/*-------------------------------------------------------------------------
*
* indxpath.c
- * Routines to determine which indices are usable for scanning a
- * given relation, and create IndexPaths accordingly.
+ * Routines to determine which indexes are usable for scanning a
+ * given relation, and create Paths accordingly.
*
* Portions Copyright (c) 1996-2005, PostgreSQL Global Development Group
* Portions Copyright (c) 1994, Regents of the University of California
*
*
* IDENTIFICATION
- * $PostgreSQL: pgsql/src/backend/optimizer/path/indxpath.c,v 1.167 2004/12/31 22:00:04 pgsql Exp $
+ * $PostgreSQL: pgsql/src/backend/optimizer/path/indxpath.c,v 1.180 2005/05/06 17:24:54 tgl Exp $
*
*-------------------------------------------------------------------------
*/
#include "utils/builtins.h"
#include "utils/catcache.h"
#include "utils/lsyscache.h"
+#include "utils/memutils.h"
#include "utils/pg_locale.h"
#include "utils/selfuncs.h"
#include "utils/syscache.h"
#define is_indexable_operator(clause,opclass,indexkey_on_left) \
(indexable_operator(clause,opclass,indexkey_on_left) != InvalidOid)
+#define IsBooleanOpclass(opclass) \
+ ((opclass) == BOOL_BTREE_OPS_OID || (opclass) == BOOL_HASH_OPS_OID)
-static List *group_clauses_by_indexkey(RelOptInfo *rel, IndexOptInfo *index);
-static List *group_clauses_by_indexkey_for_join(Query *root,
- RelOptInfo *rel, IndexOptInfo *index,
- Relids outer_relids,
- JoinType jointype, bool isouterjoin);
-static bool match_clause_to_indexcol(RelOptInfo *rel, IndexOptInfo *index,
+
+static List *find_usable_indexes(Query *root, RelOptInfo *rel,
+ List *clauses, List *outer_clauses,
+ bool istoplevel, bool isjoininner,
+ Relids outer_relids);
+static Path *choose_bitmap_and(Query *root, RelOptInfo *rel, List *paths);
+static int bitmap_path_comparator(const void *a, const void *b);
+static Cost bitmap_and_cost_est(Query *root, RelOptInfo *rel, List *paths);
+static bool match_clause_to_indexcol(IndexOptInfo *index,
int indexcol, Oid opclass,
- RestrictInfo *rinfo);
-static bool match_join_clause_to_indexcol(RelOptInfo *rel, IndexOptInfo *index,
- int indexcol, Oid opclass,
- RestrictInfo *rinfo);
+ RestrictInfo *rinfo,
+ Relids outer_relids);
static Oid indexable_operator(Expr *clause, Oid opclass,
bool indexkey_on_left);
-static bool pred_test_recurse_pred(Expr *predicate, List *restrictinfo_list);
-static bool pred_test_restrict_list(Expr *predicate, List *restrictinfo_list);
-static bool pred_test_recurse_restrict(Expr *predicate, Node *clause);
+static bool pred_test_recurse(Node *clause, Node *predicate);
static bool pred_test_simple_clause(Expr *predicate, Node *clause);
-static Relids indexable_outerrelids(RelOptInfo *rel, IndexOptInfo *index);
-static Path *make_innerjoin_index_path(Query *root,
- RelOptInfo *rel, IndexOptInfo *index,
- List *clausegroups);
-static bool match_index_to_operand(Node *operand, int indexcol,
- RelOptInfo *rel, IndexOptInfo *index);
+static Relids indexable_outerrelids(RelOptInfo *rel);
+static bool list_matches_any_index(List *clauses, RelOptInfo *rel,
+ Relids outer_relids);
+static bool matches_any_index(RestrictInfo *rinfo, RelOptInfo *rel,
+ Relids outer_relids);
+static List *find_clauses_for_join(Query *root, RelOptInfo *rel,
+ Relids outer_relids, bool isouterjoin);
+static bool match_boolean_index_clause(Node *clause, int indexcol,
+ IndexOptInfo *index);
static bool match_special_index_operator(Expr *clause, Oid opclass,
bool indexkey_on_left);
+static Expr *expand_boolean_index_clause(Node *clause, int indexcol,
+ IndexOptInfo *index);
static List *expand_indexqual_condition(RestrictInfo *rinfo, Oid opclass);
static List *prefix_quals(Node *leftop, Oid opclass,
Const *prefix, Pattern_Prefix_Status pstatus);
void
create_index_paths(Query *root, RelOptInfo *rel)
{
- Relids all_join_outerrelids = NULL;
+ List *indexpaths;
+ List *bitindexpaths;
+ ListCell *l;
+
+ /* Skip the whole mess if no indexes */
+ if (rel->indexlist == NIL)
+ {
+ rel->index_outer_relids = NULL;
+ return;
+ }
+
+ /*
+ * Examine join clauses to see which ones are potentially usable with
+ * indexes of this rel, and generate the set of all other relids that
+ * participate in such join clauses. We'll use this set later to
+ * recognize outer rels that are equivalent for joining purposes.
+ */
+ rel->index_outer_relids = indexable_outerrelids(rel);
+
+ /*
+ * Find all the index paths that are directly usable for this relation
+ * (ie, are valid without considering OR or JOIN clauses).
+ */
+ indexpaths = find_usable_indexes(root, rel,
+ rel->baserestrictinfo, NIL,
+ true, false, NULL);
+
+ /*
+ * We can submit them all to add_path. (This generates access paths for
+ * plain IndexScan plans.) However, for the next step we will only want
+ * the ones that have some selectivity; we must discard anything that was
+ * generated solely for ordering purposes.
+ */
+ bitindexpaths = NIL;
+ foreach(l, indexpaths)
+ {
+ IndexPath *ipath = (IndexPath *) lfirst(l);
+
+ add_path(rel, (Path *) ipath);
+
+ if (ipath->indexselectivity < 1.0 &&
+ !ScanDirectionIsBackward(ipath->indexscandir))
+ bitindexpaths = lappend(bitindexpaths, ipath);
+ }
+
+ /*
+ * Generate BitmapOrPaths for any suitable OR-clauses present in the
+ * restriction list. Add these to bitindexpaths.
+ */
+ indexpaths = generate_bitmap_or_paths(root, rel,
+ rel->baserestrictinfo, NIL,
+ false, NULL);
+ bitindexpaths = list_concat(bitindexpaths, indexpaths);
+
+ /*
+ * If we found anything usable, generate a BitmapHeapPath for the
+ * most promising combination of bitmap index paths.
+ */
+ if (bitindexpaths != NIL)
+ {
+ Path *bitmapqual;
+ BitmapHeapPath *bpath;
+
+ bitmapqual = choose_bitmap_and(root, rel, bitindexpaths);
+ bpath = create_bitmap_heap_path(root, rel, bitmapqual, false);
+ add_path(rel, (Path *) bpath);
+ }
+}
+
+
+/*----------
+ * find_usable_indexes
+ * Given a list of restriction clauses, find all the potentially usable
+ * indexes for the given relation, and return a list of IndexPaths.
+ *
+ * The caller actually supplies two lists of restriction clauses: some
+ * "current" ones and some "outer" ones. Both lists can be used freely
+ * to match keys of the index, but an index must use at least one of the
+ * "current" clauses to be considered usable. The motivation for this is
+ * examples like
+ * WHERE (x = 42) AND (... OR (y = 52 AND z = 77) OR ....)
+ * While we are considering the y/z subclause of the OR, we can use "x = 42"
+ * as one of the available index conditions; but we shouldn't match the
+ * subclause to any index on x alone, because such a Path would already have
+ * been generated at the upper level. So we could use an index on x,y,z
+ * or an index on x,y for the OR subclause, but not an index on just x.
+ *
+ * If istoplevel is true (indicating we are considering the top level of a
+ * rel's restriction clauses), we will include indexes in the result that
+ * have an interesting sort order, even if they have no matching restriction
+ * clauses.
+ *
+ * 'rel' is the relation for which we want to generate index paths
+ * 'clauses' is the current list of clauses (RestrictInfo nodes)
+ * 'outer_clauses' is the list of additional upper-level clauses
+ * 'istoplevel' is true if clauses are the rel's top-level restriction list
+ * 'isjoininner' is true if forming an inner indexscan (so some of the
+ * given clauses are join clauses)
+ * 'outer_relids' identifies the outer side of the join (pass NULL
+ * if not isjoininner)
+ *
+ * Note: check_partial_indexes() must have been run previously.
+ *----------
+ */
+static List *
+find_usable_indexes(Query *root, RelOptInfo *rel,
+ List *clauses, List *outer_clauses,
+ bool istoplevel, bool isjoininner,
+ Relids outer_relids)
+{
+ List *result = NIL;
+ List *all_clauses = NIL; /* not computed till needed */
ListCell *ilist;
foreach(ilist, rel->indexlist)
{
IndexOptInfo *index = (IndexOptInfo *) lfirst(ilist);
+ IndexPath *ipath;
List *restrictclauses;
List *index_pathkeys;
List *useful_pathkeys;
bool index_is_ordered;
- Relids join_outerrelids;
- /* Ignore partial indexes that do not match the query */
+ /*
+ * Ignore partial indexes that do not match the query. If a partial
+ * index is marked predOK then we know it's OK; otherwise, if we
+ * are at top level we know it's not OK (since predOK is exactly
+ * whether its predicate could be proven from the toplevel clauses).
+ * Otherwise, we have to test whether the added clauses are
+ * sufficient to imply the predicate. If so, we could use
+ * the index in the current context.
+ */
if (index->indpred != NIL && !index->predOK)
- continue;
+ {
+ if (istoplevel)
+ continue; /* no point in trying to prove it */
+
+ /* Form all_clauses if not done already */
+ if (all_clauses == NIL)
+ all_clauses = list_concat(list_copy(clauses),
+ outer_clauses);
+
+ if (!pred_test(index->indpred, all_clauses) ||
+ pred_test(index->indpred, outer_clauses))
+ continue;
+ }
/*
- * 1. Match the index against non-OR restriction clauses. (OR
- * clauses will be considered later by orindxpath.c.)
+ * 1. Match the index against the available restriction clauses.
*/
- restrictclauses = group_clauses_by_indexkey(rel, index);
+ restrictclauses = group_clauses_by_indexkey(index,
+ clauses,
+ outer_clauses,
+ outer_relids);
/*
* 2. Compute pathkeys describing index's ordering, if any, then
- * see how many of them are actually useful for this query.
+ * see how many of them are actually useful for this query. This
+ * is not relevant unless we are at top level.
*/
- index_pathkeys = build_index_pathkeys(root, rel, index,
- ForwardScanDirection);
- index_is_ordered = (index_pathkeys != NIL);
- useful_pathkeys = truncate_useless_pathkeys(root, rel,
- index_pathkeys);
+ index_is_ordered = OidIsValid(index->ordering[0]);
+ if (istoplevel && index_is_ordered && !isjoininner)
+ {
+ index_pathkeys = build_index_pathkeys(root, index,
+ ForwardScanDirection);
+ useful_pathkeys = truncate_useless_pathkeys(root, rel,
+ index_pathkeys);
+ }
+ else
+ useful_pathkeys = NIL;
/*
* 3. Generate an indexscan path if there are relevant restriction
* If there is a predicate, consider it anyway since the index
* predicate has already been found to match the query. The
* selectivity of the predicate might alone make the index useful.
+ *
+ * Note: not all index AMs support scans with no restriction clauses.
+ * We assume here that the AM does so if and only if it supports
+ * ordered scans. (It would probably be better if there were a
+ * specific flag for this in pg_am, but there's not.)
*/
if (restrictclauses != NIL ||
useful_pathkeys != NIL ||
- index->indpred != NIL)
- add_path(rel, (Path *)
- create_index_path(root, rel, index,
- restrictclauses,
- useful_pathkeys,
- index_is_ordered ?
- ForwardScanDirection :
- NoMovementScanDirection));
+ (index->indpred != NIL && index_is_ordered))
+ {
+ ipath = create_index_path(root, index,
+ restrictclauses,
+ useful_pathkeys,
+ index_is_ordered ?
+ ForwardScanDirection :
+ NoMovementScanDirection,
+ isjoininner);
+ result = lappend(result, ipath);
+ }
/*
* 4. If the index is ordered, a backwards scan might be
* interesting. Currently this is only possible for a DESC query
* result ordering.
*/
- if (index_is_ordered)
+ if (istoplevel && index_is_ordered && !isjoininner)
{
- index_pathkeys = build_index_pathkeys(root, rel, index,
+ index_pathkeys = build_index_pathkeys(root, index,
BackwardScanDirection);
useful_pathkeys = truncate_useless_pathkeys(root, rel,
index_pathkeys);
if (useful_pathkeys != NIL)
- add_path(rel, (Path *)
- create_index_path(root, rel, index,
- restrictclauses,
- useful_pathkeys,
- BackwardScanDirection));
+ {
+ ipath = create_index_path(root, index,
+ restrictclauses,
+ useful_pathkeys,
+ BackwardScanDirection,
+ false);
+ result = lappend(result, ipath);
+ }
}
-
- /*
- * 5. Examine join clauses to see which ones are potentially
- * usable with this index, and generate the set of all other
- * relids that participate in such join clauses. We'll use this
- * set later to recognize outer rels that are equivalent for
- * joining purposes. We compute both per-index and
- * overall-for-relation sets.
- */
- join_outerrelids = indexable_outerrelids(rel, index);
- index->outer_relids = join_outerrelids;
- all_join_outerrelids = bms_add_members(all_join_outerrelids,
- join_outerrelids);
}
- rel->index_outer_relids = all_join_outerrelids;
+ return result;
}
-/****************************************************************************
- * ---- ROUTINES TO CHECK RESTRICTIONS ----
- ****************************************************************************/
-
-
/*
- * group_clauses_by_indexkey
- * Find restriction clauses that can be used with an index.
- *
- * 'rel' is the node of the relation itself.
- * 'index' is a index on 'rel'.
+ * generate_bitmap_or_paths
+ * Look through the list of clauses to find OR clauses, and generate
+ * a BitmapOrPath for each one we can handle that way. Return a list
+ * of the generated BitmapOrPaths.
*
- * Returns a list of sublists of RestrictInfo nodes for clauses that can be
- * used with this index. Each sublist contains clauses that can be used
- * with one index key (in no particular order); the top list is ordered by
- * index key. (This is depended on by expand_indexqual_conditions().)
- *
- * Note that in a multi-key index, we stop if we find a key that cannot be
- * used with any clause. For example, given an index on (A,B,C), we might
- * return ((C1 C2) (C3 C4)) if we find that clauses C1 and C2 use column A,
- * clauses C3 and C4 use column B, and no clauses use column C. But if
- * no clauses match B we will return ((C1 C2)), whether or not there are
- * clauses matching column C, because the executor couldn't use them anyway.
- * Therefore, there are no empty sublists in the result.
+ * outer_clauses is a list of additional clauses that can be assumed true
+ * for the purpose of generating indexquals, but are not to be searched for
+ * ORs. (See find_usable_indexes() for motivation.)
*/
-static List *
-group_clauses_by_indexkey(RelOptInfo *rel, IndexOptInfo *index)
+List *
+generate_bitmap_or_paths(Query *root, RelOptInfo *rel,
+ List *clauses, List *outer_clauses,
+ bool isjoininner,
+ Relids outer_relids)
{
- List *clausegroup_list = NIL;
- List *restrictinfo_list = rel->baserestrictinfo;
- int indexcol = 0;
- Oid *classes = index->classlist;
+ List *result = NIL;
+ List *all_clauses;
+ ListCell *l;
- if (restrictinfo_list == NIL)
- return NIL;
+ /*
+ * We can use both the current and outer clauses as context for
+ * find_usable_indexes
+ */
+ all_clauses = list_concat(list_copy(clauses), outer_clauses);
- do
+ foreach(l, clauses)
{
- Oid curClass = classes[0];
- List *clausegroup = NIL;
- ListCell *l;
+ RestrictInfo *rinfo = (RestrictInfo *) lfirst(l);
+ List *pathlist;
+ Path *bitmapqual;
+ ListCell *j;
+
+ Assert(IsA(rinfo, RestrictInfo));
+ /* Ignore RestrictInfos that aren't ORs */
+ if (!restriction_is_or_clause(rinfo))
+ continue;
- foreach(l, restrictinfo_list)
+ /*
+ * We must be able to match at least one index to each of the arms
+ * of the OR, else we can't use it.
+ */
+ pathlist = NIL;
+ foreach(j, ((BoolExpr *) rinfo->orclause)->args)
{
- RestrictInfo *rinfo = (RestrictInfo *) lfirst(l);
+ Node *orarg = (Node *) lfirst(j);
+ List *indlist;
- if (match_clause_to_indexcol(rel,
- index,
- indexcol,
- curClass,
- rinfo))
- clausegroup = lappend(clausegroup, rinfo);
+ /* OR arguments should be ANDs or sub-RestrictInfos */
+ if (and_clause(orarg))
+ {
+ List *andargs = ((BoolExpr *) orarg)->args;
+
+ indlist = find_usable_indexes(root, rel,
+ andargs,
+ all_clauses,
+ false,
+ isjoininner,
+ outer_relids);
+ /* Recurse in case there are sub-ORs */
+ indlist = list_concat(indlist,
+ generate_bitmap_or_paths(root, rel,
+ andargs,
+ all_clauses,
+ isjoininner,
+ outer_relids));
+ }
+ else
+ {
+ Assert(IsA(orarg, RestrictInfo));
+ Assert(!restriction_is_or_clause((RestrictInfo *) orarg));
+ indlist = find_usable_indexes(root, rel,
+ list_make1(orarg),
+ all_clauses,
+ false,
+ isjoininner,
+ outer_relids);
+ }
+ /*
+ * If nothing matched this arm, we can't do anything
+ * with this OR clause.
+ */
+ if (indlist == NIL)
+ {
+ pathlist = NIL;
+ break;
+ }
+ /*
+ * OK, pick the most promising AND combination,
+ * and add it to pathlist.
+ */
+ bitmapqual = choose_bitmap_and(root, rel, indlist);
+ pathlist = lappend(pathlist, bitmapqual);
}
-
/*
- * If no clauses match this key, we're done; we don't want to look
- * at keys to its right.
+ * If we have a match for every arm, then turn them
+ * into a BitmapOrPath, and add to result list.
*/
- if (clausegroup == NIL)
- break;
-
- clausegroup_list = lappend(clausegroup_list, clausegroup);
-
- indexcol++;
- classes++;
-
- } while (!DoneMatchingIndexKeys(classes));
+ if (pathlist != NIL)
+ {
+ bitmapqual = (Path *) create_bitmap_or_path(root, rel, pathlist);
+ result = lappend(result, bitmapqual);
+ }
+ }
- return clausegroup_list;
+ return result;
}
+
/*
- * group_clauses_by_indexkey_for_join
- * Generate a list of sublists of clauses that can be used with an index
- * to scan the inner side of a nestloop join.
- *
- * This is much like group_clauses_by_indexkey(), but we consider both
- * join and restriction clauses. Any joinclause that uses only otherrels
- * in the specified outer_relids is fair game. But there must be at least
- * one such joinclause in the final list, otherwise we return NIL indicating
- * that this index isn't interesting as an inner indexscan. (A scan using
- * only restriction clauses shouldn't be created here, because a regular Path
- * will already have been generated for it.)
+ * choose_bitmap_and
+ * Given a nonempty list of bitmap paths, AND them into one path.
+ *
+ * This is a nontrivial decision since we can legally use any subset of the
+ * given path set. We want to choose a good tradeoff between selectivity
+ * and cost of computing the bitmap.
+ *
+ * The result is either a single one of the inputs, or a BitmapAndPath
+ * combining multiple inputs.
*/
-static List *
-group_clauses_by_indexkey_for_join(Query *root,
- RelOptInfo *rel, IndexOptInfo *index,
- Relids outer_relids,
- JoinType jointype, bool isouterjoin)
+static Path *
+choose_bitmap_and(Query *root, RelOptInfo *rel, List *paths)
{
- List *clausegroup_list = NIL;
- bool jfound = false;
- int indexcol = 0;
- Oid *classes = index->classlist;
+ int npaths = list_length(paths);
+ Path **patharray;
+ Cost costsofar;
+ List *qualsofar;
+ ListCell *lastcell;
+ int i;
+ ListCell *l;
- do
- {
- Oid curClass = classes[0];
- List *clausegroup = NIL;
- int numsources;
- ListCell *l;
+ Assert(npaths > 0); /* else caller error */
+ if (npaths == 1)
+ return (Path *) linitial(paths); /* easy case */
- /*
- * We can always use plain restriction clauses for the rel. We
- * scan these first because we want them first in the clausegroup
- * list for the convenience of remove_redundant_join_clauses,
- * which can never remove non-join clauses and hence won't be able
- * to get rid of a non-join clause if it appears after a join
- * clause it is redundant with.
- */
- foreach(l, rel->baserestrictinfo)
- {
- RestrictInfo *rinfo = (RestrictInfo *) lfirst(l);
+ /*
+ * In theory we should consider every nonempty subset of the given paths.
+ * In practice that seems like overkill, given the crude nature of the
+ * estimates, not to mention the possible effects of higher-level AND and
+ * OR clauses. As a compromise, we sort the paths by selectivity.
+ * We always take the first, and sequentially add on paths that result
+ * in a lower estimated cost.
+ *
+ * We also make some effort to detect directly redundant input paths,
+ * as can happen if there are multiple possibly usable indexes. For
+ * this we look only at plain IndexPath inputs, not at sub-OR clauses.
+ * And we consider an index redundant if all its index conditions were
+ * already used by earlier indexes. (We could use pred_test() to have
+ * a more intelligent, but much more expensive, check --- but in most
+ * cases simple pointer equality should suffice, since after all the
+ * index conditions are all coming from the same RestrictInfo lists.)
+ *
+ * XXX is there any risk of throwing away a useful partial index here
+ * because we don't explicitly look at indpred? At least in simple
+ * cases, the partial index will sort before competing non-partial
+ * indexes and so it makes the right choice, but perhaps we need to
+ * work harder.
+ */
- /* Can't use pushed-down clauses in outer join */
- if (isouterjoin && rinfo->is_pushed_down)
- continue;
+ /* Convert list to array so we can apply qsort */
+ patharray = (Path **) palloc(npaths * sizeof(Path *));
+ i = 0;
+ foreach(l, paths)
+ {
+ patharray[i++] = (Path *) lfirst(l);
+ }
+ qsort(patharray, npaths, sizeof(Path *), bitmap_path_comparator);
- if (match_clause_to_indexcol(rel,
- index,
- indexcol,
- curClass,
- rinfo))
- clausegroup = lappend(clausegroup, rinfo);
- }
+ paths = list_make1(patharray[0]);
+ costsofar = bitmap_and_cost_est(root, rel, paths);
+ if (IsA(patharray[0], IndexPath))
+ qualsofar = list_copy(((IndexPath *) patharray[0])->indexclauses);
+ else
+ qualsofar = NIL;
+ lastcell = list_head(paths); /* for quick deletions */
- /* found anything in base restrict list? */
- numsources = (clausegroup != NIL) ? 1 : 0;
+ for (i = 1; i < npaths; i++)
+ {
+ Path *newpath = patharray[i];
+ List *newqual = NIL;
+ Cost newcost;
- /* Look for joinclauses that are usable with given outer_relids */
- foreach(l, rel->joininfo)
+ if (IsA(newpath, IndexPath))
{
- JoinInfo *joininfo = (JoinInfo *) lfirst(l);
- bool jfoundhere = false;
- ListCell *j;
-
- if (!bms_is_subset(joininfo->unjoined_relids, outer_relids))
- continue;
-
- foreach(j, joininfo->jinfo_restrictinfo)
- {
- RestrictInfo *rinfo = (RestrictInfo *) lfirst(j);
-
- /* Can't use pushed-down clauses in outer join */
- if (isouterjoin && rinfo->is_pushed_down)
- continue;
-
- if (match_join_clause_to_indexcol(rel,
- index,
- indexcol,
- curClass,
- rinfo))
- {
- clausegroup = lappend(clausegroup, rinfo);
- if (!jfoundhere)
- {
- jfoundhere = true;
- jfound = true;
- numsources++;
- }
- }
- }
+ newqual = ((IndexPath *) newpath)->indexclauses;
+ if (list_difference_ptr(newqual, qualsofar) == NIL)
+ continue; /* redundant */
}
- /*
- * If we found clauses in more than one list, we may now have
- * clauses that are known redundant. Get rid of 'em.
- */
- if (numsources > 1)
+ paths = lappend(paths, newpath);
+ newcost = bitmap_and_cost_est(root, rel, paths);
+ if (newcost < costsofar)
{
- clausegroup = remove_redundant_join_clauses(root,
- clausegroup,
- jointype);
+ costsofar = newcost;
+ if (newqual)
+ qualsofar = list_concat(qualsofar, list_copy(newqual));
+ lastcell = lnext(lastcell);
}
+ else
+ {
+ paths = list_delete_cell(paths, lnext(lastcell), lastcell);
+ }
+ Assert(lnext(lastcell) == NULL);
+ }
- /*
- * If no clauses match this key, we're done; we don't want to look
- * at keys to its right.
- */
- if (clausegroup == NIL)
- break;
+ if (list_length(paths) == 1)
+ return (Path *) linitial(paths); /* no need for AND */
+ return (Path *) create_bitmap_and_path(root, rel, paths);
+}
- clausegroup_list = lappend(clausegroup_list, clausegroup);
+/* qsort comparator to sort in increasing selectivity order */
+static int
+bitmap_path_comparator(const void *a, const void *b)
+{
+ Path *pa = *(Path * const *) a;
+ Path *pb = *(Path * const *) b;
+ Cost acost;
+ Cost bcost;
+ Selectivity aselec;
+ Selectivity bselec;
+
+ cost_bitmap_tree_node(pa, &acost, &aselec);
+ cost_bitmap_tree_node(pb, &bcost, &bselec);
+
+ if (aselec < bselec)
+ return -1;
+ if (aselec > bselec)
+ return 1;
+ /* if identical selectivity, sort by cost */
+ if (acost < bcost)
+ return -1;
+ if (acost > bcost)
+ return 1;
+ return 0;
+}
- indexcol++;
- classes++;
+/*
+ * Estimate the cost of actually executing a BitmapAnd with the given
+ * inputs.
+ */
+static Cost
+bitmap_and_cost_est(Query *root, RelOptInfo *rel, List *paths)
+{
+ BitmapAndPath apath;
+ Path bpath;
- } while (!DoneMatchingIndexKeys(classes));
+ /* Set up a dummy BitmapAndPath */
+ apath.path.type = T_BitmapAndPath;
+ apath.path.parent = rel;
+ apath.bitmapquals = paths;
+ cost_bitmap_and_node(&apath, root);
- /* if no join clause was matched then forget it, per comments above */
- if (!jfound)
- return NIL;
+ /* Now we can do cost_bitmap_heap_scan */
+ cost_bitmap_heap_scan(&bpath, root, rel, (Path *) &apath, false);
- return clausegroup_list;
+ return bpath.total_cost;
}
+/****************************************************************************
+ * ---- ROUTINES TO CHECK RESTRICTIONS ----
+ ****************************************************************************/
+
+
/*
- * group_clauses_by_indexkey_for_or
- * Generate a list of sublists of clauses that can be used with an index
- * to find rows matching an OR subclause.
- *
- * This is essentially just like group_clauses_by_indexkey() except that
- * we can use the given clause (or any AND subclauses of it) as well as
- * top-level restriction clauses of the relation. Furthermore, we demand
- * that at least one such use be made, otherwise we fail and return NIL.
- * (Any path we made without such a use would be redundant with non-OR
- * indexscans. Compare also group_clauses_by_indexkey_for_join.)
- *
- * XXX When we generate an indexqual list that uses both the OR subclause
- * and top-level restriction clauses, we end up with a slightly inefficient
- * plan because create_indexscan_plan is not very bright about figuring out
- * which restriction clauses are implied by the generated indexqual condition.
- * Currently we'll end up rechecking both the OR clause and the top-level
- * restriction clause as qpquals. FIXME someday.
+ * group_clauses_by_indexkey
+ * Find restriction clauses that can be used with an index.
+ *
+ * As explained in the comments for find_usable_indexes(), we can use
+ * clauses from either of the given lists, but the result is required to
+ * use at least one clause from the "current clauses" list. We return
+ * NIL if we don't find any such clause.
+ *
+ * outer_relids determines what Vars will be allowed on the other side
+ * of a possible index qual; see match_clause_to_indexcol().
+ *
+ * Returns a list of sublists of RestrictInfo nodes for clauses that can be
+ * used with this index. Each sublist contains clauses that can be used
+ * with one index key (in no particular order); the top list is ordered by
+ * index key. (This is depended on by expand_indexqual_conditions().)
+ *
+ * Note that in a multi-key index, we stop if we find a key that cannot be
+ * used with any clause. For example, given an index on (A,B,C), we might
+ * return ((C1 C2) (C3 C4)) if we find that clauses C1 and C2 use column A,
+ * clauses C3 and C4 use column B, and no clauses use column C. But if
+ * no clauses match B we will return ((C1 C2)), whether or not there are
+ * clauses matching column C, because the executor couldn't use them anyway.
+ * Therefore, there are no empty sublists in the result.
*/
List *
-group_clauses_by_indexkey_for_or(RelOptInfo *rel,
- IndexOptInfo *index,
- Expr *orsubclause)
+group_clauses_by_indexkey(IndexOptInfo *index,
+ List *clauses, List *outer_clauses,
+ Relids outer_relids)
{
List *clausegroup_list = NIL;
- bool matched = false;
+ bool found_clause = false;
int indexcol = 0;
Oid *classes = index->classlist;
+ if (clauses == NIL)
+ return NIL; /* cannot succeed */
+
do
{
Oid curClass = classes[0];
List *clausegroup = NIL;
- ListCell *item;
+ ListCell *l;
- /* Try to match the OR subclause to the index key */
- if (IsA(orsubclause, RestrictInfo))
- {
- if (match_clause_to_indexcol(rel, index,
- indexcol, curClass,
- (RestrictInfo *) orsubclause))
- {
- clausegroup = lappend(clausegroup, orsubclause);
- matched = true;
- }
- }
- else if (and_clause((Node *) orsubclause))
+ /* check the current clauses */
+ foreach(l, clauses)
{
- foreach(item, ((BoolExpr *) orsubclause)->args)
- {
- RestrictInfo *subsubclause = (RestrictInfo *) lfirst(item);
+ RestrictInfo *rinfo = (RestrictInfo *) lfirst(l);
- if (IsA(subsubclause, RestrictInfo) &&
- match_clause_to_indexcol(rel, index,
- indexcol, curClass,
- subsubclause))
- {
- clausegroup = lappend(clausegroup, subsubclause);
- matched = true;
- }
+ Assert(IsA(rinfo, RestrictInfo));
+ if (match_clause_to_indexcol(index,
+ indexcol,
+ curClass,
+ rinfo,
+ outer_relids))
+ {
+ clausegroup = lappend(clausegroup, rinfo);
+ found_clause = true;
}
}
- /*
- * If we found no clauses for this indexkey in the OR subclause
- * itself, try looking in the rel's top-level restriction list.
- *
- * XXX should we always search the top-level list? Slower but could
- * sometimes yield a better plan.
- */
- if (clausegroup == NIL)
+ /* check the outer clauses */
+ foreach(l, outer_clauses)
{
- foreach(item, rel->baserestrictinfo)
- {
- RestrictInfo *rinfo = (RestrictInfo *) lfirst(item);
+ RestrictInfo *rinfo = (RestrictInfo *) lfirst(l);
- if (match_clause_to_indexcol(rel, index,
- indexcol, curClass,
- rinfo))
- clausegroup = lappend(clausegroup, rinfo);
- }
+ Assert(IsA(rinfo, RestrictInfo));
+ if (match_clause_to_indexcol(index,
+ indexcol,
+ curClass,
+ rinfo,
+ outer_relids))
+ clausegroup = lappend(clausegroup, rinfo);
}
/*
- * If still no clauses match this key, we're done; we don't want
- * to look at keys to its right.
+ * If no clauses match this key, we're done; we don't want to look
+ * at keys to its right.
*/
if (clausegroup == NIL)
break;
indexcol++;
classes++;
+
} while (!DoneMatchingIndexKeys(classes));
- /* if OR clause was not used then forget it, per comments above */
- if (!matched)
+ if (!found_clause)
return NIL;
return clausegroup_list;
* match_clause_to_indexcol()
* Determines whether a restriction clause matches a column of an index.
*
- * To match, the clause:
+ * To match a normal index, the clause:
*
* (1) must be in the form (indexkey op const) or (const op indexkey);
* and
* operator for this column, or is a "special" operator as recognized
* by match_special_index_operator().
*
+ * Our definition of "const" is pretty liberal: we allow Vars belonging
+ * to the caller-specified outer_relids relations (which had better not
+ * include the relation whose index is being tested). outer_relids should
+ * be NULL when checking simple restriction clauses, and the outer side
+ * of the join when building a join inner scan. Other than that, the
+ * only thing we don't like is volatile functions.
+ *
+ * Note: in most cases we already know that the clause as a whole uses
+ * vars from the interesting set of relations. The reason for the
+ * outer_relids test is to reject clauses like (a.f1 OP (b.f2 OP a.f3));
+ * that's not processable by an indexscan nestloop join on A, whereas
+ * (a.f1 OP (b.f2 OP c.f3)) is.
+ *
* Presently, the executor can only deal with indexquals that have the
* indexkey on the left, so we can only use clauses that have the indexkey
* on the right if we can commute the clause to put the key on the left.
* We do not actually do the commuting here, but we check whether a
* suitable commutator operator is available.
*
- * 'rel' is the relation of interest.
- * 'index' is an index on 'rel'.
+ * For boolean indexes, it is also possible to match the clause directly
+ * to the indexkey; or perhaps the clause is (NOT indexkey).
+ *
+ * 'index' is the index of interest.
* 'indexcol' is a column number of 'index' (counting from 0).
* 'opclass' is the corresponding operator class.
* 'rinfo' is the clause to be tested (as a RestrictInfo node).
* responsibility of higher-level routines to cope with those.
*/
static bool
-match_clause_to_indexcol(RelOptInfo *rel,
- IndexOptInfo *index,
+match_clause_to_indexcol(IndexOptInfo *index,
int indexcol,
Oid opclass,
- RestrictInfo *rinfo)
+ RestrictInfo *rinfo,
+ Relids outer_relids)
{
Expr *clause = rinfo->clause;
Node *leftop,
*rightop;
- /* Clause must be a binary opclause. */
+ /* First check for boolean-index cases. */
+ if (IsBooleanOpclass(opclass))
+ {
+ if (match_boolean_index_clause((Node *) clause, indexcol, index))
+ return true;
+ }
+
+ /* Else clause must be a binary opclause. */
if (!is_opclause(clause))
return false;
leftop = get_leftop(clause);
/*
* Check for clauses of the form: (indexkey operator constant) or
- * (constant operator indexkey). Anything that is a "pseudo constant"
- * expression will do.
+ * (constant operator indexkey). See above notes about const-ness.
*/
- if (match_index_to_operand(leftop, indexcol, rel, index) &&
- is_pseudo_constant_clause_relids(rightop, rinfo->right_relids))
+ if (match_index_to_operand(leftop, indexcol, index) &&
+ bms_is_subset(rinfo->right_relids, outer_relids) &&
+ !contain_volatile_functions(rightop))
{
if (is_indexable_operator(clause, opclass, true))
return true;
return false;
}
- if (match_index_to_operand(rightop, indexcol, rel, index) &&
- is_pseudo_constant_clause_relids(leftop, rinfo->left_relids))
+ if (match_index_to_operand(rightop, indexcol, index) &&
+ bms_is_subset(rinfo->left_relids, outer_relids) &&
+ !contain_volatile_functions(leftop))
{
if (is_indexable_operator(clause, opclass, false))
return true;
return false;
}
-/*
- * match_join_clause_to_indexcol()
- * Determines whether a join clause matches a column of an index.
- *
- * To match, the clause:
- *
- * (1) must be in the form (indexkey op others) or (others op indexkey),
- * where others is an expression involving only vars of the other
- * relation(s); and
- * (2) must contain an operator which is in the same class as the index
- * operator for this column, or is a "special" operator as recognized
- * by match_special_index_operator().
- *
- * As above, we must be able to commute the clause to put the indexkey
- * on the left.
- *
- * Note that we already know that the clause as a whole uses vars from
- * the interesting set of relations. But we need to defend against
- * expressions like (a.f1 OP (b.f2 OP a.f3)); that's not processable by
- * an indexscan nestloop join, whereas (a.f1 OP (b.f2 OP c.f3)) is.
- *
- * 'rel' is the relation of interest.
- * 'index' is an index on 'rel'.
- * 'indexcol' is a column number of 'index' (counting from 0).
- * 'opclass' is the corresponding operator class.
- * 'rinfo' is the clause to be tested (as a RestrictInfo node).
- *
- * Returns true if the clause can be used with this index key.
- *
- * NOTE: returns false if clause is an OR or AND clause; it is the
- * responsibility of higher-level routines to cope with those.
- */
-static bool
-match_join_clause_to_indexcol(RelOptInfo *rel,
- IndexOptInfo *index,
- int indexcol,
- Oid opclass,
- RestrictInfo *rinfo)
-{
- Expr *clause = rinfo->clause;
- Node *leftop,
- *rightop;
-
- /* Clause must be a binary opclause. */
- if (!is_opclause(clause))
- return false;
- leftop = get_leftop(clause);
- rightop = get_rightop(clause);
- if (!leftop || !rightop)
- return false;
-
- /*
- * Check for an indexqual that could be handled by a nestloop join. We
- * need the index key to be compared against an expression that uses
- * none of the indexed relation's vars and contains no volatile
- * functions.
- */
- if (match_index_to_operand(leftop, indexcol, rel, index))
- {
- Relids othervarnos = rinfo->right_relids;
- bool isIndexable;
-
- isIndexable =
- !bms_overlap(rel->relids, othervarnos) &&
- !contain_volatile_functions(rightop) &&
- is_indexable_operator(clause, opclass, true);
- return isIndexable;
- }
-
- if (match_index_to_operand(rightop, indexcol, rel, index))
- {
- Relids othervarnos = rinfo->left_relids;
- bool isIndexable;
-
- isIndexable =
- !bms_overlap(rel->relids, othervarnos) &&
- !contain_volatile_functions(leftop) &&
- is_indexable_operator(clause, opclass, false);
- return isIndexable;
- }
-
- return false;
-}
-
/*
* indexable_operator
* Does a binary opclause contain an operator matching the index opclass?
* Recursively checks whether the clauses in restrictinfo_list imply
* that the given predicate is true.
*
- * This routine (together with the routines it calls) first breaks down
- * the predicate to its constituent AND/OR elements, then similarly
- * breaks down the restriction clauses to AND/OR elements in an effort
- * to prove that each predicate element is implied. The top-level
- * List structure of each list corresponds to an AND list.
+ * The top-level List structure of each list corresponds to an AND list.
+ * We assume that eval_const_expressions() has been applied and so there
+ * are no un-flattened ANDs or ORs (e.g., no AND immediately within an AND,
+ * including AND just below the top-level List structure).
+ * If this is not true we might fail to prove an implication that is
+ * valid, but no worse consequences will ensue.
*/
bool
pred_test(List *predicate_list, List *restrictinfo_list)
{
- ListCell *pred;
+ ListCell *item;
/*
* Note: if Postgres tried to optimize queries by forming equivalence
return false; /* no restriction clauses: the test must
* fail */
- foreach(pred, predicate_list)
+ /*
+ * In all cases where the predicate is an AND-clause, pred_test_recurse()
+ * will prefer to iterate over the predicate's components. So we can
+ * just do that to start with here, and eliminate the need for
+ * pred_test_recurse() to handle a bare List on the predicate side.
+ *
+ * Logic is: restriction must imply each of the AND'ed predicate items.
+ */
+ foreach(item, predicate_list)
{
- /*
- * if any clause is not implied, the whole predicate is not
- * implied.
- */
- if (!pred_test_recurse_pred(lfirst(pred), restrictinfo_list))
+ if (!pred_test_recurse((Node *) restrictinfo_list, lfirst(item)))
return false;
}
return true;
}
-/*
- * pred_test_recurse_pred
- * Does the "predicate inclusion test" for one conjunct of a predicate
- * expression. Here we recursively deal with the possibility that the
- * predicate conjunct is itself an AND or OR structure.
- */
-static bool
-pred_test_recurse_pred(Expr *predicate, List *restrictinfo_list)
-{
- List *items;
- ListCell *item;
-
- Assert(predicate != NULL);
- if (or_clause((Node *) predicate))
- {
- items = ((BoolExpr *) predicate)->args;
- foreach(item, items)
- {
- /* if any item is implied, the whole predicate is implied */
- if (pred_test_recurse_pred(lfirst(item), restrictinfo_list))
- return true;
- }
- return false;
- }
- else if (and_clause((Node *) predicate))
- {
- items = ((BoolExpr *) predicate)->args;
- foreach(item, items)
- {
- /*
- * if any item is not implied, the whole predicate is not
- * implied
- */
- if (!pred_test_recurse_pred(lfirst(item), restrictinfo_list))
- return false;
- }
- return true;
- }
- else
- return pred_test_restrict_list(predicate, restrictinfo_list);
-}
-
-
-/*
- * pred_test_restrict_list
- * Does the "predicate inclusion test" for one element of a predicate
- * expression. Here we take care of the AND semantics of the top-level
- * restrictinfo list.
- */
-static bool
-pred_test_restrict_list(Expr *predicate, List *restrictinfo_list)
-{
- ListCell *item;
-
- foreach(item, restrictinfo_list)
- {
- /* if any clause implies the predicate, return true */
- if (pred_test_recurse_restrict(predicate,
- (Node *) lfirst(item)))
- return true;
- }
- return false;
-}
-
-
-/*
- * pred_test_recurse_restrict
- * Does the "predicate inclusion test" for one element of a predicate
- * expression. Here we recursively deal with the possibility that the
- * restriction-list element is itself an AND or OR structure; also,
- * we strip off RestrictInfo nodes to find bare predicate expressions.
+/*----------
+ * pred_test_recurse
+ * Does the "predicate inclusion test" for non-NULL restriction and
+ * predicate clauses.
+ *
+ * The logic followed here is ("=>" means "implies"):
+ * atom A => atom B iff: pred_test_simple_clause says so
+ * atom A => AND-expr B iff: A => each of B's components
+ * atom A => OR-expr B iff: A => any of B's components
+ * AND-expr A => atom B iff: any of A's components => B
+ * AND-expr A => AND-expr B iff: A => each of B's components
+ * AND-expr A => OR-expr B iff: A => any of B's components,
+ * *or* any of A's components => B
+ * OR-expr A => atom B iff: each of A's components => B
+ * OR-expr A => AND-expr B iff: A => each of B's components
+ * OR-expr A => OR-expr B iff: each of A's components => any of B's
+ *
+ * An "atom" is anything other than an AND or OR node. Notice that we don't
+ * have any special logic to handle NOT nodes; these should have been pushed
+ * down or eliminated where feasible by prepqual.c.
+ *
+ * We can't recursively expand either side first, but have to interleave
+ * the expansions per the above rules, to be sure we handle all of these
+ * examples:
+ * (x OR y) => (x OR y OR z)
+ * (x AND y AND z) => (x AND y)
+ * (x AND y) => ((x AND y) OR z)
+ * ((x OR y) AND z) => (x OR y)
+ * This is still not an exhaustive test, but it handles most normal cases
+ * under the assumption that both inputs have been AND/OR flattened.
+ *
+ * A bare List node on the restriction side is interpreted as an AND clause,
+ * in order to handle the top-level restriction List properly. However we
+ * need not consider a List on the predicate side since pred_test() already
+ * expanded it.
+ *
+ * We have to be prepared to handle RestrictInfo nodes in the restrictinfo
+ * tree, though not in the predicate tree.
+ *----------
*/
static bool
-pred_test_recurse_restrict(Expr *predicate, Node *clause)
+pred_test_recurse(Node *clause, Node *predicate)
{
- List *items;
ListCell *item;
Assert(clause != NULL);
+ /* skip through RestrictInfo */
if (IsA(clause, RestrictInfo))
{
- RestrictInfo *restrictinfo = (RestrictInfo *) clause;
-
- return pred_test_recurse_restrict(predicate,
- (Node *) restrictinfo->clause);
+ clause = (Node *) ((RestrictInfo *) clause)->clause;
+ Assert(clause != NULL);
+ Assert(!IsA(clause, RestrictInfo));
}
- else if (or_clause(clause))
+ Assert(predicate != NULL);
+
+ /*
+ * Since a restriction List clause is handled the same as an AND clause,
+ * we can avoid duplicate code like this:
+ */
+ if (and_clause(clause))
+ clause = (Node *) ((BoolExpr *) clause)->args;
+
+ if (IsA(clause, List))
{
- items = ((BoolExpr *) clause)->args;
- foreach(item, items)
+ if (and_clause(predicate))
{
- /* if any OR item doesn't imply the predicate, clause doesn't */
- if (!pred_test_recurse_restrict(predicate, lfirst(item)))
- return false;
+ /* AND-clause => AND-clause if A implies each of B's items */
+ foreach(item, ((BoolExpr *) predicate)->args)
+ {
+ if (!pred_test_recurse(clause, lfirst(item)))
+ return false;
+ }
+ return true;
+ }
+ else if (or_clause(predicate))
+ {
+ /* AND-clause => OR-clause if A implies any of B's items */
+ /* Needed to handle (x AND y) => ((x AND y) OR z) */
+ foreach(item, ((BoolExpr *) predicate)->args)
+ {
+ if (pred_test_recurse(clause, lfirst(item)))
+ return true;
+ }
+ /* Also check if any of A's items implies B */
+ /* Needed to handle ((x OR y) AND z) => (x OR y) */
+ foreach(item, (List *) clause)
+ {
+ if (pred_test_recurse(lfirst(item), predicate))
+ return true;
+ }
+ return false;
+ }
+ else
+ {
+ /* AND-clause => atom if any of A's items implies B */
+ foreach(item, (List *) clause)
+ {
+ if (pred_test_recurse(lfirst(item), predicate))
+ return true;
+ }
+ return false;
}
- return true;
}
- else if (and_clause(clause))
+ else if (or_clause(clause))
{
- items = ((BoolExpr *) clause)->args;
- foreach(item, items)
+ if (or_clause(predicate))
{
/*
- * if any AND item implies the predicate, the whole clause
- * does
+ * OR-clause => OR-clause if each of A's items implies any of
+ * B's items. Messy but can't do it any more simply.
*/
- if (pred_test_recurse_restrict(predicate, lfirst(item)))
- return true;
+ foreach(item, ((BoolExpr *) clause)->args)
+ {
+ Node *citem = lfirst(item);
+ ListCell *item2;
+
+ foreach(item2, ((BoolExpr *) predicate)->args)
+ {
+ if (pred_test_recurse(citem, lfirst(item2)))
+ break;
+ }
+ if (item2 == NULL)
+ return false; /* doesn't imply any of B's */
+ }
+ return true;
+ }
+ else
+ {
+ /* OR-clause => AND-clause if each of A's items implies B */
+ /* OR-clause => atom if each of A's items implies B */
+ foreach(item, ((BoolExpr *) clause)->args)
+ {
+ if (!pred_test_recurse(lfirst(item), predicate))
+ return false;
+ }
+ return true;
}
- return false;
}
else
- return pred_test_simple_clause(predicate, clause);
+ {
+ if (and_clause(predicate))
+ {
+ /* atom => AND-clause if A implies each of B's items */
+ foreach(item, ((BoolExpr *) predicate)->args)
+ {
+ if (!pred_test_recurse(clause, lfirst(item)))
+ return false;
+ }
+ return true;
+ }
+ else if (or_clause(predicate))
+ {
+ /* atom => OR-clause if A implies any of B's items */
+ foreach(item, ((BoolExpr *) predicate)->args)
+ {
+ if (pred_test_recurse(clause, lfirst(item)))
+ return true;
+ }
+ return false;
+ }
+ else
+ {
+ /* atom => atom is the base case */
+ return pred_test_simple_clause((Expr *) predicate, clause);
+ }
+ }
}
/*
* indexable_outerrelids
* Finds all other relids that participate in any indexable join clause
- * for the specified index. Returns a set of relids.
- *
- * 'rel' is the relation for which 'index' is defined
+ * for the specified table. Returns a set of relids.
*/
static Relids
-indexable_outerrelids(RelOptInfo *rel, IndexOptInfo *index)
+indexable_outerrelids(RelOptInfo *rel)
{
Relids outer_relids = NULL;
ListCell *l;
foreach(l, rel->joininfo)
{
JoinInfo *joininfo = (JoinInfo *) lfirst(l);
- bool match_found = false;
- ListCell *j;
/*
* Examine each joinclause in the JoinInfo node's list to see if
- * it matches any key of the index. If so, add the JoinInfo's
+ * it matches any key of any index. If so, add the JoinInfo's
* otherrels to the result. We can skip examining other
- * joinclauses in the same list as soon as we find a match (since
- * by definition they all have the same otherrels).
+ * joinclauses in the same list as soon as we find a match, since
+ * by definition they all have the same otherrels.
*/
- foreach(j, joininfo->jinfo_restrictinfo)
- {
- RestrictInfo *rinfo = (RestrictInfo *) lfirst(j);
- int indexcol = 0;
- Oid *classes = index->classlist;
+ if (list_matches_any_index(joininfo->jinfo_restrictinfo,
+ rel,
+ joininfo->unjoined_relids))
+ outer_relids = bms_add_members(outer_relids,
+ joininfo->unjoined_relids);
+ }
- do
- {
- Oid curClass = classes[0];
+ return outer_relids;
+}
- if (match_join_clause_to_indexcol(rel,
- index,
- indexcol,
- curClass,
- rinfo))
- {
- match_found = true;
- break;
- }
+/*
+ * list_matches_any_index
+ * Workhorse for indexable_outerrelids: given a list of RestrictInfos,
+ * see if any of them match any index of the given rel.
+ *
+ * We define it like this so that we can recurse into OR subclauses.
+ */
+static bool
+list_matches_any_index(List *clauses, RelOptInfo *rel, Relids outer_relids)
+{
+ ListCell *l;
- indexcol++;
- classes++;
+ foreach(l, clauses)
+ {
+ RestrictInfo *rinfo = (RestrictInfo *) lfirst(l);
+ ListCell *j;
- } while (!DoneMatchingIndexKeys(classes));
+ Assert(IsA(rinfo, RestrictInfo));
- if (match_found)
- break;
+ /* RestrictInfos that aren't ORs are easy */
+ if (!restriction_is_or_clause(rinfo))
+ {
+ if (matches_any_index(rinfo, rel, outer_relids))
+ return true;
+ continue;
}
- if (match_found)
+ foreach(j, ((BoolExpr *) rinfo->orclause)->args)
{
- outer_relids = bms_add_members(outer_relids,
- joininfo->unjoined_relids);
+ Node *orarg = (Node *) lfirst(j);
+
+ /* OR arguments should be ANDs or sub-RestrictInfos */
+ if (and_clause(orarg))
+ {
+ List *andargs = ((BoolExpr *) orarg)->args;
+
+ /* Recurse to examine AND items and sub-ORs */
+ if (list_matches_any_index(andargs, rel, outer_relids))
+ return true;
+ }
+ else
+ {
+ Assert(IsA(orarg, RestrictInfo));
+ Assert(!restriction_is_or_clause((RestrictInfo *) orarg));
+ if (matches_any_index((RestrictInfo *) orarg, rel,
+ outer_relids))
+ return true;
+ }
}
}
- return outer_relids;
+ return false;
+}
+
+/*
+ * matches_any_index
+ * Workhorse for indexable_outerrelids: see if a simple joinclause can be
+ * matched to any index of the given rel.
+ */
+static bool
+matches_any_index(RestrictInfo *rinfo, RelOptInfo *rel, Relids outer_relids)
+{
+ ListCell *l;
+
+ /* Normal case for a simple restriction clause */
+ foreach(l, rel->indexlist)
+ {
+ IndexOptInfo *index = (IndexOptInfo *) lfirst(l);
+ int indexcol = 0;
+ Oid *classes = index->classlist;
+
+ do
+ {
+ Oid curClass = classes[0];
+
+ if (match_clause_to_indexcol(index,
+ indexcol,
+ curClass,
+ rinfo,
+ outer_relids))
+ return true;
+
+ indexcol++;
+ classes++;
+ } while (!DoneMatchingIndexKeys(classes));
+ }
+
+ return false;
}
/*
best_inner_indexscan(Query *root, RelOptInfo *rel,
Relids outer_relids, JoinType jointype)
{
- Path *cheapest = NULL;
+ Path *cheapest;
bool isouterjoin;
- ListCell *ilist;
- ListCell *jlist;
+ List *clause_list;
+ List *indexpaths;
+ List *bitindexpaths;
+ ListCell *l;
InnerIndexscanInfo *info;
MemoryContext oldcontext;
/*
* Intersect the given outer_relids with index_outer_relids to find
- * the set of outer relids actually relevant for this index. If there
+ * the set of outer relids actually relevant for this rel. If there
* are none, again we can fail immediately.
*/
outer_relids = bms_intersect(rel->index_outer_relids, outer_relids);
* necessary because it should always be the same for a given
* innerrel.)
*/
- foreach(jlist, rel->index_inner_paths)
+ foreach(l, rel->index_inner_paths)
{
- info = (InnerIndexscanInfo *) lfirst(jlist);
+ info = (InnerIndexscanInfo *) lfirst(l);
if (bms_equal(info->other_relids, outer_relids) &&
info->isouterjoin == isouterjoin)
{
}
/*
- * For each index of the rel, find the best path; then choose the best
- * overall. We cache the per-index results as well as the overall
- * result. (This is useful because different indexes may have
- * different relevant outerrel sets, so different overall outerrel
- * sets might still map to the same computation for a given index.)
+ * Find all the relevant restriction and join clauses.
*/
- foreach(ilist, rel->indexlist)
- {
- IndexOptInfo *index = (IndexOptInfo *) lfirst(ilist);
- Relids index_outer_relids;
- Path *path = NULL;
+ clause_list = find_clauses_for_join(root, rel, outer_relids, isouterjoin);
- /* identify set of relevant outer relids for this index */
- index_outer_relids = bms_intersect(index->outer_relids, outer_relids);
- /* skip if none */
- if (bms_is_empty(index_outer_relids))
- {
- bms_free(index_outer_relids);
- continue;
- }
+ /*
+ * Find all the index paths that are usable for this join, except for
+ * stuff involving OR clauses.
+ */
+ indexpaths = find_usable_indexes(root, rel,
+ clause_list, NIL,
+ false, true,
+ outer_relids);
- /*
- * Look to see if we already computed the result for this index.
- */
- foreach(jlist, index->inner_paths)
- {
- info = (InnerIndexscanInfo *) lfirst(jlist);
- if (bms_equal(info->other_relids, index_outer_relids) &&
- info->isouterjoin == isouterjoin)
- {
- path = info->best_innerpath;
- bms_free(index_outer_relids); /* not needed anymore */
- break;
- }
- }
+ /*
+ * Generate BitmapOrPaths for any suitable OR-clauses present in the
+ * clause list.
+ */
+ bitindexpaths = generate_bitmap_or_paths(root, rel,
+ clause_list, NIL,
+ true,
+ outer_relids);
- if (jlist == NULL) /* failed to find a match? */
- {
- List *clausegroups;
-
- /* find useful clauses for this index and outerjoin set */
- clausegroups = group_clauses_by_indexkey_for_join(root,
- rel,
- index,
- index_outer_relids,
- jointype,
- isouterjoin);
- if (clausegroups)
- {
- /* make the path */
- path = make_innerjoin_index_path(root, rel, index,
- clausegroups);
- }
+ /*
+ * Include the regular index paths in bitindexpaths.
+ */
+ bitindexpaths = list_concat(bitindexpaths, list_copy(indexpaths));
- /* Cache the result --- whether positive or negative */
- info = makeNode(InnerIndexscanInfo);
- info->other_relids = index_outer_relids;
- info->isouterjoin = isouterjoin;
- info->best_innerpath = path;
- index->inner_paths = lcons(info, index->inner_paths);
- }
+ /*
+ * If we found anything usable, generate a BitmapHeapPath for the
+ * most promising combination of bitmap index paths.
+ */
+ if (bitindexpaths != NIL)
+ {
+ Path *bitmapqual;
+ BitmapHeapPath *bpath;
+
+ bitmapqual = choose_bitmap_and(root, rel, bitindexpaths);
+ bpath = create_bitmap_heap_path(root, rel, bitmapqual, true);
+ indexpaths = lappend(indexpaths, bpath);
+ }
+
+ /*
+ * Now choose the cheapest member of indexpaths.
+ */
+ cheapest = NULL;
+ foreach(l, indexpaths)
+ {
+ Path *path = (Path *) lfirst(l);
- if (path != NULL &&
- (cheapest == NULL ||
- compare_path_costs(path, cheapest, TOTAL_COST) < 0))
+ if (cheapest == NULL ||
+ compare_path_costs(path, cheapest, TOTAL_COST) < 0)
cheapest = path;
}
return cheapest;
}
-/****************************************************************************
- * ---- PATH CREATION UTILITIES ----
- ****************************************************************************/
-
/*
- * make_innerjoin_index_path
- * Create an index path node for a path to be used as an inner
- * relation in a nestloop join.
+ * find_clauses_for_join
+ * Generate a list of clauses that are potentially useful for
+ * scanning rel as the inner side of a nestloop join.
*
- * 'rel' is the relation for which 'index' is defined
- * 'clausegroups' is a list of lists of RestrictInfos that can use 'index'
+ * We consider both join and restriction clauses. Any joinclause that uses
+ * only otherrels in the specified outer_relids is fair game. But there must
+ * be at least one such joinclause in the final list, otherwise we return NIL
+ * indicating that there isn't any potential win here.
*/
-static Path *
-make_innerjoin_index_path(Query *root,
- RelOptInfo *rel, IndexOptInfo *index,
- List *clausegroups)
+static List *
+find_clauses_for_join(Query *root, RelOptInfo *rel,
+ Relids outer_relids, bool isouterjoin)
{
- IndexPath *pathnode = makeNode(IndexPath);
- List *indexquals,
- *allclauses;
-
- /* XXX perhaps this code should be merged with create_index_path? */
-
- pathnode->path.pathtype = T_IndexScan;
- pathnode->path.parent = rel;
+ List *clause_list = NIL;
+ bool jfound = false;
+ int numsources;
+ ListCell *l;
/*
- * There's no point in marking the path with any pathkeys, since it
- * will only ever be used as the inner path of a nestloop, and so its
- * ordering does not matter.
+ * We can always use plain restriction clauses for the rel. We
+ * scan these first because we want them first in the clause
+ * list for the convenience of remove_redundant_join_clauses,
+ * which can never remove non-join clauses and hence won't be able
+ * to get rid of a non-join clause if it appears after a join
+ * clause it is redundant with.
*/
- pathnode->path.pathkeys = NIL;
+ foreach(l, rel->baserestrictinfo)
+ {
+ RestrictInfo *rinfo = (RestrictInfo *) lfirst(l);
- /* Convert clauses to indexquals the executor can handle */
- indexquals = expand_indexqual_conditions(index, clausegroups);
+ /* Can't use pushed-down clauses in outer join */
+ if (isouterjoin && rinfo->is_pushed_down)
+ continue;
+ clause_list = lappend(clause_list, rinfo);
+ }
- /* Flatten the clausegroups list to produce indexclauses list */
- allclauses = flatten_clausegroups_list(clausegroups);
+ /* found anything in base restrict list? */
+ numsources = (clause_list != NIL) ? 1 : 0;
- /*
- * Note that we are making a pathnode for a single-scan indexscan;
- * therefore, indexinfo etc should be single-element lists.
- */
- pathnode->indexinfo = list_make1(index);
- pathnode->indexclauses = list_make1(allclauses);
- pathnode->indexquals = list_make1(indexquals);
+ /* Look for joinclauses that are usable with given outer_relids */
+ foreach(l, rel->joininfo)
+ {
+ JoinInfo *joininfo = (JoinInfo *) lfirst(l);
+ bool jfoundhere = false;
+ ListCell *j;
+
+ if (!bms_is_subset(joininfo->unjoined_relids, outer_relids))
+ continue;
+
+ foreach(j, joininfo->jinfo_restrictinfo)
+ {
+ RestrictInfo *rinfo = (RestrictInfo *) lfirst(j);
+
+ /* Can't use pushed-down clauses in outer join */
+ if (isouterjoin && rinfo->is_pushed_down)
+ continue;
- pathnode->isjoininner = true;
+ clause_list = lappend(clause_list, rinfo);
+ if (!jfoundhere)
+ {
+ jfoundhere = true;
+ jfound = true;
+ numsources++;
+ }
+ }
+ }
- /* We don't actually care what order the index scans in ... */
- pathnode->indexscandir = NoMovementScanDirection;
+ /* if no join clause was matched then forget it, per comments above */
+ if (!jfound)
+ return NIL;
/*
- * We must compute the estimated number of output rows for the
- * indexscan. This is less than rel->rows because of the additional
- * selectivity of the join clauses. Since clausegroups may contain
- * both restriction and join clauses, we have to do a set union to get
- * the full set of clauses that must be considered to compute the
- * correct selectivity. (Without the union operation, we might have
- * some restriction clauses appearing twice, which'd mislead
- * clauselist_selectivity into double-counting their selectivity.
- * However, since RestrictInfo nodes aren't copied when linking them
- * into different lists, it should be sufficient to use pointer
- * comparison to remove duplicates.)
- *
- * Always assume the join type is JOIN_INNER; even if some of the join
- * clauses come from other contexts, that's not our problem.
+ * If we found clauses in more than one list, we may now have
+ * clauses that are known redundant. Get rid of 'em.
*/
- allclauses = list_union_ptr(rel->baserestrictinfo, allclauses);
- pathnode->rows = rel->tuples *
- clauselist_selectivity(root,
- allclauses,
- rel->relid, /* do not use 0! */
- JOIN_INNER);
- /* Like costsize.c, force estimate to be at least one row */
- pathnode->rows = clamp_row_est(pathnode->rows);
-
- cost_index(&pathnode->path, root, rel, index, indexquals, true);
-
- return (Path *) pathnode;
+ if (numsources > 1)
+ {
+ clause_list = remove_redundant_join_clauses(root,
+ clause_list,
+ isouterjoin);
+ }
+
+ return clause_list;
}
+/****************************************************************************
+ * ---- PATH CREATION UTILITIES ----
+ ****************************************************************************/
+
/*
* flatten_clausegroups_list
* Given a list of lists of RestrictInfos, flatten it to a list
* of RestrictInfos.
*
* This is used to flatten out the result of group_clauses_by_indexkey()
- * or one of its sibling routines, to produce an indexclauses list.
+ * to produce an indexclauses list.
*/
List *
flatten_clausegroups_list(List *clausegroups)
return allclauses;
}
-/*
- * make_expr_from_indexclauses()
- * Given an indexclauses structure, produce an ordinary boolean expression.
- *
- * This consists of stripping out the RestrictInfo nodes and inserting
- * explicit AND and OR nodes as needed. There's not much to it, but
- * the functionality is needed in a few places, so centralize the logic.
- */
-Expr *
-make_expr_from_indexclauses(List *indexclauses)
-{
- List *orclauses = NIL;
- ListCell *orlist;
-
- /* There's no such thing as an indexpath with zero scans */
- Assert(indexclauses != NIL);
-
- foreach(orlist, indexclauses)
- {
- List *andlist = (List *) lfirst(orlist);
-
- /* Strip RestrictInfos */
- andlist = get_actual_clauses(andlist);
- /* Insert AND node if needed, and add to orclauses list */
- orclauses = lappend(orclauses, make_ands_explicit(andlist));
- }
-
- if (list_length(orclauses) > 1)
- return make_orclause(orclauses);
- else
- return (Expr *) linitial(orclauses);
-}
-
/****************************************************************************
* ---- ROUTINES TO CHECK OPERANDS ----
*
* operand: the nodetree to be compared to the index
* indexcol: the column number of the index (counting from 0)
- * rel: the parent relation
* index: the index of interest
*/
-static bool
+bool
match_index_to_operand(Node *operand,
int indexcol,
- RelOptInfo *rel,
IndexOptInfo *index)
{
int indkey;
* Simple index column; operand must be a matching Var.
*/
if (operand && IsA(operand, Var) &&
- rel->relid == ((Var *) operand)->varno &&
+ index->rel->relid == ((Var *) operand)->varno &&
indkey == ((Var *) operand)->varattno)
return true;
}
* from LIKE to indexscan limits rather harder than one might think ...
* but that's the basic idea.)
*
- * Two routines are provided here, match_special_index_operator() and
- * expand_indexqual_conditions(). match_special_index_operator() is
- * just an auxiliary function for match_clause_to_indexcol(); after
- * the latter fails to recognize a restriction opclause's operator
- * as a member of an index's opclass, it asks match_special_index_operator()
- * whether the clause should be considered an indexqual anyway.
+ * Another thing that we do with this machinery is to provide special
+ * smarts for "boolean" indexes (that is, indexes on boolean columns
+ * that support boolean equality). We can transform a plain reference
+ * to the indexkey into "indexkey = true", or "NOT indexkey" into
+ * "indexkey = false", so as to make the expression indexable using the
+ * regular index operators. (As of Postgres 8.1, we must do this here
+ * because constant simplification does the reverse transformation;
+ * without this code there'd be no way to use such an index at all.)
+ *
+ * Three routines are provided here:
+ *
+ * match_special_index_operator() is just an auxiliary function for
+ * match_clause_to_indexcol(); after the latter fails to recognize a
+ * restriction opclause's operator as a member of an index's opclass,
+ * it asks match_special_index_operator() whether the clause should be
+ * considered an indexqual anyway.
+ *
+ * match_boolean_index_clause() similarly detects clauses that can be
+ * converted into boolean equality operators.
+ *
* expand_indexqual_conditions() converts a list of lists of RestrictInfo
* nodes (with implicit AND semantics across list elements) into
* a list of clauses that the executor can actually handle. For operators
* that are members of the index's opclass this transformation is a no-op,
- * but operators recognized by match_special_index_operator() must be
- * converted into one or more "regular" indexqual conditions.
+ * but clauses recognized by match_special_index_operator() or
+ * match_boolean_index_clause() must be converted into one or more "regular"
+ * indexqual conditions.
*----------
*/
+/*
+ * match_boolean_index_clause
+ * Recognize restriction clauses that can be matched to a boolean index.
+ *
+ * This should be called only when IsBooleanOpclass() recognizes the
+ * index's operator class. We check to see if the clause matches the
+ * index's key.
+ */
+static bool
+match_boolean_index_clause(Node *clause,
+ int indexcol,
+ IndexOptInfo *index)
+{
+ /* Direct match? */
+ if (match_index_to_operand(clause, indexcol, index))
+ return true;
+ /* NOT clause? */
+ if (not_clause(clause))
+ {
+ if (match_index_to_operand((Node *) get_notclausearg((Expr *) clause),
+ indexcol, index))
+ return true;
+ }
+ /*
+ * Since we only consider clauses at top level of WHERE, we can convert
+ * indexkey IS TRUE and indexkey IS FALSE to index searches as well.
+ * The different meaning for NULL isn't important.
+ */
+ else if (clause && IsA(clause, BooleanTest))
+ {
+ BooleanTest *btest = (BooleanTest *) clause;
+
+ if (btest->booltesttype == IS_TRUE ||
+ btest->booltesttype == IS_FALSE)
+ if (match_index_to_operand((Node *) btest->arg,
+ indexcol, index))
+ return true;
+ }
+ return false;
+}
+
/*
* match_special_index_operator
* Recognize restriction clauses that can be used to generate
* expand_indexqual_conditions
* Given a list of sublists of RestrictInfo nodes, produce a flat list
* of index qual clauses. Standard qual clauses (those in the index's
- * opclass) are passed through unchanged. "Special" index operators
- * are expanded into clauses that the indexscan machinery will know
- * what to do with.
+ * opclass) are passed through unchanged. Boolean clauses and "special"
+ * index operators are expanded into clauses that the indexscan machinery
+ * will know what to do with.
*
* The input list is ordered by index key, and so the output list is too.
* (The latter is not depended on by any part of the planner, so far as I can
- * tell; but some parts of the executor do assume that the indxqual list
+ * tell; but some parts of the executor do assume that the indexqual list
* ultimately delivered to the executor is so ordered. One such place is
* _bt_preprocess_keys() in the btree support. Perhaps that ought to be fixed
* someday --- tgl 7/00)
{
List *resultquals = NIL;
ListCell *clausegroup_item;
+ int indexcol = 0;
Oid *classes = index->classlist;
if (clausegroups == NIL)
{
RestrictInfo *rinfo = (RestrictInfo *) lfirst(l);
+ /* First check for boolean cases */
+ if (IsBooleanOpclass(curClass))
+ {
+ Expr *boolqual;
+
+ boolqual = expand_boolean_index_clause((Node *) rinfo->clause,
+ indexcol,
+ index);
+ if (boolqual)
+ {
+ resultquals = lappend(resultquals,
+ make_restrictinfo(boolqual,
+ true, true));
+ continue;
+ }
+ }
+
resultquals = list_concat(resultquals,
expand_indexqual_condition(rinfo,
- curClass));
+ curClass));
}
clausegroup_item = lnext(clausegroup_item);
+
+ indexcol++;
classes++;
} while (clausegroup_item != NULL && !DoneMatchingIndexKeys(classes));
return resultquals;
}
+/*
+ * expand_boolean_index_clause
+ * Convert a clause recognized by match_boolean_index_clause into
+ * a boolean equality operator clause.
+ *
+ * Returns NULL if the clause isn't a boolean index qual.
+ */
+static Expr *
+expand_boolean_index_clause(Node *clause,
+ int indexcol,
+ IndexOptInfo *index)
+{
+ /* Direct match? */
+ if (match_index_to_operand(clause, indexcol, index))
+ {
+ /* convert to indexkey = TRUE */
+ return make_opclause(BooleanEqualOperator, BOOLOID, false,
+ (Expr *) clause,
+ (Expr *) makeBoolConst(true, false));
+ }
+ /* NOT clause? */
+ if (not_clause(clause))
+ {
+ Node *arg = (Node *) get_notclausearg((Expr *) clause);
+
+ /* It must have matched the indexkey */
+ Assert(match_index_to_operand(arg, indexcol, index));
+ /* convert to indexkey = FALSE */
+ return make_opclause(BooleanEqualOperator, BOOLOID, false,
+ (Expr *) arg,
+ (Expr *) makeBoolConst(false, false));
+ }
+ if (clause && IsA(clause, BooleanTest))
+ {
+ BooleanTest *btest = (BooleanTest *) clause;
+ Node *arg = (Node *) btest->arg;
+
+ /* It must have matched the indexkey */
+ Assert(match_index_to_operand(arg, indexcol, index));
+ if (btest->booltesttype == IS_TRUE)
+ {
+ /* convert to indexkey = TRUE */
+ return make_opclause(BooleanEqualOperator, BOOLOID, false,
+ (Expr *) arg,
+ (Expr *) makeBoolConst(true, false));
+ }
+ if (btest->booltesttype == IS_FALSE)
+ {
+ /* convert to indexkey = FALSE */
+ return make_opclause(BooleanEqualOperator, BOOLOID, false,
+ (Expr *) arg,
+ (Expr *) makeBoolConst(false, false));
+ }
+ /* Oops */
+ Assert(false);
+ }
+
+ return NULL;
+}
+
/*
* expand_indexqual_condition --- expand a single indexqual condition
+ * (other than a boolean-qual case)
*
* The input is a single RestrictInfo, the output a list of RestrictInfos
*/
expand_indexqual_condition(RestrictInfo *rinfo, Oid opclass)
{
Expr *clause = rinfo->clause;
-
/* we know these will succeed */
Node *leftop = get_leftop(clause);
Node *rightop = get_rightop(clause);