1021                         /* Always use left var */
1022                         res_node = l_node;
1023                         break;
1024                 case JOIN_RIGHT:
1025                         /* Always use right var */
1026                         res_node = r_node;
1027                         break;
1028                 case JOIN_FULL:
1029                         {
1030                                 /*
1031                                  * Here we must build a COALESCE expression to ensure that the
1032                                  * join output is non-null if either input is.
1033                                  */
1034                                 CoalesceExpr *c = makeNode(CoalesceExpr);
1035
1036                                 c->coalescetype = outcoltype;
1037                                 c->args = list_make2(l_node, r_node);
1038                                 res_node = (Node *) c;
1039                                 break;
1040                         }
1041                 default:
1042                         elog(ERROR, "unrecognized join type: %d", (int) jointype);
1043                         res_node = NULL;        /* keep compiler quiet */
1044                         break;
1045         }
1046
1047         return res_node;
1048 }
1049
1050
1051 /*
1052  * transformWhereClause -
1053  *        Transform the qualification and make sure it is of type boolean.
1054  *        Used for WHERE and allied clauses.
1055  *
1056  * constructName does not affect the semantics, but is used in error messages
1057  */
1058 Node *
1059 transformWhereClause(ParseState *pstate, Node *clause,
1060                                          const char *constructName)
1061 {
1062         Node       *qual;
1063
1064         if (clause == NULL)
1065                 return NULL;
1066
1067         qual = transformExpr(pstate, clause);
1068
1069         qual = coerce_to_boolean(pstate, qual, constructName);
1070
1071         return qual;
1072 }
1073
1074
1075 /*
1076  * transformLimitClause -
1077  *        Transform the expression and make sure it is of type integer.
1078  *        Used for LIMIT and allied clauses.
1079  *
1080  * constructName does not affect the semantics, but is used in error messages
1081  */
1082 Node *
1083 transformLimitClause(ParseState *pstate, Node *clause,
1084                                          const char *constructName)
1085 {
1086         Node       *qual;
1087
1088         if (clause == NULL)
1089                 return NULL;
1090
1091         qual = transformExpr(pstate, clause);
1092
1093         qual = coerce_to_integer64(pstate, qual, constructName);
1094
1095         /*
1096          * LIMIT can't refer to any vars or aggregates of the current query; we
1097          * don't allow subselects either (though that case would at least be
1098          * sensible)
1099          */
1100         if (contain_vars_of_level(qual, 0))
1101         {
1102                 ereport(ERROR,
1103                                 (errcode(ERRCODE_INVALID_COLUMN_REFERENCE),
1104                 /* translator: %s is name of a SQL construct, eg LIMIT */
1105                                  errmsg("argument of %s must not contain variables",
1106                                                 constructName)));
1107         }
1108         if (checkExprHasAggs(qual))
1109         {
1110                 ereport(ERROR,
1111                                 (errcode(ERRCODE_GROUPING_ERROR),
1112                 /* translator: %s is name of a SQL construct, eg LIMIT */
1113                                  errmsg("argument of %s must not contain aggregates",
1114                                                 constructName)));
1115         }
1116         if (contain_subplans(qual))
1117         {
1118                 ereport(ERROR,
1119                                 (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
1120                 /* translator: %s is name of a SQL construct, eg LIMIT */
1121                                  errmsg("argument of %s must not contain subqueries",
1122                                                 constructName)));
1123         }
1124
1125         return qual;
1126 }
1127
1128
1129 /*
1130  *      findTargetlistEntry -
1131  *        Returns the targetlist entry matching the given (untransformed) node.
1132  *        If no matching entry exists, one is created and appended to the target
1133  *        list as a "resjunk" node.
1134  *
1135  * node         the ORDER BY, GROUP BY, or DISTINCT ON expression to be matched
1136  * tlist        the target list (passed by reference so we can append to it)
1137  * clause       identifies clause type being processed
1138  */
1139 static TargetEntry *
1140 findTargetlistEntry(ParseState *pstate, Node *node, List **tlist, int clause)
1141 {
1142         TargetEntry *target_result = NULL;
1143         ListCell   *tl;
1144         Node       *expr;
1145
1146         /*----------
1147          * Handle two special cases as mandated by the SQL92 spec:
1148          *
1149          * 1. Bare ColumnName (no qualifier or subscripts)
1150          *        For a bare identifier, we search for a matching column name
1151          *        in the existing target list.  Multiple matches are an error
1152          *        unless they refer to identical values; for example,
1153          *        we allow      SELECT a, a FROM table ORDER BY a
1154          *        but not       SELECT a AS b, b FROM table ORDER BY b
1155          *        If no match is found, we fall through and treat the identifier
1156          *        as an expression.
1157          *        For GROUP BY, it is incorrect to match the grouping item against
1158          *        targetlist entries: according to SQL92, an identifier in GROUP BY
1159          *        is a reference to a column name exposed by FROM, not to a target
1160          *        list column.  However, many implementations (including pre-7.0
1161          *        PostgreSQL) accept this anyway.  So for GROUP BY, we look first
1162          *        to see if the identifier matches any FROM column name, and only
1163          *        try for a targetlist name if it doesn't.  This ensures that we
1164          *        adhere to the spec in the case where the name could be both.
1165          *        DISTINCT ON isn't in the standard, so we can do what we like there;
1166          *        we choose to make it work like ORDER BY, on the rather flimsy
1167          *        grounds that ordinary DISTINCT works on targetlist entries.
1168          *
1169          * 2. IntegerConstant
1170          *        This means to use the n'th item in the existing target list.
1171          *        Note that it would make no sense to order/group/distinct by an
1172          *        actual constant, so this does not create a conflict with our
1173          *        extension to order/group by an expression.
1174          *        GROUP BY column-number is not allowed by SQL92, but since
1175          *        the standard has no other behavior defined for this syntax,
1176          *        we may as well accept this common extension.
1177          *
1178          * Note that pre-existing resjunk targets must not be used in either case,
1179          * since the user didn't write them in his SELECT list.
1180          *
1181          * If neither special case applies, fall through to treat the item as
1182          * an expression.
1183          *----------
1184          */
1185         if (IsA(node, ColumnRef) &&
1186                 list_length(((ColumnRef *) node)->fields) == 1)
1187         {
1188                 char       *name = strVal(linitial(((ColumnRef *) node)->fields));
1189                 int                     location = ((ColumnRef *) node)->location;
1190
1191                 if (clause == GROUP_CLAUSE)
1192                 {
1193                         /*
1194                          * In GROUP BY, we must prefer a match against a FROM-clause
1195                          * column to one against the targetlist.  Look to see if there is
1196                          * a matching column.  If so, fall through to let transformExpr()
1197                          * do the rest.  NOTE: if name could refer ambiguously to more
1198                          * than one column name exposed by FROM, colNameToVar will
1199                          * ereport(ERROR).      That's just what we want here.
1200                          *
1201                          * Small tweak for 7.4.3: ignore matches in upper query levels.
1202                          * This effectively changes the search order for bare names to (1)
1203                          * local FROM variables, (2) local targetlist aliases, (3) outer
1204                          * FROM variables, whereas before it was (1) (3) (2). SQL92 and
1205                          * SQL99 do not allow GROUPing BY an outer reference, so this
1206                          * breaks no cases that are legal per spec, and it seems a more
1207                          * self-consistent behavior.
1208                          */
1209                         if (colNameToVar(pstate, name, true, location) != NULL)
1210                                 name = NULL;
1211                 }
1212
1213                 if (name != NULL)
1214                 {
1215                         foreach(tl, *tlist)
1216                         {
1217                                 TargetEntry *tle = (TargetEntry *) lfirst(tl);
1218
1219                                 if (!tle->resjunk &&
1220                                         strcmp(tle->resname, name) == 0)
1221                                 {
1222                                         if (target_result != NULL)
1223                                         {
1224                                                 if (!equal(target_result->expr, tle->expr))
1225                                                         ereport(ERROR,
1226                                                                         (errcode(ERRCODE_AMBIGUOUS_COLUMN),
1227
1228                                                         /*
1229                                                          * translator: first %s is name of a SQL
1230                                                          * construct, eg ORDER BY
1231                                                          */
1232                                                                          errmsg("%s \"%s\" is ambiguous",
1233                                                                                         clauseText[clause], name),
1234                                                                          parser_errposition(pstate, location)));
1235                                         }
1236                                         else
1237                                                 target_result = tle;
1238                                         /* Stay in loop to check for ambiguity */
1239                                 }
1240                         }
1241                         if (target_result != NULL)
1242                                 return target_result;   /* return the first match */
1243                 }
1244         }
1245         if (IsA(node, A_Const))
1246         {
1247                 Value      *val = &((A_Const *) node)->val;
1248                 int                     targetlist_pos = 0;
1249                 int                     target_pos;
1250
1251                 if (!IsA(val, Integer))
1252                         ereport(ERROR,
1253                                         (errcode(ERRCODE_SYNTAX_ERROR),
1254                         /* translator: %s is name of a SQL construct, eg ORDER BY */
1255                                          errmsg("non-integer constant in %s",
1256                                                         clauseText[clause])));
1257                 target_pos = intVal(val);
1258                 foreach(tl, *tlist)
1259                 {
1260                         TargetEntry *tle = (TargetEntry *) lfirst(tl);
1261
1262                         if (!tle->resjunk)
1263                         {
1264                                 if (++targetlist_pos == target_pos)
1265                                         return tle; /* return the unique match */
1266                         }
1267                 }
1268                 ereport(ERROR,
1269                                 (errcode(ERRCODE_INVALID_COLUMN_REFERENCE),
1270                 /* translator: %s is name of a SQL construct, eg ORDER BY */
1271                                  errmsg("%s position %d is not in select list",
1272                                                 clauseText[clause], target_pos)));
1273         }
1274
1275         /*
1276          * Otherwise, we have an expression (this is a Postgres extension not
1277          * found in SQL92).  Convert the untransformed node to a transformed
1278          * expression, and search for a match in the tlist. NOTE: it doesn't
1279          * really matter whether there is more than one match.  Also, we are
1280          * willing to match a resjunk target here, though the above cases must
1281          * ignore resjunk targets.
1282          */
1283         expr = transformExpr(pstate, node);
1284
1285         foreach(tl, *tlist)
1286         {
1287                 TargetEntry *tle = (TargetEntry *) lfirst(tl);
1288
1289                 if (equal(expr, tle->expr))
1290                         return tle;
1291         }
1292
1293         /*
1294          * If no matches, construct a new target entry which is appended to the
1295          * end of the target list.      This target is given resjunk = TRUE so that it
1296          * will not be projected into the final tuple.
1297          */
1298         target_result = transformTargetEntry(pstate, node, expr, NULL, true);
1299
1300         *tlist = lappend(*tlist, target_result);
1301
1302         return target_result;
1303 }
1304
1305 static GroupClause *
1306 make_group_clause(TargetEntry *tle, List *targetlist, Oid sortop)
1307 {
1308         GroupClause *result;
1309
1310         result = makeNode(GroupClause);
1311         result->tleSortGroupRef = assignSortGroupRef(tle, targetlist);
1312         result->sortop = sortop;
1313         return result;
1314 }
1315
1316 /*
1317  * transformGroupClause -
1318  *        transform a GROUP BY clause
1319  *
1320  * GROUP BY items will be added to the targetlist (as resjunk columns)
1321  * if not already present, so the targetlist must be passed by reference.
1322  *
1323  * The order of the elements of the grouping clause does not affect
1324  * the semantics of the query. However, the optimizer is not currently
1325  * smart enough to reorder the grouping clause, so we try to do some
1326  * primitive reordering here.
1327  */
1328 List *
1329 transformGroupClause(ParseState *pstate, List *grouplist,
1330                                          List **targetlist, List *sortClause)
1331 {
1332         List       *result = NIL;
1333         List       *tle_list = NIL;
1334         ListCell   *l;
1335
1336         /* Preprocess the grouping clause, lookup TLEs */
1337         foreach (l, grouplist)
1338         {
1339                 TargetEntry *tle;
1340                 Oid                      restype;
1341
1342                 tle = findTargetlistEntry(pstate, lfirst(l),
1343                                                                   targetlist, GROUP_CLAUSE);
1344
1345                 /* if tlist item is an UNKNOWN literal, change it to TEXT */
1346                 restype = exprType((Node *) tle->expr);
1347
1348                 if (restype == UNKNOWNOID)
1349                         tle->expr = (Expr *) coerce_type(pstate, (Node *) tle->expr,
1350                                                                                          restype, TEXTOID, -1,
1351                                                                                          COERCION_IMPLICIT,
1352                                                                                          COERCE_IMPLICIT_CAST);
1353
1354                 tle_list = lappend(tle_list, tle);
1355         }
1356
1357         /*
1358          * Now iterate through the ORDER BY clause. If we find a grouping
1359          * element that matches the ORDER BY element, append the grouping
1360          * element to the result set immediately. Otherwise, stop
1361          * iterating. The effect of this is to look for a prefix of the
1362          * ORDER BY list in the grouping clauses, and to move that prefix
1363          * to the front of the GROUP BY.
1364          */
1365         foreach (l, sortClause)
1366         {
1367                 SortClause      *sc = (SortClause *) lfirst(l);
1368                 ListCell        *prev = NULL;
1369                 ListCell        *tl;
1370                 bool             found = false;
1371
1372                 foreach (tl, tle_list)
1373                 {
1374                         TargetEntry *tle = (TargetEntry *) lfirst(tl);
1375
1376                         if (sc->tleSortGroupRef == tle->ressortgroupref)
1377                         {
1378                                 GroupClause *gc;
1379
1380                                 tle_list = list_delete_cell(tle_list, tl, prev);
1381
1382                                 /* Use the sort clause's sorting operator */
1383                                 gc = make_group_clause(tle, *targetlist, sc->sortop);
1384                                 result = lappend(result, gc);
1385                                 found = true;
1386                                 break;
1387                         }
1388
1389                         prev = tl;
1390                 }
1391
1392                 /* As soon as we've failed to match an ORDER BY element, stop */
1393                 if (!found)
1394                         break;
1395         }
1396
1397         /*
1398          * Now add any remaining elements of the GROUP BY list in the
1399          * order we received them.
1400          *
1401          * XXX: are there any additional criteria to consider when
1402          * ordering grouping clauses?
1403          */
1404         foreach(l, tle_list)
1405         {
1406                 TargetEntry *tle = (TargetEntry *) lfirst(l);
1407                 GroupClause *gc;
1408                 Oid                      sort_op;
1409
1410                 /* avoid making duplicate grouplist entries */
1411                 if (targetIsInSortList(tle, result))
1412                         continue;
1413
1414                 sort_op = ordering_oper_opid(exprType((Node *) tle->expr));
1415                 gc = make_group_clause(tle, *targetlist, sort_op);
1416                 result = lappend(result, gc);
1417         }
1418
1419         list_free(tle_list);
1420         return result;
1421 }
1422
1423 /*
1424  * transformSortClause -
1425  *        transform an ORDER BY clause
1426  *
1427  * ORDER BY items will be added to the targetlist (as resjunk columns)
1428  * if not already present, so the targetlist must be passed by reference.
1429  */
1430 List *
1431 transformSortClause(ParseState *pstate,
1432                                         List *orderlist,
1433                                         List **targetlist,
1434                                         bool resolveUnknown)
1435 {
1436         List       *sortlist = NIL;
1437         ListCell   *olitem;
1438
1439         foreach(olitem, orderlist)
1440         {
1441                 SortBy     *sortby = lfirst(olitem);
1442                 TargetEntry *tle;
1443
1444                 tle = findTargetlistEntry(pstate, sortby->node,
1445                                                                   targetlist, ORDER_CLAUSE);
1446
1447                 sortlist = addTargetToSortList(pstate, tle,
1448                                                                            sortlist, *targetlist,
1449                                                                            sortby->sortby_kind,
1450                                                                            sortby->useOp,
1451                                                                            resolveUnknown);
1452         }
1453
1454         return sortlist;
1455 }
1456
1457 /*
1458  * transformDistinctClause -
1459  *        transform a DISTINCT or DISTINCT ON clause
1460  *
1461  * Since we may need to add items to the query's sortClause list, that list
1462  * is passed by reference.      Likewise for the targetlist.
1463  */
1464 List *
1465 transformDistinctClause(ParseState *pstate, List *distinctlist,
1466                                                 List **targetlist, List **sortClause)
1467 {
1468         List       *result = NIL;
1469         ListCell   *slitem;
1470         ListCell   *dlitem;
1471
1472         /* No work if there was no DISTINCT clause */
1473         if (distinctlist == NIL)
1474                 return NIL;
1475
1476         if (linitial(distinctlist) == NULL)
1477         {
1478                 /* We had SELECT DISTINCT */
1479
1480                 /*
1481                  * All non-resjunk elements from target list that are not already in
1482                  * the sort list should be added to it.  (We don't really care what
1483                  * order the DISTINCT fields are checked in, so we can leave the
1484                  * user's ORDER BY spec alone, and just add additional sort keys to it
1485                  * to ensure that all targetlist items get sorted.)
1486                  */
1487                 *sortClause = addAllTargetsToSortList(pstate,
1488                                                                                           *sortClause,
1489                                                                                           *targetlist,
1490                                                                                           true);
1491
1492                 /*
1493                  * Now, DISTINCT list consists of all non-resjunk sortlist items.
1494                  * Actually, all the sortlist items had better be non-resjunk!
1495                  * Otherwise, user wrote SELECT DISTINCT with an ORDER BY item that
1496                  * does not appear anywhere in the SELECT targetlist, and we can't
1497                  * implement that with only one sorting pass...
1498                  */
1499                 foreach(slitem, *sortClause)
1500                 {
1501                         SortClause *scl = (SortClause *) lfirst(slitem);
1502                         TargetEntry *tle = get_sortgroupclause_tle(scl, *targetlist);
1503
1504                         if (tle->resjunk)
1505                                 ereport(ERROR,
1506                                                 (errcode(ERRCODE_INVALID_COLUMN_REFERENCE),
1507                                                  errmsg("for SELECT DISTINCT, ORDER BY expressions must appear in select list")));
1508                         else
1509                                 result = lappend(result, copyObject(scl));
1510                 }
1511         }
1512         else
1513         {
1514                 /* We had SELECT DISTINCT ON (expr, ...) */
1515
1516                 /*
1517                  * If the user writes both DISTINCT ON and ORDER BY, then the two
1518                  * expression lists must match (until one or the other runs out).
1519                  * Otherwise the ORDER BY requires a different sort order than the
1520                  * DISTINCT does, and we can't implement that with only one sort pass
1521                  * (and if we do two passes, the results will be rather
1522                  * unpredictable). However, it's OK to have more DISTINCT ON
1523                  * expressions than ORDER BY expressions; we can just add the extra
1524                  * DISTINCT values to the sort list, much as we did above for ordinary
1525                  * DISTINCT fields.
1526                  *
1527                  * Actually, it'd be OK for the common prefixes of the two lists to
1528                  * match in any order, but implementing that check seems like more
1529                  * trouble than it's worth.
1530                  */
1531                 ListCell   *nextsortlist = list_head(*sortClause);
1532
1533                 foreach(dlitem, distinctlist)
1534                 {
1535                         TargetEntry *tle;
1536
1537                         tle = findTargetlistEntry(pstate, lfirst(dlitem),
1538                                                                           targetlist, DISTINCT_ON_CLAUSE);
1539
1540                         if (nextsortlist != NULL)
1541                         {
1542                                 SortClause *scl = (SortClause *) lfirst(nextsortlist);
1543
1544                                 if (tle->ressortgroupref != scl->tleSortGroupRef)
1545                                         ereport(ERROR,
1546                                                         (errcode(ERRCODE_INVALID_COLUMN_REFERENCE),
1547                                                          errmsg("SELECT DISTINCT ON expressions must match initial ORDER BY expressions")));
1548                                 result = lappend(result, copyObject(scl));
1549                                 nextsortlist = lnext(nextsortlist);
1550                         }
1551                         else
1552                         {
1553                                 *sortClause = addTargetToSortList(pstate, tle,
1554                                                                                                   *sortClause, *targetlist,
1555                                                                                                   SORTBY_ASC, NIL, true);
1556
1557                                 /*
1558                                  * Probably, the tle should always have been added at the end
1559                                  * of the sort list ... but search to be safe.
1560                                  */
1561                                 foreach(slitem, *sortClause)
1562                                 {
1563                                         SortClause *scl = (SortClause *) lfirst(slitem);
1564
1565                                         if (tle->ressortgroupref == scl->tleSortGroupRef)
1566                                         {
1567                                                 result = lappend(result, copyObject(scl));
1568                                                 break;
1569                                         }
1570                                 }
1571                                 if (slitem == NULL)             /* should not happen */
1572                                         elog(ERROR, "failed to add DISTINCT ON clause to target list");
1573                         }
1574                 }
1575         }
1576
1577         return result;
1578 }
1579
1580 /*
1581  * addAllTargetsToSortList
1582  *              Make sure all non-resjunk targets in the targetlist are in the
1583  *              ORDER BY list, adding the not-yet-sorted ones to the end of the list.
1584  *              This is typically used to help implement SELECT DISTINCT.
1585  *
1586  * See addTargetToSortList for info about pstate and resolveUnknown inputs.
1587  *
1588  * Returns the updated ORDER BY list.
1589  */
1590 List *
1591 addAllTargetsToSortList(ParseState *pstate, List *sortlist,
1592                                                 List *targetlist, bool resolveUnknown)
1593 {
1594         ListCell   *l;
1595
1596         foreach(l, targetlist)
1597         {
1598                 TargetEntry *tle = (TargetEntry *) lfirst(l);
1599
1600                 if (!tle->resjunk)
1601                         sortlist = addTargetToSortList(pstate, tle,
1602                                                                                    sortlist, targetlist,
1603                                                                                    SORTBY_ASC, NIL,
1604                                                                                    resolveUnknown);
1605         }
1606         return sortlist;
1607 }
1608
1609 /*
1610  * addTargetToSortList
1611  *              If the given targetlist entry isn't already in the ORDER BY list,
1612  *              add it to the end of the list, using the sortop with given name
1613  *              or the default sort operator if opname == NIL.
1614  *
1615  * If resolveUnknown is TRUE, convert TLEs of type UNKNOWN to TEXT.  If not,
1616  * do nothing (which implies the search for a sort operator will fail).
1617  * pstate should be provided if resolveUnknown is TRUE, but can be NULL
1618  * otherwise.
1619  *
1620  * Returns the updated ORDER BY list.
1621  */
1622 List *
1623 addTargetToSortList(ParseState *pstate, TargetEntry *tle,
1624                                         List *sortlist, List *targetlist,
1625                                         int sortby_kind, List *sortby_opname,
1626                                         bool resolveUnknown)
1627 {
1628         /* avoid making duplicate sortlist entries */
1629         if (!targetIsInSortList(tle, sortlist))
1630         {
1631                 SortClause *sortcl = makeNode(SortClause);
1632                 Oid                     restype = exprType((Node *) tle->expr);
1633
1634                 /* if tlist item is an UNKNOWN literal, change it to TEXT */
1635                 if (restype == UNKNOWNOID && resolveUnknown)
1636                 {
1637                         tle->expr = (Expr *) coerce_type(pstate, (Node *) tle->expr,
1638                                                                                          restype, TEXTOID, -1,
1639                                                                                          COERCION_IMPLICIT,
1640                                                                                          COERCE_IMPLICIT_CAST);
1641                         restype = TEXTOID;
1642                 }
1643
1644                 sortcl->tleSortGroupRef = assignSortGroupRef(tle, targetlist);
1645
1646                 switch (sortby_kind)
1647                 {
1648                         case SORTBY_ASC:
1649                                 sortcl->sortop = ordering_oper_opid(restype);
1650                                 break;
1651                         case SORTBY_DESC:
1652                                 sortcl->sortop = reverse_ordering_oper_opid(restype);
1653                                 break;
1654                         case SORTBY_USING:
1655                                 Assert(sortby_opname != NIL);
1656                                 sortcl->sortop = compatible_oper_opid(sortby_opname,
1657                                                                                                           restype,
1658                                                                                                           restype,
1659                                                                                                           false);
1660                                 break;
1661                         default:
1662                                 elog(ERROR, "unrecognized sortby_kind: %d", sortby_kind);
1663                                 break;
1664                 }
1665
1666                 sortlist = lappend(sortlist, sortcl);
1667         }
1668         return sortlist;
1669 }
1670
1671 /*
1672  * assignSortGroupRef
1673  *        Assign the targetentry an unused ressortgroupref, if it doesn't
1674  *        already have one.  Return the assigned or pre-existing refnumber.
1675  *
1676  * 'tlist' is the targetlist containing (or to contain) the given targetentry.
1677  */
1678 Index
1679 assignSortGroupRef(TargetEntry *tle, List *tlist)
1680 {
1681         Index           maxRef;
1682         ListCell   *l;
1683
1684         if (tle->ressortgroupref)       /* already has one? */
1685                 return tle->ressortgroupref;
1686
1687         /* easiest way to pick an unused refnumber: max used + 1 */
1688         maxRef = 0;
1689         foreach(l, tlist)
1690         {
1691                 Index           ref = ((TargetEntry *) lfirst(l))->ressortgroupref;
1692
1693                 if (ref > maxRef)
1694                         maxRef = ref;
1695         }
1696         tle->ressortgroupref = maxRef + 1;
1697         return tle->ressortgroupref;
1698 }
1699
1700 /*
1701  * targetIsInSortList
1702  *              Is the given target item already in the sortlist?
1703  *
1704  * Works for both SortClause and GroupClause lists.  Note that the main
1705  * reason we need this routine (and not just a quick test for nonzeroness
1706  * of ressortgroupref) is that a TLE might be in only one of the lists.
1707  */
1708 bool
1709 targetIsInSortList(TargetEntry *tle, List *sortList)
1710 {
1711         Index           ref = tle->ressortgroupref;
1712         ListCell   *l;
1713
1714         /* no need to scan list if tle has no marker */
1715         if (ref == 0)
1716                 return false;
1717
1718         foreach(l, sortList)
1719         {
1720                 SortClause *scl = (SortClause *) lfirst(l);
1721
1722                 if (scl->tleSortGroupRef == ref)
1723                         return true;
1724         }
1725         return false;
1726 }