1 /*-------------------------------------------------------------------------
4 * Definitions for planner's internal data structures.
7 * Portions Copyright (c) 1996-2007, PostgreSQL Global Development Group
8 * Portions Copyright (c) 1994, Regents of the University of California
10 * $PostgreSQL: pgsql/src/include/nodes/relation.h,v 1.135 2007/02/16 20:57:19 tgl Exp $
12 *-------------------------------------------------------------------------
17 #include "access/sdir.h"
18 #include "nodes/bitmapset.h"
19 #include "nodes/parsenodes.h"
20 #include "storage/block.h"
25 * Set of relation identifiers (indexes into the rangetable).
27 typedef Bitmapset *Relids;
30 * When looking for a "cheapest path", this enum specifies whether we want
31 * cheapest startup cost or cheapest total cost.
33 typedef enum CostSelector
35 STARTUP_COST, TOTAL_COST
39 * The cost estimate produced by cost_qual_eval() includes both a one-time
40 * (startup) cost, and a per-tuple cost.
42 typedef struct QualCost
44 Cost startup; /* one-time cost */
45 Cost per_tuple; /* per-evaluation cost */
51 * Per-query information for planning/optimization
53 * This struct is conventionally called "root" in all the planner routines.
54 * It holds links to all of the planner's working state, in addition to the
55 * original Query. Note that at present the planner extensively modifies
56 * the passed-in Query data structure; someday that should stop.
59 typedef struct PlannerInfo
63 Query *parse; /* the Query being planned */
66 * simple_rel_array holds pointers to "base rels" and "other rels" (see
67 * comments for RelOptInfo for more info). It is indexed by rangetable
68 * index (so entry 0 is always wasted). Entries can be NULL when an RTE
69 * does not correspond to a base relation, such as a join RTE or an
70 * unreferenced view RTE; or if the RelOptInfo hasn't been made yet.
72 struct RelOptInfo **simple_rel_array; /* All 1-rel RelOptInfos */
73 int simple_rel_array_size; /* allocated size of array */
76 * join_rel_list is a list of all join-relation RelOptInfos we have
77 * considered in this planning run. For small problems we just scan the
78 * list to do lookups, but when there are many join relations we build a
79 * hash table for faster lookups. The hash table is present and valid
80 * when join_rel_hash is not NULL. Note that we still maintain the list
81 * even when using the hash table for lookups; this simplifies life for
84 List *join_rel_list; /* list of join-relation RelOptInfos */
85 struct HTAB *join_rel_hash; /* optional hashtable for join relations */
87 List *eq_classes; /* list of active EquivalenceClasses */
89 List *canon_pathkeys; /* list of "canonical" PathKeys */
91 List *left_join_clauses; /* list of RestrictInfos for
92 * mergejoinable outer join clauses
93 * w/nonnullable var on left */
95 List *right_join_clauses; /* list of RestrictInfos for
96 * mergejoinable outer join clauses
97 * w/nonnullable var on right */
99 List *full_join_clauses; /* list of RestrictInfos for
100 * mergejoinable full join clauses */
102 List *oj_info_list; /* list of OuterJoinInfos */
104 List *in_info_list; /* list of InClauseInfos */
106 List *append_rel_list; /* list of AppendRelInfos */
108 List *query_pathkeys; /* desired pathkeys for query_planner(), and
109 * actual pathkeys afterwards */
111 List *group_pathkeys; /* groupClause pathkeys, if any */
112 List *sort_pathkeys; /* sortClause pathkeys, if any */
114 MemoryContext planner_cxt; /* context holding PlannerInfo */
116 double total_table_pages; /* # of pages in all tables of query */
118 double tuple_fraction; /* tuple_fraction passed to query_planner */
120 bool hasJoinRTEs; /* true if any RTEs are RTE_JOIN kind */
121 bool hasOuterJoins; /* true if any RTEs are outer joins */
122 bool hasHavingQual; /* true if havingQual was non-null */
123 bool hasPseudoConstantQuals; /* true if any RestrictInfo has
124 * pseudoconstant = true */
130 * Per-relation information for planning/optimization
132 * For planning purposes, a "base rel" is either a plain relation (a table)
133 * or the output of a sub-SELECT or function that appears in the range table.
134 * In either case it is uniquely identified by an RT index. A "joinrel"
135 * is the joining of two or more base rels. A joinrel is identified by
136 * the set of RT indexes for its component baserels. We create RelOptInfo
137 * nodes for each baserel and joinrel, and store them in the PlannerInfo's
138 * simple_rel_array and join_rel_list respectively.
140 * Note that there is only one joinrel for any given set of component
141 * baserels, no matter what order we assemble them in; so an unordered
142 * set is the right datatype to identify it with.
144 * We also have "other rels", which are like base rels in that they refer to
145 * single RT indexes; but they are not part of the join tree, and are given
146 * a different RelOptKind to identify them.
148 * Currently the only kind of otherrels are those made for member relations
149 * of an "append relation", that is an inheritance set or UNION ALL subquery.
150 * An append relation has a parent RTE that is a base rel, which represents
151 * the entire append relation. The member RTEs are otherrels. The parent
152 * is present in the query join tree but the members are not. The member
153 * RTEs and otherrels are used to plan the scans of the individual tables or
154 * subqueries of the append set; then the parent baserel is given an Append
155 * plan comprising the best plans for the individual member rels. (See
156 * comments for AppendRelInfo for more information.)
158 * At one time we also made otherrels to represent join RTEs, for use in
159 * handling join alias Vars. Currently this is not needed because all join
160 * alias Vars are expanded to non-aliased form during preprocess_expression.
162 * Parts of this data structure are specific to various scan and join
163 * mechanisms. It didn't seem worth creating new node types for them.
165 * relids - Set of base-relation identifiers; it is a base relation
166 * if there is just one, a join relation if more than one
167 * rows - estimated number of tuples in the relation after restriction
168 * clauses have been applied (ie, output rows of a plan for it)
169 * width - avg. number of bytes per tuple in the relation after the
170 * appropriate projections have been done (ie, output width)
171 * reltargetlist - List of Var nodes for the attributes we need to
172 * output from this relation (in no particular order)
173 * NOTE: in a child relation, may contain RowExprs
174 * pathlist - List of Path nodes, one for each potentially useful
175 * method of generating the relation
176 * cheapest_startup_path - the pathlist member with lowest startup cost
177 * (regardless of its ordering)
178 * cheapest_total_path - the pathlist member with lowest total cost
179 * (regardless of its ordering)
180 * cheapest_unique_path - for caching cheapest path to produce unique
181 * (no duplicates) output from relation
183 * If the relation is a base relation it will have these fields set:
185 * relid - RTE index (this is redundant with the relids field, but
186 * is provided for convenience of access)
187 * rtekind - distinguishes plain relation, subquery, or function RTE
188 * min_attr, max_attr - range of valid AttrNumbers for rel
189 * attr_needed - array of bitmapsets indicating the highest joinrel
190 * in which each attribute is needed; if bit 0 is set then
191 * the attribute is needed as part of final targetlist
192 * attr_widths - cache space for per-attribute width estimates;
193 * zero means not computed yet
194 * indexlist - list of IndexOptInfo nodes for relation's indexes
195 * (always NIL if it's not a table)
196 * pages - number of disk pages in relation (zero if not a table)
197 * tuples - number of tuples in relation (not considering restrictions)
198 * subplan - plan for subquery (NULL if it's not a subquery)
200 * Note: for a subquery, tuples and subplan are not set immediately
201 * upon creation of the RelOptInfo object; they are filled in when
202 * set_base_rel_pathlist processes the object.
204 * For otherrels that are appendrel members, these fields are filled
205 * in just as for a baserel.
207 * The presence of the remaining fields depends on the restrictions
208 * and joins that the relation participates in:
210 * baserestrictinfo - List of RestrictInfo nodes, containing info about
211 * each non-join qualification clause in which this relation
212 * participates (only used for base rels)
213 * baserestrictcost - Estimated cost of evaluating the baserestrictinfo
214 * clauses at a single tuple (only used for base rels)
215 * joininfo - List of RestrictInfo nodes, containing info about each
216 * join clause in which this relation participates (but
217 * note this excludes clauses that might be derivable from
218 * EquivalenceClasses)
219 * has_eclass_joins - flag that EquivalenceClass joins are possible
220 * index_outer_relids - only used for base rels; set of outer relids
221 * that participate in indexable joinclauses for this rel
222 * index_inner_paths - only used for base rels; list of InnerIndexscanInfo
223 * nodes showing best indexpaths for various subsets of
224 * index_outer_relids.
226 * Note: Keeping a restrictinfo list in the RelOptInfo is useful only for
227 * base rels, because for a join rel the set of clauses that are treated as
228 * restrict clauses varies depending on which sub-relations we choose to join.
229 * (For example, in a 3-base-rel join, a clause relating rels 1 and 2 must be
230 * treated as a restrictclause if we join {1} and {2 3} to make {1 2 3}; but
231 * if we join {1 2} and {3} then that clause will be a restrictclause in {1 2}
232 * and should not be processed again at the level of {1 2 3}.) Therefore,
233 * the restrictinfo list in the join case appears in individual JoinPaths
234 * (field joinrestrictinfo), not in the parent relation. But it's OK for
235 * the RelOptInfo to store the joininfo list, because that is the same
236 * for a given rel no matter how we form it.
238 * We store baserestrictcost in the RelOptInfo (for base relations) because
239 * we know we will need it at least once (to price the sequential scan)
240 * and may need it multiple times to price index scans.
243 typedef enum RelOptKind
247 RELOPT_OTHER_MEMBER_REL
250 typedef struct RelOptInfo
254 RelOptKind reloptkind;
256 /* all relations included in this RelOptInfo */
257 Relids relids; /* set of base relids (rangetable indexes) */
259 /* size estimates generated by planner */
260 double rows; /* estimated number of result tuples */
261 int width; /* estimated avg width of result tuples */
263 /* materialization information */
264 List *reltargetlist; /* needed Vars */
265 List *pathlist; /* Path structures */
266 struct Path *cheapest_startup_path;
267 struct Path *cheapest_total_path;
268 struct Path *cheapest_unique_path;
270 /* information about a base rel (not set for join rels!) */
272 RTEKind rtekind; /* RELATION, SUBQUERY, or FUNCTION */
273 AttrNumber min_attr; /* smallest attrno of rel (often <0) */
274 AttrNumber max_attr; /* largest attrno of rel */
275 Relids *attr_needed; /* array indexed [min_attr .. max_attr] */
276 int32 *attr_widths; /* array indexed [min_attr .. max_attr] */
280 struct Plan *subplan; /* if subquery */
282 /* used by various scans and joins: */
283 List *baserestrictinfo; /* RestrictInfo structures (if base
285 QualCost baserestrictcost; /* cost of evaluating the above */
286 List *joininfo; /* RestrictInfo structures for join clauses
287 * involving this rel */
288 bool has_eclass_joins; /* T means joininfo is incomplete */
290 /* cached info about inner indexscan paths for relation: */
291 Relids index_outer_relids; /* other relids in indexable join
293 List *index_inner_paths; /* InnerIndexscanInfo nodes */
296 * Inner indexscans are not in the main pathlist because they are not
297 * usable except in specific join contexts. We use the index_inner_paths
298 * list just to avoid recomputing the best inner indexscan repeatedly for
299 * similar outer relations. See comments for InnerIndexscanInfo.
305 * Per-index information for planning/optimization
307 * Prior to Postgres 7.0, RelOptInfo was used to describe both relations
308 * and indexes, but that created confusion without actually doing anything
309 * useful. So now we have a separate IndexOptInfo struct for indexes.
311 * opfamily[], indexkeys[], fwdsortop[], revsortop[], and nulls_first[]
312 * each have ncolumns entries. Note: for historical reasons, the
313 * opfamily array has an extra entry that is always zero. Some code
314 * scans until it sees a zero entry, rather than looking at ncolumns.
316 * Zeroes in the indexkeys[] array indicate index columns that are
317 * expressions; there is one element in indexprs for each such column.
319 * For an unordered index, the sortop arrays contains zeroes. Note that
320 * fwdsortop[] and nulls_first[] describe the sort ordering of a forward
321 * indexscan; we can also consider a backward indexscan, which will
322 * generate sort order described by revsortop/!nulls_first.
324 * The indexprs and indpred expressions have been run through
325 * prepqual.c and eval_const_expressions() for ease of matching to
326 * WHERE clauses. indpred is in implicit-AND form.
328 typedef struct IndexOptInfo
332 Oid indexoid; /* OID of the index relation */
333 RelOptInfo *rel; /* back-link to index's table */
335 /* statistics from pg_class */
336 BlockNumber pages; /* number of disk pages in index */
337 double tuples; /* number of index tuples in index */
339 /* index descriptor information */
340 int ncolumns; /* number of columns in index */
341 Oid *opfamily; /* OIDs of operator families for columns */
342 int *indexkeys; /* column numbers of index's keys, or 0 */
343 Oid *fwdsortop; /* OIDs of sort operators for each column */
344 Oid *revsortop; /* OIDs of sort operators for backward scan */
345 bool *nulls_first; /* do NULLs come first in the sort order? */
346 Oid relam; /* OID of the access method (in pg_am) */
348 RegProcedure amcostestimate; /* OID of the access method's cost fcn */
350 List *indexprs; /* expressions for non-simple index columns */
351 List *indpred; /* predicate if a partial index, else NIL */
353 bool predOK; /* true if predicate matches query */
354 bool unique; /* true if a unique index */
355 bool amoptionalkey; /* can query omit key for the first column? */
362 * Whenever we can determine that a mergejoinable equality clause A = B is
363 * not delayed by any outer join, we create an EquivalenceClass containing
364 * the expressions A and B to record this knowledge. If we later find another
365 * equivalence B = C, we add C to the existing EquivalenceClass; this may
366 * require merging two existing EquivalenceClasses. At the end of the qual
367 * distribution process, we have sets of values that are known all transitively
368 * equal to each other, where "equal" is according to the rules of the btree
369 * operator family(s) shown in ec_opfamilies. (We restrict an EC to contain
370 * only equalities whose operators belong to the same set of opfamilies. This
371 * could probably be relaxed, but for now it's not worth the trouble, since
372 * nearly all equality operators belong to only one btree opclass anyway.)
374 * We also use EquivalenceClasses as the base structure for PathKeys, letting
375 * us represent knowledge about different sort orderings being equivalent.
376 * Since every PathKey must reference an EquivalenceClass, we will end up
377 * with single-member EquivalenceClasses whenever a sort key expression has
378 * not been equivalenced to anything else. It is also possible that such an
379 * EquivalenceClass will contain a volatile expression ("ORDER BY random()"),
380 * which is a case that can't arise otherwise since clauses containing
381 * volatile functions are never considered mergejoinable. We mark such
382 * EquivalenceClasses specially to prevent them from being merged with
383 * ordinary EquivalenceClasses.
385 * We allow equality clauses appearing below the nullable side of an outer join
386 * to form EquivalenceClasses, but these have a slightly different meaning:
387 * the included values might be all NULL rather than all the same non-null
388 * values. See src/backend/optimizer/README for more on that point.
390 * NB: if ec_merged isn't NULL, this class has been merged into another, and
391 * should be ignored in favor of using the pointed-to class.
393 typedef struct EquivalenceClass
397 List *ec_opfamilies; /* btree operator family OIDs */
398 List *ec_members; /* list of EquivalenceMembers */
399 List *ec_sources; /* list of generating RestrictInfos */
400 List *ec_derives; /* list of derived RestrictInfos */
401 Relids ec_relids; /* all relids appearing in ec_members */
402 bool ec_has_const; /* any pseudoconstants in ec_members? */
403 bool ec_has_volatile; /* the (sole) member is a volatile expr */
404 bool ec_below_outer_join; /* equivalence applies below an OJ */
405 bool ec_broken; /* failed to generate needed clauses? */
406 struct EquivalenceClass *ec_merged; /* set if merged into another EC */
410 * EquivalenceMember - one member expression of an EquivalenceClass
412 * em_is_child signifies that this element was built by transposing a member
413 * for an inheritance parent relation to represent the corresponding expression
414 * on an inheritance child. The element should be ignored for all purposes
415 * except constructing inner-indexscan paths for the child relation. (Other
416 * types of join are driven from transposed joininfo-list entries.) Note
417 * that the EC's ec_relids field does NOT include the child relation.
419 * em_datatype is usually the same as exprType(em_expr), but can be
420 * different when dealing with a binary-compatible opfamily; in particular
421 * anyarray_ops would never work without this. Use em_datatype when
422 * looking up a specific btree operator to work with this expression.
424 typedef struct EquivalenceMember
428 Expr *em_expr; /* the expression represented */
429 Relids em_relids; /* all relids appearing in em_expr */
430 bool em_is_const; /* expression is pseudoconstant? */
431 bool em_is_child; /* derived version for a child relation? */
432 Oid em_datatype; /* the "nominal type" used by the opfamily */
438 * The sort ordering of a path is represented by a list of PathKey nodes.
439 * An empty list implies no known ordering. Otherwise the first item
440 * represents the primary sort key, the second the first secondary sort key,
441 * etc. The value being sorted is represented by linking to an
442 * EquivalenceClass containing that value and including pk_opfamily among its
443 * ec_opfamilies. This is a convenient method because it makes it trivial
444 * to detect equivalent and closely-related orderings. (See optimizer/README
445 * for more information.)
447 * Note: pk_strategy is either BTLessStrategyNumber (for ASC) or
448 * BTGreaterStrategyNumber (for DESC). We assume that all ordering-capable
449 * index types will use btree-compatible strategy numbers.
452 typedef struct PathKey
456 EquivalenceClass *pk_eclass; /* the value that is ordered */
457 Oid pk_opfamily; /* btree opfamily defining the ordering */
458 int pk_strategy; /* sort direction (ASC or DESC) */
459 bool pk_nulls_first; /* do NULLs come before normal values? */
463 * Type "Path" is used as-is for sequential-scan paths. For other
464 * path types it is the first component of a larger struct.
466 * Note: "pathtype" is the NodeTag of the Plan node we could build from this
467 * Path. It is partially redundant with the Path's NodeTag, but allows us
468 * to use the same Path type for multiple Plan types where there is no need
469 * to distinguish the Plan type during path processing.
476 NodeTag pathtype; /* tag identifying scan/join method */
478 RelOptInfo *parent; /* the relation this path can build */
480 /* estimated execution costs for path (see costsize.c for more info) */
481 Cost startup_cost; /* cost expended before fetching any tuples */
482 Cost total_cost; /* total cost (assuming all tuples fetched) */
484 List *pathkeys; /* sort ordering of path's output */
485 /* pathkeys is a List of PathKey nodes; see above */
489 * IndexPath represents an index scan over a single index.
491 * 'indexinfo' is the index to be scanned.
493 * 'indexclauses' is a list of index qualification clauses, with implicit
494 * AND semantics across the list. Each clause is a RestrictInfo node from
495 * the query's WHERE or JOIN conditions.
497 * 'indexquals' has the same structure as 'indexclauses', but it contains
498 * the actual indexqual conditions that can be used with the index.
499 * In simple cases this is identical to 'indexclauses', but when special
500 * indexable operators appear in 'indexclauses', they are replaced by the
501 * derived indexscannable conditions in 'indexquals'.
503 * 'isjoininner' is TRUE if the path is a nestloop inner scan (that is,
504 * some of the index conditions are join rather than restriction clauses).
505 * Note that the path costs will be calculated differently from a plain
506 * indexscan in this case, and in addition there's a special 'rows' value
507 * different from the parent RelOptInfo's (see below).
509 * 'indexscandir' is one of:
510 * ForwardScanDirection: forward scan of an ordered index
511 * BackwardScanDirection: backward scan of an ordered index
512 * NoMovementScanDirection: scan of an unordered index, or don't care
513 * (The executor doesn't care whether it gets ForwardScanDirection or
514 * NoMovementScanDirection for an indexscan, but the planner wants to
515 * distinguish ordered from unordered indexes for building pathkeys.)
517 * 'indextotalcost' and 'indexselectivity' are saved in the IndexPath so that
518 * we need not recompute them when considering using the same index in a
519 * bitmap index/heap scan (see BitmapHeapPath). The costs of the IndexPath
520 * itself represent the costs of an IndexScan plan type.
522 * 'rows' is the estimated result tuple count for the indexscan. This
523 * is the same as path.parent->rows for a simple indexscan, but it is
524 * different for a nestloop inner scan, because the additional indexquals
525 * coming from join clauses make the scan more selective than the parent
526 * rel's restrict clauses alone would do.
529 typedef struct IndexPath
532 IndexOptInfo *indexinfo;
536 ScanDirection indexscandir;
538 Selectivity indexselectivity;
539 double rows; /* estimated number of result tuples */
543 * BitmapHeapPath represents one or more indexscans that generate TID bitmaps
544 * instead of directly accessing the heap, followed by AND/OR combinations
545 * to produce a single bitmap, followed by a heap scan that uses the bitmap.
546 * Note that the output is always considered unordered, since it will come
547 * out in physical heap order no matter what the underlying indexes did.
549 * The individual indexscans are represented by IndexPath nodes, and any
550 * logic on top of them is represented by a tree of BitmapAndPath and
551 * BitmapOrPath nodes. Notice that we can use the same IndexPath node both
552 * to represent a regular IndexScan plan, and as the child of a BitmapHeapPath
553 * that represents scanning the same index using a BitmapIndexScan. The
554 * startup_cost and total_cost figures of an IndexPath always represent the
555 * costs to use it as a regular IndexScan. The costs of a BitmapIndexScan
556 * can be computed using the IndexPath's indextotalcost and indexselectivity.
558 * BitmapHeapPaths can be nestloop inner indexscans. The isjoininner and
559 * rows fields serve the same purpose as for plain IndexPaths.
561 typedef struct BitmapHeapPath
564 Path *bitmapqual; /* IndexPath, BitmapAndPath, BitmapOrPath */
565 bool isjoininner; /* T if it's a nestloop inner scan */
566 double rows; /* estimated number of result tuples */
570 * BitmapAndPath represents a BitmapAnd plan node; it can only appear as
571 * part of the substructure of a BitmapHeapPath. The Path structure is
572 * a bit more heavyweight than we really need for this, but for simplicity
573 * we make it a derivative of Path anyway.
575 typedef struct BitmapAndPath
578 List *bitmapquals; /* IndexPaths and BitmapOrPaths */
579 Selectivity bitmapselectivity;
583 * BitmapOrPath represents a BitmapOr plan node; it can only appear as
584 * part of the substructure of a BitmapHeapPath. The Path structure is
585 * a bit more heavyweight than we really need for this, but for simplicity
586 * we make it a derivative of Path anyway.
588 typedef struct BitmapOrPath
591 List *bitmapquals; /* IndexPaths and BitmapAndPaths */
592 Selectivity bitmapselectivity;
596 * TidPath represents a scan by TID
598 * tidquals is an implicitly OR'ed list of qual expressions of the form
599 * "CTID = pseudoconstant" or "CTID = ANY(pseudoconstant_array)".
600 * Note they are bare expressions, not RestrictInfos.
602 typedef struct TidPath
605 List *tidquals; /* qual(s) involving CTID = something */
609 * AppendPath represents an Append plan, ie, successive execution of
610 * several member plans.
612 * Note: it is possible for "subpaths" to contain only one, or even no,
613 * elements. These cases are optimized during create_append_plan.
615 typedef struct AppendPath
618 List *subpaths; /* list of component Paths */
622 * ResultPath represents use of a Result plan node to compute a variable-free
623 * targetlist with no underlying tables (a "SELECT expressions" query).
624 * The query could have a WHERE clause, too, represented by "quals".
626 * Note that quals is a list of bare clauses, not RestrictInfos.
628 typedef struct ResultPath
635 * MaterialPath represents use of a Material plan node, i.e., caching of
636 * the output of its subpath. This is used when the subpath is expensive
637 * and needs to be scanned repeatedly, or when we need mark/restore ability
638 * and the subpath doesn't have it.
640 typedef struct MaterialPath
647 * UniquePath represents elimination of distinct rows from the output of
650 * This is unlike the other Path nodes in that it can actually generate
651 * different plans: either hash-based or sort-based implementation, or a
652 * no-op if the input path can be proven distinct already. The decision
653 * is sufficiently localized that it's not worth having separate Path node
654 * types. (Note: in the no-op case, we could eliminate the UniquePath node
655 * entirely and just return the subpath; but it's convenient to have a
656 * UniquePath in the path tree to signal upper-level routines that the input
657 * is known distinct.)
661 UNIQUE_PATH_NOOP, /* input is known unique already */
662 UNIQUE_PATH_HASH, /* use hashing */
663 UNIQUE_PATH_SORT /* use sorting */
666 typedef struct UniquePath
670 UniquePathMethod umethod;
671 double rows; /* estimated number of result tuples */
675 * All join-type paths share these fields.
678 typedef struct JoinPath
684 Path *outerjoinpath; /* path for the outer side of the join */
685 Path *innerjoinpath; /* path for the inner side of the join */
687 List *joinrestrictinfo; /* RestrictInfos to apply to join */
690 * See the notes for RelOptInfo to understand why joinrestrictinfo is
691 * needed in JoinPath, and can't be merged into the parent RelOptInfo.
696 * A nested-loop path needs no special fields.
699 typedef JoinPath NestPath;
702 * A mergejoin path has these fields.
704 * path_mergeclauses lists the clauses (in the form of RestrictInfos)
705 * that will be used in the merge.
707 * Note that the mergeclauses are a subset of the parent relation's
708 * restriction-clause list. Any join clauses that are not mergejoinable
709 * appear only in the parent's restrict list, and must be checked by a
710 * qpqual at execution time.
712 * outersortkeys (resp. innersortkeys) is NIL if the outer path
713 * (resp. inner path) is already ordered appropriately for the
714 * mergejoin. If it is not NIL then it is a PathKeys list describing
715 * the ordering that must be created by an explicit sort step.
718 typedef struct MergePath
721 List *path_mergeclauses; /* join clauses to be used for merge */
722 List *outersortkeys; /* keys for explicit sort, if any */
723 List *innersortkeys; /* keys for explicit sort, if any */
727 * A hashjoin path has these fields.
729 * The remarks above for mergeclauses apply for hashclauses as well.
731 * Hashjoin does not care what order its inputs appear in, so we have
732 * no need for sortkeys.
735 typedef struct HashPath
738 List *path_hashclauses; /* join clauses used for hashing */
742 * Restriction clause info.
744 * We create one of these for each AND sub-clause of a restriction condition
745 * (WHERE or JOIN/ON clause). Since the restriction clauses are logically
746 * ANDed, we can use any one of them or any subset of them to filter out
747 * tuples, without having to evaluate the rest. The RestrictInfo node itself
748 * stores data used by the optimizer while choosing the best query plan.
750 * If a restriction clause references a single base relation, it will appear
751 * in the baserestrictinfo list of the RelOptInfo for that base rel.
753 * If a restriction clause references more than one base rel, it will
754 * appear in the joininfo list of every RelOptInfo that describes a strict
755 * subset of the base rels mentioned in the clause. The joininfo lists are
756 * used to drive join tree building by selecting plausible join candidates.
757 * The clause cannot actually be applied until we have built a join rel
758 * containing all the base rels it references, however.
760 * When we construct a join rel that includes all the base rels referenced
761 * in a multi-relation restriction clause, we place that clause into the
762 * joinrestrictinfo lists of paths for the join rel, if neither left nor
763 * right sub-path includes all base rels referenced in the clause. The clause
764 * will be applied at that join level, and will not propagate any further up
765 * the join tree. (Note: the "predicate migration" code was once intended to
766 * push restriction clauses up and down the plan tree based on evaluation
767 * costs, but it's dead code and is unlikely to be resurrected in the
768 * foreseeable future.)
770 * Note that in the presence of more than two rels, a multi-rel restriction
771 * might reach different heights in the join tree depending on the join
772 * sequence we use. So, these clauses cannot be associated directly with
773 * the join RelOptInfo, but must be kept track of on a per-join-path basis.
775 * RestrictInfos that represent equivalence conditions (i.e., mergejoinable
776 * equalities that are not outerjoin-delayed) are handled a bit differently.
777 * Initially we attach them to the EquivalenceClasses that are derived from
778 * them. When we construct a scan or join path, we look through all the
779 * EquivalenceClasses and generate derived RestrictInfos representing the
780 * minimal set of conditions that need to be checked for this particular scan
781 * or join to enforce that all members of each EquivalenceClass are in fact
782 * equal in all rows emitted by the scan or join.
784 * When dealing with outer joins we have to be very careful about pushing qual
785 * clauses up and down the tree. An outer join's own JOIN/ON conditions must
786 * be evaluated exactly at that join node, and any quals appearing in WHERE or
787 * in a JOIN above the outer join cannot be pushed down below the outer join.
788 * Otherwise the outer join will produce wrong results because it will see the
789 * wrong sets of input rows. All quals are stored as RestrictInfo nodes
790 * during planning, but there's a flag to indicate whether a qual has been
791 * pushed down to a lower level than its original syntactic placement in the
792 * join tree would suggest. If an outer join prevents us from pushing a qual
793 * down to its "natural" semantic level (the level associated with just the
794 * base rels used in the qual) then we mark the qual with a "required_relids"
795 * value including more than just the base rels it actually uses. By
796 * pretending that the qual references all the rels appearing in the outer
797 * join, we prevent it from being evaluated below the outer join's joinrel.
798 * When we do form the outer join's joinrel, we still need to distinguish
799 * those quals that are actually in that join's JOIN/ON condition from those
800 * that appeared elsewhere in the tree and were pushed down to the join rel
801 * because they used no other rels. That's what the is_pushed_down flag is
802 * for; it tells us that a qual is not an OUTER JOIN qual for the set of base
803 * rels listed in required_relids. A clause that originally came from WHERE
804 * or an INNER JOIN condition will *always* have its is_pushed_down flag set.
805 * It's possible for an OUTER JOIN clause to be marked is_pushed_down too,
806 * if we decide that it can be pushed down into the nullable side of the join.
807 * In that case it acts as a plain filter qual for wherever it gets evaluated.
809 * When application of a qual must be delayed by outer join, we also mark it
810 * with outerjoin_delayed = true. This isn't redundant with required_relids
811 * because that might equal clause_relids whether or not it's an outer-join
814 * In general, the referenced clause might be arbitrarily complex. The
815 * kinds of clauses we can handle as indexscan quals, mergejoin clauses,
816 * or hashjoin clauses are limited (e.g., no volatile functions). The code
817 * for each kind of path is responsible for identifying the restrict clauses
818 * it can use and ignoring the rest. Clauses not implemented by an indexscan,
819 * mergejoin, or hashjoin will be placed in the plan qual or joinqual field
820 * of the finished Plan node, where they will be enforced by general-purpose
821 * qual-expression-evaluation code. (But we are still entitled to count
822 * their selectivity when estimating the result tuple count, if we
823 * can guess what it is...)
825 * When the referenced clause is an OR clause, we generate a modified copy
826 * in which additional RestrictInfo nodes are inserted below the top-level
827 * OR/AND structure. This is a convenience for OR indexscan processing:
828 * indexquals taken from either the top level or an OR subclause will have
829 * associated RestrictInfo nodes.
831 * The can_join flag is set true if the clause looks potentially useful as
832 * a merge or hash join clause, that is if it is a binary opclause with
833 * nonoverlapping sets of relids referenced in the left and right sides.
834 * (Whether the operator is actually merge or hash joinable isn't checked,
837 * The pseudoconstant flag is set true if the clause contains no Vars of
838 * the current query level and no volatile functions. Such a clause can be
839 * pulled out and used as a one-time qual in a gating Result node. We keep
840 * pseudoconstant clauses in the same lists as other RestrictInfos so that
841 * the regular clause-pushing machinery can assign them to the correct join
842 * level, but they need to be treated specially for cost and selectivity
843 * estimates. Note that a pseudoconstant clause can never be an indexqual
844 * or merge or hash join clause, so it's of no interest to large parts of
847 * When join clauses are generated from EquivalenceClasses, there may be
848 * several equally valid ways to enforce join equivalence, of which we need
849 * apply only one. We mark clauses of this kind by setting parent_ec to
850 * point to the generating EquivalenceClass. Multiple clauses with the same
851 * parent_ec in the same join are redundant.
854 typedef struct RestrictInfo
858 Expr *clause; /* the represented clause of WHERE or JOIN */
860 bool is_pushed_down; /* TRUE if clause was pushed down in level */
862 bool outerjoin_delayed; /* TRUE if delayed by outer join */
864 bool can_join; /* see comment above */
866 bool pseudoconstant; /* see comment above */
868 /* The set of relids (varnos) actually referenced in the clause: */
869 Relids clause_relids;
871 /* The set of relids required to evaluate the clause: */
872 Relids required_relids;
874 /* These fields are set for any binary opclause: */
875 Relids left_relids; /* relids in left side of clause */
876 Relids right_relids; /* relids in right side of clause */
878 /* This field is NULL unless clause is an OR clause: */
879 Expr *orclause; /* modified clause with RestrictInfos */
881 /* This field is NULL unless clause is potentially redundant: */
882 EquivalenceClass *parent_ec; /* generating EquivalenceClass */
884 /* cache space for cost and selectivity */
885 QualCost eval_cost; /* eval cost of clause; -1 if not yet set */
886 Selectivity this_selec; /* selectivity; -1 if not yet set */
888 /* valid if clause is mergejoinable, else NIL */
889 List *mergeopfamilies; /* opfamilies containing clause operator */
891 /* cache space for mergeclause processing; NULL if not yet set */
892 EquivalenceClass *left_ec; /* EquivalenceClass containing lefthand */
893 EquivalenceClass *right_ec; /* EquivalenceClass containing righthand */
894 EquivalenceMember *left_em; /* EquivalenceMember for lefthand */
895 EquivalenceMember *right_em; /* EquivalenceMember for righthand */
896 List *scansel_cache; /* list of MergeScanSelCache structs */
898 /* transient workspace for use while considering a specific join path */
899 bool outer_is_left; /* T = outer var on left, F = on right */
901 /* valid if clause is hashjoinable, else InvalidOid: */
902 Oid hashjoinoperator; /* copy of clause operator */
904 /* cache space for hashclause processing; -1 if not yet set */
905 Selectivity left_bucketsize; /* avg bucketsize of left side */
906 Selectivity right_bucketsize; /* avg bucketsize of right side */
910 * Since mergejoinscansel() is a relatively expensive function, and would
911 * otherwise be invoked many times while planning a large join tree,
912 * we go out of our way to cache its results. Each mergejoinable
913 * RestrictInfo carries a list of the specific sort orderings that have
914 * been considered for use with it, and the resulting selectivities.
916 typedef struct MergeScanSelCache
918 /* Ordering details (cache lookup key) */
919 Oid opfamily; /* btree opfamily defining the ordering */
920 int strategy; /* sort direction (ASC or DESC) */
921 bool nulls_first; /* do NULLs come before normal values? */
923 Selectivity leftscansel; /* scan fraction for clause left side */
924 Selectivity rightscansel; /* scan fraction for clause right side */
928 * Inner indexscan info.
930 * An inner indexscan is one that uses one or more joinclauses as index
931 * conditions (perhaps in addition to plain restriction clauses). So it
932 * can only be used as the inner path of a nestloop join where the outer
933 * relation includes all other relids appearing in those joinclauses.
934 * The set of usable joinclauses, and thus the best inner indexscan,
935 * thus varies depending on which outer relation we consider; so we have
936 * to recompute the best such path for every join. To avoid lots of
937 * redundant computation, we cache the results of such searches. For
938 * each relation we compute the set of possible otherrelids (all relids
939 * appearing in joinquals that could become indexquals for this table).
940 * Two outer relations whose relids have the same intersection with this
941 * set will have the same set of available joinclauses and thus the same
942 * best inner indexscan for the inner relation. By taking the intersection
943 * before scanning the cache, we avoid recomputing when considering
944 * join rels that differ only by the inclusion of irrelevant other rels.
946 * The search key also includes a bool showing whether the join being
947 * considered is an outer join. Since we constrain the join order for
948 * outer joins, I believe that this bool can only have one possible value
949 * for any particular base relation; but store it anyway to avoid confusion.
952 typedef struct InnerIndexscanInfo
955 /* The lookup key: */
956 Relids other_relids; /* a set of relevant other relids */
957 bool isouterjoin; /* true if join is outer */
958 /* Best path for this lookup key: */
959 Path *best_innerpath; /* best inner indexscan, or NULL if none */
960 } InnerIndexscanInfo;
965 * One-sided outer joins constrain the order of joining partially but not
966 * completely. We flatten such joins into the planner's top-level list of
967 * relations to join, but record information about each outer join in an
968 * OuterJoinInfo struct. These structs are kept in the PlannerInfo node's
971 * min_lefthand and min_righthand are the sets of base relids that must be
972 * available on each side when performing the outer join. lhs_strict is
973 * true if the outer join's condition cannot succeed when the LHS variables
974 * are all NULL (this means that the outer join can commute with upper-level
975 * outer joins even if it appears in their RHS). We don't bother to set
976 * lhs_strict for FULL JOINs, however.
978 * It is not valid for either min_lefthand or min_righthand to be empty sets;
979 * if they were, this would break the logic that enforces join order.
981 * Note: OuterJoinInfo directly represents only LEFT JOIN and FULL JOIN;
982 * RIGHT JOIN is handled by switching the inputs to make it a LEFT JOIN.
983 * We make an OuterJoinInfo for FULL JOINs even though there is no flexibility
984 * of planning for them, because this simplifies make_join_rel()'s API.
987 typedef struct OuterJoinInfo
990 Relids min_lefthand; /* base relids in minimum LHS for join */
991 Relids min_righthand; /* base relids in minimum RHS for join */
992 bool is_full_join; /* it's a FULL OUTER JOIN */
993 bool lhs_strict; /* joinclause is strict for some LHS rel */
999 * When we convert top-level IN quals into join operations, we must restrict
1000 * the order of joining and use special join methods at some join points.
1001 * We record information about each such IN clause in an InClauseInfo struct.
1002 * These structs are kept in the PlannerInfo node's in_info_list.
1004 * Note: sub_targetlist is just a list of Vars or expressions; it does not
1005 * contain TargetEntry nodes.
1008 typedef struct InClauseInfo
1011 Relids lefthand; /* base relids in lefthand expressions */
1012 Relids righthand; /* base relids coming from the subselect */
1013 List *sub_targetlist; /* targetlist of original RHS subquery */
1014 List *in_operators; /* OIDs of the IN's equality operator(s) */
1018 * Append-relation info.
1020 * When we expand an inheritable table or a UNION-ALL subselect into an
1021 * "append relation" (essentially, a list of child RTEs), we build an
1022 * AppendRelInfo for each child RTE. The list of AppendRelInfos indicates
1023 * which child RTEs must be included when expanding the parent, and each
1024 * node carries information needed to translate Vars referencing the parent
1025 * into Vars referencing that child.
1027 * These structs are kept in the PlannerInfo node's append_rel_list.
1028 * Note that we just throw all the structs into one list, and scan the
1029 * whole list when desiring to expand any one parent. We could have used
1030 * a more complex data structure (eg, one list per parent), but this would
1031 * be harder to update during operations such as pulling up subqueries,
1032 * and not really any easier to scan. Considering that typical queries
1033 * will not have many different append parents, it doesn't seem worthwhile
1034 * to complicate things.
1036 * Note: after completion of the planner prep phase, any given RTE is an
1037 * append parent having entries in append_rel_list if and only if its
1038 * "inh" flag is set. We clear "inh" for plain tables that turn out not
1039 * to have inheritance children, and (in an abuse of the original meaning
1040 * of the flag) we set "inh" for subquery RTEs that turn out to be
1041 * flattenable UNION ALL queries. This lets us avoid useless searches
1042 * of append_rel_list.
1044 * Note: the data structure assumes that append-rel members are single
1045 * baserels. This is OK for inheritance, but it prevents us from pulling
1046 * up a UNION ALL member subquery if it contains a join. While that could
1047 * be fixed with a more complex data structure, at present there's not much
1048 * point because no improvement in the plan could result.
1051 typedef struct AppendRelInfo
1056 * These fields uniquely identify this append relationship. There can be
1057 * (in fact, always should be) multiple AppendRelInfos for the same
1058 * parent_relid, but never more than one per child_relid, since a given
1059 * RTE cannot be a child of more than one append parent.
1061 Index parent_relid; /* RT index of append parent rel */
1062 Index child_relid; /* RT index of append child rel */
1065 * For an inheritance appendrel, the parent and child are both regular
1066 * relations, and we store their rowtype OIDs here for use in translating
1067 * whole-row Vars. For a UNION-ALL appendrel, the parent and child are
1068 * both subqueries with no named rowtype, and we store InvalidOid here.
1070 Oid parent_reltype; /* OID of parent's composite type */
1071 Oid child_reltype; /* OID of child's composite type */
1074 * The N'th element of this list is the integer column number of the child
1075 * column corresponding to the N'th column of the parent. A list element
1076 * is zero if it corresponds to a dropped column of the parent (this is
1077 * only possible for inheritance cases, not UNION ALL).
1079 List *col_mappings; /* list of child attribute numbers */
1082 * The N'th element of this list is a Var or expression representing the
1083 * child column corresponding to the N'th column of the parent. This is
1084 * used to translate Vars referencing the parent rel into references to
1085 * the child. A list element is NULL if it corresponds to a dropped
1086 * column of the parent (this is only possible for inheritance cases, not
1089 * This might seem redundant with the col_mappings data, but it is handy
1090 * because flattening of sub-SELECTs that are members of a UNION ALL will
1091 * cause changes in the expressions that need to be substituted for a
1092 * parent Var. Adjusting this data structure lets us track what really
1093 * needs to be substituted.
1095 * Notice we only store entries for user columns (attno > 0). Whole-row
1096 * Vars are special-cased, and system columns (attno < 0) need no special
1097 * translation since their attnos are the same for all tables.
1099 * Caution: the Vars have varlevelsup = 0. Be careful to adjust as needed
1100 * when copying into a subquery.
1102 List *translated_vars; /* Expressions in the child's Vars */
1105 * We store the parent table's OID here for inheritance, or InvalidOid for
1106 * UNION ALL. This is only needed to help in generating error messages if
1107 * an attempt is made to reference a dropped parent column.
1109 Oid parent_reloid; /* OID of parent relation */
1112 #endif /* RELATION_H */