* Definitions for planner's internal data structures.
*
*
- * Portions Copyright (c) 1996-2003, PostgreSQL Global Development Group
+ * Portions Copyright (c) 1996-2007, PostgreSQL Global Development Group
* Portions Copyright (c) 1994, Regents of the University of California
*
- * $PostgreSQL: pgsql/src/include/nodes/relation.h,v 1.88 2003/12/30 23:53:15 tgl Exp $
+ * $PostgreSQL: pgsql/src/include/nodes/relation.h,v 1.133 2007/01/20 20:45:40 tgl Exp $
*
*-------------------------------------------------------------------------
*/
#include "access/sdir.h"
#include "nodes/bitmapset.h"
#include "nodes/parsenodes.h"
+#include "storage/block.h"
/*
* Relids
* Set of relation identifiers (indexes into the rangetable).
*/
-
typedef Bitmapset *Relids;
/*
Cost per_tuple; /* per-evaluation cost */
} QualCost;
+
+/*----------
+ * PlannerInfo
+ * Per-query information for planning/optimization
+ *
+ * This struct is conventionally called "root" in all the planner routines.
+ * It holds links to all of the planner's working state, in addition to the
+ * original Query. Note that at present the planner extensively modifies
+ * the passed-in Query data structure; someday that should stop.
+ *----------
+ */
+typedef struct PlannerInfo
+{
+ NodeTag type;
+
+ Query *parse; /* the Query being planned */
+
+ /*
+ * simple_rel_array holds pointers to "base rels" and "other rels" (see
+ * comments for RelOptInfo for more info). It is indexed by rangetable
+ * index (so entry 0 is always wasted). Entries can be NULL when an RTE
+ * does not correspond to a base relation, such as a join RTE or an
+ * unreferenced view RTE; or if the RelOptInfo hasn't been made yet.
+ */
+ struct RelOptInfo **simple_rel_array; /* All 1-rel RelOptInfos */
+ int simple_rel_array_size; /* allocated size of array */
+
+ /*
+ * join_rel_list is a list of all join-relation RelOptInfos we have
+ * considered in this planning run. For small problems we just scan the
+ * list to do lookups, but when there are many join relations we build a
+ * hash table for faster lookups. The hash table is present and valid
+ * when join_rel_hash is not NULL. Note that we still maintain the list
+ * even when using the hash table for lookups; this simplifies life for
+ * GEQO.
+ */
+ List *join_rel_list; /* list of join-relation RelOptInfos */
+ struct HTAB *join_rel_hash; /* optional hashtable for join relations */
+
+ List *eq_classes; /* list of active EquivalenceClasses */
+
+ List *canon_pathkeys; /* list of "canonical" PathKeys */
+
+ List *left_join_clauses; /* list of RestrictInfos for
+ * mergejoinable outer join clauses
+ * w/nonnullable var on left */
+
+ List *right_join_clauses; /* list of RestrictInfos for
+ * mergejoinable outer join clauses
+ * w/nonnullable var on right */
+
+ List *full_join_clauses; /* list of RestrictInfos for
+ * mergejoinable full join clauses */
+
+ List *oj_info_list; /* list of OuterJoinInfos */
+
+ List *in_info_list; /* list of InClauseInfos */
+
+ List *append_rel_list; /* list of AppendRelInfos */
+
+ List *query_pathkeys; /* desired pathkeys for query_planner(), and
+ * actual pathkeys afterwards */
+
+ List *group_pathkeys; /* groupClause pathkeys, if any */
+ List *sort_pathkeys; /* sortClause pathkeys, if any */
+
+ MemoryContext planner_cxt; /* context holding PlannerInfo */
+
+ double total_table_pages; /* # of pages in all tables of query */
+
+ double tuple_fraction; /* tuple_fraction passed to query_planner */
+
+ bool hasJoinRTEs; /* true if any RTEs are RTE_JOIN kind */
+ bool hasOuterJoins; /* true if any RTEs are outer joins */
+ bool hasHavingQual; /* true if havingQual was non-null */
+ bool hasPseudoConstantQuals; /* true if any RestrictInfo has
+ * pseudoconstant = true */
+} PlannerInfo;
+
+
/*----------
* RelOptInfo
* Per-relation information for planning/optimization
* In either case it is uniquely identified by an RT index. A "joinrel"
* is the joining of two or more base rels. A joinrel is identified by
* the set of RT indexes for its component baserels. We create RelOptInfo
- * nodes for each baserel and joinrel, and store them in the Query's
- * base_rel_list and join_rel_list respectively.
+ * nodes for each baserel and joinrel, and store them in the PlannerInfo's
+ * simple_rel_array and join_rel_list respectively.
*
* Note that there is only one joinrel for any given set of component
* baserels, no matter what order we assemble them in; so an unordered
* set is the right datatype to identify it with.
*
* We also have "other rels", which are like base rels in that they refer to
- * single RT indexes; but they are not part of the join tree, and are stored
- * in other_rel_list not base_rel_list.
- *
- * Currently the only kind of otherrels are those made for child relations
- * of an inheritance scan (SELECT FROM foo*). The parent table's RTE and
- * corresponding baserel represent the whole result of the inheritance scan.
- * The planner creates separate RTEs and associated RelOptInfos for each child
- * table (including the parent table, in its capacity as a member of the
- * inheritance set). These RelOptInfos are physically identical to baserels,
- * but are otherrels because they are not in the main join tree. These added
- * RTEs and otherrels are used to plan the scans of the individual tables in
- * the inheritance set; then the parent baserel is given an Append plan
- * comprising the best plans for the individual child tables.
+ * single RT indexes; but they are not part of the join tree, and are given
+ * a different RelOptKind to identify them.
+ *
+ * Currently the only kind of otherrels are those made for member relations
+ * of an "append relation", that is an inheritance set or UNION ALL subquery.
+ * An append relation has a parent RTE that is a base rel, which represents
+ * the entire append relation. The member RTEs are otherrels. The parent
+ * is present in the query join tree but the members are not. The member
+ * RTEs and otherrels are used to plan the scans of the individual tables or
+ * subqueries of the append set; then the parent baserel is given an Append
+ * plan comprising the best plans for the individual member rels. (See
+ * comments for AppendRelInfo for more information.)
*
* At one time we also made otherrels to represent join RTEs, for use in
* handling join alias Vars. Currently this is not needed because all join
* appropriate projections have been done (ie, output width)
* reltargetlist - List of Var nodes for the attributes we need to
* output from this relation (in no particular order)
+ * NOTE: in a child relation, may contain RowExprs
* pathlist - List of Path nodes, one for each potentially useful
* method of generating the relation
* cheapest_startup_path - the pathlist member with lowest startup cost
* (regardless of its ordering)
* cheapest_unique_path - for caching cheapest path to produce unique
* (no duplicates) output from relation
- * pruneable - flag to let the planner know whether it can prune the
- * pathlist of this RelOptInfo or not.
*
* If the relation is a base relation it will have these fields set:
*
* upon creation of the RelOptInfo object; they are filled in when
* set_base_rel_pathlist processes the object.
*
- * For otherrels that are inheritance children, these fields are filled
+ * For otherrels that are appendrel members, these fields are filled
* in just as for a baserel.
*
* The presence of the remaining fields depends on the restrictions
* and joins that the relation participates in:
*
* baserestrictinfo - List of RestrictInfo nodes, containing info about
- * each qualification clause in which this relation
+ * each non-join qualification clause in which this relation
* participates (only used for base rels)
* baserestrictcost - Estimated cost of evaluating the baserestrictinfo
* clauses at a single tuple (only used for base rels)
- * outerjoinset - For a base rel: if the rel appears within the nullable
- * side of an outer join, the set of all relids
- * participating in the highest such outer join; else NULL.
- * Otherwise, unused.
- * joininfo - List of JoinInfo nodes, containing info about each join
- * clause in which this relation participates
+ * joininfo - List of RestrictInfo nodes, containing info about each
+ * join clause in which this relation participates (but
+ * note this excludes clauses that might be derivable from
+ * EquivalenceClasses)
+ * has_eclass_joins - flag that EquivalenceClass joins are possible
* index_outer_relids - only used for base rels; set of outer relids
* that participate in indexable joinclauses for this rel
* index_inner_paths - only used for base rels; list of InnerIndexscanInfo
* and should not be processed again at the level of {1 2 3}.) Therefore,
* the restrictinfo list in the join case appears in individual JoinPaths
* (field joinrestrictinfo), not in the parent relation. But it's OK for
- * the RelOptInfo to store the joininfo lists, because those are the same
+ * the RelOptInfo to store the joininfo list, because that is the same
* for a given rel no matter how we form it.
*
* We store baserestrictcost in the RelOptInfo (for base relations) because
* we know we will need it at least once (to price the sequential scan)
* and may need it multiple times to price index scans.
- *
- * outerjoinset is used to ensure correct placement of WHERE clauses that
- * apply to outer-joined relations; we must not apply such WHERE clauses
- * until after the outer join is performed.
*----------
*/
typedef enum RelOptKind
{
RELOPT_BASEREL,
RELOPT_JOINREL,
- RELOPT_OTHER_CHILD_REL
+ RELOPT_OTHER_MEMBER_REL
} RelOptKind;
typedef struct RelOptInfo
int width; /* estimated avg width of result tuples */
/* materialization information */
- FastList reltargetlist;
+ List *reltargetlist; /* needed Vars */
List *pathlist; /* Path structures */
struct Path *cheapest_startup_path;
struct Path *cheapest_total_path;
struct Path *cheapest_unique_path;
- bool pruneable;
/* information about a base rel (not set for join rels!) */
Index relid;
Relids *attr_needed; /* array indexed [min_attr .. max_attr] */
int32 *attr_widths; /* array indexed [min_attr .. max_attr] */
List *indexlist;
- long pages;
+ BlockNumber pages;
double tuples;
struct Plan *subplan; /* if subquery */
/* used by various scans and joins: */
- List *baserestrictinfo; /* RestrictInfo structures (if
- * base rel) */
+ List *baserestrictinfo; /* RestrictInfo structures (if base
+ * rel) */
QualCost baserestrictcost; /* cost of evaluating the above */
- Relids outerjoinset; /* set of base relids */
- List *joininfo; /* JoinInfo structures */
+ List *joininfo; /* RestrictInfo structures for join clauses
+ * involving this rel */
+ bool has_eclass_joins; /* T means joininfo is incomplete */
/* cached info about inner indexscan paths for relation: */
Relids index_outer_relids; /* other relids in indexable join
/*
* Inner indexscans are not in the main pathlist because they are not
- * usable except in specific join contexts. We use the
- * index_inner_paths list just to avoid recomputing the best inner
- * indexscan repeatedly for similar outer relations. See comments for
- * InnerIndexscanInfo.
+ * usable except in specific join contexts. We use the index_inner_paths
+ * list just to avoid recomputing the best inner indexscan repeatedly for
+ * similar outer relations. See comments for InnerIndexscanInfo.
*/
} RelOptInfo;
* and indexes, but that created confusion without actually doing anything
* useful. So now we have a separate IndexOptInfo struct for indexes.
*
- * classlist[], indexkeys[], and ordering[] have ncolumns entries.
+ * opfamily[], indexkeys[], fwdsortop[], revsortop[], and nulls_first[]
+ * each have ncolumns entries. Note: for historical reasons, the
+ * opfamily array has an extra entry that is always zero. Some code
+ * scans until it sees a zero entry, rather than looking at ncolumns.
+ *
* Zeroes in the indexkeys[] array indicate index columns that are
* expressions; there is one element in indexprs for each such column.
*
- * Note: for historical reasons, the classlist and ordering arrays have
- * an extra entry that is always zero. Some code scans until it sees a
- * zero entry, rather than looking at ncolumns.
+ * For an unordered index, the sortop arrays contains zeroes. Note that
+ * fwdsortop[] and nulls_first[] describe the sort ordering of a forward
+ * indexscan; we can also consider a backward indexscan, which will
+ * generate sort order described by revsortop/!nulls_first.
*
* The indexprs and indpred expressions have been run through
* prepqual.c and eval_const_expressions() for ease of matching to
- * WHERE clauses. indpred is in implicit-AND form.
+ * WHERE clauses. indpred is in implicit-AND form.
*/
-
typedef struct IndexOptInfo
{
NodeTag type;
Oid indexoid; /* OID of the index relation */
+ RelOptInfo *rel; /* back-link to index's table */
/* statistics from pg_class */
- long pages; /* number of disk pages in index */
+ BlockNumber pages; /* number of disk pages in index */
double tuples; /* number of index tuples in index */
/* index descriptor information */
int ncolumns; /* number of columns in index */
- Oid *classlist; /* OIDs of operator classes for columns */
+ Oid *opfamily; /* OIDs of operator families for columns */
int *indexkeys; /* column numbers of index's keys, or 0 */
- Oid *ordering; /* OIDs of sort operators for each column */
+ Oid *fwdsortop; /* OIDs of sort operators for each column */
+ Oid *revsortop; /* OIDs of sort operators for backward scan */
+ bool *nulls_first; /* do NULLs come first in the sort order? */
Oid relam; /* OID of the access method (in pg_am) */
RegProcedure amcostestimate; /* OID of the access method's cost fcn */
- List *indexprs; /* expressions for non-simple index
- * columns */
+ List *indexprs; /* expressions for non-simple index columns */
List *indpred; /* predicate if a partial index, else NIL */
- bool unique; /* true if a unique index */
- /* cached info about inner indexscan paths for index */
- Relids outer_relids; /* other relids in usable join clauses */
- List *inner_paths; /* List of InnerIndexscanInfo nodes */
+ bool predOK; /* true if predicate matches query */
+ bool unique; /* true if a unique index */
+ bool amoptionalkey; /* can query omit key for the first column? */
} IndexOptInfo;
/*
- * PathKeys
+ * EquivalenceClasses
+ *
+ * Whenever we can determine that a mergejoinable equality clause A = B is
+ * not delayed by any outer join, we create an EquivalenceClass containing
+ * the expressions A and B to record this knowledge. If we later find another
+ * equivalence B = C, we add C to the existing EquivalenceClass; this may
+ * require merging two existing EquivalenceClasses. At the end of the qual
+ * distribution process, we have sets of values that are known all transitively
+ * equal to each other, where "equal" is according to the rules of the btree
+ * operator family(s) shown in ec_opfamilies. (We restrict an EC to contain
+ * only equalities whose operators belong to the same set of opfamilies. This
+ * could probably be relaxed, but for now it's not worth the trouble, since
+ * nearly all equality operators belong to only one btree opclass anyway.)
+ *
+ * We also use EquivalenceClasses as the base structure for PathKeys, letting
+ * us represent knowledge about different sort orderings being equivalent.
+ * Since every PathKey must reference an EquivalenceClass, we will end up
+ * with single-member EquivalenceClasses whenever a sort key expression has
+ * not been equivalenced to anything else. It is also possible that such an
+ * EquivalenceClass will contain a volatile expression ("ORDER BY random()"),
+ * which is a case that can't arise otherwise since clauses containing
+ * volatile functions are never considered mergejoinable. We mark such
+ * EquivalenceClasses specially to prevent them from being merged with
+ * ordinary EquivalenceClasses.
*
- * The sort ordering of a path is represented by a list of sublists of
- * PathKeyItem nodes. An empty list implies no known ordering. Otherwise
- * the first sublist represents the primary sort key, the second the
- * first secondary sort key, etc. Each sublist contains one or more
- * PathKeyItem nodes, each of which can be taken as the attribute that
- * appears at that sort position. (See the top of optimizer/path/pathkeys.c
- * for more information.)
+ * We allow equality clauses appearing below the nullable side of an outer join
+ * to form EquivalenceClasses, but these have a slightly different meaning:
+ * the included values might be all NULL rather than all the same non-null
+ * values. See src/backend/optimizer/README for more on that point.
+ *
+ * NB: if ec_merged isn't NULL, this class has been merged into another, and
+ * should be ignored in favor of using the pointed-to class.
*/
+typedef struct EquivalenceClass
+{
+ NodeTag type;
+
+ List *ec_opfamilies; /* btree operator family OIDs */
+ List *ec_members; /* list of EquivalenceMembers */
+ List *ec_sources; /* list of generating RestrictInfos */
+ Relids ec_relids; /* all relids appearing in ec_members */
+ bool ec_has_const; /* any pseudoconstants in ec_members? */
+ bool ec_has_volatile; /* the (sole) member is a volatile expr */
+ bool ec_below_outer_join; /* equivalence applies below an OJ */
+ bool ec_broken; /* failed to generate needed clauses? */
+ struct EquivalenceClass *ec_merged; /* set if merged into another EC */
+} EquivalenceClass;
-typedef struct PathKeyItem
+/*
+ * EquivalenceMember - one member expression of an EquivalenceClass
+ *
+ * em_is_child signifies that this element was built by transposing a member
+ * for an inheritance parent relation to represent the corresponding expression
+ * on an inheritance child. The element should be ignored for all purposes
+ * except constructing inner-indexscan paths for the child relation. (Other
+ * types of join are driven from transposed joininfo-list entries.) Note
+ * that the EC's ec_relids field does NOT include the child relation.
+ *
+ * em_datatype is usually the same as exprType(em_expr), but can be
+ * different when dealing with a binary-compatible opfamily; in particular
+ * anyarray_ops would never work without this. Use em_datatype when
+ * looking up a specific btree operator to work with this expression.
+ */
+typedef struct EquivalenceMember
{
NodeTag type;
- Node *key; /* the item that is ordered */
- Oid sortop; /* the ordering operator ('<' op) */
+ Expr *em_expr; /* the expression represented */
+ Relids em_relids; /* all relids appearing in em_expr */
+ bool em_is_const; /* expression is pseudoconstant? */
+ bool em_is_child; /* derived version for a child relation? */
+ Oid em_datatype; /* the "nominal type" used by the opfamily */
+} EquivalenceMember;
- /*
- * key typically points to a Var node, ie a relation attribute, but it
- * can also point to an arbitrary expression representing the value
- * indexed by an index expression.
- */
-} PathKeyItem;
+/*
+ * PathKeys
+ *
+ * The sort ordering of a path is represented by a list of PathKey nodes.
+ * An empty list implies no known ordering. Otherwise the first item
+ * represents the primary sort key, the second the first secondary sort key,
+ * etc. The value being sorted is represented by linking to an
+ * EquivalenceClass containing that value and including pk_opfamily among its
+ * ec_opfamilies. This is a convenient method because it makes it trivial
+ * to detect equivalent and closely-related orderings. (See optimizer/README
+ * for more information.)
+ *
+ * Note: pk_strategy is either BTLessStrategyNumber (for ASC) or
+ * BTGreaterStrategyNumber (for DESC). We assume that all ordering-capable
+ * index types will use btree-compatible strategy numbers.
+ */
+
+typedef struct PathKey
+{
+ NodeTag type;
+
+ EquivalenceClass *pk_eclass; /* the value that is ordered */
+ Oid pk_opfamily; /* btree opfamily defining the ordering */
+ int pk_strategy; /* sort direction (ASC or DESC) */
+ bool pk_nulls_first; /* do NULLs come before normal values? */
+} PathKey;
/*
* Type "Path" is used as-is for sequential-scan paths. For other
{
NodeTag type;
+ NodeTag pathtype; /* tag identifying scan/join method */
+
RelOptInfo *parent; /* the relation this path can build */
/* estimated execution costs for path (see costsize.c for more info) */
- Cost startup_cost; /* cost expended before fetching any
- * tuples */
- Cost total_cost; /* total cost (assuming all tuples
- * fetched) */
-
- NodeTag pathtype; /* tag identifying scan/join method */
+ Cost startup_cost; /* cost expended before fetching any tuples */
+ Cost total_cost; /* total cost (assuming all tuples fetched) */
List *pathkeys; /* sort ordering of path's output */
- /* pathkeys is a List of Lists of PathKeyItem nodes; see above */
+ /* pathkeys is a List of PathKey nodes; see above */
} Path;
/*----------
- * IndexPath represents an index scan. Although an indexscan can only read
- * a single relation, it can scan it more than once, potentially using a
- * different index during each scan. The result is the union (OR) of all the
- * tuples matched during any scan. (The executor is smart enough not to return
- * the same tuple more than once, even if it is matched in multiple scans.)
- *
- * 'indexinfo' is a list of IndexOptInfo nodes, one per scan to be performed.
- *
- * 'indexqual' is a list of index qualifications, also one per scan.
- * Each entry in 'indexqual' is a sublist of qualification expressions with
- * implicit AND semantics across the sublist items. Only expressions that
- * are usable as indexquals (as determined by indxpath.c) may appear here.
- * NOTE that the semantics of the top-level list in 'indexqual' is OR
- * combination, while the sublists are implicitly AND combinations!
- * Also note that indexquals lists do not contain RestrictInfo nodes,
- * just bare clause expressions.
- *
- * 'indexjoinclauses' is NIL for an ordinary indexpath (one that does not
- * use any join clauses in the index conditions). For an innerjoin indexpath,
- * it has the same structure as 'indexqual', but references the RestrictInfo
- * nodes from which the indexqual was built, rather than the bare clause
- * expressions. (Note: there isn't necessarily a one-to-one correspondence
- * between RestrictInfos and expressions, because of expansion of special
- * indexable operators.) We need this so that we can eliminate redundant
- * join clauses when plans are built.
+ * IndexPath represents an index scan over a single index.
+ *
+ * 'indexinfo' is the index to be scanned.
+ *
+ * 'indexclauses' is a list of index qualification clauses, with implicit
+ * AND semantics across the list. Each clause is a RestrictInfo node from
+ * the query's WHERE or JOIN conditions.
+ *
+ * 'indexquals' has the same structure as 'indexclauses', but it contains
+ * the actual indexqual conditions that can be used with the index.
+ * In simple cases this is identical to 'indexclauses', but when special
+ * indexable operators appear in 'indexclauses', they are replaced by the
+ * derived indexscannable conditions in 'indexquals'.
+ *
+ * 'isjoininner' is TRUE if the path is a nestloop inner scan (that is,
+ * some of the index conditions are join rather than restriction clauses).
+ * Note that the path costs will be calculated differently from a plain
+ * indexscan in this case, and in addition there's a special 'rows' value
+ * different from the parent RelOptInfo's (see below).
*
* 'indexscandir' is one of:
* ForwardScanDirection: forward scan of an ordered index
* NoMovementScanDirection for an indexscan, but the planner wants to
* distinguish ordered from unordered indexes for building pathkeys.)
*
+ * 'indextotalcost' and 'indexselectivity' are saved in the IndexPath so that
+ * we need not recompute them when considering using the same index in a
+ * bitmap index/heap scan (see BitmapHeapPath). The costs of the IndexPath
+ * itself represent the costs of an IndexScan plan type.
+ *
* 'rows' is the estimated result tuple count for the indexscan. This
* is the same as path.parent->rows for a simple indexscan, but it is
* different for a nestloop inner scan, because the additional indexquals
typedef struct IndexPath
{
Path path;
- List *indexinfo;
- List *indexqual;
- List *indexjoinclauses;
+ IndexOptInfo *indexinfo;
+ List *indexclauses;
+ List *indexquals;
+ bool isjoininner;
ScanDirection indexscandir;
+ Cost indextotalcost;
+ Selectivity indexselectivity;
double rows; /* estimated number of result tuples */
} IndexPath;
+/*
+ * BitmapHeapPath represents one or more indexscans that generate TID bitmaps
+ * instead of directly accessing the heap, followed by AND/OR combinations
+ * to produce a single bitmap, followed by a heap scan that uses the bitmap.
+ * Note that the output is always considered unordered, since it will come
+ * out in physical heap order no matter what the underlying indexes did.
+ *
+ * The individual indexscans are represented by IndexPath nodes, and any
+ * logic on top of them is represented by a tree of BitmapAndPath and
+ * BitmapOrPath nodes. Notice that we can use the same IndexPath node both
+ * to represent a regular IndexScan plan, and as the child of a BitmapHeapPath
+ * that represents scanning the same index using a BitmapIndexScan. The
+ * startup_cost and total_cost figures of an IndexPath always represent the
+ * costs to use it as a regular IndexScan. The costs of a BitmapIndexScan
+ * can be computed using the IndexPath's indextotalcost and indexselectivity.
+ *
+ * BitmapHeapPaths can be nestloop inner indexscans. The isjoininner and
+ * rows fields serve the same purpose as for plain IndexPaths.
+ */
+typedef struct BitmapHeapPath
+{
+ Path path;
+ Path *bitmapqual; /* IndexPath, BitmapAndPath, BitmapOrPath */
+ bool isjoininner; /* T if it's a nestloop inner scan */
+ double rows; /* estimated number of result tuples */
+} BitmapHeapPath;
+
+/*
+ * BitmapAndPath represents a BitmapAnd plan node; it can only appear as
+ * part of the substructure of a BitmapHeapPath. The Path structure is
+ * a bit more heavyweight than we really need for this, but for simplicity
+ * we make it a derivative of Path anyway.
+ */
+typedef struct BitmapAndPath
+{
+ Path path;
+ List *bitmapquals; /* IndexPaths and BitmapOrPaths */
+ Selectivity bitmapselectivity;
+} BitmapAndPath;
+
+/*
+ * BitmapOrPath represents a BitmapOr plan node; it can only appear as
+ * part of the substructure of a BitmapHeapPath. The Path structure is
+ * a bit more heavyweight than we really need for this, but for simplicity
+ * we make it a derivative of Path anyway.
+ */
+typedef struct BitmapOrPath
+{
+ Path path;
+ List *bitmapquals; /* IndexPaths and BitmapAndPaths */
+ Selectivity bitmapselectivity;
+} BitmapOrPath;
+
/*
* TidPath represents a scan by TID
+ *
+ * tidquals is an implicitly OR'ed list of qual expressions of the form
+ * "CTID = pseudoconstant" or "CTID = ANY(pseudoconstant_array)".
+ * Note they are bare expressions, not RestrictInfos.
*/
typedef struct TidPath
{
Path path;
- List *tideval; /* qual(s) involving CTID = something */
+ List *tidquals; /* qual(s) involving CTID = something */
} TidPath;
/*
* AppendPath represents an Append plan, ie, successive execution of
- * several member plans. Currently it is only used to handle expansion
- * of inheritance trees.
+ * several member plans.
+ *
+ * Note: it is possible for "subpaths" to contain only one, or even no,
+ * elements. These cases are optimized during create_append_plan.
*/
typedef struct AppendPath
{
} AppendPath;
/*
- * ResultPath represents use of a Result plan node, either to compute a
- * variable-free targetlist or to gate execution of a subplan with a
- * one-time (variable-free) qual condition. Note that in the former case
- * path.parent will be NULL; in the latter case it is copied from the subpath.
+ * ResultPath represents use of a Result plan node to compute a variable-free
+ * targetlist with no underlying tables (a "SELECT expressions" query).
+ * The query could have a WHERE clause, too, represented by "quals".
+ *
+ * Note that quals is a list of bare clauses, not RestrictInfos.
*/
typedef struct ResultPath
{
Path path;
- Path *subpath;
- List *constantqual;
+ List *quals;
} ResultPath;
/*
* its subpath.
*
* This is unlike the other Path nodes in that it can actually generate
- * two different plans: either hash-based or sort-based implementation.
- * The decision is sufficiently localized that it's not worth having two
- * separate Path node types.
+ * different plans: either hash-based or sort-based implementation, or a
+ * no-op if the input path can be proven distinct already. The decision
+ * is sufficiently localized that it's not worth having separate Path node
+ * types. (Note: in the no-op case, we could eliminate the UniquePath node
+ * entirely and just return the subpath; but it's convenient to have a
+ * UniquePath in the path tree to signal upper-level routines that the input
+ * is known distinct.)
*/
+typedef enum
+{
+ UNIQUE_PATH_NOOP, /* input is known unique already */
+ UNIQUE_PATH_HASH, /* use hashing */
+ UNIQUE_PATH_SORT /* use sorting */
+} UniquePathMethod;
+
typedef struct UniquePath
{
Path path;
Path *subpath;
- bool use_hash;
+ UniquePathMethod umethod;
double rows; /* estimated number of result tuples */
} UniquePath;
* A mergejoin path has these fields.
*
* path_mergeclauses lists the clauses (in the form of RestrictInfos)
- * that will be used in the merge. (Before 7.0, this was a list of bare
- * clause expressions, but we can save on list memory and cost_qual_eval
- * work by leaving it in the form of a RestrictInfo list.)
+ * that will be used in the merge.
*
* Note that the mergeclauses are a subset of the parent relation's
* restriction-clause list. Any join clauses that are not mergejoinable
typedef struct MergePath
{
JoinPath jpath;
- List *path_mergeclauses; /* join clauses to be used for
- * merge */
+ List *path_mergeclauses; /* join clauses to be used for merge */
List *outersortkeys; /* keys for explicit sort, if any */
List *innersortkeys; /* keys for explicit sort, if any */
} MergePath;
* in the baserestrictinfo list of the RelOptInfo for that base rel.
*
* If a restriction clause references more than one base rel, it will
- * appear in the JoinInfo lists of every RelOptInfo that describes a strict
- * subset of the base rels mentioned in the clause. The JoinInfo lists are
+ * appear in the joininfo list of every RelOptInfo that describes a strict
+ * subset of the base rels mentioned in the clause. The joininfo lists are
* used to drive join tree building by selecting plausible join candidates.
* The clause cannot actually be applied until we have built a join rel
* containing all the base rels it references, however.
* sequence we use. So, these clauses cannot be associated directly with
* the join RelOptInfo, but must be kept track of on a per-join-path basis.
*
+ * RestrictInfos that represent equivalence conditions (i.e., mergejoinable
+ * equalities that are not outerjoin-delayed) are handled a bit differently.
+ * Initially we attach them to the EquivalenceClasses that are derived from
+ * them. When we construct a scan or join path, we look through all the
+ * EquivalenceClasses and generate derived RestrictInfos representing the
+ * minimal set of conditions that need to be checked for this particular scan
+ * or join to enforce that all members of each EquivalenceClass are in fact
+ * equal in all rows emitted by the scan or join.
+ *
* When dealing with outer joins we have to be very careful about pushing qual
* clauses up and down the tree. An outer join's own JOIN/ON conditions must
* be evaluated exactly at that join node, and any quals appearing in WHERE or
* pushed down to a lower level than its original syntactic placement in the
* join tree would suggest. If an outer join prevents us from pushing a qual
* down to its "natural" semantic level (the level associated with just the
- * base rels used in the qual) then the qual will appear in JoinInfo lists
- * that reference more than just the base rels it actually uses. By
+ * base rels used in the qual) then we mark the qual with a "required_relids"
+ * value including more than just the base rels it actually uses. By
* pretending that the qual references all the rels appearing in the outer
* join, we prevent it from being evaluated below the outer join's joinrel.
* When we do form the outer join's joinrel, we still need to distinguish
* those quals that are actually in that join's JOIN/ON condition from those
* that appeared higher in the tree and were pushed down to the join rel
- * because they used no other rels. That's what the ispusheddown flag is for;
- * it tells us that a qual came from a point above the join of the specific
- * set of base rels that it uses (or that the JoinInfo structures claim it
- * uses). A clause that originally came from WHERE will *always* have its
- * ispusheddown flag set; a clause that came from an INNER JOIN condition,
- * but doesn't use all the rels being joined, will also have ispusheddown set
- * because it will get attached to some lower joinrel.
+ * because they used no other rels. That's what the is_pushed_down flag is
+ * for; it tells us that a qual came from a point above the join of the
+ * set of base rels listed in required_relids. A clause that originally came
+ * from WHERE will *always* have its is_pushed_down flag set; a clause that
+ * came from an INNER JOIN condition, but doesn't use all the rels being
+ * joined, will also have is_pushed_down set because it will get attached to
+ * some lower joinrel.
+ *
+ * When application of a qual must be delayed by outer join, we also mark it
+ * with outerjoin_delayed = true. This isn't redundant with required_relids
+ * because that might equal clause_relids whether or not it's an outer-join
+ * clause.
*
* In general, the referenced clause might be arbitrarily complex. The
* kinds of clauses we can handle as indexscan quals, mergejoin clauses,
- * or hashjoin clauses are fairly limited --- the code for each kind of
- * path is responsible for identifying the restrict clauses it can use
- * and ignoring the rest. Clauses not implemented by an indexscan,
+ * or hashjoin clauses are limited (e.g., no volatile functions). The code
+ * for each kind of path is responsible for identifying the restrict clauses
+ * it can use and ignoring the rest. Clauses not implemented by an indexscan,
* mergejoin, or hashjoin will be placed in the plan qual or joinqual field
* of the finished Plan node, where they will be enforced by general-purpose
* qual-expression-evaluation code. (But we are still entitled to count
* their selectivity when estimating the result tuple count, if we
* can guess what it is...)
+ *
+ * When the referenced clause is an OR clause, we generate a modified copy
+ * in which additional RestrictInfo nodes are inserted below the top-level
+ * OR/AND structure. This is a convenience for OR indexscan processing:
+ * indexquals taken from either the top level or an OR subclause will have
+ * associated RestrictInfo nodes.
+ *
+ * The can_join flag is set true if the clause looks potentially useful as
+ * a merge or hash join clause, that is if it is a binary opclause with
+ * nonoverlapping sets of relids referenced in the left and right sides.
+ * (Whether the operator is actually merge or hash joinable isn't checked,
+ * however.)
+ *
+ * The pseudoconstant flag is set true if the clause contains no Vars of
+ * the current query level and no volatile functions. Such a clause can be
+ * pulled out and used as a one-time qual in a gating Result node. We keep
+ * pseudoconstant clauses in the same lists as other RestrictInfos so that
+ * the regular clause-pushing machinery can assign them to the correct join
+ * level, but they need to be treated specially for cost and selectivity
+ * estimates. Note that a pseudoconstant clause can never be an indexqual
+ * or merge or hash join clause, so it's of no interest to large parts of
+ * the planner.
+ *
+ * When join clauses are generated from EquivalenceClasses, there may be
+ * several equally valid ways to enforce join equivalence, of which we need
+ * apply only one. We mark clauses of this kind by setting parent_ec to
+ * point to the generating EquivalenceClass. Multiple clauses with the same
+ * parent_ec in the same join are redundant.
*/
typedef struct RestrictInfo
Expr *clause; /* the represented clause of WHERE or JOIN */
- bool ispusheddown; /* TRUE if clause was pushed down in level */
+ bool is_pushed_down; /* TRUE if clause was pushed down in level */
- /*
- * This flag is set true if the clause looks potentially useful as a
- * merge or hash join clause, that is if it is a binary opclause with
- * nonoverlapping sets of relids referenced in the left and right sides.
- * (Whether the operator is actually merge or hash joinable isn't
- * checked, however.)
- */
- bool canjoin;
+ bool outerjoin_delayed; /* TRUE if delayed by outer join */
+
+ bool can_join; /* see comment above */
+
+ bool pseudoconstant; /* see comment above */
+
+ /* The set of relids (varnos) actually referenced in the clause: */
+ Relids clause_relids;
+
+ /* The set of relids required to evaluate the clause: */
+ Relids required_relids;
/* These fields are set for any binary opclause: */
Relids left_relids; /* relids in left side of clause */
Relids right_relids; /* relids in right side of clause */
- /* only used if clause is an OR clause: */
- List *subclauseindices; /* indexes matching subclauses */
- /* subclauseindices is a List of Lists of IndexOptInfos */
+ /* This field is NULL unless clause is an OR clause: */
+ Expr *orclause; /* modified clause with RestrictInfos */
- /* cache space for costs (currently only used for join clauses) */
+ /* This field is NULL unless clause is potentially redundant: */
+ EquivalenceClass *parent_ec; /* generating EquivalenceClass */
+
+ /* cache space for cost and selectivity */
QualCost eval_cost; /* eval cost of clause; -1 if not yet set */
Selectivity this_selec; /* selectivity; -1 if not yet set */
- /* valid if clause is mergejoinable, else InvalidOid: */
- Oid mergejoinoperator; /* copy of clause operator */
- Oid left_sortop; /* leftside sortop needed for mergejoin */
- Oid right_sortop; /* rightside sortop needed for mergejoin */
+ /* valid if clause is mergejoinable, else NIL */
+ List *mergeopfamilies; /* opfamilies containing clause operator */
- /* cache space for mergeclause processing; NIL if not yet set */
- List *left_pathkey; /* canonical pathkey for left side */
- List *right_pathkey; /* canonical pathkey for right side */
+ /* cache space for mergeclause processing; NULL if not yet set */
+ EquivalenceClass *left_ec; /* EquivalenceClass containing lefthand */
+ EquivalenceClass *right_ec; /* EquivalenceClass containing righthand */
- /* cache space for mergeclause processing; -1 if not yet set */
- Selectivity left_mergescansel; /* fraction of left side to scan */
- Selectivity right_mergescansel; /* fraction of right side to scan */
+ /* transient workspace for use while considering a specific join path */
+ bool outer_is_left; /* T = outer var on left, F = on right */
/* valid if clause is hashjoinable, else InvalidOid: */
Oid hashjoinoperator; /* copy of clause operator */
Selectivity right_bucketsize; /* avg bucketsize of right side */
} RestrictInfo;
-/*
- * Join clause info.
- *
- * We make a list of these for each RelOptInfo, containing info about
- * all the join clauses this RelOptInfo participates in. (For this
- * purpose, a "join clause" is a WHERE clause that mentions both vars
- * belonging to this relation and vars belonging to relations not yet
- * joined to it.) We group these clauses according to the set of
- * other base relations (unjoined relations) mentioned in them.
- * There is one JoinInfo for each distinct set of unjoined_relids,
- * and its jinfo_restrictinfo lists the clause(s) that use that set
- * of other relations.
- */
-
-typedef struct JoinInfo
-{
- NodeTag type;
- Relids unjoined_relids; /* some rels not yet part of my RelOptInfo */
- List *jinfo_restrictinfo; /* relevant RestrictInfos */
-} JoinInfo;
-
/*
* Inner indexscan info.
*
* thus varies depending on which outer relation we consider; so we have
* to recompute the best such path for every join. To avoid lots of
* redundant computation, we cache the results of such searches. For
- * each index we compute the set of possible otherrelids (all relids
- * appearing in joinquals that could become indexquals for this index).
+ * each relation we compute the set of possible otherrelids (all relids
+ * appearing in joinquals that could become indexquals for this table).
* Two outer relations whose relids have the same intersection with this
* set will have the same set of available joinclauses and thus the same
- * best inner indexscan for that index. Similarly, for each base relation,
- * we form the union of the per-index otherrelids sets. Two outer relations
- * with the same intersection with that set will have the same best overall
- * inner indexscan for the base relation. We use lists of InnerIndexscanInfo
- * nodes to cache the results of these searches at both the index and
- * relation level.
+ * best inner indexscan for the inner relation. By taking the intersection
+ * before scanning the cache, we avoid recomputing when considering
+ * join rels that differ only by the inclusion of irrelevant other rels.
*
* The search key also includes a bool showing whether the join being
* considered is an outer join. Since we constrain the join order for
Path *best_innerpath; /* best inner indexscan, or NULL if none */
} InnerIndexscanInfo;
+/*
+ * Outer join info.
+ *
+ * One-sided outer joins constrain the order of joining partially but not
+ * completely. We flatten such joins into the planner's top-level list of
+ * relations to join, but record information about each outer join in an
+ * OuterJoinInfo struct. These structs are kept in the PlannerInfo node's
+ * oj_info_list.
+ *
+ * min_lefthand and min_righthand are the sets of base relids that must be
+ * available on each side when performing the outer join. lhs_strict is
+ * true if the outer join's condition cannot succeed when the LHS variables
+ * are all NULL (this means that the outer join can commute with upper-level
+ * outer joins even if it appears in their RHS). We don't bother to set
+ * lhs_strict for FULL JOINs, however.
+ *
+ * It is not valid for either min_lefthand or min_righthand to be empty sets;
+ * if they were, this would break the logic that enforces join order.
+ *
+ * Note: OuterJoinInfo directly represents only LEFT JOIN and FULL JOIN;
+ * RIGHT JOIN is handled by switching the inputs to make it a LEFT JOIN.
+ * We make an OuterJoinInfo for FULL JOINs even though there is no flexibility
+ * of planning for them, because this simplifies make_join_rel()'s API.
+ */
+
+typedef struct OuterJoinInfo
+{
+ NodeTag type;
+ Relids min_lefthand; /* base relids in minimum LHS for join */
+ Relids min_righthand; /* base relids in minimum RHS for join */
+ bool is_full_join; /* it's a FULL OUTER JOIN */
+ bool lhs_strict; /* joinclause is strict for some LHS rel */
+} OuterJoinInfo;
+
/*
* IN clause info.
*
* When we convert top-level IN quals into join operations, we must restrict
* the order of joining and use special join methods at some join points.
* We record information about each such IN clause in an InClauseInfo struct.
- * These structs are kept in the Query node's in_info_list.
+ * These structs are kept in the PlannerInfo node's in_info_list.
+ *
+ * Note: sub_targetlist is just a list of Vars or expressions; it does not
+ * contain TargetEntry nodes.
*/
typedef struct InClauseInfo
Relids lefthand; /* base relids in lefthand expressions */
Relids righthand; /* base relids coming from the subselect */
List *sub_targetlist; /* targetlist of original RHS subquery */
+ List *in_operators; /* OIDs of the IN's equality operator(s) */
+} InClauseInfo;
+
+/*
+ * Append-relation info.
+ *
+ * When we expand an inheritable table or a UNION-ALL subselect into an
+ * "append relation" (essentially, a list of child RTEs), we build an
+ * AppendRelInfo for each child RTE. The list of AppendRelInfos indicates
+ * which child RTEs must be included when expanding the parent, and each
+ * node carries information needed to translate Vars referencing the parent
+ * into Vars referencing that child.
+ *
+ * These structs are kept in the PlannerInfo node's append_rel_list.
+ * Note that we just throw all the structs into one list, and scan the
+ * whole list when desiring to expand any one parent. We could have used
+ * a more complex data structure (eg, one list per parent), but this would
+ * be harder to update during operations such as pulling up subqueries,
+ * and not really any easier to scan. Considering that typical queries
+ * will not have many different append parents, it doesn't seem worthwhile
+ * to complicate things.
+ *
+ * Note: after completion of the planner prep phase, any given RTE is an
+ * append parent having entries in append_rel_list if and only if its
+ * "inh" flag is set. We clear "inh" for plain tables that turn out not
+ * to have inheritance children, and (in an abuse of the original meaning
+ * of the flag) we set "inh" for subquery RTEs that turn out to be
+ * flattenable UNION ALL queries. This lets us avoid useless searches
+ * of append_rel_list.
+ *
+ * Note: the data structure assumes that append-rel members are single
+ * baserels. This is OK for inheritance, but it prevents us from pulling
+ * up a UNION ALL member subquery if it contains a join. While that could
+ * be fixed with a more complex data structure, at present there's not much
+ * point because no improvement in the plan could result.
+ */
+
+typedef struct AppendRelInfo
+{
+ NodeTag type;
/*
- * Note: sub_targetlist is just a list of Vars or expressions; it does
- * not contain TargetEntry nodes.
+ * These fields uniquely identify this append relationship. There can be
+ * (in fact, always should be) multiple AppendRelInfos for the same
+ * parent_relid, but never more than one per child_relid, since a given
+ * RTE cannot be a child of more than one append parent.
*/
-} InClauseInfo;
+ Index parent_relid; /* RT index of append parent rel */
+ Index child_relid; /* RT index of append child rel */
+
+ /*
+ * For an inheritance appendrel, the parent and child are both regular
+ * relations, and we store their rowtype OIDs here for use in translating
+ * whole-row Vars. For a UNION-ALL appendrel, the parent and child are
+ * both subqueries with no named rowtype, and we store InvalidOid here.
+ */
+ Oid parent_reltype; /* OID of parent's composite type */
+ Oid child_reltype; /* OID of child's composite type */
+
+ /*
+ * The N'th element of this list is the integer column number of the child
+ * column corresponding to the N'th column of the parent. A list element
+ * is zero if it corresponds to a dropped column of the parent (this is
+ * only possible for inheritance cases, not UNION ALL).
+ */
+ List *col_mappings; /* list of child attribute numbers */
+
+ /*
+ * The N'th element of this list is a Var or expression representing the
+ * child column corresponding to the N'th column of the parent. This is
+ * used to translate Vars referencing the parent rel into references to
+ * the child. A list element is NULL if it corresponds to a dropped
+ * column of the parent (this is only possible for inheritance cases, not
+ * UNION ALL).
+ *
+ * This might seem redundant with the col_mappings data, but it is handy
+ * because flattening of sub-SELECTs that are members of a UNION ALL will
+ * cause changes in the expressions that need to be substituted for a
+ * parent Var. Adjusting this data structure lets us track what really
+ * needs to be substituted.
+ *
+ * Notice we only store entries for user columns (attno > 0). Whole-row
+ * Vars are special-cased, and system columns (attno < 0) need no special
+ * translation since their attnos are the same for all tables.
+ *
+ * Caution: the Vars have varlevelsup = 0. Be careful to adjust as needed
+ * when copying into a subquery.
+ */
+ List *translated_vars; /* Expressions in the child's Vars */
+
+ /*
+ * We store the parent table's OID here for inheritance, or InvalidOid for
+ * UNION ALL. This is only needed to help in generating error messages if
+ * an attempt is made to reference a dropped parent column.
+ */
+ Oid parent_reloid; /* OID of parent relation */
+} AppendRelInfo;
#endif /* RELATION_H */