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  *     set_append_rel_pathlist()
  19  *     add_paths_to_append_rel()
  20  *     generate_mergeappend_paths()
  21  *     get_cheapest_parameterized_child_path()
  22  *     accumulate_append_subpath()
  23  *
  24  *  public functions:
  25  *     standard_join_search(): This funcion is not static. The reason for
  26  *        including this function is make_rels_by_clause_joins. In order to
  27  *        avoid generating apparently unwanted join combination, we decided to
  28  *        change the behavior of make_join_rel, which is called under this
  29  *        function.
  30  *
  31  * src/backend/optimizer/path/joinrels.c
  32  *
  33  *      public functions:
  34  *     join_search_one_level(): We have to modify this to call my definition of
  35  *                  make_rels_by_clause_joins.
  36  *
  37  *      static functions:
  38  *     make_rels_by_clause_joins()
  39  *     make_rels_by_clauseless_joins()
  40  *     join_is_legal()
  41  *     has_join_restriction()
  42  *     is_dummy_rel()
  43  *     mark_dummy_rel()
  44  *     restriction_is_constant_false()
  45  *
  46  *
  47  * Portions Copyright (c) 1996-2017, 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  * set_append_rel_pathlist
  87  *        Build access paths for an "append relation"
  88  */
  89 static void
  90 set_append_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
  91                                                 Index rti, RangeTblEntry *rte)
  92 {
  93         int                     parentRTindex = rti;
  94         List       *live_childrels = NIL;
  95         ListCell   *l;
  96
  97         /*
  98          * Generate access paths for each member relation, and remember the
  99          * non-dummy children.
 100          */
 101         foreach(l, root->append_rel_list)
 102         {
 103                 AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
 104                 int                     childRTindex;
 105                 RangeTblEntry *childRTE;
 106                 RelOptInfo *childrel;
 107
 108                 /* append_rel_list contains all append rels; ignore others */
 109                 if (appinfo->parent_relid != parentRTindex)
 110                         continue;
 111
 112                 /* Re-locate the child RTE and RelOptInfo */
 113                 childRTindex = appinfo->child_relid;
 114                 childRTE = root->simple_rte_array[childRTindex];
 115                 childrel = root->simple_rel_array[childRTindex];
 116
 117                 /*
 118                  * If set_append_rel_size() decided the parent appendrel was
 119                  * parallel-unsafe at some point after visiting this child rel, we
 120                  * need to propagate the unsafety marking down to the child, so that
 121                  * we don't generate useless partial paths for it.
 122                  */
 123                 if (!rel->consider_parallel)
 124                         childrel->consider_parallel = false;
 125
 126                 /*
 127                  * Compute the child's access paths.
 128                  */
 129                 set_rel_pathlist(root, childrel, childRTindex, childRTE);
 130
 131                 /*
 132                  * If child is dummy, ignore it.
 133                  */
 134                 if (IS_DUMMY_REL(childrel))
 135                         continue;
 136
 137                 /*
 138                  * Child is live, so add it to the live_childrels list for use below.
 139                  */
 140                 live_childrels = lappend(live_childrels, childrel);
 141         }
 142
 143         /* Add paths to the "append" relation. */
 144         add_paths_to_append_rel(root, rel, live_childrels);
 145 }
 146
 147 /*
 148  * add_paths_to_append_rel
 149  *              Generate paths for given "append" relation given the set of non-dummy
 150  *              child rels.
 151  *
 152  * The function collects all parameterizations and orderings supported by the
 153  * non-dummy children. For every such parameterization or ordering, it creates
 154  * an append path collecting one path from each non-dummy child with given
 155  * parameterization or ordering. Similarly it collects partial paths from
 156  * non-dummy children to create partial append paths.
 157  */
 158 static void
 159 add_paths_to_append_rel(PlannerInfo *root, RelOptInfo *rel,
 160                                                 List *live_childrels)
 161 {
 162         List       *subpaths = NIL;
 163         bool            subpaths_valid = true;
 164         List       *partial_subpaths = NIL;
 165         bool            partial_subpaths_valid = true;
 166         List       *all_child_pathkeys = NIL;
 167         List       *all_child_outers = NIL;
 168         ListCell   *l;
 169         List       *partitioned_rels = NIL;
 170         RangeTblEntry *rte;
 171         bool            build_partitioned_rels = false;
 172
 173         /*
 174          * A plain relation will already have a PartitionedChildRelInfo if it is
 175          * partitioned.  For a subquery RTE, no PartitionedChildRelInfo exists; we
 176          * collect all partitioned_rels associated with any child.  (This assumes
 177          * that we don't need to look through multiple levels of subquery RTEs; if
 178          * we ever do, we could create a PartitionedChildRelInfo with the
 179          * accumulated list of partitioned_rels which would then be found when
 180          * populated our parent rel with paths.  For the present, that appears to
 181          * be unnecessary.)
 182          */
 183         rte = planner_rt_fetch(rel->relid, root);
 184         switch (rte->rtekind)
 185         {
 186                 case RTE_RELATION:
 187                         if (rte->relkind == RELKIND_PARTITIONED_TABLE)
 188                         {
 189                                 partitioned_rels =
 190                                         get_partitioned_child_rels(root, rel->relid);
 191                                 Assert(list_length(partitioned_rels) >= 1);
 192                         }
 193                         break;
 194                 case RTE_SUBQUERY:
 195                         build_partitioned_rels = true;
 196                         break;
 197                 default:
 198                         elog(ERROR, "unexpcted rtekind: %d", (int) rte->rtekind);
 199         }
 200
 201         /*
 202          * For every non-dummy child, remember the cheapest path.  Also, identify
 203          * all pathkeys (orderings) and parameterizations (required_outer sets)
 204          * available for the non-dummy member relations.
 205          */
 206         foreach(l, live_childrels)
 207         {
 208                 RelOptInfo *childrel = lfirst(l);
 209                 ListCell   *lcp;
 210
 211                 /*
 212                  * If we need to build partitioned_rels, accumulate the partitioned
 213                  * rels for this child.
 214                  */
 215                 if (build_partitioned_rels)
 216                 {
 217                         List       *cprels;
 218
 219                         cprels = get_partitioned_child_rels(root, childrel->relid);
 220                         partitioned_rels = list_concat(partitioned_rels,
 221                                                                                    list_copy(cprels));
 222                 }
 223
 224                 /*
 225                  * If child has an unparameterized cheapest-total path, add that to
 226                  * the unparameterized Append path we are constructing for the parent.
 227                  * If not, there's no workable unparameterized path.
 228                  */
 229                 if (childrel->cheapest_total_path->param_info == NULL)
 230                         subpaths = accumulate_append_subpath(subpaths,
 231                                                                                                  childrel->cheapest_total_path);
 232                 else
 233                         subpaths_valid = false;
 234
 235                 /* Same idea, but for a partial plan. */
 236                 if (childrel->partial_pathlist != NIL)
 237                         partial_subpaths = accumulate_append_subpath(partial_subpaths,
 238                                                                                                                  linitial(childrel->partial_pathlist));
 239                 else
 240                         partial_subpaths_valid = false;
 241
 242                 /*
 243                  * Collect lists of all the available path orderings and
 244                  * parameterizations for all the children.  We use these as a
 245                  * heuristic to indicate which sort orderings and parameterizations we
 246                  * should build Append and MergeAppend paths for.
 247                  */
 248                 foreach(lcp, childrel->pathlist)
 249                 {
 250                         Path       *childpath = (Path *) lfirst(lcp);
 251                         List       *childkeys = childpath->pathkeys;
 252                         Relids          childouter = PATH_REQ_OUTER(childpath);
 253
 254                         /* Unsorted paths don't contribute to pathkey list */
 255                         if (childkeys != NIL)
 256                         {
 257                                 ListCell   *lpk;
 258                                 bool            found = false;
 259
 260                                 /* Have we already seen this ordering? */
 261                                 foreach(lpk, all_child_pathkeys)
 262                                 {
 263                                         List       *existing_pathkeys = (List *) lfirst(lpk);
 264
 265                                         if (compare_pathkeys(existing_pathkeys,
 266                                                                                  childkeys) == PATHKEYS_EQUAL)
 267                                         {
 268                                                 found = true;
 269                                                 break;
 270                                         }
 271                                 }
 272                                 if (!found)
 273                                 {
 274                                         /* No, so add it to all_child_pathkeys */
 275                                         all_child_pathkeys = lappend(all_child_pathkeys,
 276                                                                                                  childkeys);
 277                                 }
 278                         }
 279
 280                         /* Unparameterized paths don't contribute to param-set list */
 281                         if (childouter)
 282                         {
 283                                 ListCell   *lco;
 284                                 bool            found = false;
 285
 286                                 /* Have we already seen this param set? */
 287                                 foreach(lco, all_child_outers)
 288                                 {
 289                                         Relids          existing_outers = (Relids) lfirst(lco);
 290
 291                                         if (bms_equal(existing_outers, childouter))
 292                                         {
 293                                                 found = true;
 294                                                 break;
 295                                         }
 296                                 }
 297                                 if (!found)
 298                                 {
 299                                         /* No, so add it to all_child_outers */
 300                                         all_child_outers = lappend(all_child_outers,
 301                                                                                            childouter);
 302                                 }
 303                         }
 304                 }
 305         }
 306
 307         /*
 308          * If we found unparameterized paths for all children, build an unordered,
 309          * unparameterized Append path for the rel.  (Note: this is correct even
 310          * if we have zero or one live subpath due to constraint exclusion.)
 311          */
 312         if (subpaths_valid)
 313                 add_path(rel, (Path *) create_append_path(rel, subpaths, NULL, 0,
 314                                                                                                   partitioned_rels));
 315
 316         /*
 317          * Consider an append of partial unordered, unparameterized partial paths.
 318          */
 319         if (partial_subpaths_valid)
 320         {
 321                 AppendPath *appendpath;
 322                 ListCell   *lc;
 323                 int                     parallel_workers = 0;
 324
 325                 /*
 326                  * Decide on the number of workers to request for this append path.
 327                  * For now, we just use the maximum value from among the members.  It
 328                  * might be useful to use a higher number if the Append node were
 329                  * smart enough to spread out the workers, but it currently isn't.
 330                  */
 331                 foreach(lc, partial_subpaths)
 332                 {
 333                         Path       *path = lfirst(lc);
 334
 335                         parallel_workers = Max(parallel_workers, path->parallel_workers);
 336                 }
 337                 Assert(parallel_workers > 0);
 338
 339                 /* Generate a partial append path. */
 340                 appendpath = create_append_path(rel, partial_subpaths, NULL,
 341                                                                                 parallel_workers, partitioned_rels);
 342                 add_partial_path(rel, (Path *) appendpath);
 343         }
 344
 345         /*
 346          * Also build unparameterized MergeAppend paths based on the collected
 347          * list of child pathkeys.
 348          */
 349         if (subpaths_valid)
 350                 generate_mergeappend_paths(root, rel, live_childrels,
 351                                                                    all_child_pathkeys,
 352                                                                    partitioned_rels);
 353
 354         /*
 355          * Build Append paths for each parameterization seen among the child rels.
 356          * (This may look pretty expensive, but in most cases of practical
 357          * interest, the child rels will expose mostly the same parameterizations,
 358          * so that not that many cases actually get considered here.)
 359          *
 360          * The Append node itself cannot enforce quals, so all qual checking must
 361          * be done in the child paths.  This means that to have a parameterized
 362          * Append path, we must have the exact same parameterization for each
 363          * child path; otherwise some children might be failing to check the
 364          * moved-down quals.  To make them match up, we can try to increase the
 365          * parameterization of lesser-parameterized paths.
 366          */
 367         foreach(l, all_child_outers)
 368         {
 369                 Relids          required_outer = (Relids) lfirst(l);
 370                 ListCell   *lcr;
 371
 372                 /* Select the child paths for an Append with this parameterization */
 373                 subpaths = NIL;
 374                 subpaths_valid = true;
 375                 foreach(lcr, live_childrels)
 376                 {
 377                         RelOptInfo *childrel = (RelOptInfo *) lfirst(lcr);
 378                         Path       *subpath;
 379
 380                         subpath = get_cheapest_parameterized_child_path(root,
 381                                                                                                                         childrel,
 382                                                                                                                         required_outer);
 383                         if (subpath == NULL)
 384                         {
 385                                 /* failed to make a suitable path for this child */
 386                                 subpaths_valid = false;
 387                                 break;
 388                         }
 389                         subpaths = accumulate_append_subpath(subpaths, subpath);
 390                 }
 391
 392                 if (subpaths_valid)
 393                         add_path(rel, (Path *)
 394                                          create_append_path(rel, subpaths, required_outer, 0,
 395                                                                                 partitioned_rels));
 396         }
 397 }
 398
 399
 400 /*
 401  * generate_mergeappend_paths
 402  *              Generate MergeAppend paths for an append relation
 403  *
 404  * Generate a path for each ordering (pathkey list) appearing in
 405  * all_child_pathkeys.
 406  *
 407  * We consider both cheapest-startup and cheapest-total cases, ie, for each
 408  * interesting ordering, collect all the cheapest startup subpaths and all the
 409  * cheapest total paths, and build a MergeAppend path for each case.
 410  *
 411  * We don't currently generate any parameterized MergeAppend paths.  While
 412  * it would not take much more code here to do so, it's very unclear that it
 413  * is worth the planning cycles to investigate such paths: there's little
 414  * use for an ordered path on the inside of a nestloop.  In fact, it's likely
 415  * that the current coding of add_path would reject such paths out of hand,
 416  * because add_path gives no credit for sort ordering of parameterized paths,
 417  * and a parameterized MergeAppend is going to be more expensive than the
 418  * corresponding parameterized Append path.  If we ever try harder to support
 419  * parameterized mergejoin plans, it might be worth adding support for
 420  * parameterized MergeAppends to feed such joins.  (See notes in
 421  * optimizer/README for why that might not ever happen, though.)
 422  */
 423 static void
 424 generate_mergeappend_paths(PlannerInfo *root, RelOptInfo *rel,
 425                                                    List *live_childrels,
 426                                                    List *all_child_pathkeys,
 427                                                    List *partitioned_rels)
 428 {
 429         ListCell   *lcp;
 430
 431         foreach(lcp, all_child_pathkeys)
 432         {
 433                 List       *pathkeys = (List *) lfirst(lcp);
 434                 List       *startup_subpaths = NIL;
 435                 List       *total_subpaths = NIL;
 436                 bool            startup_neq_total = false;
 437                 ListCell   *lcr;
 438
 439                 /* Select the child paths for this ordering... */
 440                 foreach(lcr, live_childrels)
 441                 {
 442                         RelOptInfo *childrel = (RelOptInfo *) lfirst(lcr);
 443                         Path       *cheapest_startup,
 444                                            *cheapest_total;
 445
 446                         /* Locate the right paths, if they are available. */
 447                         cheapest_startup =
 448                                 get_cheapest_path_for_pathkeys(childrel->pathlist,
 449                                                                                            pathkeys,
 450                                                                                            NULL,
 451                                                                                            STARTUP_COST,
 452                                                                                            false);
 453                         cheapest_total =
 454                                 get_cheapest_path_for_pathkeys(childrel->pathlist,
 455                                                                                            pathkeys,
 456                                                                                            NULL,
 457                                                                                            TOTAL_COST,
 458                                                                                            false);
 459
 460                         /*
 461                          * If we can't find any paths with the right order just use the
 462                          * cheapest-total path; we'll have to sort it later.
 463                          */
 464                         if (cheapest_startup == NULL || cheapest_total == NULL)
 465                         {
 466                                 cheapest_startup = cheapest_total =
 467                                         childrel->cheapest_total_path;
 468                                 /* Assert we do have an unparameterized path for this child */
 469                                 Assert(cheapest_total->param_info == NULL);
 470                         }
 471
 472                         /*
 473                          * Notice whether we actually have different paths for the
 474                          * "cheapest" and "total" cases; frequently there will be no point
 475                          * in two create_merge_append_path() calls.
 476                          */
 477                         if (cheapest_startup != cheapest_total)
 478                                 startup_neq_total = true;
 479
 480                         startup_subpaths =
 481                                 accumulate_append_subpath(startup_subpaths, cheapest_startup);
 482                         total_subpaths =
 483                                 accumulate_append_subpath(total_subpaths, cheapest_total);
 484                 }
 485
 486                 /* ... and build the MergeAppend paths */
 487                 add_path(rel, (Path *) create_merge_append_path(root,
 488                                                                                                                 rel,
 489                                                                                                                 startup_subpaths,
 490                                                                                                                 pathkeys,
 491                                                                                                                 NULL,
 492                                                                                                                 partitioned_rels));
 493                 if (startup_neq_total)
 494                         add_path(rel, (Path *) create_merge_append_path(root,
 495                                                                                                                         rel,
 496                                                                                                                         total_subpaths,
 497                                                                                                                         pathkeys,
 498                                                                                                                         NULL,
 499                                                                                                                         partitioned_rels));
 500         }
 501 }
 502
 503
 504 /*
 505  * get_cheapest_parameterized_child_path
 506  *              Get cheapest path for this relation that has exactly the requested
 507  *              parameterization.
 508  *
 509  * Returns NULL if unable to create such a path.
 510  */
 511 static Path *
 512 get_cheapest_parameterized_child_path(PlannerInfo *root, RelOptInfo *rel,
 513                                                                           Relids required_outer)
 514 {
 515         Path       *cheapest;
 516         ListCell   *lc;
 517
 518         /*
 519          * Look up the cheapest existing path with no more than the needed
 520          * parameterization.  If it has exactly the needed parameterization, we're
 521          * done.
 522          */
 523         cheapest = get_cheapest_path_for_pathkeys(rel->pathlist,
 524                                                                                           NIL,
 525                                                                                           required_outer,
 526                                                                                           TOTAL_COST,
 527                                                                                           false);
 528         Assert(cheapest != NULL);
 529         if (bms_equal(PATH_REQ_OUTER(cheapest), required_outer))
 530                 return cheapest;
 531
 532         /*
 533          * Otherwise, we can "reparameterize" an existing path to match the given
 534          * parameterization, which effectively means pushing down additional
 535          * joinquals to be checked within the path's scan.  However, some existing
 536          * paths might check the available joinquals already while others don't;
 537          * therefore, it's not clear which existing path will be cheapest after
 538          * reparameterization.  We have to go through them all and find out.
 539          */
 540         cheapest = NULL;
 541         foreach(lc, rel->pathlist)
 542         {
 543                 Path       *path = (Path *) lfirst(lc);
 544
 545                 /* Can't use it if it needs more than requested parameterization */
 546                 if (!bms_is_subset(PATH_REQ_OUTER(path), required_outer))
 547                         continue;
 548
 549                 /*
 550                  * Reparameterization can only increase the path's cost, so if it's
 551                  * already more expensive than the current cheapest, forget it.
 552                  */
 553                 if (cheapest != NULL &&
 554                         compare_path_costs(cheapest, path, TOTAL_COST) <= 0)
 555                         continue;
 556
 557                 /* Reparameterize if needed, then recheck cost */
 558                 if (!bms_equal(PATH_REQ_OUTER(path), required_outer))
 559                 {
 560                         path = reparameterize_path(root, path, required_outer, 1.0);
 561                         if (path == NULL)
 562                                 continue;               /* failed to reparameterize this one */
 563                         Assert(bms_equal(PATH_REQ_OUTER(path), required_outer));
 564
 565                         if (cheapest != NULL &&
 566                                 compare_path_costs(cheapest, path, TOTAL_COST) <= 0)
 567                                 continue;
 568                 }
 569
 570                 /* We have a new best path */
 571                 cheapest = path;
 572         }
 573
 574         /* Return the best path, or NULL if we found no suitable candidate */
 575         return cheapest;
 576 }
 577
 578
 579 /*
 580  * accumulate_append_subpath
 581  *              Add a subpath to the list being built for an Append or MergeAppend
 582  *
 583  * It's possible that the child is itself an Append or MergeAppend path, in
 584  * which case we can "cut out the middleman" and just add its child paths to
 585  * our own list.  (We don't try to do this earlier because we need to apply
 586  * both levels of transformation to the quals.)
 587  *
 588  * Note that if we omit a child MergeAppend in this way, we are effectively
 589  * omitting a sort step, which seems fine: if the parent is to be an Append,
 590  * its result would be unsorted anyway, while if the parent is to be a
 591  * MergeAppend, there's no point in a separate sort on a child.
 592  */
 593 static List *
 594 accumulate_append_subpath(List *subpaths, Path *path)
 595 {
 596         if (IsA(path, AppendPath))
 597         {
 598                 AppendPath *apath = (AppendPath *) path;
 599
 600                 /* list_copy is important here to avoid sharing list substructure */
 601                 return list_concat(subpaths, list_copy(apath->subpaths));
 602         }
 603         else if (IsA(path, MergeAppendPath))
 604         {
 605                 MergeAppendPath *mpath = (MergeAppendPath *) path;
 606
 607                 /* list_copy is important here to avoid sharing list substructure */
 608                 return list_concat(subpaths, list_copy(mpath->subpaths));
 609         }
 610         else
 611                 return lappend(subpaths, path);
 612 }
 613
 614
 615 /*
 616  * standard_join_search
 617  *        Find possible joinpaths for a query by successively finding ways
 618  *        to join component relations into join relations.
 619  *
 620  * 'levels_needed' is the number of iterations needed, ie, the number of
 621  *              independent jointree items in the query.  This is > 1.
 622  *
 623  * 'initial_rels' is a list of RelOptInfo nodes for each independent
 624  *              jointree item.  These are the components to be joined together.
 625  *              Note that levels_needed == list_length(initial_rels).
 626  *
 627  * Returns the final level of join relations, i.e., the relation that is
 628  * the result of joining all the original relations together.
 629  * At least one implementation path must be provided for this relation and
 630  * all required sub-relations.
 631  *
 632  * To support loadable plugins that modify planner behavior by changing the
 633  * join searching algorithm, we provide a hook variable that lets a plugin
 634  * replace or supplement this function.  Any such hook must return the same
 635  * final join relation as the standard code would, but it might have a
 636  * different set of implementation paths attached, and only the sub-joinrels
 637  * needed for these paths need have been instantiated.
 638  *
 639  * Note to plugin authors: the functions invoked during standard_join_search()
 640  * modify root->join_rel_list and root->join_rel_hash.  If you want to do more
 641  * than one join-order search, you'll probably need to save and restore the
 642  * original states of those data structures.  See geqo_eval() for an example.
 643  */
 644 RelOptInfo *
 645 standard_join_search(PlannerInfo *root, int levels_needed, List *initial_rels)
 646 {
 647         int                     lev;
 648         RelOptInfo *rel;
 649
 650         /*
 651          * This function cannot be invoked recursively within any one planning
 652          * problem, so join_rel_level[] can't be in use already.
 653          */
 654         Assert(root->join_rel_level == NULL);
 655
 656         /*
 657          * We employ a simple "dynamic programming" algorithm: we first find all
 658          * ways to build joins of two jointree items, then all ways to build joins
 659          * of three items (from two-item joins and single items), then four-item
 660          * joins, and so on until we have considered all ways to join all the
 661          * items into one rel.
 662          *
 663          * root->join_rel_level[j] is a list of all the j-item rels.  Initially we
 664          * set root->join_rel_level[1] to represent all the single-jointree-item
 665          * relations.
 666          */
 667         root->join_rel_level = (List **) palloc0((levels_needed + 1) * sizeof(List *));
 668
 669         root->join_rel_level[1] = initial_rels;
 670
 671         for (lev = 2; lev <= levels_needed; lev++)
 672         {
 673                 ListCell   *lc;
 674
 675                 /*
 676                  * Determine all possible pairs of relations to be joined at this
 677                  * level, and build paths for making each one from every available
 678                  * pair of lower-level relations.
 679                  */
 680                 join_search_one_level(root, lev);
 681
 682                 /*
 683                  * Run generate_gather_paths() for each just-processed joinrel.  We
 684                  * could not do this earlier because both regular and partial paths
 685                  * can get added to a particular joinrel at multiple times within
 686                  * join_search_one_level.  After that, we're done creating paths for
 687                  * the joinrel, so run set_cheapest().
 688                  */
 689                 foreach(lc, root->join_rel_level[lev])
 690                 {
 691                         rel = (RelOptInfo *) lfirst(lc);
 692
 693                         /* Create GatherPaths for any useful partial paths for rel */
 694                         generate_gather_paths(root, rel);
 695
 696                         /* Find and save the cheapest paths for this rel */
 697                         set_cheapest(rel);
 698
 699 #ifdef OPTIMIZER_DEBUG
 700                         debug_print_rel(root, rel);
 701 #endif
 702                 }
 703         }
 704
 705         /*
 706          * We should have a single rel at the final level.
 707          */
 708         if (root->join_rel_level[levels_needed] == NIL)
 709                 elog(ERROR, "failed to build any %d-way joins", levels_needed);
 710         Assert(list_length(root->join_rel_level[levels_needed]) == 1);
 711
 712         rel = (RelOptInfo *) linitial(root->join_rel_level[levels_needed]);
 713
 714         root->join_rel_level = NULL;
 715
 716         return rel;
 717 }
 718
 719 /*
 720  * create_plain_partial_paths
 721  *        Build partial access paths for parallel scan of a plain relation
 722  */
 723 static void
 724 create_plain_partial_paths(PlannerInfo *root, RelOptInfo *rel)
 725 {
 726         int                     parallel_workers;
 727
 728         parallel_workers = compute_parallel_worker(rel, rel->pages, -1);
 729
 730         /* If any limit was set to zero, the user doesn't want a parallel scan. */
 731         if (parallel_workers <= 0)
 732                 return;
 733
 734         /* Add an unordered partial path based on a parallel sequential scan. */
 735         add_partial_path(rel, create_seqscan_path(root, rel, NULL, parallel_workers));
 736 }
 737
 738
 739 /*
 740  * join_search_one_level
 741  *        Consider ways to produce join relations containing exactly 'level'
 742  *        jointree items.  (This is one step of the dynamic-programming method
 743  *        embodied in standard_join_search.)  Join rel nodes for each feasible
 744  *        combination of lower-level rels are created and returned in a list.
 745  *        Implementation paths are created for each such joinrel, too.
 746  *
 747  * level: level of rels we want to make this time
 748  * root->join_rel_level[j], 1 <= j < level, is a list of rels containing j items
 749  *
 750  * The result is returned in root->join_rel_level[level].
 751  */
 752 void
 753 join_search_one_level(PlannerInfo *root, int level)
 754 {
 755         List      **joinrels = root->join_rel_level;
 756         ListCell   *r;
 757         int                     k;
 758
 759         Assert(joinrels[level] == NIL);
 760
 761         /* Set join_cur_level so that new joinrels are added to proper list */
 762         root->join_cur_level = level;
 763
 764         /*
 765          * First, consider left-sided and right-sided plans, in which rels of
 766          * exactly level-1 member relations are joined against initial relations.
 767          * We prefer to join using join clauses, but if we find a rel of level-1
 768          * members that has no join clauses, we will generate Cartesian-product
 769          * joins against all initial rels not already contained in it.
 770          */
 771         foreach(r, joinrels[level - 1])
 772         {
 773                 RelOptInfo *old_rel = (RelOptInfo *) lfirst(r);
 774
 775                 if (old_rel->joininfo != NIL || old_rel->has_eclass_joins ||
 776                         has_join_restriction(root, old_rel))
 777                 {
 778                         /*
 779                          * There are join clauses or join order restrictions relevant to
 780                          * this rel, so consider joins between this rel and (only) those
 781                          * initial rels it is linked to by a clause or restriction.
 782                          *
 783                          * At level 2 this condition is symmetric, so there is no need to
 784                          * look at initial rels before this one in the list; we already
 785                          * considered such joins when we were at the earlier rel.  (The
 786                          * mirror-image joins are handled automatically by make_join_rel.)
 787                          * In later passes (level > 2), we join rels of the previous level
 788                          * to each initial rel they don't already include but have a join
 789                          * clause or restriction with.
 790                          */
 791                         ListCell   *other_rels;
 792
 793                         if (level == 2)         /* consider remaining initial rels */
 794                                 other_rels = lnext(r);
 795                         else                            /* consider all initial rels */
 796                                 other_rels = list_head(joinrels[1]);
 797
 798                         make_rels_by_clause_joins(root,
 799                                                                           old_rel,
 800                                                                           other_rels);
 801                 }
 802                 else
 803                 {
 804                         /*
 805                          * Oops, we have a relation that is not joined to any other
 806                          * relation, either directly or by join-order restrictions.
 807                          * Cartesian product time.
 808                          *
 809                          * We consider a cartesian product with each not-already-included
 810                          * initial rel, whether it has other join clauses or not.  At
 811                          * level 2, if there are two or more clauseless initial rels, we
 812                          * will redundantly consider joining them in both directions; but
 813                          * such cases aren't common enough to justify adding complexity to
 814                          * avoid the duplicated effort.
 815                          */
 816                         make_rels_by_clauseless_joins(root,
 817                                                                                   old_rel,
 818                                                                                   list_head(joinrels[1]));
 819                 }
 820         }
 821
 822         /*
 823          * Now, consider "bushy plans" in which relations of k initial rels are
 824          * joined to relations of level-k initial rels, for 2 <= k <= level-2.
 825          *
 826          * We only consider bushy-plan joins for pairs of rels where there is a
 827          * suitable join clause (or join order restriction), in order to avoid
 828          * unreasonable growth of planning time.
 829          */
 830         for (k = 2;; k++)
 831         {
 832                 int                     other_level = level - k;
 833
 834                 /*
 835                  * Since make_join_rel(x, y) handles both x,y and y,x cases, we only
 836                  * need to go as far as the halfway point.
 837                  */
 838                 if (k > other_level)
 839                         break;
 840
 841                 foreach(r, joinrels[k])
 842                 {
 843                         RelOptInfo *old_rel = (RelOptInfo *) lfirst(r);
 844                         ListCell   *other_rels;
 845                         ListCell   *r2;
 846
 847                         /*
 848                          * We can ignore relations without join clauses here, unless they
 849                          * participate in join-order restrictions --- then we might have
 850                          * to force a bushy join plan.
 851                          */
 852                         if (old_rel->joininfo == NIL && !old_rel->has_eclass_joins &&
 853                                 !has_join_restriction(root, old_rel))
 854                                 continue;
 855
 856                         if (k == other_level)
 857                                 other_rels = lnext(r);  /* only consider remaining rels */
 858                         else
 859                                 other_rels = list_head(joinrels[other_level]);
 860
 861                         for_each_cell(r2, other_rels)
 862                         {
 863                                 RelOptInfo *new_rel = (RelOptInfo *) lfirst(r2);
 864
 865                                 if (!bms_overlap(old_rel->relids, new_rel->relids))
 866                                 {
 867                                         /*
 868                                          * OK, we can build a rel of the right level from this
 869                                          * pair of rels.  Do so if there is at least one relevant
 870                                          * join clause or join order restriction.
 871                                          */
 872                                         if (have_relevant_joinclause(root, old_rel, new_rel) ||
 873                                                 have_join_order_restriction(root, old_rel, new_rel))
 874                                         {
 875                                                 (void) make_join_rel(root, old_rel, new_rel);
 876                                         }
 877                                 }
 878                         }
 879                 }
 880         }
 881
 882         /*----------
 883          * Last-ditch effort: if we failed to find any usable joins so far, force
 884          * a set of cartesian-product joins to be generated.  This handles the
 885          * special case where all the available rels have join clauses but we
 886          * cannot use any of those clauses yet.  This can only happen when we are
 887          * considering a join sub-problem (a sub-joinlist) and all the rels in the
 888          * sub-problem have only join clauses with rels outside the sub-problem.
 889          * An example is
 890          *
 891          *              SELECT ... FROM a INNER JOIN b ON TRUE, c, d, ...
 892          *              WHERE a.w = c.x and b.y = d.z;
 893          *
 894          * If the "a INNER JOIN b" sub-problem does not get flattened into the
 895          * upper level, we must be willing to make a cartesian join of a and b;
 896          * but the code above will not have done so, because it thought that both
 897          * a and b have joinclauses.  We consider only left-sided and right-sided
 898          * cartesian joins in this case (no bushy).
 899          *----------
 900          */
 901         if (joinrels[level] == NIL)
 902         {
 903                 /*
 904                  * This loop is just like the first one, except we always call
 905                  * make_rels_by_clauseless_joins().
 906                  */
 907                 foreach(r, joinrels[level - 1])
 908                 {
 909                         RelOptInfo *old_rel = (RelOptInfo *) lfirst(r);
 910
 911                         make_rels_by_clauseless_joins(root,
 912                                                                                   old_rel,
 913                                                                                   list_head(joinrels[1]));
 914                 }
 915
 916                 /*----------
 917                  * When special joins are involved, there may be no legal way
 918                  * to make an N-way join for some values of N.  For example consider
 919                  *
 920                  * SELECT ... FROM t1 WHERE
 921                  *       x IN (SELECT ... FROM t2,t3 WHERE ...) AND
 922                  *       y IN (SELECT ... FROM t4,t5 WHERE ...)
 923                  *
 924                  * We will flatten this query to a 5-way join problem, but there are
 925                  * no 4-way joins that join_is_legal() will consider legal.  We have
 926                  * to accept failure at level 4 and go on to discover a workable
 927                  * bushy plan at level 5.
 928                  *
 929                  * However, if there are no special joins and no lateral references
 930                  * then join_is_legal() should never fail, and so the following sanity
 931                  * check is useful.
 932                  *----------
 933                  */
 934                 if (joinrels[level] == NIL &&
 935                         root->join_info_list == NIL &&
 936                         !root->hasLateralRTEs)
 937                         elog(ERROR, "failed to build any %d-way joins", level);
 938         }
 939 }
 940
 941
 942 /*
 943  * make_rels_by_clause_joins
 944  *        Build joins between the given relation 'old_rel' and other relations
 945  *        that participate in join clauses that 'old_rel' also participates in
 946  *        (or participate in join-order restrictions with it).
 947  *        The join rels are returned in root->join_rel_level[join_cur_level].
 948  *
 949  * Note: at levels above 2 we will generate the same joined relation in
 950  * multiple ways --- for example (a join b) join c is the same RelOptInfo as
 951  * (b join c) join a, though the second case will add a different set of Paths
 952  * to it.  This is the reason for using the join_rel_level mechanism, which
 953  * automatically ensures that each new joinrel is only added to the list once.
 954  *
 955  * 'old_rel' is the relation entry for the relation to be joined
 956  * 'other_rels': the first cell in a linked list containing the other
 957  * rels to be considered for joining
 958  *
 959  * Currently, this is only used with initial rels in other_rels, but it
 960  * will work for joining to joinrels too.
 961  */
 962 static void
 963 make_rels_by_clause_joins(PlannerInfo *root,
 964                                                   RelOptInfo *old_rel,
 965                                                   ListCell *other_rels)
 966 {
 967         ListCell   *l;
 968
 969         for_each_cell(l, other_rels)
 970         {
 971                 RelOptInfo *other_rel = (RelOptInfo *) lfirst(l);
 972
 973                 if (!bms_overlap(old_rel->relids, other_rel->relids) &&
 974                         (have_relevant_joinclause(root, old_rel, other_rel) ||
 975                          have_join_order_restriction(root, old_rel, other_rel)))
 976                 {
 977                         (void) make_join_rel(root, old_rel, other_rel);
 978                 }
 979         }
 980 }
 981
 982
 983 /*
 984  * make_rels_by_clauseless_joins
 985  *        Given a relation 'old_rel' and a list of other relations
 986  *        'other_rels', create a join relation between 'old_rel' and each
 987  *        member of 'other_rels' that isn't already included in 'old_rel'.
 988  *        The join rels are returned in root->join_rel_level[join_cur_level].
 989  *
 990  * 'old_rel' is the relation entry for the relation to be joined
 991  * 'other_rels': the first cell of a linked list containing the
 992  * other rels to be considered for joining
 993  *
 994  * Currently, this is only used with initial rels in other_rels, but it would
 995  * work for joining to joinrels too.
 996  */
 997 static void
 998 make_rels_by_clauseless_joins(PlannerInfo *root,
 999                                                           RelOptInfo *old_rel,
1000                                                           ListCell *other_rels)
1001 {
1002         ListCell   *l;
1003
1004         for_each_cell(l, other_rels)
1005         {
1006                 RelOptInfo *other_rel = (RelOptInfo *) lfirst(l);
1007
1008                 if (!bms_overlap(other_rel->relids, old_rel->relids))
1009                 {
1010                         (void) make_join_rel(root, old_rel, other_rel);
1011                 }
1012         }
1013 }
1014
1015
1016 /*
1017  * join_is_legal
1018  *         Determine whether a proposed join is legal given the query's
1019  *         join order constraints; and if it is, determine the join type.
1020  *
1021  * Caller must supply not only the two rels, but the union of their relids.
1022  * (We could simplify the API by computing joinrelids locally, but this
1023  * would be redundant work in the normal path through make_join_rel.)
1024  *
1025  * On success, *sjinfo_p is set to NULL if this is to be a plain inner join,
1026  * else it's set to point to the associated SpecialJoinInfo node.  Also,
1027  * *reversed_p is set TRUE if the given relations need to be swapped to
1028  * match the SpecialJoinInfo node.
1029  */
1030 static bool
1031 join_is_legal(PlannerInfo *root, RelOptInfo *rel1, RelOptInfo *rel2,
1032                           Relids joinrelids,
1033                           SpecialJoinInfo **sjinfo_p, bool *reversed_p)
1034 {
1035         SpecialJoinInfo *match_sjinfo;
1036         bool            reversed;
1037         bool            unique_ified;
1038         bool            must_be_leftjoin;
1039         ListCell   *l;
1040
1041         /*
1042          * Ensure output params are set on failure return.  This is just to
1043          * suppress uninitialized-variable warnings from overly anal compilers.
1044          */
1045         *sjinfo_p = NULL;
1046         *reversed_p = false;
1047
1048         /*
1049          * If we have any special joins, the proposed join might be illegal; and
1050          * in any case we have to determine its join type.  Scan the join info
1051          * list for matches and conflicts.
1052          */
1053         match_sjinfo = NULL;
1054         reversed = false;
1055         unique_ified = false;
1056         must_be_leftjoin = false;
1057
1058         foreach(l, root->join_info_list)
1059         {
1060                 SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
1061
1062                 /*
1063                  * This special join is not relevant unless its RHS overlaps the
1064                  * proposed join.  (Check this first as a fast path for dismissing
1065                  * most irrelevant SJs quickly.)
1066                  */
1067                 if (!bms_overlap(sjinfo->min_righthand, joinrelids))
1068                         continue;
1069
1070                 /*
1071                  * Also, not relevant if proposed join is fully contained within RHS
1072                  * (ie, we're still building up the RHS).
1073                  */
1074                 if (bms_is_subset(joinrelids, sjinfo->min_righthand))
1075                         continue;
1076
1077                 /*
1078                  * Also, not relevant if SJ is already done within either input.
1079                  */
1080                 if (bms_is_subset(sjinfo->min_lefthand, rel1->relids) &&
1081                         bms_is_subset(sjinfo->min_righthand, rel1->relids))
1082                         continue;
1083                 if (bms_is_subset(sjinfo->min_lefthand, rel2->relids) &&
1084                         bms_is_subset(sjinfo->min_righthand, rel2->relids))
1085                         continue;
1086
1087                 /*
1088                  * If it's a semijoin and we already joined the RHS to any other rels
1089                  * within either input, then we must have unique-ified the RHS at that
1090                  * point (see below).  Therefore the semijoin is no longer relevant in
1091                  * this join path.
1092                  */
1093                 if (sjinfo->jointype == JOIN_SEMI)
1094                 {
1095                         if (bms_is_subset(sjinfo->syn_righthand, rel1->relids) &&
1096                                 !bms_equal(sjinfo->syn_righthand, rel1->relids))
1097                                 continue;
1098                         if (bms_is_subset(sjinfo->syn_righthand, rel2->relids) &&
1099                                 !bms_equal(sjinfo->syn_righthand, rel2->relids))
1100                                 continue;
1101                 }
1102
1103                 /*
1104                  * If one input contains min_lefthand and the other contains
1105                  * min_righthand, then we can perform the SJ at this join.
1106                  *
1107                  * Reject if we get matches to more than one SJ; that implies we're
1108                  * considering something that's not really valid.
1109                  */
1110                 if (bms_is_subset(sjinfo->min_lefthand, rel1->relids) &&
1111                         bms_is_subset(sjinfo->min_righthand, rel2->relids))
1112                 {
1113                         if (match_sjinfo)
1114                                 return false;   /* invalid join path */
1115                         match_sjinfo = sjinfo;
1116                         reversed = false;
1117                 }
1118                 else if (bms_is_subset(sjinfo->min_lefthand, rel2->relids) &&
1119                                  bms_is_subset(sjinfo->min_righthand, rel1->relids))
1120                 {
1121                         if (match_sjinfo)
1122                                 return false;   /* invalid join path */
1123                         match_sjinfo = sjinfo;
1124                         reversed = true;
1125                 }
1126                 else if (sjinfo->jointype == JOIN_SEMI &&
1127                                  bms_equal(sjinfo->syn_righthand, rel2->relids) &&
1128                                  create_unique_path(root, rel2, rel2->cheapest_total_path,
1129                                                                         sjinfo) != NULL)
1130                 {
1131                         /*----------
1132                          * For a semijoin, we can join the RHS to anything else by
1133                          * unique-ifying the RHS (if the RHS can be unique-ified).
1134                          * We will only get here if we have the full RHS but less
1135                          * than min_lefthand on the LHS.
1136                          *
1137                          * The reason to consider such a join path is exemplified by
1138                          *      SELECT ... FROM a,b WHERE (a.x,b.y) IN (SELECT c1,c2 FROM c)
1139                          * If we insist on doing this as a semijoin we will first have
1140                          * to form the cartesian product of A*B.  But if we unique-ify
1141                          * C then the semijoin becomes a plain innerjoin and we can join
1142                          * in any order, eg C to A and then to B.  When C is much smaller
1143                          * than A and B this can be a huge win.  So we allow C to be
1144                          * joined to just A or just B here, and then make_join_rel has
1145                          * to handle the case properly.
1146                          *
1147                          * Note that actually we'll allow unique-ified C to be joined to
1148                          * some other relation D here, too.  That is legal, if usually not
1149                          * very sane, and this routine is only concerned with legality not
1150                          * with whether the join is good strategy.
1151                          *----------
1152                          */
1153                         if (match_sjinfo)
1154                                 return false;   /* invalid join path */
1155                         match_sjinfo = sjinfo;
1156                         reversed = false;
1157                         unique_ified = true;
1158                 }
1159                 else if (sjinfo->jointype == JOIN_SEMI &&
1160                                  bms_equal(sjinfo->syn_righthand, rel1->relids) &&
1161                                  create_unique_path(root, rel1, rel1->cheapest_total_path,
1162                                                                         sjinfo) != NULL)
1163                 {
1164                         /* Reversed semijoin case */
1165                         if (match_sjinfo)
1166                                 return false;   /* invalid join path */
1167                         match_sjinfo = sjinfo;
1168                         reversed = true;
1169                         unique_ified = true;
1170                 }
1171                 else
1172                 {
1173                         /*
1174                          * Otherwise, the proposed join overlaps the RHS but isn't a valid
1175                          * implementation of this SJ.  But don't panic quite yet: the RHS
1176                          * violation might have occurred previously, in one or both input
1177                          * relations, in which case we must have previously decided that
1178                          * it was OK to commute some other SJ with this one.  If we need
1179                          * to perform this join to finish building up the RHS, rejecting
1180                          * it could lead to not finding any plan at all.  (This can occur
1181                          * because of the heuristics elsewhere in this file that postpone
1182                          * clauseless joins: we might not consider doing a clauseless join
1183                          * within the RHS until after we've performed other, validly
1184                          * commutable SJs with one or both sides of the clauseless join.)
1185                          * This consideration boils down to the rule that if both inputs
1186                          * overlap the RHS, we can allow the join --- they are either
1187                          * fully within the RHS, or represent previously-allowed joins to
1188                          * rels outside it.
1189                          */
1190                         if (bms_overlap(rel1->relids, sjinfo->min_righthand) &&
1191                                 bms_overlap(rel2->relids, sjinfo->min_righthand))
1192                                 continue;               /* assume valid previous violation of RHS */
1193
1194                         /*
1195                          * The proposed join could still be legal, but only if we're
1196                          * allowed to associate it into the RHS of this SJ.  That means
1197                          * this SJ must be a LEFT join (not SEMI or ANTI, and certainly
1198                          * not FULL) and the proposed join must not overlap the LHS.
1199                          */
1200                         if (sjinfo->jointype != JOIN_LEFT ||
1201                                 bms_overlap(joinrelids, sjinfo->min_lefthand))
1202                                 return false;   /* invalid join path */
1203
1204                         /*
1205                          * To be valid, the proposed join must be a LEFT join; otherwise
1206                          * it can't associate into this SJ's RHS.  But we may not yet have
1207                          * found the SpecialJoinInfo matching the proposed join, so we
1208                          * can't test that yet.  Remember the requirement for later.
1209                          */
1210                         must_be_leftjoin = true;
1211                 }
1212         }
1213
1214         /*
1215          * Fail if violated any SJ's RHS and didn't match to a LEFT SJ: the
1216          * proposed join can't associate into an SJ's RHS.
1217          *
1218          * Also, fail if the proposed join's predicate isn't strict; we're
1219          * essentially checking to see if we can apply outer-join identity 3, and
1220          * that's a requirement.  (This check may be redundant with checks in
1221          * make_outerjoininfo, but I'm not quite sure, and it's cheap to test.)
1222          */
1223         if (must_be_leftjoin &&
1224                 (match_sjinfo == NULL ||
1225                  match_sjinfo->jointype != JOIN_LEFT ||
1226                  !match_sjinfo->lhs_strict))
1227                 return false;                   /* invalid join path */
1228
1229         /*
1230          * We also have to check for constraints imposed by LATERAL references.
1231          */
1232         if (root->hasLateralRTEs)
1233         {
1234                 bool            lateral_fwd;
1235                 bool            lateral_rev;
1236                 Relids          join_lateral_rels;
1237
1238                 /*
1239                  * The proposed rels could each contain lateral references to the
1240                  * other, in which case the join is impossible.  If there are lateral
1241                  * references in just one direction, then the join has to be done with
1242                  * a nestloop with the lateral referencer on the inside.  If the join
1243                  * matches an SJ that cannot be implemented by such a nestloop, the
1244                  * join is impossible.
1245                  *
1246                  * Also, if the lateral reference is only indirect, we should reject
1247                  * the join; whatever rel(s) the reference chain goes through must be
1248                  * joined to first.
1249                  *
1250                  * Another case that might keep us from building a valid plan is the
1251                  * implementation restriction described by have_dangerous_phv().
1252                  */
1253                 lateral_fwd = bms_overlap(rel1->relids, rel2->lateral_relids);
1254                 lateral_rev = bms_overlap(rel2->relids, rel1->lateral_relids);
1255                 if (lateral_fwd && lateral_rev)
1256                         return false;           /* have lateral refs in both directions */
1257                 if (lateral_fwd)
1258                 {
1259                         /* has to be implemented as nestloop with rel1 on left */
1260                         if (match_sjinfo &&
1261                                 (reversed ||
1262                                  unique_ified ||
1263                                  match_sjinfo->jointype == JOIN_FULL))
1264                                 return false;   /* not implementable as nestloop */
1265                         /* check there is a direct reference from rel2 to rel1 */
1266                         if (!bms_overlap(rel1->relids, rel2->direct_lateral_relids))
1267                                 return false;   /* only indirect refs, so reject */
1268                         /* check we won't have a dangerous PHV */
1269                         if (have_dangerous_phv(root, rel1->relids, rel2->lateral_relids))
1270                                 return false;   /* might be unable to handle required PHV */
1271                 }
1272                 else if (lateral_rev)
1273                 {
1274                         /* has to be implemented as nestloop with rel2 on left */
1275                         if (match_sjinfo &&
1276                                 (!reversed ||
1277                                  unique_ified ||
1278                                  match_sjinfo->jointype == JOIN_FULL))
1279                                 return false;   /* not implementable as nestloop */
1280                         /* check there is a direct reference from rel1 to rel2 */
1281                         if (!bms_overlap(rel2->relids, rel1->direct_lateral_relids))
1282                                 return false;   /* only indirect refs, so reject */
1283                         /* check we won't have a dangerous PHV */
1284                         if (have_dangerous_phv(root, rel2->relids, rel1->lateral_relids))
1285                                 return false;   /* might be unable to handle required PHV */
1286                 }
1287
1288                 /*
1289                  * LATERAL references could also cause problems later on if we accept
1290                  * this join: if the join's minimum parameterization includes any rels
1291                  * that would have to be on the inside of an outer join with this join
1292                  * rel, then it's never going to be possible to build the complete
1293                  * query using this join.  We should reject this join not only because
1294                  * it'll save work, but because if we don't, the clauseless-join
1295                  * heuristics might think that legality of this join means that some
1296                  * other join rel need not be formed, and that could lead to failure
1297                  * to find any plan at all.  We have to consider not only rels that
1298                  * are directly on the inner side of an OJ with the joinrel, but also
1299                  * ones that are indirectly so, so search to find all such rels.
1300                  */
1301                 join_lateral_rels = min_join_parameterization(root, joinrelids,
1302                                                                                                           rel1, rel2);
1303                 if (join_lateral_rels)
1304                 {
1305                         Relids          join_plus_rhs = bms_copy(joinrelids);
1306                         bool            more;
1307
1308                         do
1309                         {
1310                                 more = false;
1311                                 foreach(l, root->join_info_list)
1312                                 {
1313                                         SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
1314
1315                                         if (bms_overlap(sjinfo->min_lefthand, join_plus_rhs) &&
1316                                                 !bms_is_subset(sjinfo->min_righthand, join_plus_rhs))
1317                                         {
1318                                                 join_plus_rhs = bms_add_members(join_plus_rhs,
1319                                                                                                                 sjinfo->min_righthand);
1320                                                 more = true;
1321                                         }
1322                                         /* full joins constrain both sides symmetrically */
1323                                         if (sjinfo->jointype == JOIN_FULL &&
1324                                                 bms_overlap(sjinfo->min_righthand, join_plus_rhs) &&
1325                                                 !bms_is_subset(sjinfo->min_lefthand, join_plus_rhs))
1326                                         {
1327                                                 join_plus_rhs = bms_add_members(join_plus_rhs,
1328                                                                                                                 sjinfo->min_lefthand);
1329                                                 more = true;
1330                                         }
1331                                 }
1332                         } while (more);
1333                         if (bms_overlap(join_plus_rhs, join_lateral_rels))
1334                                 return false;   /* will not be able to join to some RHS rel */
1335                 }
1336         }
1337
1338         /* Otherwise, it's a valid join */
1339         *sjinfo_p = match_sjinfo;
1340         *reversed_p = reversed;
1341         return true;
1342 }
1343
1344
1345 /*
1346  * has_join_restriction
1347  *              Detect whether the specified relation has join-order restrictions,
1348  *              due to being inside an outer join or an IN (sub-SELECT),
1349  *              or participating in any LATERAL references or multi-rel PHVs.
1350  *
1351  * Essentially, this tests whether have_join_order_restriction() could
1352  * succeed with this rel and some other one.  It's OK if we sometimes
1353  * say "true" incorrectly.  (Therefore, we don't bother with the relatively
1354  * expensive has_legal_joinclause test.)
1355  */
1356 static bool
1357 has_join_restriction(PlannerInfo *root, RelOptInfo *rel)
1358 {
1359         ListCell   *l;
1360
1361         if (rel->lateral_relids != NULL || rel->lateral_referencers != NULL)
1362                 return true;
1363
1364         foreach(l, root->placeholder_list)
1365         {
1366                 PlaceHolderInfo *phinfo = (PlaceHolderInfo *) lfirst(l);
1367
1368                 if (bms_is_subset(rel->relids, phinfo->ph_eval_at) &&
1369                         !bms_equal(rel->relids, phinfo->ph_eval_at))
1370                         return true;
1371         }
1372
1373         foreach(l, root->join_info_list)
1374         {
1375                 SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
1376
1377                 /* ignore full joins --- other mechanisms preserve their ordering */
1378                 if (sjinfo->jointype == JOIN_FULL)
1379                         continue;
1380
1381                 /* ignore if SJ is already contained in rel */
1382                 if (bms_is_subset(sjinfo->min_lefthand, rel->relids) &&
1383                         bms_is_subset(sjinfo->min_righthand, rel->relids))
1384                         continue;
1385
1386                 /* restricted if it overlaps LHS or RHS, but doesn't contain SJ */
1387                 if (bms_overlap(sjinfo->min_lefthand, rel->relids) ||
1388                         bms_overlap(sjinfo->min_righthand, rel->relids))
1389                         return true;
1390         }
1391
1392         return false;
1393 }
1394
1395
1396 /*
1397  * is_dummy_rel --- has relation been proven empty?
1398  */
1399 static bool
1400 is_dummy_rel(RelOptInfo *rel)
1401 {
1402         return IS_DUMMY_REL(rel);
1403 }
1404
1405
1406 /*
1407  * Mark a relation as proven empty.
1408  *
1409  * During GEQO planning, this can get invoked more than once on the same
1410  * baserel struct, so it's worth checking to see if the rel is already marked
1411  * dummy.
1412  *
1413  * Also, when called during GEQO join planning, we are in a short-lived
1414  * memory context.  We must make sure that the dummy path attached to a
1415  * baserel survives the GEQO cycle, else the baserel is trashed for future
1416  * GEQO cycles.  On the other hand, when we are marking a joinrel during GEQO,
1417  * we don't want the dummy path to clutter the main planning context.  Upshot
1418  * is that the best solution is to explicitly make the dummy path in the same
1419  * context the given RelOptInfo is in.
1420  */
1421 static void
1422 mark_dummy_rel(RelOptInfo *rel)
1423 {
1424         MemoryContext oldcontext;
1425
1426         /* Already marked? */
1427         if (is_dummy_rel(rel))
1428                 return;
1429
1430         /* No, so choose correct context to make the dummy path in */
1431         oldcontext = MemoryContextSwitchTo(GetMemoryChunkContext(rel));
1432
1433         /* Set dummy size estimate */
1434         rel->rows = 0;
1435
1436         /* Evict any previously chosen paths */
1437         rel->pathlist = NIL;
1438         rel->partial_pathlist = NIL;
1439
1440         /* Set up the dummy path */
1441         add_path(rel, (Path *) create_append_path(rel, NIL, NULL, 0, NIL));
1442
1443         /* Set or update cheapest_total_path and related fields */
1444         set_cheapest(rel);
1445
1446         MemoryContextSwitchTo(oldcontext);
1447 }
1448
1449
1450 /*
1451  * restriction_is_constant_false --- is a restrictlist just FALSE?
1452  *
1453  * In cases where a qual is provably constant FALSE, eval_const_expressions
1454  * will generally have thrown away anything that's ANDed with it.  In outer
1455  * join situations this will leave us computing cartesian products only to
1456  * decide there's no match for an outer row, which is pretty stupid.  So,
1457  * we need to detect the case.
1458  *
1459  * If only_pushed_down is TRUE, then consider only pushed-down quals.
1460  */
1461 static bool
1462 restriction_is_constant_false(List *restrictlist, bool only_pushed_down)
1463 {
1464         ListCell   *lc;
1465
1466         /*
1467          * Despite the above comment, the restriction list we see here might
1468          * possibly have other members besides the FALSE constant, since other
1469          * quals could get "pushed down" to the outer join level.  So we check
1470          * each member of the list.
1471          */
1472         foreach(lc, restrictlist)
1473         {
1474                 RestrictInfo *rinfo = lfirst_node(RestrictInfo, lc);
1475
1476                 if (only_pushed_down && !rinfo->is_pushed_down)
1477                         continue;
1478
1479                 if (rinfo->clause && IsA(rinfo->clause, Const))
1480                 {
1481                         Const      *con = (Const *) rinfo->clause;
1482
1483                         /* constant NULL is as good as constant FALSE for our purposes */
1484                         if (con->constisnull)
1485                                 return true;
1486                         if (!DatumGetBool(con->constvalue))
1487                                 return true;
1488                 }
1489         }
1490         return false;
1491 }