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 1.142 2007/05/22 01:40:33 tgl Exp $
12 *-------------------------------------------------------------------------
17 #include "access/sdir.h"
18 #include "nodes/bitmapset.h"
19 #include "nodes/params.h"
20 #include "nodes/parsenodes.h"
21 #include "storage/block.h"
26 * Set of relation identifiers (indexes into the rangetable).
28 typedef Bitmapset *Relids;
31 * When looking for a "cheapest path", this enum specifies whether we want
32 * cheapest startup cost or cheapest total cost.
34 typedef enum CostSelector
36 STARTUP_COST, TOTAL_COST
40 * The cost estimate produced by cost_qual_eval() includes both a one-time
41 * (startup) cost, and a per-tuple cost.
43 typedef struct QualCost
45 Cost startup; /* one-time cost */
46 Cost per_tuple; /* per-evaluation cost */
52 * Global information for planning/optimization
54 * PlannerGlobal holds state for an entire planner invocation; this state
55 * is shared across all levels of sub-Queries that exist in the command being
59 typedef struct PlannerGlobal
63 ParamListInfo boundParams; /* Param values provided to planner() */
65 List *paramlist; /* to keep track of cross-level Params */
67 List *subplans; /* Plans for SubPlan nodes */
69 List *subrtables; /* Rangetables for SubPlan nodes */
71 Bitmapset *rewindPlanIDs; /* indices of subplans that require REWIND */
73 List *finalrtable; /* "flat" rangetable for executor */
76 /* macro for fetching the Plan associated with a SubPlan node */
77 #define planner_subplan_get_plan(root, subplan) \
78 ((Plan *) list_nth((root)->glob->subplans, (subplan)->plan_id - 1))
83 * Per-query information for planning/optimization
85 * This struct is conventionally called "root" in all the planner routines.
86 * It holds links to all of the planner's working state, in addition to the
87 * original Query. Note that at present the planner extensively modifies
88 * the passed-in Query data structure; someday that should stop.
91 typedef struct PlannerInfo
95 Query *parse; /* the Query being planned */
97 PlannerGlobal *glob; /* global info for current planner run */
99 Index query_level; /* 1 at the outermost Query */
102 * simple_rel_array holds pointers to "base rels" and "other rels" (see
103 * comments for RelOptInfo for more info). It is indexed by rangetable
104 * index (so entry 0 is always wasted). Entries can be NULL when an RTE
105 * does not correspond to a base relation, such as a join RTE or an
106 * unreferenced view RTE; or if the RelOptInfo hasn't been made yet.
108 struct RelOptInfo **simple_rel_array; /* All 1-rel RelOptInfos */
109 int simple_rel_array_size; /* allocated size of array */
112 * simple_rte_array is the same length as simple_rel_array and holds
113 * pointers to the associated rangetable entries. This lets us avoid
114 * rt_fetch(), which can be a bit slow once large inheritance sets have
117 RangeTblEntry **simple_rte_array; /* rangetable as an array */
120 * join_rel_list is a list of all join-relation RelOptInfos we have
121 * considered in this planning run. For small problems we just scan the
122 * list to do lookups, but when there are many join relations we build a
123 * hash table for faster lookups. The hash table is present and valid
124 * when join_rel_hash is not NULL. Note that we still maintain the list
125 * even when using the hash table for lookups; this simplifies life for
128 List *join_rel_list; /* list of join-relation RelOptInfos */
129 struct HTAB *join_rel_hash; /* optional hashtable for join relations */
131 List *resultRelations; /* integer list of RT indexes, or NIL */
133 List *returningLists; /* list of lists of TargetEntry, or NIL */
135 List *init_plans; /* init subplans for query */
137 List *eq_classes; /* list of active EquivalenceClasses */
139 List *canon_pathkeys; /* list of "canonical" PathKeys */
141 List *left_join_clauses; /* list of RestrictInfos for
142 * mergejoinable outer join clauses
143 * w/nonnullable var on left */
145 List *right_join_clauses; /* list of RestrictInfos for
146 * mergejoinable outer join clauses
147 * w/nonnullable var on right */
149 List *full_join_clauses; /* list of RestrictInfos for
150 * mergejoinable full join clauses */
152 List *oj_info_list; /* list of OuterJoinInfos */
154 List *in_info_list; /* list of InClauseInfos */
156 List *append_rel_list; /* list of AppendRelInfos */
158 List *query_pathkeys; /* desired pathkeys for query_planner(), and
159 * actual pathkeys afterwards */
161 List *group_pathkeys; /* groupClause pathkeys, if any */
162 List *sort_pathkeys; /* sortClause pathkeys, if any */
164 MemoryContext planner_cxt; /* context holding PlannerInfo */
166 double total_table_pages; /* # of pages in all tables of query */
168 double tuple_fraction; /* tuple_fraction passed to query_planner */
170 bool hasJoinRTEs; /* true if any RTEs are RTE_JOIN kind */
171 bool hasOuterJoins; /* true if any RTEs are outer joins */
172 bool hasHavingQual; /* true if havingQual was non-null */
173 bool hasPseudoConstantQuals; /* true if any RestrictInfo has
174 * pseudoconstant = true */
179 * In places where it's known that simple_rte_array[] must have been prepared
180 * already, we just index into it to fetch RTEs. In code that might be
181 * executed before or after entering query_planner(), use this macro.
183 #define planner_rt_fetch(rti, root) \
184 ((root)->simple_rte_array ? (root)->simple_rte_array[rti] : \
185 rt_fetch(rti, (root)->parse->rtable))
190 * Per-relation information for planning/optimization
192 * For planning purposes, a "base rel" is either a plain relation (a table)
193 * or the output of a sub-SELECT or function that appears in the range table.
194 * In either case it is uniquely identified by an RT index. A "joinrel"
195 * is the joining of two or more base rels. A joinrel is identified by
196 * the set of RT indexes for its component baserels. We create RelOptInfo
197 * nodes for each baserel and joinrel, and store them in the PlannerInfo's
198 * simple_rel_array and join_rel_list respectively.
200 * Note that there is only one joinrel for any given set of component
201 * baserels, no matter what order we assemble them in; so an unordered
202 * set is the right datatype to identify it with.
204 * We also have "other rels", which are like base rels in that they refer to
205 * single RT indexes; but they are not part of the join tree, and are given
206 * a different RelOptKind to identify them.
208 * Currently the only kind of otherrels are those made for member relations
209 * of an "append relation", that is an inheritance set or UNION ALL subquery.
210 * An append relation has a parent RTE that is a base rel, which represents
211 * the entire append relation. The member RTEs are otherrels. The parent
212 * is present in the query join tree but the members are not. The member
213 * RTEs and otherrels are used to plan the scans of the individual tables or
214 * subqueries of the append set; then the parent baserel is given an Append
215 * plan comprising the best plans for the individual member rels. (See
216 * comments for AppendRelInfo for more information.)
218 * At one time we also made otherrels to represent join RTEs, for use in
219 * handling join alias Vars. Currently this is not needed because all join
220 * alias Vars are expanded to non-aliased form during preprocess_expression.
222 * Parts of this data structure are specific to various scan and join
223 * mechanisms. It didn't seem worth creating new node types for them.
225 * relids - Set of base-relation identifiers; it is a base relation
226 * if there is just one, a join relation if more than one
227 * rows - estimated number of tuples in the relation after restriction
228 * clauses have been applied (ie, output rows of a plan for it)
229 * width - avg. number of bytes per tuple in the relation after the
230 * appropriate projections have been done (ie, output width)
231 * reltargetlist - List of Var nodes for the attributes we need to
232 * output from this relation (in no particular order)
233 * NOTE: in a child relation, may contain RowExprs
234 * pathlist - List of Path nodes, one for each potentially useful
235 * method of generating the relation
236 * cheapest_startup_path - the pathlist member with lowest startup cost
237 * (regardless of its ordering)
238 * cheapest_total_path - the pathlist member with lowest total cost
239 * (regardless of its ordering)
240 * cheapest_unique_path - for caching cheapest path to produce unique
241 * (no duplicates) output from relation
243 * If the relation is a base relation it will have these fields set:
245 * relid - RTE index (this is redundant with the relids field, but
246 * is provided for convenience of access)
247 * rtekind - distinguishes plain relation, subquery, or function RTE
248 * min_attr, max_attr - range of valid AttrNumbers for rel
249 * attr_needed - array of bitmapsets indicating the highest joinrel
250 * in which each attribute is needed; if bit 0 is set then
251 * the attribute is needed as part of final targetlist
252 * attr_widths - cache space for per-attribute width estimates;
253 * zero means not computed yet
254 * indexlist - list of IndexOptInfo nodes for relation's indexes
255 * (always NIL if it's not a table)
256 * pages - number of disk pages in relation (zero if not a table)
257 * tuples - number of tuples in relation (not considering restrictions)
258 * subplan - plan for subquery (NULL if it's not a subquery)
259 * subrtable - rangetable for subquery (NIL if it's not a subquery)
261 * Note: for a subquery, tuples and subplan are not set immediately
262 * upon creation of the RelOptInfo object; they are filled in when
263 * set_base_rel_pathlist processes the object.
265 * For otherrels that are appendrel members, these fields are filled
266 * in just as for a baserel.
268 * The presence of the remaining fields depends on the restrictions
269 * and joins that the relation participates in:
271 * baserestrictinfo - List of RestrictInfo nodes, containing info about
272 * each non-join qualification clause in which this relation
273 * participates (only used for base rels)
274 * baserestrictcost - Estimated cost of evaluating the baserestrictinfo
275 * clauses at a single tuple (only used for base rels)
276 * joininfo - List of RestrictInfo nodes, containing info about each
277 * join clause in which this relation participates (but
278 * note this excludes clauses that might be derivable from
279 * EquivalenceClasses)
280 * has_eclass_joins - flag that EquivalenceClass joins are possible
281 * index_outer_relids - only used for base rels; set of outer relids
282 * that participate in indexable joinclauses for this rel
283 * index_inner_paths - only used for base rels; list of InnerIndexscanInfo
284 * nodes showing best indexpaths for various subsets of
285 * index_outer_relids.
287 * Note: Keeping a restrictinfo list in the RelOptInfo is useful only for
288 * base rels, because for a join rel the set of clauses that are treated as
289 * restrict clauses varies depending on which sub-relations we choose to join.
290 * (For example, in a 3-base-rel join, a clause relating rels 1 and 2 must be
291 * treated as a restrictclause if we join {1} and {2 3} to make {1 2 3}; but
292 * if we join {1 2} and {3} then that clause will be a restrictclause in {1 2}
293 * and should not be processed again at the level of {1 2 3}.) Therefore,
294 * the restrictinfo list in the join case appears in individual JoinPaths
295 * (field joinrestrictinfo), not in the parent relation. But it's OK for
296 * the RelOptInfo to store the joininfo list, because that is the same
297 * for a given rel no matter how we form it.
299 * We store baserestrictcost in the RelOptInfo (for base relations) because
300 * we know we will need it at least once (to price the sequential scan)
301 * and may need it multiple times to price index scans.
304 typedef enum RelOptKind
308 RELOPT_OTHER_MEMBER_REL
311 typedef struct RelOptInfo
315 RelOptKind reloptkind;
317 /* all relations included in this RelOptInfo */
318 Relids relids; /* set of base relids (rangetable indexes) */
320 /* size estimates generated by planner */
321 double rows; /* estimated number of result tuples */
322 int width; /* estimated avg width of result tuples */
324 /* materialization information */
325 List *reltargetlist; /* needed Vars */
326 List *pathlist; /* Path structures */
327 struct Path *cheapest_startup_path;
328 struct Path *cheapest_total_path;
329 struct Path *cheapest_unique_path;
331 /* information about a base rel (not set for join rels!) */
333 RTEKind rtekind; /* RELATION, SUBQUERY, or FUNCTION */
334 AttrNumber min_attr; /* smallest attrno of rel (often <0) */
335 AttrNumber max_attr; /* largest attrno of rel */
336 Relids *attr_needed; /* array indexed [min_attr .. max_attr] */
337 int32 *attr_widths; /* array indexed [min_attr .. max_attr] */
341 struct Plan *subplan; /* if subquery */
342 List *subrtable; /* if subquery */
344 /* used by various scans and joins: */
345 List *baserestrictinfo; /* RestrictInfo structures (if base
347 QualCost baserestrictcost; /* cost of evaluating the above */
348 List *joininfo; /* RestrictInfo structures for join clauses
349 * involving this rel */
350 bool has_eclass_joins; /* T means joininfo is incomplete */
352 /* cached info about inner indexscan paths for relation: */
353 Relids index_outer_relids; /* other relids in indexable join
355 List *index_inner_paths; /* InnerIndexscanInfo nodes */
358 * Inner indexscans are not in the main pathlist because they are not
359 * usable except in specific join contexts. We use the index_inner_paths
360 * list just to avoid recomputing the best inner indexscan repeatedly for
361 * similar outer relations. See comments for InnerIndexscanInfo.
367 * Per-index information for planning/optimization
369 * Prior to Postgres 7.0, RelOptInfo was used to describe both relations
370 * and indexes, but that created confusion without actually doing anything
371 * useful. So now we have a separate IndexOptInfo struct for indexes.
373 * opfamily[], indexkeys[], fwdsortop[], revsortop[], and nulls_first[]
374 * each have ncolumns entries. Note: for historical reasons, the
375 * opfamily array has an extra entry that is always zero. Some code
376 * scans until it sees a zero entry, rather than looking at ncolumns.
378 * Zeroes in the indexkeys[] array indicate index columns that are
379 * expressions; there is one element in indexprs for each such column.
381 * For an unordered index, the sortop arrays contains zeroes. Note that
382 * fwdsortop[] and nulls_first[] describe the sort ordering of a forward
383 * indexscan; we can also consider a backward indexscan, which will
384 * generate sort order described by revsortop/!nulls_first.
386 * The indexprs and indpred expressions have been run through
387 * prepqual.c and eval_const_expressions() for ease of matching to
388 * WHERE clauses. indpred is in implicit-AND form.
390 typedef struct IndexOptInfo
394 Oid indexoid; /* OID of the index relation */
395 RelOptInfo *rel; /* back-link to index's table */
397 /* statistics from pg_class */
398 BlockNumber pages; /* number of disk pages in index */
399 double tuples; /* number of index tuples in index */
401 /* index descriptor information */
402 int ncolumns; /* number of columns in index */
403 Oid *opfamily; /* OIDs of operator families for columns */
404 int *indexkeys; /* column numbers of index's keys, or 0 */
405 Oid *fwdsortop; /* OIDs of sort operators for each column */
406 Oid *revsortop; /* OIDs of sort operators for backward scan */
407 bool *nulls_first; /* do NULLs come first in the sort order? */
408 Oid relam; /* OID of the access method (in pg_am) */
410 RegProcedure amcostestimate; /* OID of the access method's cost fcn */
412 List *indexprs; /* expressions for non-simple index columns */
413 List *indpred; /* predicate if a partial index, else NIL */
415 bool predOK; /* true if predicate matches query */
416 bool unique; /* true if a unique index */
417 bool amoptionalkey; /* can query omit key for the first column? */
418 bool amsearchnulls; /* can AM search for NULL index entries? */
425 * Whenever we can determine that a mergejoinable equality clause A = B is
426 * not delayed by any outer join, we create an EquivalenceClass containing
427 * the expressions A and B to record this knowledge. If we later find another
428 * equivalence B = C, we add C to the existing EquivalenceClass; this may
429 * require merging two existing EquivalenceClasses. At the end of the qual
430 * distribution process, we have sets of values that are known all transitively
431 * equal to each other, where "equal" is according to the rules of the btree
432 * operator family(s) shown in ec_opfamilies. (We restrict an EC to contain
433 * only equalities whose operators belong to the same set of opfamilies. This
434 * could probably be relaxed, but for now it's not worth the trouble, since
435 * nearly all equality operators belong to only one btree opclass anyway.)
437 * We also use EquivalenceClasses as the base structure for PathKeys, letting
438 * us represent knowledge about different sort orderings being equivalent.
439 * Since every PathKey must reference an EquivalenceClass, we will end up
440 * with single-member EquivalenceClasses whenever a sort key expression has
441 * not been equivalenced to anything else. It is also possible that such an
442 * EquivalenceClass will contain a volatile expression ("ORDER BY random()"),
443 * which is a case that can't arise otherwise since clauses containing
444 * volatile functions are never considered mergejoinable. We mark such
445 * EquivalenceClasses specially to prevent them from being merged with
446 * ordinary EquivalenceClasses.
448 * We allow equality clauses appearing below the nullable side of an outer join
449 * to form EquivalenceClasses, but these have a slightly different meaning:
450 * the included values might be all NULL rather than all the same non-null
451 * values. See src/backend/optimizer/README for more on that point.
453 * NB: if ec_merged isn't NULL, this class has been merged into another, and
454 * should be ignored in favor of using the pointed-to class.
456 typedef struct EquivalenceClass
460 List *ec_opfamilies; /* btree operator family OIDs */
461 List *ec_members; /* list of EquivalenceMembers */
462 List *ec_sources; /* list of generating RestrictInfos */
463 List *ec_derives; /* list of derived RestrictInfos */
464 Relids ec_relids; /* all relids appearing in ec_members */
465 bool ec_has_const; /* any pseudoconstants in ec_members? */
466 bool ec_has_volatile; /* the (sole) member is a volatile expr */
467 bool ec_below_outer_join; /* equivalence applies below an OJ */
468 bool ec_broken; /* failed to generate needed clauses? */
469 struct EquivalenceClass *ec_merged; /* set if merged into another EC */
473 * EquivalenceMember - one member expression of an EquivalenceClass
475 * em_is_child signifies that this element was built by transposing a member
476 * for an inheritance parent relation to represent the corresponding expression
477 * on an inheritance child. The element should be ignored for all purposes
478 * except constructing inner-indexscan paths for the child relation. (Other
479 * types of join are driven from transposed joininfo-list entries.) Note
480 * that the EC's ec_relids field does NOT include the child relation.
482 * em_datatype is usually the same as exprType(em_expr), but can be
483 * different when dealing with a binary-compatible opfamily; in particular
484 * anyarray_ops would never work without this. Use em_datatype when
485 * looking up a specific btree operator to work with this expression.
487 typedef struct EquivalenceMember
491 Expr *em_expr; /* the expression represented */
492 Relids em_relids; /* all relids appearing in em_expr */
493 bool em_is_const; /* expression is pseudoconstant? */
494 bool em_is_child; /* derived version for a child relation? */
495 Oid em_datatype; /* the "nominal type" used by the opfamily */
501 * The sort ordering of a path is represented by a list of PathKey nodes.
502 * An empty list implies no known ordering. Otherwise the first item
503 * represents the primary sort key, the second the first secondary sort key,
504 * etc. The value being sorted is represented by linking to an
505 * EquivalenceClass containing that value and including pk_opfamily among its
506 * ec_opfamilies. This is a convenient method because it makes it trivial
507 * to detect equivalent and closely-related orderings. (See optimizer/README
508 * for more information.)
510 * Note: pk_strategy is either BTLessStrategyNumber (for ASC) or
511 * BTGreaterStrategyNumber (for DESC). We assume that all ordering-capable
512 * index types will use btree-compatible strategy numbers.
515 typedef struct PathKey
519 EquivalenceClass *pk_eclass; /* the value that is ordered */
520 Oid pk_opfamily; /* btree opfamily defining the ordering */
521 int pk_strategy; /* sort direction (ASC or DESC) */
522 bool pk_nulls_first; /* do NULLs come before normal values? */
526 * Type "Path" is used as-is for sequential-scan paths. For other
527 * path types it is the first component of a larger struct.
529 * Note: "pathtype" is the NodeTag of the Plan node we could build from this
530 * Path. It is partially redundant with the Path's NodeTag, but allows us
531 * to use the same Path type for multiple Plan types where there is no need
532 * to distinguish the Plan type during path processing.
539 NodeTag pathtype; /* tag identifying scan/join method */
541 RelOptInfo *parent; /* the relation this path can build */
543 /* estimated execution costs for path (see costsize.c for more info) */
544 Cost startup_cost; /* cost expended before fetching any tuples */
545 Cost total_cost; /* total cost (assuming all tuples fetched) */
547 List *pathkeys; /* sort ordering of path's output */
548 /* pathkeys is a List of PathKey nodes; see above */
552 * IndexPath represents an index scan over a single index.
554 * 'indexinfo' is the index to be scanned.
556 * 'indexclauses' is a list of index qualification clauses, with implicit
557 * AND semantics across the list. Each clause is a RestrictInfo node from
558 * the query's WHERE or JOIN conditions.
560 * 'indexquals' has the same structure as 'indexclauses', but it contains
561 * the actual indexqual conditions that can be used with the index.
562 * In simple cases this is identical to 'indexclauses', but when special
563 * indexable operators appear in 'indexclauses', they are replaced by the
564 * derived indexscannable conditions in 'indexquals'.
566 * 'isjoininner' is TRUE if the path is a nestloop inner scan (that is,
567 * some of the index conditions are join rather than restriction clauses).
568 * Note that the path costs will be calculated differently from a plain
569 * indexscan in this case, and in addition there's a special 'rows' value
570 * different from the parent RelOptInfo's (see below).
572 * 'indexscandir' is one of:
573 * ForwardScanDirection: forward scan of an ordered index
574 * BackwardScanDirection: backward scan of an ordered index
575 * NoMovementScanDirection: scan of an unordered index, or don't care
576 * (The executor doesn't care whether it gets ForwardScanDirection or
577 * NoMovementScanDirection for an indexscan, but the planner wants to
578 * distinguish ordered from unordered indexes for building pathkeys.)
580 * 'indextotalcost' and 'indexselectivity' are saved in the IndexPath so that
581 * we need not recompute them when considering using the same index in a
582 * bitmap index/heap scan (see BitmapHeapPath). The costs of the IndexPath
583 * itself represent the costs of an IndexScan plan type.
585 * 'rows' is the estimated result tuple count for the indexscan. This
586 * is the same as path.parent->rows for a simple indexscan, but it is
587 * different for a nestloop inner scan, because the additional indexquals
588 * coming from join clauses make the scan more selective than the parent
589 * rel's restrict clauses alone would do.
592 typedef struct IndexPath
595 IndexOptInfo *indexinfo;
599 ScanDirection indexscandir;
601 Selectivity indexselectivity;
602 double rows; /* estimated number of result tuples */
606 * BitmapHeapPath represents one or more indexscans that generate TID bitmaps
607 * instead of directly accessing the heap, followed by AND/OR combinations
608 * to produce a single bitmap, followed by a heap scan that uses the bitmap.
609 * Note that the output is always considered unordered, since it will come
610 * out in physical heap order no matter what the underlying indexes did.
612 * The individual indexscans are represented by IndexPath nodes, and any
613 * logic on top of them is represented by a tree of BitmapAndPath and
614 * BitmapOrPath nodes. Notice that we can use the same IndexPath node both
615 * to represent a regular IndexScan plan, and as the child of a BitmapHeapPath
616 * that represents scanning the same index using a BitmapIndexScan. The
617 * startup_cost and total_cost figures of an IndexPath always represent the
618 * costs to use it as a regular IndexScan. The costs of a BitmapIndexScan
619 * can be computed using the IndexPath's indextotalcost and indexselectivity.
621 * BitmapHeapPaths can be nestloop inner indexscans. The isjoininner and
622 * rows fields serve the same purpose as for plain IndexPaths.
624 typedef struct BitmapHeapPath
627 Path *bitmapqual; /* IndexPath, BitmapAndPath, BitmapOrPath */
628 bool isjoininner; /* T if it's a nestloop inner scan */
629 double rows; /* estimated number of result tuples */
633 * BitmapAndPath represents a BitmapAnd plan node; it can only appear as
634 * part of the substructure of a BitmapHeapPath. The Path structure is
635 * a bit more heavyweight than we really need for this, but for simplicity
636 * we make it a derivative of Path anyway.
638 typedef struct BitmapAndPath
641 List *bitmapquals; /* IndexPaths and BitmapOrPaths */
642 Selectivity bitmapselectivity;
646 * BitmapOrPath represents a BitmapOr plan node; it can only appear as
647 * part of the substructure of a BitmapHeapPath. The Path structure is
648 * a bit more heavyweight than we really need for this, but for simplicity
649 * we make it a derivative of Path anyway.
651 typedef struct BitmapOrPath
654 List *bitmapquals; /* IndexPaths and BitmapAndPaths */
655 Selectivity bitmapselectivity;
659 * TidPath represents a scan by TID
661 * tidquals is an implicitly OR'ed list of qual expressions of the form
662 * "CTID = pseudoconstant" or "CTID = ANY(pseudoconstant_array)".
663 * Note they are bare expressions, not RestrictInfos.
665 typedef struct TidPath
668 List *tidquals; /* qual(s) involving CTID = something */
672 * AppendPath represents an Append plan, ie, successive execution of
673 * several member plans.
675 * Note: it is possible for "subpaths" to contain only one, or even no,
676 * elements. These cases are optimized during create_append_plan.
678 typedef struct AppendPath
681 List *subpaths; /* list of component Paths */
685 * ResultPath represents use of a Result plan node to compute a variable-free
686 * targetlist with no underlying tables (a "SELECT expressions" query).
687 * The query could have a WHERE clause, too, represented by "quals".
689 * Note that quals is a list of bare clauses, not RestrictInfos.
691 typedef struct ResultPath
698 * MaterialPath represents use of a Material plan node, i.e., caching of
699 * the output of its subpath. This is used when the subpath is expensive
700 * and needs to be scanned repeatedly, or when we need mark/restore ability
701 * and the subpath doesn't have it.
703 typedef struct MaterialPath
710 * UniquePath represents elimination of distinct rows from the output of
713 * This is unlike the other Path nodes in that it can actually generate
714 * different plans: either hash-based or sort-based implementation, or a
715 * no-op if the input path can be proven distinct already. The decision
716 * is sufficiently localized that it's not worth having separate Path node
717 * types. (Note: in the no-op case, we could eliminate the UniquePath node
718 * entirely and just return the subpath; but it's convenient to have a
719 * UniquePath in the path tree to signal upper-level routines that the input
720 * is known distinct.)
724 UNIQUE_PATH_NOOP, /* input is known unique already */
725 UNIQUE_PATH_HASH, /* use hashing */
726 UNIQUE_PATH_SORT /* use sorting */
729 typedef struct UniquePath
733 UniquePathMethod umethod;
734 double rows; /* estimated number of result tuples */
738 * All join-type paths share these fields.
741 typedef struct JoinPath
747 Path *outerjoinpath; /* path for the outer side of the join */
748 Path *innerjoinpath; /* path for the inner side of the join */
750 List *joinrestrictinfo; /* RestrictInfos to apply to join */
753 * See the notes for RelOptInfo to understand why joinrestrictinfo is
754 * needed in JoinPath, and can't be merged into the parent RelOptInfo.
759 * A nested-loop path needs no special fields.
762 typedef JoinPath NestPath;
765 * A mergejoin path has these fields.
767 * path_mergeclauses lists the clauses (in the form of RestrictInfos)
768 * that will be used in the merge.
770 * Note that the mergeclauses are a subset of the parent relation's
771 * restriction-clause list. Any join clauses that are not mergejoinable
772 * appear only in the parent's restrict list, and must be checked by a
773 * qpqual at execution time.
775 * outersortkeys (resp. innersortkeys) is NIL if the outer path
776 * (resp. inner path) is already ordered appropriately for the
777 * mergejoin. If it is not NIL then it is a PathKeys list describing
778 * the ordering that must be created by an explicit sort step.
781 typedef struct MergePath
784 List *path_mergeclauses; /* join clauses to be used for merge */
785 List *outersortkeys; /* keys for explicit sort, if any */
786 List *innersortkeys; /* keys for explicit sort, if any */
790 * A hashjoin path has these fields.
792 * The remarks above for mergeclauses apply for hashclauses as well.
794 * Hashjoin does not care what order its inputs appear in, so we have
795 * no need for sortkeys.
798 typedef struct HashPath
801 List *path_hashclauses; /* join clauses used for hashing */
805 * Restriction clause info.
807 * We create one of these for each AND sub-clause of a restriction condition
808 * (WHERE or JOIN/ON clause). Since the restriction clauses are logically
809 * ANDed, we can use any one of them or any subset of them to filter out
810 * tuples, without having to evaluate the rest. The RestrictInfo node itself
811 * stores data used by the optimizer while choosing the best query plan.
813 * If a restriction clause references a single base relation, it will appear
814 * in the baserestrictinfo list of the RelOptInfo for that base rel.
816 * If a restriction clause references more than one base rel, it will
817 * appear in the joininfo list of every RelOptInfo that describes a strict
818 * subset of the base rels mentioned in the clause. The joininfo lists are
819 * used to drive join tree building by selecting plausible join candidates.
820 * The clause cannot actually be applied until we have built a join rel
821 * containing all the base rels it references, however.
823 * When we construct a join rel that includes all the base rels referenced
824 * in a multi-relation restriction clause, we place that clause into the
825 * joinrestrictinfo lists of paths for the join rel, if neither left nor
826 * right sub-path includes all base rels referenced in the clause. The clause
827 * will be applied at that join level, and will not propagate any further up
828 * the join tree. (Note: the "predicate migration" code was once intended to
829 * push restriction clauses up and down the plan tree based on evaluation
830 * costs, but it's dead code and is unlikely to be resurrected in the
831 * foreseeable future.)
833 * Note that in the presence of more than two rels, a multi-rel restriction
834 * might reach different heights in the join tree depending on the join
835 * sequence we use. So, these clauses cannot be associated directly with
836 * the join RelOptInfo, but must be kept track of on a per-join-path basis.
838 * RestrictInfos that represent equivalence conditions (i.e., mergejoinable
839 * equalities that are not outerjoin-delayed) are handled a bit differently.
840 * Initially we attach them to the EquivalenceClasses that are derived from
841 * them. When we construct a scan or join path, we look through all the
842 * EquivalenceClasses and generate derived RestrictInfos representing the
843 * minimal set of conditions that need to be checked for this particular scan
844 * or join to enforce that all members of each EquivalenceClass are in fact
845 * equal in all rows emitted by the scan or join.
847 * When dealing with outer joins we have to be very careful about pushing qual
848 * clauses up and down the tree. An outer join's own JOIN/ON conditions must
849 * be evaluated exactly at that join node, and any quals appearing in WHERE or
850 * in a JOIN above the outer join cannot be pushed down below the outer join.
851 * Otherwise the outer join will produce wrong results because it will see the
852 * wrong sets of input rows. All quals are stored as RestrictInfo nodes
853 * during planning, but there's a flag to indicate whether a qual has been
854 * pushed down to a lower level than its original syntactic placement in the
855 * join tree would suggest. If an outer join prevents us from pushing a qual
856 * down to its "natural" semantic level (the level associated with just the
857 * base rels used in the qual) then we mark the qual with a "required_relids"
858 * value including more than just the base rels it actually uses. By
859 * pretending that the qual references all the rels appearing in the outer
860 * join, we prevent it from being evaluated below the outer join's joinrel.
861 * When we do form the outer join's joinrel, we still need to distinguish
862 * those quals that are actually in that join's JOIN/ON condition from those
863 * that appeared elsewhere in the tree and were pushed down to the join rel
864 * because they used no other rels. That's what the is_pushed_down flag is
865 * for; it tells us that a qual is not an OUTER JOIN qual for the set of base
866 * rels listed in required_relids. A clause that originally came from WHERE
867 * or an INNER JOIN condition will *always* have its is_pushed_down flag set.
868 * It's possible for an OUTER JOIN clause to be marked is_pushed_down too,
869 * if we decide that it can be pushed down into the nullable side of the join.
870 * In that case it acts as a plain filter qual for wherever it gets evaluated.
872 * When application of a qual must be delayed by outer join, we also mark it
873 * with outerjoin_delayed = true. This isn't redundant with required_relids
874 * because that might equal clause_relids whether or not it's an outer-join
877 * In general, the referenced clause might be arbitrarily complex. The
878 * kinds of clauses we can handle as indexscan quals, mergejoin clauses,
879 * or hashjoin clauses are limited (e.g., no volatile functions). The code
880 * for each kind of path is responsible for identifying the restrict clauses
881 * it can use and ignoring the rest. Clauses not implemented by an indexscan,
882 * mergejoin, or hashjoin will be placed in the plan qual or joinqual field
883 * of the finished Plan node, where they will be enforced by general-purpose
884 * qual-expression-evaluation code. (But we are still entitled to count
885 * their selectivity when estimating the result tuple count, if we
886 * can guess what it is...)
888 * When the referenced clause is an OR clause, we generate a modified copy
889 * in which additional RestrictInfo nodes are inserted below the top-level
890 * OR/AND structure. This is a convenience for OR indexscan processing:
891 * indexquals taken from either the top level or an OR subclause will have
892 * associated RestrictInfo nodes.
894 * The can_join flag is set true if the clause looks potentially useful as
895 * a merge or hash join clause, that is if it is a binary opclause with
896 * nonoverlapping sets of relids referenced in the left and right sides.
897 * (Whether the operator is actually merge or hash joinable isn't checked,
900 * The pseudoconstant flag is set true if the clause contains no Vars of
901 * the current query level and no volatile functions. Such a clause can be
902 * pulled out and used as a one-time qual in a gating Result node. We keep
903 * pseudoconstant clauses in the same lists as other RestrictInfos so that
904 * the regular clause-pushing machinery can assign them to the correct join
905 * level, but they need to be treated specially for cost and selectivity
906 * estimates. Note that a pseudoconstant clause can never be an indexqual
907 * or merge or hash join clause, so it's of no interest to large parts of
910 * When join clauses are generated from EquivalenceClasses, there may be
911 * several equally valid ways to enforce join equivalence, of which we need
912 * apply only one. We mark clauses of this kind by setting parent_ec to
913 * point to the generating EquivalenceClass. Multiple clauses with the same
914 * parent_ec in the same join are redundant.
917 typedef struct RestrictInfo
921 Expr *clause; /* the represented clause of WHERE or JOIN */
923 bool is_pushed_down; /* TRUE if clause was pushed down in level */
925 bool outerjoin_delayed; /* TRUE if delayed by outer join */
927 bool can_join; /* see comment above */
929 bool pseudoconstant; /* see comment above */
931 /* The set of relids (varnos) actually referenced in the clause: */
932 Relids clause_relids;
934 /* The set of relids required to evaluate the clause: */
935 Relids required_relids;
937 /* These fields are set for any binary opclause: */
938 Relids left_relids; /* relids in left side of clause */
939 Relids right_relids; /* relids in right side of clause */
941 /* This field is NULL unless clause is an OR clause: */
942 Expr *orclause; /* modified clause with RestrictInfos */
944 /* This field is NULL unless clause is potentially redundant: */
945 EquivalenceClass *parent_ec; /* generating EquivalenceClass */
947 /* cache space for cost and selectivity */
948 QualCost eval_cost; /* eval cost of clause; -1 if not yet set */
949 Selectivity this_selec; /* selectivity; -1 if not yet set */
951 /* valid if clause is mergejoinable, else NIL */
952 List *mergeopfamilies; /* opfamilies containing clause operator */
954 /* cache space for mergeclause processing; NULL if not yet set */
955 EquivalenceClass *left_ec; /* EquivalenceClass containing lefthand */
956 EquivalenceClass *right_ec; /* EquivalenceClass containing righthand */
957 EquivalenceMember *left_em; /* EquivalenceMember for lefthand */
958 EquivalenceMember *right_em; /* EquivalenceMember for righthand */
959 List *scansel_cache; /* list of MergeScanSelCache structs */
961 /* transient workspace for use while considering a specific join path */
962 bool outer_is_left; /* T = outer var on left, F = on right */
964 /* valid if clause is hashjoinable, else InvalidOid: */
965 Oid hashjoinoperator; /* copy of clause operator */
967 /* cache space for hashclause processing; -1 if not yet set */
968 Selectivity left_bucketsize; /* avg bucketsize of left side */
969 Selectivity right_bucketsize; /* avg bucketsize of right side */
973 * Since mergejoinscansel() is a relatively expensive function, and would
974 * otherwise be invoked many times while planning a large join tree,
975 * we go out of our way to cache its results. Each mergejoinable
976 * RestrictInfo carries a list of the specific sort orderings that have
977 * been considered for use with it, and the resulting selectivities.
979 typedef struct MergeScanSelCache
981 /* Ordering details (cache lookup key) */
982 Oid opfamily; /* btree opfamily defining the ordering */
983 int strategy; /* sort direction (ASC or DESC) */
984 bool nulls_first; /* do NULLs come before normal values? */
986 Selectivity leftscansel; /* scan fraction for clause left side */
987 Selectivity rightscansel; /* scan fraction for clause right side */
991 * Inner indexscan info.
993 * An inner indexscan is one that uses one or more joinclauses as index
994 * conditions (perhaps in addition to plain restriction clauses). So it
995 * can only be used as the inner path of a nestloop join where the outer
996 * relation includes all other relids appearing in those joinclauses.
997 * The set of usable joinclauses, and thus the best inner indexscan,
998 * thus varies depending on which outer relation we consider; so we have
999 * to recompute the best such paths for every join. To avoid lots of
1000 * redundant computation, we cache the results of such searches. For
1001 * each relation we compute the set of possible otherrelids (all relids
1002 * appearing in joinquals that could become indexquals for this table).
1003 * Two outer relations whose relids have the same intersection with this
1004 * set will have the same set of available joinclauses and thus the same
1005 * best inner indexscans for the inner relation. By taking the intersection
1006 * before scanning the cache, we avoid recomputing when considering
1007 * join rels that differ only by the inclusion of irrelevant other rels.
1009 * The search key also includes a bool showing whether the join being
1010 * considered is an outer join. Since we constrain the join order for
1011 * outer joins, I believe that this bool can only have one possible value
1012 * for any particular lookup key; but store it anyway to avoid confusion.
1015 typedef struct InnerIndexscanInfo
1018 /* The lookup key: */
1019 Relids other_relids; /* a set of relevant other relids */
1020 bool isouterjoin; /* true if join is outer */
1021 /* Best paths for this lookup key (NULL if no available indexscans): */
1022 Path *cheapest_startup_innerpath; /* cheapest startup cost */
1023 Path *cheapest_total_innerpath; /* cheapest total cost */
1024 } InnerIndexscanInfo;
1029 * One-sided outer joins constrain the order of joining partially but not
1030 * completely. We flatten such joins into the planner's top-level list of
1031 * relations to join, but record information about each outer join in an
1032 * OuterJoinInfo struct. These structs are kept in the PlannerInfo node's
1035 * min_lefthand and min_righthand are the sets of base relids that must be
1036 * available on each side when performing the outer join. lhs_strict is
1037 * true if the outer join's condition cannot succeed when the LHS variables
1038 * are all NULL (this means that the outer join can commute with upper-level
1039 * outer joins even if it appears in their RHS). We don't bother to set
1040 * lhs_strict for FULL JOINs, however.
1042 * It is not valid for either min_lefthand or min_righthand to be empty sets;
1043 * if they were, this would break the logic that enforces join order.
1045 * Note: OuterJoinInfo directly represents only LEFT JOIN and FULL JOIN;
1046 * RIGHT JOIN is handled by switching the inputs to make it a LEFT JOIN.
1047 * We make an OuterJoinInfo for FULL JOINs even though there is no flexibility
1048 * of planning for them, because this simplifies make_join_rel()'s API.
1051 typedef struct OuterJoinInfo
1054 Relids min_lefthand; /* base relids in minimum LHS for join */
1055 Relids min_righthand; /* base relids in minimum RHS for join */
1056 bool is_full_join; /* it's a FULL OUTER JOIN */
1057 bool lhs_strict; /* joinclause is strict for some LHS rel */
1063 * When we convert top-level IN quals into join operations, we must restrict
1064 * the order of joining and use special join methods at some join points.
1065 * We record information about each such IN clause in an InClauseInfo struct.
1066 * These structs are kept in the PlannerInfo node's in_info_list.
1068 * Note: sub_targetlist is just a list of Vars or expressions; it does not
1069 * contain TargetEntry nodes.
1072 typedef struct InClauseInfo
1075 Relids lefthand; /* base relids in lefthand expressions */
1076 Relids righthand; /* base relids coming from the subselect */
1077 List *sub_targetlist; /* targetlist of original RHS subquery */
1078 List *in_operators; /* OIDs of the IN's equality operator(s) */
1082 * Append-relation info.
1084 * When we expand an inheritable table or a UNION-ALL subselect into an
1085 * "append relation" (essentially, a list of child RTEs), we build an
1086 * AppendRelInfo for each child RTE. The list of AppendRelInfos indicates
1087 * which child RTEs must be included when expanding the parent, and each
1088 * node carries information needed to translate Vars referencing the parent
1089 * into Vars referencing that child.
1091 * These structs are kept in the PlannerInfo node's append_rel_list.
1092 * Note that we just throw all the structs into one list, and scan the
1093 * whole list when desiring to expand any one parent. We could have used
1094 * a more complex data structure (eg, one list per parent), but this would
1095 * be harder to update during operations such as pulling up subqueries,
1096 * and not really any easier to scan. Considering that typical queries
1097 * will not have many different append parents, it doesn't seem worthwhile
1098 * to complicate things.
1100 * Note: after completion of the planner prep phase, any given RTE is an
1101 * append parent having entries in append_rel_list if and only if its
1102 * "inh" flag is set. We clear "inh" for plain tables that turn out not
1103 * to have inheritance children, and (in an abuse of the original meaning
1104 * of the flag) we set "inh" for subquery RTEs that turn out to be
1105 * flattenable UNION ALL queries. This lets us avoid useless searches
1106 * of append_rel_list.
1108 * Note: the data structure assumes that append-rel members are single
1109 * baserels. This is OK for inheritance, but it prevents us from pulling
1110 * up a UNION ALL member subquery if it contains a join. While that could
1111 * be fixed with a more complex data structure, at present there's not much
1112 * point because no improvement in the plan could result.
1115 typedef struct AppendRelInfo
1120 * These fields uniquely identify this append relationship. There can be
1121 * (in fact, always should be) multiple AppendRelInfos for the same
1122 * parent_relid, but never more than one per child_relid, since a given
1123 * RTE cannot be a child of more than one append parent.
1125 Index parent_relid; /* RT index of append parent rel */
1126 Index child_relid; /* RT index of append child rel */
1129 * For an inheritance appendrel, the parent and child are both regular
1130 * relations, and we store their rowtype OIDs here for use in translating
1131 * whole-row Vars. For a UNION-ALL appendrel, the parent and child are
1132 * both subqueries with no named rowtype, and we store InvalidOid here.
1134 Oid parent_reltype; /* OID of parent's composite type */
1135 Oid child_reltype; /* OID of child's composite type */
1138 * The N'th element of this list is the integer column number of the child
1139 * column corresponding to the N'th column of the parent. A list element
1140 * is zero if it corresponds to a dropped column of the parent (this is
1141 * only possible for inheritance cases, not UNION ALL).
1143 List *col_mappings; /* list of child attribute numbers */
1146 * The N'th element of this list is a Var or expression representing the
1147 * child column corresponding to the N'th column of the parent. This is
1148 * used to translate Vars referencing the parent rel into references to
1149 * the child. A list element is NULL if it corresponds to a dropped
1150 * column of the parent (this is only possible for inheritance cases, not
1153 * This might seem redundant with the col_mappings data, but it is handy
1154 * because flattening of sub-SELECTs that are members of a UNION ALL will
1155 * cause changes in the expressions that need to be substituted for a
1156 * parent Var. Adjusting this data structure lets us track what really
1157 * needs to be substituted.
1159 * Notice we only store entries for user columns (attno > 0). Whole-row
1160 * Vars are special-cased, and system columns (attno < 0) need no special
1161 * translation since their attnos are the same for all tables.
1163 * Caution: the Vars have varlevelsup = 0. Be careful to adjust as needed
1164 * when copying into a subquery.
1166 List *translated_vars; /* Expressions in the child's Vars */
1169 * We store the parent table's OID here for inheritance, or InvalidOid for
1170 * UNION ALL. This is only needed to help in generating error messages if
1171 * an attempt is made to reference a dropped parent column.
1173 Oid parent_reloid; /* OID of parent relation */
1177 * glob->paramlist keeps track of the PARAM_EXEC slots that we have decided
1178 * we need for the query. At runtime these slots are used to pass values
1179 * either down into subqueries (for outer references in subqueries) or up out
1180 * of subqueries (for the results of a subplan). The n'th entry in the list
1181 * (n counts from 0) corresponds to Param->paramid = n.
1183 * Each paramlist item shows the absolute query level it is associated with,
1184 * where the outermost query is level 1 and nested subqueries have higher
1185 * numbers. The item the parameter slot represents can be one of three kinds:
1187 * A Var: the slot represents a variable of that level that must be passed
1188 * down because subqueries have outer references to it. The varlevelsup
1189 * value in the Var will always be zero.
1191 * An Aggref (with an expression tree representing its argument): the slot
1192 * represents an aggregate expression that is an outer reference for some
1193 * subquery. The Aggref itself has agglevelsup = 0, and its argument tree
1194 * is adjusted to match in level.
1196 * A Param: the slot holds the result of a subplan (it is a setParam item
1197 * for that subplan). The absolute level shown for such items corresponds
1198 * to the parent query of the subplan.
1200 * Note: we detect duplicate Var parameters and coalesce them into one slot,
1201 * but we do not do this for Aggref or Param slots.
1203 typedef struct PlannerParamItem
1207 Node *item; /* the Var, Aggref, or Param */
1208 Index abslevel; /* its absolute query level */
1211 #endif /* RELATION_H */