core.c

   1 /*-------------------------------------------------------------------------
   2  *
   3  * core.c
   4  *        Routines copied from PostgreSQL core distribution.
   5  *
   6
   7  * The main purpose of this files is having access to static functions in core.
   8  * Another purpose is tweaking functions behavior by replacing part of them by
   9  * macro definitions. See at the end of pg_hint_plan.c for details. Anyway,
  10  * this file *must* contain required functions without making any change.
  11  *
  12  * This file contains the following functions from corresponding files.
  13  *
  14  * src/backend/optimizer/path/allpaths.c
  15  *
  16  *      static functions:
  17  *     set_plain_rel_pathlist()
  18  *     create_plain_partial_paths()
  19  *     set_append_rel_pathlist()
  20  *     add_paths_to_append_rel()
  21  *     generate_mergeappend_paths()
  22  *     get_cheapest_parameterized_child_path()
  23  *     accumulate_append_subpath()
  24  *
  25  *  public functions:
  26  *     standard_join_search(): This funcion is not static. The reason for
  27  *        including this function is make_rels_by_clause_joins. In order to
  28  *        avoid generating apparently unwanted join combination, we decided to
  29  *        change the behavior of make_join_rel, which is called under this
  30  *        function.
  31  *
  32  * src/backend/optimizer/path/joinrels.c
  33  *
  34  *      public functions:
  35  *     join_search_one_level(): We have to modify this to call my definition of
  36  *                  make_rels_by_clause_joins.
  37  *
  38  *      static functions:
  39  *     make_rels_by_clause_joins()
  40  *     make_rels_by_clauseless_joins()
  41  *     join_is_legal()
  42  *     has_join_restriction()
  43  *     mark_dummy_rel()
  44  *     restriction_is_constant_false()
  45  *
  46  *
  47  * Portions Copyright (c) 1996-2020, PostgreSQL Global Development Group
  48  * Portions Copyright (c) 1994, Regents of the University of California
  49  *
  50  *-------------------------------------------------------------------------
  51  */
  52
  53
  54 /*
  55  * set_plain_rel_pathlist
  56  *        Build access paths for a plain relation (no subquery, no inheritance)
  57  */
  58 static void
  59 set_plain_rel_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
  60 {
  61         Relids          required_outer;
  62
  63         /*
  64          * We don't support pushing join clauses into the quals of a seqscan, but
  65          * it could still have required parameterization due to LATERAL refs in
  66          * its tlist.
  67          */
  68         required_outer = rel->lateral_relids;
  69
  70         /* Consider sequential scan */
  71         add_path(rel, create_seqscan_path(root, rel, required_outer, 0));
  72
  73         /* If appropriate, consider parallel sequential scan */
  74         if (rel->consider_parallel && required_outer == NULL)
  75                 create_plain_partial_paths(root, rel);
  76
  77         /* Consider index scans */
  78         create_index_paths(root, rel);
  79
  80         /* Consider TID scans */
  81         create_tidscan_paths(root, rel);
  82 }
  83
  84
  85 /*
  86  * create_plain_partial_paths
  87  *        Build partial access paths for parallel scan of a plain relation
  88  */
  89 static void
  90 create_plain_partial_paths(PlannerInfo *root, RelOptInfo *rel)
  91 {
  92         int                     parallel_workers;
  93
  94         parallel_workers = compute_parallel_worker(rel, rel->pages, -1);
  95
  96         /* If any limit was set to zero, the user doesn't want a parallel scan. */
  97         if (parallel_workers <= 0)
  98                 return;
  99
 100         /* Add an unordered partial path based on a parallel sequential scan. */
 101         add_partial_path(rel, create_seqscan_path(root, rel, NULL, parallel_workers));
 102 }
 103
 104
 105 /*
 106  * set_append_rel_pathlist
 107  *        Build access paths for an "append relation"
 108  */
 109 static void
 110 set_append_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
 111                                                 Index rti, RangeTblEntry *rte)
 112 {
 113         int                     parentRTindex = rti;
 114         List       *live_childrels = NIL;
 115         ListCell   *l;
 116
 117         /*
 118          * Generate access paths for each member relation, and remember the
 119          * non-dummy children.
 120          */
 121         foreach(l, root->append_rel_list)
 122         {
 123                 AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
 124                 int                     childRTindex;
 125                 RangeTblEntry *childRTE;
 126                 RelOptInfo *childrel;
 127
 128                 /* append_rel_list contains all append rels; ignore others */
 129                 if (appinfo->parent_relid != parentRTindex)
 130                         continue;
 131
 132                 /* Re-locate the child RTE and RelOptInfo */
 133                 childRTindex = appinfo->child_relid;
 134                 childRTE = root->simple_rte_array[childRTindex];
 135                 childrel = root->simple_rel_array[childRTindex];
 136
 137                 /*
 138                  * If set_append_rel_size() decided the parent appendrel was
 139                  * parallel-unsafe at some point after visiting this child rel, we
 140                  * need to propagate the unsafety marking down to the child, so that
 141                  * we don't generate useless partial paths for it.
 142                  */
 143                 if (!rel->consider_parallel)
 144                         childrel->consider_parallel = false;
 145
 146                 /*
 147                  * Compute the child's access paths.
 148                  */
 149                 set_rel_pathlist(root, childrel, childRTindex, childRTE);
 150
 151                 /*
 152                  * If child is dummy, ignore it.
 153                  */
 154                 if (IS_DUMMY_REL(childrel))
 155                         continue;
 156
 157                 /*
 158                  * Child is live, so add it to the live_childrels list for use below.
 159                  */
 160                 live_childrels = lappend(live_childrels, childrel);
 161         }
 162
 163         /* Add paths to the "append" relation. */
 164         add_paths_to_append_rel(root, rel, live_childrels);
 165 }
 166
 167 /*
 168  * add_paths_to_append_rel
 169  *              Generate paths for given "append" relation given the set of non-dummy
 170  *              child rels.
 171  *
 172  * The function collects all parameterizations and orderings supported by the
 173  * non-dummy children. For every such parameterization or ordering, it creates
 174  * an append path collecting one path from each non-dummy child with given
 175  * parameterization or ordering. Similarly it collects partial paths from
 176  * non-dummy children to create partial append paths.
 177  */
 178 static void
 179 add_paths_to_append_rel(PlannerInfo *root, RelOptInfo *rel,
 180                                                 List *live_childrels)
 181 {
 182         List       *subpaths = NIL;
 183         bool            subpaths_valid = true;
 184         List       *partial_subpaths = NIL;
 185         bool            partial_subpaths_valid = true;
 186         List       *all_child_pathkeys = NIL;
 187         List       *all_child_outers = NIL;
 188         ListCell   *l;
 189         List       *partitioned_rels = NIL;
 190         RangeTblEntry *rte;
 191         bool            build_partitioned_rels = false;
 192
 193         /*
 194          * A plain relation will already have a PartitionedChildRelInfo if it is
 195          * partitioned.  For a subquery RTE, no PartitionedChildRelInfo exists; we
 196          * collect all partitioned_rels associated with any child.  (This assumes
 197          * that we don't need to look through multiple levels of subquery RTEs; if
 198          * we ever do, we could create a PartitionedChildRelInfo with the
 199          * accumulated list of partitioned_rels which would then be found when
 200          * populated our parent rel with paths.  For the present, that appears to
 201          * be unnecessary.)
 202          */
 203         rte = planner_rt_fetch(rel->relid, root);
 204         switch (rte->rtekind)
 205         {
 206                 case RTE_RELATION:
 207                         if (rte->relkind == RELKIND_PARTITIONED_TABLE)
 208                         {
 209                                 partitioned_rels =
 210                                         get_partitioned_child_rels(root, rel->relid);
 211                                 Assert(list_length(partitioned_rels) >= 1);
 212                         }
 213                         break;
 214                 case RTE_SUBQUERY:
 215                         build_partitioned_rels = true;
 216                         break;
 217                 default:
 218                         elog(ERROR, "unexpected rtekind: %d", (int) rte->rtekind);
 219         }
 220
 221         /*
 222          * For every non-dummy child, remember the cheapest path.  Also, identify
 223          * all pathkeys (orderings) and parameterizations (required_outer sets)
 224          * available for the non-dummy member relations.
 225          */
 226         foreach(l, live_childrels)
 227         {
 228                 RelOptInfo *childrel = lfirst(l);
 229                 ListCell   *lcp;
 230
 231                 /*
 232                  * If we need to build partitioned_rels, accumulate the partitioned
 233                  * rels for this child.
 234                  */
 235                 if (build_partitioned_rels)
 236                 {
 237                         List       *cprels;
 238
 239                         cprels = get_partitioned_child_rels(root, childrel->relid);
 240                         partitioned_rels = list_concat(partitioned_rels,
 241                                                                                    list_copy(cprels));
 242                 }
 243
 244                 /*
 245                  * If child has an unparameterized cheapest-total path, add that to
 246                  * the unparameterized Append path we are constructing for the parent.
 247                  * If not, there's no workable unparameterized path.
 248                  */
 249                 if (childrel->cheapest_total_path->param_info == NULL)
 250                         subpaths = accumulate_append_subpath(subpaths,
 251                                                                                                  childrel->cheapest_total_path);
 252                 else
 253                         subpaths_valid = false;
 254
 255                 /* Same idea, but for a partial plan. */
 256                 if (childrel->partial_pathlist != NIL)
 257                         partial_subpaths = accumulate_append_subpath(partial_subpaths,
 258                                                                                                                  linitial(childrel->partial_pathlist));
 259                 else
 260                         partial_subpaths_valid = false;
 261
 262                 /*
 263                  * Collect lists of all the available path orderings and
 264                  * parameterizations for all the children.  We use these as a
 265                  * heuristic to indicate which sort orderings and parameterizations we
 266                  * should build Append and MergeAppend paths for.
 267                  */
 268                 foreach(lcp, childrel->pathlist)
 269                 {
 270                         Path       *childpath = (Path *) lfirst(lcp);
 271                         List       *childkeys = childpath->pathkeys;
 272                         Relids          childouter = PATH_REQ_OUTER(childpath);
 273
 274                         /* Unsorted paths don't contribute to pathkey list */
 275                         if (childkeys != NIL)
 276                         {
 277                                 ListCell   *lpk;
 278                                 bool            found = false;
 279
 280                                 /* Have we already seen this ordering? */
 281                                 foreach(lpk, all_child_pathkeys)
 282                                 {
 283                                         List       *existing_pathkeys = (List *) lfirst(lpk);
 284
 285                                         if (compare_pathkeys(existing_pathkeys,
 286                                                                                  childkeys) == PATHKEYS_EQUAL)
 287                                         {
 288                                                 found = true;
 289                                                 break;
 290                                         }
 291                                 }
 292                                 if (!found)
 293                                 {
 294                                         /* No, so add it to all_child_pathkeys */
 295                                         all_child_pathkeys = lappend(all_child_pathkeys,
 296                                                                                                  childkeys);
 297                                 }
 298                         }
 299
 300                         /* Unparameterized paths don't contribute to param-set list */
 301                         if (childouter)
 302                         {
 303                                 ListCell   *lco;
 304                                 bool            found = false;
 305
 306                                 /* Have we already seen this param set? */
 307                                 foreach(lco, all_child_outers)
 308                                 {
 309                                         Relids          existing_outers = (Relids) lfirst(lco);
 310
 311                                         if (bms_equal(existing_outers, childouter))
 312                                         {
 313                                                 found = true;
 314                                                 break;
 315                                         }
 316                                 }
 317                                 if (!found)
 318                                 {
 319                                         /* No, so add it to all_child_outers */
 320                                         all_child_outers = lappend(all_child_outers,
 321                                                                                            childouter);
 322                                 }
 323                         }
 324                 }
 325         }
 326
 327         /*
 328          * If we found unparameterized paths for all children, build an unordered,
 329          * unparameterized Append path for the rel.  (Note: this is correct even
 330          * if we have zero or one live subpath due to constraint exclusion.)
 331          */
 332         if (subpaths_valid)
 333                 add_path(rel, (Path *) create_append_path(rel, subpaths, NULL, 0,
 334                                                                                                   partitioned_rels));
 335
 336         /*
 337          * Consider an append of partial unordered, unparameterized partial paths.
 338          */
 339         if (partial_subpaths_valid && partial_subpaths != NIL)
 340         {
 341                 AppendPath *appendpath;
 342                 ListCell   *lc;
 343                 int                     parallel_workers = 0;
 344
 345                 /*
 346                  * Decide on the number of workers to request for this append path.
 347                  * For now, we just use the maximum value from among the members.  It
 348                  * might be useful to use a higher number if the Append node were
 349                  * smart enough to spread out the workers, but it currently isn't.
 350                  */
 351                 foreach(lc, partial_subpaths)
 352                 {
 353                         Path       *path = lfirst(lc);
 354
 355                         parallel_workers = Max(parallel_workers, path->parallel_workers);
 356                 }
 357                 Assert(parallel_workers > 0);
 358
 359                 /* Generate a partial append path. */
 360                 appendpath = create_append_path(rel, partial_subpaths, NULL,
 361                                                                                 parallel_workers, partitioned_rels);
 362                 add_partial_path(rel, (Path *) appendpath);
 363         }
 364
 365         /*
 366          * Also build unparameterized MergeAppend paths based on the collected
 367          * list of child pathkeys.
 368          */
 369         if (subpaths_valid)
 370                 generate_mergeappend_paths(root, rel, live_childrels,
 371                                                                    all_child_pathkeys,
 372                                                                    partitioned_rels);
 373
 374         /*
 375          * Build Append paths for each parameterization seen among the child rels.
 376          * (This may look pretty expensive, but in most cases of practical
 377          * interest, the child rels will expose mostly the same parameterizations,
 378          * so that not that many cases actually get considered here.)
 379          *
 380          * The Append node itself cannot enforce quals, so all qual checking must
 381          * be done in the child paths.  This means that to have a parameterized
 382          * Append path, we must have the exact same parameterization for each
 383          * child path; otherwise some children might be failing to check the
 384          * moved-down quals.  To make them match up, we can try to increase the
 385          * parameterization of lesser-parameterized paths.
 386          */
 387         foreach(l, all_child_outers)
 388         {
 389                 Relids          required_outer = (Relids) lfirst(l);
 390                 ListCell   *lcr;
 391
 392                 /* Select the child paths for an Append with this parameterization */
 393                 subpaths = NIL;
 394                 subpaths_valid = true;
 395                 foreach(lcr, live_childrels)
 396                 {
 397                         RelOptInfo *childrel = (RelOptInfo *) lfirst(lcr);
 398                         Path       *subpath;
 399
 400                         subpath = get_cheapest_parameterized_child_path(root,
 401                                                                                                                         childrel,
 402                                                                                                                         required_outer);
 403                         if (subpath == NULL)
 404                         {
 405                                 /* failed to make a suitable path for this child */
 406                                 subpaths_valid = false;
 407                                 break;
 408                         }
 409                         subpaths = accumulate_append_subpath(subpaths, subpath);
 410                 }
 411
 412                 if (subpaths_valid)
 413                         add_path(rel, (Path *)
 414                                          create_append_path(rel, subpaths, required_outer, 0,
 415                                                                                 partitioned_rels));
 416         }
 417 }
 418
 419
 420 /*
 421  * generate_mergeappend_paths
 422  *              Generate MergeAppend paths for an append relation
 423  *
 424  * Generate a path for each ordering (pathkey list) appearing in
 425  * all_child_pathkeys.
 426  *
 427  * We consider both cheapest-startup and cheapest-total cases, ie, for each
 428  * interesting ordering, collect all the cheapest startup subpaths and all the
 429  * cheapest total paths, and build a MergeAppend path for each case.
 430  *
 431  * We don't currently generate any parameterized MergeAppend paths.  While
 432  * it would not take much more code here to do so, it's very unclear that it
 433  * is worth the planning cycles to investigate such paths: there's little
 434  * use for an ordered path on the inside of a nestloop.  In fact, it's likely
 435  * that the current coding of add_path would reject such paths out of hand,
 436  * because add_path gives no credit for sort ordering of parameterized paths,
 437  * and a parameterized MergeAppend is going to be more expensive than the
 438  * corresponding parameterized Append path.  If we ever try harder to support
 439  * parameterized mergejoin plans, it might be worth adding support for
 440  * parameterized MergeAppends to feed such joins.  (See notes in
 441  * optimizer/README for why that might not ever happen, though.)
 442  */
 443 static void
 444 generate_mergeappend_paths(PlannerInfo *root, RelOptInfo *rel,
 445                                                    List *live_childrels,
 446                                                    List *all_child_pathkeys,
 447                                                    List *partitioned_rels)
 448 {
 449         ListCell   *lcp;
 450
 451         foreach(lcp, all_child_pathkeys)
 452         {
 453                 List       *pathkeys = (List *) lfirst(lcp);
 454                 List       *startup_subpaths = NIL;
 455                 List       *total_subpaths = NIL;
 456                 bool            startup_neq_total = false;
 457                 ListCell   *lcr;
 458
 459                 /* Select the child paths for this ordering... */
 460                 foreach(lcr, live_childrels)
 461                 {
 462                         RelOptInfo *childrel = (RelOptInfo *) lfirst(lcr);
 463                         Path       *cheapest_startup,
 464                                            *cheapest_total;
 465
 466                         /* Locate the right paths, if they are available. */
 467                         cheapest_startup =
 468                                 get_cheapest_path_for_pathkeys(childrel->pathlist,
 469                                                                                            pathkeys,
 470                                                                                            NULL,
 471                                                                                            STARTUP_COST,
 472                                                                                            false);
 473                         cheapest_total =
 474                                 get_cheapest_path_for_pathkeys(childrel->pathlist,
 475                                                                                            pathkeys,
 476                                                                                            NULL,
 477                                                                                            TOTAL_COST,
 478                                                                                            false);
 479
 480                         /*
 481                          * If we can't find any paths with the right order just use the
 482                          * cheapest-total path; we'll have to sort it later.
 483                          */
 484                         if (cheapest_startup == NULL || cheapest_total == NULL)
 485                         {
 486                                 cheapest_startup = cheapest_total =
 487                                         childrel->cheapest_total_path;
 488                                 /* Assert we do have an unparameterized path for this child */
 489                                 Assert(cheapest_total->param_info == NULL);
 490                         }
 491
 492                         /*
 493                          * Notice whether we actually have different paths for the
 494                          * "cheapest" and "total" cases; frequently there will be no point
 495                          * in two create_merge_append_path() calls.
 496                          */
 497                         if (cheapest_startup != cheapest_total)
 498                                 startup_neq_total = true;
 499
 500                         startup_subpaths =
 501                                 accumulate_append_subpath(startup_subpaths, cheapest_startup);
 502                         total_subpaths =
 503                                 accumulate_append_subpath(total_subpaths, cheapest_total);
 504                 }
 505
 506                 /* ... and build the MergeAppend paths */
 507                 add_path(rel, (Path *) create_merge_append_path(root,
 508                                                                                                                 rel,
 509                                                                                                                 startup_subpaths,
 510                                                                                                                 pathkeys,
 511                                                                                                                 NULL,
 512                                                                                                                 partitioned_rels));
 513                 if (startup_neq_total)
 514                         add_path(rel, (Path *) create_merge_append_path(root,
 515                                                                                                                         rel,
 516                                                                                                                         total_subpaths,
 517                                                                                                                         pathkeys,
 518                                                                                                                         NULL,
 519                                                                                                                         partitioned_rels));
 520         }
 521 }
 522
 523
 524 /*
 525  * get_cheapest_parameterized_child_path
 526  *              Get cheapest path for this relation that has exactly the requested
 527  *              parameterization.
 528  *
 529  * Returns NULL if unable to create such a path.
 530  */
 531 static Path *
 532 get_cheapest_parameterized_child_path(PlannerInfo *root, RelOptInfo *rel,
 533                                                                           Relids required_outer)
 534 {
 535         Path       *cheapest;
 536         ListCell   *lc;
 537
 538         /*
 539          * Look up the cheapest existing path with no more than the needed
 540          * parameterization.  If it has exactly the needed parameterization, we're
 541          * done.
 542          */
 543         cheapest = get_cheapest_path_for_pathkeys(rel->pathlist,
 544                                                                                           NIL,
 545                                                                                           required_outer,
 546                                                                                           TOTAL_COST,
 547                                                                                           false);
 548         Assert(cheapest != NULL);
 549         if (bms_equal(PATH_REQ_OUTER(cheapest), required_outer))
 550                 return cheapest;
 551
 552         /*
 553          * Otherwise, we can "reparameterize" an existing path to match the given
 554          * parameterization, which effectively means pushing down additional
 555          * joinquals to be checked within the path's scan.  However, some existing
 556          * paths might check the available joinquals already while others don't;
 557          * therefore, it's not clear which existing path will be cheapest after
 558          * reparameterization.  We have to go through them all and find out.
 559          */
 560         cheapest = NULL;
 561         foreach(lc, rel->pathlist)
 562         {
 563                 Path       *path = (Path *) lfirst(lc);
 564
 565                 /* Can't use it if it needs more than requested parameterization */
 566                 if (!bms_is_subset(PATH_REQ_OUTER(path), required_outer))
 567                         continue;
 568
 569                 /*
 570                  * Reparameterization can only increase the path's cost, so if it's
 571                  * already more expensive than the current cheapest, forget it.
 572                  */
 573                 if (cheapest != NULL &&
 574                         compare_path_costs(cheapest, path, TOTAL_COST) <= 0)
 575                         continue;
 576
 577                 /* Reparameterize if needed, then recheck cost */
 578                 if (!bms_equal(PATH_REQ_OUTER(path), required_outer))
 579                 {
 580                         path = reparameterize_path(root, path, required_outer, 1.0);
 581                         if (path == NULL)
 582                                 continue;               /* failed to reparameterize this one */
 583                         Assert(bms_equal(PATH_REQ_OUTER(path), required_outer));
 584
 585                         if (cheapest != NULL &&
 586                                 compare_path_costs(cheapest, path, TOTAL_COST) <= 0)
 587                                 continue;
 588                 }
 589
 590                 /* We have a new best path */
 591                 cheapest = path;
 592         }
 593
 594         /* Return the best path, or NULL if we found no suitable candidate */
 595         return cheapest;
 596 }
 597
 598
 599 /*
 600  * accumulate_append_subpath
 601  *              Add a subpath to the list being built for an Append or MergeAppend
 602  *
 603  * It's possible that the child is itself an Append or MergeAppend path, in
 604  * which case we can "cut out the middleman" and just add its child paths to
 605  * our own list.  (We don't try to do this earlier because we need to apply
 606  * both levels of transformation to the quals.)
 607  *
 608  * Note that if we omit a child MergeAppend in this way, we are effectively
 609  * omitting a sort step, which seems fine: if the parent is to be an Append,
 610  * its result would be unsorted anyway, while if the parent is to be a
 611  * MergeAppend, there's no point in a separate sort on a child.
 612  */
 613 static List *
 614 accumulate_append_subpath(List *subpaths, Path *path)
 615 {
 616         if (IsA(path, AppendPath))
 617         {
 618                 AppendPath *apath = (AppendPath *) path;
 619
 620                 /* list_copy is important here to avoid sharing list substructure */
 621                 return list_concat(subpaths, list_copy(apath->subpaths));
 622         }
 623         else if (IsA(path, MergeAppendPath))
 624         {
 625                 MergeAppendPath *mpath = (MergeAppendPath *) path;
 626
 627                 /* list_copy is important here to avoid sharing list substructure */
 628                 return list_concat(subpaths, list_copy(mpath->subpaths));
 629         }
 630         else
 631                 return lappend(subpaths, path);
 632 }
 633
 634
 635 /*
 636  * standard_join_search
 637  *        Find possible joinpaths for a query by successively finding ways
 638  *        to join component relations into join relations.
 639  *
 640  * 'levels_needed' is the number of iterations needed, ie, the number of
 641  *              independent jointree items in the query.  This is > 1.
 642  *
 643  * 'initial_rels' is a list of RelOptInfo nodes for each independent
 644  *              jointree item.  These are the components to be joined together.
 645  *              Note that levels_needed == list_length(initial_rels).
 646  *
 647  * Returns the final level of join relations, i.e., the relation that is
 648  * the result of joining all the original relations together.
 649  * At least one implementation path must be provided for this relation and
 650  * all required sub-relations.
 651  *
 652  * To support loadable plugins that modify planner behavior by changing the
 653  * join searching algorithm, we provide a hook variable that lets a plugin
 654  * replace or supplement this function.  Any such hook must return the same
 655  * final join relation as the standard code would, but it might have a
 656  * different set of implementation paths attached, and only the sub-joinrels
 657  * needed for these paths need have been instantiated.
 658  *
 659  * Note to plugin authors: the functions invoked during standard_join_search()
 660  * modify root->join_rel_list and root->join_rel_hash.  If you want to do more
 661  * than one join-order search, you'll probably need to save and restore the
 662  * original states of those data structures.  See geqo_eval() for an example.
 663  */
 664 RelOptInfo *
 665 standard_join_search(PlannerInfo *root, int levels_needed, List *initial_rels)
 666 {
 667         int                     lev;
 668         RelOptInfo *rel;
 669
 670         /*
 671          * This function cannot be invoked recursively within any one planning
 672          * problem, so join_rel_level[] can't be in use already.
 673          */
 674         Assert(root->join_rel_level == NULL);
 675
 676         /*
 677          * We employ a simple "dynamic programming" algorithm: we first find all
 678          * ways to build joins of two jointree items, then all ways to build joins
 679          * of three items (from two-item joins and single items), then four-item
 680          * joins, and so on until we have considered all ways to join all the
 681          * items into one rel.
 682          *
 683          * root->join_rel_level[j] is a list of all the j-item rels.  Initially we
 684          * set root->join_rel_level[1] to represent all the single-jointree-item
 685          * relations.
 686          */
 687         root->join_rel_level = (List **) palloc0((levels_needed + 1) * sizeof(List *));
 688
 689         root->join_rel_level[1] = initial_rels;
 690
 691         for (lev = 2; lev <= levels_needed; lev++)
 692         {
 693                 ListCell   *lc;
 694
 695                 /*
 696                  * Determine all possible pairs of relations to be joined at this
 697                  * level, and build paths for making each one from every available
 698                  * pair of lower-level relations.
 699                  */
 700                 join_search_one_level(root, lev);
 701
 702                 /*
 703                  * Run generate_gather_paths() for each just-processed joinrel.  We
 704                  * could not do this earlier because both regular and partial paths
 705                  * can get added to a particular joinrel at multiple times within
 706                  * join_search_one_level.  After that, we're done creating paths for
 707                  * the joinrel, so run set_cheapest().
 708                  */
 709                 foreach(lc, root->join_rel_level[lev])
 710                 {
 711                         rel = (RelOptInfo *) lfirst(lc);
 712
 713                         /* Create GatherPaths for any useful partial paths for rel */
 714                         generate_gather_paths(root, rel);
 715
 716                         /* Find and save the cheapest paths for this rel */
 717                         set_cheapest(rel);
 718
 719 #ifdef OPTIMIZER_DEBUG
 720                         debug_print_rel(root, rel);
 721 #endif
 722                 }
 723         }
 724
 725         /*
 726          * We should have a single rel at the final level.
 727          */
 728         if (root->join_rel_level[levels_needed] == NIL)
 729                 elog(ERROR, "failed to build any %d-way joins", levels_needed);
 730         Assert(list_length(root->join_rel_level[levels_needed]) == 1);
 731
 732         rel = (RelOptInfo *) linitial(root->join_rel_level[levels_needed]);
 733
 734         root->join_rel_level = NULL;
 735
 736         return rel;
 737 }
 738
 739
 740 /*
 741  * join_search_one_level
 742  *        Consider ways to produce join relations containing exactly 'level'
 743  *        jointree items.  (This is one step of the dynamic-programming method
 744  *        embodied in standard_join_search.)  Join rel nodes for each feasible
 745  *        combination of lower-level rels are created and returned in a list.
 746  *        Implementation paths are created for each such joinrel, too.
 747  *
 748  * level: level of rels we want to make this time
 749  * root->join_rel_level[j], 1 <= j < level, is a list of rels containing j items
 750  *
 751  * The result is returned in root->join_rel_level[level].
 752  */
 753 void
 754 join_search_one_level(PlannerInfo *root, int level)
 755 {
 756         List      **joinrels = root->join_rel_level;
 757         ListCell   *r;
 758         int                     k;
 759
 760         Assert(joinrels[level] == NIL);
 761
 762         /* Set join_cur_level so that new joinrels are added to proper list */
 763         root->join_cur_level = level;
 764
 765         /*
 766          * First, consider left-sided and right-sided plans, in which rels of
 767          * exactly level-1 member relations are joined against initial relations.
 768          * We prefer to join using join clauses, but if we find a rel of level-1
 769          * members that has no join clauses, we will generate Cartesian-product
 770          * joins against all initial rels not already contained in it.
 771          */
 772         foreach(r, joinrels[level - 1])
 773         {
 774                 RelOptInfo *old_rel = (RelOptInfo *) lfirst(r);
 775
 776                 if (old_rel->joininfo != NIL || old_rel->has_eclass_joins ||
 777                         has_join_restriction(root, old_rel))
 778                 {
 779                         /*
 780                          * There are join clauses or join order restrictions relevant to
 781                          * this rel, so consider joins between this rel and (only) those
 782                          * initial rels it is linked to by a clause or restriction.
 783                          *
 784                          * At level 2 this condition is symmetric, so there is no need to
 785                          * look at initial rels before this one in the list; we already
 786                          * considered such joins when we were at the earlier rel.  (The
 787                          * mirror-image joins are handled automatically by make_join_rel.)
 788                          * In later passes (level > 2), we join rels of the previous level
 789                          * to each initial rel they don't already include but have a join
 790                          * clause or restriction with.
 791                          */
 792                         ListCell   *other_rels;
 793
 794                         if (level == 2)         /* consider remaining initial rels */
 795                                 other_rels = lnext(r);
 796                         else                            /* consider all initial rels */
 797                                 other_rels = list_head(joinrels[1]);
 798
 799                         make_rels_by_clause_joins(root,
 800                                                                           old_rel,
 801                                                                           other_rels);
 802                 }
 803                 else
 804                 {
 805                         /*
 806                          * Oops, we have a relation that is not joined to any other
 807                          * relation, either directly or by join-order restrictions.
 808                          * Cartesian product time.
 809                          *
 810                          * We consider a cartesian product with each not-already-included
 811                          * initial rel, whether it has other join clauses or not.  At
 812                          * level 2, if there are two or more clauseless initial rels, we
 813                          * will redundantly consider joining them in both directions; but
 814                          * such cases aren't common enough to justify adding complexity to
 815                          * avoid the duplicated effort.
 816                          */
 817                         make_rels_by_clauseless_joins(root,
 818                                                                                   old_rel,
 819                                                                                   list_head(joinrels[1]));
 820                 }
 821         }
 822
 823         /*
 824          * Now, consider "bushy plans" in which relations of k initial rels are
 825          * joined to relations of level-k initial rels, for 2 <= k <= level-2.
 826          *
 827          * We only consider bushy-plan joins for pairs of rels where there is a
 828          * suitable join clause (or join order restriction), in order to avoid
 829          * unreasonable growth of planning time.
 830          */
 831         for (k = 2;; k++)
 832         {
 833                 int                     other_level = level - k;
 834
 835                 /*
 836                  * Since make_join_rel(x, y) handles both x,y and y,x cases, we only
 837                  * need to go as far as the halfway point.
 838                  */
 839                 if (k > other_level)
 840                         break;
 841
 842                 foreach(r, joinrels[k])
 843                 {
 844                         RelOptInfo *old_rel = (RelOptInfo *) lfirst(r);
 845                         ListCell   *other_rels;
 846                         ListCell   *r2;
 847
 848                         /*
 849                          * We can ignore relations without join clauses here, unless they
 850                          * participate in join-order restrictions --- then we might have
 851                          * to force a bushy join plan.
 852                          */
 853                         if (old_rel->joininfo == NIL && !old_rel->has_eclass_joins &&
 854                                 !has_join_restriction(root, old_rel))
 855                                 continue;
 856
 857                         if (k == other_level)
 858                                 other_rels = lnext(r);  /* only consider remaining rels */
 859                         else
 860                                 other_rels = list_head(joinrels[other_level]);
 861
 862                         for_each_cell(r2, other_rels)
 863                         {
 864                                 RelOptInfo *new_rel = (RelOptInfo *) lfirst(r2);
 865
 866                                 if (!bms_overlap(old_rel->relids, new_rel->relids))
 867                                 {
 868                                         /*
 869                                          * OK, we can build a rel of the right level from this
 870                                          * pair of rels.  Do so if there is at least one relevant
 871                                          * join clause or join order restriction.
 872                                          */
 873                                         if (have_relevant_joinclause(root, old_rel, new_rel) ||
 874                                                 have_join_order_restriction(root, old_rel, new_rel))
 875                                         {
 876                                                 (void) make_join_rel(root, old_rel, new_rel);
 877                                         }
 878                                 }
 879                         }
 880                 }
 881         }
 882
 883         /*----------
 884          * Last-ditch effort: if we failed to find any usable joins so far, force
 885          * a set of cartesian-product joins to be generated.  This handles the
 886          * special case where all the available rels have join clauses but we
 887          * cannot use any of those clauses yet.  This can only happen when we are
 888          * considering a join sub-problem (a sub-joinlist) and all the rels in the
 889          * sub-problem have only join clauses with rels outside the sub-problem.
 890          * An example is
 891          *
 892          *              SELECT ... FROM a INNER JOIN b ON TRUE, c, d, ...
 893          *              WHERE a.w = c.x and b.y = d.z;
 894          *
 895          * If the "a INNER JOIN b" sub-problem does not get flattened into the
 896          * upper level, we must be willing to make a cartesian join of a and b;
 897          * but the code above will not have done so, because it thought that both
 898          * a and b have joinclauses.  We consider only left-sided and right-sided
 899          * cartesian joins in this case (no bushy).
 900          *----------
 901          */
 902         if (joinrels[level] == NIL)
 903         {
 904                 /*
 905                  * This loop is just like the first one, except we always call
 906                  * make_rels_by_clauseless_joins().
 907                  */
 908                 foreach(r, joinrels[level - 1])
 909                 {
 910                         RelOptInfo *old_rel = (RelOptInfo *) lfirst(r);
 911
 912                         make_rels_by_clauseless_joins(root,
 913                                                                                   old_rel,
 914                                                                                   list_head(joinrels[1]));
 915                 }
 916
 917                 /*----------
 918                  * When special joins are involved, there may be no legal way
 919                  * to make an N-way join for some values of N.  For example consider
 920                  *
 921                  * SELECT ... FROM t1 WHERE
 922                  *       x IN (SELECT ... FROM t2,t3 WHERE ...) AND
 923                  *       y IN (SELECT ... FROM t4,t5 WHERE ...)
 924                  *
 925                  * We will flatten this query to a 5-way join problem, but there are
 926                  * no 4-way joins that join_is_legal() will consider legal.  We have
 927                  * to accept failure at level 4 and go on to discover a workable
 928                  * bushy plan at level 5.
 929                  *
 930                  * However, if there are no special joins and no lateral references
 931                  * then join_is_legal() should never fail, and so the following sanity
 932                  * check is useful.
 933                  *----------
 934                  */
 935                 if (joinrels[level] == NIL &&
 936                         root->join_info_list == NIL &&
 937                         !root->hasLateralRTEs)
 938                         elog(ERROR, "failed to build any %d-way joins", level);
 939         }
 940 }
 941
 942
 943 /*
 944  * make_rels_by_clause_joins
 945  *        Build joins between the given relation 'old_rel' and other relations
 946  *        that participate in join clauses that 'old_rel' also participates in
 947  *        (or participate in join-order restrictions with it).
 948  *        The join rels are returned in root->join_rel_level[join_cur_level].
 949  *
 950  * Note: at levels above 2 we will generate the same joined relation in
 951  * multiple ways --- for example (a join b) join c is the same RelOptInfo as
 952  * (b join c) join a, though the second case will add a different set of Paths
 953  * to it.  This is the reason for using the join_rel_level mechanism, which
 954  * automatically ensures that each new joinrel is only added to the list once.
 955  *
 956  * 'old_rel' is the relation entry for the relation to be joined
 957  * 'other_rels': the first cell in a linked list containing the other
 958  * rels to be considered for joining
 959  *
 960  * Currently, this is only used with initial rels in other_rels, but it
 961  * will work for joining to joinrels too.
 962  */
 963 static void
 964 make_rels_by_clause_joins(PlannerInfo *root,
 965                                                   RelOptInfo *old_rel,
 966                                                   ListCell *other_rels)
 967 {
 968         ListCell   *l;
 969
 970         for_each_cell(l, other_rels)
 971         {
 972                 RelOptInfo *other_rel = (RelOptInfo *) lfirst(l);
 973
 974                 if (!bms_overlap(old_rel->relids, other_rel->relids) &&
 975                         (have_relevant_joinclause(root, old_rel, other_rel) ||
 976                          have_join_order_restriction(root, old_rel, other_rel)))
 977                 {
 978                         (void) make_join_rel(root, old_rel, other_rel);
 979                 }
 980         }
 981 }
 982
 983
 984 /*
 985  * make_rels_by_clauseless_joins
 986  *        Given a relation 'old_rel' and a list of other relations
 987  *        'other_rels', create a join relation between 'old_rel' and each
 988  *        member of 'other_rels' that isn't already included in 'old_rel'.
 989  *        The join rels are returned in root->join_rel_level[join_cur_level].
 990  *
 991  * 'old_rel' is the relation entry for the relation to be joined
 992  * 'other_rels': the first cell of a linked list containing the
 993  * other rels to be considered for joining
 994  *
 995  * Currently, this is only used with initial rels in other_rels, but it would
 996  * work for joining to joinrels too.
 997  */
 998 static void
 999 make_rels_by_clauseless_joins(PlannerInfo *root,
1000                                                           RelOptInfo *old_rel,
1001                                                           ListCell *other_rels)
1002 {
1003         ListCell   *l;
1004
1005         for_each_cell(l, other_rels)
1006         {
1007                 RelOptInfo *other_rel = (RelOptInfo *) lfirst(l);
1008
1009                 if (!bms_overlap(other_rel->relids, old_rel->relids))
1010                 {
1011                         (void) make_join_rel(root, old_rel, other_rel);
1012                 }
1013         }
1014 }
1015
1016
1017 /*
1018  * join_is_legal
1019  *         Determine whether a proposed join is legal given the query's
1020  *         join order constraints; and if it is, determine the join type.
1021  *
1022  * Caller must supply not only the two rels, but the union of their relids.
1023  * (We could simplify the API by computing joinrelids locally, but this
1024  * would be redundant work in the normal path through make_join_rel.)
1025  *
1026  * On success, *sjinfo_p is set to NULL if this is to be a plain inner join,
1027  * else it's set to point to the associated SpecialJoinInfo node.  Also,
1028  * *reversed_p is set TRUE if the given relations need to be swapped to
1029  * match the SpecialJoinInfo node.
1030  */
1031 static bool
1032 join_is_legal(PlannerInfo *root, RelOptInfo *rel1, RelOptInfo *rel2,
1033                           Relids joinrelids,
1034                           SpecialJoinInfo **sjinfo_p, bool *reversed_p)
1035 {
1036         SpecialJoinInfo *match_sjinfo;
1037         bool            reversed;
1038         bool            unique_ified;
1039         bool            must_be_leftjoin;
1040         ListCell   *l;
1041
1042         /*
1043          * Ensure output params are set on failure return.  This is just to
1044          * suppress uninitialized-variable warnings from overly anal compilers.
1045          */
1046         *sjinfo_p = NULL;
1047         *reversed_p = false;
1048
1049         /*
1050          * If we have any special joins, the proposed join might be illegal; and
1051          * in any case we have to determine its join type.  Scan the join info
1052          * list for matches and conflicts.
1053          */
1054         match_sjinfo = NULL;
1055         reversed = false;
1056         unique_ified = false;
1057         must_be_leftjoin = false;
1058
1059         foreach(l, root->join_info_list)
1060         {
1061                 SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
1062
1063                 /*
1064                  * This special join is not relevant unless its RHS overlaps the
1065                  * proposed join.  (Check this first as a fast path for dismissing
1066                  * most irrelevant SJs quickly.)
1067                  */
1068                 if (!bms_overlap(sjinfo->min_righthand, joinrelids))
1069                         continue;
1070
1071                 /*
1072                  * Also, not relevant if proposed join is fully contained within RHS
1073                  * (ie, we're still building up the RHS).
1074                  */
1075                 if (bms_is_subset(joinrelids, sjinfo->min_righthand))
1076                         continue;
1077
1078                 /*
1079                  * Also, not relevant if SJ is already done within either input.
1080                  */
1081                 if (bms_is_subset(sjinfo->min_lefthand, rel1->relids) &&
1082                         bms_is_subset(sjinfo->min_righthand, rel1->relids))
1083                         continue;
1084                 if (bms_is_subset(sjinfo->min_lefthand, rel2->relids) &&
1085                         bms_is_subset(sjinfo->min_righthand, rel2->relids))
1086                         continue;
1087
1088                 /*
1089                  * If it's a semijoin and we already joined the RHS to any other rels
1090                  * within either input, then we must have unique-ified the RHS at that
1091                  * point (see below).  Therefore the semijoin is no longer relevant in
1092                  * this join path.
1093                  */
1094                 if (sjinfo->jointype == JOIN_SEMI)
1095                 {
1096                         if (bms_is_subset(sjinfo->syn_righthand, rel1->relids) &&
1097                                 !bms_equal(sjinfo->syn_righthand, rel1->relids))
1098                                 continue;
1099                         if (bms_is_subset(sjinfo->syn_righthand, rel2->relids) &&
1100                                 !bms_equal(sjinfo->syn_righthand, rel2->relids))
1101                                 continue;
1102                 }
1103
1104                 /*
1105                  * If one input contains min_lefthand and the other contains
1106                  * min_righthand, then we can perform the SJ at this join.
1107                  *
1108                  * Reject if we get matches to more than one SJ; that implies we're
1109                  * considering something that's not really valid.
1110                  */
1111                 if (bms_is_subset(sjinfo->min_lefthand, rel1->relids) &&
1112                         bms_is_subset(sjinfo->min_righthand, rel2->relids))
1113                 {
1114                         if (match_sjinfo)
1115                                 return false;   /* invalid join path */
1116                         match_sjinfo = sjinfo;
1117                         reversed = false;
1118                 }
1119                 else if (bms_is_subset(sjinfo->min_lefthand, rel2->relids) &&
1120                                  bms_is_subset(sjinfo->min_righthand, rel1->relids))
1121                 {
1122                         if (match_sjinfo)
1123                                 return false;   /* invalid join path */
1124                         match_sjinfo = sjinfo;
1125                         reversed = true;
1126                 }
1127                 else if (sjinfo->jointype == JOIN_SEMI &&
1128                                  bms_equal(sjinfo->syn_righthand, rel2->relids) &&
1129                                  create_unique_path(root, rel2, rel2->cheapest_total_path,
1130                                                                         sjinfo) != NULL)
1131                 {
1132                         /*----------
1133                          * For a semijoin, we can join the RHS to anything else by
1134                          * unique-ifying the RHS (if the RHS can be unique-ified).
1135                          * We will only get here if we have the full RHS but less
1136                          * than min_lefthand on the LHS.
1137                          *
1138                          * The reason to consider such a join path is exemplified by
1139                          *      SELECT ... FROM a,b WHERE (a.x,b.y) IN (SELECT c1,c2 FROM c)
1140                          * If we insist on doing this as a semijoin we will first have
1141                          * to form the cartesian product of A*B.  But if we unique-ify
1142                          * C then the semijoin becomes a plain innerjoin and we can join
1143                          * in any order, eg C to A and then to B.  When C is much smaller
1144                          * than A and B this can be a huge win.  So we allow C to be
1145                          * joined to just A or just B here, and then make_join_rel has
1146                          * to handle the case properly.
1147                          *
1148                          * Note that actually we'll allow unique-ified C to be joined to
1149                          * some other relation D here, too.  That is legal, if usually not
1150                          * very sane, and this routine is only concerned with legality not
1151                          * with whether the join is good strategy.
1152                          *----------
1153                          */
1154                         if (match_sjinfo)
1155                                 return false;   /* invalid join path */
1156                         match_sjinfo = sjinfo;
1157                         reversed = false;
1158                         unique_ified = true;
1159                 }
1160                 else if (sjinfo->jointype == JOIN_SEMI &&
1161                                  bms_equal(sjinfo->syn_righthand, rel1->relids) &&
1162                                  create_unique_path(root, rel1, rel1->cheapest_total_path,
1163                                                                         sjinfo) != NULL)
1164                 {
1165                         /* Reversed semijoin case */
1166                         if (match_sjinfo)
1167                                 return false;   /* invalid join path */
1168                         match_sjinfo = sjinfo;
1169                         reversed = true;
1170                         unique_ified = true;
1171                 }
1172                 else
1173                 {
1174                         /*
1175                          * Otherwise, the proposed join overlaps the RHS but isn't a valid
1176                          * implementation of this SJ.  But don't panic quite yet: the RHS
1177                          * violation might have occurred previously, in one or both input
1178                          * relations, in which case we must have previously decided that
1179                          * it was OK to commute some other SJ with this one.  If we need
1180                          * to perform this join to finish building up the RHS, rejecting
1181                          * it could lead to not finding any plan at all.  (This can occur
1182                          * because of the heuristics elsewhere in this file that postpone
1183                          * clauseless joins: we might not consider doing a clauseless join
1184                          * within the RHS until after we've performed other, validly
1185                          * commutable SJs with one or both sides of the clauseless join.)
1186                          * This consideration boils down to the rule that if both inputs
1187                          * overlap the RHS, we can allow the join --- they are either
1188                          * fully within the RHS, or represent previously-allowed joins to
1189                          * rels outside it.
1190                          */
1191                         if (bms_overlap(rel1->relids, sjinfo->min_righthand) &&
1192                                 bms_overlap(rel2->relids, sjinfo->min_righthand))
1193                                 continue;               /* assume valid previous violation of RHS */
1194
1195                         /*
1196                          * The proposed join could still be legal, but only if we're
1197                          * allowed to associate it into the RHS of this SJ.  That means
1198                          * this SJ must be a LEFT join (not SEMI or ANTI, and certainly
1199                          * not FULL) and the proposed join must not overlap the LHS.
1200                          */
1201                         if (sjinfo->jointype != JOIN_LEFT ||
1202                                 bms_overlap(joinrelids, sjinfo->min_lefthand))
1203                                 return false;   /* invalid join path */
1204
1205                         /*
1206                          * To be valid, the proposed join must be a LEFT join; otherwise
1207                          * it can't associate into this SJ's RHS.  But we may not yet have
1208                          * found the SpecialJoinInfo matching the proposed join, so we
1209                          * can't test that yet.  Remember the requirement for later.
1210                          */
1211                         must_be_leftjoin = true;
1212                 }
1213         }
1214
1215         /*
1216          * Fail if violated any SJ's RHS and didn't match to a LEFT SJ: the
1217          * proposed join can't associate into an SJ's RHS.
1218          *
1219          * Also, fail if the proposed join's predicate isn't strict; we're
1220          * essentially checking to see if we can apply outer-join identity 3, and
1221          * that's a requirement.  (This check may be redundant with checks in
1222          * make_outerjoininfo, but I'm not quite sure, and it's cheap to test.)
1223          */
1224         if (must_be_leftjoin &&
1225                 (match_sjinfo == NULL ||
1226                  match_sjinfo->jointype != JOIN_LEFT ||
1227                  !match_sjinfo->lhs_strict))
1228                 return false;                   /* invalid join path */
1229
1230         /*
1231          * We also have to check for constraints imposed by LATERAL references.
1232          */
1233         if (root->hasLateralRTEs)
1234         {
1235                 bool            lateral_fwd;
1236                 bool            lateral_rev;
1237                 Relids          join_lateral_rels;
1238
1239                 /*
1240                  * The proposed rels could each contain lateral references to the
1241                  * other, in which case the join is impossible.  If there are lateral
1242                  * references in just one direction, then the join has to be done with
1243                  * a nestloop with the lateral referencer on the inside.  If the join
1244                  * matches an SJ that cannot be implemented by such a nestloop, the
1245                  * join is impossible.
1246                  *
1247                  * Also, if the lateral reference is only indirect, we should reject
1248                  * the join; whatever rel(s) the reference chain goes through must be
1249                  * joined to first.
1250                  *
1251                  * Another case that might keep us from building a valid plan is the
1252                  * implementation restriction described by have_dangerous_phv().
1253                  */
1254                 lateral_fwd = bms_overlap(rel1->relids, rel2->lateral_relids);
1255                 lateral_rev = bms_overlap(rel2->relids, rel1->lateral_relids);
1256                 if (lateral_fwd && lateral_rev)
1257                         return false;           /* have lateral refs in both directions */
1258                 if (lateral_fwd)
1259                 {
1260                         /* has to be implemented as nestloop with rel1 on left */
1261                         if (match_sjinfo &&
1262                                 (reversed ||
1263                                  unique_ified ||
1264                                  match_sjinfo->jointype == JOIN_FULL))
1265                                 return false;   /* not implementable as nestloop */
1266                         /* check there is a direct reference from rel2 to rel1 */
1267                         if (!bms_overlap(rel1->relids, rel2->direct_lateral_relids))
1268                                 return false;   /* only indirect refs, so reject */
1269                         /* check we won't have a dangerous PHV */
1270                         if (have_dangerous_phv(root, rel1->relids, rel2->lateral_relids))
1271                                 return false;   /* might be unable to handle required PHV */
1272                 }
1273                 else if (lateral_rev)
1274                 {
1275                         /* has to be implemented as nestloop with rel2 on left */
1276                         if (match_sjinfo &&
1277                                 (!reversed ||
1278                                  unique_ified ||
1279                                  match_sjinfo->jointype == JOIN_FULL))
1280                                 return false;   /* not implementable as nestloop */
1281                         /* check there is a direct reference from rel1 to rel2 */
1282                         if (!bms_overlap(rel2->relids, rel1->direct_lateral_relids))
1283                                 return false;   /* only indirect refs, so reject */
1284                         /* check we won't have a dangerous PHV */
1285                         if (have_dangerous_phv(root, rel2->relids, rel1->lateral_relids))
1286                                 return false;   /* might be unable to handle required PHV */
1287                 }
1288
1289                 /*
1290                  * LATERAL references could also cause problems later on if we accept
1291                  * this join: if the join's minimum parameterization includes any rels
1292                  * that would have to be on the inside of an outer join with this join
1293                  * rel, then it's never going to be possible to build the complete
1294                  * query using this join.  We should reject this join not only because
1295                  * it'll save work, but because if we don't, the clauseless-join
1296                  * heuristics might think that legality of this join means that some
1297                  * other join rel need not be formed, and that could lead to failure
1298                  * to find any plan at all.  We have to consider not only rels that
1299                  * are directly on the inner side of an OJ with the joinrel, but also
1300                  * ones that are indirectly so, so search to find all such rels.
1301                  */
1302                 join_lateral_rels = min_join_parameterization(root, joinrelids,
1303                                                                                                           rel1, rel2);
1304                 if (join_lateral_rels)
1305                 {
1306                         Relids          join_plus_rhs = bms_copy(joinrelids);
1307                         bool            more;
1308
1309                         do
1310                         {
1311                                 more = false;
1312                                 foreach(l, root->join_info_list)
1313                                 {
1314                                         SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
1315
1316                                         /* ignore full joins --- their ordering is predetermined */
1317                                         if (sjinfo->jointype == JOIN_FULL)
1318                                                 continue;
1319
1320                                         if (bms_overlap(sjinfo->min_lefthand, join_plus_rhs) &&
1321                                                 !bms_is_subset(sjinfo->min_righthand, join_plus_rhs))
1322                                         {
1323                                                 join_plus_rhs = bms_add_members(join_plus_rhs,
1324                                                                                                                 sjinfo->min_righthand);
1325                                                 more = true;
1326                                         }
1327                                 }
1328                         } while (more);
1329                         if (bms_overlap(join_plus_rhs, join_lateral_rels))
1330                                 return false;   /* will not be able to join to some RHS rel */
1331                 }
1332         }
1333
1334         /* Otherwise, it's a valid join */
1335         *sjinfo_p = match_sjinfo;
1336         *reversed_p = reversed;
1337         return true;
1338 }
1339
1340
1341 /*
1342  * has_join_restriction
1343  *              Detect whether the specified relation has join-order restrictions,
1344  *              due to being inside an outer join or an IN (sub-SELECT),
1345  *              or participating in any LATERAL references or multi-rel PHVs.
1346  *
1347  * Essentially, this tests whether have_join_order_restriction() could
1348  * succeed with this rel and some other one.  It's OK if we sometimes
1349  * say "true" incorrectly.  (Therefore, we don't bother with the relatively
1350  * expensive has_legal_joinclause test.)
1351  */
1352 static bool
1353 has_join_restriction(PlannerInfo *root, RelOptInfo *rel)
1354 {
1355         ListCell   *l;
1356
1357         if (rel->lateral_relids != NULL || rel->lateral_referencers != NULL)
1358                 return true;
1359
1360         foreach(l, root->placeholder_list)
1361         {
1362                 PlaceHolderInfo *phinfo = (PlaceHolderInfo *) lfirst(l);
1363
1364                 if (bms_is_subset(rel->relids, phinfo->ph_eval_at) &&
1365                         !bms_equal(rel->relids, phinfo->ph_eval_at))
1366                         return true;
1367         }
1368
1369         foreach(l, root->join_info_list)
1370         {
1371                 SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
1372
1373                 /* ignore full joins --- other mechanisms preserve their ordering */
1374                 if (sjinfo->jointype == JOIN_FULL)
1375                         continue;
1376
1377                 /* ignore if SJ is already contained in rel */
1378                 if (bms_is_subset(sjinfo->min_lefthand, rel->relids) &&
1379                         bms_is_subset(sjinfo->min_righthand, rel->relids))
1380                         continue;
1381
1382                 /* restricted if it overlaps LHS or RHS, but doesn't contain SJ */
1383                 if (bms_overlap(sjinfo->min_lefthand, rel->relids) ||
1384                         bms_overlap(sjinfo->min_righthand, rel->relids))
1385                         return true;
1386         }
1387
1388         return false;
1389 }
1390
1391
1392 /*
1393  * Mark a relation as proven empty.
1394  *
1395  * During GEQO planning, this can get invoked more than once on the same
1396  * baserel struct, so it's worth checking to see if the rel is already marked
1397  * dummy.
1398  *
1399  * Also, when called during GEQO join planning, we are in a short-lived
1400  * memory context.  We must make sure that the dummy path attached to a
1401  * baserel survives the GEQO cycle, else the baserel is trashed for future
1402  * GEQO cycles.  On the other hand, when we are marking a joinrel during GEQO,
1403  * we don't want the dummy path to clutter the main planning context.  Upshot
1404  * is that the best solution is to explicitly make the dummy path in the same
1405  * context the given RelOptInfo is in.
1406  */
1407 static void
1408 mark_dummy_rel(RelOptInfo *rel)
1409 {
1410         MemoryContext oldcontext;
1411
1412         /* Already marked? */
1413         if (is_dummy_rel(rel))
1414                 return;
1415
1416         /* No, so choose correct context to make the dummy path in */
1417         oldcontext = MemoryContextSwitchTo(GetMemoryChunkContext(rel));
1418
1419         /* Set dummy size estimate */
1420         rel->rows = 0;
1421
1422         /* Evict any previously chosen paths */
1423         rel->pathlist = NIL;
1424         rel->partial_pathlist = NIL;
1425
1426         /* Set up the dummy path */
1427         add_path(rel, (Path *) create_append_path(rel, NIL,
1428                                                                                           rel->lateral_relids,
1429                                                                                           0, NIL));
1430
1431         /* Set or update cheapest_total_path and related fields */
1432         set_cheapest(rel);
1433
1434         MemoryContextSwitchTo(oldcontext);
1435 }
1436
1437
1438 /*
1439  * restriction_is_constant_false --- is a restrictlist just false?
1440  *
1441  * In cases where a qual is provably constant false, eval_const_expressions
1442  * will generally have thrown away anything that's ANDed with it.  In outer
1443  * join situations this will leave us computing cartesian products only to
1444  * decide there's no match for an outer row, which is pretty stupid.  So,
1445  * we need to detect the case.
1446  *
1447  * If only_pushed_down is true, then consider only quals that are pushed-down
1448  * from the point of view of the joinrel.
1449  */
1450 static bool
1451 restriction_is_constant_false(List *restrictlist,
1452                                                           RelOptInfo *joinrel,
1453                                                           bool only_pushed_down)
1454 {
1455         ListCell   *lc;
1456
1457         /*
1458          * Despite the above comment, the restriction list we see here might
1459          * possibly have other members besides the FALSE constant, since other
1460          * quals could get "pushed down" to the outer join level.  So we check
1461          * each member of the list.
1462          */
1463         foreach(lc, restrictlist)
1464         {
1465                 RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc);
1466
1467                 if (only_pushed_down && !RINFO_IS_PUSHED_DOWN(rinfo, joinrel->relids))
1468                         continue;
1469
1470                 if (rinfo->clause && IsA(rinfo->clause, Const))
1471                 {
1472                         Const      *con = (Const *) rinfo->clause;
1473
1474                         /* constant NULL is as good as constant FALSE for our purposes */
1475                         if (con->constisnull)
1476                                 return true;
1477                         if (!DatumGetBool(con->constvalue))
1478                                 return true;
1479                 }
1480         }
1481         return false;
1482 }