1 /*-------------------------------------------------------------------------
4 * Definitions for planner's internal data structures.
7 * Portions Copyright (c) 1996-2008, PostgreSQL Global Development Group
8 * Portions Copyright (c) 1994, Regents of the University of California
10 * $PostgreSQL: pgsql/src/include/nodes/relation.h,v 1.154 2008/01/11 04:02:18 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 */
75 List *relationOids; /* OIDs of relations the plan depends on */
77 bool transientPlan; /* redo plan when TransactionXmin changes? */
80 /* macro for fetching the Plan associated with a SubPlan node */
81 #define planner_subplan_get_plan(root, subplan) \
82 ((Plan *) list_nth((root)->glob->subplans, (subplan)->plan_id - 1))
87 * Per-query information for planning/optimization
89 * This struct is conventionally called "root" in all the planner routines.
90 * It holds links to all of the planner's working state, in addition to the
91 * original Query. Note that at present the planner extensively modifies
92 * the passed-in Query data structure; someday that should stop.
95 typedef struct PlannerInfo
99 Query *parse; /* the Query being planned */
101 PlannerGlobal *glob; /* global info for current planner run */
103 Index query_level; /* 1 at the outermost Query */
106 * simple_rel_array holds pointers to "base rels" and "other rels" (see
107 * comments for RelOptInfo for more info). It is indexed by rangetable
108 * index (so entry 0 is always wasted). Entries can be NULL when an RTE
109 * does not correspond to a base relation, such as a join RTE or an
110 * unreferenced view RTE; or if the RelOptInfo hasn't been made yet.
112 struct RelOptInfo **simple_rel_array; /* All 1-rel RelOptInfos */
113 int simple_rel_array_size; /* allocated size of array */
116 * simple_rte_array is the same length as simple_rel_array and holds
117 * pointers to the associated rangetable entries. This lets us avoid
118 * rt_fetch(), which can be a bit slow once large inheritance sets have
121 RangeTblEntry **simple_rte_array; /* rangetable as an array */
124 * join_rel_list is a list of all join-relation RelOptInfos we have
125 * considered in this planning run. For small problems we just scan the
126 * list to do lookups, but when there are many join relations we build a
127 * hash table for faster lookups. The hash table is present and valid
128 * when join_rel_hash is not NULL. Note that we still maintain the list
129 * even when using the hash table for lookups; this simplifies life for
132 List *join_rel_list; /* list of join-relation RelOptInfos */
133 struct HTAB *join_rel_hash; /* optional hashtable for join relations */
135 List *resultRelations; /* integer list of RT indexes, or NIL */
137 List *returningLists; /* list of lists of TargetEntry, or NIL */
139 List *init_plans; /* init subplans for query */
141 List *eq_classes; /* list of active EquivalenceClasses */
143 List *canon_pathkeys; /* list of "canonical" PathKeys */
145 List *left_join_clauses; /* list of RestrictInfos for
146 * mergejoinable outer join clauses
147 * w/nonnullable var on left */
149 List *right_join_clauses; /* list of RestrictInfos for
150 * mergejoinable outer join clauses
151 * w/nonnullable var on right */
153 List *full_join_clauses; /* list of RestrictInfos for
154 * mergejoinable full join clauses */
156 List *oj_info_list; /* list of OuterJoinInfos */
158 List *in_info_list; /* list of InClauseInfos */
160 List *append_rel_list; /* list of AppendRelInfos */
162 List *query_pathkeys; /* desired pathkeys for query_planner(), and
163 * actual pathkeys afterwards */
165 List *group_pathkeys; /* groupClause pathkeys, if any */
166 List *sort_pathkeys; /* sortClause pathkeys, if any */
168 List *initial_rels; /* RelOptInfos we are now trying to join */
170 MemoryContext planner_cxt; /* context holding PlannerInfo */
172 double total_table_pages; /* # of pages in all tables of query */
174 double tuple_fraction; /* tuple_fraction passed to query_planner */
176 bool hasJoinRTEs; /* true if any RTEs are RTE_JOIN kind */
177 bool hasOuterJoins; /* true if any RTEs are outer joins */
178 bool hasHavingQual; /* true if havingQual was non-null */
179 bool hasPseudoConstantQuals; /* true if any RestrictInfo has
180 * pseudoconstant = true */
185 * In places where it's known that simple_rte_array[] must have been prepared
186 * already, we just index into it to fetch RTEs. In code that might be
187 * executed before or after entering query_planner(), use this macro.
189 #define planner_rt_fetch(rti, root) \
190 ((root)->simple_rte_array ? (root)->simple_rte_array[rti] : \
191 rt_fetch(rti, (root)->parse->rtable))
196 * Per-relation information for planning/optimization
198 * For planning purposes, a "base rel" is either a plain relation (a table)
199 * or the output of a sub-SELECT or function that appears in the range table.
200 * In either case it is uniquely identified by an RT index. A "joinrel"
201 * is the joining of two or more base rels. A joinrel is identified by
202 * the set of RT indexes for its component baserels. We create RelOptInfo
203 * nodes for each baserel and joinrel, and store them in the PlannerInfo's
204 * simple_rel_array and join_rel_list respectively.
206 * Note that there is only one joinrel for any given set of component
207 * baserels, no matter what order we assemble them in; so an unordered
208 * set is the right datatype to identify it with.
210 * We also have "other rels", which are like base rels in that they refer to
211 * single RT indexes; but they are not part of the join tree, and are given
212 * a different RelOptKind to identify them.
214 * Currently the only kind of otherrels are those made for member relations
215 * of an "append relation", that is an inheritance set or UNION ALL subquery.
216 * An append relation has a parent RTE that is a base rel, which represents
217 * the entire append relation. The member RTEs are otherrels. The parent
218 * is present in the query join tree but the members are not. The member
219 * RTEs and otherrels are used to plan the scans of the individual tables or
220 * subqueries of the append set; then the parent baserel is given an Append
221 * plan comprising the best plans for the individual member rels. (See
222 * comments for AppendRelInfo for more information.)
224 * At one time we also made otherrels to represent join RTEs, for use in
225 * handling join alias Vars. Currently this is not needed because all join
226 * alias Vars are expanded to non-aliased form during preprocess_expression.
228 * Parts of this data structure are specific to various scan and join
229 * mechanisms. It didn't seem worth creating new node types for them.
231 * relids - Set of base-relation identifiers; it is a base relation
232 * if there is just one, a join relation if more than one
233 * rows - estimated number of tuples in the relation after restriction
234 * clauses have been applied (ie, output rows of a plan for it)
235 * width - avg. number of bytes per tuple in the relation after the
236 * appropriate projections have been done (ie, output width)
237 * reltargetlist - List of Var nodes for the attributes we need to
238 * output from this relation (in no particular order)
239 * NOTE: in a child relation, may contain RowExprs
240 * pathlist - List of Path nodes, one for each potentially useful
241 * method of generating the relation
242 * cheapest_startup_path - the pathlist member with lowest startup cost
243 * (regardless of its ordering)
244 * cheapest_total_path - the pathlist member with lowest total cost
245 * (regardless of its ordering)
246 * cheapest_unique_path - for caching cheapest path to produce unique
247 * (no duplicates) output from relation
249 * If the relation is a base relation it will have these fields set:
251 * relid - RTE index (this is redundant with the relids field, but
252 * is provided for convenience of access)
253 * rtekind - distinguishes plain relation, subquery, or function RTE
254 * min_attr, max_attr - range of valid AttrNumbers for rel
255 * attr_needed - array of bitmapsets indicating the highest joinrel
256 * in which each attribute is needed; if bit 0 is set then
257 * the attribute is needed as part of final targetlist
258 * attr_widths - cache space for per-attribute width estimates;
259 * zero means not computed yet
260 * indexlist - list of IndexOptInfo nodes for relation's indexes
261 * (always NIL if it's not a table)
262 * pages - number of disk pages in relation (zero if not a table)
263 * tuples - number of tuples in relation (not considering restrictions)
264 * subplan - plan for subquery (NULL if it's not a subquery)
265 * subrtable - rangetable for subquery (NIL if it's not a subquery)
267 * Note: for a subquery, tuples and subplan are not set immediately
268 * upon creation of the RelOptInfo object; they are filled in when
269 * set_base_rel_pathlist processes the object.
271 * For otherrels that are appendrel members, these fields are filled
272 * in just as for a baserel.
274 * The presence of the remaining fields depends on the restrictions
275 * and joins that the relation participates in:
277 * baserestrictinfo - List of RestrictInfo nodes, containing info about
278 * each non-join qualification clause in which this relation
279 * participates (only used for base rels)
280 * baserestrictcost - Estimated cost of evaluating the baserestrictinfo
281 * clauses at a single tuple (only used for base rels)
282 * joininfo - List of RestrictInfo nodes, containing info about each
283 * join clause in which this relation participates (but
284 * note this excludes clauses that might be derivable from
285 * EquivalenceClasses)
286 * has_eclass_joins - flag that EquivalenceClass joins are possible
287 * index_outer_relids - only used for base rels; set of outer relids
288 * that participate in indexable joinclauses for this rel
289 * index_inner_paths - only used for base rels; list of InnerIndexscanInfo
290 * nodes showing best indexpaths for various subsets of
291 * index_outer_relids.
293 * Note: Keeping a restrictinfo list in the RelOptInfo is useful only for
294 * base rels, because for a join rel the set of clauses that are treated as
295 * restrict clauses varies depending on which sub-relations we choose to join.
296 * (For example, in a 3-base-rel join, a clause relating rels 1 and 2 must be
297 * treated as a restrictclause if we join {1} and {2 3} to make {1 2 3}; but
298 * if we join {1 2} and {3} then that clause will be a restrictclause in {1 2}
299 * and should not be processed again at the level of {1 2 3}.) Therefore,
300 * the restrictinfo list in the join case appears in individual JoinPaths
301 * (field joinrestrictinfo), not in the parent relation. But it's OK for
302 * the RelOptInfo to store the joininfo list, because that is the same
303 * for a given rel no matter how we form it.
305 * We store baserestrictcost in the RelOptInfo (for base relations) because
306 * we know we will need it at least once (to price the sequential scan)
307 * and may need it multiple times to price index scans.
310 typedef enum RelOptKind
314 RELOPT_OTHER_MEMBER_REL
317 typedef struct RelOptInfo
321 RelOptKind reloptkind;
323 /* all relations included in this RelOptInfo */
324 Relids relids; /* set of base relids (rangetable indexes) */
326 /* size estimates generated by planner */
327 double rows; /* estimated number of result tuples */
328 int width; /* estimated avg width of result tuples */
330 /* materialization information */
331 List *reltargetlist; /* needed Vars */
332 List *pathlist; /* Path structures */
333 struct Path *cheapest_startup_path;
334 struct Path *cheapest_total_path;
335 struct Path *cheapest_unique_path;
337 /* information about a base rel (not set for join rels!) */
339 RTEKind rtekind; /* RELATION, SUBQUERY, or FUNCTION */
340 AttrNumber min_attr; /* smallest attrno of rel (often <0) */
341 AttrNumber max_attr; /* largest attrno of rel */
342 Relids *attr_needed; /* array indexed [min_attr .. max_attr] */
343 int32 *attr_widths; /* array indexed [min_attr .. max_attr] */
347 struct Plan *subplan; /* if subquery */
348 List *subrtable; /* if subquery */
350 /* used by various scans and joins: */
351 List *baserestrictinfo; /* RestrictInfo structures (if base
353 QualCost baserestrictcost; /* cost of evaluating the above */
354 List *joininfo; /* RestrictInfo structures for join clauses
355 * involving this rel */
356 bool has_eclass_joins; /* T means joininfo is incomplete */
358 /* cached info about inner indexscan paths for relation: */
359 Relids index_outer_relids; /* other relids in indexable join
361 List *index_inner_paths; /* InnerIndexscanInfo nodes */
364 * Inner indexscans are not in the main pathlist because they are not
365 * usable except in specific join contexts. We use the index_inner_paths
366 * list just to avoid recomputing the best inner indexscan repeatedly for
367 * similar outer relations. See comments for InnerIndexscanInfo.
373 * Per-index information for planning/optimization
375 * Prior to Postgres 7.0, RelOptInfo was used to describe both relations
376 * and indexes, but that created confusion without actually doing anything
377 * useful. So now we have a separate IndexOptInfo struct for indexes.
379 * opfamily[], indexkeys[], opcintype[], fwdsortop[], revsortop[],
380 * and nulls_first[] each have ncolumns entries.
381 * Note: for historical reasons, the opfamily array has an extra entry
382 * that is always zero. Some code scans until it sees a zero entry,
383 * rather than looking at ncolumns.
385 * Zeroes in the indexkeys[] array indicate index columns that are
386 * expressions; there is one element in indexprs for each such column.
388 * For an unordered index, the sortop arrays contains zeroes. Note that
389 * fwdsortop[] and nulls_first[] describe the sort ordering of a forward
390 * indexscan; we can also consider a backward indexscan, which will
391 * generate sort order described by revsortop/!nulls_first.
393 * The indexprs and indpred expressions have been run through
394 * prepqual.c and eval_const_expressions() for ease of matching to
395 * WHERE clauses. indpred is in implicit-AND form.
397 typedef struct IndexOptInfo
401 Oid indexoid; /* OID of the index relation */
402 RelOptInfo *rel; /* back-link to index's table */
404 /* statistics from pg_class */
405 BlockNumber pages; /* number of disk pages in index */
406 double tuples; /* number of index tuples in index */
408 /* index descriptor information */
409 int ncolumns; /* number of columns in index */
410 Oid *opfamily; /* OIDs of operator families for columns */
411 int *indexkeys; /* column numbers of index's keys, or 0 */
412 Oid *opcintype; /* OIDs of opclass declared input data types */
413 Oid *fwdsortop; /* OIDs of sort operators for each column */
414 Oid *revsortop; /* OIDs of sort operators for backward scan */
415 bool *nulls_first; /* do NULLs come first in the sort order? */
416 Oid relam; /* OID of the access method (in pg_am) */
418 RegProcedure amcostestimate; /* OID of the access method's cost fcn */
420 List *indexprs; /* expressions for non-simple index columns */
421 List *indpred; /* predicate if a partial index, else NIL */
423 bool predOK; /* true if predicate matches query */
424 bool unique; /* true if a unique index */
425 bool amoptionalkey; /* can query omit key for the first column? */
426 bool amsearchnulls; /* can AM search for NULL index entries? */
433 * Whenever we can determine that a mergejoinable equality clause A = B is
434 * not delayed by any outer join, we create an EquivalenceClass containing
435 * the expressions A and B to record this knowledge. If we later find another
436 * equivalence B = C, we add C to the existing EquivalenceClass; this may
437 * require merging two existing EquivalenceClasses. At the end of the qual
438 * distribution process, we have sets of values that are known all transitively
439 * equal to each other, where "equal" is according to the rules of the btree
440 * operator family(s) shown in ec_opfamilies. (We restrict an EC to contain
441 * only equalities whose operators belong to the same set of opfamilies. This
442 * could probably be relaxed, but for now it's not worth the trouble, since
443 * nearly all equality operators belong to only one btree opclass anyway.)
445 * We also use EquivalenceClasses as the base structure for PathKeys, letting
446 * us represent knowledge about different sort orderings being equivalent.
447 * Since every PathKey must reference an EquivalenceClass, we will end up
448 * with single-member EquivalenceClasses whenever a sort key expression has
449 * not been equivalenced to anything else. It is also possible that such an
450 * EquivalenceClass will contain a volatile expression ("ORDER BY random()"),
451 * which is a case that can't arise otherwise since clauses containing
452 * volatile functions are never considered mergejoinable. We mark such
453 * EquivalenceClasses specially to prevent them from being merged with
454 * ordinary EquivalenceClasses. Also, for volatile expressions we have
455 * to be careful to match the EquivalenceClass to the correct targetlist
456 * entry: consider SELECT random() AS a, random() AS b ... ORDER BY b,a.
457 * So we record the SortGroupRef of the originating sort clause.
459 * We allow equality clauses appearing below the nullable side of an outer join
460 * to form EquivalenceClasses, but these have a slightly different meaning:
461 * the included values might be all NULL rather than all the same non-null
462 * values. See src/backend/optimizer/README for more on that point.
464 * NB: if ec_merged isn't NULL, this class has been merged into another, and
465 * should be ignored in favor of using the pointed-to class.
467 typedef struct EquivalenceClass
471 List *ec_opfamilies; /* btree operator family OIDs */
472 List *ec_members; /* list of EquivalenceMembers */
473 List *ec_sources; /* list of generating RestrictInfos */
474 List *ec_derives; /* list of derived RestrictInfos */
475 Relids ec_relids; /* all relids appearing in ec_members */
476 bool ec_has_const; /* any pseudoconstants in ec_members? */
477 bool ec_has_volatile; /* the (sole) member is a volatile expr */
478 bool ec_below_outer_join; /* equivalence applies below an OJ */
479 bool ec_broken; /* failed to generate needed clauses? */
480 Index ec_sortref; /* originating sortclause label, or 0 */
481 struct EquivalenceClass *ec_merged; /* set if merged into another EC */
485 * If an EC contains a const and isn't below-outer-join, any PathKey depending
486 * on it must be redundant, since there's only one possible value of the key.
488 #define EC_MUST_BE_REDUNDANT(eclass) \
489 ((eclass)->ec_has_const && !(eclass)->ec_below_outer_join)
492 * EquivalenceMember - one member expression of an EquivalenceClass
494 * em_is_child signifies that this element was built by transposing a member
495 * for an inheritance parent relation to represent the corresponding expression
496 * on an inheritance child. The element should be ignored for all purposes
497 * except constructing inner-indexscan paths for the child relation. (Other
498 * types of join are driven from transposed joininfo-list entries.) Note
499 * that the EC's ec_relids field does NOT include the child relation.
501 * em_datatype is usually the same as exprType(em_expr), but can be
502 * different when dealing with a binary-compatible opfamily; in particular
503 * anyarray_ops would never work without this. Use em_datatype when
504 * looking up a specific btree operator to work with this expression.
506 typedef struct EquivalenceMember
510 Expr *em_expr; /* the expression represented */
511 Relids em_relids; /* all relids appearing in em_expr */
512 bool em_is_const; /* expression is pseudoconstant? */
513 bool em_is_child; /* derived version for a child relation? */
514 Oid em_datatype; /* the "nominal type" used by the opfamily */
520 * The sort ordering of a path is represented by a list of PathKey nodes.
521 * An empty list implies no known ordering. Otherwise the first item
522 * represents the primary sort key, the second the first secondary sort key,
523 * etc. The value being sorted is represented by linking to an
524 * EquivalenceClass containing that value and including pk_opfamily among its
525 * ec_opfamilies. This is a convenient method because it makes it trivial
526 * to detect equivalent and closely-related orderings. (See optimizer/README
527 * for more information.)
529 * Note: pk_strategy is either BTLessStrategyNumber (for ASC) or
530 * BTGreaterStrategyNumber (for DESC). We assume that all ordering-capable
531 * index types will use btree-compatible strategy numbers.
534 typedef struct PathKey
538 EquivalenceClass *pk_eclass; /* the value that is ordered */
539 Oid pk_opfamily; /* btree opfamily defining the ordering */
540 int pk_strategy; /* sort direction (ASC or DESC) */
541 bool pk_nulls_first; /* do NULLs come before normal values? */
545 * Type "Path" is used as-is for sequential-scan paths. For other
546 * path types it is the first component of a larger struct.
548 * Note: "pathtype" is the NodeTag of the Plan node we could build from this
549 * Path. It is partially redundant with the Path's NodeTag, but allows us
550 * to use the same Path type for multiple Plan types where there is no need
551 * to distinguish the Plan type during path processing.
558 NodeTag pathtype; /* tag identifying scan/join method */
560 RelOptInfo *parent; /* the relation this path can build */
562 /* estimated execution costs for path (see costsize.c for more info) */
563 Cost startup_cost; /* cost expended before fetching any tuples */
564 Cost total_cost; /* total cost (assuming all tuples fetched) */
566 List *pathkeys; /* sort ordering of path's output */
567 /* pathkeys is a List of PathKey nodes; see above */
571 * IndexPath represents an index scan over a single index.
573 * 'indexinfo' is the index to be scanned.
575 * 'indexclauses' is a list of index qualification clauses, with implicit
576 * AND semantics across the list. Each clause is a RestrictInfo node from
577 * the query's WHERE or JOIN conditions.
579 * 'indexquals' has the same structure as 'indexclauses', but it contains
580 * the actual indexqual conditions that can be used with the index.
581 * In simple cases this is identical to 'indexclauses', but when special
582 * indexable operators appear in 'indexclauses', they are replaced by the
583 * derived indexscannable conditions in 'indexquals'.
585 * 'isjoininner' is TRUE if the path is a nestloop inner scan (that is,
586 * some of the index conditions are join rather than restriction clauses).
587 * Note that the path costs will be calculated differently from a plain
588 * indexscan in this case, and in addition there's a special 'rows' value
589 * different from the parent RelOptInfo's (see below).
591 * 'indexscandir' is one of:
592 * ForwardScanDirection: forward scan of an ordered index
593 * BackwardScanDirection: backward scan of an ordered index
594 * NoMovementScanDirection: scan of an unordered index, or don't care
595 * (The executor doesn't care whether it gets ForwardScanDirection or
596 * NoMovementScanDirection for an indexscan, but the planner wants to
597 * distinguish ordered from unordered indexes for building pathkeys.)
599 * 'indextotalcost' and 'indexselectivity' are saved in the IndexPath so that
600 * we need not recompute them when considering using the same index in a
601 * bitmap index/heap scan (see BitmapHeapPath). The costs of the IndexPath
602 * itself represent the costs of an IndexScan plan type.
604 * 'rows' is the estimated result tuple count for the indexscan. This
605 * is the same as path.parent->rows for a simple indexscan, but it is
606 * different for a nestloop inner scan, because the additional indexquals
607 * coming from join clauses make the scan more selective than the parent
608 * rel's restrict clauses alone would do.
611 typedef struct IndexPath
614 IndexOptInfo *indexinfo;
618 ScanDirection indexscandir;
620 Selectivity indexselectivity;
621 double rows; /* estimated number of result tuples */
625 * BitmapHeapPath represents one or more indexscans that generate TID bitmaps
626 * instead of directly accessing the heap, followed by AND/OR combinations
627 * to produce a single bitmap, followed by a heap scan that uses the bitmap.
628 * Note that the output is always considered unordered, since it will come
629 * out in physical heap order no matter what the underlying indexes did.
631 * The individual indexscans are represented by IndexPath nodes, and any
632 * logic on top of them is represented by a tree of BitmapAndPath and
633 * BitmapOrPath nodes. Notice that we can use the same IndexPath node both
634 * to represent a regular IndexScan plan, and as the child of a BitmapHeapPath
635 * that represents scanning the same index using a BitmapIndexScan. The
636 * startup_cost and total_cost figures of an IndexPath always represent the
637 * costs to use it as a regular IndexScan. The costs of a BitmapIndexScan
638 * can be computed using the IndexPath's indextotalcost and indexselectivity.
640 * BitmapHeapPaths can be nestloop inner indexscans. The isjoininner and
641 * rows fields serve the same purpose as for plain IndexPaths.
643 typedef struct BitmapHeapPath
646 Path *bitmapqual; /* IndexPath, BitmapAndPath, BitmapOrPath */
647 bool isjoininner; /* T if it's a nestloop inner scan */
648 double rows; /* estimated number of result tuples */
652 * BitmapAndPath represents a BitmapAnd plan node; it can only appear as
653 * part of the substructure of a BitmapHeapPath. The Path structure is
654 * a bit more heavyweight than we really need for this, but for simplicity
655 * we make it a derivative of Path anyway.
657 typedef struct BitmapAndPath
660 List *bitmapquals; /* IndexPaths and BitmapOrPaths */
661 Selectivity bitmapselectivity;
665 * BitmapOrPath represents a BitmapOr plan node; it can only appear as
666 * part of the substructure of a BitmapHeapPath. The Path structure is
667 * a bit more heavyweight than we really need for this, but for simplicity
668 * we make it a derivative of Path anyway.
670 typedef struct BitmapOrPath
673 List *bitmapquals; /* IndexPaths and BitmapAndPaths */
674 Selectivity bitmapselectivity;
678 * TidPath represents a scan by TID
680 * tidquals is an implicitly OR'ed list of qual expressions of the form
681 * "CTID = pseudoconstant" or "CTID = ANY(pseudoconstant_array)".
682 * Note they are bare expressions, not RestrictInfos.
684 typedef struct TidPath
687 List *tidquals; /* qual(s) involving CTID = something */
691 * AppendPath represents an Append plan, ie, successive execution of
692 * several member plans.
694 * Note: it is possible for "subpaths" to contain only one, or even no,
695 * elements. These cases are optimized during create_append_plan.
697 typedef struct AppendPath
700 List *subpaths; /* list of component Paths */
704 * ResultPath represents use of a Result plan node to compute a variable-free
705 * targetlist with no underlying tables (a "SELECT expressions" query).
706 * The query could have a WHERE clause, too, represented by "quals".
708 * Note that quals is a list of bare clauses, not RestrictInfos.
710 typedef struct ResultPath
717 * MaterialPath represents use of a Material plan node, i.e., caching of
718 * the output of its subpath. This is used when the subpath is expensive
719 * and needs to be scanned repeatedly, or when we need mark/restore ability
720 * and the subpath doesn't have it.
722 typedef struct MaterialPath
729 * UniquePath represents elimination of distinct rows from the output of
732 * This is unlike the other Path nodes in that it can actually generate
733 * different plans: either hash-based or sort-based implementation, or a
734 * no-op if the input path can be proven distinct already. The decision
735 * is sufficiently localized that it's not worth having separate Path node
736 * types. (Note: in the no-op case, we could eliminate the UniquePath node
737 * entirely and just return the subpath; but it's convenient to have a
738 * UniquePath in the path tree to signal upper-level routines that the input
739 * is known distinct.)
743 UNIQUE_PATH_NOOP, /* input is known unique already */
744 UNIQUE_PATH_HASH, /* use hashing */
745 UNIQUE_PATH_SORT /* use sorting */
748 typedef struct UniquePath
752 UniquePathMethod umethod;
753 double rows; /* estimated number of result tuples */
757 * All join-type paths share these fields.
760 typedef struct JoinPath
766 Path *outerjoinpath; /* path for the outer side of the join */
767 Path *innerjoinpath; /* path for the inner side of the join */
769 List *joinrestrictinfo; /* RestrictInfos to apply to join */
772 * See the notes for RelOptInfo to understand why joinrestrictinfo is
773 * needed in JoinPath, and can't be merged into the parent RelOptInfo.
778 * A nested-loop path needs no special fields.
781 typedef JoinPath NestPath;
784 * A mergejoin path has these fields.
786 * path_mergeclauses lists the clauses (in the form of RestrictInfos)
787 * that will be used in the merge.
789 * Note that the mergeclauses are a subset of the parent relation's
790 * restriction-clause list. Any join clauses that are not mergejoinable
791 * appear only in the parent's restrict list, and must be checked by a
792 * qpqual at execution time.
794 * outersortkeys (resp. innersortkeys) is NIL if the outer path
795 * (resp. inner path) is already ordered appropriately for the
796 * mergejoin. If it is not NIL then it is a PathKeys list describing
797 * the ordering that must be created by an explicit sort step.
800 typedef struct MergePath
803 List *path_mergeclauses; /* join clauses to be used for merge */
804 List *outersortkeys; /* keys for explicit sort, if any */
805 List *innersortkeys; /* keys for explicit sort, if any */
809 * A hashjoin path has these fields.
811 * The remarks above for mergeclauses apply for hashclauses as well.
813 * Hashjoin does not care what order its inputs appear in, so we have
814 * no need for sortkeys.
817 typedef struct HashPath
820 List *path_hashclauses; /* join clauses used for hashing */
824 * Restriction clause info.
826 * We create one of these for each AND sub-clause of a restriction condition
827 * (WHERE or JOIN/ON clause). Since the restriction clauses are logically
828 * ANDed, we can use any one of them or any subset of them to filter out
829 * tuples, without having to evaluate the rest. The RestrictInfo node itself
830 * stores data used by the optimizer while choosing the best query plan.
832 * If a restriction clause references a single base relation, it will appear
833 * in the baserestrictinfo list of the RelOptInfo for that base rel.
835 * If a restriction clause references more than one base rel, it will
836 * appear in the joininfo list of every RelOptInfo that describes a strict
837 * subset of the base rels mentioned in the clause. The joininfo lists are
838 * used to drive join tree building by selecting plausible join candidates.
839 * The clause cannot actually be applied until we have built a join rel
840 * containing all the base rels it references, however.
842 * When we construct a join rel that includes all the base rels referenced
843 * in a multi-relation restriction clause, we place that clause into the
844 * joinrestrictinfo lists of paths for the join rel, if neither left nor
845 * right sub-path includes all base rels referenced in the clause. The clause
846 * will be applied at that join level, and will not propagate any further up
847 * the join tree. (Note: the "predicate migration" code was once intended to
848 * push restriction clauses up and down the plan tree based on evaluation
849 * costs, but it's dead code and is unlikely to be resurrected in the
850 * foreseeable future.)
852 * Note that in the presence of more than two rels, a multi-rel restriction
853 * might reach different heights in the join tree depending on the join
854 * sequence we use. So, these clauses cannot be associated directly with
855 * the join RelOptInfo, but must be kept track of on a per-join-path basis.
857 * RestrictInfos that represent equivalence conditions (i.e., mergejoinable
858 * equalities that are not outerjoin-delayed) are handled a bit differently.
859 * Initially we attach them to the EquivalenceClasses that are derived from
860 * them. When we construct a scan or join path, we look through all the
861 * EquivalenceClasses and generate derived RestrictInfos representing the
862 * minimal set of conditions that need to be checked for this particular scan
863 * or join to enforce that all members of each EquivalenceClass are in fact
864 * equal in all rows emitted by the scan or join.
866 * When dealing with outer joins we have to be very careful about pushing qual
867 * clauses up and down the tree. An outer join's own JOIN/ON conditions must
868 * be evaluated exactly at that join node, unless they are "degenerate"
869 * conditions that reference only Vars from the nullable side of the join.
870 * Quals appearing in WHERE or in a JOIN above the outer join cannot be pushed
871 * down below the outer join, if they reference any nullable Vars.
872 * RestrictInfo nodes contain a flag to indicate whether a qual has been
873 * pushed down to a lower level than its original syntactic placement in the
874 * join tree would suggest. If an outer join prevents us from pushing a qual
875 * down to its "natural" semantic level (the level associated with just the
876 * base rels used in the qual) then we mark the qual with a "required_relids"
877 * value including more than just the base rels it actually uses. By
878 * pretending that the qual references all the rels required to form the outer
879 * join, we prevent it from being evaluated below the outer join's joinrel.
880 * When we do form the outer join's joinrel, we still need to distinguish
881 * those quals that are actually in that join's JOIN/ON condition from those
882 * that appeared elsewhere in the tree and were pushed down to the join rel
883 * because they used no other rels. That's what the is_pushed_down flag is
884 * for; it tells us that a qual is not an OUTER JOIN qual for the set of base
885 * rels listed in required_relids. A clause that originally came from WHERE
886 * or an INNER JOIN condition will *always* have its is_pushed_down flag set.
887 * It's possible for an OUTER JOIN clause to be marked is_pushed_down too,
888 * if we decide that it can be pushed down into the nullable side of the join.
889 * In that case it acts as a plain filter qual for wherever it gets evaluated.
890 * (In short, is_pushed_down is only false for non-degenerate outer join
891 * conditions. Possibly we should rename it to reflect that meaning?)
893 * RestrictInfo nodes also contain an outerjoin_delayed flag, which is true
894 * if the clause's applicability must be delayed due to any outer joins
895 * appearing below its own syntactic level (ie, it references any Vars from
896 * the nullable side of any lower outer join).
898 * In general, the referenced clause might be arbitrarily complex. The
899 * kinds of clauses we can handle as indexscan quals, mergejoin clauses,
900 * or hashjoin clauses are limited (e.g., no volatile functions). The code
901 * for each kind of path is responsible for identifying the restrict clauses
902 * it can use and ignoring the rest. Clauses not implemented by an indexscan,
903 * mergejoin, or hashjoin will be placed in the plan qual or joinqual field
904 * of the finished Plan node, where they will be enforced by general-purpose
905 * qual-expression-evaluation code. (But we are still entitled to count
906 * their selectivity when estimating the result tuple count, if we
907 * can guess what it is...)
909 * When the referenced clause is an OR clause, we generate a modified copy
910 * in which additional RestrictInfo nodes are inserted below the top-level
911 * OR/AND structure. This is a convenience for OR indexscan processing:
912 * indexquals taken from either the top level or an OR subclause will have
913 * associated RestrictInfo nodes.
915 * The can_join flag is set true if the clause looks potentially useful as
916 * a merge or hash join clause, that is if it is a binary opclause with
917 * nonoverlapping sets of relids referenced in the left and right sides.
918 * (Whether the operator is actually merge or hash joinable isn't checked,
921 * The pseudoconstant flag is set true if the clause contains no Vars of
922 * the current query level and no volatile functions. Such a clause can be
923 * pulled out and used as a one-time qual in a gating Result node. We keep
924 * pseudoconstant clauses in the same lists as other RestrictInfos so that
925 * the regular clause-pushing machinery can assign them to the correct join
926 * level, but they need to be treated specially for cost and selectivity
927 * estimates. Note that a pseudoconstant clause can never be an indexqual
928 * or merge or hash join clause, so it's of no interest to large parts of
931 * When join clauses are generated from EquivalenceClasses, there may be
932 * several equally valid ways to enforce join equivalence, of which we need
933 * apply only one. We mark clauses of this kind by setting parent_ec to
934 * point to the generating EquivalenceClass. Multiple clauses with the same
935 * parent_ec in the same join are redundant.
938 typedef struct RestrictInfo
942 Expr *clause; /* the represented clause of WHERE or JOIN */
944 bool is_pushed_down; /* TRUE if clause was pushed down in level */
946 bool outerjoin_delayed; /* TRUE if delayed by lower outer join */
948 bool can_join; /* see comment above */
950 bool pseudoconstant; /* see comment above */
952 /* The set of relids (varnos) actually referenced in the clause: */
953 Relids clause_relids;
955 /* The set of relids required to evaluate the clause: */
956 Relids required_relids;
958 /* These fields are set for any binary opclause: */
959 Relids left_relids; /* relids in left side of clause */
960 Relids right_relids; /* relids in right side of clause */
962 /* This field is NULL unless clause is an OR clause: */
963 Expr *orclause; /* modified clause with RestrictInfos */
965 /* This field is NULL unless clause is potentially redundant: */
966 EquivalenceClass *parent_ec; /* generating EquivalenceClass */
968 /* cache space for cost and selectivity */
969 QualCost eval_cost; /* eval cost of clause; -1 if not yet set */
970 Selectivity this_selec; /* selectivity; -1 if not yet set */
972 /* valid if clause is mergejoinable, else NIL */
973 List *mergeopfamilies; /* opfamilies containing clause operator */
975 /* cache space for mergeclause processing; NULL if not yet set */
976 EquivalenceClass *left_ec; /* EquivalenceClass containing lefthand */
977 EquivalenceClass *right_ec; /* EquivalenceClass containing righthand */
978 EquivalenceMember *left_em; /* EquivalenceMember for lefthand */
979 EquivalenceMember *right_em; /* EquivalenceMember for righthand */
980 List *scansel_cache; /* list of MergeScanSelCache structs */
982 /* transient workspace for use while considering a specific join path */
983 bool outer_is_left; /* T = outer var on left, F = on right */
985 /* valid if clause is hashjoinable, else InvalidOid: */
986 Oid hashjoinoperator; /* copy of clause operator */
988 /* cache space for hashclause processing; -1 if not yet set */
989 Selectivity left_bucketsize; /* avg bucketsize of left side */
990 Selectivity right_bucketsize; /* avg bucketsize of right side */
994 * Since mergejoinscansel() is a relatively expensive function, and would
995 * otherwise be invoked many times while planning a large join tree,
996 * we go out of our way to cache its results. Each mergejoinable
997 * RestrictInfo carries a list of the specific sort orderings that have
998 * been considered for use with it, and the resulting selectivities.
1000 typedef struct MergeScanSelCache
1002 /* Ordering details (cache lookup key) */
1003 Oid opfamily; /* btree opfamily defining the ordering */
1004 int strategy; /* sort direction (ASC or DESC) */
1005 bool nulls_first; /* do NULLs come before normal values? */
1007 Selectivity leftstartsel; /* first-join fraction for clause left side */
1008 Selectivity leftendsel; /* last-join fraction for clause left side */
1009 Selectivity rightstartsel; /* first-join fraction for clause right side */
1010 Selectivity rightendsel; /* last-join fraction for clause right side */
1011 } MergeScanSelCache;
1014 * Inner indexscan info.
1016 * An inner indexscan is one that uses one or more joinclauses as index
1017 * conditions (perhaps in addition to plain restriction clauses). So it
1018 * can only be used as the inner path of a nestloop join where the outer
1019 * relation includes all other relids appearing in those joinclauses.
1020 * The set of usable joinclauses, and thus the best inner indexscan,
1021 * thus varies depending on which outer relation we consider; so we have
1022 * to recompute the best such paths for every join. To avoid lots of
1023 * redundant computation, we cache the results of such searches. For
1024 * each relation we compute the set of possible otherrelids (all relids
1025 * appearing in joinquals that could become indexquals for this table).
1026 * Two outer relations whose relids have the same intersection with this
1027 * set will have the same set of available joinclauses and thus the same
1028 * best inner indexscans for the inner relation. By taking the intersection
1029 * before scanning the cache, we avoid recomputing when considering
1030 * join rels that differ only by the inclusion of irrelevant other rels.
1032 * The search key also includes a bool showing whether the join being
1033 * considered is an outer join. Since we constrain the join order for
1034 * outer joins, I believe that this bool can only have one possible value
1035 * for any particular lookup key; but store it anyway to avoid confusion.
1038 typedef struct InnerIndexscanInfo
1041 /* The lookup key: */
1042 Relids other_relids; /* a set of relevant other relids */
1043 bool isouterjoin; /* true if join is outer */
1044 /* Best paths for this lookup key (NULL if no available indexscans): */
1045 Path *cheapest_startup_innerpath; /* cheapest startup cost */
1046 Path *cheapest_total_innerpath; /* cheapest total cost */
1047 } InnerIndexscanInfo;
1052 * One-sided outer joins constrain the order of joining partially but not
1053 * completely. We flatten such joins into the planner's top-level list of
1054 * relations to join, but record information about each outer join in an
1055 * OuterJoinInfo struct. These structs are kept in the PlannerInfo node's
1058 * min_lefthand and min_righthand are the sets of base relids that must be
1059 * available on each side when performing the outer join. lhs_strict is
1060 * true if the outer join's condition cannot succeed when the LHS variables
1061 * are all NULL (this means that the outer join can commute with upper-level
1062 * outer joins even if it appears in their RHS). We don't bother to set
1063 * lhs_strict for FULL JOINs, however.
1065 * It is not valid for either min_lefthand or min_righthand to be empty sets;
1066 * if they were, this would break the logic that enforces join order.
1068 * syn_lefthand and syn_righthand are the sets of base relids that are
1069 * syntactically below this outer join. (These are needed to help compute
1070 * min_lefthand and min_righthand for higher joins, but are not used
1073 * delay_upper_joins is set TRUE if we detect a pushed-down clause that has
1074 * to be evaluated after this join is formed (because it references the RHS).
1075 * Any outer joins that have such a clause and this join in their RHS cannot
1076 * commute with this join, because that would leave noplace to check the
1077 * pushed-down clause. (We don't track this for FULL JOINs, either.)
1079 * Note: OuterJoinInfo directly represents only LEFT JOIN and FULL JOIN;
1080 * RIGHT JOIN is handled by switching the inputs to make it a LEFT JOIN.
1081 * We make an OuterJoinInfo for FULL JOINs even though there is no flexibility
1082 * of planning for them, because this simplifies make_join_rel()'s API.
1085 typedef struct OuterJoinInfo
1088 Relids min_lefthand; /* base relids in minimum LHS for join */
1089 Relids min_righthand; /* base relids in minimum RHS for join */
1090 Relids syn_lefthand; /* base relids syntactically within LHS */
1091 Relids syn_righthand; /* base relids syntactically within RHS */
1092 bool is_full_join; /* it's a FULL OUTER JOIN */
1093 bool lhs_strict; /* joinclause is strict for some LHS rel */
1094 bool delay_upper_joins; /* can't commute with upper RHS */
1100 * When we convert top-level IN quals into join operations, we must restrict
1101 * the order of joining and use special join methods at some join points.
1102 * We record information about each such IN clause in an InClauseInfo struct.
1103 * These structs are kept in the PlannerInfo node's in_info_list.
1105 * Note: sub_targetlist is just a list of Vars or expressions; it does not
1106 * contain TargetEntry nodes.
1109 typedef struct InClauseInfo
1112 Relids lefthand; /* base relids in lefthand expressions */
1113 Relids righthand; /* base relids coming from the subselect */
1114 List *sub_targetlist; /* targetlist of original RHS subquery */
1115 List *in_operators; /* OIDs of the IN's equality operator(s) */
1119 * Append-relation info.
1121 * When we expand an inheritable table or a UNION-ALL subselect into an
1122 * "append relation" (essentially, a list of child RTEs), we build an
1123 * AppendRelInfo for each child RTE. The list of AppendRelInfos indicates
1124 * which child RTEs must be included when expanding the parent, and each
1125 * node carries information needed to translate Vars referencing the parent
1126 * into Vars referencing that child.
1128 * These structs are kept in the PlannerInfo node's append_rel_list.
1129 * Note that we just throw all the structs into one list, and scan the
1130 * whole list when desiring to expand any one parent. We could have used
1131 * a more complex data structure (eg, one list per parent), but this would
1132 * be harder to update during operations such as pulling up subqueries,
1133 * and not really any easier to scan. Considering that typical queries
1134 * will not have many different append parents, it doesn't seem worthwhile
1135 * to complicate things.
1137 * Note: after completion of the planner prep phase, any given RTE is an
1138 * append parent having entries in append_rel_list if and only if its
1139 * "inh" flag is set. We clear "inh" for plain tables that turn out not
1140 * to have inheritance children, and (in an abuse of the original meaning
1141 * of the flag) we set "inh" for subquery RTEs that turn out to be
1142 * flattenable UNION ALL queries. This lets us avoid useless searches
1143 * of append_rel_list.
1145 * Note: the data structure assumes that append-rel members are single
1146 * baserels. This is OK for inheritance, but it prevents us from pulling
1147 * up a UNION ALL member subquery if it contains a join. While that could
1148 * be fixed with a more complex data structure, at present there's not much
1149 * point because no improvement in the plan could result.
1152 typedef struct AppendRelInfo
1157 * These fields uniquely identify this append relationship. There can be
1158 * (in fact, always should be) multiple AppendRelInfos for the same
1159 * parent_relid, but never more than one per child_relid, since a given
1160 * RTE cannot be a child of more than one append parent.
1162 Index parent_relid; /* RT index of append parent rel */
1163 Index child_relid; /* RT index of append child rel */
1166 * For an inheritance appendrel, the parent and child are both regular
1167 * relations, and we store their rowtype OIDs here for use in translating
1168 * whole-row Vars. For a UNION-ALL appendrel, the parent and child are
1169 * both subqueries with no named rowtype, and we store InvalidOid here.
1171 Oid parent_reltype; /* OID of parent's composite type */
1172 Oid child_reltype; /* OID of child's composite type */
1175 * The N'th element of this list is the integer column number of the child
1176 * column corresponding to the N'th column of the parent. A list element
1177 * is zero if it corresponds to a dropped column of the parent (this is
1178 * only possible for inheritance cases, not UNION ALL).
1180 List *col_mappings; /* list of child attribute numbers */
1183 * The N'th element of this list is a Var or expression representing the
1184 * child column corresponding to the N'th column of the parent. This is
1185 * used to translate Vars referencing the parent rel into references to
1186 * the child. A list element is NULL if it corresponds to a dropped
1187 * column of the parent (this is only possible for inheritance cases, not
1190 * This might seem redundant with the col_mappings data, but it is handy
1191 * because flattening of sub-SELECTs that are members of a UNION ALL will
1192 * cause changes in the expressions that need to be substituted for a
1193 * parent Var. Adjusting this data structure lets us track what really
1194 * needs to be substituted.
1196 * Notice we only store entries for user columns (attno > 0). Whole-row
1197 * Vars are special-cased, and system columns (attno < 0) need no special
1198 * translation since their attnos are the same for all tables.
1200 * Caution: the Vars have varlevelsup = 0. Be careful to adjust as needed
1201 * when copying into a subquery.
1203 List *translated_vars; /* Expressions in the child's Vars */
1206 * We store the parent table's OID here for inheritance, or InvalidOid for
1207 * UNION ALL. This is only needed to help in generating error messages if
1208 * an attempt is made to reference a dropped parent column.
1210 Oid parent_reloid; /* OID of parent relation */
1214 * glob->paramlist keeps track of the PARAM_EXEC slots that we have decided
1215 * we need for the query. At runtime these slots are used to pass values
1216 * either down into subqueries (for outer references in subqueries) or up out
1217 * of subqueries (for the results of a subplan). The n'th entry in the list
1218 * (n counts from 0) corresponds to Param->paramid = n.
1220 * Each paramlist item shows the absolute query level it is associated with,
1221 * where the outermost query is level 1 and nested subqueries have higher
1222 * numbers. The item the parameter slot represents can be one of three kinds:
1224 * A Var: the slot represents a variable of that level that must be passed
1225 * down because subqueries have outer references to it. The varlevelsup
1226 * value in the Var will always be zero.
1228 * An Aggref (with an expression tree representing its argument): the slot
1229 * represents an aggregate expression that is an outer reference for some
1230 * subquery. The Aggref itself has agglevelsup = 0, and its argument tree
1231 * is adjusted to match in level.
1233 * A Param: the slot holds the result of a subplan (it is a setParam item
1234 * for that subplan). The absolute level shown for such items corresponds
1235 * to the parent query of the subplan.
1237 * Note: we detect duplicate Var parameters and coalesce them into one slot,
1238 * but we do not do this for Aggref or Param slots.
1240 typedef struct PlannerParamItem
1244 Node *item; /* the Var, Aggref, or Param */
1245 Index abslevel; /* its absolute query level */
1248 #endif /* RELATION_H */