mirrors/postgres
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Planner speedup hacking. Avoid saving useless pathkeys, so that path

Browse Source 
comparison does not consider paths different when they differ only in
uninteresting aspects of sort order.  (We had a special case of this
consideration for indexscans already, but generalize it to apply to
ordered join paths too.)  Be stricter about what is a canonical pathkey
to allow faster pathkey comparison.  Cache canonical pathkeys and
dispersion stats for left and right sides of a RestrictInfo's clause,
to avoid repeated computation.  Total speedup will depend on number of
tables in a query, but I see about 4x speedup of planning phase for
a sample seven-table query.
This commit is contained in:
Tom Lane 2000-12-14 22:30:45 +00:00
parent db11f4382a
commit ea166f1146
16 changed files with 622 additions and 365 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
src/backend/nodes/copyfuncs.c
View File

@ -15,7 +15,7 @@
* Portions Copyright (c) 1994, Regents of the University of California * Portions Copyright (c) 1994, Regents of the University of California
* *
* IDENTIFICATION * IDENTIFICATION
* $Header: /cvsroot/pgsql/src/backend/nodes/copyfuncs.c,v 1.134 2000/12/12 23:33:32 tgl Exp $ * $Header: /cvsroot/pgsql/src/backend/nodes/copyfuncs.c,v 1.135 2000/12/14 22:30:42 tgl Exp $
* *
*------------------------------------------------------------------------- *-------------------------------------------------------------------------
*/ */
@ -1424,7 +1424,12 @@ _copyRestrictInfo(RestrictInfo *from)
newnode->mergejoinoperator = from->mergejoinoperator; newnode->mergejoinoperator = from->mergejoinoperator;
newnode->left_sortop = from->left_sortop; newnode->left_sortop = from->left_sortop;
newnode->right_sortop = from->right_sortop; newnode->right_sortop = from->right_sortop;
/* Do not copy pathkeys, since they'd not be canonical in a copied query */
newnode->left_pathkey = NIL;
newnode->right_pathkey = NIL;
newnode->hashjoinoperator = from->hashjoinoperator; newnode->hashjoinoperator = from->hashjoinoperator;
newnode->left_dispersion = from->left_dispersion;
newnode->right_dispersion = from->right_dispersion;
return newnode; return newnode;
} }

7
src/backend/nodes/equalfuncs.c
View File

@ -20,7 +20,7 @@
* Portions Copyright (c) 1994, Regents of the University of California * Portions Copyright (c) 1994, Regents of the University of California
* *
* IDENTIFICATION * IDENTIFICATION
* $Header: /cvsroot/pgsql/src/backend/nodes/equalfuncs.c,v 1.84 2000/12/12 23:33:33 tgl Exp $ * $Header: /cvsroot/pgsql/src/backend/nodes/equalfuncs.c,v 1.85 2000/12/14 22:30:42 tgl Exp $
* *
*------------------------------------------------------------------------- *-------------------------------------------------------------------------
*/ */
@ -515,8 +515,9 @@ _equalRestrictInfo(RestrictInfo *a, RestrictInfo *b)
if (!equal(a->clause, b->clause)) if (!equal(a->clause, b->clause))
return false; return false;
/* /*
* ignore eval_cost, since it may not be set yet, and should be * ignore eval_cost, left/right_pathkey, and left/right_dispersion,
* derivable from the clause anyway * since they may not be set yet, and should be derivable from the
* clause anyway
*/ */
if (a->ispusheddown != b->ispusheddown) if (a->ispusheddown != b->ispusheddown)
return false; return false;

7
src/backend/nodes/readfuncs.c
View File

@ -8,7 +8,7 @@
* *
* *
* IDENTIFICATION * IDENTIFICATION
* $Header: /cvsroot/pgsql/src/backend/nodes/readfuncs.c,v 1.101 2000/12/12 23:33:33 tgl Exp $ * $Header: /cvsroot/pgsql/src/backend/nodes/readfuncs.c,v 1.102 2000/12/14 22:30:42 tgl Exp $
* *
* NOTES * NOTES
* Most of the read functions for plan nodes are tested. (In fact, they * Most of the read functions for plan nodes are tested. (In fact, they
@ -1848,6 +1848,11 @@ _readRestrictInfo(void)
/* eval_cost is not part of saved representation; compute on first use */ /* eval_cost is not part of saved representation; compute on first use */
local_node->eval_cost = -1; local_node->eval_cost = -1;
/* ditto for cached pathkeys and dispersion */
local_node->left_pathkey = NIL;
local_node->right_pathkey = NIL;
local_node->left_dispersion = -1;
local_node->right_dispersion = -1;
return local_node; return local_node;
} }

39
src/backend/optimizer/README
View File

@ -356,9 +356,13 @@ relation the pathkey is for, *no matter how we formed the join*. It works
as long as the clause has been applied at some point while forming the as long as the clause has been applied at some point while forming the
join relation. (In the current implementation, we always apply qual join relation. (In the current implementation, we always apply qual
clauses as soon as possible, ie, as far down in the plan tree as possible. clauses as soon as possible, ie, as far down in the plan tree as possible.
So we can always make this deduction. If we postponed filtering by qual So we can treat the pathkeys as equivalent everywhere. The exception is
clauses then we'd not be able to assume pathkey equivalence until after when the relations A and B are joined inside the nullable side of an
the equality check(s) had been applied.) OUTER JOIN and the equijoin clause comes from above the OUTER JOIN. In this
case we cannot apply the qual as soon as A and B are joined, so we do not
consider the pathkeys to be equivalent. This could be improved if we wanted
to go to the trouble of making pathkey equivalence be context-dependent,
but that seems much more complex than it's worth.)
In short, then: when producing the pathkeys for a merge or nestloop join, In short, then: when producing the pathkeys for a merge or nestloop join,
we can keep all of the keys of the outer path, since the ordering of the we can keep all of the keys of the outer path, since the ordering of the
@ -415,18 +419,14 @@ whenever we make a pathkey sublist that mentions any var appearing in an
equivalence set, we instantly add all the other vars equivalenced to it, equivalence set, we instantly add all the other vars equivalenced to it,
whether they appear yet in the pathkey's relation or not. And we also whether they appear yet in the pathkey's relation or not. And we also
mandate that the pathkey sublist appear in the same order as the mandate that the pathkey sublist appear in the same order as the
equivalence set it comes from. (In practice, we simply return a pointer equivalence set it comes from.
to the relevant equivalence set without building any new sublist at all.
Each equivalence set becomes a "canonical pathkey" for all its members.)
This makes comparing pathkeys very simple and fast, and saves a lot of
work and memory space for pathkey construction as well.
Note that pathkey sublists having just one item still exist, and are In fact, we can go even further, and say that the canonical representation
not expected to be equal() to any equivalence set. This occurs when of a pathkey sublist is a pointer directly to the relevant equivalence set,
we describe a sort order that involves a var that's not mentioned in which is kept in a list of pathkey equivalence sets for the query. Then
any equijoin clause of the WHERE. We could add singleton sets containing pathkey sublist comparison reduces to pointer-equality checking! To do this
such vars to the query's list of equivalence sets, but there's little we also have to add single-element pathkey sublists to the query's list of
point in doing so. equivalence sets, but that's a small price to pay.
By the way, it's OK and even useful for us to build equivalence sets By the way, it's OK and even useful for us to build equivalence sets
that mention multiple vars from the same relation. For example, if that mention multiple vars from the same relation. For example, if
@ -469,4 +469,15 @@ we're given WHERE int2var = int4var AND int4var = int8var, we'll fail
while trying to create a representation of the implied clause while trying to create a representation of the implied clause
int2var = int8var. int2var = int8var.
An additional refinement we can make is to insist that canonical pathkey
lists (sort orderings) do not mention the same pathkey set more than once.
For example, a pathkey list ((A) (B) (A)) is redundant --- the second
occurrence of (A) does not change the ordering, since the data must already
be sorted by A. Although a user probably wouldn't write ORDER BY A,B,A
directly, such redundancies are more probable once equijoin equivalences
have been considered. Also, the system is likely to generate redundant
pathkey lists when computing the sort ordering needed for a mergejoin. By
eliminating the redundancy, we save time and improve planning, since the
planner will more easily recognize equivalent orderings as being equivalent.
-- bjm & tgl -- bjm & tgl

6
src/backend/optimizer/path/allpaths.c
View File

@ -8,7 +8,7 @@
* *
* *
* IDENTIFICATION * IDENTIFICATION
* $Header: /cvsroot/pgsql/src/backend/optimizer/path/allpaths.c,v 1.67 2000/11/12 00:36:58 tgl Exp $ * $Header: /cvsroot/pgsql/src/backend/optimizer/path/allpaths.c,v 1.68 2000/12/14 22:30:43 tgl Exp $
* *
*------------------------------------------------------------------------- *-------------------------------------------------------------------------
*/ */
@ -162,9 +162,7 @@ set_plain_rel_pathlist(Query *root, RelOptInfo *rel, RangeTblEntry *rte)
create_tidscan_paths(root, rel); create_tidscan_paths(root, rel);
/* Consider index paths for both simple and OR index clauses */ /* Consider index paths for both simple and OR index clauses */
create_index_paths(root, rel, indices, create_index_paths(root, rel, indices);
rel->baserestrictinfo,
rel->joininfo);
/* /*
* Note: create_or_index_paths depends on create_index_paths to * Note: create_or_index_paths depends on create_index_paths to

172
src/backend/optimizer/path/indxpath.c
View File

@ -9,7 +9,7 @@
* *
* *
* IDENTIFICATION * IDENTIFICATION
* $Header: /cvsroot/pgsql/src/backend/optimizer/path/indxpath.c,v 1.99 2000/11/25 20:33:51 tgl Exp $ * $Header: /cvsroot/pgsql/src/backend/optimizer/path/indxpath.c,v 1.100 2000/12/14 22:30:43 tgl Exp $
* *
*------------------------------------------------------------------------- *-------------------------------------------------------------------------
*/ */
@ -87,11 +87,6 @@ static void indexable_joinclauses(RelOptInfo *rel, IndexOptInfo *index,
List **clausegroups, List **outerrelids); List **clausegroups, List **outerrelids);
static List *index_innerjoin(Query *root, RelOptInfo *rel, IndexOptInfo *index, static List *index_innerjoin(Query *root, RelOptInfo *rel, IndexOptInfo *index,
List *clausegroup_list, List *outerrelids_list); List *clausegroup_list, List *outerrelids_list);
static bool useful_for_mergejoin(RelOptInfo *rel, IndexOptInfo *index,
List *joininfo_list);
static bool useful_for_ordering(Query *root, RelOptInfo *rel,
IndexOptInfo *index,
ScanDirection scandir);
static bool match_index_to_operand(int indexkey, Var *operand, static bool match_index_to_operand(int indexkey, Var *operand,
RelOptInfo *rel, IndexOptInfo *index); RelOptInfo *rel, IndexOptInfo *index);
static bool function_index_operand(Expr *funcOpnd, RelOptInfo *rel, static bool function_index_operand(Expr *funcOpnd, RelOptInfo *rel,
@ -125,31 +120,31 @@ static Const *string_to_const(const char *str, Oid datatype);
* attributes are available and fixed during any one scan of the indexpath. * attributes are available and fixed during any one scan of the indexpath.
* *
* An IndexPath is generated and submitted to add_path() for each index * An IndexPath is generated and submitted to add_path() for each index
* this routine deems potentially interesting for the current query * this routine deems potentially interesting for the current query.
* (at most one IndexPath per index on the given relation). An innerjoin * An innerjoin path is also generated for each interesting combination of
* path is also generated for each interesting combination of outer join * outer join relations. The innerjoin paths are *not* passed to add_path(),
* relations. The innerjoin paths are *not* passed to add_path(), but are * but are appended to the "innerjoin" list of the relation for later
* appended to the "innerjoin" list of the relation for later consideration * consideration in nested-loop joins.
* in nested-loop joins.
* *
* 'rel' is the relation for which we want to generate index paths * 'rel' is the relation for which we want to generate index paths
* 'indices' is a list of available indexes for 'rel' * 'indices' is a list of available indexes for 'rel'
* 'restrictinfo_list' is a list of restrictinfo nodes for 'rel'
* 'joininfo_list' is a list of joininfo nodes for 'rel'
*/ */
void void
create_index_paths(Query *root, create_index_paths(Query *root,
RelOptInfo *rel, RelOptInfo *rel,
List *indices, List *indices)
List *restrictinfo_list,
List *joininfo_list)
{ {
List *restrictinfo_list = rel->baserestrictinfo;
List *joininfo_list = rel->joininfo;
List *ilist; List *ilist;
foreach(ilist, indices) foreach(ilist, indices)
{ {
IndexOptInfo *index = (IndexOptInfo *) lfirst(ilist); IndexOptInfo *index = (IndexOptInfo *) lfirst(ilist);
List *restrictclauses; List *restrictclauses;
List *index_pathkeys;
List *useful_pathkeys;
bool index_is_ordered;
List *joinclausegroups; List *joinclausegroups;
List *joinouterrelids; List *joinouterrelids;
@ -179,9 +174,7 @@ create_index_paths(Query *root,
match_index_orclauses(rel, index, restrictinfo_list); match_index_orclauses(rel, index, restrictinfo_list);
/* /*
* 2. If the keys of this index match any of the available * 2. Match the index against non-'or' restriction clauses.
* non-'or' restriction clauses, then create a path using those
* clauses as indexquals.
*/ */
restrictclauses = group_clauses_by_indexkey(rel, restrictclauses = group_clauses_by_indexkey(rel,
index, index,
@ -189,43 +182,50 @@ create_index_paths(Query *root,
index->classlist, index->classlist,
restrictinfo_list); restrictinfo_list);
if (restrictclauses != NIL) /*
add_path(rel, (Path *) create_index_path(root, rel, index, * 3. Compute pathkeys describing index's ordering, if any,
* then see how many of them are actually useful for this query.
*/
index_pathkeys = build_index_pathkeys(root, rel, index,
ForwardScanDirection);
index_is_ordered = (index_pathkeys != NIL);
useful_pathkeys = truncate_useless_pathkeys(root, rel,
index_pathkeys);
/*
* 4. Generate an indexscan path if there are relevant restriction
* clauses OR the index ordering is potentially useful for later
* merging or final output ordering.
*/
if (restrictclauses != NIL || useful_pathkeys != NIL)
add_path(rel, (Path *)
create_index_path(root, rel, index,
restrictclauses, restrictclauses,
useful_pathkeys,
index_is_ordered ?
ForwardScanDirection :
NoMovementScanDirection)); NoMovementScanDirection));
/* /*
* 3. If this index can be used for a mergejoin, then create an * 5. If the index is ordered, a backwards scan might be interesting.
* index path for it even if there were no restriction clauses. * Currently this is only possible for a DESC query result ordering.
* (If there were, there is no need to make another index path.)
* This will allow the index to be considered as a base for a
* mergejoin in later processing. Similarly, if the index matches
* the ordering that is needed for the overall query result, make
* an index path for it even if there is no other reason to do so.
*/ */
if (restrictclauses == NIL) if (index_is_ordered)
{ {
if (useful_for_mergejoin(rel, index, joininfo_list) || index_pathkeys = build_index_pathkeys(root, rel, index,
useful_for_ordering(root, rel, index, ForwardScanDirection)) BackwardScanDirection);
useful_pathkeys = truncate_useless_pathkeys(root, rel,
index_pathkeys);
if (useful_pathkeys != NIL)
add_path(rel, (Path *) add_path(rel, (Path *)
create_index_path(root, rel, index, create_index_path(root, rel, index,
restrictclauses, restrictclauses,
ForwardScanDirection)); useful_pathkeys,
BackwardScanDirection));
} }
/* /*
* Currently, backwards scan is never considered except for the * 6. Create an innerjoin index path for each combination of other
* case of matching a query result ordering. Possibly should
* consider it in other places?
*/
if (useful_for_ordering(root, rel, index, BackwardScanDirection))
add_path(rel, (Path *)
create_index_path(root, rel, index,
restrictclauses,
BackwardScanDirection));
/*
* 4. Create an innerjoin index path for each combination of other
* rels used in available join clauses. These paths will be * rels used in available join clauses. These paths will be
* considered as the inner side of nestloop joins against those * considered as the inner side of nestloop joins against those
* sets of other rels. indexable_joinclauses() finds sets of * sets of other rels. indexable_joinclauses() finds sets of
@ -904,88 +904,6 @@ indexable_operator(Expr *clause, Oid opclass, Oid relam,
return InvalidOid; return InvalidOid;
} }
/*
* useful_for_mergejoin
* Determine whether the given index can support a mergejoin based
* on any available join clause.
*
* We look to see whether the first indexkey of the index matches the
* left or right sides of any of the mergejoinable clauses and provides
* the ordering needed for that side. If so, the index is useful.
* Matching a second or later indexkey is not useful unless there is
* also a mergeclause for the first indexkey, so we need not consider
* secondary indexkeys at this stage.
*
* 'rel' is the relation for which 'index' is defined
* 'joininfo_list' is the list of JoinInfo nodes for 'rel'
*/
static bool
useful_for_mergejoin(RelOptInfo *rel,
IndexOptInfo *index,
List *joininfo_list)
{
int *indexkeys = index->indexkeys;
Oid *ordering = index->ordering;
List *i;
if (!indexkeys || indexkeys[0] == 0 ||
!ordering || ordering[0] == InvalidOid)
return false; /* unordered index is not useful */
foreach(i, joininfo_list)
{
JoinInfo *joininfo = (JoinInfo *) lfirst(i);
List *j;
foreach(j, joininfo->jinfo_restrictinfo)
{
RestrictInfo *restrictinfo = (RestrictInfo *) lfirst(j);
if (restrictinfo->mergejoinoperator)
{
if (restrictinfo->left_sortop == ordering[0] &&
match_index_to_operand(indexkeys[0],
get_leftop(restrictinfo->clause),
rel, index))
return true;
if (restrictinfo->right_sortop == ordering[0] &&
match_index_to_operand(indexkeys[0],
get_rightop(restrictinfo->clause),
rel, index))
return true;
}
}
}
return false;
}
/*
* useful_for_ordering
* Determine whether the given index can produce an ordering matching
* the order that is wanted for the query result.
*
* 'rel' is the relation for which 'index' is defined
* 'scandir' is the contemplated scan direction
*/
static bool
useful_for_ordering(Query *root,
RelOptInfo *rel,
IndexOptInfo *index,
ScanDirection scandir)
{
List *index_pathkeys;
if (root->query_pathkeys == NIL)
return false; /* no special ordering requested */
index_pathkeys = build_index_pathkeys(root, rel, index, scandir);
if (index_pathkeys == NIL)
return false; /* unordered index */
return pathkeys_contained_in(root->query_pathkeys, index_pathkeys);
}
/**************************************************************************** /****************************************************************************
* ---- ROUTINES TO DO PARTIAL INDEX PREDICATE TESTS ---- * ---- ROUTINES TO DO PARTIAL INDEX PREDICATE TESTS ----
****************************************************************************/ ****************************************************************************/

182
src/backend/optimizer/path/joinpath.c
View File

@ -8,7 +8,7 @@
* *
* *
* IDENTIFICATION * IDENTIFICATION
* $Header: /cvsroot/pgsql/src/backend/optimizer/path/joinpath.c,v 1.59 2000/11/23 03:57:31 tgl Exp $ * $Header: /cvsroot/pgsql/src/backend/optimizer/path/joinpath.c,v 1.60 2000/12/14 22:30:43 tgl Exp $
* *
*------------------------------------------------------------------------- *-------------------------------------------------------------------------
*/ */
@ -152,6 +152,7 @@ sort_inner_and_outer(Query *root,
List *mergeclause_list, List *mergeclause_list,
JoinType jointype) JoinType jointype)
{ {
List *all_pathkeys;
List *i; List *i;
/* /*
@ -159,36 +160,57 @@ sort_inner_and_outer(Query *root,
* generate a differently-sorted result path at essentially the same * generate a differently-sorted result path at essentially the same
* cost. We have no basis for choosing one over another at this level * cost. We have no basis for choosing one over another at this level
* of joining, but some sort orders may be more useful than others for * of joining, but some sort orders may be more useful than others for
* higher-level mergejoins. Generating a path here for *every* * higher-level mergejoins, so it's worth considering multiple orderings.
* permutation of mergejoin clauses doesn't seem like a winning
* strategy, however; the cost in planning time is too high.
* *
* For now, we generate one path for each mergejoin clause, listing that * Actually, it's not quite true that every mergeclause ordering will
* clause first and the rest in random order. This should allow at * generate a different path order, because some of the clauses may be
* redundant. Therefore, what we do is convert the mergeclause list to
* a list of canonical pathkeys, and then consider different orderings
* of the pathkeys.
*
* Generating a path for *every* permutation of the pathkeys doesn't
* seem like a winning strategy; the cost in planning time is too high.
* For now, we generate one path for each pathkey, listing that pathkey
* first and the rest in random order. This should allow at
* least a one-clause mergejoin without re-sorting against any other * least a one-clause mergejoin without re-sorting against any other
* possible mergejoin partner path. But if we've not guessed the * possible mergejoin partner path. But if we've not guessed the
* right ordering of secondary clauses, we may end up evaluating * right ordering of secondary keys, we may end up evaluating
* clauses as qpquals when they could have been done as mergeclauses. * clauses as qpquals when they could have been done as mergeclauses.
* We need to figure out a better way. (Two possible approaches: look * We need to figure out a better way. (Two possible approaches: look
* at all the relevant index relations to suggest plausible sort * at all the relevant index relations to suggest plausible sort
* orders, or make just one output path and somehow mark it as having * orders, or make just one output path and somehow mark it as having
* a sort-order that can be rearranged freely.) * a sort-order that can be rearranged freely.)
*/ */
foreach(i, mergeclause_list) all_pathkeys = make_pathkeys_for_mergeclauses(root,
mergeclause_list,
outerrel);
foreach(i, all_pathkeys)
{ {
RestrictInfo *restrictinfo = lfirst(i); List *front_pathkey = lfirst(i);
List *curclause_list; List *cur_pathkeys;
List *cur_mergeclauses;
List *outerkeys; List *outerkeys;
List *innerkeys; List *innerkeys;
List *merge_pathkeys; List *merge_pathkeys;
/* Make a mergeclause list with this guy first. */ /* Make a pathkey list with this guy first. */
if (i != mergeclause_list) if (i != all_pathkeys)
curclause_list = lcons(restrictinfo, cur_pathkeys = lcons(front_pathkey,
lremove(restrictinfo, lremove(front_pathkey,
listCopy(mergeclause_list))); listCopy(all_pathkeys)));
else else
curclause_list = mergeclause_list; /* no work at first one... */ cur_pathkeys = all_pathkeys; /* no work at first one... */
/*
* Select mergeclause(s) that match this sort ordering. If we had
* redundant merge clauses then we will get a subset of the original
* clause list. There had better be some match, however...
*/
cur_mergeclauses = find_mergeclauses_for_pathkeys(root,
cur_pathkeys,
mergeclause_list);
Assert(cur_mergeclauses != NIL);
/* /*
* Build sort pathkeys for both sides. * Build sort pathkeys for both sides.
@ -198,15 +220,13 @@ sort_inner_and_outer(Query *root,
* suppress an explicit sort step, so we needn't do so here. * suppress an explicit sort step, so we needn't do so here.
*/ */
outerkeys = make_pathkeys_for_mergeclauses(root, outerkeys = make_pathkeys_for_mergeclauses(root,
curclause_list, cur_mergeclauses,
outerrel); outerrel);
innerkeys = make_pathkeys_for_mergeclauses(root, innerkeys = make_pathkeys_for_mergeclauses(root,
curclause_list, cur_mergeclauses,
innerrel); innerrel);
/* Build pathkeys representing output sort order. */ /* Build pathkeys representing output sort order. */
merge_pathkeys = build_join_pathkeys(outerkeys, merge_pathkeys = build_join_pathkeys(root, joinrel, outerkeys);
joinrel->targetlist,
root->equi_key_list);
/* /*
* And now we can make the path. We only consider the cheapest- * And now we can make the path. We only consider the cheapest-
@ -221,7 +241,7 @@ sort_inner_and_outer(Query *root,
innerrel->cheapest_total_path, innerrel->cheapest_total_path,
restrictlist, restrictlist,
merge_pathkeys, merge_pathkeys,
curclause_list, cur_mergeclauses,
outerkeys, outerkeys,
innerkeys)); innerkeys));
} }
@ -301,17 +321,16 @@ match_unsorted_outer(Query *root,
List *trialsortkeys; List *trialsortkeys;
Path *cheapest_startup_inner; Path *cheapest_startup_inner;
Path *cheapest_total_inner; Path *cheapest_total_inner;
int num_mergeclauses; int num_sortkeys;
int clausecnt; int sortkeycnt;
/* /*
* The result will have this sort order (even if it is implemented * The result will have this sort order (even if it is implemented
* as a nestloop, and even if some of the mergeclauses are * as a nestloop, and even if some of the mergeclauses are
* implemented by qpquals rather than as true mergeclauses): * implemented by qpquals rather than as true mergeclauses):
*/ */
merge_pathkeys = build_join_pathkeys(outerpath->pathkeys, merge_pathkeys = build_join_pathkeys(root, joinrel,
joinrel->targetlist, outerpath->pathkeys);
root->equi_key_list);
if (nestjoinOK) if (nestjoinOK)
{ {
@ -347,7 +366,8 @@ match_unsorted_outer(Query *root,
} }
/* Look for useful mergeclauses (if any) */ /* Look for useful mergeclauses (if any) */
mergeclauses = find_mergeclauses_for_pathkeys(outerpath->pathkeys, mergeclauses = find_mergeclauses_for_pathkeys(root,
outerpath->pathkeys,
mergeclause_list); mergeclause_list);
/* Done with this outer path if no chance for a mergejoin */ /* Done with this outer path if no chance for a mergejoin */
@ -362,7 +382,8 @@ match_unsorted_outer(Query *root,
/* /*
* Generate a mergejoin on the basis of sorting the cheapest * Generate a mergejoin on the basis of sorting the cheapest
* inner. Since a sort will be needed, only cheapest total cost * inner. Since a sort will be needed, only cheapest total cost
* matters. * matters. (But create_mergejoin_path will do the right thing
* if innerrel->cheapest_total_path is already correctly sorted.)
*/ */
add_path(joinrel, (Path *) add_path(joinrel, (Path *)
create_mergejoin_path(joinrel, create_mergejoin_path(joinrel,
@ -376,38 +397,49 @@ match_unsorted_outer(Query *root,
innersortkeys)); innersortkeys));
/* /*
* Look for presorted inner paths that satisfy the mergeclause * Look for presorted inner paths that satisfy the innersortkey
* list or any truncation thereof. Here, we consider both cheap * list or any truncation thereof. Here, we consider both cheap
* startup cost and cheap total cost. * startup cost and cheap total cost. Ignore
* innerrel->cheapest_total_path, since we already made a path with it.
*/ */
trialsortkeys = listCopy(innersortkeys); /* modifiable copy */ num_sortkeys = length(innersortkeys);
if (num_sortkeys > 1)
trialsortkeys = listCopy(innersortkeys); /* need modifiable copy */
else
trialsortkeys = innersortkeys; /* won't really truncate */
cheapest_startup_inner = NULL; cheapest_startup_inner = NULL;
cheapest_total_inner = NULL; cheapest_total_inner = NULL;
num_mergeclauses = length(mergeclauses);
for (clausecnt = num_mergeclauses; clausecnt > 0; clausecnt--) for (sortkeycnt = num_sortkeys; sortkeycnt > 0; sortkeycnt--)
{ {
Path *innerpath; Path *innerpath;
List *newclauses = NIL; List *newclauses = NIL;
/* /*
* Look for an inner path ordered well enough to merge with * Look for an inner path ordered well enough for the first
* the first 'clausecnt' mergeclauses. NB: trialsortkeys list * 'sortkeycnt' innersortkeys. NB: trialsortkeys list
* is modified destructively, which is why we made a copy... * is modified destructively, which is why we made a copy...
*/ */
trialsortkeys = ltruncate(clausecnt, trialsortkeys); trialsortkeys = ltruncate(sortkeycnt, trialsortkeys);
innerpath = get_cheapest_path_for_pathkeys(innerrel->pathlist, innerpath = get_cheapest_path_for_pathkeys(innerrel->pathlist,
trialsortkeys, trialsortkeys,
TOTAL_COST); TOTAL_COST);
if (innerpath != NULL && if (innerpath != NULL &&
innerpath != innerrel->cheapest_total_path &&
(cheapest_total_inner == NULL || (cheapest_total_inner == NULL ||
compare_path_costs(innerpath, cheapest_total_inner, compare_path_costs(innerpath, cheapest_total_inner,
TOTAL_COST) < 0)) TOTAL_COST) < 0))
{ {
/* Found a cheap (or even-cheaper) sorted path */ /* Found a cheap (or even-cheaper) sorted path */
if (clausecnt < num_mergeclauses) /* Select the right mergeclauses, if we didn't already */
newclauses = ltruncate(clausecnt, if (sortkeycnt < num_sortkeys)
listCopy(mergeclauses)); {
newclauses =
find_mergeclauses_for_pathkeys(root,
trialsortkeys,
mergeclauses);
Assert(newclauses != NIL);
}
else else
newclauses = mergeclauses; newclauses = mergeclauses;
add_path(joinrel, (Path *) add_path(joinrel, (Path *)
@ -427,6 +459,7 @@ match_unsorted_outer(Query *root,
trialsortkeys, trialsortkeys,
STARTUP_COST); STARTUP_COST);
if (innerpath != NULL && if (innerpath != NULL &&
innerpath != innerrel->cheapest_total_path &&
(cheapest_startup_inner == NULL || (cheapest_startup_inner == NULL ||
compare_path_costs(innerpath, cheapest_startup_inner, compare_path_costs(innerpath, cheapest_startup_inner,
STARTUP_COST) < 0)) STARTUP_COST) < 0))
@ -441,9 +474,14 @@ match_unsorted_outer(Query *root,
*/ */
if (newclauses == NIL) if (newclauses == NIL)
{ {
if (clausecnt < num_mergeclauses) if (sortkeycnt < num_sortkeys)
newclauses = ltruncate(clausecnt, {
listCopy(mergeclauses)); newclauses =
find_mergeclauses_for_pathkeys(root,
trialsortkeys,
mergeclauses);
Assert(newclauses != NIL);
}
else else
newclauses = mergeclauses; newclauses = mergeclauses;
} }
@ -501,7 +539,8 @@ match_unsorted_inner(Query *root,
Path *startupouterpath; Path *startupouterpath;
/* Look for useful mergeclauses (if any) */ /* Look for useful mergeclauses (if any) */
mergeclauses = find_mergeclauses_for_pathkeys(innerpath->pathkeys, mergeclauses = find_mergeclauses_for_pathkeys(root,
innerpath->pathkeys,
mergeclause_list); mergeclause_list);
if (mergeclauses == NIL) if (mergeclauses == NIL)
continue; continue;
@ -516,9 +555,7 @@ match_unsorted_inner(Query *root,
* outer. Since a sort will be needed, only cheapest total cost * outer. Since a sort will be needed, only cheapest total cost
* matters. * matters.
*/ */
merge_pathkeys = build_join_pathkeys(outersortkeys, merge_pathkeys = build_join_pathkeys(root, joinrel, outersortkeys);
joinrel->targetlist,
root->equi_key_list);
add_path(joinrel, (Path *) add_path(joinrel, (Path *)
create_mergejoin_path(joinrel, create_mergejoin_path(joinrel,
jointype, jointype,
@ -545,9 +582,8 @@ match_unsorted_inner(Query *root,
continue; /* there won't be a startup-cost path continue; /* there won't be a startup-cost path
* either */ * either */
merge_pathkeys = build_join_pathkeys(totalouterpath->pathkeys, merge_pathkeys = build_join_pathkeys(root, joinrel,
joinrel->targetlist, totalouterpath->pathkeys);
root->equi_key_list);
add_path(joinrel, (Path *) add_path(joinrel, (Path *)
create_mergejoin_path(joinrel, create_mergejoin_path(joinrel,
jointype, jointype,
@ -564,9 +600,8 @@ match_unsorted_inner(Query *root,
STARTUP_COST); STARTUP_COST);
if (startupouterpath != NULL && startupouterpath != totalouterpath) if (startupouterpath != NULL && startupouterpath != totalouterpath)
{ {
merge_pathkeys = build_join_pathkeys(startupouterpath->pathkeys, merge_pathkeys = build_join_pathkeys(root, joinrel,
joinrel->targetlist, startupouterpath->pathkeys);
root->equi_key_list);
add_path(joinrel, (Path *) add_path(joinrel, (Path *)
create_mergejoin_path(joinrel, create_mergejoin_path(joinrel,
jointype, jointype,
@ -637,10 +672,9 @@ hash_inner_and_outer(Query *root,
RestrictInfo *restrictinfo = (RestrictInfo *) lfirst(i); RestrictInfo *restrictinfo = (RestrictInfo *) lfirst(i);
Expr *clause; Expr *clause;
Var *left, Var *left,
*right, *right;
*inner;
List *hashclauses;
Selectivity innerdispersion; Selectivity innerdispersion;
List *hashclauses;
if (restrictinfo->hashjoinoperator == InvalidOid) if (restrictinfo->hashjoinoperator == InvalidOid)
continue; /* not hashjoinable */ continue; /* not hashjoinable */
@ -657,26 +691,48 @@ hash_inner_and_outer(Query *root,
left = get_leftop(clause); left = get_leftop(clause);
right = get_rightop(clause); right = get_rightop(clause);
/* check if clause is usable with these sub-rels, find inner var */ /*
* Check if clause is usable with these sub-rels, find inner side,
* estimate dispersion of inner var for costing purposes.
*
* Since we tend to visit the same clauses over and over when
* planning a large query, we cache the dispersion estimates in the
* RestrictInfo node to avoid repeated lookups of statistics.
*/
if (intMember(left->varno, outerrelids) && if (intMember(left->varno, outerrelids) &&
intMember(right->varno, innerrelids)) intMember(right->varno, innerrelids))
inner = right; {
/* righthand side is inner */
innerdispersion = restrictinfo->right_dispersion;
if (innerdispersion < 0)
{
/* not cached yet */
innerdispersion = estimate_dispersion(root, right);
restrictinfo->right_dispersion = innerdispersion;
}
}
else if (intMember(left->varno, innerrelids) && else if (intMember(left->varno, innerrelids) &&
intMember(right->varno, outerrelids)) intMember(right->varno, outerrelids))
inner = left; {
/* lefthand side is inner */
innerdispersion = restrictinfo->left_dispersion;
if (innerdispersion < 0)
{
/* not cached yet */
innerdispersion = estimate_dispersion(root, left);
restrictinfo->left_dispersion = innerdispersion;
}
}
else else
continue; /* no good for these input relations */ continue; /* no good for these input relations */
/* always a one-element list of hash clauses */ /* always a one-element list of hash clauses */
hashclauses = makeList1(restrictinfo); hashclauses = makeList1(restrictinfo);
/* estimate dispersion of inner var for costing purposes */
innerdispersion = estimate_dispersion(root, inner);
/* /*
* We consider both the cheapest-total-cost and * We consider both the cheapest-total-cost and
* cheapest-startup-cost outer paths. There's no need to consider * cheapest-startup-cost outer paths. There's no need to consider
* any but the cheapest- total-cost inner path, however. * any but the cheapest-total-cost inner path, however.
*/ */
add_path(joinrel, (Path *) add_path(joinrel, (Path *)
create_hashjoin_path(joinrel, create_hashjoin_path(joinrel,

413
src/backend/optimizer/path/pathkeys.c
View File

@ -11,7 +11,7 @@
* Portions Copyright (c) 1994, Regents of the University of California * Portions Copyright (c) 1994, Regents of the University of California
* *
* IDENTIFICATION * IDENTIFICATION
* $Header: /cvsroot/pgsql/src/backend/optimizer/path/pathkeys.c,v 1.27 2000/11/12 00:36:58 tgl Exp $ * $Header: /cvsroot/pgsql/src/backend/optimizer/path/pathkeys.c,v 1.28 2000/12/14 22:30:43 tgl Exp $
* *
*------------------------------------------------------------------------- *-------------------------------------------------------------------------
*/ */
@ -195,9 +195,8 @@ generate_implied_equalities(Query *root)
* Given a PathKeyItem, find the equi_key_list subset it is a member of, * Given a PathKeyItem, find the equi_key_list subset it is a member of,
* if any. If so, return a pointer to that sublist, which is the * if any. If so, return a pointer to that sublist, which is the
* canonical representation (for this query) of that PathKeyItem's * canonical representation (for this query) of that PathKeyItem's
* equivalence set. If it is not found, return a single-element list * equivalence set. If it is not found, add a singleton "equivalence set"
* containing the PathKeyItem (when the item has no equivalence peers, * to the equi_key_list and return that --- see compare_pathkeys.
* we just allow it to be a standalone list).
* *
* Note that this function must not be used until after we have completed * Note that this function must not be used until after we have completed
* scanning the WHERE clause for equijoin operators. * scanning the WHERE clause for equijoin operators.
@ -206,6 +205,7 @@ static List *
make_canonical_pathkey(Query *root, PathKeyItem *item) make_canonical_pathkey(Query *root, PathKeyItem *item)
{ {
List *cursetlink; List *cursetlink;
List *newset;
foreach(cursetlink, root->equi_key_list) foreach(cursetlink, root->equi_key_list)
{ {
@ -214,7 +214,9 @@ make_canonical_pathkey(Query *root, PathKeyItem *item)
if (member(item, curset)) if (member(item, curset))
return curset; return curset;
} }
return lcons(item, NIL); newset = makeList1(item);
root->equi_key_list = lcons(newset, root->equi_key_list);
return newset;
} }
/* /*
@ -234,6 +236,7 @@ canonicalize_pathkeys(Query *root, List *pathkeys)
{ {
List *pathkey = (List *) lfirst(i); List *pathkey = (List *) lfirst(i);
PathKeyItem *item; PathKeyItem *item;
List *cpathkey;
/* /*
* It's sufficient to look at the first entry in the sublist; if * It's sufficient to look at the first entry in the sublist; if
@ -242,8 +245,15 @@ canonicalize_pathkeys(Query *root, List *pathkeys)
*/ */
Assert(pathkey != NIL); Assert(pathkey != NIL);
item = (PathKeyItem *) lfirst(pathkey); item = (PathKeyItem *) lfirst(pathkey);
new_pathkeys = lappend(new_pathkeys, cpathkey = make_canonical_pathkey(root, item);
make_canonical_pathkey(root, item)); /*
* Eliminate redundant ordering requests --- ORDER BY A,A
* is the same as ORDER BY A. We want to check this only
* after we have canonicalized the keys, so that equivalent-key
* knowledge is used when deciding if an item is redundant.
*/
if (!ptrMember(cpathkey, new_pathkeys))
new_pathkeys = lappend(new_pathkeys, cpathkey);
} }
return new_pathkeys; return new_pathkeys;
} }
@ -257,19 +267,9 @@ canonicalize_pathkeys(Query *root, List *pathkeys)
* Compare two pathkeys to see if they are equivalent, and if not whether * Compare two pathkeys to see if they are equivalent, and if not whether
* one is "better" than the other. * one is "better" than the other.
* *
* A pathkey can be considered better than another if it is a superset: * This function may only be applied to canonicalized pathkey lists.
* it contains all the keys of the other plus more. For example, either * In the canonical representation, sublists can be checked for equality
* ((A) (B)) or ((A B)) is better than ((A)). * by simple pointer comparison.
*
* Because we actually only expect to see canonicalized pathkey sublists,
* we don't have to do the full two-way-subset-inclusion test on each
* pair of sublists that is implied by the above statement. Instead we
* just do an equal(). In the normal case where multi-element sublists
* are pointers into the root's equi_key_list, equal() will be very fast:
* it will recognize pointer equality when the sublists are the same,
* and will fail at the first sublist element when they are not.
*
* Yes, this gets called enough to be worth coding it this tensely.
*/ */
PathKeysComparison PathKeysComparison
compare_pathkeys(List *keys1, List *keys2) compare_pathkeys(List *keys1, List *keys2)
@ -285,10 +285,70 @@ compare_pathkeys(List *keys1, List *keys2)
List *subkey2 = lfirst(key2); List *subkey2 = lfirst(key2);
/* /*
* We will never have two subkeys where one is a subset of the * XXX would like to check that we've been given canonicalized input,
* other, because of the canonicalization explained above. Either * but query root not accessible here...
* they are equal or they ain't.
*/ */
#ifdef NOT_USED
Assert(ptrMember(subkey1, root->equi_key_list));
Assert(ptrMember(subkey2, root->equi_key_list));
#endif
/*
* We will never have two subkeys where one is a subset of the
* other, because of the canonicalization process. Either they
* are equal or they ain't. Furthermore, we only need pointer
* comparison to detect equality.
*/
if (subkey1 != subkey2)
return PATHKEYS_DIFFERENT; /* no need to keep looking */
}
/*
* If we reached the end of only one list, the other is longer and
* therefore not a subset. (We assume the additional sublist(s) of
* the other list are not NIL --- no pathkey list should ever have a
* NIL sublist.)
*/
if (key1 == NIL && key2 == NIL)
return PATHKEYS_EQUAL;
if (key1 != NIL)
return PATHKEYS_BETTER1;/* key1 is longer */
return PATHKEYS_BETTER2; /* key2 is longer */
}
/*
* compare_noncanonical_pathkeys
* Compare two pathkeys to see if they are equivalent, and if not whether
* one is "better" than the other. This is used when we must compare
* non-canonicalized pathkeys.
*
* A pathkey can be considered better than another if it is a superset:
* it contains all the keys of the other plus more. For example, either
* ((A) (B)) or ((A B)) is better than ((A)).
*
* Currently, the only user of this routine is grouping_planner(),
* and it will only pass single-element sublists (from
* make_pathkeys_for_sortclauses). Therefore we don't have to do the
* full two-way-subset-inclusion test on each pair of sublists that is
* implied by the above statement. Instead we just verify they are
* singleton lists and then do an equal(). This could be improved if
* necessary.
*/
PathKeysComparison
compare_noncanonical_pathkeys(List *keys1, List *keys2)
{
List *key1,
*key2;
for (key1 = keys1, key2 = keys2;
key1 != NIL && key2 != NIL;
key1 = lnext(key1), key2 = lnext(key2))
{
List *subkey1 = lfirst(key1);
List *subkey2 = lfirst(key2);
Assert(length(subkey1) == 1);
Assert(length(subkey2) == 1);
if (!equal(subkey1, subkey2)) if (!equal(subkey1, subkey2))
return PATHKEYS_DIFFERENT; /* no need to keep looking */ return PATHKEYS_DIFFERENT; /* no need to keep looking */
} }
@ -325,6 +385,24 @@ pathkeys_contained_in(List *keys1, List *keys2)
return false; return false;
} }
/*
* noncanonical_pathkeys_contained_in
* The same, when we don't have canonical pathkeys.
*/
bool
noncanonical_pathkeys_contained_in(List *keys1, List *keys2)
{
switch (compare_noncanonical_pathkeys(keys1, keys2))
{
case PATHKEYS_EQUAL:
case PATHKEYS_BETTER2:
return true;
default:
break;
}
return false;
}
/* /*
* get_cheapest_path_for_pathkeys * get_cheapest_path_for_pathkeys
* Find the cheapest path (according to the specified criterion) that * Find the cheapest path (according to the specified criterion) that
@ -464,6 +542,7 @@ build_index_pathkeys(Query *root,
while (*indexkeys != 0 && *ordering != InvalidOid) while (*indexkeys != 0 && *ordering != InvalidOid)
{ {
Var *relvar = find_indexkey_var(root, rel, *indexkeys); Var *relvar = find_indexkey_var(root, rel, *indexkeys);
List *cpathkey;
sortop = *ordering; sortop = *ordering;
if (ScanDirectionIsBackward(scandir)) if (ScanDirectionIsBackward(scandir))
@ -475,8 +554,13 @@ build_index_pathkeys(Query *root,
/* OK, make a sublist for this sort key */ /* OK, make a sublist for this sort key */
item = makePathKeyItem((Node *) relvar, sortop); item = makePathKeyItem((Node *) relvar, sortop);
retval = lappend(retval, make_canonical_pathkey(root, item)); cpathkey = make_canonical_pathkey(root, item);
/*
* Eliminate redundant ordering info; could happen if query
* is such that index keys are equijoined...
*/
if (!ptrMember(cpathkey, retval))
retval = lappend(retval, cpathkey);
indexkeys++; indexkeys++;
ordering++; ordering++;
} }
@ -526,21 +610,20 @@ find_indexkey_var(Query *root, RelOptInfo *rel, AttrNumber varattno)
* outer path (since the join will retain the ordering of the outer path) * outer path (since the join will retain the ordering of the outer path)
* plus any vars of the inner path that are equijoined to the outer vars. * plus any vars of the inner path that are equijoined to the outer vars.
* *
* Per the discussion at the top of this file, equijoined inner vars * Per the discussion in backend/optimizer/README, equijoined inner vars
* can be considered path keys of the result, just the same as the outer * can be considered path keys of the result, just the same as the outer
* vars they were joined with; furthermore, it doesn't matter what kind * vars they were joined with; furthermore, it doesn't matter what kind
* of join algorithm is actually used. * of join algorithm is actually used.
* *
* 'outer_pathkeys' is the list of the outer path's path keys * 'joinrel' is the join relation that paths are being formed for
* 'join_rel_tlist' is the target list of the join relation * 'outer_pathkeys' is the list of the current outer path's path keys
* 'equi_key_list' is the query's list of pathkeyitem equivalence sets
* *
* Returns the list of new path keys. * Returns the list of new path keys.
*/ */
List * List *
build_join_pathkeys(List *outer_pathkeys, build_join_pathkeys(Query *root,
List *join_rel_tlist, RelOptInfo *joinrel,
List *equi_key_list) List *outer_pathkeys)
{ {
/* /*
@ -549,9 +632,11 @@ build_join_pathkeys(List *outer_pathkeys,
* a darn thing here! The inner-rel vars we used to need to add are * a darn thing here! The inner-rel vars we used to need to add are
* *already* part of the outer pathkey! * *already* part of the outer pathkey!
* *
* I'd remove the routine entirely, but maybe someday we'll need it... * We do, however, need to truncate the pathkeys list, since it may
* contain pathkeys that were useful for forming this joinrel but are
* uninteresting to higher levels.
*/ */
return outer_pathkeys; return truncate_useless_pathkeys(root, joinrel, outer_pathkeys);
} }
/**************************************************************************** /****************************************************************************
@ -602,6 +687,39 @@ make_pathkeys_for_sortclauses(List *sortclauses,
* PATHKEYS AND MERGECLAUSES * PATHKEYS AND MERGECLAUSES
****************************************************************************/ ****************************************************************************/
/*
* cache_mergeclause_pathkeys
* Make the cached pathkeys valid in a mergeclause restrictinfo.
*
* RestrictInfo contains fields in which we may cache the result
* of looking up the canonical pathkeys for the left and right sides
* of the mergeclause. (Note that in normal cases they will be the
* same, but not if the mergeclause appears above an OUTER JOIN.)
* This is a worthwhile savings because these routines will be invoked
* many times when dealing with a many-relation query.
*/
static void
cache_mergeclause_pathkeys(Query *root, RestrictInfo *restrictinfo)
{
Node *key;
PathKeyItem *item;
Assert(restrictinfo->mergejoinoperator != InvalidOid);
if (restrictinfo->left_pathkey == NIL)
{
key = (Node *) get_leftop(restrictinfo->clause);
item = makePathKeyItem(key, restrictinfo->left_sortop);
restrictinfo->left_pathkey = make_canonical_pathkey(root, item);
}
if (restrictinfo->right_pathkey == NIL)
{
key = (Node *) get_rightop(restrictinfo->clause);
item = makePathKeyItem(key, restrictinfo->right_sortop);
restrictinfo->right_pathkey = make_canonical_pathkey(root, item);
}
}
/* /*
* find_mergeclauses_for_pathkeys * find_mergeclauses_for_pathkeys
* This routine attempts to find a set of mergeclauses that can be * This routine attempts to find a set of mergeclauses that can be
@ -618,11 +736,13 @@ make_pathkeys_for_sortclauses(List *sortclauses,
* *
* XXX Ideally we ought to be considering context, ie what path orderings * XXX Ideally we ought to be considering context, ie what path orderings
* are available on the other side of the join, rather than just making * are available on the other side of the join, rather than just making
* an arbitrary choice among the mergeclause orders that will work for * an arbitrary choice among the mergeclauses that will work for this side
* this side of the join. * of the join.
*/ */
List * List *
find_mergeclauses_for_pathkeys(List *pathkeys, List *restrictinfos) find_mergeclauses_for_pathkeys(Query *root,
List *pathkeys,
List *restrictinfos)
{ {
List *mergeclauses = NIL; List *mergeclauses = NIL;
List *i; List *i;
@ -634,39 +754,29 @@ find_mergeclauses_for_pathkeys(List *pathkeys, List *restrictinfos)
List *j; List *j;
/* /*
* We can match any of the keys in this pathkey sublist, since * We can match a pathkey against either left or right side of any
* they're all equivalent. And we can match against either left * mergejoin clause we haven't used yet. For the moment we use a
* or right side of any mergejoin clause we haven't used yet. For * dumb "greedy" algorithm with no backtracking. Is it worth being
* the moment we use a dumb "greedy" algorithm with no * any smarter to make a longer list of usable mergeclauses?
* backtracking. Is it worth being any smarter to make a longer * Probably not.
* list of usable mergeclauses? Probably not.
*/ */
foreach(j, pathkey) foreach(j, restrictinfos)
{ {
PathKeyItem *keyitem = lfirst(j); RestrictInfo *restrictinfo = lfirst(j);
Node *key = keyitem->key;
Oid keyop = keyitem->sortop;
List *k;
foreach(k, restrictinfos) cache_mergeclause_pathkeys(root, restrictinfo);
{ /*
RestrictInfo *restrictinfo = lfirst(k); * We can compare canonical pathkey sublists by simple
* pointer equality; see compare_pathkeys.
Assert(restrictinfo->mergejoinoperator != InvalidOid); */
if ((pathkey == restrictinfo->left_pathkey ||
if (((keyop == restrictinfo->left_sortop && pathkey == restrictinfo->right_pathkey) &&
equal(key, get_leftop(restrictinfo->clause))) || !ptrMember(restrictinfo, mergeclauses))
(keyop == restrictinfo->right_sortop &&
equal(key, get_rightop(restrictinfo->clause)))) &&
!member(restrictinfo, mergeclauses))
{ {
matched_restrictinfo = restrictinfo; matched_restrictinfo = restrictinfo;
break; break;
} }
} }
if (matched_restrictinfo)
break;
}
/* /*
* If we didn't find a mergeclause, we're done --- any additional * If we didn't find a mergeclause, we're done --- any additional
@ -715,47 +825,170 @@ make_pathkeys_for_mergeclauses(Query *root,
{ {
RestrictInfo *restrictinfo = (RestrictInfo *) lfirst(i); RestrictInfo *restrictinfo = (RestrictInfo *) lfirst(i);
Node *key; Node *key;
Oid sortop;
PathKeyItem *item;
List *pathkey; List *pathkey;
Assert(restrictinfo->mergejoinoperator != InvalidOid); cache_mergeclause_pathkeys(root, restrictinfo);
/*
* Which key and sortop is needed for this relation?
*/
key = (Node *) get_leftop(restrictinfo->clause); key = (Node *) get_leftop(restrictinfo->clause);
sortop = restrictinfo->left_sortop; if (IsA(key, Var) && intMember(((Var *) key)->varno, rel->relids))
if (!IsA(key, Var) || {
!intMember(((Var *) key)->varno, rel->relids)) /* Rel is left side of mergeclause */
pathkey = restrictinfo->left_pathkey;
}
else
{ {
key = (Node *) get_rightop(restrictinfo->clause); key = (Node *) get_rightop(restrictinfo->clause);
sortop = restrictinfo->right_sortop; if (IsA(key, Var) && intMember(((Var *) key)->varno, rel->relids))
if (!IsA(key, Var) || {
!intMember(((Var *) key)->varno, rel->relids)) /* Rel is right side of mergeclause */
pathkey = restrictinfo->right_pathkey;
}
else
{
elog(ERROR, "make_pathkeys_for_mergeclauses: can't identify which side of mergeclause to use"); elog(ERROR, "make_pathkeys_for_mergeclauses: can't identify which side of mergeclause to use");
pathkey = NIL; /* keep compiler quiet */
}
} }
/* /*
* Find or create canonical pathkey sublist for this sort item. * When we are given multiple merge clauses, it's possible that some
* clauses refer to the same vars as earlier clauses. There's no
* reason for us to specify sort keys like (A,B,A) when (A,B) will
* do --- and adding redundant sort keys makes add_path think that
* this sort order is different from ones that are really the same,
* so don't do it. Since we now have a canonicalized pathkey,
* a simple ptrMember test is sufficient to detect redundant keys.
*/ */
item = makePathKeyItem(key, sortop); if (!ptrMember(pathkey, pathkeys))
pathkey = make_canonical_pathkey(root, item);
/*
* Most of the time we will get back a canonical pathkey set
* including both the mergeclause's left and right sides (the only
* case where we don't is if the mergeclause appeared in an OUTER
* JOIN, which causes us not to generate an equijoin set from it).
* Therefore, most of the time the item we just made is not part
* of the returned structure, and we can free it. This check
* saves a useful amount of storage in a big join tree.
*/
if (item != (PathKeyItem *) lfirst(pathkey))
pfree(item);
pathkeys = lappend(pathkeys, pathkey); pathkeys = lappend(pathkeys, pathkey);
} }
return pathkeys; return pathkeys;
} }
/****************************************************************************
* PATHKEY USEFULNESS CHECKS
*
* We only want to remember as many of the pathkeys of a path as have some
* potential use, either for subsequent mergejoins or for meeting the query's
* requested output ordering. This ensures that add_path() won't consider
* a path to have a usefully different ordering unless it really is useful.
* These routines check for usefulness of given pathkeys.
****************************************************************************/
/*
* pathkeys_useful_for_merging
* Count the number of pathkeys that may be useful for mergejoins
* above the given relation (by looking at its joininfo lists).
*
* We consider a pathkey potentially useful if it corresponds to the merge
* ordering of either side of any joinclause for the rel. This might be
* overoptimistic, since joinclauses that appear in different join lists
* might never be usable at the same time, but trying to be exact is likely
* to be more trouble than it's worth.
*/
int
pathkeys_useful_for_merging(Query *root, RelOptInfo *rel, List *pathkeys)
{
int useful = 0;
List *i;
foreach(i, pathkeys)
{
List *pathkey = lfirst(i);
bool matched = false;
List *j;
foreach(j, rel->joininfo)
{
JoinInfo *joininfo = (JoinInfo *) lfirst(j);
List *k;
foreach(k, joininfo->jinfo_restrictinfo)
{
RestrictInfo *restrictinfo = (RestrictInfo *) lfirst(k);
if (restrictinfo->mergejoinoperator == InvalidOid)
continue;
cache_mergeclause_pathkeys(root, restrictinfo);
/*
* We can compare canonical pathkey sublists by simple
* pointer equality; see compare_pathkeys.
*/
if (pathkey == restrictinfo->left_pathkey ||
pathkey == restrictinfo->right_pathkey)
{
matched = true;
break;
}
}
if (matched)
break;
}
/*
* If we didn't find a mergeclause, we're done --- any additional
* sort-key positions in the pathkeys are useless. (But we can
* still mergejoin if we found at least one mergeclause.)
*/
if (matched)
useful++;
else
break;
}
return useful;
}
/*
* pathkeys_useful_for_ordering
* Count the number of pathkeys that are useful for meeting the
* query's requested output ordering.
*
* Unlike merge pathkeys, this is an all-or-nothing affair: it does us
* no good to order by just the first key(s) of the requested ordering.
* So the result is always either 0 or length(root->query_pathkeys).
*/
int
pathkeys_useful_for_ordering(Query *root, List *pathkeys)
{
if (root->query_pathkeys == NIL)
return 0; /* no special ordering requested */
if (pathkeys == NIL)
return 0; /* unordered path */
if (pathkeys_contained_in(root->query_pathkeys, pathkeys))
{
/* It's useful ... or at least the first N keys are */
return length(root->query_pathkeys);
}
return 0; /* path ordering not useful */
}
/*
* truncate_useless_pathkeys
* Shorten the given pathkey list to just the useful pathkeys.
*/
List *
truncate_useless_pathkeys(Query *root,
RelOptInfo *rel,
List *pathkeys)
{
int nuseful;
int nuseful2;
nuseful = pathkeys_useful_for_merging(root, rel, pathkeys);
nuseful2 = pathkeys_useful_for_ordering(root, pathkeys);
if (nuseful2 > nuseful)
nuseful = nuseful2;
/* Note: not safe to modify input list destructively, but we can avoid
* copying the list if we're not actually going to change it
*/
if (nuseful == length(pathkeys))
return pathkeys;
else
return ltruncate(nuseful, listCopy(pathkeys));
}

6
src/backend/optimizer/plan/initsplan.c
View File

@ -8,7 +8,7 @@
* *
* *
* IDENTIFICATION * IDENTIFICATION
* $Header: /cvsroot/pgsql/src/backend/optimizer/plan/initsplan.c,v 1.54 2000/12/12 23:33:33 tgl Exp $ * $Header: /cvsroot/pgsql/src/backend/optimizer/plan/initsplan.c,v 1.55 2000/12/14 22:30:43 tgl Exp $
* *
*------------------------------------------------------------------------- *-------------------------------------------------------------------------
*/ */
@ -343,7 +343,11 @@ distribute_qual_to_rels(Query *root, Node *clause,
restrictinfo->mergejoinoperator = InvalidOid; restrictinfo->mergejoinoperator = InvalidOid;
restrictinfo->left_sortop = InvalidOid; restrictinfo->left_sortop = InvalidOid;
restrictinfo->right_sortop = InvalidOid; restrictinfo->right_sortop = InvalidOid;
restrictinfo->left_pathkey = NIL; /* not computable yet */
restrictinfo->right_pathkey = NIL;
restrictinfo->hashjoinoperator = InvalidOid; restrictinfo->hashjoinoperator = InvalidOid;
restrictinfo->left_dispersion = -1; /* not computed until needed */
restrictinfo->right_dispersion = -1;
/* /*
* Retrieve all relids and vars contained within the clause. * Retrieve all relids and vars contained within the clause.

11
src/backend/optimizer/plan/planner.c
View File

@ -8,7 +8,7 @@
* *
* *
* IDENTIFICATION * IDENTIFICATION
* $Header: /cvsroot/pgsql/src/backend/optimizer/plan/planner.c,v 1.97 2000/12/06 23:55:17 tgl Exp $ * $Header: /cvsroot/pgsql/src/backend/optimizer/plan/planner.c,v 1.98 2000/12/14 22:30:43 tgl Exp $
* *
*------------------------------------------------------------------------- *-------------------------------------------------------------------------
*/ */
@ -931,13 +931,12 @@ grouping_planner(Query *parse, double tuple_fraction)
* If both GROUP BY and ORDER BY are specified, we will need * If both GROUP BY and ORDER BY are specified, we will need
* two levels of sort --- and, therefore, certainly need to * two levels of sort --- and, therefore, certainly need to
* read all the input tuples --- unless ORDER BY is a subset * read all the input tuples --- unless ORDER BY is a subset
* of GROUP BY. (Although we are comparing non-canonicalized * of GROUP BY. (We have not yet canonicalized the pathkeys,
* pathkeys here, it should be OK since they will both contain * so must use the slower noncanonical comparison method.)
* only single-element sublists at this point. See
* pathkeys.c.)
*/ */
if (parse->groupClause && parse->sortClause && if (parse->groupClause && parse->sortClause &&
!pathkeys_contained_in(sort_pathkeys, group_pathkeys)) !noncanonical_pathkeys_contained_in(sort_pathkeys,
group_pathkeys))
tuple_fraction = 0.0; tuple_fraction = 0.0;
} }
else if (parse->hasAggs) else if (parse->hasAggs)

6
src/backend/optimizer/prep/prepunion.c
View File

@ -14,7 +14,7 @@
* *
* *
* IDENTIFICATION * IDENTIFICATION
* $Header: /cvsroot/pgsql/src/backend/optimizer/prep/prepunion.c,v 1.57 2000/12/12 23:33:34 tgl Exp $ * $Header: /cvsroot/pgsql/src/backend/optimizer/prep/prepunion.c,v 1.58 2000/12/14 22:30:44 tgl Exp $
* *
*------------------------------------------------------------------------- *-------------------------------------------------------------------------
*/ */
@ -667,6 +667,10 @@ adjust_inherited_attrs_mutator(Node *node,
newinfo->subclauseindices = NIL; newinfo->subclauseindices = NIL;
newinfo->eval_cost = -1; /* reset this too */ newinfo->eval_cost = -1; /* reset this too */
newinfo->left_pathkey = NIL; /* and these */
newinfo->right_pathkey = NIL;
newinfo->left_dispersion = -1;
newinfo->right_dispersion = -1;
return (Node *) newinfo; return (Node *) newinfo;
} }

67
src/backend/optimizer/util/pathnode.c
View File

@ -8,7 +8,7 @@
* *
* *
* IDENTIFICATION * IDENTIFICATION
* $Header: /cvsroot/pgsql/src/backend/optimizer/util/pathnode.c,v 1.68 2000/11/12 00:36:59 tgl Exp $ * $Header: /cvsroot/pgsql/src/backend/optimizer/util/pathnode.c,v 1.69 2000/12/14 22:30:44 tgl Exp $
* *
*------------------------------------------------------------------------- *-------------------------------------------------------------------------
*/ */
@ -161,11 +161,17 @@ set_cheapest(RelOptInfo *parent_rel)
* pathlist any old paths that are dominated by new_path --- that is, * pathlist any old paths that are dominated by new_path --- that is,
* new_path is both cheaper and at least as well ordered. * new_path is both cheaper and at least as well ordered.
* *
* The pathlist is kept sorted by TOTAL_COST metric, with cheaper paths
* at the front. No code depends on that for correctness; it's simply
* a speed hack within this routine. Doing it that way makes it more
* likely that we will reject an inferior path after a few comparisons,
* rather than many comparisons.
*
* NOTE: discarded Path objects are immediately pfree'd to reduce planner * NOTE: discarded Path objects are immediately pfree'd to reduce planner
* memory consumption. We dare not try to free the substructure of a Path, * memory consumption. We dare not try to free the substructure of a Path,
* since much of it may be shared with other Paths or the query tree itself; * since much of it may be shared with other Paths or the query tree itself;
* but just recycling discarded Path nodes is a very useful savings in * but just recycling discarded Path nodes is a very useful savings in
* a large join tree. * a large join tree. We can recycle the List nodes of pathlist, too.
* *
* 'parent_rel' is the relation entry to which the path corresponds. * 'parent_rel' is the relation entry to which the path corresponds.
* 'new_path' is a potential path for parent_rel. * 'new_path' is a potential path for parent_rel.
@ -177,6 +183,7 @@ add_path(RelOptInfo *parent_rel, Path *new_path)
{ {
bool accept_new = true; /* unless we find a superior old bool accept_new = true; /* unless we find a superior old
* path */ * path */
List *insert_after = NIL; /* where to insert new item */
List *p1_prev = NIL; List *p1_prev = NIL;
List *p1; List *p1;
@ -185,7 +192,8 @@ add_path(RelOptInfo *parent_rel, Path *new_path)
* possible for more than one old path to be tossed out because * possible for more than one old path to be tossed out because
* new_path dominates it. * new_path dominates it.
*/ */
foreach(p1, parent_rel->pathlist) p1 = parent_rel->pathlist; /* cannot use foreach here */
while (p1 != NIL)
{ {
Path *old_path = (Path *) lfirst(p1); Path *old_path = (Path *) lfirst(p1);
bool remove_old = false; /* unless new proves superior */ bool remove_old = false; /* unless new proves superior */
@ -197,13 +205,14 @@ add_path(RelOptInfo *parent_rel, Path *new_path)
* If the two paths compare differently for startup and total * If the two paths compare differently for startup and total
* cost, then we want to keep both, and we can skip the (much * cost, then we want to keep both, and we can skip the (much
* slower) comparison of pathkeys. If they compare the same, * slower) comparison of pathkeys. If they compare the same,
* proceed with the pathkeys comparison. Note this test relies on * proceed with the pathkeys comparison. Note: this test relies
* the fact that compare_path_costs will only return 0 if both * on the fact that compare_path_costs will only return 0 if both
* costs are equal (and, therefore, there's no need to call it * costs are equal (and, therefore, there's no need to call it
* twice in that case). * twice in that case).
*/ */
if (costcmp == 0 || if (costcmp == 0 ||
costcmp == compare_path_costs(new_path, old_path, STARTUP_COST)) costcmp == compare_path_costs(new_path, old_path,
STARTUP_COST))
{ {
switch (compare_pathkeys(new_path->pathkeys, old_path->pathkeys)) switch (compare_pathkeys(new_path->pathkeys, old_path->pathkeys))
{ {
@ -234,14 +243,24 @@ add_path(RelOptInfo *parent_rel, Path *new_path)
*/ */
if (remove_old && parent_rel->pruneable) if (remove_old && parent_rel->pruneable)
{ {
List *p1_next = lnext(p1);
if (p1_prev) if (p1_prev)
lnext(p1_prev) = lnext(p1); lnext(p1_prev) = p1_next;
else else
parent_rel->pathlist = lnext(p1); parent_rel->pathlist = p1_next;
pfree(old_path); pfree(old_path);
pfree(p1); /* this is why we can't use foreach */
p1 = p1_next;
} }
else else
{
/* new belongs after this old path if it has cost >= old's */
if (costcmp >= 0)
insert_after = p1;
p1_prev = p1; p1_prev = p1;
p1 = lnext(p1);
}
/* /*
* If we found an old path that dominates new_path, we can quit * If we found an old path that dominates new_path, we can quit
@ -254,12 +273,15 @@ add_path(RelOptInfo *parent_rel, Path *new_path)
if (accept_new) if (accept_new)
{ {
/* Accept the path */ /* Accept the new path: insert it at proper place in pathlist */
if (insert_after)
lnext(insert_after) = lcons(new_path, lnext(insert_after));
else
parent_rel->pathlist = lcons(new_path, parent_rel->pathlist); parent_rel->pathlist = lcons(new_path, parent_rel->pathlist);
} }
else else
{ {
/* Reject and recycle the path */ /* Reject and recycle the new path */
pfree(new_path); pfree(new_path);
} }
} }
@ -296,9 +318,9 @@ create_seqscan_path(RelOptInfo *rel)
* 'index' is an index on 'rel' * 'index' is an index on 'rel'
* 'restriction_clauses' is a list of RestrictInfo nodes * 'restriction_clauses' is a list of RestrictInfo nodes
* to be used as index qual conditions in the scan. * to be used as index qual conditions in the scan.
* 'pathkeys' describes the ordering of the path.
* 'indexscandir' is ForwardScanDirection or BackwardScanDirection * 'indexscandir' is ForwardScanDirection or BackwardScanDirection
* if the caller expects a specific scan direction, * for an ordered index, or NoMovementScanDirection for
* or NoMovementScanDirection if the caller is willing to accept
* an unordered index. * an unordered index.
* *
* Returns the new path node. * Returns the new path node.
@ -308,6 +330,7 @@ create_index_path(Query *root,
RelOptInfo *rel, RelOptInfo *rel,
IndexOptInfo *index, IndexOptInfo *index,
List *restriction_clauses, List *restriction_clauses,
List *pathkeys,
ScanDirection indexscandir) ScanDirection indexscandir)
{ {
IndexPath *pathnode = makeNode(IndexPath); IndexPath *pathnode = makeNode(IndexPath);
@ -315,25 +338,7 @@ create_index_path(Query *root,
pathnode->path.pathtype = T_IndexScan; pathnode->path.pathtype = T_IndexScan;
pathnode->path.parent = rel; pathnode->path.parent = rel;
pathnode->path.pathkeys = pathkeys;
pathnode->path.pathkeys = build_index_pathkeys(root, rel, index,
indexscandir);
if (pathnode->path.pathkeys == NIL)
{
/* No ordering available from index, is that OK? */
if (!ScanDirectionIsNoMovement(indexscandir))
elog(ERROR, "create_index_path: failed to create ordered index scan");
}
else
{
/*
* The index is ordered, and build_index_pathkeys defaulted to
* forward scan, so make sure we mark the pathnode properly.
*/
if (ScanDirectionIsNoMovement(indexscandir))
indexscandir = ForwardScanDirection;
}
indexquals = get_actual_clauses(restriction_clauses); indexquals = get_actual_clauses(restriction_clauses);
/* expand special operators to indexquals the executor can handle */ /* expand special operators to indexquals the executor can handle */

10
src/include/nodes/relation.h
View File

@ -7,7 +7,7 @@
* Portions Copyright (c) 1996-2000, PostgreSQL, Inc * Portions Copyright (c) 1996-2000, PostgreSQL, Inc
* Portions Copyright (c) 1994, Regents of the University of California * Portions Copyright (c) 1994, Regents of the University of California
* *
* $Id: relation.h,v 1.51 2000/12/12 23:33:32 tgl Exp $ * $Id: relation.h,v 1.52 2000/12/14 22:30:44 tgl Exp $
* *
*------------------------------------------------------------------------- *-------------------------------------------------------------------------
*/ */
@ -504,8 +504,16 @@ typedef struct RestrictInfo
Oid left_sortop; /* leftside sortop needed for mergejoin */ Oid left_sortop; /* leftside sortop needed for mergejoin */
Oid right_sortop; /* rightside sortop needed for mergejoin */ Oid right_sortop; /* rightside sortop needed for mergejoin */
/* 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 */
/* valid if clause is hashjoinable, else InvalidOid: */ /* valid if clause is hashjoinable, else InvalidOid: */
Oid hashjoinoperator; /* copy of clause operator */ Oid hashjoinoperator; /* copy of clause operator */
/* cache space for hashclause processing; -1 if not yet set */
Selectivity left_dispersion; /* dispersion of left side */
Selectivity right_dispersion; /* dispersion of right side */
} RestrictInfo; } RestrictInfo;
/* /*

3
src/include/optimizer/pathnode.h
View File

@ -7,7 +7,7 @@
* Portions Copyright (c) 1996-2000, PostgreSQL, Inc * Portions Copyright (c) 1996-2000, PostgreSQL, Inc
* Portions Copyright (c) 1994, Regents of the University of California * Portions Copyright (c) 1994, Regents of the University of California
* *
* $Id: pathnode.h,v 1.31 2000/11/12 00:37:01 tgl Exp $ * $Id: pathnode.h,v 1.32 2000/12/14 22:30:45 tgl Exp $
* *
*------------------------------------------------------------------------- *-------------------------------------------------------------------------
*/ */
@ -30,6 +30,7 @@ extern Path *create_seqscan_path(RelOptInfo *rel);
extern IndexPath *create_index_path(Query *root, RelOptInfo *rel, extern IndexPath *create_index_path(Query *root, RelOptInfo *rel,
IndexOptInfo *index, IndexOptInfo *index,
List *restriction_clauses, List *restriction_clauses,
List *pathkeys,
ScanDirection indexscandir); ScanDirection indexscandir);
extern TidPath *create_tidscan_path(RelOptInfo *rel, List *tideval); extern TidPath *create_tidscan_path(RelOptInfo *rel, List *tideval);
extern AppendPath *create_append_path(RelOptInfo *rel, List *subpaths); extern AppendPath *create_append_path(RelOptInfo *rel, List *subpaths);

25
src/include/optimizer/paths.h
View File

@ -8,7 +8,7 @@
* Portions Copyright (c) 1996-2000, PostgreSQL, Inc * Portions Copyright (c) 1996-2000, PostgreSQL, Inc
* Portions Copyright (c) 1994, Regents of the University of California * Portions Copyright (c) 1994, Regents of the University of California
* *
* $Id: paths.h,v 1.48 2000/09/29 18:21:40 tgl Exp $ * $Id: paths.h,v 1.49 2000/12/14 22:30:45 tgl Exp $
* *
*------------------------------------------------------------------------- *-------------------------------------------------------------------------
*/ */
@ -34,9 +34,7 @@ extern RelOptInfo *make_fromexpr_rel(Query *root, FromExpr *from);
* indxpath.c * indxpath.c
* routines to generate index paths * routines to generate index paths
*/ */
extern void create_index_paths(Query *root, RelOptInfo *rel, List *indices, extern void create_index_paths(Query *root, RelOptInfo *rel, List *indices);
List *restrictinfo_list,
List *joininfo_list);
extern Oid indexable_operator(Expr *clause, Oid opclass, Oid relam, extern Oid indexable_operator(Expr *clause, Oid opclass, Oid relam,
bool indexkey_on_left); bool indexkey_on_left);
extern List *extract_or_indexqual_conditions(RelOptInfo *rel, extern List *extract_or_indexqual_conditions(RelOptInfo *rel,
@ -97,6 +95,9 @@ extern void generate_implied_equalities(Query *root);
extern List *canonicalize_pathkeys(Query *root, List *pathkeys); extern List *canonicalize_pathkeys(Query *root, List *pathkeys);
extern PathKeysComparison compare_pathkeys(List *keys1, List *keys2); extern PathKeysComparison compare_pathkeys(List *keys1, List *keys2);
extern bool pathkeys_contained_in(List *keys1, List *keys2); extern bool pathkeys_contained_in(List *keys1, List *keys2);
extern PathKeysComparison compare_noncanonical_pathkeys(List *keys1,
List *keys2);
extern bool noncanonical_pathkeys_contained_in(List *keys1, List *keys2);
extern Path *get_cheapest_path_for_pathkeys(List *paths, List *pathkeys, extern Path *get_cheapest_path_for_pathkeys(List *paths, List *pathkeys,
CostSelector cost_criterion); CostSelector cost_criterion);
extern Path *get_cheapest_fractional_path_for_pathkeys(List *paths, extern Path *get_cheapest_fractional_path_for_pathkeys(List *paths,
@ -105,15 +106,23 @@ extern Path *get_cheapest_fractional_path_for_pathkeys(List *paths,
extern List *build_index_pathkeys(Query *root, RelOptInfo *rel, extern List *build_index_pathkeys(Query *root, RelOptInfo *rel,
IndexOptInfo *index, IndexOptInfo *index,
ScanDirection scandir); ScanDirection scandir);
extern List *build_join_pathkeys(List *outer_pathkeys, extern List *build_join_pathkeys(Query *root,
List *join_rel_tlist, RelOptInfo *joinrel,
List *equi_key_list); List *outer_pathkeys);
extern List *make_pathkeys_for_sortclauses(List *sortclauses, extern List *make_pathkeys_for_sortclauses(List *sortclauses,
List *tlist); List *tlist);
extern List *find_mergeclauses_for_pathkeys(List *pathkeys, extern List *find_mergeclauses_for_pathkeys(Query *root,
List *pathkeys,
List *restrictinfos); List *restrictinfos);
extern List *make_pathkeys_for_mergeclauses(Query *root, extern List *make_pathkeys_for_mergeclauses(Query *root,
List *mergeclauses, List *mergeclauses,
RelOptInfo *rel); RelOptInfo *rel);
extern int pathkeys_useful_for_merging(Query *root,
RelOptInfo *rel,
List *pathkeys);
extern int pathkeys_useful_for_ordering(Query *root, List *pathkeys);
extern List *truncate_useless_pathkeys(Query *root,
RelOptInfo *rel,
List *pathkeys);
#endif /* PATHS_H */ #endif /* PATHS_H */

8
src/test/regress/expected/join.out
View File

@ -1796,13 +1796,13 @@ SELECT '' AS "xxx", *
| 4 | 1 | four | | 4 | 1 | four |
| 5 | 0 | five | -5 | 5 | 0 | five | -5
| 5 | 0 | five | -5 | 5 | 0 | five | -5
| | | |
| | | | 0
| 6 | 6 | six | | 6 | 6 | six |
| 7 | 7 | seven | | 7 | 7 | seven |
| 8 | 8 | eight | | 8 | 8 | eight |
| | | null | | | | null |
| | 0 | zero | | | 0 | zero |
| | | |
| | | | 0
(15 rows) (15 rows)
SELECT '' AS "xxx", * SELECT '' AS "xxx", *
@ -1817,13 +1817,13 @@ SELECT '' AS "xxx", *
| 4 | 1 | four | | 4 | 1 | four |
| 5 | 0 | five | -5 | 5 | 0 | five | -5
| 5 | 0 | five | -5 | 5 | 0 | five | -5
| | | |
| | | | 0
| 6 | 6 | six | | 6 | 6 | six |
| 7 | 7 | seven | | 7 | 7 | seven |
| 8 | 8 | eight | | 8 | 8 | eight |
| | | null | | | | null |
| | 0 | zero | | | 0 | zero |
| | | |
| | | | 0
(15 rows) (15 rows)
SELECT '' AS "xxx", * SELECT '' AS "xxx", *