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  *     generate_mergeappend_paths()
  20  *     get_cheapest_parameterized_child_path()
  21  *     accumulate_append_subpath()
  22  *
  23  *  public functions:
  24  *     standard_join_search(): This funcion is not static. The reason for
  25  *        including this function is make_rels_by_clause_joins. In order to
  26  *        avoid generating apparently unwanted join combination, we decided to
  27  *        change the behavior of make_join_rel, which is called under this
  28  *        function.
  29  *
  30  * src/backend/optimizer/path/joinrels.c
  31  *
  32  *      public functions:
  33  *     join_search_one_level(): We have to modify this to call my definition of
  34  *                  make_rels_by_clause_joins.
  35  *
  36  *      static functions:
  37  *     make_rels_by_clause_joins()
  38  *     make_rels_by_clauseless_joins()
  39  *     join_is_legal()
  40  *     has_join_restriction()
  41  *     is_dummy_rel()
  42  *     mark_dummy_rel()
  43  *     restriction_is_constant_false()
  44  *
  45  *
  46  * Portions Copyright (c) 1996-2017, PostgreSQL Global Development Group
  47  * Portions Copyright (c) 1994, Regents of the University of California
  48  *
  49  *-------------------------------------------------------------------------
  50  */
  51
  52
  53 /*
  54  * set_plain_rel_pathlist
  55  *        Build access paths for a plain relation (no subquery, no inheritance)
  56  */
  57 static void
  58 set_plain_rel_pathlist(PlannerInfo *root, RelOptInfo *rel, RangeTblEntry *rte)
  59 {
  60         Relids          required_outer;
  61
  62         /*
  63          * We don't support pushing join clauses into the quals of a seqscan, but
  64          * it could still have required parameterization due to LATERAL refs in
  65          * its tlist.
  66          */
  67         required_outer = rel->lateral_relids;
  68
  69         /* Consider sequential scan */
  70         add_path(rel, create_seqscan_path(root, rel, required_outer, 0));
  71
  72         /* If appropriate, consider parallel sequential scan */
  73         if (rel->consider_parallel && required_outer == NULL)
  74                 create_plain_partial_paths(root, rel);
  75
  76         /* Consider index scans */
  77         create_index_paths(root, rel);
  78
  79         /* Consider TID scans */
  80         create_tidscan_paths(root, rel);
  81 }
  82
  83 /*
  84  * set_append_rel_pathlist
  85  *        Build access paths for an "append relation"
  86  */
  87 static void
  88 set_append_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
  89                                                 Index rti, RangeTblEntry *rte)
  90 {
  91         int                     parentRTindex = rti;
  92         List       *live_childrels = NIL;
  93         List       *subpaths = NIL;
  94         bool            subpaths_valid = true;
  95         List       *partial_subpaths = NIL;
  96         bool            partial_subpaths_valid = true;
  97         List       *all_child_pathkeys = NIL;
  98         List       *all_child_outers = NIL;
  99         ListCell   *l;
 100
 101         /*
 102          * Generate access paths for each member relation, and remember the
 103          * cheapest path for each one.  Also, identify all pathkeys (orderings)
 104          * and parameterizations (required_outer sets) available for the member
 105          * relations.
 106          */
 107         foreach(l, root->append_rel_list)
 108         {
 109                 AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
 110                 int                     childRTindex;
 111                 RangeTblEntry *childRTE;
 112                 RelOptInfo *childrel;
 113                 ListCell   *lcp;
 114
 115                 /* append_rel_list contains all append rels; ignore others */
 116                 if (appinfo->parent_relid != parentRTindex)
 117                         continue;
 118
 119                 /* Re-locate the child RTE and RelOptInfo */
 120                 childRTindex = appinfo->child_relid;
 121                 childRTE = root->simple_rte_array[childRTindex];
 122                 childrel = root->simple_rel_array[childRTindex];
 123
 124                 /*
 125                  * If set_append_rel_size() decided the parent appendrel was
 126                  * parallel-unsafe at some point after visiting this child rel, we
 127                  * need to propagate the unsafety marking down to the child, so that
 128                  * we don't generate useless partial paths for it.
 129                  */
 130                 if (!rel->consider_parallel)
 131                         childrel->consider_parallel = false;
 132
 133                 /*
 134                  * Compute the child's access paths.
 135                  */
 136                 set_rel_pathlist(root, childrel, childRTindex, childRTE);
 137
 138                 /*
 139                  * If child is dummy, ignore it.
 140                  */
 141                 if (IS_DUMMY_REL(childrel))
 142                         continue;
 143
 144                 /*
 145                  * Child is live, so add it to the live_childrels list for use below.
 146                  */
 147                 live_childrels = lappend(live_childrels, childrel);
 148
 149                 /*
 150                  * If child has an unparameterized cheapest-total path, add that to
 151                  * the unparameterized Append path we are constructing for the parent.
 152                  * If not, there's no workable unparameterized path.
 153                  */
 154                 if (childrel->cheapest_total_path->param_info == NULL)
 155                         subpaths = accumulate_append_subpath(subpaths,
 156                                                                                           childrel->cheapest_total_path);
 157                 else
 158                         subpaths_valid = false;
 159
 160                 /* Same idea, but for a partial plan. */
 161                 if (childrel->partial_pathlist != NIL)
 162                         partial_subpaths = accumulate_append_subpath(partial_subpaths,
 163                                                                            linitial(childrel->partial_pathlist));
 164                 else
 165                         partial_subpaths_valid = false;
 166
 167                 /*
 168                  * Collect lists of all the available path orderings and
 169                  * parameterizations for all the children.  We use these as a
 170                  * heuristic to indicate which sort orderings and parameterizations we
 171                  * should build Append and MergeAppend paths for.
 172                  */
 173                 foreach(lcp, childrel->pathlist)
 174                 {
 175                         Path       *childpath = (Path *) lfirst(lcp);
 176                         List       *childkeys = childpath->pathkeys;
 177                         Relids          childouter = PATH_REQ_OUTER(childpath);
 178
 179                         /* Unsorted paths don't contribute to pathkey list */
 180                         if (childkeys != NIL)
 181                         {
 182                                 ListCell   *lpk;
 183                                 bool            found = false;
 184
 185                                 /* Have we already seen this ordering? */
 186                                 foreach(lpk, all_child_pathkeys)
 187                                 {
 188                                         List       *existing_pathkeys = (List *) lfirst(lpk);
 189
 190                                         if (compare_pathkeys(existing_pathkeys,
 191                                                                                  childkeys) == PATHKEYS_EQUAL)
 192                                         {
 193                                                 found = true;
 194                                                 break;
 195                                         }
 196                                 }
 197                                 if (!found)
 198                                 {
 199                                         /* No, so add it to all_child_pathkeys */
 200                                         all_child_pathkeys = lappend(all_child_pathkeys,
 201                                                                                                  childkeys);
 202                                 }
 203                         }
 204
 205                         /* Unparameterized paths don't contribute to param-set list */
 206                         if (childouter)
 207                         {
 208                                 ListCell   *lco;
 209                                 bool            found = false;
 210
 211                                 /* Have we already seen this param set? */
 212                                 foreach(lco, all_child_outers)
 213                                 {
 214                                         Relids          existing_outers = (Relids) lfirst(lco);
 215
 216                                         if (bms_equal(existing_outers, childouter))
 217                                         {
 218                                                 found = true;
 219                                                 break;
 220                                         }
 221                                 }
 222                                 if (!found)
 223                                 {
 224                                         /* No, so add it to all_child_outers */
 225                                         all_child_outers = lappend(all_child_outers,
 226                                                                                            childouter);
 227                                 }
 228                         }
 229                 }
 230         }
 231
 232         /*
 233          * If we found unparameterized paths for all children, build an unordered,
 234          * unparameterized Append path for the rel.  (Note: this is correct even
 235          * if we have zero or one live subpath due to constraint exclusion.)
 236          */
 237         if (subpaths_valid)
 238                 add_path(rel, (Path *) create_append_path(rel, subpaths, NULL, 0));
 239
 240         /*
 241          * Consider an append of partial unordered, unparameterized partial paths.
 242          */
 243         if (partial_subpaths_valid)
 244         {
 245                 AppendPath *appendpath;
 246                 ListCell   *lc;
 247                 int                     parallel_workers = 0;
 248
 249                 /*
 250                  * Decide on the number of workers to request for this append path.
 251                  * For now, we just use the maximum value from among the members.  It
 252                  * might be useful to use a higher number if the Append node were
 253                  * smart enough to spread out the workers, but it currently isn't.
 254                  */
 255                 foreach(lc, partial_subpaths)
 256                 {
 257                         Path       *path = lfirst(lc);
 258
 259                         parallel_workers = Max(parallel_workers, path->parallel_workers);
 260                 }
 261                 Assert(parallel_workers > 0);
 262
 263                 /* Generate a partial append path. */
 264                 appendpath = create_append_path(rel, partial_subpaths, NULL,
 265                                                                                 parallel_workers);
 266                 add_partial_path(rel, (Path *) appendpath);
 267         }
 268
 269         /*
 270          * Also build unparameterized MergeAppend paths based on the collected
 271          * list of child pathkeys.
 272          */
 273         if (subpaths_valid)
 274                 generate_mergeappend_paths(root, rel, live_childrels,
 275                                                                    all_child_pathkeys);
 276
 277         /*
 278          * Build Append paths for each parameterization seen among the child rels.
 279          * (This may look pretty expensive, but in most cases of practical
 280          * interest, the child rels will expose mostly the same parameterizations,
 281          * so that not that many cases actually get considered here.)
 282          *
 283          * The Append node itself cannot enforce quals, so all qual checking must
 284          * be done in the child paths.  This means that to have a parameterized
 285          * Append path, we must have the exact same parameterization for each
 286          * child path; otherwise some children might be failing to check the
 287          * moved-down quals.  To make them match up, we can try to increase the
 288          * parameterization of lesser-parameterized paths.
 289          */
 290         foreach(l, all_child_outers)
 291         {
 292                 Relids          required_outer = (Relids) lfirst(l);
 293                 ListCell   *lcr;
 294
 295                 /* Select the child paths for an Append with this parameterization */
 296                 subpaths = NIL;
 297                 subpaths_valid = true;
 298                 foreach(lcr, live_childrels)
 299                 {
 300                         RelOptInfo *childrel = (RelOptInfo *) lfirst(lcr);
 301                         Path       *subpath;
 302
 303                         subpath = get_cheapest_parameterized_child_path(root,
 304                                                                                                                         childrel,
 305                                                                                                                         required_outer);
 306                         if (subpath == NULL)
 307                         {
 308                                 /* failed to make a suitable path for this child */
 309                                 subpaths_valid = false;
 310                                 break;
 311                         }
 312                         subpaths = accumulate_append_subpath(subpaths, subpath);
 313                 }
 314
 315                 if (subpaths_valid)
 316                         add_path(rel, (Path *)
 317                                          create_append_path(rel, subpaths, required_outer, 0));
 318         }
 319 }
 320
 321 /*
 322  * generate_mergeappend_paths
 323  *              Generate MergeAppend paths for an append relation
 324  *
 325  * Generate a path for each ordering (pathkey list) appearing in
 326  * all_child_pathkeys.
 327  *
 328  * We consider both cheapest-startup and cheapest-total cases, ie, for each
 329  * interesting ordering, collect all the cheapest startup subpaths and all the
 330  * cheapest total paths, and build a MergeAppend path for each case.
 331  *
 332  * We don't currently generate any parameterized MergeAppend paths.  While
 333  * it would not take much more code here to do so, it's very unclear that it
 334  * is worth the planning cycles to investigate such paths: there's little
 335  * use for an ordered path on the inside of a nestloop.  In fact, it's likely
 336  * that the current coding of add_path would reject such paths out of hand,
 337  * because add_path gives no credit for sort ordering of parameterized paths,
 338  * and a parameterized MergeAppend is going to be more expensive than the
 339  * corresponding parameterized Append path.  If we ever try harder to support
 340  * parameterized mergejoin plans, it might be worth adding support for
 341  * parameterized MergeAppends to feed such joins.  (See notes in
 342  * optimizer/README for why that might not ever happen, though.)
 343  */
 344 static void
 345 generate_mergeappend_paths(PlannerInfo *root, RelOptInfo *rel,
 346                                                    List *live_childrels,
 347                                                    List *all_child_pathkeys)
 348 {
 349         ListCell   *lcp;
 350
 351         foreach(lcp, all_child_pathkeys)
 352         {
 353                 List       *pathkeys = (List *) lfirst(lcp);
 354                 List       *startup_subpaths = NIL;
 355                 List       *total_subpaths = NIL;
 356                 bool            startup_neq_total = false;
 357                 ListCell   *lcr;
 358
 359                 /* Select the child paths for this ordering... */
 360                 foreach(lcr, live_childrels)
 361                 {
 362                         RelOptInfo *childrel = (RelOptInfo *) lfirst(lcr);
 363                         Path       *cheapest_startup,
 364                                            *cheapest_total;
 365
 366                         /* Locate the right paths, if they are available. */
 367                         cheapest_startup =
 368                                 get_cheapest_path_for_pathkeys(childrel->pathlist,
 369                                                                                            pathkeys,
 370                                                                                            NULL,
 371                                                                                            STARTUP_COST);
 372                         cheapest_total =
 373                                 get_cheapest_path_for_pathkeys(childrel->pathlist,
 374                                                                                            pathkeys,
 375                                                                                            NULL,
 376                                                                                            TOTAL_COST);
 377
 378                         /*
 379                          * If we can't find any paths with the right order just use the
 380                          * cheapest-total path; we'll have to sort it later.
 381                          */
 382                         if (cheapest_startup == NULL || cheapest_total == NULL)
 383                         {
 384                                 cheapest_startup = cheapest_total =
 385                                         childrel->cheapest_total_path;
 386                                 /* Assert we do have an unparameterized path for this child */
 387                                 Assert(cheapest_total->param_info == NULL);
 388                         }
 389
 390                         /*
 391                          * Notice whether we actually have different paths for the
 392                          * "cheapest" and "total" cases; frequently there will be no point
 393                          * in two create_merge_append_path() calls.
 394                          */
 395                         if (cheapest_startup != cheapest_total)
 396                                 startup_neq_total = true;
 397
 398                         startup_subpaths =
 399                                 accumulate_append_subpath(startup_subpaths, cheapest_startup);
 400                         total_subpaths =
 401                                 accumulate_append_subpath(total_subpaths, cheapest_total);
 402                 }
 403
 404                 /* ... and build the MergeAppend paths */
 405                 add_path(rel, (Path *) create_merge_append_path(root,
 406                                                                                                                 rel,
 407                                                                                                                 startup_subpaths,
 408                                                                                                                 pathkeys,
 409                                                                                                                 NULL));
 410                 if (startup_neq_total)
 411                         add_path(rel, (Path *) create_merge_append_path(root,
 412                                                                                                                         rel,
 413                                                                                                                         total_subpaths,
 414                                                                                                                         pathkeys,
 415                                                                                                                         NULL));
 416         }
 417 }
 418
 419 /*
 420  * get_cheapest_parameterized_child_path
 421  *              Get cheapest path for this relation that has exactly the requested
 422  *              parameterization.
 423  *
 424  * Returns NULL if unable to create such a path.
 425  */
 426 static Path *
 427 get_cheapest_parameterized_child_path(PlannerInfo *root, RelOptInfo *rel,
 428                                                                           Relids required_outer)
 429 {
 430         Path       *cheapest;
 431         ListCell   *lc;
 432
 433         /*
 434          * Look up the cheapest existing path with no more than the needed
 435          * parameterization.  If it has exactly the needed parameterization, we're
 436          * done.
 437          */
 438         cheapest = get_cheapest_path_for_pathkeys(rel->pathlist,
 439                                                                                           NIL,
 440                                                                                           required_outer,
 441                                                                                           TOTAL_COST);
 442         Assert(cheapest != NULL);
 443         if (bms_equal(PATH_REQ_OUTER(cheapest), required_outer))
 444                 return cheapest;
 445
 446         /*
 447          * Otherwise, we can "reparameterize" an existing path to match the given
 448          * parameterization, which effectively means pushing down additional
 449          * joinquals to be checked within the path's scan.  However, some existing
 450          * paths might check the available joinquals already while others don't;
 451          * therefore, it's not clear which existing path will be cheapest after
 452          * reparameterization.  We have to go through them all and find out.
 453          */
 454         cheapest = NULL;
 455         foreach(lc, rel->pathlist)
 456         {
 457                 Path       *path = (Path *) lfirst(lc);
 458
 459                 /* Can't use it if it needs more than requested parameterization */
 460                 if (!bms_is_subset(PATH_REQ_OUTER(path), required_outer))
 461                         continue;
 462
 463                 /*
 464                  * Reparameterization can only increase the path's cost, so if it's
 465                  * already more expensive than the current cheapest, forget it.
 466                  */
 467                 if (cheapest != NULL &&
 468                         compare_path_costs(cheapest, path, TOTAL_COST) <= 0)
 469                         continue;
 470
 471                 /* Reparameterize if needed, then recheck cost */
 472                 if (!bms_equal(PATH_REQ_OUTER(path), required_outer))
 473                 {
 474                         path = reparameterize_path(root, path, required_outer, 1.0);
 475                         if (path == NULL)
 476                                 continue;               /* failed to reparameterize this one */
 477                         Assert(bms_equal(PATH_REQ_OUTER(path), required_outer));
 478
 479                         if (cheapest != NULL &&
 480                                 compare_path_costs(cheapest, path, TOTAL_COST) <= 0)
 481                                 continue;
 482                 }
 483
 484                 /* We have a new best path */
 485                 cheapest = path;
 486         }
 487
 488         /* Return the best path, or NULL if we found no suitable candidate */
 489         return cheapest;
 490 }
 491
 492 /*
 493  * accumulate_append_subpath
 494  *              Add a subpath to the list being built for an Append or MergeAppend
 495  *
 496  * It's possible that the child is itself an Append or MergeAppend path, in
 497  * which case we can "cut out the middleman" and just add its child paths to
 498  * our own list.  (We don't try to do this earlier because we need to apply
 499  * both levels of transformation to the quals.)
 500  *
 501  * Note that if we omit a child MergeAppend in this way, we are effectively
 502  * omitting a sort step, which seems fine: if the parent is to be an Append,
 503  * its result would be unsorted anyway, while if the parent is to be a
 504  * MergeAppend, there's no point in a separate sort on a child.
 505  */
 506 static List *
 507 accumulate_append_subpath(List *subpaths, Path *path)
 508 {
 509         if (IsA(path, AppendPath))
 510         {
 511                 AppendPath *apath = (AppendPath *) path;
 512
 513                 /* list_copy is important here to avoid sharing list substructure */
 514                 return list_concat(subpaths, list_copy(apath->subpaths));
 515         }
 516         else if (IsA(path, MergeAppendPath))
 517         {
 518                 MergeAppendPath *mpath = (MergeAppendPath *) path;
 519
 520                 /* list_copy is important here to avoid sharing list substructure */
 521                 return list_concat(subpaths, list_copy(mpath->subpaths));
 522         }
 523         else
 524                 return lappend(subpaths, path);
 525 }
 526
 527 /*
 528  * standard_join_search
 529  *        Find possible joinpaths for a query by successively finding ways
 530  *        to join component relations into join relations.
 531  *
 532  * 'levels_needed' is the number of iterations needed, ie, the number of
 533  *              independent jointree items in the query.  This is > 1.
 534  *
 535  * 'initial_rels' is a list of RelOptInfo nodes for each independent
 536  *              jointree item.  These are the components to be joined together.
 537  *              Note that levels_needed == list_length(initial_rels).
 538  *
 539  * Returns the final level of join relations, i.e., the relation that is
 540  * the result of joining all the original relations together.
 541  * At least one implementation path must be provided for this relation and
 542  * all required sub-relations.
 543  *
 544  * To support loadable plugins that modify planner behavior by changing the
 545  * join searching algorithm, we provide a hook variable that lets a plugin
 546  * replace or supplement this function.  Any such hook must return the same
 547  * final join relation as the standard code would, but it might have a
 548  * different set of implementation paths attached, and only the sub-joinrels
 549  * needed for these paths need have been instantiated.
 550  *
 551  * Note to plugin authors: the functions invoked during standard_join_search()
 552  * modify root->join_rel_list and root->join_rel_hash.  If you want to do more
 553  * than one join-order search, you'll probably need to save and restore the
 554  * original states of those data structures.  See geqo_eval() for an example.
 555  */
 556 RelOptInfo *
 557 standard_join_search(PlannerInfo *root, int levels_needed, List *initial_rels)
 558 {
 559         int                     lev;
 560         RelOptInfo *rel;
 561
 562         /*
 563          * This function cannot be invoked recursively within any one planning
 564          * problem, so join_rel_level[] can't be in use already.
 565          */
 566         Assert(root->join_rel_level == NULL);
 567
 568         /*
 569          * We employ a simple "dynamic programming" algorithm: we first find all
 570          * ways to build joins of two jointree items, then all ways to build joins
 571          * of three items (from two-item joins and single items), then four-item
 572          * joins, and so on until we have considered all ways to join all the
 573          * items into one rel.
 574          *
 575          * root->join_rel_level[j] is a list of all the j-item rels.  Initially we
 576          * set root->join_rel_level[1] to represent all the single-jointree-item
 577          * relations.
 578          */
 579         root->join_rel_level = (List **) palloc0((levels_needed + 1) * sizeof(List *));
 580
 581         root->join_rel_level[1] = initial_rels;
 582
 583         for (lev = 2; lev <= levels_needed; lev++)
 584         {
 585                 ListCell   *lc;
 586
 587                 /*
 588                  * Determine all possible pairs of relations to be joined at this
 589                  * level, and build paths for making each one from every available
 590                  * pair of lower-level relations.
 591                  */
 592                 join_search_one_level(root, lev);
 593
 594                 /*
 595                  * Run generate_gather_paths() for each just-processed joinrel.  We
 596                  * could not do this earlier because both regular and partial paths
 597                  * can get added to a particular joinrel at multiple times within
 598                  * join_search_one_level.  After that, we're done creating paths for
 599                  * the joinrel, so run set_cheapest().
 600                  */
 601                 foreach(lc, root->join_rel_level[lev])
 602                 {
 603                         rel = (RelOptInfo *) lfirst(lc);
 604
 605                         /* Create GatherPaths for any useful partial paths for rel */
 606                         generate_gather_paths(root, rel);
 607
 608                         /* Find and save the cheapest paths for this rel */
 609                         set_cheapest(rel);
 610
 611 #ifdef OPTIMIZER_DEBUG
 612                         debug_print_rel(root, rel);
 613 #endif
 614                 }
 615         }
 616
 617         /*
 618          * We should have a single rel at the final level.
 619          */
 620         if (root->join_rel_level[levels_needed] == NIL)
 621                 elog(ERROR, "failed to build any %d-way joins", levels_needed);
 622         Assert(list_length(root->join_rel_level[levels_needed]) == 1);
 623
 624         rel = (RelOptInfo *) linitial(root->join_rel_level[levels_needed]);
 625
 626         root->join_rel_level = NULL;
 627
 628         return rel;
 629 }
 630
 631 /*
 632  * create_plain_partial_paths
 633  *        Build partial access paths for parallel scan of a plain relation
 634  */
 635 static void
 636 create_plain_partial_paths(PlannerInfo *root, RelOptInfo *rel)
 637 {
 638         int                     parallel_workers;
 639
 640         /*
 641          * If the user has set the parallel_workers reloption, use that; otherwise
 642          * select a default number of workers.
 643          */
 644         if (rel->rel_parallel_workers != -1)
 645                 parallel_workers = rel->rel_parallel_workers;
 646         else
 647         {
 648                 int                     parallel_threshold;
 649
 650                 /*
 651                  * If this relation is too small to be worth a parallel scan, just
 652                  * return without doing anything ... unless it's an inheritance child.
 653                  * In that case, we want to generate a parallel path here anyway.  It
 654                  * might not be worthwhile just for this relation, but when combined
 655                  * with all of its inheritance siblings it may well pay off.
 656                  */
 657                 if (rel->pages < (BlockNumber) min_parallel_relation_size &&
 658                         rel->reloptkind == RELOPT_BASEREL)
 659                         return;
 660
 661                 /*
 662                  * Select the number of workers based on the log of the size of the
 663                  * relation.  This probably needs to be a good deal more
 664                  * sophisticated, but we need something here for now.  Note that the
 665                  * upper limit of the min_parallel_relation_size GUC is chosen to
 666                  * prevent overflow here.
 667                  */
 668                 parallel_workers = 1;
 669                 parallel_threshold = Max(min_parallel_relation_size, 1);
 670                 while (rel->pages >= (BlockNumber) (parallel_threshold * 3))
 671                 {
 672                         parallel_workers++;
 673                         parallel_threshold *= 3;
 674                         if (parallel_threshold > INT_MAX / 3)
 675                                 break;                  /* avoid overflow */
 676                 }
 677         }
 678
 679         /*
 680          * In no case use more than max_parallel_workers_per_gather workers.
 681          */
 682         parallel_workers = Min(parallel_workers, max_parallel_workers_per_gather);
 683
 684         /* If any limit was set to zero, the user doesn't want a parallel scan. */
 685         if (parallel_workers <= 0)
 686                 return;
 687
 688         /* Add an unordered partial path based on a parallel sequential scan. */
 689         add_partial_path(rel, create_seqscan_path(root, rel, NULL, parallel_workers));
 690 }
 691
 692 /*
 693  * join_search_one_level
 694  *        Consider ways to produce join relations containing exactly 'level'
 695  *        jointree items.  (This is one step of the dynamic-programming method
 696  *        embodied in standard_join_search.)  Join rel nodes for each feasible
 697  *        combination of lower-level rels are created and returned in a list.
 698  *        Implementation paths are created for each such joinrel, too.
 699  *
 700  * level: level of rels we want to make this time
 701  * root->join_rel_level[j], 1 <= j < level, is a list of rels containing j items
 702  *
 703  * The result is returned in root->join_rel_level[level].
 704  */
 705 void
 706 join_search_one_level(PlannerInfo *root, int level)
 707 {
 708         List      **joinrels = root->join_rel_level;
 709         ListCell   *r;
 710         int                     k;
 711
 712         Assert(joinrels[level] == NIL);
 713
 714         /* Set join_cur_level so that new joinrels are added to proper list */
 715         root->join_cur_level = level;
 716
 717         /*
 718          * First, consider left-sided and right-sided plans, in which rels of
 719          * exactly level-1 member relations are joined against initial relations.
 720          * We prefer to join using join clauses, but if we find a rel of level-1
 721          * members that has no join clauses, we will generate Cartesian-product
 722          * joins against all initial rels not already contained in it.
 723          */
 724         foreach(r, joinrels[level - 1])
 725         {
 726                 RelOptInfo *old_rel = (RelOptInfo *) lfirst(r);
 727
 728                 if (old_rel->joininfo != NIL || old_rel->has_eclass_joins ||
 729                         has_join_restriction(root, old_rel))
 730                 {
 731                         /*
 732                          * There are join clauses or join order restrictions relevant to
 733                          * this rel, so consider joins between this rel and (only) those
 734                          * initial rels it is linked to by a clause or restriction.
 735                          *
 736                          * At level 2 this condition is symmetric, so there is no need to
 737                          * look at initial rels before this one in the list; we already
 738                          * considered such joins when we were at the earlier rel.  (The
 739                          * mirror-image joins are handled automatically by make_join_rel.)
 740                          * In later passes (level > 2), we join rels of the previous level
 741                          * to each initial rel they don't already include but have a join
 742                          * clause or restriction with.
 743                          */
 744                         ListCell   *other_rels;
 745
 746                         if (level == 2)         /* consider remaining initial rels */
 747                                 other_rels = lnext(r);
 748                         else    /* consider all initial rels */
 749                                 other_rels = list_head(joinrels[1]);
 750
 751                         make_rels_by_clause_joins(root,
 752                                                                           old_rel,
 753                                                                           other_rels);
 754                 }
 755                 else
 756                 {
 757                         /*
 758                          * Oops, we have a relation that is not joined to any other
 759                          * relation, either directly or by join-order restrictions.
 760                          * Cartesian product time.
 761                          *
 762                          * We consider a cartesian product with each not-already-included
 763                          * initial rel, whether it has other join clauses or not.  At
 764                          * level 2, if there are two or more clauseless initial rels, we
 765                          * will redundantly consider joining them in both directions; but
 766                          * such cases aren't common enough to justify adding complexity to
 767                          * avoid the duplicated effort.
 768                          */
 769                         make_rels_by_clauseless_joins(root,
 770                                                                                   old_rel,
 771                                                                                   list_head(joinrels[1]));
 772                 }
 773         }
 774
 775         /*
 776          * Now, consider "bushy plans" in which relations of k initial rels are
 777          * joined to relations of level-k initial rels, for 2 <= k <= level-2.
 778          *
 779          * We only consider bushy-plan joins for pairs of rels where there is a
 780          * suitable join clause (or join order restriction), in order to avoid
 781          * unreasonable growth of planning time.
 782          */
 783         for (k = 2;; k++)
 784         {
 785                 int                     other_level = level - k;
 786
 787                 /*
 788                  * Since make_join_rel(x, y) handles both x,y and y,x cases, we only
 789                  * need to go as far as the halfway point.
 790                  */
 791                 if (k > other_level)
 792                         break;
 793
 794                 foreach(r, joinrels[k])
 795                 {
 796                         RelOptInfo *old_rel = (RelOptInfo *) lfirst(r);
 797                         ListCell   *other_rels;
 798                         ListCell   *r2;
 799
 800                         /*
 801                          * We can ignore relations without join clauses here, unless they
 802                          * participate in join-order restrictions --- then we might have
 803                          * to force a bushy join plan.
 804                          */
 805                         if (old_rel->joininfo == NIL && !old_rel->has_eclass_joins &&
 806                                 !has_join_restriction(root, old_rel))
 807                                 continue;
 808
 809                         if (k == other_level)
 810                                 other_rels = lnext(r);  /* only consider remaining rels */
 811                         else
 812                                 other_rels = list_head(joinrels[other_level]);
 813
 814                         for_each_cell(r2, other_rels)
 815                         {
 816                                 RelOptInfo *new_rel = (RelOptInfo *) lfirst(r2);
 817
 818                                 if (!bms_overlap(old_rel->relids, new_rel->relids))
 819                                 {
 820                                         /*
 821                                          * OK, we can build a rel of the right level from this
 822                                          * pair of rels.  Do so if there is at least one relevant
 823                                          * join clause or join order restriction.
 824                                          */
 825                                         if (have_relevant_joinclause(root, old_rel, new_rel) ||
 826                                                 have_join_order_restriction(root, old_rel, new_rel))
 827                                         {
 828                                                 (void) make_join_rel(root, old_rel, new_rel);
 829                                         }
 830                                 }
 831                         }
 832                 }
 833         }
 834
 835         /*----------
 836          * Last-ditch effort: if we failed to find any usable joins so far, force
 837          * a set of cartesian-product joins to be generated.  This handles the
 838          * special case where all the available rels have join clauses but we
 839          * cannot use any of those clauses yet.  This can only happen when we are
 840          * considering a join sub-problem (a sub-joinlist) and all the rels in the
 841          * sub-problem have only join clauses with rels outside the sub-problem.
 842          * An example is
 843          *
 844          *              SELECT ... FROM a INNER JOIN b ON TRUE, c, d, ...
 845          *              WHERE a.w = c.x and b.y = d.z;
 846          *
 847          * If the "a INNER JOIN b" sub-problem does not get flattened into the
 848          * upper level, we must be willing to make a cartesian join of a and b;
 849          * but the code above will not have done so, because it thought that both
 850          * a and b have joinclauses.  We consider only left-sided and right-sided
 851          * cartesian joins in this case (no bushy).
 852          *----------
 853          */
 854         if (joinrels[level] == NIL)
 855         {
 856                 /*
 857                  * This loop is just like the first one, except we always call
 858                  * make_rels_by_clauseless_joins().
 859                  */
 860                 foreach(r, joinrels[level - 1])
 861                 {
 862                         RelOptInfo *old_rel = (RelOptInfo *) lfirst(r);
 863
 864                         make_rels_by_clauseless_joins(root,
 865                                                                                   old_rel,
 866                                                                                   list_head(joinrels[1]));
 867                 }
 868
 869                 /*----------
 870                  * When special joins are involved, there may be no legal way
 871                  * to make an N-way join for some values of N.  For example consider
 872                  *
 873                  * SELECT ... FROM t1 WHERE
 874                  *       x IN (SELECT ... FROM t2,t3 WHERE ...) AND
 875                  *       y IN (SELECT ... FROM t4,t5 WHERE ...)
 876                  *
 877                  * We will flatten this query to a 5-way join problem, but there are
 878                  * no 4-way joins that join_is_legal() will consider legal.  We have
 879                  * to accept failure at level 4 and go on to discover a workable
 880                  * bushy plan at level 5.
 881                  *
 882                  * However, if there are no special joins and no lateral references
 883                  * then join_is_legal() should never fail, and so the following sanity
 884                  * check is useful.
 885                  *----------
 886                  */
 887                 if (joinrels[level] == NIL &&
 888                         root->join_info_list == NIL &&
 889                         !root->hasLateralRTEs)
 890                         elog(ERROR, "failed to build any %d-way joins", level);
 891         }
 892 }
 893
 894 /*
 895  * make_rels_by_clause_joins
 896  *        Build joins between the given relation 'old_rel' and other relations
 897  *        that participate in join clauses that 'old_rel' also participates in
 898  *        (or participate in join-order restrictions with it).
 899  *        The join rels are returned in root->join_rel_level[join_cur_level].
 900  *
 901  * Note: at levels above 2 we will generate the same joined relation in
 902  * multiple ways --- for example (a join b) join c is the same RelOptInfo as
 903  * (b join c) join a, though the second case will add a different set of Paths
 904  * to it.  This is the reason for using the join_rel_level mechanism, which
 905  * automatically ensures that each new joinrel is only added to the list once.
 906  *
 907  * 'old_rel' is the relation entry for the relation to be joined
 908  * 'other_rels': the first cell in a linked list containing the other
 909  * rels to be considered for joining
 910  *
 911  * Currently, this is only used with initial rels in other_rels, but it
 912  * will work for joining to joinrels too.
 913  */
 914 static void
 915 make_rels_by_clause_joins(PlannerInfo *root,
 916                                                   RelOptInfo *old_rel,
 917                                                   ListCell *other_rels)
 918 {
 919         ListCell   *l;
 920
 921         for_each_cell(l, other_rels)
 922         {
 923                 RelOptInfo *other_rel = (RelOptInfo *) lfirst(l);
 924
 925                 if (!bms_overlap(old_rel->relids, other_rel->relids) &&
 926                         (have_relevant_joinclause(root, old_rel, other_rel) ||
 927                          have_join_order_restriction(root, old_rel, other_rel)))
 928                 {
 929                         (void) make_join_rel(root, old_rel, other_rel);
 930                 }
 931         }
 932 }
 933
 934 /*
 935  * make_rels_by_clauseless_joins
 936  *        Given a relation 'old_rel' and a list of other relations
 937  *        'other_rels', create a join relation between 'old_rel' and each
 938  *        member of 'other_rels' that isn't already included in 'old_rel'.
 939  *        The join rels are returned in root->join_rel_level[join_cur_level].
 940  *
 941  * 'old_rel' is the relation entry for the relation to be joined
 942  * 'other_rels': the first cell of a linked list containing the
 943  * other rels to be considered for joining
 944  *
 945  * Currently, this is only used with initial rels in other_rels, but it would
 946  * work for joining to joinrels too.
 947  */
 948 static void
 949 make_rels_by_clauseless_joins(PlannerInfo *root,
 950                                                           RelOptInfo *old_rel,
 951                                                           ListCell *other_rels)
 952 {
 953         ListCell   *l;
 954
 955         for_each_cell(l, other_rels)
 956         {
 957                 RelOptInfo *other_rel = (RelOptInfo *) lfirst(l);
 958
 959                 if (!bms_overlap(other_rel->relids, old_rel->relids))
 960                 {
 961                         (void) make_join_rel(root, old_rel, other_rel);
 962                 }
 963         }
 964 }
 965
 966 /*
 967  * join_is_legal
 968  *         Determine whether a proposed join is legal given the query's
 969  *         join order constraints; and if it is, determine the join type.
 970  *
 971  * Caller must supply not only the two rels, but the union of their relids.
 972  * (We could simplify the API by computing joinrelids locally, but this
 973  * would be redundant work in the normal path through make_join_rel.)
 974  *
 975  * On success, *sjinfo_p is set to NULL if this is to be a plain inner join,
 976  * else it's set to point to the associated SpecialJoinInfo node.  Also,
 977  * *reversed_p is set TRUE if the given relations need to be swapped to
 978  * match the SpecialJoinInfo node.
 979  */
 980 static bool
 981 join_is_legal(PlannerInfo *root, RelOptInfo *rel1, RelOptInfo *rel2,
 982                           Relids joinrelids,
 983                           SpecialJoinInfo **sjinfo_p, bool *reversed_p)
 984 {
 985         SpecialJoinInfo *match_sjinfo;
 986         bool            reversed;
 987         bool            unique_ified;
 988         bool            must_be_leftjoin;
 989         ListCell   *l;
 990
 991         /*
 992          * Ensure output params are set on failure return.  This is just to
 993          * suppress uninitialized-variable warnings from overly anal compilers.
 994          */
 995         *sjinfo_p = NULL;
 996         *reversed_p = false;
 997
 998         /*
 999          * If we have any special joins, the proposed join might be illegal; and
1000          * in any case we have to determine its join type.  Scan the join info
1001          * list for matches and conflicts.
1002          */
1003         match_sjinfo = NULL;
1004         reversed = false;
1005         unique_ified = false;
1006         must_be_leftjoin = false;
1007
1008         foreach(l, root->join_info_list)
1009         {
1010                 SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
1011
1012                 /*
1013                  * This special join is not relevant unless its RHS overlaps the
1014                  * proposed join.  (Check this first as a fast path for dismissing
1015                  * most irrelevant SJs quickly.)
1016                  */
1017                 if (!bms_overlap(sjinfo->min_righthand, joinrelids))
1018                         continue;
1019
1020                 /*
1021                  * Also, not relevant if proposed join is fully contained within RHS
1022                  * (ie, we're still building up the RHS).
1023                  */
1024                 if (bms_is_subset(joinrelids, sjinfo->min_righthand))
1025                         continue;
1026
1027                 /*
1028                  * Also, not relevant if SJ is already done within either input.
1029                  */
1030                 if (bms_is_subset(sjinfo->min_lefthand, rel1->relids) &&
1031                         bms_is_subset(sjinfo->min_righthand, rel1->relids))
1032                         continue;
1033                 if (bms_is_subset(sjinfo->min_lefthand, rel2->relids) &&
1034                         bms_is_subset(sjinfo->min_righthand, rel2->relids))
1035                         continue;
1036
1037                 /*
1038                  * If it's a semijoin and we already joined the RHS to any other rels
1039                  * within either input, then we must have unique-ified the RHS at that
1040                  * point (see below).  Therefore the semijoin is no longer relevant in
1041                  * this join path.
1042                  */
1043                 if (sjinfo->jointype == JOIN_SEMI)
1044                 {
1045                         if (bms_is_subset(sjinfo->syn_righthand, rel1->relids) &&
1046                                 !bms_equal(sjinfo->syn_righthand, rel1->relids))
1047                                 continue;
1048                         if (bms_is_subset(sjinfo->syn_righthand, rel2->relids) &&
1049                                 !bms_equal(sjinfo->syn_righthand, rel2->relids))
1050                                 continue;
1051                 }
1052
1053                 /*
1054                  * If one input contains min_lefthand and the other contains
1055                  * min_righthand, then we can perform the SJ at this join.
1056                  *
1057                  * Reject if we get matches to more than one SJ; that implies we're
1058                  * considering something that's not really valid.
1059                  */
1060                 if (bms_is_subset(sjinfo->min_lefthand, rel1->relids) &&
1061                         bms_is_subset(sjinfo->min_righthand, rel2->relids))
1062                 {
1063                         if (match_sjinfo)
1064                                 return false;   /* invalid join path */
1065                         match_sjinfo = sjinfo;
1066                         reversed = false;
1067                 }
1068                 else if (bms_is_subset(sjinfo->min_lefthand, rel2->relids) &&
1069                                  bms_is_subset(sjinfo->min_righthand, rel1->relids))
1070                 {
1071                         if (match_sjinfo)
1072                                 return false;   /* invalid join path */
1073                         match_sjinfo = sjinfo;
1074                         reversed = true;
1075                 }
1076                 else if (sjinfo->jointype == JOIN_SEMI &&
1077                                  bms_equal(sjinfo->syn_righthand, rel2->relids) &&
1078                                  create_unique_path(root, rel2, rel2->cheapest_total_path,
1079                                                                         sjinfo) != NULL)
1080                 {
1081                         /*----------
1082                          * For a semijoin, we can join the RHS to anything else by
1083                          * unique-ifying the RHS (if the RHS can be unique-ified).
1084                          * We will only get here if we have the full RHS but less
1085                          * than min_lefthand on the LHS.
1086                          *
1087                          * The reason to consider such a join path is exemplified by
1088                          *      SELECT ... FROM a,b WHERE (a.x,b.y) IN (SELECT c1,c2 FROM c)
1089                          * If we insist on doing this as a semijoin we will first have
1090                          * to form the cartesian product of A*B.  But if we unique-ify
1091                          * C then the semijoin becomes a plain innerjoin and we can join
1092                          * in any order, eg C to A and then to B.  When C is much smaller
1093                          * than A and B this can be a huge win.  So we allow C to be
1094                          * joined to just A or just B here, and then make_join_rel has
1095                          * to handle the case properly.
1096                          *
1097                          * Note that actually we'll allow unique-ified C to be joined to
1098                          * some other relation D here, too.  That is legal, if usually not
1099                          * very sane, and this routine is only concerned with legality not
1100                          * with whether the join is good strategy.
1101                          *----------
1102                          */
1103                         if (match_sjinfo)
1104                                 return false;   /* invalid join path */
1105                         match_sjinfo = sjinfo;
1106                         reversed = false;
1107                         unique_ified = true;
1108                 }
1109                 else if (sjinfo->jointype == JOIN_SEMI &&
1110                                  bms_equal(sjinfo->syn_righthand, rel1->relids) &&
1111                                  create_unique_path(root, rel1, rel1->cheapest_total_path,
1112                                                                         sjinfo) != NULL)
1113                 {
1114                         /* Reversed semijoin case */
1115                         if (match_sjinfo)
1116                                 return false;   /* invalid join path */
1117                         match_sjinfo = sjinfo;
1118                         reversed = true;
1119                         unique_ified = true;
1120                 }
1121                 else
1122                 {
1123                         /*
1124                          * Otherwise, the proposed join overlaps the RHS but isn't a valid
1125                          * implementation of this SJ.  But don't panic quite yet: the RHS
1126                          * violation might have occurred previously, in one or both input
1127                          * relations, in which case we must have previously decided that
1128                          * it was OK to commute some other SJ with this one.  If we need
1129                          * to perform this join to finish building up the RHS, rejecting
1130                          * it could lead to not finding any plan at all.  (This can occur
1131                          * because of the heuristics elsewhere in this file that postpone
1132                          * clauseless joins: we might not consider doing a clauseless join
1133                          * within the RHS until after we've performed other, validly
1134                          * commutable SJs with one or both sides of the clauseless join.)
1135                          * This consideration boils down to the rule that if both inputs
1136                          * overlap the RHS, we can allow the join --- they are either
1137                          * fully within the RHS, or represent previously-allowed joins to
1138                          * rels outside it.
1139                          */
1140                         if (bms_overlap(rel1->relids, sjinfo->min_righthand) &&
1141                                 bms_overlap(rel2->relids, sjinfo->min_righthand))
1142                                 continue;               /* assume valid previous violation of RHS */
1143
1144                         /*
1145                          * The proposed join could still be legal, but only if we're
1146                          * allowed to associate it into the RHS of this SJ.  That means
1147                          * this SJ must be a LEFT join (not SEMI or ANTI, and certainly
1148                          * not FULL) and the proposed join must not overlap the LHS.
1149                          */
1150                         if (sjinfo->jointype != JOIN_LEFT ||
1151                                 bms_overlap(joinrelids, sjinfo->min_lefthand))
1152                                 return false;   /* invalid join path */
1153
1154                         /*
1155                          * To be valid, the proposed join must be a LEFT join; otherwise
1156                          * it can't associate into this SJ's RHS.  But we may not yet have
1157                          * found the SpecialJoinInfo matching the proposed join, so we
1158                          * can't test that yet.  Remember the requirement for later.
1159                          */
1160                         must_be_leftjoin = true;
1161                 }
1162         }
1163
1164         /*
1165          * Fail if violated any SJ's RHS and didn't match to a LEFT SJ: the
1166          * proposed join can't associate into an SJ's RHS.
1167          *
1168          * Also, fail if the proposed join's predicate isn't strict; we're
1169          * essentially checking to see if we can apply outer-join identity 3, and
1170          * that's a requirement.  (This check may be redundant with checks in
1171          * make_outerjoininfo, but I'm not quite sure, and it's cheap to test.)
1172          */
1173         if (must_be_leftjoin &&
1174                 (match_sjinfo == NULL ||
1175                  match_sjinfo->jointype != JOIN_LEFT ||
1176                  !match_sjinfo->lhs_strict))
1177                 return false;                   /* invalid join path */
1178
1179         /*
1180          * We also have to check for constraints imposed by LATERAL references.
1181          */
1182         if (root->hasLateralRTEs)
1183         {
1184                 bool            lateral_fwd;
1185                 bool            lateral_rev;
1186                 Relids          join_lateral_rels;
1187
1188                 /*
1189                  * The proposed rels could each contain lateral references to the
1190                  * other, in which case the join is impossible.  If there are lateral
1191                  * references in just one direction, then the join has to be done with
1192                  * a nestloop with the lateral referencer on the inside.  If the join
1193                  * matches an SJ that cannot be implemented by such a nestloop, the
1194                  * join is impossible.
1195                  *
1196                  * Also, if the lateral reference is only indirect, we should reject
1197                  * the join; whatever rel(s) the reference chain goes through must be
1198                  * joined to first.
1199                  *
1200                  * Another case that might keep us from building a valid plan is the
1201                  * implementation restriction described by have_dangerous_phv().
1202                  */
1203                 lateral_fwd = bms_overlap(rel1->relids, rel2->lateral_relids);
1204                 lateral_rev = bms_overlap(rel2->relids, rel1->lateral_relids);
1205                 if (lateral_fwd && lateral_rev)
1206                         return false;           /* have lateral refs in both directions */
1207                 if (lateral_fwd)
1208                 {
1209                         /* has to be implemented as nestloop with rel1 on left */
1210                         if (match_sjinfo &&
1211                                 (reversed ||
1212                                  unique_ified ||
1213                                  match_sjinfo->jointype == JOIN_FULL))
1214                                 return false;   /* not implementable as nestloop */
1215                         /* check there is a direct reference from rel2 to rel1 */
1216                         if (!bms_overlap(rel1->relids, rel2->direct_lateral_relids))
1217                                 return false;   /* only indirect refs, so reject */
1218                         /* check we won't have a dangerous PHV */
1219                         if (have_dangerous_phv(root, rel1->relids, rel2->lateral_relids))
1220                                 return false;   /* might be unable to handle required PHV */
1221                 }
1222                 else if (lateral_rev)
1223                 {
1224                         /* has to be implemented as nestloop with rel2 on left */
1225                         if (match_sjinfo &&
1226                                 (!reversed ||
1227                                  unique_ified ||
1228                                  match_sjinfo->jointype == JOIN_FULL))
1229                                 return false;   /* not implementable as nestloop */
1230                         /* check there is a direct reference from rel1 to rel2 */
1231                         if (!bms_overlap(rel2->relids, rel1->direct_lateral_relids))
1232                                 return false;   /* only indirect refs, so reject */
1233                         /* check we won't have a dangerous PHV */
1234                         if (have_dangerous_phv(root, rel2->relids, rel1->lateral_relids))
1235                                 return false;   /* might be unable to handle required PHV */
1236                 }
1237
1238                 /*
1239                  * LATERAL references could also cause problems later on if we accept
1240                  * this join: if the join's minimum parameterization includes any rels
1241                  * that would have to be on the inside of an outer join with this join
1242                  * rel, then it's never going to be possible to build the complete
1243                  * query using this join.  We should reject this join not only because
1244                  * it'll save work, but because if we don't, the clauseless-join
1245                  * heuristics might think that legality of this join means that some
1246                  * other join rel need not be formed, and that could lead to failure
1247                  * to find any plan at all.  We have to consider not only rels that
1248                  * are directly on the inner side of an OJ with the joinrel, but also
1249                  * ones that are indirectly so, so search to find all such rels.
1250                  */
1251                 join_lateral_rels = min_join_parameterization(root, joinrelids,
1252                                                                                                           rel1, rel2);
1253                 if (join_lateral_rels)
1254                 {
1255                         Relids          join_plus_rhs = bms_copy(joinrelids);
1256                         bool            more;
1257
1258                         do
1259                         {
1260                                 more = false;
1261                                 foreach(l, root->join_info_list)
1262                                 {
1263                                         SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
1264
1265                                         if (bms_overlap(sjinfo->min_lefthand, join_plus_rhs) &&
1266                                                 !bms_is_subset(sjinfo->min_righthand, join_plus_rhs))
1267                                         {
1268                                                 join_plus_rhs = bms_add_members(join_plus_rhs,
1269                                                                                                           sjinfo->min_righthand);
1270                                                 more = true;
1271                                         }
1272                                         /* full joins constrain both sides symmetrically */
1273                                         if (sjinfo->jointype == JOIN_FULL &&
1274                                                 bms_overlap(sjinfo->min_righthand, join_plus_rhs) &&
1275                                                 !bms_is_subset(sjinfo->min_lefthand, join_plus_rhs))
1276                                         {
1277                                                 join_plus_rhs = bms_add_members(join_plus_rhs,
1278                                                                                                                 sjinfo->min_lefthand);
1279                                                 more = true;
1280                                         }
1281                                 }
1282                         } while (more);
1283                         if (bms_overlap(join_plus_rhs, join_lateral_rels))
1284                                 return false;   /* will not be able to join to some RHS rel */
1285                 }
1286         }
1287
1288         /* Otherwise, it's a valid join */
1289         *sjinfo_p = match_sjinfo;
1290         *reversed_p = reversed;
1291         return true;
1292 }
1293
1294 /*
1295  * has_join_restriction
1296  *              Detect whether the specified relation has join-order restrictions,
1297  *              due to being inside an outer join or an IN (sub-SELECT),
1298  *              or participating in any LATERAL references or multi-rel PHVs.
1299  *
1300  * Essentially, this tests whether have_join_order_restriction() could
1301  * succeed with this rel and some other one.  It's OK if we sometimes
1302  * say "true" incorrectly.  (Therefore, we don't bother with the relatively
1303  * expensive has_legal_joinclause test.)
1304  */
1305 static bool
1306 has_join_restriction(PlannerInfo *root, RelOptInfo *rel)
1307 {
1308         ListCell   *l;
1309
1310         if (rel->lateral_relids != NULL || rel->lateral_referencers != NULL)
1311                 return true;
1312
1313         foreach(l, root->placeholder_list)
1314         {
1315                 PlaceHolderInfo *phinfo = (PlaceHolderInfo *) lfirst(l);
1316
1317                 if (bms_is_subset(rel->relids, phinfo->ph_eval_at) &&
1318                         !bms_equal(rel->relids, phinfo->ph_eval_at))
1319                         return true;
1320         }
1321
1322         foreach(l, root->join_info_list)
1323         {
1324                 SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
1325
1326                 /* ignore full joins --- other mechanisms preserve their ordering */
1327                 if (sjinfo->jointype == JOIN_FULL)
1328                         continue;
1329
1330                 /* ignore if SJ is already contained in rel */
1331                 if (bms_is_subset(sjinfo->min_lefthand, rel->relids) &&
1332                         bms_is_subset(sjinfo->min_righthand, rel->relids))
1333                         continue;
1334
1335                 /* restricted if it overlaps LHS or RHS, but doesn't contain SJ */
1336                 if (bms_overlap(sjinfo->min_lefthand, rel->relids) ||
1337                         bms_overlap(sjinfo->min_righthand, rel->relids))
1338                         return true;
1339         }
1340
1341         return false;
1342 }
1343
1344 /*
1345  * is_dummy_rel --- has relation been proven empty?
1346  */
1347 static bool
1348 is_dummy_rel(RelOptInfo *rel)
1349 {
1350         return IS_DUMMY_REL(rel);
1351 }
1352
1353 /*
1354  * Mark a relation as proven empty.
1355  *
1356  * During GEQO planning, this can get invoked more than once on the same
1357  * baserel struct, so it's worth checking to see if the rel is already marked
1358  * dummy.
1359  *
1360  * Also, when called during GEQO join planning, we are in a short-lived
1361  * memory context.  We must make sure that the dummy path attached to a
1362  * baserel survives the GEQO cycle, else the baserel is trashed for future
1363  * GEQO cycles.  On the other hand, when we are marking a joinrel during GEQO,
1364  * we don't want the dummy path to clutter the main planning context.  Upshot
1365  * is that the best solution is to explicitly make the dummy path in the same
1366  * context the given RelOptInfo is in.
1367  */
1368 static void
1369 mark_dummy_rel(RelOptInfo *rel)
1370 {
1371         MemoryContext oldcontext;
1372
1373         /* Already marked? */
1374         if (is_dummy_rel(rel))
1375                 return;
1376
1377         /* No, so choose correct context to make the dummy path in */
1378         oldcontext = MemoryContextSwitchTo(GetMemoryChunkContext(rel));
1379
1380         /* Set dummy size estimate */
1381         rel->rows = 0;
1382
1383         /* Evict any previously chosen paths */
1384         rel->pathlist = NIL;
1385         rel->partial_pathlist = NIL;
1386
1387         /* Set up the dummy path */
1388         add_path(rel, (Path *) create_append_path(rel, NIL, NULL, 0));
1389
1390         /* Set or update cheapest_total_path and related fields */
1391         set_cheapest(rel);
1392
1393         MemoryContextSwitchTo(oldcontext);
1394 }
1395
1396 /*
1397  * restriction_is_constant_false --- is a restrictlist just FALSE?
1398  *
1399  * In cases where a qual is provably constant FALSE, eval_const_expressions
1400  * will generally have thrown away anything that's ANDed with it.  In outer
1401  * join situations this will leave us computing cartesian products only to
1402  * decide there's no match for an outer row, which is pretty stupid.  So,
1403  * we need to detect the case.
1404  *
1405  * If only_pushed_down is TRUE, then consider only pushed-down quals.
1406  */
1407 static bool
1408 restriction_is_constant_false(List *restrictlist, bool only_pushed_down)
1409 {
1410         ListCell   *lc;
1411
1412         /*
1413          * Despite the above comment, the restriction list we see here might
1414          * possibly have other members besides the FALSE constant, since other
1415          * quals could get "pushed down" to the outer join level.  So we check
1416          * each member of the list.
1417          */
1418         foreach(lc, restrictlist)
1419         {
1420                 RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
1421
1422                 Assert(IsA(rinfo, RestrictInfo));
1423                 if (only_pushed_down && !rinfo->is_pushed_down)
1424                         continue;
1425
1426                 if (rinfo->clause && IsA(rinfo->clause, Const))
1427                 {
1428                         Const      *con = (Const *) rinfo->clause;
1429
1430                         /* constant NULL is as good as constant FALSE for our purposes */
1431                         if (con->constisnull)
1432                                 return true;
1433                         if (!DatumGetBool(con->constvalue))
1434                                 return true;
1435                 }
1436         }
1437         return false;
1438 }