core-9.2.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  *     accumulate_append_subpath()
  10  *     standard_join_search()
  11  *
  12  * src/backend/optimizer/path/joinrels.c
  13  *     join_search_one_level()
  14  *     make_rels_by_clause_joins()
  15  *     make_rels_by_clauseless_joins()
  16  *     join_is_legal()
  17  *     has_join_restriction()
  18  *     is_dummy_rel()
  19  *     mark_dummy_rel()
  20  *     restriction_is_constant_false()
  21  *
  22  * Portions Copyright (c) 1996-2012, PostgreSQL Global Development Group
  23  * Portions Copyright (c) 1994, Regents of the University of California
  24  *
  25  *-------------------------------------------------------------------------
  26  */
  27
  28 /*
  29  * set_append_rel_pathlist
  30  *        Build access paths for an "append relation"
  31  */
  32 static void
  33 set_append_rel_pathlist(PlannerInfo *root, RelOptInfo *rel,
  34                                                 Index rti, RangeTblEntry *rte)
  35 {
  36         int                     parentRTindex = rti;
  37         List       *live_childrels = NIL;
  38         List       *subpaths = NIL;
  39         List       *all_child_pathkeys = NIL;
  40         List       *all_child_outers = NIL;
  41         ListCell   *l;
  42
  43         /*
  44          * Generate access paths for each member relation, and remember the
  45          * cheapest path for each one.  Also, identify all pathkeys (orderings)
  46          * and parameterizations (required_outer sets) available for the member
  47          * relations.
  48          */
  49         foreach(l, root->append_rel_list)
  50         {
  51                 AppendRelInfo *appinfo = (AppendRelInfo *) lfirst(l);
  52                 int                     childRTindex;
  53                 RangeTblEntry *childRTE;
  54                 RelOptInfo *childrel;
  55                 ListCell   *lcp;
  56
  57                 /* append_rel_list contains all append rels; ignore others */
  58                 if (appinfo->parent_relid != parentRTindex)
  59                         continue;
  60
  61                 /* Re-locate the child RTE and RelOptInfo */
  62                 childRTindex = appinfo->child_relid;
  63                 childRTE = root->simple_rte_array[childRTindex];
  64                 childrel = root->simple_rel_array[childRTindex];
  65
  66                 /*
  67                  * Compute the child's access paths.
  68                  */
  69                 set_rel_pathlist(root, childrel, childRTindex, childRTE);
  70
  71                 /*
  72                  * If child is dummy, ignore it.
  73                  */
  74                 if (IS_DUMMY_REL(childrel))
  75                         continue;
  76
  77                 /*
  78                  * Child is live, so add its cheapest access path to the Append path
  79                  * we are constructing for the parent.
  80                  */
  81                 subpaths = accumulate_append_subpath(subpaths,
  82                                                                                          childrel->cheapest_total_path);
  83
  84                 /* Remember which childrels are live, for logic below */
  85                 live_childrels = lappend(live_childrels, childrel);
  86
  87                 /*
  88                  * Collect lists of all the available path orderings and
  89                  * parameterizations for all the children.      We use these as a
  90                  * heuristic to indicate which sort orderings and parameterizations we
  91                  * should build Append and MergeAppend paths for.
  92                  */
  93                 foreach(lcp, childrel->pathlist)
  94                 {
  95                         Path       *childpath = (Path *) lfirst(lcp);
  96                         List       *childkeys = childpath->pathkeys;
  97                         Relids          childouter = PATH_REQ_OUTER(childpath);
  98
  99                         /* Unsorted paths don't contribute to pathkey list */
 100                         if (childkeys != NIL)
 101                         {
 102                                 ListCell   *lpk;
 103                                 bool            found = false;
 104
 105                                 /* Have we already seen this ordering? */
 106                                 foreach(lpk, all_child_pathkeys)
 107                                 {
 108                                         List       *existing_pathkeys = (List *) lfirst(lpk);
 109
 110                                         if (compare_pathkeys(existing_pathkeys,
 111                                                                                  childkeys) == PATHKEYS_EQUAL)
 112                                         {
 113                                                 found = true;
 114                                                 break;
 115                                         }
 116                                 }
 117                                 if (!found)
 118                                 {
 119                                         /* No, so add it to all_child_pathkeys */
 120                                         all_child_pathkeys = lappend(all_child_pathkeys,
 121                                                                                                  childkeys);
 122                                 }
 123                         }
 124
 125                         /* Unparameterized paths don't contribute to param-set list */
 126                         if (childouter)
 127                         {
 128                                 ListCell   *lco;
 129                                 bool            found = false;
 130
 131                                 /* Have we already seen this param set? */
 132                                 foreach(lco, all_child_outers)
 133                                 {
 134                                         Relids          existing_outers = (Relids) lfirst(lco);
 135
 136                                         if (bms_equal(existing_outers, childouter))
 137                                         {
 138                                                 found = true;
 139                                                 break;
 140                                         }
 141                                 }
 142                                 if (!found)
 143                                 {
 144                                         /* No, so add it to all_child_outers */
 145                                         all_child_outers = lappend(all_child_outers,
 146                                                                                            childouter);
 147                                 }
 148                         }
 149                 }
 150         }
 151
 152         /*
 153          * Next, build an unordered, unparameterized Append path for the rel.
 154          * (Note: this is correct even if we have zero or one live subpath due to
 155          * constraint exclusion.)
 156          */
 157         add_path(rel, (Path *) create_append_path(rel, subpaths, NULL));
 158
 159         /*
 160          * Build unparameterized MergeAppend paths based on the collected list of
 161          * child pathkeys.
 162          */
 163         generate_mergeappend_paths(root, rel, live_childrels, all_child_pathkeys);
 164
 165         /*
 166          * Build Append paths for each parameterization seen among the child rels.
 167          * (This may look pretty expensive, but in most cases of practical
 168          * interest, the child rels will expose mostly the same parameterizations,
 169          * so that not that many cases actually get considered here.)
 170          *
 171          * The Append node itself cannot enforce quals, so all qual checking must
 172          * be done in the child paths.  This means that to have a parameterized
 173          * Append path, we must have the exact same parameterization for each
 174          * child path; otherwise some children might be failing to check the
 175          * moved-down quals.  To make them match up, we can try to increase the
 176          * parameterization of lesser-parameterized paths.
 177          */
 178         foreach(l, all_child_outers)
 179         {
 180                 Relids          required_outer = (Relids) lfirst(l);
 181                 bool            ok = true;
 182                 ListCell   *lcr;
 183
 184                 /* Select the child paths for an Append with this parameterization */
 185                 subpaths = NIL;
 186                 foreach(lcr, live_childrels)
 187                 {
 188                         RelOptInfo *childrel = (RelOptInfo *) lfirst(lcr);
 189                         Path       *cheapest_total;
 190
 191                         cheapest_total =
 192                                 get_cheapest_path_for_pathkeys(childrel->pathlist,
 193                                                                                            NIL,
 194                                                                                            required_outer,
 195                                                                                            TOTAL_COST);
 196                         Assert(cheapest_total != NULL);
 197
 198                         /* Children must have exactly the desired parameterization */
 199                         if (!bms_equal(PATH_REQ_OUTER(cheapest_total), required_outer))
 200                         {
 201                                 cheapest_total = reparameterize_path(root, cheapest_total,
 202                                                                                                          required_outer, 1.0);
 203                                 if (cheapest_total == NULL)
 204                                 {
 205                                         ok = false;
 206                                         break;
 207                                 }
 208                         }
 209
 210                         subpaths = accumulate_append_subpath(subpaths, cheapest_total);
 211                 }
 212
 213                 if (ok)
 214                         add_path(rel, (Path *)
 215                                          create_append_path(rel, subpaths, required_outer));
 216         }
 217
 218         /* Select cheapest paths */
 219         set_cheapest(rel);
 220 }
 221
 222 /*
 223  * generate_mergeappend_paths
 224  *              Generate MergeAppend paths for an append relation
 225  *
 226  * Generate a path for each ordering (pathkey list) appearing in
 227  * all_child_pathkeys.
 228  *
 229  * We consider both cheapest-startup and cheapest-total cases, ie, for each
 230  * interesting ordering, collect all the cheapest startup subpaths and all the
 231  * cheapest total paths, and build a MergeAppend path for each case.
 232  *
 233  * We don't currently generate any parameterized MergeAppend paths.  While
 234  * it would not take much more code here to do so, it's very unclear that it
 235  * is worth the planning cycles to investigate such paths: there's little
 236  * use for an ordered path on the inside of a nestloop.  In fact, it's likely
 237  * that the current coding of add_path would reject such paths out of hand,
 238  * because add_path gives no credit for sort ordering of parameterized paths,
 239  * and a parameterized MergeAppend is going to be more expensive than the
 240  * corresponding parameterized Append path.  If we ever try harder to support
 241  * parameterized mergejoin plans, it might be worth adding support for
 242  * parameterized MergeAppends to feed such joins.  (See notes in
 243  * optimizer/README for why that might not ever happen, though.)
 244  */
 245 static void
 246 generate_mergeappend_paths(PlannerInfo *root, RelOptInfo *rel,
 247                                                    List *live_childrels,
 248                                                    List *all_child_pathkeys)
 249 {
 250         ListCell   *lcp;
 251
 252         foreach(lcp, all_child_pathkeys)
 253         {
 254                 List       *pathkeys = (List *) lfirst(lcp);
 255                 List       *startup_subpaths = NIL;
 256                 List       *total_subpaths = NIL;
 257                 bool            startup_neq_total = false;
 258                 ListCell   *lcr;
 259
 260                 /* Select the child paths for this ordering... */
 261                 foreach(lcr, live_childrels)
 262                 {
 263                         RelOptInfo *childrel = (RelOptInfo *) lfirst(lcr);
 264                         Path       *cheapest_startup,
 265                                            *cheapest_total;
 266
 267                         /* Locate the right paths, if they are available. */
 268                         cheapest_startup =
 269                                 get_cheapest_path_for_pathkeys(childrel->pathlist,
 270                                                                                            pathkeys,
 271                                                                                            NULL,
 272                                                                                            STARTUP_COST);
 273                         cheapest_total =
 274                                 get_cheapest_path_for_pathkeys(childrel->pathlist,
 275                                                                                            pathkeys,
 276                                                                                            NULL,
 277                                                                                            TOTAL_COST);
 278
 279                         /*
 280                          * If we can't find any paths with the right order just use the
 281                          * cheapest-total path; we'll have to sort it later.
 282                          */
 283                         if (cheapest_startup == NULL || cheapest_total == NULL)
 284                         {
 285                                 cheapest_startup = cheapest_total =
 286                                         childrel->cheapest_total_path;
 287                                 Assert(cheapest_total != NULL);
 288                         }
 289
 290                         /*
 291                          * Notice whether we actually have different paths for the
 292                          * "cheapest" and "total" cases; frequently there will be no point
 293                          * in two create_merge_append_path() calls.
 294                          */
 295                         if (cheapest_startup != cheapest_total)
 296                                 startup_neq_total = true;
 297
 298                         startup_subpaths =
 299                                 accumulate_append_subpath(startup_subpaths, cheapest_startup);
 300                         total_subpaths =
 301                                 accumulate_append_subpath(total_subpaths, cheapest_total);
 302                 }
 303
 304                 /* ... and build the MergeAppend paths */
 305                 add_path(rel, (Path *) create_merge_append_path(root,
 306                                                                                                                 rel,
 307                                                                                                                 startup_subpaths,
 308                                                                                                                 pathkeys,
 309                                                                                                                 NULL));
 310                 if (startup_neq_total)
 311                         add_path(rel, (Path *) create_merge_append_path(root,
 312                                                                                                                         rel,
 313                                                                                                                         total_subpaths,
 314                                                                                                                         pathkeys,
 315                                                                                                                         NULL));
 316         }
 317 }
 318
 319 /*
 320  * accumulate_append_subpath
 321  *              Add a subpath to the list being built for an Append or MergeAppend
 322  *
 323  * It's possible that the child is itself an Append path, in which case
 324  * we can "cut out the middleman" and just add its child paths to our
 325  * own list.  (We don't try to do this earlier because we need to
 326  * apply both levels of transformation to the quals.)
 327  */
 328 static List *
 329 accumulate_append_subpath(List *subpaths, Path *path)
 330 {
 331         if (IsA(path, AppendPath))
 332         {
 333                 AppendPath *apath = (AppendPath *) path;
 334
 335                 /* list_copy is important here to avoid sharing list substructure */
 336                 return list_concat(subpaths, list_copy(apath->subpaths));
 337         }
 338         else
 339                 return lappend(subpaths, path);
 340 }
 341
 342 /*
 343  * standard_join_search
 344  *        Find possible joinpaths for a query by successively finding ways
 345  *        to join component relations into join relations.
 346  *
 347  * 'levels_needed' is the number of iterations needed, ie, the number of
 348  *              independent jointree items in the query.  This is > 1.
 349  *
 350  * 'initial_rels' is a list of RelOptInfo nodes for each independent
 351  *              jointree item.  These are the components to be joined together.
 352  *              Note that levels_needed == list_length(initial_rels).
 353  *
 354  * Returns the final level of join relations, i.e., the relation that is
 355  * the result of joining all the original relations together.
 356  * At least one implementation path must be provided for this relation and
 357  * all required sub-relations.
 358  *
 359  * To support loadable plugins that modify planner behavior by changing the
 360  * join searching algorithm, we provide a hook variable that lets a plugin
 361  * replace or supplement this function.  Any such hook must return the same
 362  * final join relation as the standard code would, but it might have a
 363  * different set of implementation paths attached, and only the sub-joinrels
 364  * needed for these paths need have been instantiated.
 365  *
 366  * Note to plugin authors: the functions invoked during standard_join_search()
 367  * modify root->join_rel_list and root->join_rel_hash.  If you want to do more
 368  * than one join-order search, you'll probably need to save and restore the
 369  * original states of those data structures.  See geqo_eval() for an example.
 370  */
 371 RelOptInfo *
 372 standard_join_search(PlannerInfo *root, int levels_needed, List *initial_rels)
 373 {
 374         int                     lev;
 375         RelOptInfo *rel;
 376
 377         /*
 378          * This function cannot be invoked recursively within any one planning
 379          * problem, so join_rel_level[] can't be in use already.
 380          */
 381         Assert(root->join_rel_level == NULL);
 382
 383         /*
 384          * We employ a simple "dynamic programming" algorithm: we first find all
 385          * ways to build joins of two jointree items, then all ways to build joins
 386          * of three items (from two-item joins and single items), then four-item
 387          * joins, and so on until we have considered all ways to join all the
 388          * items into one rel.
 389          *
 390          * root->join_rel_level[j] is a list of all the j-item rels.  Initially we
 391          * set root->join_rel_level[1] to represent all the single-jointree-item
 392          * relations.
 393          */
 394         root->join_rel_level = (List **) palloc0((levels_needed + 1) * sizeof(List *));
 395
 396         root->join_rel_level[1] = initial_rels;
 397
 398         for (lev = 2; lev <= levels_needed; lev++)
 399         {
 400                 ListCell   *lc;
 401
 402                 /*
 403                  * Determine all possible pairs of relations to be joined at this
 404                  * level, and build paths for making each one from every available
 405                  * pair of lower-level relations.
 406                  */
 407                 join_search_one_level(root, lev);
 408
 409                 /*
 410                  * Do cleanup work on each just-processed rel.
 411                  */
 412                 foreach(lc, root->join_rel_level[lev])
 413                 {
 414                         rel = (RelOptInfo *) lfirst(lc);
 415
 416                         /* Find and save the cheapest paths for this rel */
 417                         set_cheapest(rel);
 418
 419 #ifdef OPTIMIZER_DEBUG
 420                         debug_print_rel(root, rel);
 421 #endif
 422                 }
 423         }
 424
 425         /*
 426          * We should have a single rel at the final level.
 427          */
 428         if (root->join_rel_level[levels_needed] == NIL)
 429                 elog(ERROR, "failed to build any %d-way joins", levels_needed);
 430         Assert(list_length(root->join_rel_level[levels_needed]) == 1);
 431
 432         rel = (RelOptInfo *) linitial(root->join_rel_level[levels_needed]);
 433
 434         root->join_rel_level = NULL;
 435
 436         return rel;
 437 }
 438
 439 /*
 440  * join_search_one_level
 441  *        Consider ways to produce join relations containing exactly 'level'
 442  *        jointree items.  (This is one step of the dynamic-programming method
 443  *        embodied in standard_join_search.)  Join rel nodes for each feasible
 444  *        combination of lower-level rels are created and returned in a list.
 445  *        Implementation paths are created for each such joinrel, too.
 446  *
 447  * level: level of rels we want to make this time
 448  * root->join_rel_level[j], 1 <= j < level, is a list of rels containing j items
 449  *
 450  * The result is returned in root->join_rel_level[level].
 451  */
 452 void
 453 join_search_one_level(PlannerInfo *root, int level)
 454 {
 455         List      **joinrels = root->join_rel_level;
 456         ListCell   *r;
 457         int                     k;
 458
 459         Assert(joinrels[level] == NIL);
 460
 461         /* Set join_cur_level so that new joinrels are added to proper list */
 462         root->join_cur_level = level;
 463
 464         /*
 465          * First, consider left-sided and right-sided plans, in which rels of
 466          * exactly level-1 member relations are joined against initial relations.
 467          * We prefer to join using join clauses, but if we find a rel of level-1
 468          * members that has no join clauses, we will generate Cartesian-product
 469          * joins against all initial rels not already contained in it.
 470          */
 471         foreach(r, joinrels[level - 1])
 472         {
 473                 RelOptInfo *old_rel = (RelOptInfo *) lfirst(r);
 474
 475                 if (old_rel->joininfo != NIL || old_rel->has_eclass_joins ||
 476                         has_join_restriction(root, old_rel))
 477                 {
 478                         /*
 479                          * There are join clauses or join order restrictions relevant to
 480                          * this rel, so consider joins between this rel and (only) those
 481                          * initial rels it is linked to by a clause or restriction.
 482                          *
 483                          * At level 2 this condition is symmetric, so there is no need to
 484                          * look at initial rels before this one in the list; we already
 485                          * considered such joins when we were at the earlier rel.  (The
 486                          * mirror-image joins are handled automatically by make_join_rel.)
 487                          * In later passes (level > 2), we join rels of the previous level
 488                          * to each initial rel they don't already include but have a join
 489                          * clause or restriction with.
 490                          */
 491                         ListCell   *other_rels;
 492
 493                         if (level == 2)         /* consider remaining initial rels */
 494                                 other_rels = lnext(r);
 495                         else    /* consider all initial rels */
 496                                 other_rels = list_head(joinrels[1]);
 497
 498                         make_rels_by_clause_joins(root,
 499                                                                           old_rel,
 500                                                                           other_rels);
 501                 }
 502                 else
 503                 {
 504                         /*
 505                          * Oops, we have a relation that is not joined to any other
 506                          * relation, either directly or by join-order restrictions.
 507                          * Cartesian product time.
 508                          *
 509                          * We consider a cartesian product with each not-already-included
 510                          * initial rel, whether it has other join clauses or not.  At
 511                          * level 2, if there are two or more clauseless initial rels, we
 512                          * will redundantly consider joining them in both directions; but
 513                          * such cases aren't common enough to justify adding complexity to
 514                          * avoid the duplicated effort.
 515                          */
 516                         make_rels_by_clauseless_joins(root,
 517                                                                                   old_rel,
 518                                                                                   list_head(joinrels[1]));
 519                 }
 520         }
 521
 522         /*
 523          * Now, consider "bushy plans" in which relations of k initial rels are
 524          * joined to relations of level-k initial rels, for 2 <= k <= level-2.
 525          *
 526          * We only consider bushy-plan joins for pairs of rels where there is a
 527          * suitable join clause (or join order restriction), in order to avoid
 528          * unreasonable growth of planning time.
 529          */
 530         for (k = 2;; k++)
 531         {
 532                 int                     other_level = level - k;
 533
 534                 /*
 535                  * Since make_join_rel(x, y) handles both x,y and y,x cases, we only
 536                  * need to go as far as the halfway point.
 537                  */
 538                 if (k > other_level)
 539                         break;
 540
 541                 foreach(r, joinrels[k])
 542                 {
 543                         RelOptInfo *old_rel = (RelOptInfo *) lfirst(r);
 544                         ListCell   *other_rels;
 545                         ListCell   *r2;
 546
 547                         /*
 548                          * We can ignore relations without join clauses here, unless they
 549                          * participate in join-order restrictions --- then we might have
 550                          * to force a bushy join plan.
 551                          */
 552                         if (old_rel->joininfo == NIL && !old_rel->has_eclass_joins &&
 553                                 !has_join_restriction(root, old_rel))
 554                                 continue;
 555
 556                         if (k == other_level)
 557                                 other_rels = lnext(r);  /* only consider remaining rels */
 558                         else
 559                                 other_rels = list_head(joinrels[other_level]);
 560
 561                         for_each_cell(r2, other_rels)
 562                         {
 563                                 RelOptInfo *new_rel = (RelOptInfo *) lfirst(r2);
 564
 565                                 if (!bms_overlap(old_rel->relids, new_rel->relids))
 566                                 {
 567                                         /*
 568                                          * OK, we can build a rel of the right level from this
 569                                          * pair of rels.  Do so if there is at least one relevant
 570                                          * join clause or join order restriction.
 571                                          */
 572                                         if (have_relevant_joinclause(root, old_rel, new_rel) ||
 573                                                 have_join_order_restriction(root, old_rel, new_rel))
 574                                         {
 575                                                 (void) make_join_rel(root, old_rel, new_rel);
 576                                         }
 577                                 }
 578                         }
 579                 }
 580         }
 581
 582         /*----------
 583          * Last-ditch effort: if we failed to find any usable joins so far, force
 584          * a set of cartesian-product joins to be generated.  This handles the
 585          * special case where all the available rels have join clauses but we
 586          * cannot use any of those clauses yet.  This can only happen when we are
 587          * considering a join sub-problem (a sub-joinlist) and all the rels in the
 588          * sub-problem have only join clauses with rels outside the sub-problem.
 589          * An example is
 590          *
 591          *              SELECT ... FROM a INNER JOIN b ON TRUE, c, d, ...
 592          *              WHERE a.w = c.x and b.y = d.z;
 593          *
 594          * If the "a INNER JOIN b" sub-problem does not get flattened into the
 595          * upper level, we must be willing to make a cartesian join of a and b;
 596          * but the code above will not have done so, because it thought that both
 597          * a and b have joinclauses.  We consider only left-sided and right-sided
 598          * cartesian joins in this case (no bushy).
 599          *----------
 600          */
 601         if (joinrels[level] == NIL)
 602         {
 603                 /*
 604                  * This loop is just like the first one, except we always call
 605                  * make_rels_by_clauseless_joins().
 606                  */
 607                 foreach(r, joinrels[level - 1])
 608                 {
 609                         RelOptInfo *old_rel = (RelOptInfo *) lfirst(r);
 610
 611                         make_rels_by_clauseless_joins(root,
 612                                                                                   old_rel,
 613                                                                                   list_head(joinrels[1]));
 614                 }
 615
 616                 /*----------
 617                  * When special joins are involved, there may be no legal way
 618                  * to make an N-way join for some values of N.  For example consider
 619                  *
 620                  * SELECT ... FROM t1 WHERE
 621                  *       x IN (SELECT ... FROM t2,t3 WHERE ...) AND
 622                  *       y IN (SELECT ... FROM t4,t5 WHERE ...)
 623                  *
 624                  * We will flatten this query to a 5-way join problem, but there are
 625                  * no 4-way joins that join_is_legal() will consider legal.  We have
 626                  * to accept failure at level 4 and go on to discover a workable
 627                  * bushy plan at level 5.
 628                  *
 629                  * However, if there are no special joins then join_is_legal() should
 630                  * never fail, and so the following sanity check is useful.
 631                  *----------
 632                  */
 633                 if (joinrels[level] == NIL && root->join_info_list == NIL)
 634                         elog(ERROR, "failed to build any %d-way joins", level);
 635         }
 636 }
 637
 638 /*
 639  * make_rels_by_clause_joins
 640  *        Build joins between the given relation 'old_rel' and other relations
 641  *        that participate in join clauses that 'old_rel' also participates in
 642  *        (or participate in join-order restrictions with it).
 643  *        The join rels are returned in root->join_rel_level[join_cur_level].
 644  *
 645  * Note: at levels above 2 we will generate the same joined relation in
 646  * multiple ways --- for example (a join b) join c is the same RelOptInfo as
 647  * (b join c) join a, though the second case will add a different set of Paths
 648  * to it.  This is the reason for using the join_rel_level mechanism, which
 649  * automatically ensures that each new joinrel is only added to the list once.
 650  *
 651  * 'old_rel' is the relation entry for the relation to be joined
 652  * 'other_rels': the first cell in a linked list containing the other
 653  * rels to be considered for joining
 654  *
 655  * Currently, this is only used with initial rels in other_rels, but it
 656  * will work for joining to joinrels too.
 657  */
 658 static void
 659 make_rels_by_clause_joins(PlannerInfo *root,
 660                                                   RelOptInfo *old_rel,
 661                                                   ListCell *other_rels)
 662 {
 663         ListCell   *l;
 664
 665         for_each_cell(l, other_rels)
 666         {
 667                 RelOptInfo *other_rel = (RelOptInfo *) lfirst(l);
 668
 669                 if (!bms_overlap(old_rel->relids, other_rel->relids) &&
 670                         (have_relevant_joinclause(root, old_rel, other_rel) ||
 671                          have_join_order_restriction(root, old_rel, other_rel)))
 672                 {
 673                         (void) make_join_rel(root, old_rel, other_rel);
 674                 }
 675         }
 676 }
 677
 678 /*
 679  * make_rels_by_clauseless_joins
 680  *        Given a relation 'old_rel' and a list of other relations
 681  *        'other_rels', create a join relation between 'old_rel' and each
 682  *        member of 'other_rels' that isn't already included in 'old_rel'.
 683  *        The join rels are returned in root->join_rel_level[join_cur_level].
 684  *
 685  * 'old_rel' is the relation entry for the relation to be joined
 686  * 'other_rels': the first cell of a linked list containing the
 687  * other rels to be considered for joining
 688  *
 689  * Currently, this is only used with initial rels in other_rels, but it would
 690  * work for joining to joinrels too.
 691  */
 692 static void
 693 make_rels_by_clauseless_joins(PlannerInfo *root,
 694                                                           RelOptInfo *old_rel,
 695                                                           ListCell *other_rels)
 696 {
 697         ListCell   *l;
 698
 699         for_each_cell(l, other_rels)
 700         {
 701                 RelOptInfo *other_rel = (RelOptInfo *) lfirst(l);
 702
 703                 if (!bms_overlap(other_rel->relids, old_rel->relids))
 704                 {
 705                         (void) make_join_rel(root, old_rel, other_rel);
 706                 }
 707         }
 708 }
 709
 710 /*
 711  * join_is_legal
 712  *         Determine whether a proposed join is legal given the query's
 713  *         join order constraints; and if it is, determine the join type.
 714  *
 715  * Caller must supply not only the two rels, but the union of their relids.
 716  * (We could simplify the API by computing joinrelids locally, but this
 717  * would be redundant work in the normal path through make_join_rel.)
 718  *
 719  * On success, *sjinfo_p is set to NULL if this is to be a plain inner join,
 720  * else it's set to point to the associated SpecialJoinInfo node.  Also,
 721  * *reversed_p is set TRUE if the given relations need to be swapped to
 722  * match the SpecialJoinInfo node.
 723  */
 724 static bool
 725 join_is_legal(PlannerInfo *root, RelOptInfo *rel1, RelOptInfo *rel2,
 726                           Relids joinrelids,
 727                           SpecialJoinInfo **sjinfo_p, bool *reversed_p)
 728 {
 729         SpecialJoinInfo *match_sjinfo;
 730         bool            reversed;
 731         bool            unique_ified;
 732         bool            is_valid_inner;
 733         ListCell   *l;
 734
 735         /*
 736          * Ensure output params are set on failure return.      This is just to
 737          * suppress uninitialized-variable warnings from overly anal compilers.
 738          */
 739         *sjinfo_p = NULL;
 740         *reversed_p = false;
 741
 742         /*
 743          * If we have any special joins, the proposed join might be illegal; and
 744          * in any case we have to determine its join type.      Scan the join info
 745          * list for conflicts.
 746          */
 747         match_sjinfo = NULL;
 748         reversed = false;
 749         unique_ified = false;
 750         is_valid_inner = true;
 751
 752         foreach(l, root->join_info_list)
 753         {
 754                 SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
 755
 756                 /*
 757                  * This special join is not relevant unless its RHS overlaps the
 758                  * proposed join.  (Check this first as a fast path for dismissing
 759                  * most irrelevant SJs quickly.)
 760                  */
 761                 if (!bms_overlap(sjinfo->min_righthand, joinrelids))
 762                         continue;
 763
 764                 /*
 765                  * Also, not relevant if proposed join is fully contained within RHS
 766                  * (ie, we're still building up the RHS).
 767                  */
 768                 if (bms_is_subset(joinrelids, sjinfo->min_righthand))
 769                         continue;
 770
 771                 /*
 772                  * Also, not relevant if SJ is already done within either input.
 773                  */
 774                 if (bms_is_subset(sjinfo->min_lefthand, rel1->relids) &&
 775                         bms_is_subset(sjinfo->min_righthand, rel1->relids))
 776                         continue;
 777                 if (bms_is_subset(sjinfo->min_lefthand, rel2->relids) &&
 778                         bms_is_subset(sjinfo->min_righthand, rel2->relids))
 779                         continue;
 780
 781                 /*
 782                  * If it's a semijoin and we already joined the RHS to any other rels
 783                  * within either input, then we must have unique-ified the RHS at that
 784                  * point (see below).  Therefore the semijoin is no longer relevant in
 785                  * this join path.
 786                  */
 787                 if (sjinfo->jointype == JOIN_SEMI)
 788                 {
 789                         if (bms_is_subset(sjinfo->syn_righthand, rel1->relids) &&
 790                                 !bms_equal(sjinfo->syn_righthand, rel1->relids))
 791                                 continue;
 792                         if (bms_is_subset(sjinfo->syn_righthand, rel2->relids) &&
 793                                 !bms_equal(sjinfo->syn_righthand, rel2->relids))
 794                                 continue;
 795                 }
 796
 797                 /*
 798                  * If one input contains min_lefthand and the other contains
 799                  * min_righthand, then we can perform the SJ at this join.
 800                  *
 801                  * Barf if we get matches to more than one SJ (is that possible?)
 802                  */
 803                 if (bms_is_subset(sjinfo->min_lefthand, rel1->relids) &&
 804                         bms_is_subset(sjinfo->min_righthand, rel2->relids))
 805                 {
 806                         if (match_sjinfo)
 807                                 return false;   /* invalid join path */
 808                         match_sjinfo = sjinfo;
 809                         reversed = false;
 810                 }
 811                 else if (bms_is_subset(sjinfo->min_lefthand, rel2->relids) &&
 812                                  bms_is_subset(sjinfo->min_righthand, rel1->relids))
 813                 {
 814                         if (match_sjinfo)
 815                                 return false;   /* invalid join path */
 816                         match_sjinfo = sjinfo;
 817                         reversed = true;
 818                 }
 819                 else if (sjinfo->jointype == JOIN_SEMI &&
 820                                  bms_equal(sjinfo->syn_righthand, rel2->relids) &&
 821                                  create_unique_path(root, rel2, rel2->cheapest_total_path,
 822                                                                         sjinfo) != NULL)
 823                 {
 824                         /*----------
 825                          * For a semijoin, we can join the RHS to anything else by
 826                          * unique-ifying the RHS (if the RHS can be unique-ified).
 827                          * We will only get here if we have the full RHS but less
 828                          * than min_lefthand on the LHS.
 829                          *
 830                          * The reason to consider such a join path is exemplified by
 831                          *      SELECT ... FROM a,b WHERE (a.x,b.y) IN (SELECT c1,c2 FROM c)
 832                          * If we insist on doing this as a semijoin we will first have
 833                          * to form the cartesian product of A*B.  But if we unique-ify
 834                          * C then the semijoin becomes a plain innerjoin and we can join
 835                          * in any order, eg C to A and then to B.  When C is much smaller
 836                          * than A and B this can be a huge win.  So we allow C to be
 837                          * joined to just A or just B here, and then make_join_rel has
 838                          * to handle the case properly.
 839                          *
 840                          * Note that actually we'll allow unique-ified C to be joined to
 841                          * some other relation D here, too.  That is legal, if usually not
 842                          * very sane, and this routine is only concerned with legality not
 843                          * with whether the join is good strategy.
 844                          *----------
 845                          */
 846                         if (match_sjinfo)
 847                                 return false;   /* invalid join path */
 848                         match_sjinfo = sjinfo;
 849                         reversed = false;
 850                         unique_ified = true;
 851                 }
 852                 else if (sjinfo->jointype == JOIN_SEMI &&
 853                                  bms_equal(sjinfo->syn_righthand, rel1->relids) &&
 854                                  create_unique_path(root, rel1, rel1->cheapest_total_path,
 855                                                                         sjinfo) != NULL)
 856                 {
 857                         /* Reversed semijoin case */
 858                         if (match_sjinfo)
 859                                 return false;   /* invalid join path */
 860                         match_sjinfo = sjinfo;
 861                         reversed = true;
 862                         unique_ified = true;
 863                 }
 864                 else
 865                 {
 866                         /*----------
 867                          * Otherwise, the proposed join overlaps the RHS but isn't
 868                          * a valid implementation of this SJ.  It might still be
 869                          * a legal join, however.  If both inputs overlap the RHS,
 870                          * assume that it's OK.  Since the inputs presumably got past
 871                          * this function's checks previously, they can't overlap the
 872                          * LHS and their violations of the RHS boundary must represent
 873                          * SJs that have been determined to commute with this one.
 874                          * We have to allow this to work correctly in cases like
 875                          *              (a LEFT JOIN (b JOIN (c LEFT JOIN d)))
 876                          * when the c/d join has been determined to commute with the join
 877                          * to a, and hence d is not part of min_righthand for the upper
 878                          * join.  It should be legal to join b to c/d but this will appear
 879                          * as a violation of the upper join's RHS.
 880                          * Furthermore, if one input overlaps the RHS and the other does
 881                          * not, we should still allow the join if it is a valid
 882                          * implementation of some other SJ.  We have to allow this to
 883                          * support the associative identity
 884                          *              (a LJ b on Pab) LJ c ON Pbc = a LJ (b LJ c ON Pbc) on Pab
 885                          * since joining B directly to C violates the lower SJ's RHS.
 886                          * We assume that make_outerjoininfo() set things up correctly
 887                          * so that we'll only match to some SJ if the join is valid.
 888                          * Set flag here to check at bottom of loop.
 889                          *----------
 890                          */
 891                         if (sjinfo->jointype != JOIN_SEMI &&
 892                                 bms_overlap(rel1->relids, sjinfo->min_righthand) &&
 893                                 bms_overlap(rel2->relids, sjinfo->min_righthand))
 894                         {
 895                                 /* seems OK */
 896                                 Assert(!bms_overlap(joinrelids, sjinfo->min_lefthand));
 897                         }
 898                         else
 899                                 is_valid_inner = false;
 900                 }
 901         }
 902
 903         /*
 904          * Fail if violated some SJ's RHS and didn't match to another SJ. However,
 905          * "matching" to a semijoin we are implementing by unique-ification
 906          * doesn't count (think: it's really an inner join).
 907          */
 908         if (!is_valid_inner &&
 909                 (match_sjinfo == NULL || unique_ified))
 910                 return false;                   /* invalid join path */
 911
 912         /* Otherwise, it's a valid join */
 913         *sjinfo_p = match_sjinfo;
 914         *reversed_p = reversed;
 915         return true;
 916 }
 917
 918 /*
 919  * has_join_restriction
 920  *              Detect whether the specified relation has join-order restrictions
 921  *              due to being inside an outer join or an IN (sub-SELECT).
 922  *
 923  * Essentially, this tests whether have_join_order_restriction() could
 924  * succeed with this rel and some other one.  It's OK if we sometimes
 925  * say "true" incorrectly.      (Therefore, we don't bother with the relatively
 926  * expensive has_legal_joinclause test.)
 927  */
 928 static bool
 929 has_join_restriction(PlannerInfo *root, RelOptInfo *rel)
 930 {
 931         ListCell   *l;
 932
 933         foreach(l, root->join_info_list)
 934         {
 935                 SpecialJoinInfo *sjinfo = (SpecialJoinInfo *) lfirst(l);
 936
 937                 /* ignore full joins --- other mechanisms preserve their ordering */
 938                 if (sjinfo->jointype == JOIN_FULL)
 939                         continue;
 940
 941                 /* ignore if SJ is already contained in rel */
 942                 if (bms_is_subset(sjinfo->min_lefthand, rel->relids) &&
 943                         bms_is_subset(sjinfo->min_righthand, rel->relids))
 944                         continue;
 945
 946                 /* restricted if it overlaps LHS or RHS, but doesn't contain SJ */
 947                 if (bms_overlap(sjinfo->min_lefthand, rel->relids) ||
 948                         bms_overlap(sjinfo->min_righthand, rel->relids))
 949                         return true;
 950         }
 951
 952         return false;
 953 }
 954
 955 /*
 956  * is_dummy_rel --- has relation been proven empty?
 957  */
 958 static bool
 959 is_dummy_rel(RelOptInfo *rel)
 960 {
 961         return IS_DUMMY_REL(rel);
 962 }
 963
 964 /*
 965  * Mark a relation as proven empty.
 966  *
 967  * During GEQO planning, this can get invoked more than once on the same
 968  * baserel struct, so it's worth checking to see if the rel is already marked
 969  * dummy.
 970  *
 971  * Also, when called during GEQO join planning, we are in a short-lived
 972  * memory context.      We must make sure that the dummy path attached to a
 973  * baserel survives the GEQO cycle, else the baserel is trashed for future
 974  * GEQO cycles.  On the other hand, when we are marking a joinrel during GEQO,
 975  * we don't want the dummy path to clutter the main planning context.  Upshot
 976  * is that the best solution is to explicitly make the dummy path in the same
 977  * context the given RelOptInfo is in.
 978  */
 979 static void
 980 mark_dummy_rel(RelOptInfo *rel)
 981 {
 982         MemoryContext oldcontext;
 983
 984         /* Already marked? */
 985         if (is_dummy_rel(rel))
 986                 return;
 987
 988         /* No, so choose correct context to make the dummy path in */
 989         oldcontext = MemoryContextSwitchTo(GetMemoryChunkContext(rel));
 990
 991         /* Set dummy size estimate */
 992         rel->rows = 0;
 993
 994         /* Evict any previously chosen paths */
 995         rel->pathlist = NIL;
 996
 997         /* Set up the dummy path */
 998         add_path(rel, (Path *) create_append_path(rel, NIL, NULL));
 999
1000         /* Set or update cheapest_total_path and related fields */
1001         set_cheapest(rel);
1002
1003         MemoryContextSwitchTo(oldcontext);
1004 }
1005
1006 /*
1007  * restriction_is_constant_false --- is a restrictlist just FALSE?
1008  *
1009  * In cases where a qual is provably constant FALSE, eval_const_expressions
1010  * will generally have thrown away anything that's ANDed with it.  In outer
1011  * join situations this will leave us computing cartesian products only to
1012  * decide there's no match for an outer row, which is pretty stupid.  So,
1013  * we need to detect the case.
1014  *
1015  * If only_pushed_down is TRUE, then consider only pushed-down quals.
1016  */
1017 static bool
1018 restriction_is_constant_false(List *restrictlist, bool only_pushed_down)
1019 {
1020         ListCell   *lc;
1021
1022         /*
1023          * Despite the above comment, the restriction list we see here might
1024          * possibly have other members besides the FALSE constant, since other
1025          * quals could get "pushed down" to the outer join level.  So we check
1026          * each member of the list.
1027          */
1028         foreach(lc, restrictlist)
1029         {
1030                 RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
1031
1032                 Assert(IsA(rinfo, RestrictInfo));
1033                 if (only_pushed_down && !rinfo->is_pushed_down)
1034                         continue;
1035
1036                 if (rinfo->clause && IsA(rinfo->clause, Const))
1037                 {
1038                         Const      *con = (Const *) rinfo->clause;
1039
1040                         /* constant NULL is as good as constant FALSE for our purposes */
1041                         if (con->constisnull)
1042                                 return true;
1043                         if (!DatumGetBool(con->constvalue))
1044                                 return true;
1045                 }
1046         }
1047         return false;
1048 }