* planner.c
* The query optimizer external interface.
*
- * Portions Copyright (c) 1996-2004, PostgreSQL Global Development Group
+ * 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/plan/planner.c,v 1.173 2004/08/29 04:12:33 momjian Exp $
+ * $PostgreSQL: pgsql/src/backend/optimizer/plan/planner.c,v 1.184 2005/04/11 23:06:55 tgl Exp $
*
*-------------------------------------------------------------------------
*/
#include "catalog/pg_operator.h"
#include "catalog/pg_type.h"
#include "executor/executor.h"
+#include "executor/nodeAgg.h"
#include "miscadmin.h"
#include "nodes/makefuncs.h"
#ifdef OPTIMIZER_DEBUG
#include "optimizer/subselect.h"
#include "optimizer/tlist.h"
#include "optimizer/var.h"
-#include "parser/analyze.h"
#include "parser/parsetree.h"
#include "parser/parse_expr.h"
#include "parser/parse_oper.h"
#include "utils/syscache.h"
-ParamListInfo PlannerBoundParamList = NULL; /* current boundParams */
+ParamListInfo PlannerBoundParamList = NULL; /* current boundParams */
/* Expression kind codes for preprocess_expression */
static void preprocess_qual_conditions(Query *parse, Node *jtnode);
static Plan *inheritance_planner(Query *parse, List *inheritlist);
static Plan *grouping_planner(Query *parse, double tuple_fraction);
+static bool choose_hashed_grouping(Query *parse, double tuple_fraction,
+ Path *cheapest_path, Path *sorted_path,
+ List *sort_pathkeys, List *group_pathkeys,
+ double dNumGroups, AggClauseCounts *agg_counts);
static bool hash_safe_grouping(Query *parse);
static List *make_subplanTargetList(Query *parse, List *tlist,
AttrNumber **groupColIdx, bool *need_tlist_eval);
* eval_const_expressions tries to pre-evaluate an SQL function). So,
* these global state variables must be saved and restored.
*
- * Query level and the param list cannot be moved into the Query structure
- * since their whole purpose is communication across multiple sub-Queries.
- * Also, boundParams is explicitly info from outside the Query, and so
- * is likewise better handled as a global variable.
+ * Query level and the param list cannot be moved into the Query
+ * structure since their whole purpose is communication across
+ * multiple sub-Queries. Also, boundParams is explicitly info from
+ * outside the Query, and so is likewise better handled as a global
+ * variable.
*
* Note we do NOT save and restore PlannerPlanId: it exists to assign
* unique IDs to SubPlan nodes, and we want those IDs to be unique for
}
}
+ /*
+ * Set hasHavingQual to remember if HAVING clause is present. Needed
+ * because preprocess_expression will reduce a constant-true condition
+ * to an empty qual list ... but "HAVING TRUE" is not a semantic no-op.
+ */
+ parse->hasHavingQual = (parse->havingQual != NULL);
+
/*
* Do expression preprocessing on targetlist and quals.
*/
}
/*
- * A HAVING clause without aggregates is equivalent to a WHERE clause
- * (except it can only refer to grouped fields). Transfer any
- * agg-free clauses of the HAVING qual into WHERE. This may seem like
- * wasting cycles to cater to stupidly-written queries, but there are
- * other reasons for doing it. Firstly, if the query contains no aggs
- * at all, then we aren't going to generate an Agg plan node, and so
- * there'll be no place to execute HAVING conditions; without this
- * transfer, we'd lose the HAVING condition entirely, which is wrong.
- * Secondly, when we push down a qual condition into a sub-query, it's
- * easiest to push the qual into HAVING always, in case it contains
- * aggs, and then let this code sort it out.
+ * In some cases we may want to transfer a HAVING clause into WHERE.
+ * We cannot do so if the HAVING clause contains aggregates (obviously)
+ * or volatile functions (since a HAVING clause is supposed to be executed
+ * only once per group). Also, it may be that the clause is so expensive
+ * to execute that we're better off doing it only once per group, despite
+ * the loss of selectivity. This is hard to estimate short of doing the
+ * entire planning process twice, so we use a heuristic: clauses
+ * containing subplans are left in HAVING. Otherwise, we move or copy
+ * the HAVING clause into WHERE, in hopes of eliminating tuples before
+ * aggregation instead of after.
+ *
+ * If the query has explicit grouping then we can simply move such a
+ * clause into WHERE; any group that fails the clause will not be
+ * in the output because none of its tuples will reach the grouping
+ * or aggregation stage. Otherwise we must have a degenerate
+ * (variable-free) HAVING clause, which we put in WHERE so that
+ * query_planner() can use it in a gating Result node, but also keep
+ * in HAVING to ensure that we don't emit a bogus aggregated row.
+ * (This could be done better, but it seems not worth optimizing.)
*
* Note that both havingQual and parse->jointree->quals are in
* implicitly-ANDed-list form at this point, even though they are
{
Node *havingclause = (Node *) lfirst(l);
- if (contain_agg_clause(havingclause))
+ if (contain_agg_clause(havingclause) ||
+ contain_volatile_functions(havingclause) ||
+ contain_subplans(havingclause))
+ {
+ /* keep it in HAVING */
newHaving = lappend(newHaving, havingclause);
- else
+ }
+ else if (parse->groupClause)
+ {
+ /* move it to WHERE */
parse->jointree->quals = (Node *)
lappend((List *) parse->jointree->quals, havingclause);
+ }
+ else
+ {
+ /* put a copy in WHERE, keep it in HAVING */
+ parse->jointree->quals = (Node *)
+ lappend((List *) parse->jointree->quals,
+ copyObject(havingclause));
+ newHaving = lappend(newHaving, havingclause);
+ }
}
parse->havingQual = (Node *) newHaving;
* grouping_planner.
*/
if (parse->resultRelation &&
- (lst = expand_inherited_rtentry(parse, parse->resultRelation,
- false)) != NIL)
+ (lst = expand_inherited_rtentry(parse, parse->resultRelation)) != NIL)
plan = inheritance_planner(parse, lst);
else
plan = grouping_planner(parse, tuple_fraction);
/*
* If any subplans were generated, or if we're inside a subplan, build
- * initPlan list and extParam/allParam sets for plan nodes.
+ * initPlan list and extParam/allParam sets for plan nodes, and attach
+ * the initPlans to the top plan node.
*/
if (PlannerPlanId != saved_planid || PlannerQueryLevel > 1)
- {
- Cost initplan_cost = 0;
-
- /* Prepare extParam/allParam sets for all nodes in tree */
SS_finalize_plan(plan, parse->rtable);
- /*
- * SS_finalize_plan doesn't handle initPlans, so we have to
- * manually attach them to the topmost plan node, and add their
- * extParams to the topmost node's, too.
- *
- * We also add the total_cost of each initPlan to the startup cost of
- * the top node. This is a conservative overestimate, since in
- * fact each initPlan might be executed later than plan startup,
- * or even not at all.
- */
- plan->initPlan = PlannerInitPlan;
-
- foreach(l, plan->initPlan)
- {
- SubPlan *initplan = (SubPlan *) lfirst(l);
-
- plan->extParam = bms_add_members(plan->extParam,
- initplan->plan->extParam);
- /* allParam must include all members of extParam */
- plan->allParam = bms_add_members(plan->allParam,
- plan->extParam);
- initplan_cost += initplan->plan->total_cost;
- }
-
- plan->startup_cost += initplan_cost;
- plan->total_cost += initplan_cost;
- }
-
/* Return to outer subquery context */
PlannerQueryLevel--;
PlannerInitPlan = saved_initplan;
expr = flatten_join_alias_vars(parse, expr);
/*
- * If it's a qual or havingQual, canonicalize it. It seems most useful
- * to do this before applying eval_const_expressions, since the latter
- * can optimize flattened AND/ORs better than unflattened ones.
+ * Simplify constant expressions.
*
- * Note: all processing of a qual expression after this point must be
- * careful to maintain AND/OR flatness --- that is, do not generate a
- * tree with AND directly under AND, nor OR directly under OR.
+ * Note: this also flattens nested AND and OR expressions into N-argument
+ * form. All processing of a qual expression after this point must be
+ * careful to maintain AND/OR flatness --- that is, do not generate a tree
+ * with AND directly under AND, nor OR directly under OR.
+ */
+ expr = eval_const_expressions(expr);
+
+ /*
+ * If it's a qual or havingQual, canonicalize it.
*/
if (kind == EXPRKIND_QUAL)
{
#endif
}
- /*
- * Simplify constant expressions.
- */
- expr = eval_const_expressions(expr);
-
/* Expand SubLinks to SubPlans */
if (parse->hasSubLinks)
expr = SS_process_sublinks(expr, (kind == EXPRKIND_QUAL));
/*
* If it's a qual or havingQual, convert it to implicit-AND format.
* (We don't want to do this before eval_const_expressions, since the
- * latter would be unable to simplify a top-level AND correctly. Also,
+ * latter would be unable to simplify a top-level AND correctly. Also,
* SS_process_sublinks expects explicit-AND format.)
*/
if (kind == EXPRKIND_QUAL)
{
int childRTindex = lfirst_int(l);
Oid childOID = getrelid(childRTindex, parse->rtable);
- int subrtlength;
Query *subquery;
Plan *subplan;
subplans = lappend(subplans, subplan);
/*
- * It's possible that additional RTEs got added to the rangetable
- * due to expansion of inherited source tables (see allpaths.c).
- * If so, we must copy 'em back to the main parse tree's rtable.
+ * XXX my goodness this next bit is ugly. Really need to think about
+ * ways to rein in planner's habit of scribbling on its input.
*
- * XXX my goodness this is ugly. Really need to think about ways to
- * rein in planner's habit of scribbling on its input.
+ * Planning of the subquery might have modified the rangetable,
+ * either by addition of RTEs due to expansion of inherited source
+ * tables, or by changes of the Query structures inside subquery
+ * RTEs. We have to ensure that this gets propagated back to the
+ * master copy. However, if we aren't done planning yet, we also
+ * need to ensure that subsequent calls to grouping_planner have
+ * virgin sub-Queries to work from. So, if we are at the last
+ * list entry, just copy the subquery rangetable back to the master
+ * copy; if we are not, then extend the master copy by adding
+ * whatever the subquery added. (We assume these added entries
+ * will go untouched by the future grouping_planner calls. We are
+ * also effectively assuming that sub-Queries will get planned
+ * identically each time, or at least that the impacts on their
+ * rangetables will be the same each time. Did I say this is ugly?)
*/
- subrtlength = list_length(subquery->rtable);
- if (subrtlength > mainrtlength)
+ if (lnext(l) == NULL)
+ parse->rtable = subquery->rtable;
+ else
{
- List *subrt;
+ int subrtlength = list_length(subquery->rtable);
+
+ if (subrtlength > mainrtlength)
+ {
+ List *subrt;
- subrt = list_copy_tail(subquery->rtable, mainrtlength);
- parse->rtable = list_concat(parse->rtable, subrt);
- mainrtlength = subrtlength;
+ subrt = list_copy_tail(subquery->rtable, mainrtlength);
+ parse->rtable = list_concat(parse->rtable, subrt);
+ mainrtlength = subrtlength;
+ }
}
+
/* Save preprocessed tlist from first rel for use in Append */
if (tlist == NIL)
tlist = subplan->targetlist;
if (parse->setOperations)
{
- List *set_sortclauses;
+ List *set_sortclauses;
/*
* Construct the plan for set operations. The result will not
* the sort key information...
*/
current_pathkeys = make_pathkeys_for_sortclauses(set_sortclauses,
- result_plan->targetlist);
+ result_plan->targetlist);
current_pathkeys = canonicalize_pathkeys(parse, current_pathkeys);
/*
double sub_tuple_fraction;
Path *cheapest_path;
Path *sorted_path;
+ Path *best_path;
double dNumGroups = 0;
long numGroups = 0;
- int numAggs = 0;
+ AggClauseCounts agg_counts;
int numGroupCols = list_length(parse->groupClause);
bool use_hashed_grouping = false;
- /* Preprocess targetlist in case we are inside an INSERT/UPDATE. */
- tlist = preprocess_targetlist(tlist,
- parse->commandType,
- parse->resultRelation,
- parse->rtable);
+ MemSet(&agg_counts, 0, sizeof(AggClauseCounts));
- /*
- * Add TID targets for rels selected FOR UPDATE (should this be
- * done in preprocess_targetlist?). The executor uses the TID to
- * know which rows to lock, much as for UPDATE or DELETE.
- */
- if (parse->rowMarks)
- {
- ListCell *l;
-
- /*
- * We've got trouble if the FOR UPDATE appears inside
- * grouping, since grouping renders a reference to individual
- * tuple CTIDs invalid. This is also checked at parse time,
- * but that's insufficient because of rule substitution, query
- * pullup, etc.
- */
- CheckSelectForUpdate(parse);
-
- /*
- * Currently the executor only supports FOR UPDATE at top
- * level
- */
- if (PlannerQueryLevel > 1)
- ereport(ERROR,
- (errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
- errmsg("SELECT FOR UPDATE is not allowed in subqueries")));
-
- foreach(l, parse->rowMarks)
- {
- Index rti = lfirst_int(l);
- char *resname;
- Resdom *resdom;
- Var *var;
- TargetEntry *ctid;
-
- resname = (char *) palloc(32);
- snprintf(resname, 32, "ctid%u", rti);
- resdom = makeResdom(list_length(tlist) + 1,
- TIDOID,
- -1,
- resname,
- true);
-
- var = makeVar(rti,
- SelfItemPointerAttributeNumber,
- TIDOID,
- -1,
- 0);
-
- ctid = makeTargetEntry(resdom, (Expr *) var);
- tlist = lappend(tlist, ctid);
- }
- }
+ /* Preprocess targetlist */
+ tlist = preprocess_targetlist(parse, tlist);
/*
* Generate appropriate target list for subplan; may be different
*
* Note: think not that we can turn off hasAggs if we find no aggs.
* It is possible for constant-expression simplification to remove
- * all explicit references to aggs, but we still have to follow the
- * aggregate semantics (eg, producing only one output row).
+ * all explicit references to aggs, but we still have to follow
+ * the aggregate semantics (eg, producing only one output row).
*/
if (parse->hasAggs)
- numAggs = count_agg_clause((Node *) tlist) +
- count_agg_clause(parse->havingQual);
+ {
+ count_agg_clauses((Node *) tlist, &agg_counts);
+ count_agg_clauses(parse->havingQual, &agg_counts);
+ }
/*
* Figure out whether we need a sorted result from query_planner.
sort_pathkeys = canonicalize_pathkeys(parse, sort_pathkeys);
/*
- * Consider whether we might want to use hashed grouping.
+ * If grouping, estimate the number of groups. (We can't do this
+ * until after running query_planner(), either.) Then decide
+ * whether we want to use hashed grouping.
*/
if (parse->groupClause)
{
List *groupExprs;
double cheapest_path_rows;
- int cheapest_path_width;
/*
- * Beware in this section of the possibility that
- * cheapest_path->parent is NULL. This could happen if user
- * does something silly like SELECT 'foo' GROUP BY 1;
+ * Beware of the possibility that cheapest_path->parent is NULL.
+ * This could happen if user does something silly like
+ * SELECT 'foo' GROUP BY 1;
*/
if (cheapest_path->parent)
- {
cheapest_path_rows = cheapest_path->parent->rows;
- cheapest_path_width = cheapest_path->parent->width;
- }
else
- {
cheapest_path_rows = 1; /* assume non-set result */
- cheapest_path_width = 100; /* arbitrary */
- }
- /*
- * Always estimate the number of groups. We can't do this
- * until after running query_planner(), either.
- */
groupExprs = get_sortgrouplist_exprs(parse->groupClause,
parse->targetList);
dNumGroups = estimate_num_groups(parse,
/* Also want it as a long int --- but 'ware overflow! */
numGroups = (long) Min(dNumGroups, (double) LONG_MAX);
- /*
- * Check can't-do-it conditions, including whether the
- * grouping operators are hashjoinable.
- *
- * Executor doesn't support hashed aggregation with DISTINCT
- * aggregates. (Doing so would imply storing *all* the input
- * values in the hash table, which seems like a certain
- * loser.)
- */
- if (!enable_hashagg || !hash_safe_grouping(parse))
- use_hashed_grouping = false;
- else if (parse->hasAggs &&
- (contain_distinct_agg_clause((Node *) tlist) ||
- contain_distinct_agg_clause(parse->havingQual)))
- use_hashed_grouping = false;
- else
- {
- /*
- * Use hashed grouping if (a) we think we can fit the
- * hashtable into work_mem, *and* (b) the estimated cost is
- * no more than doing it the other way. While avoiding
- * the need for sorted input is usually a win, the fact
- * that the output won't be sorted may be a loss; so we
- * need to do an actual cost comparison.
- *
- * In most cases we have no good way to estimate the size of
- * the transition value needed by an aggregate;
- * arbitrarily assume it is 100 bytes. Also set the
- * overhead per hashtable entry at 64 bytes.
- */
- int hashentrysize = cheapest_path_width + 64 + numAggs * 100;
-
- if (hashentrysize * dNumGroups <= work_mem * 1024L)
- {
- /*
- * Okay, do the cost comparison. We need to consider
- * cheapest_path + hashagg [+ final sort] versus
- * either cheapest_path [+ sort] + group or agg [+
- * final sort] or presorted_path + group or agg [+
- * final sort] where brackets indicate a step that may
- * not be needed. We assume query_planner() will have
- * returned a presorted path only if it's a winner
- * compared to cheapest_path for this purpose.
- *
- * These path variables are dummies that just hold cost
- * fields; we don't make actual Paths for these steps.
- */
- Path hashed_p;
- Path sorted_p;
-
- cost_agg(&hashed_p, parse,
- AGG_HASHED, numAggs,
- numGroupCols, dNumGroups,
- cheapest_path->startup_cost,
- cheapest_path->total_cost,
- cheapest_path_rows);
- /* Result of hashed agg is always unsorted */
- if (sort_pathkeys)
- cost_sort(&hashed_p, parse, sort_pathkeys,
- hashed_p.total_cost,
- dNumGroups,
- cheapest_path_width);
-
- if (sorted_path)
- {
- sorted_p.startup_cost = sorted_path->startup_cost;
- sorted_p.total_cost = sorted_path->total_cost;
- current_pathkeys = sorted_path->pathkeys;
- }
- else
- {
- sorted_p.startup_cost = cheapest_path->startup_cost;
- sorted_p.total_cost = cheapest_path->total_cost;
- current_pathkeys = cheapest_path->pathkeys;
- }
- if (!pathkeys_contained_in(group_pathkeys,
- current_pathkeys))
- {
- cost_sort(&sorted_p, parse, group_pathkeys,
- sorted_p.total_cost,
- cheapest_path_rows,
- cheapest_path_width);
- current_pathkeys = group_pathkeys;
- }
- if (parse->hasAggs)
- cost_agg(&sorted_p, parse,
- AGG_SORTED, numAggs,
- numGroupCols, dNumGroups,
- sorted_p.startup_cost,
- sorted_p.total_cost,
- cheapest_path_rows);
- else
- cost_group(&sorted_p, parse,
- numGroupCols, dNumGroups,
- sorted_p.startup_cost,
- sorted_p.total_cost,
- cheapest_path_rows);
- /* The Agg or Group node will preserve ordering */
- if (sort_pathkeys &&
- !pathkeys_contained_in(sort_pathkeys,
- current_pathkeys))
- {
- cost_sort(&sorted_p, parse, sort_pathkeys,
- sorted_p.total_cost,
- dNumGroups,
- cheapest_path_width);
- }
-
- /*
- * Now make the decision using the top-level tuple
- * fraction. First we have to convert an absolute
- * count (LIMIT) into fractional form.
- */
- if (tuple_fraction >= 1.0)
- tuple_fraction /= dNumGroups;
-
- if (compare_fractional_path_costs(&hashed_p, &sorted_p,
- tuple_fraction) < 0)
- {
- /* Hashed is cheaper, so use it */
- use_hashed_grouping = true;
- }
- }
- }
+ use_hashed_grouping =
+ choose_hashed_grouping(parse, tuple_fraction,
+ cheapest_path, sorted_path,
+ sort_pathkeys, group_pathkeys,
+ dNumGroups, &agg_counts);
}
/*
- * Select the best path and create a plan to execute it.
- *
- * If we are doing hashed grouping, we will always read all the input
- * tuples, so use the cheapest-total path. Otherwise, trust
- * query_planner's decision about which to use.
+ * Select the best path. If we are doing hashed grouping, we will
+ * always read all the input tuples, so use the cheapest-total
+ * path. Otherwise, trust query_planner's decision about which to use.
*/
- if (sorted_path && !use_hashed_grouping)
- {
- result_plan = create_plan(parse, sorted_path);
- current_pathkeys = sorted_path->pathkeys;
- }
+ if (use_hashed_grouping || !sorted_path)
+ best_path = cheapest_path;
else
- {
- result_plan = create_plan(parse, cheapest_path);
- current_pathkeys = cheapest_path->pathkeys;
- }
+ best_path = sorted_path;
/*
- * create_plan() returns a plan with just a "flat" tlist of
- * required Vars. Usually we need to insert the sub_tlist as the
- * tlist of the top plan node. However, we can skip that if we
- * determined that whatever query_planner chose to return will be
- * good enough.
+ * Check to see if it's possible to optimize MIN/MAX aggregates.
+ * If so, we will forget all the work we did so far to choose a
+ * "regular" path ... but we had to do it anyway to be able to
+ * tell which way is cheaper.
*/
- if (need_tlist_eval)
+ result_plan = optimize_minmax_aggregates(parse,
+ tlist,
+ best_path);
+ if (result_plan != NULL)
{
/*
- * If the top-level plan node is one that cannot do expression
- * evaluation, we must insert a Result node to project the
- * desired tlist.
+ * optimize_minmax_aggregates generated the full plan, with
+ * the right tlist, and it has no sort order.
*/
- if (!is_projection_capable_plan(result_plan))
- {
- result_plan = (Plan *) make_result(sub_tlist, NULL,
- result_plan);
- }
- else
- {
- /*
- * Otherwise, just replace the subplan's flat tlist with
- * the desired tlist.
- */
- result_plan->targetlist = sub_tlist;
- }
-
- /*
- * Also, account for the cost of evaluation of the sub_tlist.
- *
- * Up to now, we have only been dealing with "flat" tlists,
- * containing just Vars. So their evaluation cost is zero
- * according to the model used by cost_qual_eval() (or if you
- * prefer, the cost is factored into cpu_tuple_cost). Thus we
- * can avoid accounting for tlist cost throughout
- * query_planner() and subroutines. But now we've inserted a
- * tlist that might contain actual operators, sub-selects, etc
- * --- so we'd better account for its cost.
- *
- * Below this point, any tlist eval cost for added-on nodes
- * should be accounted for as we create those nodes.
- * Presently, of the node types we can add on, only Agg and
- * Group project new tlists (the rest just copy their input
- * tuples) --- so make_agg() and make_group() are responsible
- * for computing the added cost.
- */
- cost_qual_eval(&tlist_cost, sub_tlist);
- result_plan->startup_cost += tlist_cost.startup;
- result_plan->total_cost += tlist_cost.startup +
- tlist_cost.per_tuple * result_plan->plan_rows;
+ current_pathkeys = NIL;
}
else
{
/*
- * Since we're using query_planner's tlist and not the one
- * make_subplanTargetList calculated, we have to refigure any
- * grouping-column indexes make_subplanTargetList computed.
+ * Normal case --- create a plan according to query_planner's
+ * results.
*/
- locate_grouping_columns(parse, tlist, result_plan->targetlist,
- groupColIdx);
- }
-
- /*
- * Insert AGG or GROUP node if needed, plus an explicit sort step
- * if necessary.
- *
- * HAVING clause, if any, becomes qual of the Agg node
- */
- if (use_hashed_grouping)
- {
- /* Hashed aggregate plan --- no sort needed */
- result_plan = (Plan *) make_agg(parse,
- tlist,
- (List *) parse->havingQual,
- AGG_HASHED,
- numGroupCols,
- groupColIdx,
- numGroups,
- numAggs,
- result_plan);
- /* Hashed aggregation produces randomly-ordered results */
- current_pathkeys = NIL;
- }
- else if (parse->hasAggs)
- {
- /* Plain aggregate plan --- sort if needed */
- AggStrategy aggstrategy;
+ result_plan = create_plan(parse, best_path);
+ current_pathkeys = best_path->pathkeys;
- if (parse->groupClause)
+ /*
+ * create_plan() returns a plan with just a "flat" tlist of
+ * required Vars. Usually we need to insert the sub_tlist as the
+ * tlist of the top plan node. However, we can skip that if we
+ * determined that whatever query_planner chose to return will be
+ * good enough.
+ */
+ if (need_tlist_eval)
{
- if (!pathkeys_contained_in(group_pathkeys, current_pathkeys))
+ /*
+ * If the top-level plan node is one that cannot do expression
+ * evaluation, we must insert a Result node to project the
+ * desired tlist.
+ */
+ if (!is_projection_capable_plan(result_plan))
{
- result_plan = (Plan *)
- make_sort_from_groupcols(parse,
- parse->groupClause,
- groupColIdx,
- result_plan);
- current_pathkeys = group_pathkeys;
+ result_plan = (Plan *) make_result(sub_tlist, NULL,
+ result_plan);
+ }
+ else
+ {
+ /*
+ * Otherwise, just replace the subplan's flat tlist with
+ * the desired tlist.
+ */
+ result_plan->targetlist = sub_tlist;
}
- aggstrategy = AGG_SORTED;
/*
- * The AGG node will not change the sort ordering of its
- * groups, so current_pathkeys describes the result too.
+ * Also, account for the cost of evaluation of the sub_tlist.
+ *
+ * Up to now, we have only been dealing with "flat" tlists,
+ * containing just Vars. So their evaluation cost is zero
+ * according to the model used by cost_qual_eval() (or if you
+ * prefer, the cost is factored into cpu_tuple_cost). Thus we
+ * can avoid accounting for tlist cost throughout
+ * query_planner() and subroutines. But now we've inserted a
+ * tlist that might contain actual operators, sub-selects, etc
+ * --- so we'd better account for its cost.
+ *
+ * Below this point, any tlist eval cost for added-on nodes
+ * should be accounted for as we create those nodes.
+ * Presently, of the node types we can add on, only Agg and
+ * Group project new tlists (the rest just copy their input
+ * tuples) --- so make_agg() and make_group() are responsible
+ * for computing the added cost.
*/
+ cost_qual_eval(&tlist_cost, sub_tlist);
+ result_plan->startup_cost += tlist_cost.startup;
+ result_plan->total_cost += tlist_cost.startup +
+ tlist_cost.per_tuple * result_plan->plan_rows;
}
else
{
- aggstrategy = AGG_PLAIN;
- /* Result will be only one row anyway; no sort order */
- current_pathkeys = NIL;
+ /*
+ * Since we're using query_planner's tlist and not the one
+ * make_subplanTargetList calculated, we have to refigure any
+ * grouping-column indexes make_subplanTargetList computed.
+ */
+ locate_grouping_columns(parse, tlist, result_plan->targetlist,
+ groupColIdx);
}
- result_plan = (Plan *) make_agg(parse,
- tlist,
- (List *) parse->havingQual,
- aggstrategy,
- numGroupCols,
- groupColIdx,
- numGroups,
- numAggs,
- result_plan);
- }
- else
- {
/*
- * If there are no Aggs, we shouldn't have any HAVING qual
- * anymore
+ * Insert AGG or GROUP node if needed, plus an explicit sort step
+ * if necessary.
+ *
+ * HAVING clause, if any, becomes qual of the Agg or Group node.
*/
- Assert(parse->havingQual == NULL);
+ if (use_hashed_grouping)
+ {
+ /* Hashed aggregate plan --- no sort needed */
+ result_plan = (Plan *) make_agg(parse,
+ tlist,
+ (List *) parse->havingQual,
+ AGG_HASHED,
+ numGroupCols,
+ groupColIdx,
+ numGroups,
+ agg_counts.numAggs,
+ result_plan);
+ /* Hashed aggregation produces randomly-ordered results */
+ current_pathkeys = NIL;
+ }
+ else if (parse->hasAggs)
+ {
+ /* Plain aggregate plan --- sort if needed */
+ AggStrategy aggstrategy;
- /*
- * If we have a GROUP BY clause, insert a group node (plus the
- * appropriate sort node, if necessary).
- */
- if (parse->groupClause)
+ if (parse->groupClause)
+ {
+ if (!pathkeys_contained_in(group_pathkeys,
+ current_pathkeys))
+ {
+ result_plan = (Plan *)
+ make_sort_from_groupcols(parse,
+ parse->groupClause,
+ groupColIdx,
+ result_plan);
+ current_pathkeys = group_pathkeys;
+ }
+ aggstrategy = AGG_SORTED;
+
+ /*
+ * The AGG node will not change the sort ordering of its
+ * groups, so current_pathkeys describes the result too.
+ */
+ }
+ else
+ {
+ aggstrategy = AGG_PLAIN;
+ /* Result will be only one row anyway; no sort order */
+ current_pathkeys = NIL;
+ }
+
+ result_plan = (Plan *) make_agg(parse,
+ tlist,
+ (List *) parse->havingQual,
+ aggstrategy,
+ numGroupCols,
+ groupColIdx,
+ numGroups,
+ agg_counts.numAggs,
+ result_plan);
+ }
+ else if (parse->groupClause)
{
/*
+ * GROUP BY without aggregation, so insert a group node (plus
+ * the appropriate sort node, if necessary).
+ *
* Add an explicit sort if we couldn't make the path come
* out the way the GROUP node needs it.
*/
result_plan = (Plan *) make_group(parse,
tlist,
+ (List *) parse->havingQual,
numGroupCols,
groupColIdx,
dNumGroups,
result_plan);
/* The Group node won't change sort ordering */
}
- }
+ else if (parse->hasHavingQual)
+ {
+ /*
+ * No aggregates, and no GROUP BY, but we have a HAVING qual.
+ * This is a degenerate case in which we are supposed to emit
+ * either 0 or 1 row depending on whether HAVING succeeds.
+ * Furthermore, there cannot be any variables in either HAVING
+ * or the targetlist, so we actually do not need the FROM table
+ * at all! We can just throw away the plan-so-far and generate
+ * a Result node. This is a sufficiently unusual corner case
+ * that it's not worth contorting the structure of this routine
+ * to avoid having to generate the plan in the first place.
+ */
+ result_plan = (Plan *) make_result(tlist,
+ parse->havingQual,
+ NULL);
+ }
+ } /* end of non-minmax-aggregate case */
} /* end of if (setOperations) */
/*
* it's reasonable to assume the UNIQUE filter has effects
* comparable to GROUP BY.
*/
- if (!parse->groupClause && !parse->hasAggs)
+ if (!parse->groupClause && !parse->hasHavingQual && !parse->hasAggs)
{
List *distinctExprs;
return result_plan;
}
+/*
+ * choose_hashed_grouping - should we use hashed grouping?
+ */
+static bool
+choose_hashed_grouping(Query *parse, double tuple_fraction,
+ Path *cheapest_path, Path *sorted_path,
+ List *sort_pathkeys, List *group_pathkeys,
+ double dNumGroups, AggClauseCounts *agg_counts)
+{
+ int numGroupCols = list_length(parse->groupClause);
+ double cheapest_path_rows;
+ int cheapest_path_width;
+ Size hashentrysize;
+ List *current_pathkeys;
+ Path hashed_p;
+ Path sorted_p;
+
+ /*
+ * Check can't-do-it conditions, including whether the grouping operators
+ * are hashjoinable.
+ *
+ * Executor doesn't support hashed aggregation with DISTINCT aggregates.
+ * (Doing so would imply storing *all* the input values in the hash table,
+ * which seems like a certain loser.)
+ */
+ if (!enable_hashagg)
+ return false;
+ if (agg_counts->numDistinctAggs != 0)
+ return false;
+ if (!hash_safe_grouping(parse))
+ return false;
+
+ /*
+ * Don't do it if it doesn't look like the hashtable will fit into
+ * work_mem.
+ *
+ * Beware here of the possibility that cheapest_path->parent is NULL.
+ * This could happen if user does something silly like
+ * SELECT 'foo' GROUP BY 1;
+ */
+ if (cheapest_path->parent)
+ {
+ cheapest_path_rows = cheapest_path->parent->rows;
+ cheapest_path_width = cheapest_path->parent->width;
+ }
+ else
+ {
+ cheapest_path_rows = 1; /* assume non-set result */
+ cheapest_path_width = 100; /* arbitrary */
+ }
+
+ /* Estimate per-hash-entry space at tuple width... */
+ hashentrysize = cheapest_path_width;
+ /* plus space for pass-by-ref transition values... */
+ hashentrysize += agg_counts->transitionSpace;
+ /* plus the per-hash-entry overhead */
+ hashentrysize += hash_agg_entry_size(agg_counts->numAggs);
+
+ if (hashentrysize * dNumGroups > work_mem * 1024L)
+ return false;
+
+ /*
+ * See if the estimated cost is no more than doing it the other way.
+ * While avoiding the need for sorted input is usually a win, the fact
+ * that the output won't be sorted may be a loss; so we need to do an
+ * actual cost comparison.
+ *
+ * We need to consider
+ * cheapest_path + hashagg [+ final sort]
+ * versus either
+ * cheapest_path [+ sort] + group or agg [+ final sort]
+ * or
+ * presorted_path + group or agg [+ final sort]
+ * where brackets indicate a step that may not be needed. We assume
+ * query_planner() will have returned a presorted path only if it's a
+ * winner compared to cheapest_path for this purpose.
+ *
+ * These path variables are dummies that just hold cost fields; we don't
+ * make actual Paths for these steps.
+ */
+ cost_agg(&hashed_p, parse, AGG_HASHED, agg_counts->numAggs,
+ numGroupCols, dNumGroups,
+ cheapest_path->startup_cost, cheapest_path->total_cost,
+ cheapest_path_rows);
+ /* Result of hashed agg is always unsorted */
+ if (sort_pathkeys)
+ cost_sort(&hashed_p, parse, sort_pathkeys, hashed_p.total_cost,
+ dNumGroups, cheapest_path_width);
+
+ if (sorted_path)
+ {
+ sorted_p.startup_cost = sorted_path->startup_cost;
+ sorted_p.total_cost = sorted_path->total_cost;
+ current_pathkeys = sorted_path->pathkeys;
+ }
+ else
+ {
+ sorted_p.startup_cost = cheapest_path->startup_cost;
+ sorted_p.total_cost = cheapest_path->total_cost;
+ current_pathkeys = cheapest_path->pathkeys;
+ }
+ if (!pathkeys_contained_in(group_pathkeys,
+ current_pathkeys))
+ {
+ cost_sort(&sorted_p, parse, group_pathkeys, sorted_p.total_cost,
+ cheapest_path_rows, cheapest_path_width);
+ current_pathkeys = group_pathkeys;
+ }
+
+ if (parse->hasAggs)
+ cost_agg(&sorted_p, parse, AGG_SORTED, agg_counts->numAggs,
+ numGroupCols, dNumGroups,
+ sorted_p.startup_cost, sorted_p.total_cost,
+ cheapest_path_rows);
+ else
+ cost_group(&sorted_p, parse, numGroupCols, dNumGroups,
+ sorted_p.startup_cost, sorted_p.total_cost,
+ cheapest_path_rows);
+ /* The Agg or Group node will preserve ordering */
+ if (sort_pathkeys &&
+ !pathkeys_contained_in(sort_pathkeys, current_pathkeys))
+ cost_sort(&sorted_p, parse, sort_pathkeys, sorted_p.total_cost,
+ dNumGroups, cheapest_path_width);
+
+ /*
+ * Now make the decision using the top-level tuple fraction. First we
+ * have to convert an absolute count (LIMIT) into fractional form.
+ */
+ if (tuple_fraction >= 1.0)
+ tuple_fraction /= dNumGroups;
+
+ if (compare_fractional_path_costs(&hashed_p, &sorted_p,
+ tuple_fraction) < 0)
+ {
+ /* Hashed is cheaper, so use it */
+ return true;
+ }
+ return false;
+}
+
/*
* hash_safe_grouping - are grouping operators hashable?
*
Operator optup;
bool oprcanhash;
- optup = equality_oper(tle->resdom->restype, true);
+ optup = equality_oper(exprType((Node *) tle->expr), true);
if (!optup)
return false;
oprcanhash = ((Form_pg_operator) GETSTRUCT(optup))->oprcanhash;
* make_subplanTargetList
* Generate appropriate target list when grouping is required.
*
- * When grouping_planner inserts Aggregate or Group plan nodes above
- * the result of query_planner, we typically want to pass a different
+ * When grouping_planner inserts Aggregate, Group, or Result plan nodes
+ * above the result of query_planner, we typically want to pass a different
* target list to query_planner than the outer plan nodes should have.
* This routine generates the correct target list for the subplan.
*
* The initial target list passed from the parser already contains entries
* for all ORDER BY and GROUP BY expressions, but it will not have entries
* for variables used only in HAVING clauses; so we need to add those
- * variables to the subplan target list. Also, if we are doing either
- * grouping or aggregation, we flatten all expressions except GROUP BY items
- * into their component variables; the other expressions will be computed by
- * the inserted nodes rather than by the subplan. For example,
- * given a query like
+ * variables to the subplan target list. Also, we flatten all expressions
+ * except GROUP BY items into their component variables; the other expressions
+ * will be computed by the inserted nodes rather than by the subplan.
+ * For example, given a query like
* SELECT a+b,SUM(c+d) FROM table GROUP BY a+b;
* we want to pass this targetlist to the subplan:
* a,b,c,d,a+b
*groupColIdx = NULL;
/*
- * If we're not grouping or aggregating, nothing to do here;
+ * If we're not grouping or aggregating, there's nothing to do here;
* query_planner should receive the unmodified target list.
*/
- if (!parse->hasAggs && !parse->groupClause)
+ if (!parse->hasAggs && !parse->groupClause && !parse->hasHavingQual)
{
*need_tlist_eval = true;
return tlist;
foreach(gl, parse->groupClause)
{
- GroupClause *grpcl = (GroupClause *) lfirst(gl);
- Node *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
- TargetEntry *te = NULL;
- ListCell *sl;
+ GroupClause *grpcl = (GroupClause *) lfirst(gl);
+ Node *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
+ TargetEntry *te = NULL;
+ ListCell *sl;
/* Find or make a matching sub_tlist entry */
foreach(sl, sub_tlist)
}
if (!sl)
{
- te = makeTargetEntry(makeResdom(list_length(sub_tlist) + 1,
- exprType(groupexpr),
- exprTypmod(groupexpr),
- NULL,
- false),
- (Expr *) groupexpr);
+ te = makeTargetEntry((Expr *) groupexpr,
+ list_length(sub_tlist) + 1,
+ NULL,
+ false);
sub_tlist = lappend(sub_tlist, te);
*need_tlist_eval = true; /* it's not flat anymore */
}
/* and save its resno */
- grpColIdx[keyno++] = te->resdom->resno;
+ grpColIdx[keyno++] = te->resno;
}
}
foreach(gl, parse->groupClause)
{
- GroupClause *grpcl = (GroupClause *) lfirst(gl);
- Node *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
- TargetEntry *te = NULL;
- ListCell *sl;
+ GroupClause *grpcl = (GroupClause *) lfirst(gl);
+ Node *groupexpr = get_sortgroupclause_expr(grpcl, tlist);
+ TargetEntry *te = NULL;
+ ListCell *sl;
foreach(sl, sub_tlist)
{
if (!sl)
elog(ERROR, "failed to locate grouping columns");
- groupColIdx[keyno++] = te->resdom->resno;
+ groupColIdx[keyno++] = te->resno;
}
}
TargetEntry *orig_tle;
/* ignore resjunk columns in setop result */
- if (new_tle->resdom->resjunk)
+ if (new_tle->resjunk)
continue;
Assert(orig_tlist_item != NULL);
orig_tle = (TargetEntry *) lfirst(orig_tlist_item);
orig_tlist_item = lnext(orig_tlist_item);
- if (orig_tle->resdom->resjunk) /* should not happen */
+ if (orig_tle->resjunk) /* should not happen */
elog(ERROR, "resjunk output columns are not implemented");
- Assert(new_tle->resdom->resno == orig_tle->resdom->resno);
- Assert(new_tle->resdom->restype == orig_tle->resdom->restype);
- new_tle->resdom->ressortgroupref = orig_tle->resdom->ressortgroupref;
+ Assert(new_tle->resno == orig_tle->resno);
+ new_tle->ressortgroupref = orig_tle->ressortgroupref;
}
if (orig_tlist_item != NULL)
elog(ERROR, "resjunk output columns are not implemented");
projects / postgresql / blobdiff