core.c

   1 /*-------------------------------------------------------------------------
   2  *
   3  * core.c
   4  *        Routines copied from PostgreSQL core distribution.
   5  *
   6  * src/backend/optimizer/path/allpaths.c
   7  *     set_append_rel_pathlist()
   8  *     generate_mergeappend_paths()
   9  *     get_cheapest_parameterized_child_path()
  10  *     accumulate_append_subpath()
  11  *     standard_join_search()
  12  *
  13  * src/backend/optimizer/path/joinrels.c
  14  *     join_search_one_level()
  15  *     make_rels_by_clause_joins()
  16  *     make_rels_by_clauseless_joins()
  17  *     join_is_legal()
  18  *     has_join_restriction()
  19  *     is_dummy_rel()
  20  *     mark_dummy_rel()
  21  *     restriction_is_constant_false()
  22  *
  23  * Portions Copyright (c) 1996-2013, PostgreSQL Global Development Group
  24  * Portions Copyright (c) 1994, Regents of the University of California
  25  *
  26  *-------------------------------------------------------------------------
  27  */
  28
  29 /*
  30  * set_append_rel_pathlist
  31  *        Build access paths for an "append relation"
  32  */
  33 static void
  34 set_append_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
  35                                                 Index rti, RangeTblEntry *rte)
  36 {
  37         int                     parentRTindex = rti;
  38         List       *live_childrels = NIL;
  39         List       *subpaths = NIL;
  40         bool            subpaths_valid = true;
  41         List       *all_child_pathkeys = NIL;
  42         List       *all_child_outers = NIL;
  43         ListCell   *l;
  44
  45         /*
  46          * Generate access paths for each member relation, and remember the
  47          * cheapest path for each one.  Also, identify all pathkeys (orderings)
  48          * and parameterizations (required_outer sets) available for the member
  49          * relations.
  50          */
  51         foreach(l, root->append_rel_list)
  52         {
  53                 AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
  54                 int                     childRTindex;
  55                 RangeTblEntry *childRTE;
  56                 RelOptInfo *childrel;
  57                 ListCell   *lcp;
  58
  59                 /* append_rel_list contains all append rels; ignore others */
  60                 if (appinfo->parent_relid != parentRTindex)
  61                         continue;
  62
  63                 /* Re-locate the child RTE and RelOptInfo */
  64                 childRTindex = appinfo->child_relid;
  65                 childRTE = root->simple_rte_array[childRTindex];
  66                 childrel = root->simple_rel_array[childRTindex];
  67
  68                 /*
  69                  * Compute the child's access paths.
  70                  */
  71                 set_rel_pathlist(root, childrel, childRTindex, childRTE);
  72
  73                 /*
  74                  * If child is dummy, ignore it.
  75                  */
  76                 if (IS_DUMMY_REL(childrel))
  77                         continue;
  78
  79                 /*
  80                  * Child is live, so add it to the live_childrels list for use below.
  81                  */
  82                 live_childrels = lappend(live_childrels, childrel);
  83
  84                 /*
  85                  * If child has an unparameterized cheapest-total path, add that to
  86                  * the unparameterized Append path we are constructing for the parent.
  87                  * If not, there's no workable unparameterized path.
  88                  */
  89                 if (childrel->cheapest_total_path->param_info == NULL)
  90                         subpaths = accumulate_append_subpath(subpaths,
  91                                                                                           childrel->cheapest_total_path);
  92                 else
  93                         subpaths_valid = false;
  94
  95                 /*
  96                  * Collect lists of all the available path orderings and
  97                  * parameterizations for all the children.      We use these as a
  98                  * heuristic to indicate which sort orderings and parameterizations we
  99                  * should build Append and MergeAppend paths for.
 100                  */
 101                 foreach(lcp, childrel->pathlist)
 102                 {
 103                         Path       *childpath = (Path *) lfirst(lcp);
 104                         List       *childkeys = childpath->pathkeys;
 105                         Relids          childouter = PATH_REQ_OUTER(childpath);
 106
 107                         /* Unsorted paths don't contribute to pathkey list */
 108                         if (childkeys != NIL)
 109                         {
 110                                 ListCell   *lpk;
 111                                 bool            found = false;
 112
 113                                 /* Have we already seen this ordering? */
 114                                 foreach(lpk, all_child_pathkeys)
 115                                 {
 116                                         List       *existing_pathkeys = (List *) lfirst(lpk);
 117
 118                                         if (compare_pathkeys(existing_pathkeys,
 119                                                                                  childkeys) == PATHKEYS_EQUAL)
 120                                         {
 121                                                 found = true;
 122                                                 break;
 123                                         }
 124                                 }
 125                                 if (!found)
 126                                 {
 127                                         /* No, so add it to all_child_pathkeys */
 128                                         all_child_pathkeys = lappend(all_child_pathkeys,
 129                                                                                                  childkeys);
 130                                 }
 131                         }
 132
 133                         /* Unparameterized paths don't contribute to param-set list */
 134                         if (childouter)
 135                         {
 136                                 ListCell   *lco;
 137                                 bool            found = false;
 138
 139                                 /* Have we already seen this param set? */
 140                                 foreach(lco, all_child_outers)
 141                                 {
 142                                         Relids          existing_outers = (Relids) lfirst(lco);
 143
 144                                         if (bms_equal(existing_outers, childouter))
 145                                         {
 146                                                 found = true;
 147                                                 break;
 148                                         }
 149                                 }
 150                                 if (!found)
 151                                 {
 152                                         /* No, so add it to all_child_outers */
 153                                         all_child_outers = lappend(all_child_outers,
 154                                                                                            childouter);
 155                                 }
 156                         }
 157                 }
 158         }
 159
 160         /*
 161          * If we found unparameterized paths for all children, build an unordered,
 162          * unparameterized Append path for the rel.  (Note: this is correct even
 163          * if we have zero or one live subpath due to constraint exclusion.)
 164          */
 165         if (subpaths_valid)
 166                 add_path(rel, (Path *) create_append_path(rel, subpaths, NULL));
 167
 168         /*
 169          * Also build unparameterized MergeAppend paths based on the collected
 170          * list of child pathkeys.
 171          */
 172         if (subpaths_valid)
 173                 generate_mergeappend_paths(root, rel, live_childrels,
 174                                                                    all_child_pathkeys);
 175
 176         /*
 177          * Build Append paths for each parameterization seen among the child rels.
 178          * (This may look pretty expensive, but in most cases of practical
 179          * interest, the child rels will expose mostly the same parameterizations,
 180          * so that not that many cases actually get considered here.)
 181          *
 182          * The Append node itself cannot enforce quals, so all qual checking must
 183          * be done in the child paths.  This means that to have a parameterized
 184          * Append path, we must have the exact same parameterization for each
 185          * child path; otherwise some children might be failing to check the
 186          * moved-down quals.  To make them match up, we can try to increase the
 187          * parameterization of lesser-parameterized paths.
 188          */
 189         foreach(l, all_child_outers)
 190         {
 191                 Relids          required_outer = (Relids) lfirst(l);
 192                 ListCell   *lcr;
 193
 194                 /* Select the child paths for an Append with this parameterization */
 195                 subpaths = NIL;
 196                 subpaths_valid = true;
 197                 foreach(lcr, live_childrels)
 198                 {
 199                         RelOptInfo *childrel = (RelOptInfo *) lfirst(lcr);
 200                         Path       *subpath;
 201
 202                         subpath = get_cheapest_parameterized_child_path(root,
 203                                                                                                                         childrel,
 204                                                                                                                         required_outer);
 205                         if (subpath == NULL)
 206                         {
 207                                 /* failed to make a suitable path for this child */
 208                                 subpaths_valid = false;
 209                                 break;
 210                         }
 211                         subpaths = accumulate_append_subpath(subpaths, subpath);
 212                 }
 213
 214                 if (subpaths_valid)
 215                         add_path(rel, (Path *)
 216                                          create_append_path(rel, subpaths, required_outer));
 217         }
 218
 219         /* Select cheapest paths */
 220         set_cheapest(rel);
 221 }
 222
 223 /*
 224  * generate_mergeappend_paths
 225  *              Generate MergeAppend paths for an append relation
 226  *
 227  * Generate a path for each ordering (pathkey list) appearing in
 228  * all_child_pathkeys.
 229  *
 230  * We consider both cheapest-startup and cheapest-total cases, ie, for each
 231  * interesting ordering, collect all the cheapest startup subpaths and all the
 232  * cheapest total paths, and build a MergeAppend path for each case.
 233  *
 234  * We don't currently generate any parameterized MergeAppend paths.  While
 235  * it would not take much more code here to do so, it's very unclear that it
 236  * is worth the planning cycles to investigate such paths: there's little
 237  * use for an ordered path on the inside of a nestloop.  In fact, it's likely
 238  * that the current coding of add_path would reject such paths out of hand,
 239  * because add_path gives no credit for sort ordering of parameterized paths,
 240  * and a parameterized MergeAppend is going to be more expensive than the
 241  * corresponding parameterized Append path.  If we ever try harder to support
 242  * parameterized mergejoin plans, it might be worth adding support for
 243  * parameterized MergeAppends to feed such joins.  (See notes in
 244  * optimizer/README for why that might not ever happen, though.)
 245  */
 246 static void
 247 generate_mergeappend_paths(PlannerInfo *root, RelOptInfo *rel,
 248                                                    List *live_childrels,
 249                                                    List *all_child_pathkeys)
 250 {
 251         ListCell   *lcp;
 252
 253         foreach(lcp, all_child_pathkeys)
 254         {
 255                 List       *pathkeys = (List *) lfirst(lcp);
 256                 List       *startup_subpaths = NIL;
 257                 List       *total_subpaths = NIL;
 258                 bool            startup_neq_total = false;
 259                 ListCell   *lcr;
 260
 261                 /* Select the child paths for this ordering... */
 262                 foreach(lcr, live_childrels)
 263                 {
 264                         RelOptInfo *childrel = (RelOptInfo *) lfirst(lcr);
 265                         Path       *cheapest_startup,
 266                                            *cheapest_total;
 267
 268                         /* Locate the right paths, if they are available. */
 269                         cheapest_startup =
 270                                 get_cheapest_path_for_pathkeys(childrel->pathlist,
 271                                                                                            pathkeys,
 272                                                                                            NULL,
 273                                                                                            STARTUP_COST);
 274                         cheapest_total =
 275                                 get_cheapest_path_for_pathkeys(childrel->pathlist,
 276                                                                                            pathkeys,
 277                                                                                            NULL,
 278                                                                                            TOTAL_COST);
 279
 280                         /*
 281                          * If we can't find any paths with the right order just use the
 282                          * cheapest-total path; we'll have to sort it later.
 283                          */
 284                         if (cheapest_startup == NULL || cheapest_total == NULL)
 285                         {
 286                                 cheapest_startup = cheapest_total =
 287                                         childrel->cheapest_total_path;
 288                                 /* Assert we do have an unparameterized path for this child */
 289                                 Assert(cheapest_total->param_info == NULL);
 290                         }
 291
 292                         /*
 293                          * Notice whether we actually have different paths for the
 294                          * "cheapest" and "total" cases; frequently there will be no point
 295                          * in two create_merge_append_path() calls.
 296                          */
 297                         if (cheapest_startup != cheapest_total)
 298                                 startup_neq_total = true;
 299
 300                         startup_subpaths =
 301                                 accumulate_append_subpath(startup_subpaths, cheapest_startup);
 302                         total_subpaths =
 303                                 accumulate_append_subpath(total_subpaths, cheapest_total);
 304                 }
 305
 306                 /* ... and build the MergeAppend paths */
 307                 add_path(rel, (Path *) create_merge_append_path(root,
 308                                                                                                                 rel,
 309                                                                                                                 startup_subpaths,
 310                                                                                                                 pathkeys,
 311                                                                                                                 NULL));
 312                 if (startup_neq_total)
 313                         add_path(rel, (Path *) create_merge_append_path(root,
 314                                                                                                                         rel,
 315                                                                                                                         total_subpaths,
 316                                                                                                                         pathkeys,
 317                                                                                                                         NULL));
 318         }
 319 }
 320
 321 /*
 322  * get_cheapest_parameterized_child_path
 323  *              Get cheapest path for this relation that has exactly the requested
 324  *              parameterization.
 325  *
 326  * Returns NULL if unable to create such a path.
 327  */
 328 static Path *
 329 get_cheapest_parameterized_child_path(PlannerInfo *root, RelOptInfo *rel,
 330                                                                           Relids required_outer)
 331 {
 332         Path       *cheapest;
 333         ListCell   *lc;
 334
 335         /*
 336          * Look up the cheapest existing path with no more than the needed
 337          * parameterization.  If it has exactly the needed parameterization, we're
 338          * done.
 339          */
 340         cheapest = get_cheapest_path_for_pathkeys(rel->pathlist,
 341                                                                                           NIL,
 342                                                                                           required_outer,
 343                                                                                           TOTAL_COST);
 344         Assert(cheapest != NULL);
 345         if (bms_equal(PATH_REQ_OUTER(cheapest), required_outer))
 346                 return cheapest;
 347
 348         /*
 349          * Otherwise, we can "reparameterize" an existing path to match the given
 350          * parameterization, which effectively means pushing down additional
 351          * joinquals to be checked within the path's scan.  However, some existing
 352          * paths might check the available joinquals already while others don't;
 353          * therefore, it's not clear which existing path will be cheapest after
 354          * reparameterization.  We have to go through them all and find out.
 355          */
 356         cheapest = NULL;
 357         foreach(lc, rel->pathlist)
 358         {
 359                 Path       *path = (Path *) lfirst(lc);
 360
 361                 /* Can't use it if it needs more than requested parameterization */
 362                 if (!bms_is_subset(PATH_REQ_OUTER(path), required_outer))
 363                         continue;
 364
 365                 /*
 366                  * Reparameterization can only increase the path's cost, so if it's
 367                  * already more expensive than the current cheapest, forget it.
 368                  */
 369                 if (cheapest != NULL &&
 370                         compare_path_costs(cheapest, path, TOTAL_COST) <= 0)
 371                         continue;
 372
 373                 /* Reparameterize if needed, then recheck cost */
 374                 if (!bms_equal(PATH_REQ_OUTER(path), required_outer))
 375                 {
 376                         path = reparameterize_path(root, path, required_outer, 1.0);
 377                         if (path == NULL)
 378                                 continue;               /* failed to reparameterize this one */
 379                         Assert(bms_equal(PATH_REQ_OUTER(path), required_outer));
 380
 381                         if (cheapest != NULL &&
 382                                 compare_path_costs(cheapest, path, TOTAL_COST) <= 0)
 383                                 continue;
 384                 }
 385
 386                 /* We have a new best path */
 387                 cheapest = path;
 388         }
 389
 390         /* Return the best path, or NULL if we found no suitable candidate */
 391         return cheapest;
 392 }
 393
 394 /*
 395  * accumulate_append_subpath
 396  *              Add a subpath to the list being built for an Append or MergeAppend
 397  *
 398  * It's possible that the child is itself an Append path, in which case
 399  * we can "cut out the middleman" and just add its child paths to our
 400  * own list.  (We don't try to do this earlier because we need to
 401  * apply both levels of transformation to the quals.)
 402  */
 403 static List *
 404 accumulate_append_subpath(List *subpaths, Path *path)
 405 {
 406         if (IsA(path, AppendPath))
 407         {
 408                 AppendPath *apath = (AppendPath *) path;
 409
 410                 /* list_copy is important here to avoid sharing list substructure */
 411                 return list_concat(subpaths, list_copy(apath->subpaths));
 412         }
 413         else
 414                 return lappend(subpaths, path);
 415 }
 416
 417 /*
 418  * standard_join_search
 419  *        Find possible joinpaths for a query by successively finding ways
 420  *        to join component relations into join relations.
 421  *
 422  * 'levels_needed' is the number of iterations needed, ie, the number of
 423  *              independent jointree items in the query.  This is > 1.
 424  *
 425  * 'initial_rels' is a list of RelOptInfo nodes for each independent
 426  *              jointree item.  These are the components to be joined together.
 427  *              Note that levels_needed == list_length(initial_rels).
 428  *
 429  * Returns the final level of join relations, i.e., the relation that is
 430  * the result of joining all the original relations together.
 431  * At least one implementation path must be provided for this relation and
 432  * all required sub-relations.
 433  *
 434  * To support loadable plugins that modify planner behavior by changing the
 435  * join searching algorithm, we provide a hook variable that lets a plugin
 436  * replace or supplement this function.  Any such hook must return the same
 437  * final join relation as the standard code would, but it might have a
 438  * different set of implementation paths attached, and only the sub-joinrels
 439  * needed for these paths need have been instantiated.
 440  *
 441  * Note to plugin authors: the functions invoked during standard_join_search()
 442  * modify root->join_rel_list and root->join_rel_hash.  If you want to do more
 443  * than one join-order search, you'll probably need to save and restore the
 444  * original states of those data structures.  See geqo_eval() for an example.
 445  */
 446 RelOptInfo *
 447 standard_join_search(PlannerInfo *root, int levels_needed, List *initial_rels)
 448 {
 449         int                     lev;
 450         RelOptInfo *rel;
 451
 452         /*
 453          * This function cannot be invoked recursively within any one planning
 454          * problem, so join_rel_level[] can't be in use already.
 455          */
 456         Assert(root->join_rel_level == NULL);
 457
 458         /*
 459          * We employ a simple "dynamic programming" algorithm: we first find all
 460          * ways to build joins of two jointree items, then all ways to build joins
 461          * of three items (from two-item joins and single items), then four-item
 462          * joins, and so on until we have considered all ways to join all the
 463          * items into one rel.
 464          *
 465          * root->join_rel_level[j] is a list of all the j-item rels.  Initially we
 466          * set root->join_rel_level[1] to represent all the single-jointree-item
 467          * relations.
 468          */
 469         root->join_rel_level = (List **) palloc0((levels_needed + 1) * sizeof(List *));
 470
 471         root->join_rel_level[1] = initial_rels;
 472
 473         for (lev = 2; lev <= levels_needed; lev++)
 474         {
 475                 ListCell   *lc;
 476
 477                 /*
 478                  * Determine all possible pairs of relations to be joined at this
 479                  * level, and build paths for making each one from every available
 480                  * pair of lower-level relations.
 481                  */
 482                 join_search_one_level(root, lev);
 483
 484                 /*
 485                  * Do cleanup work on each just-processed rel.
 486                  */
 487                 foreach(lc, root->join_rel_level[lev])
 488                 {
 489                         rel = (RelOptInfo *) lfirst(lc);
 490
 491                         /* Find and save the cheapest paths for this rel */
 492                         set_cheapest(rel);
 493
 494 #ifdef OPTIMIZER_DEBUG
 495                         debug_print_rel(root, rel);
 496 #endif
 497                 }
 498         }
 499
 500         /*
 501          * We should have a single rel at the final level.
 502          */
 503         if (root->join_rel_level[levels_needed] == NIL)
 504                 elog(ERROR, "failed to build any %d-way joins", levels_needed);
 505         Assert(list_length(root->join_rel_level[levels_needed]) == 1);
 506
 507         rel = (RelOptInfo *) linitial(root->join_rel_level[levels_needed]);
 508
 509         root->join_rel_level = NULL;
 510
 511         return rel;
 512 }
 513
 514 /*
 515  * join_search_one_level
 516  *        Consider ways to produce join relations containing exactly 'level'
 517  *        jointree items.  (This is one step of the dynamic-programming method
 518  *        embodied in standard_join_search.)  Join rel nodes for each feasible
 519  *        combination of lower-level rels are created and returned in a list.
 520  *        Implementation paths are created for each such joinrel, too.
 521  *
 522  * level: level of rels we want to make this time
 523  * root->join_rel_level[j], 1 <= j < level, is a list of rels containing j items
 524  *
 525  * The result is returned in root->join_rel_level[level].
 526  */
 527 void
 528 join_search_one_level(PlannerInfo *root, int level)
 529 {
 530         List      **joinrels = root->join_rel_level;
 531         ListCell   *r;
 532         int                     k;
 533
 534         Assert(joinrels[level] == NIL);
 535
 536         /* Set join_cur_level so that new joinrels are added to proper list */
 537         root->join_cur_level = level;
 538
 539         /*
 540          * First, consider left-sided and right-sided plans, in which rels of
 541          * exactly level-1 member relations are joined against initial relations.
 542          * We prefer to join using join clauses, but if we find a rel of level-1
 543          * members that has no join clauses, we will generate Cartesian-product
 544          * joins against all initial rels not already contained in it.
 545          */
 546         foreach(r, joinrels[level - 1])
 547         {
 548                 RelOptInfo *old_rel = (RelOptInfo *) lfirst(r);
 549
 550                 if (old_rel->joininfo != NIL || old_rel->has_eclass_joins ||
 551                         has_join_restriction(root, old_rel))
 552                 {
 553                         /*
 554                          * There are join clauses or join order restrictions relevant to
 555                          * this rel, so consider joins between this rel and (only) those
 556                          * initial rels it is linked to by a clause or restriction.
 557                          *
 558                          * At level 2 this condition is symmetric, so there is no need to
 559                          * look at initial rels before this one in the list; we already
 560                          * considered such joins when we were at the earlier rel.  (The
 561                          * mirror-image joins are handled automatically by make_join_rel.)
 562                          * In later passes (level > 2), we join rels of the previous level
 563                          * to each initial rel they don't already include but have a join
 564                          * clause or restriction with.
 565                          */
 566                         ListCell   *other_rels;
 567
 568                         if (level == 2)         /* consider remaining initial rels */
 569                                 other_rels = lnext(r);
 570                         else    /* consider all initial rels */
 571                                 other_rels = list_head(joinrels[1]);
 572
 573                         make_rels_by_clause_joins(root,
 574                                                                           old_rel,
 575                                                                           other_rels);
 576                 }
 577                 else
 578                 {
 579                         /*
 580                          * Oops, we have a relation that is not joined to any other
 581                          * relation, either directly or by join-order restrictions.
 582                          * Cartesian product time.
 583                          *
 584                          * We consider a cartesian product with each not-already-included
 585                          * initial rel, whether it has other join clauses or not.  At
 586                          * level 2, if there are two or more clauseless initial rels, we
 587                          * will redundantly consider joining them in both directions; but
 588                          * such cases aren't common enough to justify adding complexity to
 589                          * avoid the duplicated effort.
 590                          */
 591                         make_rels_by_clauseless_joins(root,
 592                                                                                   old_rel,
 593                                                                                   list_head(joinrels[1]));
 594                 }
 595         }
 596
 597         /*
 598          * Now, consider "bushy plans" in which relations of k initial rels are
 599          * joined to relations of level-k initial rels, for 2 <= k <= level-2.
 600          *
 601          * We only consider bushy-plan joins for pairs of rels where there is a
 602          * suitable join clause (or join order restriction), in order to avoid
 603          * unreasonable growth of planning time.
 604          */
 605         for (k = 2;; k++)
 606         {
 607                 int                     other_level = level - k;
 608
 609                 /*
 610                  * Since make_join_rel(x, y) handles both x,y and y,x cases, we only
 611                  * need to go as far as the halfway point.
 612                  */
 613                 if (k > other_level)
 614                         break;
 615
 616                 foreach(r, joinrels[k])
 617                 {
 618                         RelOptInfo *old_rel = (RelOptInfo *) lfirst(r);
 619                         ListCell   *other_rels;
 620                         ListCell   *r2;
 621
 622                         /*
 623                          * We can ignore relations without join clauses here, unless they
 624                          * participate in join-order restrictions --- then we might have
 625                          * to force a bushy join plan.
 626                          */
 627                         if (old_rel->joininfo == NIL && !old_rel->has_eclass_joins &&
 628                                 !has_join_restriction(root, old_rel))
 629                                 continue;
 630
 631                         if (k == other_level)
 632                                 other_rels = lnext(r);  /* only consider remaining rels */
 633                         else
 634                                 other_rels = list_head(joinrels[other_level]);
 635
 636                         for_each_cell(r2, other_rels)
 637                         {
 638                                 RelOptInfo *new_rel = (RelOptInfo *) lfirst(r2);
 639
 640                                 if (!bms_overlap(old_rel->relids, new_rel->relids))
 641                                 {
 642                                         /*
 643                                          * OK, we can build a rel of the right level from this
 644                                          * pair of rels.  Do so if there is at least one relevant
 645                                          * join clause or join order restriction.
 646                                          */
 647                                         if (have_relevant_joinclause(root, old_rel, new_rel) ||
 648                                                 have_join_order_restriction(root, old_rel, new_rel))
 649                                         {
 650                                                 (void) make_join_rel(root, old_rel, new_rel);
 651                                         }
 652                                 }
 653                         }
 654                 }
 655         }
 656
 657         /*----------
 658          * Last-ditch effort: if we failed to find any usable joins so far, force
 659          * a set of cartesian-product joins to be generated.  This handles the
 660          * special case where all the available rels have join clauses but we
 661          * cannot use any of those clauses yet.  This can only happen when we are
 662          * considering a join sub-problem (a sub-joinlist) and all the rels in the
 663          * sub-problem have only join clauses with rels outside the sub-problem.
 664          * An example is
 665          *
 666          *              SELECT ... FROM a INNER JOIN b ON TRUE, c, d, ...
 667          *              WHERE a.w = c.x and b.y = d.z;
 668          *
 669          * If the "a INNER JOIN b" sub-problem does not get flattened into the
 670          * upper level, we must be willing to make a cartesian join of a and b;
 671          * but the code above will not have done so, because it thought that both
 672          * a and b have joinclauses.  We consider only left-sided and right-sided
 673          * cartesian joins in this case (no bushy).
 674          *----------
 675          */
 676         if (joinrels[level] == NIL)
 677         {
 678                 /*
 679                  * This loop is just like the first one, except we always call
 680                  * make_rels_by_clauseless_joins().
 681                  */
 682                 foreach(r, joinrels[level - 1])
 683                 {
 684                         RelOptInfo *old_rel = (RelOptInfo *) lfirst(r);
 685
 686                         make_rels_by_clauseless_joins(root,
 687                                                                                   old_rel,
 688                                                                                   list_head(joinrels[1]));
 689                 }
 690
 691                 /*----------
 692                  * When special joins are involved, there may be no legal way
 693                  * to make an N-way join for some values of N.  For example consider
 694                  *
 695                  * SELECT ... FROM t1 WHERE
 696                  *       x IN (SELECT ... FROM t2,t3 WHERE ...) AND
 697                  *       y IN (SELECT ... FROM t4,t5 WHERE ...)
 698                  *
 699                  * We will flatten this query to a 5-way join problem, but there are
 700                  * no 4-way joins that join_is_legal() will consider legal.  We have
 701                  * to accept failure at level 4 and go on to discover a workable
 702                  * bushy plan at level 5.
 703                  *
 704                  * However, if there are no special joins and no lateral references
 705                  * then join_is_legal() should never fail, and so the following sanity
 706                  * check is useful.
 707                  *----------
 708                  */
 709                 if (joinrels[level] == NIL &&
 710                         root->join_info_list == NIL &&
 711                         root->lateral_info_list == NIL)
 712                         elog(ERROR, "failed to build any %d-way joins", level);
 713         }
 714 }
 715
 716 /*
 717  * make_rels_by_clause_joins
 718  *        Build joins between the given relation 'old_rel' and other relations
 719  *        that participate in join clauses that 'old_rel' also participates in
 720  *        (or participate in join-order restrictions with it).
 721  *        The join rels are returned in root->join_rel_level[join_cur_level].
 722  *
 723  * Note: at levels above 2 we will generate the same joined relation in
 724  * multiple ways --- for example (a join b) join c is the same RelOptInfo as
 725  * (b join c) join a, though the second case will add a different set of Paths
 726  * to it.  This is the reason for using the join_rel_level mechanism, which
 727  * automatically ensures that each new joinrel is only added to the list once.
 728  *
 729  * 'old_rel' is the relation entry for the relation to be joined
 730  * 'other_rels': the first cell in a linked list containing the other
 731  * rels to be considered for joining
 732  *
 733  * Currently, this is only used with initial rels in other_rels, but it
 734  * will work for joining to joinrels too.
 735  */
 736 static void
 737 make_rels_by_clause_joins(PlannerInfo *root,
 738                                                   RelOptInfo *old_rel,
 739                                                   ListCell *other_rels)
 740 {
 741         ListCell   *l;
 742
 743         for_each_cell(l, other_rels)
 744         {
 745                 RelOptInfo *other_rel = (RelOptInfo *) lfirst(l);
 746
 747                 if (!bms_overlap(old_rel->relids, other_rel->relids) &&
 748                         (have_relevant_joinclause(root, old_rel, other_rel) ||
 749                          have_join_order_restriction(root, old_rel, other_rel)))
 750                 {
 751                         (void) make_join_rel(root, old_rel, other_rel);
 752                 }
 753         }
 754 }
 755
 756 /*
 757  * make_rels_by_clauseless_joins
 758  *        Given a relation 'old_rel' and a list of other relations
 759  *        'other_rels', create a join relation between 'old_rel' and each
 760  *        member of 'other_rels' that isn't already included in 'old_rel'.
 761  *        The join rels are returned in root->join_rel_level[join_cur_level].
 762  *
 763  * 'old_rel' is the relation entry for the relation to be joined
 764  * 'other_rels': the first cell of a linked list containing the
 765  * other rels to be considered for joining
 766  *
 767  * Currently, this is only used with initial rels in other_rels, but it would
 768  * work for joining to joinrels too.
 769  */
 770 static void
 771 make_rels_by_clauseless_joins(PlannerInfo *root,
 772                                                           RelOptInfo *old_rel,
 773                                                           ListCell *other_rels)
 774 {
 775         ListCell   *l;
 776
 777         for_each_cell(l, other_rels)
 778         {
 779                 RelOptInfo *other_rel = (RelOptInfo *) lfirst(l);
 780
 781                 if (!bms_overlap(other_rel->relids, old_rel->relids))
 782                 {
 783                         (void) make_join_rel(root, old_rel, other_rel);
 784                 }
 785         }
 786 }
 787
 788 /*
 789  * join_is_legal
 790  *         Determine whether a proposed join is legal given the query's
 791  *         join order constraints; and if it is, determine the join type.
 792  *
 793  * Caller must supply not only the two rels, but the union of their relids.
 794  * (We could simplify the API by computing joinrelids locally, but this
 795  * would be redundant work in the normal path through make_join_rel.)
 796  *
 797  * On success, *sjinfo_p is set to NULL if this is to be a plain inner join,
 798  * else it's set to point to the associated SpecialJoinInfo node.  Also,
 799  * *reversed_p is set TRUE if the given relations need to be swapped to
 800  * match the SpecialJoinInfo node.
 801  */
 802 static bool
 803 join_is_legal(PlannerInfo *root, RelOptInfo *rel1, RelOptInfo *rel2,
 804                           Relids joinrelids,
 805                           SpecialJoinInfo **sjinfo_p, bool *reversed_p)
 806 {
 807         SpecialJoinInfo *match_sjinfo;
 808         bool            reversed;
 809         bool            unique_ified;
 810         bool            is_valid_inner;
 811         bool            lateral_fwd;
 812         bool            lateral_rev;
 813         ListCell   *l;
 814
 815         /*
 816          * Ensure output params are set on failure return.      This is just to
 817          * suppress uninitialized-variable warnings from overly anal compilers.
 818          */
 819         *sjinfo_p = NULL;
 820         *reversed_p = false;
 821
 822         /*
 823          * If we have any special joins, the proposed join might be illegal; and
 824          * in any case we have to determine its join type.      Scan the join info
 825          * list for conflicts.
 826          */
 827         match_sjinfo = NULL;
 828         reversed = false;
 829         unique_ified = false;
 830         is_valid_inner = true;
 831
 832         foreach(l, root->join_info_list)
 833         {
 834                 SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
 835
 836                 /*
 837                  * This special join is not relevant unless its RHS overlaps the
 838                  * proposed join.  (Check this first as a fast path for dismissing
 839                  * most irrelevant SJs quickly.)
 840                  */
 841                 if (!bms_overlap(sjinfo->min_righthand, joinrelids))
 842                         continue;
 843
 844                 /*
 845                  * Also, not relevant if proposed join is fully contained within RHS
 846                  * (ie, we're still building up the RHS).
 847                  */
 848                 if (bms_is_subset(joinrelids, sjinfo->min_righthand))
 849                         continue;
 850
 851                 /*
 852                  * Also, not relevant if SJ is already done within either input.
 853                  */
 854                 if (bms_is_subset(sjinfo->min_lefthand, rel1->relids) &&
 855                         bms_is_subset(sjinfo->min_righthand, rel1->relids))
 856                         continue;
 857                 if (bms_is_subset(sjinfo->min_lefthand, rel2->relids) &&
 858                         bms_is_subset(sjinfo->min_righthand, rel2->relids))
 859                         continue;
 860
 861                 /*
 862                  * If it's a semijoin and we already joined the RHS to any other rels
 863                  * within either input, then we must have unique-ified the RHS at that
 864                  * point (see below).  Therefore the semijoin is no longer relevant in
 865                  * this join path.
 866                  */
 867                 if (sjinfo->jointype == JOIN_SEMI)
 868                 {
 869                         if (bms_is_subset(sjinfo->syn_righthand, rel1->relids) &&
 870                                 !bms_equal(sjinfo->syn_righthand, rel1->relids))
 871                                 continue;
 872                         if (bms_is_subset(sjinfo->syn_righthand, rel2->relids) &&
 873                                 !bms_equal(sjinfo->syn_righthand, rel2->relids))
 874                                 continue;
 875                 }
 876
 877                 /*
 878                  * If one input contains min_lefthand and the other contains
 879                  * min_righthand, then we can perform the SJ at this join.
 880                  *
 881                  * Barf if we get matches to more than one SJ (is that possible?)
 882                  */
 883                 if (bms_is_subset(sjinfo->min_lefthand, rel1->relids) &&
 884                         bms_is_subset(sjinfo->min_righthand, rel2->relids))
 885                 {
 886                         if (match_sjinfo)
 887                                 return false;   /* invalid join path */
 888                         match_sjinfo = sjinfo;
 889                         reversed = false;
 890                 }
 891                 else if (bms_is_subset(sjinfo->min_lefthand, rel2->relids) &&
 892                                  bms_is_subset(sjinfo->min_righthand, rel1->relids))
 893                 {
 894                         if (match_sjinfo)
 895                                 return false;   /* invalid join path */
 896                         match_sjinfo = sjinfo;
 897                         reversed = true;
 898                 }
 899                 else if (sjinfo->jointype == JOIN_SEMI &&
 900                                  bms_equal(sjinfo->syn_righthand, rel2->relids) &&
 901                                  create_unique_path(root, rel2, rel2->cheapest_total_path,
 902                                                                         sjinfo) != NULL)
 903                 {
 904                         /*----------
 905                          * For a semijoin, we can join the RHS to anything else by
 906                          * unique-ifying the RHS (if the RHS can be unique-ified).
 907                          * We will only get here if we have the full RHS but less
 908                          * than min_lefthand on the LHS.
 909                          *
 910                          * The reason to consider such a join path is exemplified by
 911                          *      SELECT ... FROM a,b WHERE (a.x,b.y) IN (SELECT c1,c2 FROM c)
 912                          * If we insist on doing this as a semijoin we will first have
 913                          * to form the cartesian product of A*B.  But if we unique-ify
 914                          * C then the semijoin becomes a plain innerjoin and we can join
 915                          * in any order, eg C to A and then to B.  When C is much smaller
 916                          * than A and B this can be a huge win.  So we allow C to be
 917                          * joined to just A or just B here, and then make_join_rel has
 918                          * to handle the case properly.
 919                          *
 920                          * Note that actually we'll allow unique-ified C to be joined to
 921                          * some other relation D here, too.  That is legal, if usually not
 922                          * very sane, and this routine is only concerned with legality not
 923                          * with whether the join is good strategy.
 924                          *----------
 925                          */
 926                         if (match_sjinfo)
 927                                 return false;   /* invalid join path */
 928                         match_sjinfo = sjinfo;
 929                         reversed = false;
 930                         unique_ified = true;
 931                 }
 932                 else if (sjinfo->jointype == JOIN_SEMI &&
 933                                  bms_equal(sjinfo->syn_righthand, rel1->relids) &&
 934                                  create_unique_path(root, rel1, rel1->cheapest_total_path,
 935                                                                         sjinfo) != NULL)
 936                 {
 937                         /* Reversed semijoin case */
 938                         if (match_sjinfo)
 939                                 return false;   /* invalid join path */
 940                         match_sjinfo = sjinfo;
 941                         reversed = true;
 942                         unique_ified = true;
 943                 }
 944                 else
 945                 {
 946                         /*----------
 947                          * Otherwise, the proposed join overlaps the RHS but isn't
 948                          * a valid implementation of this SJ.  It might still be
 949                          * a legal join, however.  If both inputs overlap the RHS,
 950                          * assume that it's OK.  Since the inputs presumably got past
 951                          * this function's checks previously, they can't overlap the
 952                          * LHS and their violations of the RHS boundary must represent
 953                          * SJs that have been determined to commute with this one.
 954                          * We have to allow this to work correctly in cases like
 955                          *              (a LEFT JOIN (b JOIN (c LEFT JOIN d)))
 956                          * when the c/d join has been determined to commute with the join
 957                          * to a, and hence d is not part of min_righthand for the upper
 958                          * join.  It should be legal to join b to c/d but this will appear
 959                          * as a violation of the upper join's RHS.
 960                          * Furthermore, if one input overlaps the RHS and the other does
 961                          * not, we should still allow the join if it is a valid
 962                          * implementation of some other SJ.  We have to allow this to
 963                          * support the associative identity
 964                          *              (a LJ b on Pab) LJ c ON Pbc = a LJ (b LJ c ON Pbc) on Pab
 965                          * since joining B directly to C violates the lower SJ's RHS.
 966                          * We assume that make_outerjoininfo() set things up correctly
 967                          * so that we'll only match to some SJ if the join is valid.
 968                          * Set flag here to check at bottom of loop.
 969                          *----------
 970                          */
 971                         if (sjinfo->jointype != JOIN_SEMI &&
 972                                 bms_overlap(rel1->relids, sjinfo->min_righthand) &&
 973                                 bms_overlap(rel2->relids, sjinfo->min_righthand))
 974                         {
 975                                 /* seems OK */
 976                                 Assert(!bms_overlap(joinrelids, sjinfo->min_lefthand));
 977                         }
 978                         else
 979                                 is_valid_inner = false;
 980                 }
 981         }
 982
 983         /*
 984          * Fail if violated some SJ's RHS and didn't match to another SJ. However,
 985          * "matching" to a semijoin we are implementing by unique-ification
 986          * doesn't count (think: it's really an inner join).
 987          */
 988         if (!is_valid_inner &&
 989                 (match_sjinfo == NULL || unique_ified))
 990                 return false;                   /* invalid join path */
 991
 992         /*
 993          * We also have to check for constraints imposed by LATERAL references.
 994          * The proposed rels could each contain lateral references to the other,
 995          * in which case the join is impossible.  If there are lateral references
 996          * in just one direction, then the join has to be done with a nestloop
 997          * with the lateral referencer on the inside.  If the join matches an SJ
 998          * that cannot be implemented by such a nestloop, the join is impossible.
 999          */
1000         lateral_fwd = lateral_rev = false;
1001         foreach(l, root->lateral_info_list)
1002         {
1003                 LateralJoinInfo *ljinfo = (LateralJoinInfo *) lfirst(l);
1004
1005                 if (bms_is_subset(ljinfo->lateral_rhs, rel2->relids) &&
1006                         bms_overlap(ljinfo->lateral_lhs, rel1->relids))
1007                 {
1008                         /* has to be implemented as nestloop with rel1 on left */
1009                         if (lateral_rev)
1010                                 return false;   /* have lateral refs in both directions */
1011                         lateral_fwd = true;
1012                         if (!bms_is_subset(ljinfo->lateral_lhs, rel1->relids))
1013                                 return false;   /* rel1 can't compute the required parameter */
1014                         if (match_sjinfo &&
1015                                 (reversed || match_sjinfo->jointype == JOIN_FULL))
1016                                 return false;   /* not implementable as nestloop */
1017                 }
1018                 if (bms_is_subset(ljinfo->lateral_rhs, rel1->relids) &&
1019                         bms_overlap(ljinfo->lateral_lhs, rel2->relids))
1020                 {
1021                         /* has to be implemented as nestloop with rel2 on left */
1022                         if (lateral_fwd)
1023                                 return false;   /* have lateral refs in both directions */
1024                         lateral_rev = true;
1025                         if (!bms_is_subset(ljinfo->lateral_lhs, rel2->relids))
1026                                 return false;   /* rel2 can't compute the required parameter */
1027                         if (match_sjinfo &&
1028                                 (!reversed || match_sjinfo->jointype == JOIN_FULL))
1029                                 return false;   /* not implementable as nestloop */
1030                 }
1031         }
1032
1033         /* Otherwise, it's a valid join */
1034         *sjinfo_p = match_sjinfo;
1035         *reversed_p = reversed;
1036         return true;
1037 }
1038
1039 /*
1040  * has_join_restriction
1041  *              Detect whether the specified relation has join-order restrictions,
1042  *              due to being inside an outer join or an IN (sub-SELECT),
1043  *              or participating in any LATERAL references.
1044  *
1045  * Essentially, this tests whether have_join_order_restriction() could
1046  * succeed with this rel and some other one.  It's OK if we sometimes
1047  * say "true" incorrectly.      (Therefore, we don't bother with the relatively
1048  * expensive has_legal_joinclause test.)
1049  */
1050 static bool
1051 has_join_restriction(PlannerInfo *root, RelOptInfo *rel)
1052 {
1053         ListCell   *l;
1054
1055         foreach(l, root->lateral_info_list)
1056         {
1057                 LateralJoinInfo *ljinfo = (LateralJoinInfo *) lfirst(l);
1058
1059                 if (bms_is_subset(ljinfo->lateral_rhs, rel->relids) ||
1060                         bms_overlap(ljinfo->lateral_lhs, rel->relids))
1061                         return true;
1062         }
1063
1064         foreach(l, root->join_info_list)
1065         {
1066                 SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
1067
1068                 /* ignore full joins --- other mechanisms preserve their ordering */
1069                 if (sjinfo->jointype == JOIN_FULL)
1070                         continue;
1071
1072                 /* ignore if SJ is already contained in rel */
1073                 if (bms_is_subset(sjinfo->min_lefthand, rel->relids) &&
1074                         bms_is_subset(sjinfo->min_righthand, rel->relids))
1075                         continue;
1076
1077                 /* restricted if it overlaps LHS or RHS, but doesn't contain SJ */
1078                 if (bms_overlap(sjinfo->min_lefthand, rel->relids) ||
1079                         bms_overlap(sjinfo->min_righthand, rel->relids))
1080                         return true;
1081         }
1082
1083         return false;
1084 }
1085
1086 /*
1087  * is_dummy_rel --- has relation been proven empty?
1088  */
1089 static bool
1090 is_dummy_rel(RelOptInfo *rel)
1091 {
1092         return IS_DUMMY_REL(rel);
1093 }
1094
1095 /*
1096  * Mark a relation as proven empty.
1097  *
1098  * During GEQO planning, this can get invoked more than once on the same
1099  * baserel struct, so it's worth checking to see if the rel is already marked
1100  * dummy.
1101  *
1102  * Also, when called during GEQO join planning, we are in a short-lived
1103  * memory context.      We must make sure that the dummy path attached to a
1104  * baserel survives the GEQO cycle, else the baserel is trashed for future
1105  * GEQO cycles.  On the other hand, when we are marking a joinrel during GEQO,
1106  * we don't want the dummy path to clutter the main planning context.  Upshot
1107  * is that the best solution is to explicitly make the dummy path in the same
1108  * context the given RelOptInfo is in.
1109  */
1110 static void
1111 mark_dummy_rel(RelOptInfo *rel)
1112 {
1113         MemoryContext oldcontext;
1114
1115         /* Already marked? */
1116         if (is_dummy_rel(rel))
1117                 return;
1118
1119         /* No, so choose correct context to make the dummy path in */
1120         oldcontext = MemoryContextSwitchTo(GetMemoryChunkContext(rel));
1121
1122         /* Set dummy size estimate */
1123         rel->rows = 0;
1124
1125         /* Evict any previously chosen paths */
1126         rel->pathlist = NIL;
1127
1128         /* Set up the dummy path */
1129         add_path(rel, (Path *) create_append_path(rel, NIL, NULL));
1130
1131         /* Set or update cheapest_total_path and related fields */
1132         set_cheapest(rel);
1133
1134         MemoryContextSwitchTo(oldcontext);
1135 }
1136
1137 /*
1138  * restriction_is_constant_false --- is a restrictlist just FALSE?
1139  *
1140  * In cases where a qual is provably constant FALSE, eval_const_expressions
1141  * will generally have thrown away anything that's ANDed with it.  In outer
1142  * join situations this will leave us computing cartesian products only to
1143  * decide there's no match for an outer row, which is pretty stupid.  So,
1144  * we need to detect the case.
1145  *
1146  * If only_pushed_down is TRUE, then consider only pushed-down quals.
1147  */
1148 static bool
1149 restriction_is_constant_false(List *restrictlist, bool only_pushed_down)
1150 {
1151         ListCell   *lc;
1152
1153         /*
1154          * Despite the above comment, the restriction list we see here might
1155          * possibly have other members besides the FALSE constant, since other
1156          * quals could get "pushed down" to the outer join level.  So we check
1157          * each member of the list.
1158          */
1159         foreach(lc, restrictlist)
1160         {
1161                 RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
1162
1163                 Assert(IsA(rinfo, RestrictInfo));
1164                 if (only_pushed_down && !rinfo->is_pushed_down)
1165                         continue;
1166
1167                 if (rinfo->clause && IsA(rinfo->clause, Const))
1168                 {
1169                         Const      *con = (Const *) rinfo->clause;
1170
1171                         /* constant NULL is as good as constant FALSE for our purposes */
1172                         if (con->constisnull)
1173                                 return true;
1174                         if (!DatumGetBool(con->constvalue))
1175                                 return true;
1176                 }
1177         }
1178         return false;
1179 }