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  *     is_dummy_rel()
  44  *     mark_dummy_rel()
  45  *     restriction_is_constant_false()
  46  *
  47  *
  48  * Portions Copyright (c) 1996-2017, PostgreSQL Global Development Group
  49  * Portions Copyright (c) 1994, Regents of the University of California
  50  *
  51  *-------------------------------------------------------------------------
  52  */
  53
  54
  55 /*
  56  * set_plain_rel_pathlist
  57  *        Build access paths for a plain relation (no subquery, no inheritance)
  58  */
  59 static void
  60 set_plain_rel_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
  61 {
  62         Relids          required_outer;
  63
  64         /*
  65          * We don't support pushing join clauses into the quals of a seqscan, but
  66          * it could still have required parameterization due to LATERAL refs in
  67          * its tlist.
  68          */
  69         required_outer = rel->lateral_relids;
  70
  71         /* Consider sequential scan */
  72         add_path(rel, create_seqscan_path(root, rel, required_outer, 0));
  73
  74         /* If appropriate, consider parallel sequential scan */
  75         if (rel->consider_parallel && required_outer == NULL)
  76                 create_plain_partial_paths(root, rel);
  77
  78         /* Consider index scans */
  79         create_index_paths(root, rel);
  80
  81         /* Consider TID scans */
  82         create_tidscan_paths(root, rel);
  83 }
  84
  85
  86 /*
  87  * create_plain_partial_paths
  88  *        Build partial access paths for parallel scan of a plain relation
  89  */
  90 static void
  91 create_plain_partial_paths(PlannerInfo *root, RelOptInfo *rel)
  92 {
  93         int                     parallel_workers;
  94
  95         parallel_workers = compute_parallel_worker(rel, rel->pages, -1);
  96
  97         /* If any limit was set to zero, the user doesn't want a parallel scan. */
  98         if (parallel_workers <= 0)
  99                 return;
 100
 101         /* Add an unordered partial path based on a parallel sequential scan. */
 102         add_partial_path(rel, create_seqscan_path(root, rel, NULL, parallel_workers));
 103 }
 104
 105
 106 /*
 107  * set_append_rel_pathlist
 108  *        Build access paths for an "append relation"
 109  */
 110 static void
 111 set_append_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
 112                                                 Index rti, RangeTblEntry *rte)
 113 {
 114         int                     parentRTindex = rti;
 115         List       *live_childrels = NIL;
 116         ListCell   *l;
 117
 118         /*
 119          * Generate access paths for each member relation, and remember the
 120          * non-dummy children.
 121          */
 122         foreach(l, root->append_rel_list)
 123         {
 124                 AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
 125                 int                     childRTindex;
 126                 RangeTblEntry *childRTE;
 127                 RelOptInfo *childrel;
 128
 129                 /* append_rel_list contains all append rels; ignore others */
 130                 if (appinfo->parent_relid != parentRTindex)
 131                         continue;
 132
 133                 /* Re-locate the child RTE and RelOptInfo */
 134                 childRTindex = appinfo->child_relid;
 135                 childRTE = root->simple_rte_array[childRTindex];
 136                 childrel = root->simple_rel_array[childRTindex];
 137
 138                 /*
 139                  * If set_append_rel_size() decided the parent appendrel was
 140                  * parallel-unsafe at some point after visiting this child rel, we
 141                  * need to propagate the unsafety marking down to the child, so that
 142                  * we don't generate useless partial paths for it.
 143                  */
 144                 if (!rel->consider_parallel)
 145                         childrel->consider_parallel = false;
 146
 147                 /*
 148                  * Compute the child's access paths.
 149                  */
 150                 set_rel_pathlist(root, childrel, childRTindex, childRTE);
 151
 152                 /*
 153                  * If child is dummy, ignore it.
 154                  */
 155                 if (IS_DUMMY_REL(childrel))
 156                         continue;
 157
 158                 /*
 159                  * Child is live, so add it to the live_childrels list for use below.
 160                  */
 161                 live_childrels = lappend(live_childrels, childrel);
 162         }
 163
 164         /* Add paths to the "append" relation. */
 165         add_paths_to_append_rel(root, rel, live_childrels);
 166 }
 167
 168 /*
 169  * add_paths_to_append_rel
 170  *              Generate paths for given "append" relation given the set of non-dummy
 171  *              child rels.
 172  *
 173  * The function collects all parameterizations and orderings supported by the
 174  * non-dummy children. For every such parameterization or ordering, it creates
 175  * an append path collecting one path from each non-dummy child with given
 176  * parameterization or ordering. Similarly it collects partial paths from
 177  * non-dummy children to create partial append paths.
 178  */
 179 static void
 180 add_paths_to_append_rel(PlannerInfo *root, RelOptInfo *rel,
 181                                                 List *live_childrels)
 182 {
 183         List       *subpaths = NIL;
 184         bool            subpaths_valid = true;
 185         List       *partial_subpaths = NIL;
 186         bool            partial_subpaths_valid = true;
 187         List       *all_child_pathkeys = NIL;
 188         List       *all_child_outers = NIL;
 189         ListCell   *l;
 190         List       *partitioned_rels = NIL;
 191         RangeTblEntry *rte;
 192         bool            build_partitioned_rels = false;
 193
 194         /*
 195          * A plain relation will already have a PartitionedChildRelInfo if it is
 196          * partitioned.  For a subquery RTE, no PartitionedChildRelInfo exists; we
 197          * collect all partitioned_rels associated with any child.  (This assumes
 198          * that we don't need to look through multiple levels of subquery RTEs; if
 199          * we ever do, we could create a PartitionedChildRelInfo with the
 200          * accumulated list of partitioned_rels which would then be found when
 201          * populated our parent rel with paths.  For the present, that appears to
 202          * be unnecessary.)
 203          */
 204         rte = planner_rt_fetch(rel->relid, root);
 205         switch (rte->rtekind)
 206         {
 207                 case RTE_RELATION:
 208                         if (rte->relkind == RELKIND_PARTITIONED_TABLE)
 209                         {
 210                                 partitioned_rels =
 211                                         get_partitioned_child_rels(root, rel->relid);
 212                                 Assert(list_length(partitioned_rels) >= 1);
 213                         }
 214                         break;
 215                 case RTE_SUBQUERY:
 216                         build_partitioned_rels = true;
 217                         break;
 218                 default:
 219                         elog(ERROR, "unexpected rtekind: %d", (int) rte->rtekind);
 220         }
 221
 222         /*
 223          * For every non-dummy child, remember the cheapest path.  Also, identify
 224          * all pathkeys (orderings) and parameterizations (required_outer sets)
 225          * available for the non-dummy member relations.
 226          */
 227         foreach(l, live_childrels)
 228         {
 229                 RelOptInfo *childrel = lfirst(l);
 230                 ListCell   *lcp;
 231
 232                 /*
 233                  * If we need to build partitioned_rels, accumulate the partitioned
 234                  * rels for this child.
 235                  */
 236                 if (build_partitioned_rels)
 237                 {
 238                         List       *cprels;
 239
 240                         cprels = get_partitioned_child_rels(root, childrel->relid);
 241                         partitioned_rels = list_concat(partitioned_rels,
 242                                                                                    list_copy(cprels));
 243                 }
 244
 245                 /*
 246                  * If child has an unparameterized cheapest-total path, add that to
 247                  * the unparameterized Append path we are constructing for the parent.
 248                  * If not, there's no workable unparameterized path.
 249                  */
 250                 if (childrel->cheapest_total_path->param_info == NULL)
 251                         subpaths = accumulate_append_subpath(subpaths,
 252                                                                                                  childrel->cheapest_total_path);
 253                 else
 254                         subpaths_valid = false;
 255
 256                 /* Same idea, but for a partial plan. */
 257                 if (childrel->partial_pathlist != NIL)
 258                         partial_subpaths = accumulate_append_subpath(partial_subpaths,
 259                                                                                                                  linitial(childrel->partial_pathlist));
 260                 else
 261                         partial_subpaths_valid = false;
 262
 263                 /*
 264                  * Collect lists of all the available path orderings and
 265                  * parameterizations for all the children.  We use these as a
 266                  * heuristic to indicate which sort orderings and parameterizations we
 267                  * should build Append and MergeAppend paths for.
 268                  */
 269                 foreach(lcp, childrel->pathlist)
 270                 {
 271                         Path       *childpath = (Path *) lfirst(lcp);
 272                         List       *childkeys = childpath->pathkeys;
 273                         Relids          childouter = PATH_REQ_OUTER(childpath);
 274
 275                         /* Unsorted paths don't contribute to pathkey list */
 276                         if (childkeys != NIL)
 277                         {
 278                                 ListCell   *lpk;
 279                                 bool            found = false;
 280
 281                                 /* Have we already seen this ordering? */
 282                                 foreach(lpk, all_child_pathkeys)
 283                                 {
 284                                         List       *existing_pathkeys = (List *) lfirst(lpk);
 285
 286                                         if (compare_pathkeys(existing_pathkeys,
 287                                                                                  childkeys) == PATHKEYS_EQUAL)
 288                                         {
 289                                                 found = true;
 290                                                 break;
 291                                         }
 292                                 }
 293                                 if (!found)
 294                                 {
 295                                         /* No, so add it to all_child_pathkeys */
 296                                         all_child_pathkeys = lappend(all_child_pathkeys,
 297                                                                                                  childkeys);
 298                                 }
 299                         }
 300
 301                         /* Unparameterized paths don't contribute to param-set list */
 302                         if (childouter)
 303                         {
 304                                 ListCell   *lco;
 305                                 bool            found = false;
 306
 307                                 /* Have we already seen this param set? */
 308                                 foreach(lco, all_child_outers)
 309                                 {
 310                                         Relids          existing_outers = (Relids) lfirst(lco);
 311
 312                                         if (bms_equal(existing_outers, childouter))
 313                                         {
 314                                                 found = true;
 315                                                 break;
 316                                         }
 317                                 }
 318                                 if (!found)
 319                                 {
 320                                         /* No, so add it to all_child_outers */
 321                                         all_child_outers = lappend(all_child_outers,
 322                                                                                            childouter);
 323                                 }
 324                         }
 325                 }
 326         }
 327
 328         /*
 329          * If we found unparameterized paths for all children, build an unordered,
 330          * unparameterized Append path for the rel.  (Note: this is correct even
 331          * if we have zero or one live subpath due to constraint exclusion.)
 332          */
 333         if (subpaths_valid)
 334                 add_path(rel, (Path *) create_append_path(rel, subpaths, NULL, 0,
 335                                                                                                   partitioned_rels));
 336
 337         /*
 338          * Consider an append of partial unordered, unparameterized partial paths.
 339          */
 340         if (partial_subpaths_valid)
 341         {
 342                 AppendPath *appendpath;
 343                 ListCell   *lc;
 344                 int                     parallel_workers = 0;
 345
 346                 /*
 347                  * Decide on the number of workers to request for this append path.
 348                  * For now, we just use the maximum value from among the members.  It
 349                  * might be useful to use a higher number if the Append node were
 350                  * smart enough to spread out the workers, but it currently isn't.
 351                  */
 352                 foreach(lc, partial_subpaths)
 353                 {
 354                         Path       *path = lfirst(lc);
 355
 356                         parallel_workers = Max(parallel_workers, path->parallel_workers);
 357                 }
 358                 Assert(parallel_workers > 0);
 359
 360                 /* Generate a partial append path. */
 361                 appendpath = create_append_path(rel, partial_subpaths, NULL,
 362                                                                                 parallel_workers, partitioned_rels);
 363                 add_partial_path(rel, (Path *) appendpath);
 364         }
 365
 366         /*
 367          * Also build unparameterized MergeAppend paths based on the collected
 368          * list of child pathkeys.
 369          */
 370         if (subpaths_valid)
 371                 generate_mergeappend_paths(root, rel, live_childrels,
 372                                                                    all_child_pathkeys,
 373                                                                    partitioned_rels);
 374
 375         /*
 376          * Build Append paths for each parameterization seen among the child rels.
 377          * (This may look pretty expensive, but in most cases of practical
 378          * interest, the child rels will expose mostly the same parameterizations,
 379          * so that not that many cases actually get considered here.)
 380          *
 381          * The Append node itself cannot enforce quals, so all qual checking must
 382          * be done in the child paths.  This means that to have a parameterized
 383          * Append path, we must have the exact same parameterization for each
 384          * child path; otherwise some children might be failing to check the
 385          * moved-down quals.  To make them match up, we can try to increase the
 386          * parameterization of lesser-parameterized paths.
 387          */
 388         foreach(l, all_child_outers)
 389         {
 390                 Relids          required_outer = (Relids) lfirst(l);
 391                 ListCell   *lcr;
 392
 393                 /* Select the child paths for an Append with this parameterization */
 394                 subpaths = NIL;
 395                 subpaths_valid = true;
 396                 foreach(lcr, live_childrels)
 397                 {
 398                         RelOptInfo *childrel = (RelOptInfo *) lfirst(lcr);
 399                         Path       *subpath;
 400
 401                         subpath = get_cheapest_parameterized_child_path(root,
 402                                                                                                                         childrel,
 403                                                                                                                         required_outer);
 404                         if (subpath == NULL)
 405                         {
 406                                 /* failed to make a suitable path for this child */
 407                                 subpaths_valid = false;
 408                                 break;
 409                         }
 410                         subpaths = accumulate_append_subpath(subpaths, subpath);
 411                 }
 412
 413                 if (subpaths_valid)
 414                         add_path(rel, (Path *)
 415                                          create_append_path(rel, subpaths, required_outer, 0,
 416                                                                                 partitioned_rels));
 417         }
 418 }
 419
 420
 421 /*
 422  * generate_mergeappend_paths
 423  *              Generate MergeAppend paths for an append relation
 424  *
 425  * Generate a path for each ordering (pathkey list) appearing in
 426  * all_child_pathkeys.
 427  *
 428  * We consider both cheapest-startup and cheapest-total cases, ie, for each
 429  * interesting ordering, collect all the cheapest startup subpaths and all the
 430  * cheapest total paths, and build a MergeAppend path for each case.
 431  *
 432  * We don't currently generate any parameterized MergeAppend paths.  While
 433  * it would not take much more code here to do so, it's very unclear that it
 434  * is worth the planning cycles to investigate such paths: there's little
 435  * use for an ordered path on the inside of a nestloop.  In fact, it's likely
 436  * that the current coding of add_path would reject such paths out of hand,
 437  * because add_path gives no credit for sort ordering of parameterized paths,
 438  * and a parameterized MergeAppend is going to be more expensive than the
 439  * corresponding parameterized Append path.  If we ever try harder to support
 440  * parameterized mergejoin plans, it might be worth adding support for
 441  * parameterized MergeAppends to feed such joins.  (See notes in
 442  * optimizer/README for why that might not ever happen, though.)
 443  */
 444 static void
 445 generate_mergeappend_paths(PlannerInfo *root, RelOptInfo *rel,
 446                                                    List *live_childrels,
 447                                                    List *all_child_pathkeys,
 448                                                    List *partitioned_rels)
 449 {
 450         ListCell   *lcp;
 451
 452         foreach(lcp, all_child_pathkeys)
 453         {
 454                 List       *pathkeys = (List *) lfirst(lcp);
 455                 List       *startup_subpaths = NIL;
 456                 List       *total_subpaths = NIL;
 457                 bool            startup_neq_total = false;
 458                 ListCell   *lcr;
 459
 460                 /* Select the child paths for this ordering... */
 461                 foreach(lcr, live_childrels)
 462                 {
 463                         RelOptInfo *childrel = (RelOptInfo *) lfirst(lcr);
 464                         Path       *cheapest_startup,
 465                                            *cheapest_total;
 466
 467                         /* Locate the right paths, if they are available. */
 468                         cheapest_startup =
 469                                 get_cheapest_path_for_pathkeys(childrel->pathlist,
 470                                                                                            pathkeys,
 471                                                                                            NULL,
 472                                                                                            STARTUP_COST,
 473                                                                                            false);
 474                         cheapest_total =
 475                                 get_cheapest_path_for_pathkeys(childrel->pathlist,
 476                                                                                            pathkeys,
 477                                                                                            NULL,
 478                                                                                            TOTAL_COST,
 479                                                                                            false);
 480
 481                         /*
 482                          * If we can't find any paths with the right order just use the
 483                          * cheapest-total path; we'll have to sort it later.
 484                          */
 485                         if (cheapest_startup == NULL || cheapest_total == NULL)
 486                         {
 487                                 cheapest_startup = cheapest_total =
 488                                         childrel->cheapest_total_path;
 489                                 /* Assert we do have an unparameterized path for this child */
 490                                 Assert(cheapest_total->param_info == NULL);
 491                         }
 492
 493                         /*
 494                          * Notice whether we actually have different paths for the
 495                          * "cheapest" and "total" cases; frequently there will be no point
 496                          * in two create_merge_append_path() calls.
 497                          */
 498                         if (cheapest_startup != cheapest_total)
 499                                 startup_neq_total = true;
 500
 501                         startup_subpaths =
 502                                 accumulate_append_subpath(startup_subpaths, cheapest_startup);
 503                         total_subpaths =
 504                                 accumulate_append_subpath(total_subpaths, cheapest_total);
 505                 }
 506
 507                 /* ... and build the MergeAppend paths */
 508                 add_path(rel, (Path *) create_merge_append_path(root,
 509                                                                                                                 rel,
 510                                                                                                                 startup_subpaths,
 511                                                                                                                 pathkeys,
 512                                                                                                                 NULL,
 513                                                                                                                 partitioned_rels));
 514                 if (startup_neq_total)
 515                         add_path(rel, (Path *) create_merge_append_path(root,
 516                                                                                                                         rel,
 517                                                                                                                         total_subpaths,
 518                                                                                                                         pathkeys,
 519                                                                                                                         NULL,
 520                                                                                                                         partitioned_rels));
 521         }
 522 }
 523
 524
 525 /*
 526  * get_cheapest_parameterized_child_path
 527  *              Get cheapest path for this relation that has exactly the requested
 528  *              parameterization.
 529  *
 530  * Returns NULL if unable to create such a path.
 531  */
 532 static Path *
 533 get_cheapest_parameterized_child_path(PlannerInfo *root, RelOptInfo *rel,
 534                                                                           Relids required_outer)
 535 {
 536         Path       *cheapest;
 537         ListCell   *lc;
 538
 539         /*
 540          * Look up the cheapest existing path with no more than the needed
 541          * parameterization.  If it has exactly the needed parameterization, we're
 542          * done.
 543          */
 544         cheapest = get_cheapest_path_for_pathkeys(rel->pathlist,
 545                                                                                           NIL,
 546                                                                                           required_outer,
 547                                                                                           TOTAL_COST,
 548                                                                                           false);
 549         Assert(cheapest != NULL);
 550         if (bms_equal(PATH_REQ_OUTER(cheapest), required_outer))
 551                 return cheapest;
 552
 553         /*
 554          * Otherwise, we can "reparameterize" an existing path to match the given
 555          * parameterization, which effectively means pushing down additional
 556          * joinquals to be checked within the path's scan.  However, some existing
 557          * paths might check the available joinquals already while others don't;
 558          * therefore, it's not clear which existing path will be cheapest after
 559          * reparameterization.  We have to go through them all and find out.
 560          */
 561         cheapest = NULL;
 562         foreach(lc, rel->pathlist)
 563         {
 564                 Path       *path = (Path *) lfirst(lc);
 565
 566                 /* Can't use it if it needs more than requested parameterization */
 567                 if (!bms_is_subset(PATH_REQ_OUTER(path), required_outer))
 568                         continue;
 569
 570                 /*
 571                  * Reparameterization can only increase the path's cost, so if it's
 572                  * already more expensive than the current cheapest, forget it.
 573                  */
 574                 if (cheapest != NULL &&
 575                         compare_path_costs(cheapest, path, TOTAL_COST) <= 0)
 576                         continue;
 577
 578                 /* Reparameterize if needed, then recheck cost */
 579                 if (!bms_equal(PATH_REQ_OUTER(path), required_outer))
 580                 {
 581                         path = reparameterize_path(root, path, required_outer, 1.0);
 582                         if (path == NULL)
 583                                 continue;               /* failed to reparameterize this one */
 584                         Assert(bms_equal(PATH_REQ_OUTER(path), required_outer));
 585
 586                         if (cheapest != NULL &&
 587                                 compare_path_costs(cheapest, path, TOTAL_COST) <= 0)
 588                                 continue;
 589                 }
 590
 591                 /* We have a new best path */
 592                 cheapest = path;
 593         }
 594
 595         /* Return the best path, or NULL if we found no suitable candidate */
 596         return cheapest;
 597 }
 598
 599
 600 /*
 601  * accumulate_append_subpath
 602  *              Add a subpath to the list being built for an Append or MergeAppend
 603  *
 604  * It's possible that the child is itself an Append or MergeAppend path, in
 605  * which case we can "cut out the middleman" and just add its child paths to
 606  * our own list.  (We don't try to do this earlier because we need to apply
 607  * both levels of transformation to the quals.)
 608  *
 609  * Note that if we omit a child MergeAppend in this way, we are effectively
 610  * omitting a sort step, which seems fine: if the parent is to be an Append,
 611  * its result would be unsorted anyway, while if the parent is to be a
 612  * MergeAppend, there's no point in a separate sort on a child.
 613  */
 614 static List *
 615 accumulate_append_subpath(List *subpaths, Path *path)
 616 {
 617         if (IsA(path, AppendPath))
 618         {
 619                 AppendPath *apath = (AppendPath *) path;
 620
 621                 /* list_copy is important here to avoid sharing list substructure */
 622                 return list_concat(subpaths, list_copy(apath->subpaths));
 623         }
 624         else if (IsA(path, MergeAppendPath))
 625         {
 626                 MergeAppendPath *mpath = (MergeAppendPath *) path;
 627
 628                 /* list_copy is important here to avoid sharing list substructure */
 629                 return list_concat(subpaths, list_copy(mpath->subpaths));
 630         }
 631         else
 632                 return lappend(subpaths, path);
 633 }
 634
 635
 636 /*
 637  * standard_join_search
 638  *        Find possible joinpaths for a query by successively finding ways
 639  *        to join component relations into join relations.
 640  *
 641  * 'levels_needed' is the number of iterations needed, ie, the number of
 642  *              independent jointree items in the query.  This is > 1.
 643  *
 644  * 'initial_rels' is a list of RelOptInfo nodes for each independent
 645  *              jointree item.  These are the components to be joined together.
 646  *              Note that levels_needed == list_length(initial_rels).
 647  *
 648  * Returns the final level of join relations, i.e., the relation that is
 649  * the result of joining all the original relations together.
 650  * At least one implementation path must be provided for this relation and
 651  * all required sub-relations.
 652  *
 653  * To support loadable plugins that modify planner behavior by changing the
 654  * join searching algorithm, we provide a hook variable that lets a plugin
 655  * replace or supplement this function.  Any such hook must return the same
 656  * final join relation as the standard code would, but it might have a
 657  * different set of implementation paths attached, and only the sub-joinrels
 658  * needed for these paths need have been instantiated.
 659  *
 660  * Note to plugin authors: the functions invoked during standard_join_search()
 661  * modify root->join_rel_list and root->join_rel_hash.  If you want to do more
 662  * than one join-order search, you'll probably need to save and restore the
 663  * original states of those data structures.  See geqo_eval() for an example.
 664  */
 665 RelOptInfo *
 666 standard_join_search(PlannerInfo *root, int levels_needed, List *initial_rels)
 667 {
 668         int                     lev;
 669         RelOptInfo *rel;
 670
 671         /*
 672          * This function cannot be invoked recursively within any one planning
 673          * problem, so join_rel_level[] can't be in use already.
 674          */
 675         Assert(root->join_rel_level == NULL);
 676
 677         /*
 678          * We employ a simple "dynamic programming" algorithm: we first find all
 679          * ways to build joins of two jointree items, then all ways to build joins
 680          * of three items (from two-item joins and single items), then four-item
 681          * joins, and so on until we have considered all ways to join all the
 682          * items into one rel.
 683          *
 684          * root->join_rel_level[j] is a list of all the j-item rels.  Initially we
 685          * set root->join_rel_level[1] to represent all the single-jointree-item
 686          * relations.
 687          */
 688         root->join_rel_level = (List **) palloc0((levels_needed + 1) * sizeof(List *));
 689
 690         root->join_rel_level[1] = initial_rels;
 691
 692         for (lev = 2; lev <= levels_needed; lev++)
 693         {
 694                 ListCell   *lc;
 695
 696                 /*
 697                  * Determine all possible pairs of relations to be joined at this
 698                  * level, and build paths for making each one from every available
 699                  * pair of lower-level relations.
 700                  */
 701                 join_search_one_level(root, lev);
 702
 703                 /*
 704                  * Run generate_gather_paths() for each just-processed joinrel.  We
 705                  * could not do this earlier because both regular and partial paths
 706                  * can get added to a particular joinrel at multiple times within
 707                  * join_search_one_level.  After that, we're done creating paths for
 708                  * the joinrel, so run set_cheapest().
 709                  */
 710                 foreach(lc, root->join_rel_level[lev])
 711                 {
 712                         rel = (RelOptInfo *) lfirst(lc);
 713
 714                         /* Create GatherPaths for any useful partial paths for rel */
 715                         generate_gather_paths(root, rel);
 716
 717                         /* Find and save the cheapest paths for this rel */
 718                         set_cheapest(rel);
 719
 720 #ifdef OPTIMIZER_DEBUG
 721                         debug_print_rel(root, rel);
 722 #endif
 723                 }
 724         }
 725
 726         /*
 727          * We should have a single rel at the final level.
 728          */
 729         if (root->join_rel_level[levels_needed] == NIL)
 730                 elog(ERROR, "failed to build any %d-way joins", levels_needed);
 731         Assert(list_length(root->join_rel_level[levels_needed]) == 1);
 732
 733         rel = (RelOptInfo *) linitial(root->join_rel_level[levels_needed]);
 734
 735         root->join_rel_level = NULL;
 736
 737         return rel;
 738 }
 739
 740
 741 /*
 742  * join_search_one_level
 743  *        Consider ways to produce join relations containing exactly 'level'
 744  *        jointree items.  (This is one step of the dynamic-programming method
 745  *        embodied in standard_join_search.)  Join rel nodes for each feasible
 746  *        combination of lower-level rels are created and returned in a list.
 747  *        Implementation paths are created for each such joinrel, too.
 748  *
 749  * level: level of rels we want to make this time
 750  * root->join_rel_level[j], 1 <= j < level, is a list of rels containing j items
 751  *
 752  * The result is returned in root->join_rel_level[level].
 753  */
 754 void
 755 join_search_one_level(PlannerInfo *root, int level)
 756 {
 757         List      **joinrels = root->join_rel_level;
 758         ListCell   *r;
 759         int                     k;
 760
 761         Assert(joinrels[level] == NIL);
 762
 763         /* Set join_cur_level so that new joinrels are added to proper list */
 764         root->join_cur_level = level;
 765
 766         /*
 767          * First, consider left-sided and right-sided plans, in which rels of
 768          * exactly level-1 member relations are joined against initial relations.
 769          * We prefer to join using join clauses, but if we find a rel of level-1
 770          * members that has no join clauses, we will generate Cartesian-product
 771          * joins against all initial rels not already contained in it.
 772          */
 773         foreach(r, joinrels[level - 1])
 774         {
 775                 RelOptInfo *old_rel = (RelOptInfo *) lfirst(r);
 776
 777                 if (old_rel->joininfo != NIL || old_rel->has_eclass_joins ||
 778                         has_join_restriction(root, old_rel))
 779                 {
 780                         /*
 781                          * There are join clauses or join order restrictions relevant to
 782                          * this rel, so consider joins between this rel and (only) those
 783                          * initial rels it is linked to by a clause or restriction.
 784                          *
 785                          * At level 2 this condition is symmetric, so there is no need to
 786                          * look at initial rels before this one in the list; we already
 787                          * considered such joins when we were at the earlier rel.  (The
 788                          * mirror-image joins are handled automatically by make_join_rel.)
 789                          * In later passes (level > 2), we join rels of the previous level
 790                          * to each initial rel they don't already include but have a join
 791                          * clause or restriction with.
 792                          */
 793                         ListCell   *other_rels;
 794
 795                         if (level == 2)         /* consider remaining initial rels */
 796                                 other_rels = lnext(r);
 797                         else                            /* consider all initial rels */
 798                                 other_rels = list_head(joinrels[1]);
 799
 800                         make_rels_by_clause_joins(root,
 801                                                                           old_rel,
 802                                                                           other_rels);
 803                 }
 804                 else
 805                 {
 806                         /*
 807                          * Oops, we have a relation that is not joined to any other
 808                          * relation, either directly or by join-order restrictions.
 809                          * Cartesian product time.
 810                          *
 811                          * We consider a cartesian product with each not-already-included
 812                          * initial rel, whether it has other join clauses or not.  At
 813                          * level 2, if there are two or more clauseless initial rels, we
 814                          * will redundantly consider joining them in both directions; but
 815                          * such cases aren't common enough to justify adding complexity to
 816                          * avoid the duplicated effort.
 817                          */
 818                         make_rels_by_clauseless_joins(root,
 819                                                                                   old_rel,
 820                                                                                   list_head(joinrels[1]));
 821                 }
 822         }
 823
 824         /*
 825          * Now, consider "bushy plans" in which relations of k initial rels are
 826          * joined to relations of level-k initial rels, for 2 <= k <= level-2.
 827          *
 828          * We only consider bushy-plan joins for pairs of rels where there is a
 829          * suitable join clause (or join order restriction), in order to avoid
 830          * unreasonable growth of planning time.
 831          */
 832         for (k = 2;; k++)
 833         {
 834                 int                     other_level = level - k;
 835
 836                 /*
 837                  * Since make_join_rel(x, y) handles both x,y and y,x cases, we only
 838                  * need to go as far as the halfway point.
 839                  */
 840                 if (k > other_level)
 841                         break;
 842
 843                 foreach(r, joinrels[k])
 844                 {
 845                         RelOptInfo *old_rel = (RelOptInfo *) lfirst(r);
 846                         ListCell   *other_rels;
 847                         ListCell   *r2;
 848
 849                         /*
 850                          * We can ignore relations without join clauses here, unless they
 851                          * participate in join-order restrictions --- then we might have
 852                          * to force a bushy join plan.
 853                          */
 854                         if (old_rel->joininfo == NIL && !old_rel->has_eclass_joins &&
 855                                 !has_join_restriction(root, old_rel))
 856                                 continue;
 857
 858                         if (k == other_level)
 859                                 other_rels = lnext(r);  /* only consider remaining rels */
 860                         else
 861                                 other_rels = list_head(joinrels[other_level]);
 862
 863                         for_each_cell(r2, other_rels)
 864                         {
 865                                 RelOptInfo *new_rel = (RelOptInfo *) lfirst(r2);
 866
 867                                 if (!bms_overlap(old_rel->relids, new_rel->relids))
 868                                 {
 869                                         /*
 870                                          * OK, we can build a rel of the right level from this
 871                                          * pair of rels.  Do so if there is at least one relevant
 872                                          * join clause or join order restriction.
 873                                          */
 874                                         if (have_relevant_joinclause(root, old_rel, new_rel) ||
 875                                                 have_join_order_restriction(root, old_rel, new_rel))
 876                                         {
 877                                                 (void) make_join_rel(root, old_rel, new_rel);
 878                                         }
 879                                 }
 880                         }
 881                 }
 882         }
 883
 884         /*----------
 885          * Last-ditch effort: if we failed to find any usable joins so far, force
 886          * a set of cartesian-product joins to be generated.  This handles the
 887          * special case where all the available rels have join clauses but we
 888          * cannot use any of those clauses yet.  This can only happen when we are
 889          * considering a join sub-problem (a sub-joinlist) and all the rels in the
 890          * sub-problem have only join clauses with rels outside the sub-problem.
 891          * An example is
 892          *
 893          *              SELECT ... FROM a INNER JOIN b ON TRUE, c, d, ...
 894          *              WHERE a.w = c.x and b.y = d.z;
 895          *
 896          * If the "a INNER JOIN b" sub-problem does not get flattened into the
 897          * upper level, we must be willing to make a cartesian join of a and b;
 898          * but the code above will not have done so, because it thought that both
 899          * a and b have joinclauses.  We consider only left-sided and right-sided
 900          * cartesian joins in this case (no bushy).
 901          *----------
 902          */
 903         if (joinrels[level] == NIL)
 904         {
 905                 /*
 906                  * This loop is just like the first one, except we always call
 907                  * make_rels_by_clauseless_joins().
 908                  */
 909                 foreach(r, joinrels[level - 1])
 910                 {
 911                         RelOptInfo *old_rel = (RelOptInfo *) lfirst(r);
 912
 913                         make_rels_by_clauseless_joins(root,
 914                                                                                   old_rel,
 915                                                                                   list_head(joinrels[1]));
 916                 }
 917
 918                 /*----------
 919                  * When special joins are involved, there may be no legal way
 920                  * to make an N-way join for some values of N.  For example consider
 921                  *
 922                  * SELECT ... FROM t1 WHERE
 923                  *       x IN (SELECT ... FROM t2,t3 WHERE ...) AND
 924                  *       y IN (SELECT ... FROM t4,t5 WHERE ...)
 925                  *
 926                  * We will flatten this query to a 5-way join problem, but there are
 927                  * no 4-way joins that join_is_legal() will consider legal.  We have
 928                  * to accept failure at level 4 and go on to discover a workable
 929                  * bushy plan at level 5.
 930                  *
 931                  * However, if there are no special joins and no lateral references
 932                  * then join_is_legal() should never fail, and so the following sanity
 933                  * check is useful.
 934                  *----------
 935                  */
 936                 if (joinrels[level] == NIL &&
 937                         root->join_info_list == NIL &&
 938                         !root->hasLateralRTEs)
 939                         elog(ERROR, "failed to build any %d-way joins", level);
 940         }
 941 }
 942
 943
 944 /*
 945  * make_rels_by_clause_joins
 946  *        Build joins between the given relation 'old_rel' and other relations
 947  *        that participate in join clauses that 'old_rel' also participates in
 948  *        (or participate in join-order restrictions with it).
 949  *        The join rels are returned in root->join_rel_level[join_cur_level].
 950  *
 951  * Note: at levels above 2 we will generate the same joined relation in
 952  * multiple ways --- for example (a join b) join c is the same RelOptInfo as
 953  * (b join c) join a, though the second case will add a different set of Paths
 954  * to it.  This is the reason for using the join_rel_level mechanism, which
 955  * automatically ensures that each new joinrel is only added to the list once.
 956  *
 957  * 'old_rel' is the relation entry for the relation to be joined
 958  * 'other_rels': the first cell in a linked list containing the other
 959  * rels to be considered for joining
 960  *
 961  * Currently, this is only used with initial rels in other_rels, but it
 962  * will work for joining to joinrels too.
 963  */
 964 static void
 965 make_rels_by_clause_joins(PlannerInfo *root,
 966                                                   RelOptInfo *old_rel,
 967                                                   ListCell *other_rels)
 968 {
 969         ListCell   *l;
 970
 971         for_each_cell(l, other_rels)
 972         {
 973                 RelOptInfo *other_rel = (RelOptInfo *) lfirst(l);
 974
 975                 if (!bms_overlap(old_rel->relids, other_rel->relids) &&
 976                         (have_relevant_joinclause(root, old_rel, other_rel) ||
 977                          have_join_order_restriction(root, old_rel, other_rel)))
 978                 {
 979                         (void) make_join_rel(root, old_rel, other_rel);
 980                 }
 981         }
 982 }
 983
 984
 985 /*
 986  * make_rels_by_clauseless_joins
 987  *        Given a relation 'old_rel' and a list of other relations
 988  *        'other_rels', create a join relation between 'old_rel' and each
 989  *        member of 'other_rels' that isn't already included in 'old_rel'.
 990  *        The join rels are returned in root->join_rel_level[join_cur_level].
 991  *
 992  * 'old_rel' is the relation entry for the relation to be joined
 993  * 'other_rels': the first cell of a linked list containing the
 994  * other rels to be considered for joining
 995  *
 996  * Currently, this is only used with initial rels in other_rels, but it would
 997  * work for joining to joinrels too.
 998  */
 999 static void
1000 make_rels_by_clauseless_joins(PlannerInfo *root,
1001                                                           RelOptInfo *old_rel,
1002                                                           ListCell *other_rels)
1003 {
1004         ListCell   *l;
1005
1006         for_each_cell(l, other_rels)
1007         {
1008                 RelOptInfo *other_rel = (RelOptInfo *) lfirst(l);
1009
1010                 if (!bms_overlap(other_rel->relids, old_rel->relids))
1011                 {
1012                         (void) make_join_rel(root, old_rel, other_rel);
1013                 }
1014         }
1015 }
1016
1017
1018 /*
1019  * join_is_legal
1020  *         Determine whether a proposed join is legal given the query's
1021  *         join order constraints; and if it is, determine the join type.
1022  *
1023  * Caller must supply not only the two rels, but the union of their relids.
1024  * (We could simplify the API by computing joinrelids locally, but this
1025  * would be redundant work in the normal path through make_join_rel.)
1026  *
1027  * On success, *sjinfo_p is set to NULL if this is to be a plain inner join,
1028  * else it's set to point to the associated SpecialJoinInfo node.  Also,
1029  * *reversed_p is set TRUE if the given relations need to be swapped to
1030  * match the SpecialJoinInfo node.
1031  */
1032 static bool
1033 join_is_legal(PlannerInfo *root, RelOptInfo *rel1, RelOptInfo *rel2,
1034                           Relids joinrelids,
1035                           SpecialJoinInfo **sjinfo_p, bool *reversed_p)
1036 {
1037         SpecialJoinInfo *match_sjinfo;
1038         bool            reversed;
1039         bool            unique_ified;
1040         bool            must_be_leftjoin;
1041         ListCell   *l;
1042
1043         /*
1044          * Ensure output params are set on failure return.  This is just to
1045          * suppress uninitialized-variable warnings from overly anal compilers.
1046          */
1047         *sjinfo_p = NULL;
1048         *reversed_p = false;
1049
1050         /*
1051          * If we have any special joins, the proposed join might be illegal; and
1052          * in any case we have to determine its join type.  Scan the join info
1053          * list for matches and conflicts.
1054          */
1055         match_sjinfo = NULL;
1056         reversed = false;
1057         unique_ified = false;
1058         must_be_leftjoin = false;
1059
1060         foreach(l, root->join_info_list)
1061         {
1062                 SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
1063
1064                 /*
1065                  * This special join is not relevant unless its RHS overlaps the
1066                  * proposed join.  (Check this first as a fast path for dismissing
1067                  * most irrelevant SJs quickly.)
1068                  */
1069                 if (!bms_overlap(sjinfo->min_righthand, joinrelids))
1070                         continue;
1071
1072                 /*
1073                  * Also, not relevant if proposed join is fully contained within RHS
1074                  * (ie, we're still building up the RHS).
1075                  */
1076                 if (bms_is_subset(joinrelids, sjinfo->min_righthand))
1077                         continue;
1078
1079                 /*
1080                  * Also, not relevant if SJ is already done within either input.
1081                  */
1082                 if (bms_is_subset(sjinfo->min_lefthand, rel1->relids) &&
1083                         bms_is_subset(sjinfo->min_righthand, rel1->relids))
1084                         continue;
1085                 if (bms_is_subset(sjinfo->min_lefthand, rel2->relids) &&
1086                         bms_is_subset(sjinfo->min_righthand, rel2->relids))
1087                         continue;
1088
1089                 /*
1090                  * If it's a semijoin and we already joined the RHS to any other rels
1091                  * within either input, then we must have unique-ified the RHS at that
1092                  * point (see below).  Therefore the semijoin is no longer relevant in
1093                  * this join path.
1094                  */
1095                 if (sjinfo->jointype == JOIN_SEMI)
1096                 {
1097                         if (bms_is_subset(sjinfo->syn_righthand, rel1->relids) &&
1098                                 !bms_equal(sjinfo->syn_righthand, rel1->relids))
1099                                 continue;
1100                         if (bms_is_subset(sjinfo->syn_righthand, rel2->relids) &&
1101                                 !bms_equal(sjinfo->syn_righthand, rel2->relids))
1102                                 continue;
1103                 }
1104
1105                 /*
1106                  * If one input contains min_lefthand and the other contains
1107                  * min_righthand, then we can perform the SJ at this join.
1108                  *
1109                  * Reject if we get matches to more than one SJ; that implies we're
1110                  * considering something that's not really valid.
1111                  */
1112                 if (bms_is_subset(sjinfo->min_lefthand, rel1->relids) &&
1113                         bms_is_subset(sjinfo->min_righthand, rel2->relids))
1114                 {
1115                         if (match_sjinfo)
1116                                 return false;   /* invalid join path */
1117                         match_sjinfo = sjinfo;
1118                         reversed = false;
1119                 }
1120                 else if (bms_is_subset(sjinfo->min_lefthand, rel2->relids) &&
1121                                  bms_is_subset(sjinfo->min_righthand, rel1->relids))
1122                 {
1123                         if (match_sjinfo)
1124                                 return false;   /* invalid join path */
1125                         match_sjinfo = sjinfo;
1126                         reversed = true;
1127                 }
1128                 else if (sjinfo->jointype == JOIN_SEMI &&
1129                                  bms_equal(sjinfo->syn_righthand, rel2->relids) &&
1130                                  create_unique_path(root, rel2, rel2->cheapest_total_path,
1131                                                                         sjinfo) != NULL)
1132                 {
1133                         /*----------
1134                          * For a semijoin, we can join the RHS to anything else by
1135                          * unique-ifying the RHS (if the RHS can be unique-ified).
1136                          * We will only get here if we have the full RHS but less
1137                          * than min_lefthand on the LHS.
1138                          *
1139                          * The reason to consider such a join path is exemplified by
1140                          *      SELECT ... FROM a,b WHERE (a.x,b.y) IN (SELECT c1,c2 FROM c)
1141                          * If we insist on doing this as a semijoin we will first have
1142                          * to form the cartesian product of A*B.  But if we unique-ify
1143                          * C then the semijoin becomes a plain innerjoin and we can join
1144                          * in any order, eg C to A and then to B.  When C is much smaller
1145                          * than A and B this can be a huge win.  So we allow C to be
1146                          * joined to just A or just B here, and then make_join_rel has
1147                          * to handle the case properly.
1148                          *
1149                          * Note that actually we'll allow unique-ified C to be joined to
1150                          * some other relation D here, too.  That is legal, if usually not
1151                          * very sane, and this routine is only concerned with legality not
1152                          * with whether the join is good strategy.
1153                          *----------
1154                          */
1155                         if (match_sjinfo)
1156                                 return false;   /* invalid join path */
1157                         match_sjinfo = sjinfo;
1158                         reversed = false;
1159                         unique_ified = true;
1160                 }
1161                 else if (sjinfo->jointype == JOIN_SEMI &&
1162                                  bms_equal(sjinfo->syn_righthand, rel1->relids) &&
1163                                  create_unique_path(root, rel1, rel1->cheapest_total_path,
1164                                                                         sjinfo) != NULL)
1165                 {
1166                         /* Reversed semijoin case */
1167                         if (match_sjinfo)
1168                                 return false;   /* invalid join path */
1169                         match_sjinfo = sjinfo;
1170                         reversed = true;
1171                         unique_ified = true;
1172                 }
1173                 else
1174                 {
1175                         /*
1176                          * Otherwise, the proposed join overlaps the RHS but isn't a valid
1177                          * implementation of this SJ.  But don't panic quite yet: the RHS
1178                          * violation might have occurred previously, in one or both input
1179                          * relations, in which case we must have previously decided that
1180                          * it was OK to commute some other SJ with this one.  If we need
1181                          * to perform this join to finish building up the RHS, rejecting
1182                          * it could lead to not finding any plan at all.  (This can occur
1183                          * because of the heuristics elsewhere in this file that postpone
1184                          * clauseless joins: we might not consider doing a clauseless join
1185                          * within the RHS until after we've performed other, validly
1186                          * commutable SJs with one or both sides of the clauseless join.)
1187                          * This consideration boils down to the rule that if both inputs
1188                          * overlap the RHS, we can allow the join --- they are either
1189                          * fully within the RHS, or represent previously-allowed joins to
1190                          * rels outside it.
1191                          */
1192                         if (bms_overlap(rel1->relids, sjinfo->min_righthand) &&
1193                                 bms_overlap(rel2->relids, sjinfo->min_righthand))
1194                                 continue;               /* assume valid previous violation of RHS */
1195
1196                         /*
1197                          * The proposed join could still be legal, but only if we're
1198                          * allowed to associate it into the RHS of this SJ.  That means
1199                          * this SJ must be a LEFT join (not SEMI or ANTI, and certainly
1200                          * not FULL) and the proposed join must not overlap the LHS.
1201                          */
1202                         if (sjinfo->jointype != JOIN_LEFT ||
1203                                 bms_overlap(joinrelids, sjinfo->min_lefthand))
1204                                 return false;   /* invalid join path */
1205
1206                         /*
1207                          * To be valid, the proposed join must be a LEFT join; otherwise
1208                          * it can't associate into this SJ's RHS.  But we may not yet have
1209                          * found the SpecialJoinInfo matching the proposed join, so we
1210                          * can't test that yet.  Remember the requirement for later.
1211                          */
1212                         must_be_leftjoin = true;
1213                 }
1214         }
1215
1216         /*
1217          * Fail if violated any SJ's RHS and didn't match to a LEFT SJ: the
1218          * proposed join can't associate into an SJ's RHS.
1219          *
1220          * Also, fail if the proposed join's predicate isn't strict; we're
1221          * essentially checking to see if we can apply outer-join identity 3, and
1222          * that's a requirement.  (This check may be redundant with checks in
1223          * make_outerjoininfo, but I'm not quite sure, and it's cheap to test.)
1224          */
1225         if (must_be_leftjoin &&
1226                 (match_sjinfo == NULL ||
1227                  match_sjinfo->jointype != JOIN_LEFT ||
1228                  !match_sjinfo->lhs_strict))
1229                 return false;                   /* invalid join path */
1230
1231         /*
1232          * We also have to check for constraints imposed by LATERAL references.
1233          */
1234         if (root->hasLateralRTEs)
1235         {
1236                 bool            lateral_fwd;
1237                 bool            lateral_rev;
1238                 Relids          join_lateral_rels;
1239
1240                 /*
1241                  * The proposed rels could each contain lateral references to the
1242                  * other, in which case the join is impossible.  If there are lateral
1243                  * references in just one direction, then the join has to be done with
1244                  * a nestloop with the lateral referencer on the inside.  If the join
1245                  * matches an SJ that cannot be implemented by such a nestloop, the
1246                  * join is impossible.
1247                  *
1248                  * Also, if the lateral reference is only indirect, we should reject
1249                  * the join; whatever rel(s) the reference chain goes through must be
1250                  * joined to first.
1251                  *
1252                  * Another case that might keep us from building a valid plan is the
1253                  * implementation restriction described by have_dangerous_phv().
1254                  */
1255                 lateral_fwd = bms_overlap(rel1->relids, rel2->lateral_relids);
1256                 lateral_rev = bms_overlap(rel2->relids, rel1->lateral_relids);
1257                 if (lateral_fwd && lateral_rev)
1258                         return false;           /* have lateral refs in both directions */
1259                 if (lateral_fwd)
1260                 {
1261                         /* has to be implemented as nestloop with rel1 on left */
1262                         if (match_sjinfo &&
1263                                 (reversed ||
1264                                  unique_ified ||
1265                                  match_sjinfo->jointype == JOIN_FULL))
1266                                 return false;   /* not implementable as nestloop */
1267                         /* check there is a direct reference from rel2 to rel1 */
1268                         if (!bms_overlap(rel1->relids, rel2->direct_lateral_relids))
1269                                 return false;   /* only indirect refs, so reject */
1270                         /* check we won't have a dangerous PHV */
1271                         if (have_dangerous_phv(root, rel1->relids, rel2->lateral_relids))
1272                                 return false;   /* might be unable to handle required PHV */
1273                 }
1274                 else if (lateral_rev)
1275                 {
1276                         /* has to be implemented as nestloop with rel2 on left */
1277                         if (match_sjinfo &&
1278                                 (!reversed ||
1279                                  unique_ified ||
1280                                  match_sjinfo->jointype == JOIN_FULL))
1281                                 return false;   /* not implementable as nestloop */
1282                         /* check there is a direct reference from rel1 to rel2 */
1283                         if (!bms_overlap(rel2->relids, rel1->direct_lateral_relids))
1284                                 return false;   /* only indirect refs, so reject */
1285                         /* check we won't have a dangerous PHV */
1286                         if (have_dangerous_phv(root, rel2->relids, rel1->lateral_relids))
1287                                 return false;   /* might be unable to handle required PHV */
1288                 }
1289
1290                 /*
1291                  * LATERAL references could also cause problems later on if we accept
1292                  * this join: if the join's minimum parameterization includes any rels
1293                  * that would have to be on the inside of an outer join with this join
1294                  * rel, then it's never going to be possible to build the complete
1295                  * query using this join.  We should reject this join not only because
1296                  * it'll save work, but because if we don't, the clauseless-join
1297                  * heuristics might think that legality of this join means that some
1298                  * other join rel need not be formed, and that could lead to failure
1299                  * to find any plan at all.  We have to consider not only rels that
1300                  * are directly on the inner side of an OJ with the joinrel, but also
1301                  * ones that are indirectly so, so search to find all such rels.
1302                  */
1303                 join_lateral_rels = min_join_parameterization(root, joinrelids,
1304                                                                                                           rel1, rel2);
1305                 if (join_lateral_rels)
1306                 {
1307                         Relids          join_plus_rhs = bms_copy(joinrelids);
1308                         bool            more;
1309
1310                         do
1311                         {
1312                                 more = false;
1313                                 foreach(l, root->join_info_list)
1314                                 {
1315                                         SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
1316
1317                                         if (bms_overlap(sjinfo->min_lefthand, join_plus_rhs) &&
1318                                                 !bms_is_subset(sjinfo->min_righthand, join_plus_rhs))
1319                                         {
1320                                                 join_plus_rhs = bms_add_members(join_plus_rhs,
1321                                                                                                                 sjinfo->min_righthand);
1322                                                 more = true;
1323                                         }
1324                                         /* full joins constrain both sides symmetrically */
1325                                         if (sjinfo->jointype == JOIN_FULL &&
1326                                                 bms_overlap(sjinfo->min_righthand, join_plus_rhs) &&
1327                                                 !bms_is_subset(sjinfo->min_lefthand, join_plus_rhs))
1328                                         {
1329                                                 join_plus_rhs = bms_add_members(join_plus_rhs,
1330                                                                                                                 sjinfo->min_lefthand);
1331                                                 more = true;
1332                                         }
1333                                 }
1334                         } while (more);
1335                         if (bms_overlap(join_plus_rhs, join_lateral_rels))
1336                                 return false;   /* will not be able to join to some RHS rel */
1337                 }
1338         }
1339
1340         /* Otherwise, it's a valid join */
1341         *sjinfo_p = match_sjinfo;
1342         *reversed_p = reversed;
1343         return true;
1344 }
1345
1346
1347 /*
1348  * has_join_restriction
1349  *              Detect whether the specified relation has join-order restrictions,
1350  *              due to being inside an outer join or an IN (sub-SELECT),
1351  *              or participating in any LATERAL references or multi-rel PHVs.
1352  *
1353  * Essentially, this tests whether have_join_order_restriction() could
1354  * succeed with this rel and some other one.  It's OK if we sometimes
1355  * say "true" incorrectly.  (Therefore, we don't bother with the relatively
1356  * expensive has_legal_joinclause test.)
1357  */
1358 static bool
1359 has_join_restriction(PlannerInfo *root, RelOptInfo *rel)
1360 {
1361         ListCell   *l;
1362
1363         if (rel->lateral_relids != NULL || rel->lateral_referencers != NULL)
1364                 return true;
1365
1366         foreach(l, root->placeholder_list)
1367         {
1368                 PlaceHolderInfo *phinfo = (PlaceHolderInfo *) lfirst(l);
1369
1370                 if (bms_is_subset(rel->relids, phinfo->ph_eval_at) &&
1371                         !bms_equal(rel->relids, phinfo->ph_eval_at))
1372                         return true;
1373         }
1374
1375         foreach(l, root->join_info_list)
1376         {
1377                 SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
1378
1379                 /* ignore full joins --- other mechanisms preserve their ordering */
1380                 if (sjinfo->jointype == JOIN_FULL)
1381                         continue;
1382
1383                 /* ignore if SJ is already contained in rel */
1384                 if (bms_is_subset(sjinfo->min_lefthand, rel->relids) &&
1385                         bms_is_subset(sjinfo->min_righthand, rel->relids))
1386                         continue;
1387
1388                 /* restricted if it overlaps LHS or RHS, but doesn't contain SJ */
1389                 if (bms_overlap(sjinfo->min_lefthand, rel->relids) ||
1390                         bms_overlap(sjinfo->min_righthand, rel->relids))
1391                         return true;
1392         }
1393
1394         return false;
1395 }
1396
1397
1398 /*
1399  * is_dummy_rel --- has relation been proven empty?
1400  */
1401 static bool
1402 is_dummy_rel(RelOptInfo *rel)
1403 {
1404         return IS_DUMMY_REL(rel);
1405 }
1406
1407
1408 /*
1409  * Mark a relation as proven empty.
1410  *
1411  * During GEQO planning, this can get invoked more than once on the same
1412  * baserel struct, so it's worth checking to see if the rel is already marked
1413  * dummy.
1414  *
1415  * Also, when called during GEQO join planning, we are in a short-lived
1416  * memory context.  We must make sure that the dummy path attached to a
1417  * baserel survives the GEQO cycle, else the baserel is trashed for future
1418  * GEQO cycles.  On the other hand, when we are marking a joinrel during GEQO,
1419  * we don't want the dummy path to clutter the main planning context.  Upshot
1420  * is that the best solution is to explicitly make the dummy path in the same
1421  * context the given RelOptInfo is in.
1422  */
1423 static void
1424 mark_dummy_rel(RelOptInfo *rel)
1425 {
1426         MemoryContext oldcontext;
1427
1428         /* Already marked? */
1429         if (is_dummy_rel(rel))
1430                 return;
1431
1432         /* No, so choose correct context to make the dummy path in */
1433         oldcontext = MemoryContextSwitchTo(GetMemoryChunkContext(rel));
1434
1435         /* Set dummy size estimate */
1436         rel->rows = 0;
1437
1438         /* Evict any previously chosen paths */
1439         rel->pathlist = NIL;
1440         rel->partial_pathlist = NIL;
1441
1442         /* Set up the dummy path */
1443         add_path(rel, (Path *) create_append_path(rel, NIL, NULL, 0, NIL));
1444
1445         /* Set or update cheapest_total_path and related fields */
1446         set_cheapest(rel);
1447
1448         MemoryContextSwitchTo(oldcontext);
1449 }
1450
1451
1452 /*
1453  * restriction_is_constant_false --- is a restrictlist just false?
1454  *
1455  * In cases where a qual is provably constant false, eval_const_expressions
1456  * will generally have thrown away anything that's ANDed with it.  In outer
1457  * join situations this will leave us computing cartesian products only to
1458  * decide there's no match for an outer row, which is pretty stupid.  So,
1459  * we need to detect the case.
1460  *
1461  * If only_pushed_down is true, then consider only quals that are pushed-down
1462  * from the point of view of the joinrel.
1463  */
1464 static bool
1465 restriction_is_constant_false(List *restrictlist,
1466                                                           RelOptInfo *joinrel,
1467                                                           bool only_pushed_down)
1468 {
1469         ListCell   *lc;
1470
1471         /*
1472          * Despite the above comment, the restriction list we see here might
1473          * possibly have other members besides the FALSE constant, since other
1474          * quals could get "pushed down" to the outer join level.  So we check
1475          * each member of the list.
1476          */
1477         foreach(lc, restrictlist)
1478         {
1479                 RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc);
1480
1481                 if (only_pushed_down && !RINFO_IS_PUSHED_DOWN(rinfo, joinrel->relids))
1482                         continue;
1483
1484                 if (rinfo->clause && IsA(rinfo->clause, Const))
1485                 {
1486                         Const      *con = (Const *) rinfo->clause;
1487
1488                         /* constant NULL is as good as constant FALSE for our purposes */
1489                         if (con->constisnull)
1490                                 return true;
1491                         if (!DatumGetBool(con->constvalue))
1492                                 return true;
1493                 }
1494         }
1495         return false;
1496 }