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 1.132 2007/01/10 18:06:04 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-relation 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 *equi_key_list; /* list of lists of equijoined PathKeyItems */
89 List *left_join_clauses; /* list of RestrictInfos for outer
90 * join clauses w/nonnullable var on
93 List *right_join_clauses; /* list of RestrictInfos for outer
94 * join clauses w/nonnullable var on
97 List *full_join_clauses; /* list of RestrictInfos for full
98 * outer join clauses */
100 List *oj_info_list; /* list of OuterJoinInfos */
102 List *in_info_list; /* list of InClauseInfos */
104 List *append_rel_list; /* list of AppendRelInfos */
106 List *query_pathkeys; /* desired pathkeys for query_planner(), and
107 * actual pathkeys afterwards */
109 List *group_pathkeys; /* groupClause pathkeys, if any */
110 List *sort_pathkeys; /* sortClause pathkeys, if any */
112 double total_table_pages; /* # of pages in all tables of query */
114 double tuple_fraction; /* tuple_fraction passed to query_planner */
116 bool hasJoinRTEs; /* true if any RTEs are RTE_JOIN kind */
117 bool hasOuterJoins; /* true if any RTEs are outer joins */
118 bool hasHavingQual; /* true if havingQual was non-null */
119 bool hasPseudoConstantQuals; /* true if any RestrictInfo has
120 * pseudoconstant = true */
126 * Per-relation information for planning/optimization
128 * For planning purposes, a "base rel" is either a plain relation (a table)
129 * or the output of a sub-SELECT or function that appears in the range table.
130 * In either case it is uniquely identified by an RT index. A "joinrel"
131 * is the joining of two or more base rels. A joinrel is identified by
132 * the set of RT indexes for its component baserels. We create RelOptInfo
133 * nodes for each baserel and joinrel, and store them in the PlannerInfo's
134 * simple_rel_array and join_rel_list respectively.
136 * Note that there is only one joinrel for any given set of component
137 * baserels, no matter what order we assemble them in; so an unordered
138 * set is the right datatype to identify it with.
140 * We also have "other rels", which are like base rels in that they refer to
141 * single RT indexes; but they are not part of the join tree, and are given
142 * a different RelOptKind to identify them.
144 * Currently the only kind of otherrels are those made for member relations
145 * of an "append relation", that is an inheritance set or UNION ALL subquery.
146 * An append relation has a parent RTE that is a base rel, which represents
147 * the entire append relation. The member RTEs are otherrels. The parent
148 * is present in the query join tree but the members are not. The member
149 * RTEs and otherrels are used to plan the scans of the individual tables or
150 * subqueries of the append set; then the parent baserel is given an Append
151 * plan comprising the best plans for the individual member rels. (See
152 * comments for AppendRelInfo for more information.)
154 * At one time we also made otherrels to represent join RTEs, for use in
155 * handling join alias Vars. Currently this is not needed because all join
156 * alias Vars are expanded to non-aliased form during preprocess_expression.
158 * Parts of this data structure are specific to various scan and join
159 * mechanisms. It didn't seem worth creating new node types for them.
161 * relids - Set of base-relation identifiers; it is a base relation
162 * if there is just one, a join relation if more than one
163 * rows - estimated number of tuples in the relation after restriction
164 * clauses have been applied (ie, output rows of a plan for it)
165 * width - avg. number of bytes per tuple in the relation after the
166 * appropriate projections have been done (ie, output width)
167 * reltargetlist - List of Var nodes for the attributes we need to
168 * output from this relation (in no particular order)
169 * NOTE: in a child relation, may contain RowExprs
170 * pathlist - List of Path nodes, one for each potentially useful
171 * method of generating the relation
172 * cheapest_startup_path - the pathlist member with lowest startup cost
173 * (regardless of its ordering)
174 * cheapest_total_path - the pathlist member with lowest total cost
175 * (regardless of its ordering)
176 * cheapest_unique_path - for caching cheapest path to produce unique
177 * (no duplicates) output from relation
179 * If the relation is a base relation it will have these fields set:
181 * relid - RTE index (this is redundant with the relids field, but
182 * is provided for convenience of access)
183 * rtekind - distinguishes plain relation, subquery, or function RTE
184 * min_attr, max_attr - range of valid AttrNumbers for rel
185 * attr_needed - array of bitmapsets indicating the highest joinrel
186 * in which each attribute is needed; if bit 0 is set then
187 * the attribute is needed as part of final targetlist
188 * attr_widths - cache space for per-attribute width estimates;
189 * zero means not computed yet
190 * indexlist - list of IndexOptInfo nodes for relation's indexes
191 * (always NIL if it's not a table)
192 * pages - number of disk pages in relation (zero if not a table)
193 * tuples - number of tuples in relation (not considering restrictions)
194 * subplan - plan for subquery (NULL if it's not a subquery)
196 * Note: for a subquery, tuples and subplan are not set immediately
197 * upon creation of the RelOptInfo object; they are filled in when
198 * set_base_rel_pathlist processes the object.
200 * For otherrels that are appendrel members, these fields are filled
201 * in just as for a baserel.
203 * The presence of the remaining fields depends on the restrictions
204 * and joins that the relation participates in:
206 * baserestrictinfo - List of RestrictInfo nodes, containing info about
207 * each non-join qualification clause in which this relation
208 * participates (only used for base rels)
209 * baserestrictcost - Estimated cost of evaluating the baserestrictinfo
210 * clauses at a single tuple (only used for base rels)
211 * joininfo - List of RestrictInfo nodes, containing info about each
212 * join clause in which this relation participates
213 * index_outer_relids - only used for base rels; set of outer relids
214 * that participate in indexable joinclauses for this rel
215 * index_inner_paths - only used for base rels; list of InnerIndexscanInfo
216 * nodes showing best indexpaths for various subsets of
217 * index_outer_relids.
219 * Note: Keeping a restrictinfo list in the RelOptInfo is useful only for
220 * base rels, because for a join rel the set of clauses that are treated as
221 * restrict clauses varies depending on which sub-relations we choose to join.
222 * (For example, in a 3-base-rel join, a clause relating rels 1 and 2 must be
223 * treated as a restrictclause if we join {1} and {2 3} to make {1 2 3}; but
224 * if we join {1 2} and {3} then that clause will be a restrictclause in {1 2}
225 * and should not be processed again at the level of {1 2 3}.) Therefore,
226 * the restrictinfo list in the join case appears in individual JoinPaths
227 * (field joinrestrictinfo), not in the parent relation. But it's OK for
228 * the RelOptInfo to store the joininfo list, because that is the same
229 * for a given rel no matter how we form it.
231 * We store baserestrictcost in the RelOptInfo (for base relations) because
232 * we know we will need it at least once (to price the sequential scan)
233 * and may need it multiple times to price index scans.
236 typedef enum RelOptKind
240 RELOPT_OTHER_MEMBER_REL
243 typedef struct RelOptInfo
247 RelOptKind reloptkind;
249 /* all relations included in this RelOptInfo */
250 Relids relids; /* set of base relids (rangetable indexes) */
252 /* size estimates generated by planner */
253 double rows; /* estimated number of result tuples */
254 int width; /* estimated avg width of result tuples */
256 /* materialization information */
257 List *reltargetlist; /* needed Vars */
258 List *pathlist; /* Path structures */
259 struct Path *cheapest_startup_path;
260 struct Path *cheapest_total_path;
261 struct Path *cheapest_unique_path;
263 /* information about a base rel (not set for join rels!) */
265 RTEKind rtekind; /* RELATION, SUBQUERY, or FUNCTION */
266 AttrNumber min_attr; /* smallest attrno of rel (often <0) */
267 AttrNumber max_attr; /* largest attrno of rel */
268 Relids *attr_needed; /* array indexed [min_attr .. max_attr] */
269 int32 *attr_widths; /* array indexed [min_attr .. max_attr] */
273 struct Plan *subplan; /* if subquery */
275 /* used by various scans and joins: */
276 List *baserestrictinfo; /* RestrictInfo structures (if base
278 QualCost baserestrictcost; /* cost of evaluating the above */
279 List *joininfo; /* RestrictInfo structures for join clauses
280 * involving this rel */
282 /* cached info about inner indexscan paths for relation: */
283 Relids index_outer_relids; /* other relids in indexable join
285 List *index_inner_paths; /* InnerIndexscanInfo nodes */
288 * Inner indexscans are not in the main pathlist because they are not
289 * usable except in specific join contexts. We use the index_inner_paths
290 * list just to avoid recomputing the best inner indexscan repeatedly for
291 * similar outer relations. See comments for InnerIndexscanInfo.
297 * Per-index information for planning/optimization
299 * Prior to Postgres 7.0, RelOptInfo was used to describe both relations
300 * and indexes, but that created confusion without actually doing anything
301 * useful. So now we have a separate IndexOptInfo struct for indexes.
303 * opfamily[], indexkeys[], fwdsortop[], revsortop[], and nulls_first[]
304 * each have ncolumns entries. Note: for historical reasons, the
305 * opfamily array has an extra entry that is always zero. Some code
306 * scans until it sees a zero entry, rather than looking at ncolumns.
308 * Zeroes in the indexkeys[] array indicate index columns that are
309 * expressions; there is one element in indexprs for each such column.
311 * For an unordered index, the sortop arrays contains zeroes. Note that
312 * fwdsortop[] and nulls_first[] describe the sort ordering of a forward
313 * indexscan; we can also consider a backward indexscan, which will
314 * generate sort order described by revsortop/!nulls_first.
316 * The indexprs and indpred expressions have been run through
317 * prepqual.c and eval_const_expressions() for ease of matching to
318 * WHERE clauses. indpred is in implicit-AND form.
320 typedef struct IndexOptInfo
324 Oid indexoid; /* OID of the index relation */
325 RelOptInfo *rel; /* back-link to index's table */
327 /* statistics from pg_class */
328 BlockNumber pages; /* number of disk pages in index */
329 double tuples; /* number of index tuples in index */
331 /* index descriptor information */
332 int ncolumns; /* number of columns in index */
333 Oid *opfamily; /* OIDs of operator families for columns */
334 int *indexkeys; /* column numbers of index's keys, or 0 */
335 Oid *fwdsortop; /* OIDs of sort operators for each column */
336 Oid *revsortop; /* OIDs of sort operators for backward scan */
337 bool *nulls_first; /* do NULLs come first in the sort order? */
338 Oid relam; /* OID of the access method (in pg_am) */
340 RegProcedure amcostestimate; /* OID of the access method's cost fcn */
342 List *indexprs; /* expressions for non-simple index columns */
343 List *indpred; /* predicate if a partial index, else NIL */
345 bool predOK; /* true if predicate matches query */
346 bool unique; /* true if a unique index */
347 bool amoptionalkey; /* can query omit key for the first column? */
354 * The sort ordering of a path is represented by a list of sublists of
355 * PathKeyItem nodes. An empty list implies no known ordering. Otherwise
356 * the first sublist represents the primary sort key, the second the
357 * first secondary sort key, etc. Each sublist contains one or more
358 * PathKeyItem nodes, each of which can be taken as the attribute that
359 * appears at that sort position. (See optimizer/README for more
363 typedef struct PathKeyItem
367 Node *key; /* the item that is ordered */
368 Oid sortop; /* the ordering operator ('<' op) */
369 bool nulls_first; /* do NULLs come before normal values? */
372 * key typically points to a Var node, ie a relation attribute, but it can
373 * also point to an arbitrary expression representing the value indexed by
374 * an index expression.
379 * Type "Path" is used as-is for sequential-scan paths. For other
380 * path types it is the first component of a larger struct.
382 * Note: "pathtype" is the NodeTag of the Plan node we could build from this
383 * Path. It is partially redundant with the Path's NodeTag, but allows us
384 * to use the same Path type for multiple Plan types where there is no need
385 * to distinguish the Plan type during path processing.
392 NodeTag pathtype; /* tag identifying scan/join method */
394 RelOptInfo *parent; /* the relation this path can build */
396 /* estimated execution costs for path (see costsize.c for more info) */
397 Cost startup_cost; /* cost expended before fetching any tuples */
398 Cost total_cost; /* total cost (assuming all tuples fetched) */
400 List *pathkeys; /* sort ordering of path's output */
401 /* pathkeys is a List of Lists of PathKeyItem nodes; see above */
405 * IndexPath represents an index scan over a single index.
407 * 'indexinfo' is the index to be scanned.
409 * 'indexclauses' is a list of index qualification clauses, with implicit
410 * AND semantics across the list. Each clause is a RestrictInfo node from
411 * the query's WHERE or JOIN conditions.
413 * 'indexquals' has the same structure as 'indexclauses', but it contains
414 * the actual indexqual conditions that can be used with the index.
415 * In simple cases this is identical to 'indexclauses', but when special
416 * indexable operators appear in 'indexclauses', they are replaced by the
417 * derived indexscannable conditions in 'indexquals'.
419 * 'isjoininner' is TRUE if the path is a nestloop inner scan (that is,
420 * some of the index conditions are join rather than restriction clauses).
421 * Note that the path costs will be calculated differently from a plain
422 * indexscan in this case, and in addition there's a special 'rows' value
423 * different from the parent RelOptInfo's (see below).
425 * 'indexscandir' is one of:
426 * ForwardScanDirection: forward scan of an ordered index
427 * BackwardScanDirection: backward scan of an ordered index
428 * NoMovementScanDirection: scan of an unordered index, or don't care
429 * (The executor doesn't care whether it gets ForwardScanDirection or
430 * NoMovementScanDirection for an indexscan, but the planner wants to
431 * distinguish ordered from unordered indexes for building pathkeys.)
433 * 'indextotalcost' and 'indexselectivity' are saved in the IndexPath so that
434 * we need not recompute them when considering using the same index in a
435 * bitmap index/heap scan (see BitmapHeapPath). The costs of the IndexPath
436 * itself represent the costs of an IndexScan plan type.
438 * 'rows' is the estimated result tuple count for the indexscan. This
439 * is the same as path.parent->rows for a simple indexscan, but it is
440 * different for a nestloop inner scan, because the additional indexquals
441 * coming from join clauses make the scan more selective than the parent
442 * rel's restrict clauses alone would do.
445 typedef struct IndexPath
448 IndexOptInfo *indexinfo;
452 ScanDirection indexscandir;
454 Selectivity indexselectivity;
455 double rows; /* estimated number of result tuples */
459 * BitmapHeapPath represents one or more indexscans that generate TID bitmaps
460 * instead of directly accessing the heap, followed by AND/OR combinations
461 * to produce a single bitmap, followed by a heap scan that uses the bitmap.
462 * Note that the output is always considered unordered, since it will come
463 * out in physical heap order no matter what the underlying indexes did.
465 * The individual indexscans are represented by IndexPath nodes, and any
466 * logic on top of them is represented by a tree of BitmapAndPath and
467 * BitmapOrPath nodes. Notice that we can use the same IndexPath node both
468 * to represent a regular IndexScan plan, and as the child of a BitmapHeapPath
469 * that represents scanning the same index using a BitmapIndexScan. The
470 * startup_cost and total_cost figures of an IndexPath always represent the
471 * costs to use it as a regular IndexScan. The costs of a BitmapIndexScan
472 * can be computed using the IndexPath's indextotalcost and indexselectivity.
474 * BitmapHeapPaths can be nestloop inner indexscans. The isjoininner and
475 * rows fields serve the same purpose as for plain IndexPaths.
477 typedef struct BitmapHeapPath
480 Path *bitmapqual; /* IndexPath, BitmapAndPath, BitmapOrPath */
481 bool isjoininner; /* T if it's a nestloop inner scan */
482 double rows; /* estimated number of result tuples */
486 * BitmapAndPath represents a BitmapAnd plan node; it can only appear as
487 * part of the substructure of a BitmapHeapPath. The Path structure is
488 * a bit more heavyweight than we really need for this, but for simplicity
489 * we make it a derivative of Path anyway.
491 typedef struct BitmapAndPath
494 List *bitmapquals; /* IndexPaths and BitmapOrPaths */
495 Selectivity bitmapselectivity;
499 * BitmapOrPath represents a BitmapOr plan node; it can only appear as
500 * part of the substructure of a BitmapHeapPath. The Path structure is
501 * a bit more heavyweight than we really need for this, but for simplicity
502 * we make it a derivative of Path anyway.
504 typedef struct BitmapOrPath
507 List *bitmapquals; /* IndexPaths and BitmapAndPaths */
508 Selectivity bitmapselectivity;
512 * TidPath represents a scan by TID
514 * tidquals is an implicitly OR'ed list of qual expressions of the form
515 * "CTID = pseudoconstant" or "CTID = ANY(pseudoconstant_array)".
516 * Note they are bare expressions, not RestrictInfos.
518 typedef struct TidPath
521 List *tidquals; /* qual(s) involving CTID = something */
525 * AppendPath represents an Append plan, ie, successive execution of
526 * several member plans.
528 * Note: it is possible for "subpaths" to contain only one, or even no,
529 * elements. These cases are optimized during create_append_plan.
531 typedef struct AppendPath
534 List *subpaths; /* list of component Paths */
538 * ResultPath represents use of a Result plan node to compute a variable-free
539 * targetlist with no underlying tables (a "SELECT expressions" query).
540 * The query could have a WHERE clause, too, represented by "quals".
542 * Note that quals is a list of bare clauses, not RestrictInfos.
544 typedef struct ResultPath
551 * MaterialPath represents use of a Material plan node, i.e., caching of
552 * the output of its subpath. This is used when the subpath is expensive
553 * and needs to be scanned repeatedly, or when we need mark/restore ability
554 * and the subpath doesn't have it.
556 typedef struct MaterialPath
563 * UniquePath represents elimination of distinct rows from the output of
566 * This is unlike the other Path nodes in that it can actually generate
567 * different plans: either hash-based or sort-based implementation, or a
568 * no-op if the input path can be proven distinct already. The decision
569 * is sufficiently localized that it's not worth having separate Path node
570 * types. (Note: in the no-op case, we could eliminate the UniquePath node
571 * entirely and just return the subpath; but it's convenient to have a
572 * UniquePath in the path tree to signal upper-level routines that the input
573 * is known distinct.)
577 UNIQUE_PATH_NOOP, /* input is known unique already */
578 UNIQUE_PATH_HASH, /* use hashing */
579 UNIQUE_PATH_SORT /* use sorting */
582 typedef struct UniquePath
586 UniquePathMethod umethod;
587 double rows; /* estimated number of result tuples */
591 * All join-type paths share these fields.
594 typedef struct JoinPath
600 Path *outerjoinpath; /* path for the outer side of the join */
601 Path *innerjoinpath; /* path for the inner side of the join */
603 List *joinrestrictinfo; /* RestrictInfos to apply to join */
606 * See the notes for RelOptInfo to understand why joinrestrictinfo is
607 * needed in JoinPath, and can't be merged into the parent RelOptInfo.
612 * A nested-loop path needs no special fields.
615 typedef JoinPath NestPath;
618 * A mergejoin path has these fields.
620 * path_mergeclauses lists the clauses (in the form of RestrictInfos)
621 * that will be used in the merge. The parallel arrays path_mergeFamilies,
622 * path_mergeStrategies, and path_mergeNullsFirst specify the merge semantics
623 * for each clause (i.e., define the relevant sort ordering for each clause).
624 * (XXX is this the most reasonable path-time representation? It's at least
625 * partially redundant with the pathkeys of the input paths.)
627 * Note that the mergeclauses are a subset of the parent relation's
628 * restriction-clause list. Any join clauses that are not mergejoinable
629 * appear only in the parent's restrict list, and must be checked by a
630 * qpqual at execution time.
632 * outersortkeys (resp. innersortkeys) is NIL if the outer path
633 * (resp. inner path) is already ordered appropriately for the
634 * mergejoin. If it is not NIL then it is a PathKeys list describing
635 * the ordering that must be created by an explicit sort step.
638 typedef struct MergePath
641 List *path_mergeclauses; /* join clauses to be used for merge */
642 /* these are arrays, but have the same length as the mergeclauses list: */
643 Oid *path_mergeFamilies; /* per-clause OIDs of opfamilies */
644 int *path_mergeStrategies; /* per-clause ordering (ASC or DESC) */
645 bool *path_mergeNullsFirst; /* per-clause nulls ordering */
646 List *outersortkeys; /* keys for explicit sort, if any */
647 List *innersortkeys; /* keys for explicit sort, if any */
651 * A hashjoin path has these fields.
653 * The remarks above for mergeclauses apply for hashclauses as well.
655 * Hashjoin does not care what order its inputs appear in, so we have
656 * no need for sortkeys.
659 typedef struct HashPath
662 List *path_hashclauses; /* join clauses used for hashing */
666 * Restriction clause info.
668 * We create one of these for each AND sub-clause of a restriction condition
669 * (WHERE or JOIN/ON clause). Since the restriction clauses are logically
670 * ANDed, we can use any one of them or any subset of them to filter out
671 * tuples, without having to evaluate the rest. The RestrictInfo node itself
672 * stores data used by the optimizer while choosing the best query plan.
674 * If a restriction clause references a single base relation, it will appear
675 * in the baserestrictinfo list of the RelOptInfo for that base rel.
677 * If a restriction clause references more than one base rel, it will
678 * appear in the joininfo list of every RelOptInfo that describes a strict
679 * subset of the base rels mentioned in the clause. The joininfo lists are
680 * used to drive join tree building by selecting plausible join candidates.
681 * The clause cannot actually be applied until we have built a join rel
682 * containing all the base rels it references, however.
684 * When we construct a join rel that includes all the base rels referenced
685 * in a multi-relation restriction clause, we place that clause into the
686 * joinrestrictinfo lists of paths for the join rel, if neither left nor
687 * right sub-path includes all base rels referenced in the clause. The clause
688 * will be applied at that join level, and will not propagate any further up
689 * the join tree. (Note: the "predicate migration" code was once intended to
690 * push restriction clauses up and down the plan tree based on evaluation
691 * costs, but it's dead code and is unlikely to be resurrected in the
692 * foreseeable future.)
694 * Note that in the presence of more than two rels, a multi-rel restriction
695 * might reach different heights in the join tree depending on the join
696 * sequence we use. So, these clauses cannot be associated directly with
697 * the join RelOptInfo, but must be kept track of on a per-join-path basis.
699 * When dealing with outer joins we have to be very careful about pushing qual
700 * clauses up and down the tree. An outer join's own JOIN/ON conditions must
701 * be evaluated exactly at that join node, and any quals appearing in WHERE or
702 * in a JOIN above the outer join cannot be pushed down below the outer join.
703 * Otherwise the outer join will produce wrong results because it will see the
704 * wrong sets of input rows. All quals are stored as RestrictInfo nodes
705 * during planning, but there's a flag to indicate whether a qual has been
706 * pushed down to a lower level than its original syntactic placement in the
707 * join tree would suggest. If an outer join prevents us from pushing a qual
708 * down to its "natural" semantic level (the level associated with just the
709 * base rels used in the qual) then we mark the qual with a "required_relids"
710 * value including more than just the base rels it actually uses. By
711 * pretending that the qual references all the rels appearing in the outer
712 * join, we prevent it from being evaluated below the outer join's joinrel.
713 * When we do form the outer join's joinrel, we still need to distinguish
714 * those quals that are actually in that join's JOIN/ON condition from those
715 * that appeared higher in the tree and were pushed down to the join rel
716 * because they used no other rels. That's what the is_pushed_down flag is
717 * for; it tells us that a qual came from a point above the join of the
718 * set of base rels listed in required_relids. A clause that originally came
719 * from WHERE will *always* have its is_pushed_down flag set; a clause that
720 * came from an INNER JOIN condition, but doesn't use all the rels being
721 * joined, will also have is_pushed_down set because it will get attached to
722 * some lower joinrel.
724 * When application of a qual must be delayed by outer join, we also mark it
725 * with outerjoin_delayed = true. This isn't redundant with required_relids
726 * because that might equal clause_relids whether or not it's an outer-join
729 * In general, the referenced clause might be arbitrarily complex. The
730 * kinds of clauses we can handle as indexscan quals, mergejoin clauses,
731 * or hashjoin clauses are fairly limited --- the code for each kind of
732 * path is responsible for identifying the restrict clauses it can use
733 * and ignoring the rest. Clauses not implemented by an indexscan,
734 * mergejoin, or hashjoin will be placed in the plan qual or joinqual field
735 * of the finished Plan node, where they will be enforced by general-purpose
736 * qual-expression-evaluation code. (But we are still entitled to count
737 * their selectivity when estimating the result tuple count, if we
738 * can guess what it is...)
740 * When the referenced clause is an OR clause, we generate a modified copy
741 * in which additional RestrictInfo nodes are inserted below the top-level
742 * OR/AND structure. This is a convenience for OR indexscan processing:
743 * indexquals taken from either the top level or an OR subclause will have
744 * associated RestrictInfo nodes.
746 * The can_join flag is set true if the clause looks potentially useful as
747 * a merge or hash join clause, that is if it is a binary opclause with
748 * nonoverlapping sets of relids referenced in the left and right sides.
749 * (Whether the operator is actually merge or hash joinable isn't checked,
752 * The pseudoconstant flag is set true if the clause contains no Vars of
753 * the current query level and no volatile functions. Such a clause can be
754 * pulled out and used as a one-time qual in a gating Result node. We keep
755 * pseudoconstant clauses in the same lists as other RestrictInfos so that
756 * the regular clause-pushing machinery can assign them to the correct join
757 * level, but they need to be treated specially for cost and selectivity
758 * estimates. Note that a pseudoconstant clause can never be an indexqual
759 * or merge or hash join clause, so it's of no interest to large parts of
763 typedef struct RestrictInfo
767 Expr *clause; /* the represented clause of WHERE or JOIN */
769 bool is_pushed_down; /* TRUE if clause was pushed down in level */
771 bool outerjoin_delayed; /* TRUE if delayed by outer join */
773 bool can_join; /* see comment above */
775 bool pseudoconstant; /* see comment above */
777 /* The set of relids (varnos) actually referenced in the clause: */
778 Relids clause_relids;
780 /* The set of relids required to evaluate the clause: */
781 Relids required_relids;
783 /* These fields are set for any binary opclause: */
784 Relids left_relids; /* relids in left side of clause */
785 Relids right_relids; /* relids in right side of clause */
787 /* This field is NULL unless clause is an OR clause: */
788 Expr *orclause; /* modified clause with RestrictInfos */
790 /* cache space for cost and selectivity */
791 QualCost eval_cost; /* eval cost of clause; -1 if not yet set */
792 Selectivity this_selec; /* selectivity; -1 if not yet set */
794 /* valid if clause is mergejoinable, else InvalidOid: */
795 Oid mergejoinoperator; /* copy of clause operator */
796 Oid left_sortop; /* leftside sortop needed for mergejoin */
797 Oid right_sortop; /* rightside sortop needed for mergejoin */
798 Oid mergeopfamily; /* btree opfamily relating these ops */
800 /* cache space for mergeclause processing; NIL if not yet set */
801 List *left_pathkey; /* canonical pathkey for left side */
802 List *right_pathkey; /* canonical pathkey for right side */
804 /* cache space for mergeclause processing; -1 if not yet set */
805 Selectivity left_mergescansel; /* fraction of left side to scan */
806 Selectivity right_mergescansel; /* fraction of right side to scan */
808 /* valid if clause is hashjoinable, else InvalidOid: */
809 Oid hashjoinoperator; /* copy of clause operator */
811 /* cache space for hashclause processing; -1 if not yet set */
812 Selectivity left_bucketsize; /* avg bucketsize of left side */
813 Selectivity right_bucketsize; /* avg bucketsize of right side */
817 * Inner indexscan info.
819 * An inner indexscan is one that uses one or more joinclauses as index
820 * conditions (perhaps in addition to plain restriction clauses). So it
821 * can only be used as the inner path of a nestloop join where the outer
822 * relation includes all other relids appearing in those joinclauses.
823 * The set of usable joinclauses, and thus the best inner indexscan,
824 * thus varies depending on which outer relation we consider; so we have
825 * to recompute the best such path for every join. To avoid lots of
826 * redundant computation, we cache the results of such searches. For
827 * each relation we compute the set of possible otherrelids (all relids
828 * appearing in joinquals that could become indexquals for this table).
829 * Two outer relations whose relids have the same intersection with this
830 * set will have the same set of available joinclauses and thus the same
831 * best inner indexscan for the inner relation. By taking the intersection
832 * before scanning the cache, we avoid recomputing when considering
833 * join rels that differ only by the inclusion of irrelevant other rels.
835 * The search key also includes a bool showing whether the join being
836 * considered is an outer join. Since we constrain the join order for
837 * outer joins, I believe that this bool can only have one possible value
838 * for any particular base relation; but store it anyway to avoid confusion.
841 typedef struct InnerIndexscanInfo
844 /* The lookup key: */
845 Relids other_relids; /* a set of relevant other relids */
846 bool isouterjoin; /* true if join is outer */
847 /* Best path for this lookup key: */
848 Path *best_innerpath; /* best inner indexscan, or NULL if none */
849 } InnerIndexscanInfo;
854 * One-sided outer joins constrain the order of joining partially but not
855 * completely. We flatten such joins into the planner's top-level list of
856 * relations to join, but record information about each outer join in an
857 * OuterJoinInfo struct. These structs are kept in the PlannerInfo node's
860 * min_lefthand and min_righthand are the sets of base relids that must be
861 * available on each side when performing the outer join. lhs_strict is
862 * true if the outer join's condition cannot succeed when the LHS variables
863 * are all NULL (this means that the outer join can commute with upper-level
864 * outer joins even if it appears in their RHS). We don't bother to set
865 * lhs_strict for FULL JOINs, however.
867 * It is not valid for either min_lefthand or min_righthand to be empty sets;
868 * if they were, this would break the logic that enforces join order.
870 * Note: OuterJoinInfo directly represents only LEFT JOIN and FULL JOIN;
871 * RIGHT JOIN is handled by switching the inputs to make it a LEFT JOIN.
872 * We make an OuterJoinInfo for FULL JOINs even though there is no flexibility
873 * of planning for them, because this simplifies make_join_rel()'s API.
876 typedef struct OuterJoinInfo
879 Relids min_lefthand; /* base relids in minimum LHS for join */
880 Relids min_righthand; /* base relids in minimum RHS for join */
881 bool is_full_join; /* it's a FULL OUTER JOIN */
882 bool lhs_strict; /* joinclause is strict for some LHS rel */
888 * When we convert top-level IN quals into join operations, we must restrict
889 * the order of joining and use special join methods at some join points.
890 * We record information about each such IN clause in an InClauseInfo struct.
891 * These structs are kept in the PlannerInfo node's in_info_list.
893 * Note: sub_targetlist is just a list of Vars or expressions; it does not
894 * contain TargetEntry nodes.
897 typedef struct InClauseInfo
900 Relids lefthand; /* base relids in lefthand expressions */
901 Relids righthand; /* base relids coming from the subselect */
902 List *sub_targetlist; /* targetlist of original RHS subquery */
903 List *in_operators; /* OIDs of the IN's equality operator(s) */
907 * Append-relation info.
909 * When we expand an inheritable table or a UNION-ALL subselect into an
910 * "append relation" (essentially, a list of child RTEs), we build an
911 * AppendRelInfo for each child RTE. The list of AppendRelInfos indicates
912 * which child RTEs must be included when expanding the parent, and each
913 * node carries information needed to translate Vars referencing the parent
914 * into Vars referencing that child.
916 * These structs are kept in the PlannerInfo node's append_rel_list.
917 * Note that we just throw all the structs into one list, and scan the
918 * whole list when desiring to expand any one parent. We could have used
919 * a more complex data structure (eg, one list per parent), but this would
920 * be harder to update during operations such as pulling up subqueries,
921 * and not really any easier to scan. Considering that typical queries
922 * will not have many different append parents, it doesn't seem worthwhile
923 * to complicate things.
925 * Note: after completion of the planner prep phase, any given RTE is an
926 * append parent having entries in append_rel_list if and only if its
927 * "inh" flag is set. We clear "inh" for plain tables that turn out not
928 * to have inheritance children, and (in an abuse of the original meaning
929 * of the flag) we set "inh" for subquery RTEs that turn out to be
930 * flattenable UNION ALL queries. This lets us avoid useless searches
931 * of append_rel_list.
933 * Note: the data structure assumes that append-rel members are single
934 * baserels. This is OK for inheritance, but it prevents us from pulling
935 * up a UNION ALL member subquery if it contains a join. While that could
936 * be fixed with a more complex data structure, at present there's not much
937 * point because no improvement in the plan could result.
940 typedef struct AppendRelInfo
945 * These fields uniquely identify this append relationship. There can be
946 * (in fact, always should be) multiple AppendRelInfos for the same
947 * parent_relid, but never more than one per child_relid, since a given
948 * RTE cannot be a child of more than one append parent.
950 Index parent_relid; /* RT index of append parent rel */
951 Index child_relid; /* RT index of append child rel */
954 * For an inheritance appendrel, the parent and child are both regular
955 * relations, and we store their rowtype OIDs here for use in translating
956 * whole-row Vars. For a UNION-ALL appendrel, the parent and child are
957 * both subqueries with no named rowtype, and we store InvalidOid here.
959 Oid parent_reltype; /* OID of parent's composite type */
960 Oid child_reltype; /* OID of child's composite type */
963 * The N'th element of this list is the integer column number of the child
964 * column corresponding to the N'th column of the parent. A list element
965 * is zero if it corresponds to a dropped column of the parent (this is
966 * only possible for inheritance cases, not UNION ALL).
968 List *col_mappings; /* list of child attribute numbers */
971 * The N'th element of this list is a Var or expression representing the
972 * child column corresponding to the N'th column of the parent. This is
973 * used to translate Vars referencing the parent rel into references to
974 * the child. A list element is NULL if it corresponds to a dropped
975 * column of the parent (this is only possible for inheritance cases, not
978 * This might seem redundant with the col_mappings data, but it is handy
979 * because flattening of sub-SELECTs that are members of a UNION ALL will
980 * cause changes in the expressions that need to be substituted for a
981 * parent Var. Adjusting this data structure lets us track what really
982 * needs to be substituted.
984 * Notice we only store entries for user columns (attno > 0). Whole-row
985 * Vars are special-cased, and system columns (attno < 0) need no special
986 * translation since their attnos are the same for all tables.
988 * Caution: the Vars have varlevelsup = 0. Be careful to adjust as needed
989 * when copying into a subquery.
991 List *translated_vars; /* Expressions in the child's Vars */
994 * We store the parent table's OID here for inheritance, or InvalidOid for
995 * UNION ALL. This is only needed to help in generating error messages if
996 * an attempt is made to reference a dropped parent column.
998 Oid parent_reloid; /* OID of parent relation */
1001 #endif /* RELATION_H */