granicus.if.org Git - postgresql/blob - src/backend/optimizer/util/pathnode.c

   1 /*-------------------------------------------------------------------------
   2  *
   3  * pathnode.c
   4  *        Routines to manipulate pathlists and create path nodes
   5  *
   6  * Portions Copyright (c) 1996-2014, PostgreSQL Global Development Group
   7  * Portions Copyright (c) 1994, Regents of the University of California
   8  *
   9  *
  10  * IDENTIFICATION
  11  *        src/backend/optimizer/util/pathnode.c
  12  *
  13  *-------------------------------------------------------------------------
  14  */
  15 #include "postgres.h"
  16
  17 #include <math.h>
  18
  19 #include "miscadmin.h"
  20 #include "nodes/nodeFuncs.h"
  21 #include "optimizer/clauses.h"
  22 #include "optimizer/cost.h"
  23 #include "optimizer/pathnode.h"
  24 #include "optimizer/paths.h"
  25 #include "optimizer/planmain.h"
  26 #include "optimizer/restrictinfo.h"
  27 #include "optimizer/var.h"
  28 #include "parser/parsetree.h"
  29 #include "utils/lsyscache.h"
  30 #include "utils/memutils.h"
  31 #include "utils/selfuncs.h"
  32
  33
  34 typedef enum
  35 {
  36         COSTS_EQUAL,                            /* path costs are fuzzily equal */
  37         COSTS_BETTER1,                          /* first path is cheaper than second */
  38         COSTS_BETTER2,                          /* second path is cheaper than first */
  39         COSTS_DIFFERENT                         /* neither path dominates the other on cost */
  40 } PathCostComparison;
  41
  42 static List *translate_sub_tlist(List *tlist, int relid);
  43
  44
  45 /*****************************************************************************
  46  *              MISC. PATH UTILITIES
  47  *****************************************************************************/
  48
  49 /*
  50  * compare_path_costs
  51  *        Return -1, 0, or +1 according as path1 is cheaper, the same cost,
  52  *        or more expensive than path2 for the specified criterion.
  53  */
  54 int
  55 compare_path_costs(Path *path1, Path *path2, CostSelector criterion)
  56 {
  57         if (criterion == STARTUP_COST)
  58         {
  59                 if (path1->startup_cost < path2->startup_cost)
  60                         return -1;
  61                 if (path1->startup_cost > path2->startup_cost)
  62                         return +1;
  63
  64                 /*
  65                  * If paths have the same startup cost (not at all unlikely), order
  66                  * them by total cost.
  67                  */
  68                 if (path1->total_cost < path2->total_cost)
  69                         return -1;
  70                 if (path1->total_cost > path2->total_cost)
  71                         return +1;
  72         }
  73         else
  74         {
  75                 if (path1->total_cost < path2->total_cost)
  76                         return -1;
  77                 if (path1->total_cost > path2->total_cost)
  78                         return +1;
  79
  80                 /*
  81                  * If paths have the same total cost, order them by startup cost.
  82                  */
  83                 if (path1->startup_cost < path2->startup_cost)
  84                         return -1;
  85                 if (path1->startup_cost > path2->startup_cost)
  86                         return +1;
  87         }
  88         return 0;
  89 }
  90
  91 /*
  92  * compare_path_fractional_costs
  93  *        Return -1, 0, or +1 according as path1 is cheaper, the same cost,
  94  *        or more expensive than path2 for fetching the specified fraction
  95  *        of the total tuples.
  96  *
  97  * If fraction is <= 0 or > 1, we interpret it as 1, ie, we select the
  98  * path with the cheaper total_cost.
  99  */
 100 int
 101 compare_fractional_path_costs(Path *path1, Path *path2,
 102                                                           double fraction)
 103 {
 104         Cost            cost1,
 105                                 cost2;
 106
 107         if (fraction <= 0.0 || fraction >= 1.0)
 108                 return compare_path_costs(path1, path2, TOTAL_COST);
 109         cost1 = path1->startup_cost +
 110                 fraction * (path1->total_cost - path1->startup_cost);
 111         cost2 = path2->startup_cost +
 112                 fraction * (path2->total_cost - path2->startup_cost);
 113         if (cost1 < cost2)
 114                 return -1;
 115         if (cost1 > cost2)
 116                 return +1;
 117         return 0;
 118 }
 119
 120 /*
 121  * compare_path_costs_fuzzily
 122  *        Compare the costs of two paths to see if either can be said to
 123  *        dominate the other.
 124  *
 125  * We use fuzzy comparisons so that add_path() can avoid keeping both of
 126  * a pair of paths that really have insignificantly different cost.
 127  *
 128  * The fuzz_factor argument must be 1.0 plus delta, where delta is the
 129  * fraction of the smaller cost that is considered to be a significant
 130  * difference.  For example, fuzz_factor = 1.01 makes the fuzziness limit
 131  * be 1% of the smaller cost.
 132  *
 133  * The two paths are said to have "equal" costs if both startup and total
 134  * costs are fuzzily the same.  Path1 is said to be better than path2 if
 135  * it has fuzzily better startup cost and fuzzily no worse total cost,
 136  * or if it has fuzzily better total cost and fuzzily no worse startup cost.
 137  * Path2 is better than path1 if the reverse holds.  Finally, if one path
 138  * is fuzzily better than the other on startup cost and fuzzily worse on
 139  * total cost, we just say that their costs are "different", since neither
 140  * dominates the other across the whole performance spectrum.
 141  *
 142  * If consider_startup is false, then we don't care about keeping paths with
 143  * good startup cost, so we'll never return COSTS_DIFFERENT.
 144  *
 145  * This function also includes special hacks to support a policy enforced
 146  * by its sole caller, add_path(): paths that have any parameterization
 147  * cannot win comparisons on the grounds of having cheaper startup cost,
 148  * since we deem only total cost to be of interest for a parameterized path.
 149  * (Unparameterized paths are more common, so we check for this case last.)
 150  */
 151 static PathCostComparison
 152 compare_path_costs_fuzzily(Path *path1, Path *path2, double fuzz_factor,
 153                                                    bool consider_startup)
 154 {
 155         /*
 156          * Check total cost first since it's more likely to be different; many
 157          * paths have zero startup cost.
 158          */
 159         if (path1->total_cost > path2->total_cost * fuzz_factor)
 160         {
 161                 /* path1 fuzzily worse on total cost */
 162                 if (consider_startup &&
 163                         path2->startup_cost > path1->startup_cost * fuzz_factor &&
 164                         path1->param_info == NULL)
 165                 {
 166                         /* ... but path2 fuzzily worse on startup, so DIFFERENT */
 167                         return COSTS_DIFFERENT;
 168                 }
 169                 /* else path2 dominates */
 170                 return COSTS_BETTER2;
 171         }
 172         if (path2->total_cost > path1->total_cost * fuzz_factor)
 173         {
 174                 /* path2 fuzzily worse on total cost */
 175                 if (consider_startup &&
 176                         path1->startup_cost > path2->startup_cost * fuzz_factor &&
 177                         path2->param_info == NULL)
 178                 {
 179                         /* ... but path1 fuzzily worse on startup, so DIFFERENT */
 180                         return COSTS_DIFFERENT;
 181                 }
 182                 /* else path1 dominates */
 183                 return COSTS_BETTER1;
 184         }
 185         /* fuzzily the same on total cost */
 186         /* (so we may as well compare startup cost, even if !consider_startup) */
 187         if (path1->startup_cost > path2->startup_cost * fuzz_factor &&
 188                 path2->param_info == NULL)
 189         {
 190                 /* ... but path1 fuzzily worse on startup, so path2 wins */
 191                 return COSTS_BETTER2;
 192         }
 193         if (path2->startup_cost > path1->startup_cost * fuzz_factor &&
 194                 path1->param_info == NULL)
 195         {
 196                 /* ... but path2 fuzzily worse on startup, so path1 wins */
 197                 return COSTS_BETTER1;
 198         }
 199         /* fuzzily the same on both costs */
 200         return COSTS_EQUAL;
 201 }
 202
 203 /*
 204  * set_cheapest
 205  *        Find the minimum-cost paths from among a relation's paths,
 206  *        and save them in the rel's cheapest-path fields.
 207  *
 208  * cheapest_total_path is normally the cheapest-total-cost unparameterized
 209  * path; but if there are no unparameterized paths, we assign it to be the
 210  * best (cheapest least-parameterized) parameterized path.  However, only
 211  * unparameterized paths are considered candidates for cheapest_startup_path,
 212  * so that will be NULL if there are no unparameterized paths.
 213  *
 214  * The cheapest_parameterized_paths list collects all parameterized paths
 215  * that have survived the add_path() tournament for this relation.  (Since
 216  * add_path ignores pathkeys and startup cost for a parameterized path,
 217  * these will be paths that have best total cost or best row count for their
 218  * parameterization.)  cheapest_parameterized_paths always includes the
 219  * cheapest-total unparameterized path, too, if there is one; the users of
 220  * that list find it more convenient if that's included.
 221  *
 222  * This is normally called only after we've finished constructing the path
 223  * list for the rel node.
 224  */
 225 void
 226 set_cheapest(RelOptInfo *parent_rel)
 227 {
 228         Path       *cheapest_startup_path;
 229         Path       *cheapest_total_path;
 230         Path       *best_param_path;
 231         List       *parameterized_paths;
 232         ListCell   *p;
 233
 234         Assert(IsA(parent_rel, RelOptInfo));
 235
 236         if (parent_rel->pathlist == NIL)
 237                 elog(ERROR, "could not devise a query plan for the given query");
 238
 239         cheapest_startup_path = cheapest_total_path = best_param_path = NULL;
 240         parameterized_paths = NIL;
 241
 242         foreach(p, parent_rel->pathlist)
 243         {
 244                 Path       *path = (Path *) lfirst(p);
 245                 int                     cmp;
 246
 247                 if (path->param_info)
 248                 {
 249                         /* Parameterized path, so add it to parameterized_paths */
 250                         parameterized_paths = lappend(parameterized_paths, path);
 251
 252                         /*
 253                          * If we have an unparameterized cheapest-total, we no longer care
 254                          * about finding the best parameterized path, so move on.
 255                          */
 256                         if (cheapest_total_path)
 257                                 continue;
 258
 259                         /*
 260                          * Otherwise, track the best parameterized path, which is the one
 261                          * with least total cost among those of the minimum
 262                          * parameterization.
 263                          */
 264                         if (best_param_path == NULL)
 265                                 best_param_path = path;
 266                         else
 267                         {
 268                                 switch (bms_subset_compare(PATH_REQ_OUTER(path),
 269                                                                                    PATH_REQ_OUTER(best_param_path)))
 270                                 {
 271                                         case BMS_EQUAL:
 272                                                 /* keep the cheaper one */
 273                                                 if (compare_path_costs(path, best_param_path,
 274                                                                                            TOTAL_COST) < 0)
 275                                                         best_param_path = path;
 276                                                 break;
 277                                         case BMS_SUBSET1:
 278                                                 /* new path is less-parameterized */
 279                                                 best_param_path = path;
 280                                                 break;
 281                                         case BMS_SUBSET2:
 282                                                 /* old path is less-parameterized, keep it */
 283                                                 break;
 284                                         case BMS_DIFFERENT:
 285
 286                                                 /*
 287                                                  * This means that neither path has the least possible
 288                                                  * parameterization for the rel.  We'll sit on the old
 289                                                  * path until something better comes along.
 290                                                  */
 291                                                 break;
 292                                 }
 293                         }
 294                 }
 295                 else
 296                 {
 297                         /* Unparameterized path, so consider it for cheapest slots */
 298                         if (cheapest_total_path == NULL)
 299                         {
 300                                 cheapest_startup_path = cheapest_total_path = path;
 301                                 continue;
 302                         }
 303
 304                         /*
 305                          * If we find two paths of identical costs, try to keep the
 306                          * better-sorted one.  The paths might have unrelated sort
 307                          * orderings, in which case we can only guess which might be
 308                          * better to keep, but if one is superior then we definitely
 309                          * should keep that one.
 310                          */
 311                         cmp = compare_path_costs(cheapest_startup_path, path, STARTUP_COST);
 312                         if (cmp > 0 ||
 313                                 (cmp == 0 &&
 314                                  compare_pathkeys(cheapest_startup_path->pathkeys,
 315                                                                   path->pathkeys) == PATHKEYS_BETTER2))
 316                                 cheapest_startup_path = path;
 317
 318                         cmp = compare_path_costs(cheapest_total_path, path, TOTAL_COST);
 319                         if (cmp > 0 ||
 320                                 (cmp == 0 &&
 321                                  compare_pathkeys(cheapest_total_path->pathkeys,
 322                                                                   path->pathkeys) == PATHKEYS_BETTER2))
 323                                 cheapest_total_path = path;
 324                 }
 325         }
 326
 327         /* Add cheapest unparameterized path, if any, to parameterized_paths */
 328         if (cheapest_total_path)
 329                 parameterized_paths = lcons(cheapest_total_path, parameterized_paths);
 330
 331         /*
 332          * If there is no unparameterized path, use the best parameterized path as
 333          * cheapest_total_path (but not as cheapest_startup_path).
 334          */
 335         if (cheapest_total_path == NULL)
 336                 cheapest_total_path = best_param_path;
 337         Assert(cheapest_total_path != NULL);
 338
 339         parent_rel->cheapest_startup_path = cheapest_startup_path;
 340         parent_rel->cheapest_total_path = cheapest_total_path;
 341         parent_rel->cheapest_unique_path = NULL;        /* computed only if needed */
 342         parent_rel->cheapest_parameterized_paths = parameterized_paths;
 343 }
 344
 345 /*
 346  * add_path
 347  *        Consider a potential implementation path for the specified parent rel,
 348  *        and add it to the rel's pathlist if it is worthy of consideration.
 349  *        A path is worthy if it has a better sort order (better pathkeys) or
 350  *        cheaper cost (on either dimension), or generates fewer rows, than any
 351  *        existing path that has the same or superset parameterization rels.
 352  *
 353  *        We also remove from the rel's pathlist any old paths that are dominated
 354  *        by new_path --- that is, new_path is cheaper, at least as well ordered,
 355  *        generates no more rows, and requires no outer rels not required by the
 356  *        old path.
 357  *
 358  *        In most cases, a path with a superset parameterization will generate
 359  *        fewer rows (since it has more join clauses to apply), so that those two
 360  *        figures of merit move in opposite directions; this means that a path of
 361  *        one parameterization can seldom dominate a path of another.  But such
 362  *        cases do arise, so we make the full set of checks anyway.
 363  *
 364  *        There are two policy decisions embedded in this function, along with
 365  *        its sibling add_path_precheck: we treat all parameterized paths as
 366  *        having NIL pathkeys, and we ignore their startup costs, so that they
 367  *        compete only on parameterization, total cost and rowcount.  This is to
 368  *        reduce the number of parameterized paths that are kept.  See discussion
 369  *        in src/backend/optimizer/README.
 370  *
 371  *        Another policy that is enforced here is that we only consider cheap
 372  *        startup cost to be interesting if parent_rel->consider_startup is true.
 373  *
 374  *        The pathlist is kept sorted by total_cost, with cheaper paths
 375  *        at the front.  Within this routine, that's simply a speed hack:
 376  *        doing it that way makes it more likely that we will reject an inferior
 377  *        path after a few comparisons, rather than many comparisons.
 378  *        However, add_path_precheck relies on this ordering to exit early
 379  *        when possible.
 380  *
 381  *        NOTE: discarded Path objects are immediately pfree'd to reduce planner
 382  *        memory consumption.  We dare not try to free the substructure of a Path,
 383  *        since much of it may be shared with other Paths or the query tree itself;
 384  *        but just recycling discarded Path nodes is a very useful savings in
 385  *        a large join tree.  We can recycle the List nodes of pathlist, too.
 386  *
 387  *        BUT: we do not pfree IndexPath objects, since they may be referenced as
 388  *        children of BitmapHeapPaths as well as being paths in their own right.
 389  *
 390  * 'parent_rel' is the relation entry to which the path corresponds.
 391  * 'new_path' is a potential path for parent_rel.
 392  *
 393  * Returns nothing, but modifies parent_rel->pathlist.
 394  */
 395 void
 396 add_path(RelOptInfo *parent_rel, Path *new_path)
 397 {
 398         bool            accept_new = true;              /* unless we find a superior old path */
 399         ListCell   *insert_after = NULL;        /* where to insert new item */
 400         List       *new_path_pathkeys;
 401         ListCell   *p1;
 402         ListCell   *p1_prev;
 403         ListCell   *p1_next;
 404
 405         /*
 406          * This is a convenient place to check for query cancel --- no part of the
 407          * planner goes very long without calling add_path().
 408          */
 409         CHECK_FOR_INTERRUPTS();
 410
 411         /* Pretend parameterized paths have no pathkeys, per comment above */
 412         new_path_pathkeys = new_path->param_info ? NIL : new_path->pathkeys;
 413
 414         /*
 415          * Loop to check proposed new path against old paths.  Note it is possible
 416          * for more than one old path to be tossed out because new_path dominates
 417          * it.
 418          *
 419          * We can't use foreach here because the loop body may delete the current
 420          * list cell.
 421          */
 422         p1_prev = NULL;
 423         for (p1 = list_head(parent_rel->pathlist); p1 != NULL; p1 = p1_next)
 424         {
 425                 Path       *old_path = (Path *) lfirst(p1);
 426                 bool            remove_old = false; /* unless new proves superior */
 427                 PathCostComparison costcmp;
 428                 PathKeysComparison keyscmp;
 429                 BMS_Comparison outercmp;
 430
 431                 p1_next = lnext(p1);
 432
 433                 /*
 434                  * Do a fuzzy cost comparison with 1% fuzziness limit.  (XXX does this
 435                  * percentage need to be user-configurable?)
 436                  */
 437                 costcmp = compare_path_costs_fuzzily(new_path, old_path, 1.01,
 438                                                                                          parent_rel->consider_startup);
 439
 440                 /*
 441                  * If the two paths compare differently for startup and total cost,
 442                  * then we want to keep both, and we can skip comparing pathkeys and
 443                  * required_outer rels.  If they compare the same, proceed with the
 444                  * other comparisons.  Row count is checked last.  (We make the tests
 445                  * in this order because the cost comparison is most likely to turn
 446                  * out "different", and the pathkeys comparison next most likely.  As
 447                  * explained above, row count very seldom makes a difference, so even
 448                  * though it's cheap to compare there's not much point in checking it
 449                  * earlier.)
 450                  */
 451                 if (costcmp != COSTS_DIFFERENT)
 452                 {
 453                         /* Similarly check to see if either dominates on pathkeys */
 454                         List       *old_path_pathkeys;
 455
 456                         old_path_pathkeys = old_path->param_info ? NIL : old_path->pathkeys;
 457                         keyscmp = compare_pathkeys(new_path_pathkeys,
 458                                                                            old_path_pathkeys);
 459                         if (keyscmp != PATHKEYS_DIFFERENT)
 460                         {
 461                                 switch (costcmp)
 462                                 {
 463                                         case COSTS_EQUAL:
 464                                                 outercmp = bms_subset_compare(PATH_REQ_OUTER(new_path),
 465                                                                                                    PATH_REQ_OUTER(old_path));
 466                                                 if (keyscmp == PATHKEYS_BETTER1)
 467                                                 {
 468                                                         if ((outercmp == BMS_EQUAL ||
 469                                                                  outercmp == BMS_SUBSET1) &&
 470                                                                 new_path->rows <= old_path->rows)
 471                                                                 remove_old = true;              /* new dominates old */
 472                                                 }
 473                                                 else if (keyscmp == PATHKEYS_BETTER2)
 474                                                 {
 475                                                         if ((outercmp == BMS_EQUAL ||
 476                                                                  outercmp == BMS_SUBSET2) &&
 477                                                                 new_path->rows >= old_path->rows)
 478                                                                 accept_new = false;             /* old dominates new */
 479                                                 }
 480                                                 else    /* keyscmp == PATHKEYS_EQUAL */
 481                                                 {
 482                                                         if (outercmp == BMS_EQUAL)
 483                                                         {
 484                                                                 /*
 485                                                                  * Same pathkeys and outer rels, and fuzzily
 486                                                                  * the same cost, so keep just one; to decide
 487                                                                  * which, first check rows and then do a fuzzy
 488                                                                  * cost comparison with very small fuzz limit.
 489                                                                  * (We used to do an exact cost comparison,
 490                                                                  * but that results in annoying
 491                                                                  * platform-specific plan variations due to
 492                                                                  * roundoff in the cost estimates.)  If things
 493                                                                  * are still tied, arbitrarily keep only the
 494                                                                  * old path.  Notice that we will keep only
 495                                                                  * the old path even if the less-fuzzy
 496                                                                  * comparison decides the startup and total
 497                                                                  * costs compare differently.
 498                                                                  */
 499                                                                 if (new_path->rows < old_path->rows)
 500                                                                         remove_old = true;      /* new dominates old */
 501                                                                 else if (new_path->rows > old_path->rows)
 502                                                                         accept_new = false; /* old dominates new */
 503                                                                 else if (compare_path_costs_fuzzily(new_path,
 504                                                                                                                                         old_path,
 505                                                                                                                                 1.0000000001,
 506                                                                                                                                         parent_rel->consider_startup) == COSTS_BETTER1)
 507                                                                         remove_old = true;      /* new dominates old */
 508                                                                 else
 509                                                                         accept_new = false; /* old equals or
 510                                                                                                                  * dominates new */
 511                                                         }
 512                                                         else if (outercmp == BMS_SUBSET1 &&
 513                                                                          new_path->rows <= old_path->rows)
 514                                                                 remove_old = true;              /* new dominates old */
 515                                                         else if (outercmp == BMS_SUBSET2 &&
 516                                                                          new_path->rows >= old_path->rows)
 517                                                                 accept_new = false;             /* old dominates new */
 518                                                         /* else different parameterizations, keep both */
 519                                                 }
 520                                                 break;
 521                                         case COSTS_BETTER1:
 522                                                 if (keyscmp != PATHKEYS_BETTER2)
 523                                                 {
 524                                                         outercmp = bms_subset_compare(PATH_REQ_OUTER(new_path),
 525                                                                                                    PATH_REQ_OUTER(old_path));
 526                                                         if ((outercmp == BMS_EQUAL ||
 527                                                                  outercmp == BMS_SUBSET1) &&
 528                                                                 new_path->rows <= old_path->rows)
 529                                                                 remove_old = true;              /* new dominates old */
 530                                                 }
 531                                                 break;
 532                                         case COSTS_BETTER2:
 533                                                 if (keyscmp != PATHKEYS_BETTER1)
 534                                                 {
 535                                                         outercmp = bms_subset_compare(PATH_REQ_OUTER(new_path),
 536                                                                                                    PATH_REQ_OUTER(old_path));
 537                                                         if ((outercmp == BMS_EQUAL ||
 538                                                                  outercmp == BMS_SUBSET2) &&
 539                                                                 new_path->rows >= old_path->rows)
 540                                                                 accept_new = false;             /* old dominates new */
 541                                                 }
 542                                                 break;
 543                                         case COSTS_DIFFERENT:
 544
 545                                                 /*
 546                                                  * can't get here, but keep this case to keep compiler
 547                                                  * quiet
 548                                                  */
 549                                                 break;
 550                                 }
 551                         }
 552                 }
 553
 554                 /*
 555                  * Remove current element from pathlist if dominated by new.
 556                  */
 557                 if (remove_old)
 558                 {
 559                         parent_rel->pathlist = list_delete_cell(parent_rel->pathlist,
 560                                                                                                         p1, p1_prev);
 561
 562                         /*
 563                          * Delete the data pointed-to by the deleted cell, if possible
 564                          */
 565                         if (!IsA(old_path, IndexPath))
 566                                 pfree(old_path);
 567                         /* p1_prev does not advance */
 568                 }
 569                 else
 570                 {
 571                         /* new belongs after this old path if it has cost >= old's */
 572                         if (new_path->total_cost >= old_path->total_cost)
 573                                 insert_after = p1;
 574                         /* p1_prev advances */
 575                         p1_prev = p1;
 576                 }
 577
 578                 /*
 579                  * If we found an old path that dominates new_path, we can quit
 580                  * scanning the pathlist; we will not add new_path, and we assume
 581                  * new_path cannot dominate any other elements of the pathlist.
 582                  */
 583                 if (!accept_new)
 584                         break;
 585         }
 586
 587         if (accept_new)
 588         {
 589                 /* Accept the new path: insert it at proper place in pathlist */
 590                 if (insert_after)
 591                         lappend_cell(parent_rel->pathlist, insert_after, new_path);
 592                 else
 593                         parent_rel->pathlist = lcons(new_path, parent_rel->pathlist);
 594         }
 595         else
 596         {
 597                 /* Reject and recycle the new path */
 598                 if (!IsA(new_path, IndexPath))
 599                         pfree(new_path);
 600         }
 601 }
 602
 603 /*
 604  * add_path_precheck
 605  *        Check whether a proposed new path could possibly get accepted.
 606  *        We assume we know the path's pathkeys and parameterization accurately,
 607  *        and have lower bounds for its costs.
 608  *
 609  * Note that we do not know the path's rowcount, since getting an estimate for
 610  * that is too expensive to do before prechecking.  We assume here that paths
 611  * of a superset parameterization will generate fewer rows; if that holds,
 612  * then paths with different parameterizations cannot dominate each other
 613  * and so we can simply ignore existing paths of another parameterization.
 614  * (In the infrequent cases where that rule of thumb fails, add_path will
 615  * get rid of the inferior path.)
 616  *
 617  * At the time this is called, we haven't actually built a Path structure,
 618  * so the required information has to be passed piecemeal.
 619  */
 620 bool
 621 add_path_precheck(RelOptInfo *parent_rel,
 622                                   Cost startup_cost, Cost total_cost,
 623                                   List *pathkeys, Relids required_outer)
 624 {
 625         List       *new_path_pathkeys;
 626         ListCell   *p1;
 627
 628         /* Pretend parameterized paths have no pathkeys, per add_path policy */
 629         new_path_pathkeys = required_outer ? NIL : pathkeys;
 630
 631         foreach(p1, parent_rel->pathlist)
 632         {
 633                 Path       *old_path = (Path *) lfirst(p1);
 634                 PathKeysComparison keyscmp;
 635
 636                 /*
 637                  * We are looking for an old_path with the same parameterization (and
 638                  * by assumption the same rowcount) that dominates the new path on
 639                  * pathkeys as well as both cost metrics.  If we find one, we can
 640                  * reject the new path.
 641                  *
 642                  * For speed, we make exact rather than fuzzy cost comparisons. If an
 643                  * old path dominates the new path exactly on both costs, it will
 644                  * surely do so fuzzily.
 645                  */
 646                 if (total_cost >= old_path->total_cost)
 647                 {
 648                         /* can win on startup cost only if unparameterized */
 649                         if (startup_cost >= old_path->startup_cost || required_outer)
 650                         {
 651                                 /* new path does not win on cost, so check pathkeys... */
 652                                 List       *old_path_pathkeys;
 653
 654                                 old_path_pathkeys = old_path->param_info ? NIL : old_path->pathkeys;
 655                                 keyscmp = compare_pathkeys(new_path_pathkeys,
 656                                                                                    old_path_pathkeys);
 657                                 if (keyscmp == PATHKEYS_EQUAL ||
 658                                         keyscmp == PATHKEYS_BETTER2)
 659                                 {
 660                                         /* new path does not win on pathkeys... */
 661                                         if (bms_equal(required_outer, PATH_REQ_OUTER(old_path)))
 662                                         {
 663                                                 /* Found an old path that dominates the new one */
 664                                                 return false;
 665                                         }
 666                                 }
 667                         }
 668                 }
 669                 else
 670                 {
 671                         /*
 672                          * Since the pathlist is sorted by total_cost, we can stop looking
 673                          * once we reach a path with a total_cost larger than the new
 674                          * path's.
 675                          */
 676                         break;
 677                 }
 678         }
 679
 680         return true;
 681 }
 682
 683
 684 /*****************************************************************************
 685  *              PATH NODE CREATION ROUTINES
 686  *****************************************************************************/
 687
 688 /*
 689  * create_seqscan_path
 690  *        Creates a path corresponding to a sequential scan, returning the
 691  *        pathnode.
 692  */
 693 Path *
 694 create_seqscan_path(PlannerInfo *root, RelOptInfo *rel, Relids required_outer)
 695 {
 696         Path       *pathnode = makeNode(Path);
 697
 698         pathnode->pathtype = T_SeqScan;
 699         pathnode->parent = rel;
 700         pathnode->param_info = get_baserel_parampathinfo(root, rel,
 701                                                                                                          required_outer);
 702         pathnode->pathkeys = NIL;       /* seqscan has unordered result */
 703
 704         cost_seqscan(pathnode, root, rel, pathnode->param_info);
 705
 706         return pathnode;
 707 }
 708
 709 /*
 710  * create_index_path
 711  *        Creates a path node for an index scan.
 712  *
 713  * 'index' is a usable index.
 714  * 'indexclauses' is a list of RestrictInfo nodes representing clauses
 715  *                      to be used as index qual conditions in the scan.
 716  * 'indexclausecols' is an integer list of index column numbers (zero based)
 717  *                      the indexclauses can be used with.
 718  * 'indexorderbys' is a list of bare expressions (no RestrictInfos)
 719  *                      to be used as index ordering operators in the scan.
 720  * 'indexorderbycols' is an integer list of index column numbers (zero based)
 721  *                      the ordering operators can be used with.
 722  * 'pathkeys' describes the ordering of the path.
 723  * 'indexscandir' is ForwardScanDirection or BackwardScanDirection
 724  *                      for an ordered index, or NoMovementScanDirection for
 725  *                      an unordered index.
 726  * 'indexonly' is true if an index-only scan is wanted.
 727  * 'required_outer' is the set of outer relids for a parameterized path.
 728  * 'loop_count' is the number of repetitions of the indexscan to factor into
 729  *              estimates of caching behavior.
 730  *
 731  * Returns the new path node.
 732  */
 733 IndexPath *
 734 create_index_path(PlannerInfo *root,
 735                                   IndexOptInfo *index,
 736                                   List *indexclauses,
 737                                   List *indexclausecols,
 738                                   List *indexorderbys,
 739                                   List *indexorderbycols,
 740                                   List *pathkeys,
 741                                   ScanDirection indexscandir,
 742                                   bool indexonly,
 743                                   Relids required_outer,
 744                                   double loop_count)
 745 {
 746         IndexPath  *pathnode = makeNode(IndexPath);
 747         RelOptInfo *rel = index->rel;
 748         List       *indexquals,
 749                            *indexqualcols;
 750
 751         pathnode->path.pathtype = indexonly ? T_IndexOnlyScan : T_IndexScan;
 752         pathnode->path.parent = rel;
 753         pathnode->path.param_info = get_baserel_parampathinfo(root, rel,
 754                                                                                                                   required_outer);
 755         pathnode->path.pathkeys = pathkeys;
 756
 757         /* Convert clauses to indexquals the executor can handle */
 758         expand_indexqual_conditions(index, indexclauses, indexclausecols,
 759                                                                 &indexquals, &indexqualcols);
 760
 761         /* Fill in the pathnode */
 762         pathnode->indexinfo = index;
 763         pathnode->indexclauses = indexclauses;
 764         pathnode->indexquals = indexquals;
 765         pathnode->indexqualcols = indexqualcols;
 766         pathnode->indexorderbys = indexorderbys;
 767         pathnode->indexorderbycols = indexorderbycols;
 768         pathnode->indexscandir = indexscandir;
 769
 770         cost_index(pathnode, root, loop_count);
 771
 772         return pathnode;
 773 }
 774
 775 /*
 776  * create_bitmap_heap_path
 777  *        Creates a path node for a bitmap scan.
 778  *
 779  * 'bitmapqual' is a tree of IndexPath, BitmapAndPath, and BitmapOrPath nodes.
 780  * 'required_outer' is the set of outer relids for a parameterized path.
 781  * 'loop_count' is the number of repetitions of the indexscan to factor into
 782  *              estimates of caching behavior.
 783  *
 784  * loop_count should match the value used when creating the component
 785  * IndexPaths.
 786  */
 787 BitmapHeapPath *
 788 create_bitmap_heap_path(PlannerInfo *root,
 789                                                 RelOptInfo *rel,
 790                                                 Path *bitmapqual,
 791                                                 Relids required_outer,
 792                                                 double loop_count)
 793 {
 794         BitmapHeapPath *pathnode = makeNode(BitmapHeapPath);
 795
 796         pathnode->path.pathtype = T_BitmapHeapScan;
 797         pathnode->path.parent = rel;
 798         pathnode->path.param_info = get_baserel_parampathinfo(root, rel,
 799                                                                                                                   required_outer);
 800         pathnode->path.pathkeys = NIL;          /* always unordered */
 801
 802         pathnode->bitmapqual = bitmapqual;
 803
 804         cost_bitmap_heap_scan(&pathnode->path, root, rel,
 805                                                   pathnode->path.param_info,
 806                                                   bitmapqual, loop_count);
 807
 808         return pathnode;
 809 }
 810
 811 /*
 812  * create_bitmap_and_path
 813  *        Creates a path node representing a BitmapAnd.
 814  */
 815 BitmapAndPath *
 816 create_bitmap_and_path(PlannerInfo *root,
 817                                            RelOptInfo *rel,
 818                                            List *bitmapquals)
 819 {
 820         BitmapAndPath *pathnode = makeNode(BitmapAndPath);
 821
 822         pathnode->path.pathtype = T_BitmapAnd;
 823         pathnode->path.parent = rel;
 824         pathnode->path.param_info = NULL;       /* not used in bitmap trees */
 825         pathnode->path.pathkeys = NIL;          /* always unordered */
 826
 827         pathnode->bitmapquals = bitmapquals;
 828
 829         /* this sets bitmapselectivity as well as the regular cost fields: */
 830         cost_bitmap_and_node(pathnode, root);
 831
 832         return pathnode;
 833 }
 834
 835 /*
 836  * create_bitmap_or_path
 837  *        Creates a path node representing a BitmapOr.
 838  */
 839 BitmapOrPath *
 840 create_bitmap_or_path(PlannerInfo *root,
 841                                           RelOptInfo *rel,
 842                                           List *bitmapquals)
 843 {
 844         BitmapOrPath *pathnode = makeNode(BitmapOrPath);
 845
 846         pathnode->path.pathtype = T_BitmapOr;
 847         pathnode->path.parent = rel;
 848         pathnode->path.param_info = NULL;       /* not used in bitmap trees */
 849         pathnode->path.pathkeys = NIL;          /* always unordered */
 850
 851         pathnode->bitmapquals = bitmapquals;
 852
 853         /* this sets bitmapselectivity as well as the regular cost fields: */
 854         cost_bitmap_or_node(pathnode, root);
 855
 856         return pathnode;
 857 }
 858
 859 /*
 860  * create_tidscan_path
 861  *        Creates a path corresponding to a scan by TID, returning the pathnode.
 862  */
 863 TidPath *
 864 create_tidscan_path(PlannerInfo *root, RelOptInfo *rel, List *tidquals,
 865                                         Relids required_outer)
 866 {
 867         TidPath    *pathnode = makeNode(TidPath);
 868
 869         pathnode->path.pathtype = T_TidScan;
 870         pathnode->path.parent = rel;
 871         pathnode->path.param_info = get_baserel_parampathinfo(root, rel,
 872                                                                                                                   required_outer);
 873         pathnode->path.pathkeys = NIL;          /* always unordered */
 874
 875         pathnode->tidquals = tidquals;
 876
 877         cost_tidscan(&pathnode->path, root, rel, tidquals,
 878                                  pathnode->path.param_info);
 879
 880         return pathnode;
 881 }
 882
 883 /*
 884  * create_append_path
 885  *        Creates a path corresponding to an Append plan, returning the
 886  *        pathnode.
 887  *
 888  * Note that we must handle subpaths = NIL, representing a dummy access path.
 889  */
 890 AppendPath *
 891 create_append_path(RelOptInfo *rel, List *subpaths, Relids required_outer)
 892 {
 893         AppendPath *pathnode = makeNode(AppendPath);
 894         ListCell   *l;
 895
 896         pathnode->path.pathtype = T_Append;
 897         pathnode->path.parent = rel;
 898         pathnode->path.param_info = get_appendrel_parampathinfo(rel,
 899                                                                                                                         required_outer);
 900         pathnode->path.pathkeys = NIL;          /* result is always considered
 901                                                                                  * unsorted */
 902         pathnode->subpaths = subpaths;
 903
 904         /*
 905          * We don't bother with inventing a cost_append(), but just do it here.
 906          *
 907          * Compute rows and costs as sums of subplan rows and costs.  We charge
 908          * nothing extra for the Append itself, which perhaps is too optimistic,
 909          * but since it doesn't do any selection or projection, it is a pretty
 910          * cheap node.  If you change this, see also make_append().
 911          */
 912         pathnode->path.rows = 0;
 913         pathnode->path.startup_cost = 0;
 914         pathnode->path.total_cost = 0;
 915         foreach(l, subpaths)
 916         {
 917                 Path       *subpath = (Path *) lfirst(l);
 918
 919                 pathnode->path.rows += subpath->rows;
 920
 921                 if (l == list_head(subpaths))   /* first node? */
 922                         pathnode->path.startup_cost = subpath->startup_cost;
 923                 pathnode->path.total_cost += subpath->total_cost;
 924
 925                 /* All child paths must have same parameterization */
 926                 Assert(bms_equal(PATH_REQ_OUTER(subpath), required_outer));
 927         }
 928
 929         return pathnode;
 930 }
 931
 932 /*
 933  * create_merge_append_path
 934  *        Creates a path corresponding to a MergeAppend plan, returning the
 935  *        pathnode.
 936  */
 937 MergeAppendPath *
 938 create_merge_append_path(PlannerInfo *root,
 939                                                  RelOptInfo *rel,
 940                                                  List *subpaths,
 941                                                  List *pathkeys,
 942                                                  Relids required_outer)
 943 {
 944         MergeAppendPath *pathnode = makeNode(MergeAppendPath);
 945         Cost            input_startup_cost;
 946         Cost            input_total_cost;
 947         ListCell   *l;
 948
 949         pathnode->path.pathtype = T_MergeAppend;
 950         pathnode->path.parent = rel;
 951         pathnode->path.param_info = get_appendrel_parampathinfo(rel,
 952                                                                                                                         required_outer);
 953         pathnode->path.pathkeys = pathkeys;
 954         pathnode->subpaths = subpaths;
 955
 956         /*
 957          * Apply query-wide LIMIT if known and path is for sole base relation.
 958          * (Handling this at this low level is a bit klugy.)
 959          */
 960         if (bms_equal(rel->relids, root->all_baserels))
 961                 pathnode->limit_tuples = root->limit_tuples;
 962         else
 963                 pathnode->limit_tuples = -1.0;
 964
 965         /*
 966          * Add up the sizes and costs of the input paths.
 967          */
 968         pathnode->path.rows = 0;
 969         input_startup_cost = 0;
 970         input_total_cost = 0;
 971         foreach(l, subpaths)
 972         {
 973                 Path       *subpath = (Path *) lfirst(l);
 974
 975                 pathnode->path.rows += subpath->rows;
 976
 977                 if (pathkeys_contained_in(pathkeys, subpath->pathkeys))
 978                 {
 979                         /* Subpath is adequately ordered, we won't need to sort it */
 980                         input_startup_cost += subpath->startup_cost;
 981                         input_total_cost += subpath->total_cost;
 982                 }
 983                 else
 984                 {
 985                         /* We'll need to insert a Sort node, so include cost for that */
 986                         Path            sort_path;              /* dummy for result of cost_sort */
 987
 988                         cost_sort(&sort_path,
 989                                           root,
 990                                           pathkeys,
 991                                           subpath->total_cost,
 992                                           subpath->parent->tuples,
 993                                           subpath->parent->width,
 994                                           0.0,
 995                                           work_mem,
 996                                           pathnode->limit_tuples);
 997                         input_startup_cost += sort_path.startup_cost;
 998                         input_total_cost += sort_path.total_cost;
 999                 }
1000
1001                 /* All child paths must have same parameterization */
1002                 Assert(bms_equal(PATH_REQ_OUTER(subpath), required_outer));
1003         }
1004
1005         /* Now we can compute total costs of the MergeAppend */
1006         cost_merge_append(&pathnode->path, root,
1007                                           pathkeys, list_length(subpaths),
1008                                           input_startup_cost, input_total_cost,
1009                                           rel->tuples);
1010
1011         return pathnode;
1012 }
1013
1014 /*
1015  * create_result_path
1016  *        Creates a path representing a Result-and-nothing-else plan.
1017  *        This is only used for the case of a query with an empty jointree.
1018  */
1019 ResultPath *
1020 create_result_path(List *quals)
1021 {
1022         ResultPath *pathnode = makeNode(ResultPath);
1023
1024         pathnode->path.pathtype = T_Result;
1025         pathnode->path.parent = NULL;
1026         pathnode->path.param_info = NULL;       /* there are no other rels... */
1027         pathnode->path.pathkeys = NIL;
1028         pathnode->quals = quals;
1029
1030         /* Hardly worth defining a cost_result() function ... just do it */
1031         pathnode->path.rows = 1;
1032         pathnode->path.startup_cost = 0;
1033         pathnode->path.total_cost = cpu_tuple_cost;
1034
1035         /*
1036          * In theory we should include the qual eval cost as well, but at present
1037          * that doesn't accomplish much except duplicate work that will be done
1038          * again in make_result; since this is only used for degenerate cases,
1039          * nothing interesting will be done with the path cost values...
1040          */
1041
1042         return pathnode;
1043 }
1044
1045 /*
1046  * create_material_path
1047  *        Creates a path corresponding to a Material plan, returning the
1048  *        pathnode.
1049  */
1050 MaterialPath *
1051 create_material_path(RelOptInfo *rel, Path *subpath)
1052 {
1053         MaterialPath *pathnode = makeNode(MaterialPath);
1054
1055         Assert(subpath->parent == rel);
1056
1057         pathnode->path.pathtype = T_Material;
1058         pathnode->path.parent = rel;
1059         pathnode->path.param_info = subpath->param_info;
1060         pathnode->path.pathkeys = subpath->pathkeys;
1061
1062         pathnode->subpath = subpath;
1063
1064         cost_material(&pathnode->path,
1065                                   subpath->startup_cost,
1066                                   subpath->total_cost,
1067                                   subpath->rows,
1068                                   rel->width);
1069
1070         return pathnode;
1071 }
1072
1073 /*
1074  * create_unique_path
1075  *        Creates a path representing elimination of distinct rows from the
1076  *        input data.  Distinct-ness is defined according to the needs of the
1077  *        semijoin represented by sjinfo.  If it is not possible to identify
1078  *        how to make the data unique, NULL is returned.
1079  *
1080  * If used at all, this is likely to be called repeatedly on the same rel;
1081  * and the input subpath should always be the same (the cheapest_total path
1082  * for the rel).  So we cache the result.
1083  */
1084 UniquePath *
1085 create_unique_path(PlannerInfo *root, RelOptInfo *rel, Path *subpath,
1086                                    SpecialJoinInfo *sjinfo)
1087 {
1088         UniquePath *pathnode;
1089         Path            sort_path;              /* dummy for result of cost_sort */
1090         Path            agg_path;               /* dummy for result of cost_agg */
1091         MemoryContext oldcontext;
1092         List       *in_operators;
1093         List       *uniq_exprs;
1094         bool            all_btree;
1095         bool            all_hash;
1096         int                     numCols;
1097         ListCell   *lc;
1098
1099         /* Caller made a mistake if subpath isn't cheapest_total ... */
1100         Assert(subpath == rel->cheapest_total_path);
1101         Assert(subpath->parent == rel);
1102         /* ... or if SpecialJoinInfo is the wrong one */
1103         Assert(sjinfo->jointype == JOIN_SEMI);
1104         Assert(bms_equal(rel->relids, sjinfo->syn_righthand));
1105
1106         /* If result already cached, return it */
1107         if (rel->cheapest_unique_path)
1108                 return (UniquePath *) rel->cheapest_unique_path;
1109
1110         /* If we previously failed, return NULL quickly */
1111         if (sjinfo->join_quals == NIL)
1112                 return NULL;
1113
1114         /*
1115          * We must ensure path struct and subsidiary data are allocated in main
1116          * planning context; otherwise GEQO memory management causes trouble.
1117          */
1118         oldcontext = MemoryContextSwitchTo(root->planner_cxt);
1119
1120         /*----------
1121          * Look to see whether the semijoin's join quals consist of AND'ed
1122          * equality operators, with (only) RHS variables on only one side of
1123          * each one.  If so, we can figure out how to enforce uniqueness for
1124          * the RHS.
1125          *
1126          * Note that the input join_quals list is the list of quals that are
1127          * *syntactically* associated with the semijoin, which in practice means
1128          * the synthesized comparison list for an IN or the WHERE of an EXISTS.
1129          * Particularly in the latter case, it might contain clauses that aren't
1130          * *semantically* associated with the join, but refer to just one side or
1131          * the other.  We can ignore such clauses here, as they will just drop
1132          * down to be processed within one side or the other.  (It is okay to
1133          * consider only the syntactically-associated clauses here because for a
1134          * semijoin, no higher-level quals could refer to the RHS, and so there
1135          * can be no other quals that are semantically associated with this join.
1136          * We do things this way because it is useful to be able to run this test
1137          * before we have extracted the list of quals that are actually
1138          * semantically associated with the particular join.)
1139          *
1140          * Note that the in_operators list consists of the joinqual operators
1141          * themselves (but commuted if needed to put the RHS value on the right).
1142          * These could be cross-type operators, in which case the operator
1143          * actually needed for uniqueness is a related single-type operator.
1144          * We assume here that that operator will be available from the btree
1145          * or hash opclass when the time comes ... if not, create_unique_plan()
1146          * will fail.
1147          *----------
1148          */
1149         in_operators = NIL;
1150         uniq_exprs = NIL;
1151         all_btree = true;
1152         all_hash = enable_hashagg;      /* don't consider hash if not enabled */
1153         foreach(lc, sjinfo->join_quals)
1154         {
1155                 OpExpr     *op = (OpExpr *) lfirst(lc);
1156                 Oid                     opno;
1157                 Node       *left_expr;
1158                 Node       *right_expr;
1159                 Relids          left_varnos;
1160                 Relids          right_varnos;
1161                 Relids          all_varnos;
1162                 Oid                     opinputtype;
1163
1164                 /* Is it a binary opclause? */
1165                 if (!IsA(op, OpExpr) ||
1166                         list_length(op->args) != 2)
1167                 {
1168                         /* No, but does it reference both sides? */
1169                         all_varnos = pull_varnos((Node *) op);
1170                         if (!bms_overlap(all_varnos, sjinfo->syn_righthand) ||
1171                                 bms_is_subset(all_varnos, sjinfo->syn_righthand))
1172                         {
1173                                 /*
1174                                  * Clause refers to only one rel, so ignore it --- unless it
1175                                  * contains volatile functions, in which case we'd better
1176                                  * punt.
1177                                  */
1178                                 if (contain_volatile_functions((Node *) op))
1179                                         goto no_unique_path;
1180                                 continue;
1181                         }
1182                         /* Non-operator clause referencing both sides, must punt */
1183                         goto no_unique_path;
1184                 }
1185
1186                 /* Extract data from binary opclause */
1187                 opno = op->opno;
1188                 left_expr = linitial(op->args);
1189                 right_expr = lsecond(op->args);
1190                 left_varnos = pull_varnos(left_expr);
1191                 right_varnos = pull_varnos(right_expr);
1192                 all_varnos = bms_union(left_varnos, right_varnos);
1193                 opinputtype = exprType(left_expr);
1194
1195                 /* Does it reference both sides? */
1196                 if (!bms_overlap(all_varnos, sjinfo->syn_righthand) ||
1197                         bms_is_subset(all_varnos, sjinfo->syn_righthand))
1198                 {
1199                         /*
1200                          * Clause refers to only one rel, so ignore it --- unless it
1201                          * contains volatile functions, in which case we'd better punt.
1202                          */
1203                         if (contain_volatile_functions((Node *) op))
1204                                 goto no_unique_path;
1205                         continue;
1206                 }
1207
1208                 /* check rel membership of arguments */
1209                 if (!bms_is_empty(right_varnos) &&
1210                         bms_is_subset(right_varnos, sjinfo->syn_righthand) &&
1211                         !bms_overlap(left_varnos, sjinfo->syn_righthand))
1212                 {
1213                         /* typical case, right_expr is RHS variable */
1214                 }
1215                 else if (!bms_is_empty(left_varnos) &&
1216                                  bms_is_subset(left_varnos, sjinfo->syn_righthand) &&
1217                                  !bms_overlap(right_varnos, sjinfo->syn_righthand))
1218                 {
1219                         /* flipped case, left_expr is RHS variable */
1220                         opno = get_commutator(opno);
1221                         if (!OidIsValid(opno))
1222                                 goto no_unique_path;
1223                         right_expr = left_expr;
1224                 }
1225                 else
1226                         goto no_unique_path;
1227
1228                 /* all operators must be btree equality or hash equality */
1229                 if (all_btree)
1230                 {
1231                         /* oprcanmerge is considered a hint... */
1232                         if (!op_mergejoinable(opno, opinputtype) ||
1233                                 get_mergejoin_opfamilies(opno) == NIL)
1234                                 all_btree = false;
1235                 }
1236                 if (all_hash)
1237                 {
1238                         /* ... but oprcanhash had better be correct */
1239                         if (!op_hashjoinable(opno, opinputtype))
1240                                 all_hash = false;
1241                 }
1242                 if (!(all_btree || all_hash))
1243                         goto no_unique_path;
1244
1245                 /* so far so good, keep building lists */
1246                 in_operators = lappend_oid(in_operators, opno);
1247                 uniq_exprs = lappend(uniq_exprs, copyObject(right_expr));
1248         }
1249
1250         /* Punt if we didn't find at least one column to unique-ify */
1251         if (uniq_exprs == NIL)
1252                 goto no_unique_path;
1253
1254         /*
1255          * The expressions we'd need to unique-ify mustn't be volatile.
1256          */
1257         if (contain_volatile_functions((Node *) uniq_exprs))
1258                 goto no_unique_path;
1259
1260         /*
1261          * If we get here, we can unique-ify using at least one of sorting and
1262          * hashing.  Start building the result Path object.
1263          */
1264         pathnode = makeNode(UniquePath);
1265
1266         pathnode->path.pathtype = T_Unique;
1267         pathnode->path.parent = rel;
1268         pathnode->path.param_info = subpath->param_info;
1269
1270         /*
1271          * Assume the output is unsorted, since we don't necessarily have pathkeys
1272          * to represent it.  (This might get overridden below.)
1273          */
1274         pathnode->path.pathkeys = NIL;
1275
1276         pathnode->subpath = subpath;
1277         pathnode->in_operators = in_operators;
1278         pathnode->uniq_exprs = uniq_exprs;
1279
1280         /*
1281          * If the input is a relation and it has a unique index that proves the
1282          * uniq_exprs are unique, then we don't need to do anything.  Note that
1283          * relation_has_unique_index_for automatically considers restriction
1284          * clauses for the rel, as well.
1285          */
1286         if (rel->rtekind == RTE_RELATION && all_btree &&
1287                 relation_has_unique_index_for(root, rel, NIL,
1288                                                                           uniq_exprs, in_operators))
1289         {
1290                 pathnode->umethod = UNIQUE_PATH_NOOP;
1291                 pathnode->path.rows = rel->rows;
1292                 pathnode->path.startup_cost = subpath->startup_cost;
1293                 pathnode->path.total_cost = subpath->total_cost;
1294                 pathnode->path.pathkeys = subpath->pathkeys;
1295
1296                 rel->cheapest_unique_path = (Path *) pathnode;
1297
1298                 MemoryContextSwitchTo(oldcontext);
1299
1300                 return pathnode;
1301         }
1302
1303         /*
1304          * If the input is a subquery whose output must be unique already, then we
1305          * don't need to do anything.  The test for uniqueness has to consider
1306          * exactly which columns we are extracting; for example "SELECT DISTINCT
1307          * x,y" doesn't guarantee that x alone is distinct. So we cannot check for
1308          * this optimization unless uniq_exprs consists only of simple Vars
1309          * referencing subquery outputs.  (Possibly we could do something with
1310          * expressions in the subquery outputs, too, but for now keep it simple.)
1311          */
1312         if (rel->rtekind == RTE_SUBQUERY)
1313         {
1314                 RangeTblEntry *rte = planner_rt_fetch(rel->relid, root);
1315
1316                 if (query_supports_distinctness(rte->subquery))
1317                 {
1318                         List       *sub_tlist_colnos;
1319
1320                         sub_tlist_colnos = translate_sub_tlist(uniq_exprs, rel->relid);
1321
1322                         if (sub_tlist_colnos &&
1323                                 query_is_distinct_for(rte->subquery,
1324                                                                           sub_tlist_colnos, in_operators))
1325                         {
1326                                 pathnode->umethod = UNIQUE_PATH_NOOP;
1327                                 pathnode->path.rows = rel->rows;
1328                                 pathnode->path.startup_cost = subpath->startup_cost;
1329                                 pathnode->path.total_cost = subpath->total_cost;
1330                                 pathnode->path.pathkeys = subpath->pathkeys;
1331
1332                                 rel->cheapest_unique_path = (Path *) pathnode;
1333
1334                                 MemoryContextSwitchTo(oldcontext);
1335
1336                                 return pathnode;
1337                         }
1338                 }
1339         }
1340
1341         /* Estimate number of output rows */
1342         pathnode->path.rows = estimate_num_groups(root, uniq_exprs, rel->rows);
1343         numCols = list_length(uniq_exprs);
1344
1345         if (all_btree)
1346         {
1347                 /*
1348                  * Estimate cost for sort+unique implementation
1349                  */
1350                 cost_sort(&sort_path, root, NIL,
1351                                   subpath->total_cost,
1352                                   rel->rows,
1353                                   rel->width,
1354                                   0.0,
1355                                   work_mem,
1356                                   -1.0);
1357
1358                 /*
1359                  * Charge one cpu_operator_cost per comparison per input tuple. We
1360                  * assume all columns get compared at most of the tuples. (XXX
1361                  * probably this is an overestimate.)  This should agree with
1362                  * make_unique.
1363                  */
1364                 sort_path.total_cost += cpu_operator_cost * rel->rows * numCols;
1365         }
1366
1367         if (all_hash)
1368         {
1369                 /*
1370                  * Estimate the overhead per hashtable entry at 64 bytes (same as in
1371                  * planner.c).
1372                  */
1373                 int                     hashentrysize = rel->width + 64;
1374
1375                 if (hashentrysize * pathnode->path.rows > work_mem * 1024L)
1376                         all_hash = false;       /* don't try to hash */
1377                 else
1378                         cost_agg(&agg_path, root,
1379                                          AGG_HASHED, NULL,
1380                                          numCols, pathnode->path.rows,
1381                                          subpath->startup_cost,
1382                                          subpath->total_cost,
1383                                          rel->rows);
1384         }
1385
1386         if (all_btree && all_hash)
1387         {
1388                 if (agg_path.total_cost < sort_path.total_cost)
1389                         pathnode->umethod = UNIQUE_PATH_HASH;
1390                 else
1391                         pathnode->umethod = UNIQUE_PATH_SORT;
1392         }
1393         else if (all_btree)
1394                 pathnode->umethod = UNIQUE_PATH_SORT;
1395         else if (all_hash)
1396                 pathnode->umethod = UNIQUE_PATH_HASH;
1397         else
1398                 goto no_unique_path;
1399
1400         if (pathnode->umethod == UNIQUE_PATH_HASH)
1401         {
1402                 pathnode->path.startup_cost = agg_path.startup_cost;
1403                 pathnode->path.total_cost = agg_path.total_cost;
1404         }
1405         else
1406         {
1407                 pathnode->path.startup_cost = sort_path.startup_cost;
1408                 pathnode->path.total_cost = sort_path.total_cost;
1409         }
1410
1411         rel->cheapest_unique_path = (Path *) pathnode;
1412
1413         MemoryContextSwitchTo(oldcontext);
1414
1415         return pathnode;
1416
1417 no_unique_path:                 /* failure exit */
1418
1419         /* Mark the SpecialJoinInfo as not unique-able */
1420         sjinfo->join_quals = NIL;
1421
1422         MemoryContextSwitchTo(oldcontext);
1423
1424         return NULL;
1425 }
1426
1427 /*
1428  * translate_sub_tlist - get subquery column numbers represented by tlist
1429  *
1430  * The given targetlist usually contains only Vars referencing the given relid.
1431  * Extract their varattnos (ie, the column numbers of the subquery) and return
1432  * as an integer List.
1433  *
1434  * If any of the tlist items is not a simple Var, we cannot determine whether
1435  * the subquery's uniqueness condition (if any) matches ours, so punt and
1436  * return NIL.
1437  */
1438 static List *
1439 translate_sub_tlist(List *tlist, int relid)
1440 {
1441         List       *result = NIL;
1442         ListCell   *l;
1443
1444         foreach(l, tlist)
1445         {
1446                 Var                *var = (Var *) lfirst(l);
1447
1448                 if (!var || !IsA(var, Var) ||
1449                         var->varno != relid)
1450                         return NIL;                     /* punt */
1451
1452                 result = lappend_int(result, var->varattno);
1453         }
1454         return result;
1455 }
1456
1457 /*
1458  * create_subqueryscan_path
1459  *        Creates a path corresponding to a sequential scan of a subquery,
1460  *        returning the pathnode.
1461  */
1462 Path *
1463 create_subqueryscan_path(PlannerInfo *root, RelOptInfo *rel,
1464                                                  List *pathkeys, Relids required_outer)
1465 {
1466         Path       *pathnode = makeNode(Path);
1467
1468         pathnode->pathtype = T_SubqueryScan;
1469         pathnode->parent = rel;
1470         pathnode->param_info = get_baserel_parampathinfo(root, rel,
1471                                                                                                          required_outer);
1472         pathnode->pathkeys = pathkeys;
1473
1474         cost_subqueryscan(pathnode, root, rel, pathnode->param_info);
1475
1476         return pathnode;
1477 }
1478
1479 /*
1480  * create_functionscan_path
1481  *        Creates a path corresponding to a sequential scan of a function,
1482  *        returning the pathnode.
1483  */
1484 Path *
1485 create_functionscan_path(PlannerInfo *root, RelOptInfo *rel,
1486                                                  List *pathkeys, Relids required_outer)
1487 {
1488         Path       *pathnode = makeNode(Path);
1489
1490         pathnode->pathtype = T_FunctionScan;
1491         pathnode->parent = rel;
1492         pathnode->param_info = get_baserel_parampathinfo(root, rel,
1493                                                                                                          required_outer);
1494         pathnode->pathkeys = pathkeys;
1495
1496         cost_functionscan(pathnode, root, rel, pathnode->param_info);
1497
1498         return pathnode;
1499 }
1500
1501 /*
1502  * create_valuesscan_path
1503  *        Creates a path corresponding to a scan of a VALUES list,
1504  *        returning the pathnode.
1505  */
1506 Path *
1507 create_valuesscan_path(PlannerInfo *root, RelOptInfo *rel,
1508                                            Relids required_outer)
1509 {
1510         Path       *pathnode = makeNode(Path);
1511
1512         pathnode->pathtype = T_ValuesScan;
1513         pathnode->parent = rel;
1514         pathnode->param_info = get_baserel_parampathinfo(root, rel,
1515                                                                                                          required_outer);
1516         pathnode->pathkeys = NIL;       /* result is always unordered */
1517
1518         cost_valuesscan(pathnode, root, rel, pathnode->param_info);
1519
1520         return pathnode;
1521 }
1522
1523 /*
1524  * create_ctescan_path
1525  *        Creates a path corresponding to a scan of a non-self-reference CTE,
1526  *        returning the pathnode.
1527  */
1528 Path *
1529 create_ctescan_path(PlannerInfo *root, RelOptInfo *rel, Relids required_outer)
1530 {
1531         Path       *pathnode = makeNode(Path);
1532
1533         pathnode->pathtype = T_CteScan;
1534         pathnode->parent = rel;
1535         pathnode->param_info = get_baserel_parampathinfo(root, rel,
1536                                                                                                          required_outer);
1537         pathnode->pathkeys = NIL;       /* XXX for now, result is always unordered */
1538
1539         cost_ctescan(pathnode, root, rel, pathnode->param_info);
1540
1541         return pathnode;
1542 }
1543
1544 /*
1545  * create_worktablescan_path
1546  *        Creates a path corresponding to a scan of a self-reference CTE,
1547  *        returning the pathnode.
1548  */
1549 Path *
1550 create_worktablescan_path(PlannerInfo *root, RelOptInfo *rel,
1551                                                   Relids required_outer)
1552 {
1553         Path       *pathnode = makeNode(Path);
1554
1555         pathnode->pathtype = T_WorkTableScan;
1556         pathnode->parent = rel;
1557         pathnode->param_info = get_baserel_parampathinfo(root, rel,
1558                                                                                                          required_outer);
1559         pathnode->pathkeys = NIL;       /* result is always unordered */
1560
1561         /* Cost is the same as for a regular CTE scan */
1562         cost_ctescan(pathnode, root, rel, pathnode->param_info);
1563
1564         return pathnode;
1565 }
1566
1567 /*
1568  * create_foreignscan_path
1569  *        Creates a path corresponding to a scan of a foreign table,
1570  *        returning the pathnode.
1571  *
1572  * This function is never called from core Postgres; rather, it's expected
1573  * to be called by the GetForeignPaths function of a foreign data wrapper.
1574  * We make the FDW supply all fields of the path, since we do not have any
1575  * way to calculate them in core.
1576  */
1577 ForeignPath *
1578 create_foreignscan_path(PlannerInfo *root, RelOptInfo *rel,
1579                                                 double rows, Cost startup_cost, Cost total_cost,
1580                                                 List *pathkeys,
1581                                                 Relids required_outer,
1582                                                 List *fdw_private)
1583 {
1584         ForeignPath *pathnode = makeNode(ForeignPath);
1585
1586         pathnode->path.pathtype = T_ForeignScan;
1587         pathnode->path.parent = rel;
1588         pathnode->path.param_info = get_baserel_parampathinfo(root, rel,
1589                                                                                                                   required_outer);
1590         pathnode->path.rows = rows;
1591         pathnode->path.startup_cost = startup_cost;
1592         pathnode->path.total_cost = total_cost;
1593         pathnode->path.pathkeys = pathkeys;
1594
1595         pathnode->fdw_private = fdw_private;
1596
1597         return pathnode;
1598 }
1599
1600 /*
1601  * calc_nestloop_required_outer
1602  *        Compute the required_outer set for a nestloop join path
1603  *
1604  * Note: result must not share storage with either input
1605  */
1606 Relids
1607 calc_nestloop_required_outer(Path *outer_path, Path *inner_path)
1608 {
1609         Relids          outer_paramrels = PATH_REQ_OUTER(outer_path);
1610         Relids          inner_paramrels = PATH_REQ_OUTER(inner_path);
1611         Relids          required_outer;
1612
1613         /* inner_path can require rels from outer path, but not vice versa */
1614         Assert(!bms_overlap(outer_paramrels, inner_path->parent->relids));
1615         /* easy case if inner path is not parameterized */
1616         if (!inner_paramrels)
1617                 return bms_copy(outer_paramrels);
1618         /* else, form the union ... */
1619         required_outer = bms_union(outer_paramrels, inner_paramrels);
1620         /* ... and remove any mention of now-satisfied outer rels */
1621         required_outer = bms_del_members(required_outer,
1622                                                                          outer_path->parent->relids);
1623         /* maintain invariant that required_outer is exactly NULL if empty */
1624         if (bms_is_empty(required_outer))
1625         {
1626                 bms_free(required_outer);
1627                 required_outer = NULL;
1628         }
1629         return required_outer;
1630 }
1631
1632 /*
1633  * calc_non_nestloop_required_outer
1634  *        Compute the required_outer set for a merge or hash join path
1635  *
1636  * Note: result must not share storage with either input
1637  */
1638 Relids
1639 calc_non_nestloop_required_outer(Path *outer_path, Path *inner_path)
1640 {
1641         Relids          outer_paramrels = PATH_REQ_OUTER(outer_path);
1642         Relids          inner_paramrels = PATH_REQ_OUTER(inner_path);
1643         Relids          required_outer;
1644
1645         /* neither path can require rels from the other */
1646         Assert(!bms_overlap(outer_paramrels, inner_path->parent->relids));
1647         Assert(!bms_overlap(inner_paramrels, outer_path->parent->relids));
1648         /* form the union ... */
1649         required_outer = bms_union(outer_paramrels, inner_paramrels);
1650         /* we do not need an explicit test for empty; bms_union gets it right */
1651         return required_outer;
1652 }
1653
1654 /*
1655  * create_nestloop_path
1656  *        Creates a pathnode corresponding to a nestloop join between two
1657  *        relations.
1658  *
1659  * 'joinrel' is the join relation.
1660  * 'jointype' is the type of join required
1661  * 'workspace' is the result from initial_cost_nestloop
1662  * 'sjinfo' is extra info about the join for selectivity estimation
1663  * 'semifactors' contains valid data if jointype is SEMI or ANTI
1664  * 'outer_path' is the outer path
1665  * 'inner_path' is the inner path
1666  * 'restrict_clauses' are the RestrictInfo nodes to apply at the join
1667  * 'pathkeys' are the path keys of the new join path
1668  * 'required_outer' is the set of required outer rels
1669  *
1670  * Returns the resulting path node.
1671  */
1672 NestPath *
1673 create_nestloop_path(PlannerInfo *root,
1674                                          RelOptInfo *joinrel,
1675                                          JoinType jointype,
1676                                          JoinCostWorkspace *workspace,
1677                                          SpecialJoinInfo *sjinfo,
1678                                          SemiAntiJoinFactors *semifactors,
1679                                          Path *outer_path,
1680                                          Path *inner_path,
1681                                          List *restrict_clauses,
1682                                          List *pathkeys,
1683                                          Relids required_outer)
1684 {
1685         NestPath   *pathnode = makeNode(NestPath);
1686         Relids          inner_req_outer = PATH_REQ_OUTER(inner_path);
1687
1688         /*
1689          * If the inner path is parameterized by the outer, we must drop any
1690          * restrict_clauses that are due to be moved into the inner path.  We have
1691          * to do this now, rather than postpone the work till createplan time,
1692          * because the restrict_clauses list can affect the size and cost
1693          * estimates for this path.
1694          */
1695         if (bms_overlap(inner_req_outer, outer_path->parent->relids))
1696         {
1697                 Relids          inner_and_outer = bms_union(inner_path->parent->relids,
1698                                                                                                 inner_req_outer);
1699                 List       *jclauses = NIL;
1700                 ListCell   *lc;
1701
1702                 foreach(lc, restrict_clauses)
1703                 {
1704                         RestrictInfo *rinfo = (RestrictInfo *) lfirst(lc);
1705
1706                         if (!join_clause_is_movable_into(rinfo,
1707                                                                                          inner_path->parent->relids,
1708                                                                                          inner_and_outer))
1709                                 jclauses = lappend(jclauses, rinfo);
1710                 }
1711                 restrict_clauses = jclauses;
1712         }
1713
1714         pathnode->path.pathtype = T_NestLoop;
1715         pathnode->path.parent = joinrel;
1716         pathnode->path.param_info =
1717                 get_joinrel_parampathinfo(root,
1718                                                                   joinrel,
1719                                                                   outer_path,
1720                                                                   inner_path,
1721                                                                   sjinfo,
1722                                                                   required_outer,
1723                                                                   &restrict_clauses);
1724         pathnode->path.pathkeys = pathkeys;
1725         pathnode->jointype = jointype;
1726         pathnode->outerjoinpath = outer_path;
1727         pathnode->innerjoinpath = inner_path;
1728         pathnode->joinrestrictinfo = restrict_clauses;
1729
1730         final_cost_nestloop(root, pathnode, workspace, sjinfo, semifactors);
1731
1732         return pathnode;
1733 }
1734
1735 /*
1736  * create_mergejoin_path
1737  *        Creates a pathnode corresponding to a mergejoin join between
1738  *        two relations
1739  *
1740  * 'joinrel' is the join relation
1741  * 'jointype' is the type of join required
1742  * 'workspace' is the result from initial_cost_mergejoin
1743  * 'sjinfo' is extra info about the join for selectivity estimation
1744  * 'outer_path' is the outer path
1745  * 'inner_path' is the inner path
1746  * 'restrict_clauses' are the RestrictInfo nodes to apply at the join
1747  * 'pathkeys' are the path keys of the new join path
1748  * 'required_outer' is the set of required outer rels
1749  * 'mergeclauses' are the RestrictInfo nodes to use as merge clauses
1750  *              (this should be a subset of the restrict_clauses list)
1751  * 'outersortkeys' are the sort varkeys for the outer relation
1752  * 'innersortkeys' are the sort varkeys for the inner relation
1753  */
1754 MergePath *
1755 create_mergejoin_path(PlannerInfo *root,
1756                                           RelOptInfo *joinrel,
1757                                           JoinType jointype,
1758                                           JoinCostWorkspace *workspace,
1759                                           SpecialJoinInfo *sjinfo,
1760                                           Path *outer_path,
1761                                           Path *inner_path,
1762                                           List *restrict_clauses,
1763                                           List *pathkeys,
1764                                           Relids required_outer,
1765                                           List *mergeclauses,
1766                                           List *outersortkeys,
1767                                           List *innersortkeys)
1768 {
1769         MergePath  *pathnode = makeNode(MergePath);
1770
1771         pathnode->jpath.path.pathtype = T_MergeJoin;
1772         pathnode->jpath.path.parent = joinrel;
1773         pathnode->jpath.path.param_info =
1774                 get_joinrel_parampathinfo(root,
1775                                                                   joinrel,
1776                                                                   outer_path,
1777                                                                   inner_path,
1778                                                                   sjinfo,
1779                                                                   required_outer,
1780                                                                   &restrict_clauses);
1781         pathnode->jpath.path.pathkeys = pathkeys;
1782         pathnode->jpath.jointype = jointype;
1783         pathnode->jpath.outerjoinpath = outer_path;
1784         pathnode->jpath.innerjoinpath = inner_path;
1785         pathnode->jpath.joinrestrictinfo = restrict_clauses;
1786         pathnode->path_mergeclauses = mergeclauses;
1787         pathnode->outersortkeys = outersortkeys;
1788         pathnode->innersortkeys = innersortkeys;
1789         /* pathnode->materialize_inner will be set by final_cost_mergejoin */
1790
1791         final_cost_mergejoin(root, pathnode, workspace, sjinfo);
1792
1793         return pathnode;
1794 }
1795
1796 /*
1797  * create_hashjoin_path
1798  *        Creates a pathnode corresponding to a hash join between two relations.
1799  *
1800  * 'joinrel' is the join relation
1801  * 'jointype' is the type of join required
1802  * 'workspace' is the result from initial_cost_hashjoin
1803  * 'sjinfo' is extra info about the join for selectivity estimation
1804  * 'semifactors' contains valid data if jointype is SEMI or ANTI
1805  * 'outer_path' is the cheapest outer path
1806  * 'inner_path' is the cheapest inner path
1807  * 'restrict_clauses' are the RestrictInfo nodes to apply at the join
1808  * 'required_outer' is the set of required outer rels
1809  * 'hashclauses' are the RestrictInfo nodes to use as hash clauses
1810  *              (this should be a subset of the restrict_clauses list)
1811  */
1812 HashPath *
1813 create_hashjoin_path(PlannerInfo *root,
1814                                          RelOptInfo *joinrel,
1815                                          JoinType jointype,
1816                                          JoinCostWorkspace *workspace,
1817                                          SpecialJoinInfo *sjinfo,
1818                                          SemiAntiJoinFactors *semifactors,
1819                                          Path *outer_path,
1820                                          Path *inner_path,
1821                                          List *restrict_clauses,
1822                                          Relids required_outer,
1823                                          List *hashclauses)
1824 {
1825         HashPath   *pathnode = makeNode(HashPath);
1826
1827         pathnode->jpath.path.pathtype = T_HashJoin;
1828         pathnode->jpath.path.parent = joinrel;
1829         pathnode->jpath.path.param_info =
1830                 get_joinrel_parampathinfo(root,
1831                                                                   joinrel,
1832                                                                   outer_path,
1833                                                                   inner_path,
1834                                                                   sjinfo,
1835                                                                   required_outer,
1836                                                                   &restrict_clauses);
1837
1838         /*
1839          * A hashjoin never has pathkeys, since its output ordering is
1840          * unpredictable due to possible batching.  XXX If the inner relation is
1841          * small enough, we could instruct the executor that it must not batch,
1842          * and then we could assume that the output inherits the outer relation's
1843          * ordering, which might save a sort step.  However there is considerable
1844          * downside if our estimate of the inner relation size is badly off. For
1845          * the moment we don't risk it.  (Note also that if we wanted to take this
1846          * seriously, joinpath.c would have to consider many more paths for the
1847          * outer rel than it does now.)
1848          */
1849         pathnode->jpath.path.pathkeys = NIL;
1850         pathnode->jpath.jointype = jointype;
1851         pathnode->jpath.outerjoinpath = outer_path;
1852         pathnode->jpath.innerjoinpath = inner_path;
1853         pathnode->jpath.joinrestrictinfo = restrict_clauses;
1854         pathnode->path_hashclauses = hashclauses;
1855         /* final_cost_hashjoin will fill in pathnode->num_batches */
1856
1857         final_cost_hashjoin(root, pathnode, workspace, sjinfo, semifactors);
1858
1859         return pathnode;
1860 }
1861
1862 /*
1863  * reparameterize_path
1864  *              Attempt to modify a Path to have greater parameterization
1865  *
1866  * We use this to attempt to bring all child paths of an appendrel to the
1867  * same parameterization level, ensuring that they all enforce the same set
1868  * of join quals (and thus that that parameterization can be attributed to
1869  * an append path built from such paths).  Currently, only a few path types
1870  * are supported here, though more could be added at need.  We return NULL
1871  * if we can't reparameterize the given path.
1872  *
1873  * Note: we intentionally do not pass created paths to add_path(); it would
1874  * possibly try to delete them on the grounds of being cost-inferior to the
1875  * paths they were made from, and we don't want that.  Paths made here are
1876  * not necessarily of general-purpose usefulness, but they can be useful
1877  * as members of an append path.
1878  */
1879 Path *
1880 reparameterize_path(PlannerInfo *root, Path *path,
1881                                         Relids required_outer,
1882                                         double loop_count)
1883 {
1884         RelOptInfo *rel = path->parent;
1885
1886         /* Can only increase, not decrease, path's parameterization */
1887         if (!bms_is_subset(PATH_REQ_OUTER(path), required_outer))
1888                 return NULL;
1889         switch (path->pathtype)
1890         {
1891                 case T_SeqScan:
1892                         return create_seqscan_path(root, rel, required_outer);
1893                 case T_IndexScan:
1894                 case T_IndexOnlyScan:
1895                         {
1896                                 IndexPath  *ipath = (IndexPath *) path;
1897                                 IndexPath  *newpath = makeNode(IndexPath);
1898
1899                                 /*
1900                                  * We can't use create_index_path directly, and would not want
1901                                  * to because it would re-compute the indexqual conditions
1902                                  * which is wasted effort.  Instead we hack things a bit:
1903                                  * flat-copy the path node, revise its param_info, and redo
1904                                  * the cost estimate.
1905                                  */
1906                                 memcpy(newpath, ipath, sizeof(IndexPath));
1907                                 newpath->path.param_info =
1908                                         get_baserel_parampathinfo(root, rel, required_outer);
1909                                 cost_index(newpath, root, loop_count);
1910                                 return (Path *) newpath;
1911                         }
1912                 case T_BitmapHeapScan:
1913                         {
1914                                 BitmapHeapPath *bpath = (BitmapHeapPath *) path;
1915
1916                                 return (Path *) create_bitmap_heap_path(root,
1917                                                                                                                 rel,
1918                                                                                                                 bpath->bitmapqual,
1919                                                                                                                 required_outer,
1920                                                                                                                 loop_count);
1921                         }
1922                 case T_SubqueryScan:
1923                         return create_subqueryscan_path(root, rel, path->pathkeys,
1924                                                                                         required_outer);
1925                 default:
1926                         break;
1927         }
1928         return NULL;
1929 }
1930
1931 /*****************************************************************************
1932  *     creation of custom-plan paths
1933  *****************************************************************************/
1934
1935 static List        *custom_path_providers = NIL;
1936
1937 /*
1938  * register_custom_path_provider
1939  *
1940  * Register a table of callback functions which implements a custom-path
1941  * provider.  This allows extension to provide additional (hopefully faster)
1942  * methods of scanning a relation.
1943  */
1944 void
1945 register_custom_path_provider(CustomPathMethods *cpp_methods)
1946 {
1947         MemoryContext   oldcxt;
1948
1949         oldcxt = MemoryContextSwitchTo(TopMemoryContext);
1950         custom_path_providers = lappend(custom_path_providers, cpp_methods);
1951         MemoryContextSwitchTo(oldcxt);
1952 }
1953
1954 /*
1955  * create_customscan_paths
1956  *
1957  * Invoke custom path provider callbacks.  If the callback determines that
1958  * the custom-path provider can handle this relation, it can add one or more
1959  * paths using add_path().
1960  */
1961 void
1962 create_customscan_paths(PlannerInfo *root,
1963                                                 RelOptInfo *baserel,
1964                                                 RangeTblEntry *rte)
1965 {
1966         ListCell           *cell;
1967
1968         foreach (cell, custom_path_providers)
1969         {
1970                 const CustomPathMethods *cpp_methods = lfirst(cell);
1971
1972                 if (cpp_methods->CreateCustomScanPath)
1973                         cpp_methods->CreateCustomScanPath(root, baserel, rte);
1974         }
1975 }